Loading...
1/*
2 * linux/mm/page_alloc.c
3 *
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
6 *
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
15 */
16
17#include <linux/stddef.h>
18#include <linux/mm.h>
19#include <linux/swap.h>
20#include <linux/interrupt.h>
21#include <linux/pagemap.h>
22#include <linux/jiffies.h>
23#include <linux/bootmem.h>
24#include <linux/memblock.h>
25#include <linux/compiler.h>
26#include <linux/kernel.h>
27#include <linux/kmemcheck.h>
28#include <linux/kasan.h>
29#include <linux/module.h>
30#include <linux/suspend.h>
31#include <linux/pagevec.h>
32#include <linux/blkdev.h>
33#include <linux/slab.h>
34#include <linux/ratelimit.h>
35#include <linux/oom.h>
36#include <linux/notifier.h>
37#include <linux/topology.h>
38#include <linux/sysctl.h>
39#include <linux/cpu.h>
40#include <linux/cpuset.h>
41#include <linux/memory_hotplug.h>
42#include <linux/nodemask.h>
43#include <linux/vmalloc.h>
44#include <linux/vmstat.h>
45#include <linux/mempolicy.h>
46#include <linux/memremap.h>
47#include <linux/stop_machine.h>
48#include <linux/sort.h>
49#include <linux/pfn.h>
50#include <linux/backing-dev.h>
51#include <linux/fault-inject.h>
52#include <linux/page-isolation.h>
53#include <linux/page_ext.h>
54#include <linux/debugobjects.h>
55#include <linux/kmemleak.h>
56#include <linux/compaction.h>
57#include <trace/events/kmem.h>
58#include <linux/prefetch.h>
59#include <linux/mm_inline.h>
60#include <linux/migrate.h>
61#include <linux/page_ext.h>
62#include <linux/hugetlb.h>
63#include <linux/sched/rt.h>
64#include <linux/page_owner.h>
65#include <linux/kthread.h>
66
67#include <asm/sections.h>
68#include <asm/tlbflush.h>
69#include <asm/div64.h>
70#include "internal.h"
71
72/* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
73static DEFINE_MUTEX(pcp_batch_high_lock);
74#define MIN_PERCPU_PAGELIST_FRACTION (8)
75
76#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
77DEFINE_PER_CPU(int, numa_node);
78EXPORT_PER_CPU_SYMBOL(numa_node);
79#endif
80
81#ifdef CONFIG_HAVE_MEMORYLESS_NODES
82/*
83 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
84 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
85 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
86 * defined in <linux/topology.h>.
87 */
88DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
89EXPORT_PER_CPU_SYMBOL(_numa_mem_);
90int _node_numa_mem_[MAX_NUMNODES];
91#endif
92
93/*
94 * Array of node states.
95 */
96nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
97 [N_POSSIBLE] = NODE_MASK_ALL,
98 [N_ONLINE] = { { [0] = 1UL } },
99#ifndef CONFIG_NUMA
100 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
101#ifdef CONFIG_HIGHMEM
102 [N_HIGH_MEMORY] = { { [0] = 1UL } },
103#endif
104#ifdef CONFIG_MOVABLE_NODE
105 [N_MEMORY] = { { [0] = 1UL } },
106#endif
107 [N_CPU] = { { [0] = 1UL } },
108#endif /* NUMA */
109};
110EXPORT_SYMBOL(node_states);
111
112/* Protect totalram_pages and zone->managed_pages */
113static DEFINE_SPINLOCK(managed_page_count_lock);
114
115unsigned long totalram_pages __read_mostly;
116unsigned long totalreserve_pages __read_mostly;
117unsigned long totalcma_pages __read_mostly;
118
119int percpu_pagelist_fraction;
120gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
121
122/*
123 * A cached value of the page's pageblock's migratetype, used when the page is
124 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
125 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
126 * Also the migratetype set in the page does not necessarily match the pcplist
127 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
128 * other index - this ensures that it will be put on the correct CMA freelist.
129 */
130static inline int get_pcppage_migratetype(struct page *page)
131{
132 return page->index;
133}
134
135static inline void set_pcppage_migratetype(struct page *page, int migratetype)
136{
137 page->index = migratetype;
138}
139
140#ifdef CONFIG_PM_SLEEP
141/*
142 * The following functions are used by the suspend/hibernate code to temporarily
143 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
144 * while devices are suspended. To avoid races with the suspend/hibernate code,
145 * they should always be called with pm_mutex held (gfp_allowed_mask also should
146 * only be modified with pm_mutex held, unless the suspend/hibernate code is
147 * guaranteed not to run in parallel with that modification).
148 */
149
150static gfp_t saved_gfp_mask;
151
152void pm_restore_gfp_mask(void)
153{
154 WARN_ON(!mutex_is_locked(&pm_mutex));
155 if (saved_gfp_mask) {
156 gfp_allowed_mask = saved_gfp_mask;
157 saved_gfp_mask = 0;
158 }
159}
160
161void pm_restrict_gfp_mask(void)
162{
163 WARN_ON(!mutex_is_locked(&pm_mutex));
164 WARN_ON(saved_gfp_mask);
165 saved_gfp_mask = gfp_allowed_mask;
166 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
167}
168
169bool pm_suspended_storage(void)
170{
171 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
172 return false;
173 return true;
174}
175#endif /* CONFIG_PM_SLEEP */
176
177#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
178unsigned int pageblock_order __read_mostly;
179#endif
180
181static void __free_pages_ok(struct page *page, unsigned int order);
182
183/*
184 * results with 256, 32 in the lowmem_reserve sysctl:
185 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
186 * 1G machine -> (16M dma, 784M normal, 224M high)
187 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
188 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
189 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
190 *
191 * TBD: should special case ZONE_DMA32 machines here - in those we normally
192 * don't need any ZONE_NORMAL reservation
193 */
194int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
195#ifdef CONFIG_ZONE_DMA
196 256,
197#endif
198#ifdef CONFIG_ZONE_DMA32
199 256,
200#endif
201#ifdef CONFIG_HIGHMEM
202 32,
203#endif
204 32,
205};
206
207EXPORT_SYMBOL(totalram_pages);
208
209static char * const zone_names[MAX_NR_ZONES] = {
210#ifdef CONFIG_ZONE_DMA
211 "DMA",
212#endif
213#ifdef CONFIG_ZONE_DMA32
214 "DMA32",
215#endif
216 "Normal",
217#ifdef CONFIG_HIGHMEM
218 "HighMem",
219#endif
220 "Movable",
221#ifdef CONFIG_ZONE_DEVICE
222 "Device",
223#endif
224};
225
226char * const migratetype_names[MIGRATE_TYPES] = {
227 "Unmovable",
228 "Movable",
229 "Reclaimable",
230 "HighAtomic",
231#ifdef CONFIG_CMA
232 "CMA",
233#endif
234#ifdef CONFIG_MEMORY_ISOLATION
235 "Isolate",
236#endif
237};
238
239compound_page_dtor * const compound_page_dtors[] = {
240 NULL,
241 free_compound_page,
242#ifdef CONFIG_HUGETLB_PAGE
243 free_huge_page,
244#endif
245#ifdef CONFIG_TRANSPARENT_HUGEPAGE
246 free_transhuge_page,
247#endif
248};
249
250int min_free_kbytes = 1024;
251int user_min_free_kbytes = -1;
252int watermark_scale_factor = 10;
253
254static unsigned long __meminitdata nr_kernel_pages;
255static unsigned long __meminitdata nr_all_pages;
256static unsigned long __meminitdata dma_reserve;
257
258#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
259static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
260static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
261static unsigned long __initdata required_kernelcore;
262static unsigned long __initdata required_movablecore;
263static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
264static bool mirrored_kernelcore;
265
266/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
267int movable_zone;
268EXPORT_SYMBOL(movable_zone);
269#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
270
271#if MAX_NUMNODES > 1
272int nr_node_ids __read_mostly = MAX_NUMNODES;
273int nr_online_nodes __read_mostly = 1;
274EXPORT_SYMBOL(nr_node_ids);
275EXPORT_SYMBOL(nr_online_nodes);
276#endif
277
278int page_group_by_mobility_disabled __read_mostly;
279
280#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
281static inline void reset_deferred_meminit(pg_data_t *pgdat)
282{
283 pgdat->first_deferred_pfn = ULONG_MAX;
284}
285
286/* Returns true if the struct page for the pfn is uninitialised */
287static inline bool __meminit early_page_uninitialised(unsigned long pfn)
288{
289 if (pfn >= NODE_DATA(early_pfn_to_nid(pfn))->first_deferred_pfn)
290 return true;
291
292 return false;
293}
294
295static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
296{
297 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
298 return true;
299
300 return false;
301}
302
303/*
304 * Returns false when the remaining initialisation should be deferred until
305 * later in the boot cycle when it can be parallelised.
306 */
307static inline bool update_defer_init(pg_data_t *pgdat,
308 unsigned long pfn, unsigned long zone_end,
309 unsigned long *nr_initialised)
310{
311 unsigned long max_initialise;
312
313 /* Always populate low zones for address-contrained allocations */
314 if (zone_end < pgdat_end_pfn(pgdat))
315 return true;
316 /*
317 * Initialise at least 2G of a node but also take into account that
318 * two large system hashes that can take up 1GB for 0.25TB/node.
319 */
320 max_initialise = max(2UL << (30 - PAGE_SHIFT),
321 (pgdat->node_spanned_pages >> 8));
322
323 (*nr_initialised)++;
324 if ((*nr_initialised > max_initialise) &&
325 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
326 pgdat->first_deferred_pfn = pfn;
327 return false;
328 }
329
330 return true;
331}
332#else
333static inline void reset_deferred_meminit(pg_data_t *pgdat)
334{
335}
336
337static inline bool early_page_uninitialised(unsigned long pfn)
338{
339 return false;
340}
341
342static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
343{
344 return false;
345}
346
347static inline bool update_defer_init(pg_data_t *pgdat,
348 unsigned long pfn, unsigned long zone_end,
349 unsigned long *nr_initialised)
350{
351 return true;
352}
353#endif
354
355
356void set_pageblock_migratetype(struct page *page, int migratetype)
357{
358 if (unlikely(page_group_by_mobility_disabled &&
359 migratetype < MIGRATE_PCPTYPES))
360 migratetype = MIGRATE_UNMOVABLE;
361
362 set_pageblock_flags_group(page, (unsigned long)migratetype,
363 PB_migrate, PB_migrate_end);
364}
365
366#ifdef CONFIG_DEBUG_VM
367static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
368{
369 int ret = 0;
370 unsigned seq;
371 unsigned long pfn = page_to_pfn(page);
372 unsigned long sp, start_pfn;
373
374 do {
375 seq = zone_span_seqbegin(zone);
376 start_pfn = zone->zone_start_pfn;
377 sp = zone->spanned_pages;
378 if (!zone_spans_pfn(zone, pfn))
379 ret = 1;
380 } while (zone_span_seqretry(zone, seq));
381
382 if (ret)
383 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
384 pfn, zone_to_nid(zone), zone->name,
385 start_pfn, start_pfn + sp);
386
387 return ret;
388}
389
390static int page_is_consistent(struct zone *zone, struct page *page)
391{
392 if (!pfn_valid_within(page_to_pfn(page)))
393 return 0;
394 if (zone != page_zone(page))
395 return 0;
396
397 return 1;
398}
399/*
400 * Temporary debugging check for pages not lying within a given zone.
401 */
402static int bad_range(struct zone *zone, struct page *page)
403{
404 if (page_outside_zone_boundaries(zone, page))
405 return 1;
406 if (!page_is_consistent(zone, page))
407 return 1;
408
409 return 0;
410}
411#else
412static inline int bad_range(struct zone *zone, struct page *page)
413{
414 return 0;
415}
416#endif
417
418static void bad_page(struct page *page, const char *reason,
419 unsigned long bad_flags)
420{
421 static unsigned long resume;
422 static unsigned long nr_shown;
423 static unsigned long nr_unshown;
424
425 /* Don't complain about poisoned pages */
426 if (PageHWPoison(page)) {
427 page_mapcount_reset(page); /* remove PageBuddy */
428 return;
429 }
430
431 /*
432 * Allow a burst of 60 reports, then keep quiet for that minute;
433 * or allow a steady drip of one report per second.
434 */
435 if (nr_shown == 60) {
436 if (time_before(jiffies, resume)) {
437 nr_unshown++;
438 goto out;
439 }
440 if (nr_unshown) {
441 pr_alert(
442 "BUG: Bad page state: %lu messages suppressed\n",
443 nr_unshown);
444 nr_unshown = 0;
445 }
446 nr_shown = 0;
447 }
448 if (nr_shown++ == 0)
449 resume = jiffies + 60 * HZ;
450
451 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
452 current->comm, page_to_pfn(page));
453 __dump_page(page, reason);
454 bad_flags &= page->flags;
455 if (bad_flags)
456 pr_alert("bad because of flags: %#lx(%pGp)\n",
457 bad_flags, &bad_flags);
458 dump_page_owner(page);
459
460 print_modules();
461 dump_stack();
462out:
463 /* Leave bad fields for debug, except PageBuddy could make trouble */
464 page_mapcount_reset(page); /* remove PageBuddy */
465 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
466}
467
468/*
469 * Higher-order pages are called "compound pages". They are structured thusly:
470 *
471 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
472 *
473 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
474 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
475 *
476 * The first tail page's ->compound_dtor holds the offset in array of compound
477 * page destructors. See compound_page_dtors.
478 *
479 * The first tail page's ->compound_order holds the order of allocation.
480 * This usage means that zero-order pages may not be compound.
481 */
482
483void free_compound_page(struct page *page)
484{
485 __free_pages_ok(page, compound_order(page));
486}
487
488void prep_compound_page(struct page *page, unsigned int order)
489{
490 int i;
491 int nr_pages = 1 << order;
492
493 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
494 set_compound_order(page, order);
495 __SetPageHead(page);
496 for (i = 1; i < nr_pages; i++) {
497 struct page *p = page + i;
498 set_page_count(p, 0);
499 p->mapping = TAIL_MAPPING;
500 set_compound_head(p, page);
501 }
502 atomic_set(compound_mapcount_ptr(page), -1);
503}
504
505#ifdef CONFIG_DEBUG_PAGEALLOC
506unsigned int _debug_guardpage_minorder;
507bool _debug_pagealloc_enabled __read_mostly
508 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
509EXPORT_SYMBOL(_debug_pagealloc_enabled);
510bool _debug_guardpage_enabled __read_mostly;
511
512static int __init early_debug_pagealloc(char *buf)
513{
514 if (!buf)
515 return -EINVAL;
516
517 if (strcmp(buf, "on") == 0)
518 _debug_pagealloc_enabled = true;
519
520 if (strcmp(buf, "off") == 0)
521 _debug_pagealloc_enabled = false;
522
523 return 0;
524}
525early_param("debug_pagealloc", early_debug_pagealloc);
526
527static bool need_debug_guardpage(void)
528{
529 /* If we don't use debug_pagealloc, we don't need guard page */
530 if (!debug_pagealloc_enabled())
531 return false;
532
533 return true;
534}
535
536static void init_debug_guardpage(void)
537{
538 if (!debug_pagealloc_enabled())
539 return;
540
541 _debug_guardpage_enabled = true;
542}
543
544struct page_ext_operations debug_guardpage_ops = {
545 .need = need_debug_guardpage,
546 .init = init_debug_guardpage,
547};
548
549static int __init debug_guardpage_minorder_setup(char *buf)
550{
551 unsigned long res;
552
553 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
554 pr_err("Bad debug_guardpage_minorder value\n");
555 return 0;
556 }
557 _debug_guardpage_minorder = res;
558 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
559 return 0;
560}
561__setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
562
563static inline void set_page_guard(struct zone *zone, struct page *page,
564 unsigned int order, int migratetype)
565{
566 struct page_ext *page_ext;
567
568 if (!debug_guardpage_enabled())
569 return;
570
571 page_ext = lookup_page_ext(page);
572 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
573
574 INIT_LIST_HEAD(&page->lru);
575 set_page_private(page, order);
576 /* Guard pages are not available for any usage */
577 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
578}
579
580static inline void clear_page_guard(struct zone *zone, struct page *page,
581 unsigned int order, int migratetype)
582{
583 struct page_ext *page_ext;
584
585 if (!debug_guardpage_enabled())
586 return;
587
588 page_ext = lookup_page_ext(page);
589 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
590
591 set_page_private(page, 0);
592 if (!is_migrate_isolate(migratetype))
593 __mod_zone_freepage_state(zone, (1 << order), migratetype);
594}
595#else
596struct page_ext_operations debug_guardpage_ops = { NULL, };
597static inline void set_page_guard(struct zone *zone, struct page *page,
598 unsigned int order, int migratetype) {}
599static inline void clear_page_guard(struct zone *zone, struct page *page,
600 unsigned int order, int migratetype) {}
601#endif
602
603static inline void set_page_order(struct page *page, unsigned int order)
604{
605 set_page_private(page, order);
606 __SetPageBuddy(page);
607}
608
609static inline void rmv_page_order(struct page *page)
610{
611 __ClearPageBuddy(page);
612 set_page_private(page, 0);
613}
614
615/*
616 * This function checks whether a page is free && is the buddy
617 * we can do coalesce a page and its buddy if
618 * (a) the buddy is not in a hole &&
619 * (b) the buddy is in the buddy system &&
620 * (c) a page and its buddy have the same order &&
621 * (d) a page and its buddy are in the same zone.
622 *
623 * For recording whether a page is in the buddy system, we set ->_mapcount
624 * PAGE_BUDDY_MAPCOUNT_VALUE.
625 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
626 * serialized by zone->lock.
627 *
628 * For recording page's order, we use page_private(page).
629 */
630static inline int page_is_buddy(struct page *page, struct page *buddy,
631 unsigned int order)
632{
633 if (!pfn_valid_within(page_to_pfn(buddy)))
634 return 0;
635
636 if (page_is_guard(buddy) && page_order(buddy) == order) {
637 if (page_zone_id(page) != page_zone_id(buddy))
638 return 0;
639
640 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
641
642 return 1;
643 }
644
645 if (PageBuddy(buddy) && page_order(buddy) == order) {
646 /*
647 * zone check is done late to avoid uselessly
648 * calculating zone/node ids for pages that could
649 * never merge.
650 */
651 if (page_zone_id(page) != page_zone_id(buddy))
652 return 0;
653
654 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
655
656 return 1;
657 }
658 return 0;
659}
660
661/*
662 * Freeing function for a buddy system allocator.
663 *
664 * The concept of a buddy system is to maintain direct-mapped table
665 * (containing bit values) for memory blocks of various "orders".
666 * The bottom level table contains the map for the smallest allocatable
667 * units of memory (here, pages), and each level above it describes
668 * pairs of units from the levels below, hence, "buddies".
669 * At a high level, all that happens here is marking the table entry
670 * at the bottom level available, and propagating the changes upward
671 * as necessary, plus some accounting needed to play nicely with other
672 * parts of the VM system.
673 * At each level, we keep a list of pages, which are heads of continuous
674 * free pages of length of (1 << order) and marked with _mapcount
675 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
676 * field.
677 * So when we are allocating or freeing one, we can derive the state of the
678 * other. That is, if we allocate a small block, and both were
679 * free, the remainder of the region must be split into blocks.
680 * If a block is freed, and its buddy is also free, then this
681 * triggers coalescing into a block of larger size.
682 *
683 * -- nyc
684 */
685
686static inline void __free_one_page(struct page *page,
687 unsigned long pfn,
688 struct zone *zone, unsigned int order,
689 int migratetype)
690{
691 unsigned long page_idx;
692 unsigned long combined_idx;
693 unsigned long uninitialized_var(buddy_idx);
694 struct page *buddy;
695 unsigned int max_order;
696
697 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
698
699 VM_BUG_ON(!zone_is_initialized(zone));
700 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
701
702 VM_BUG_ON(migratetype == -1);
703 if (likely(!is_migrate_isolate(migratetype)))
704 __mod_zone_freepage_state(zone, 1 << order, migratetype);
705
706 page_idx = pfn & ((1 << MAX_ORDER) - 1);
707
708 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
709 VM_BUG_ON_PAGE(bad_range(zone, page), page);
710
711continue_merging:
712 while (order < max_order - 1) {
713 buddy_idx = __find_buddy_index(page_idx, order);
714 buddy = page + (buddy_idx - page_idx);
715 if (!page_is_buddy(page, buddy, order))
716 goto done_merging;
717 /*
718 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
719 * merge with it and move up one order.
720 */
721 if (page_is_guard(buddy)) {
722 clear_page_guard(zone, buddy, order, migratetype);
723 } else {
724 list_del(&buddy->lru);
725 zone->free_area[order].nr_free--;
726 rmv_page_order(buddy);
727 }
728 combined_idx = buddy_idx & page_idx;
729 page = page + (combined_idx - page_idx);
730 page_idx = combined_idx;
731 order++;
732 }
733 if (max_order < MAX_ORDER) {
734 /* If we are here, it means order is >= pageblock_order.
735 * We want to prevent merge between freepages on isolate
736 * pageblock and normal pageblock. Without this, pageblock
737 * isolation could cause incorrect freepage or CMA accounting.
738 *
739 * We don't want to hit this code for the more frequent
740 * low-order merging.
741 */
742 if (unlikely(has_isolate_pageblock(zone))) {
743 int buddy_mt;
744
745 buddy_idx = __find_buddy_index(page_idx, order);
746 buddy = page + (buddy_idx - page_idx);
747 buddy_mt = get_pageblock_migratetype(buddy);
748
749 if (migratetype != buddy_mt
750 && (is_migrate_isolate(migratetype) ||
751 is_migrate_isolate(buddy_mt)))
752 goto done_merging;
753 }
754 max_order++;
755 goto continue_merging;
756 }
757
758done_merging:
759 set_page_order(page, order);
760
761 /*
762 * If this is not the largest possible page, check if the buddy
763 * of the next-highest order is free. If it is, it's possible
764 * that pages are being freed that will coalesce soon. In case,
765 * that is happening, add the free page to the tail of the list
766 * so it's less likely to be used soon and more likely to be merged
767 * as a higher order page
768 */
769 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
770 struct page *higher_page, *higher_buddy;
771 combined_idx = buddy_idx & page_idx;
772 higher_page = page + (combined_idx - page_idx);
773 buddy_idx = __find_buddy_index(combined_idx, order + 1);
774 higher_buddy = higher_page + (buddy_idx - combined_idx);
775 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
776 list_add_tail(&page->lru,
777 &zone->free_area[order].free_list[migratetype]);
778 goto out;
779 }
780 }
781
782 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
783out:
784 zone->free_area[order].nr_free++;
785}
786
787static inline int free_pages_check(struct page *page)
788{
789 const char *bad_reason = NULL;
790 unsigned long bad_flags = 0;
791
792 if (unlikely(atomic_read(&page->_mapcount) != -1))
793 bad_reason = "nonzero mapcount";
794 if (unlikely(page->mapping != NULL))
795 bad_reason = "non-NULL mapping";
796 if (unlikely(page_ref_count(page) != 0))
797 bad_reason = "nonzero _count";
798 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
799 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
800 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
801 }
802#ifdef CONFIG_MEMCG
803 if (unlikely(page->mem_cgroup))
804 bad_reason = "page still charged to cgroup";
805#endif
806 if (unlikely(bad_reason)) {
807 bad_page(page, bad_reason, bad_flags);
808 return 1;
809 }
810 page_cpupid_reset_last(page);
811 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
812 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
813 return 0;
814}
815
816/*
817 * Frees a number of pages from the PCP lists
818 * Assumes all pages on list are in same zone, and of same order.
819 * count is the number of pages to free.
820 *
821 * If the zone was previously in an "all pages pinned" state then look to
822 * see if this freeing clears that state.
823 *
824 * And clear the zone's pages_scanned counter, to hold off the "all pages are
825 * pinned" detection logic.
826 */
827static void free_pcppages_bulk(struct zone *zone, int count,
828 struct per_cpu_pages *pcp)
829{
830 int migratetype = 0;
831 int batch_free = 0;
832 int to_free = count;
833 unsigned long nr_scanned;
834
835 spin_lock(&zone->lock);
836 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
837 if (nr_scanned)
838 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
839
840 while (to_free) {
841 struct page *page;
842 struct list_head *list;
843
844 /*
845 * Remove pages from lists in a round-robin fashion. A
846 * batch_free count is maintained that is incremented when an
847 * empty list is encountered. This is so more pages are freed
848 * off fuller lists instead of spinning excessively around empty
849 * lists
850 */
851 do {
852 batch_free++;
853 if (++migratetype == MIGRATE_PCPTYPES)
854 migratetype = 0;
855 list = &pcp->lists[migratetype];
856 } while (list_empty(list));
857
858 /* This is the only non-empty list. Free them all. */
859 if (batch_free == MIGRATE_PCPTYPES)
860 batch_free = to_free;
861
862 do {
863 int mt; /* migratetype of the to-be-freed page */
864
865 page = list_last_entry(list, struct page, lru);
866 /* must delete as __free_one_page list manipulates */
867 list_del(&page->lru);
868
869 mt = get_pcppage_migratetype(page);
870 /* MIGRATE_ISOLATE page should not go to pcplists */
871 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
872 /* Pageblock could have been isolated meanwhile */
873 if (unlikely(has_isolate_pageblock(zone)))
874 mt = get_pageblock_migratetype(page);
875
876 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
877 trace_mm_page_pcpu_drain(page, 0, mt);
878 } while (--to_free && --batch_free && !list_empty(list));
879 }
880 spin_unlock(&zone->lock);
881}
882
883static void free_one_page(struct zone *zone,
884 struct page *page, unsigned long pfn,
885 unsigned int order,
886 int migratetype)
887{
888 unsigned long nr_scanned;
889 spin_lock(&zone->lock);
890 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
891 if (nr_scanned)
892 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
893
894 if (unlikely(has_isolate_pageblock(zone) ||
895 is_migrate_isolate(migratetype))) {
896 migratetype = get_pfnblock_migratetype(page, pfn);
897 }
898 __free_one_page(page, pfn, zone, order, migratetype);
899 spin_unlock(&zone->lock);
900}
901
902static int free_tail_pages_check(struct page *head_page, struct page *page)
903{
904 int ret = 1;
905
906 /*
907 * We rely page->lru.next never has bit 0 set, unless the page
908 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
909 */
910 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
911
912 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
913 ret = 0;
914 goto out;
915 }
916 switch (page - head_page) {
917 case 1:
918 /* the first tail page: ->mapping is compound_mapcount() */
919 if (unlikely(compound_mapcount(page))) {
920 bad_page(page, "nonzero compound_mapcount", 0);
921 goto out;
922 }
923 break;
924 case 2:
925 /*
926 * the second tail page: ->mapping is
927 * page_deferred_list().next -- ignore value.
928 */
929 break;
930 default:
931 if (page->mapping != TAIL_MAPPING) {
932 bad_page(page, "corrupted mapping in tail page", 0);
933 goto out;
934 }
935 break;
936 }
937 if (unlikely(!PageTail(page))) {
938 bad_page(page, "PageTail not set", 0);
939 goto out;
940 }
941 if (unlikely(compound_head(page) != head_page)) {
942 bad_page(page, "compound_head not consistent", 0);
943 goto out;
944 }
945 ret = 0;
946out:
947 page->mapping = NULL;
948 clear_compound_head(page);
949 return ret;
950}
951
952static void __meminit __init_single_page(struct page *page, unsigned long pfn,
953 unsigned long zone, int nid)
954{
955 set_page_links(page, zone, nid, pfn);
956 init_page_count(page);
957 page_mapcount_reset(page);
958 page_cpupid_reset_last(page);
959
960 INIT_LIST_HEAD(&page->lru);
961#ifdef WANT_PAGE_VIRTUAL
962 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
963 if (!is_highmem_idx(zone))
964 set_page_address(page, __va(pfn << PAGE_SHIFT));
965#endif
966}
967
968static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
969 int nid)
970{
971 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
972}
973
974#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
975static void init_reserved_page(unsigned long pfn)
976{
977 pg_data_t *pgdat;
978 int nid, zid;
979
980 if (!early_page_uninitialised(pfn))
981 return;
982
983 nid = early_pfn_to_nid(pfn);
984 pgdat = NODE_DATA(nid);
985
986 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
987 struct zone *zone = &pgdat->node_zones[zid];
988
989 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
990 break;
991 }
992 __init_single_pfn(pfn, zid, nid);
993}
994#else
995static inline void init_reserved_page(unsigned long pfn)
996{
997}
998#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
999
1000/*
1001 * Initialised pages do not have PageReserved set. This function is
1002 * called for each range allocated by the bootmem allocator and
1003 * marks the pages PageReserved. The remaining valid pages are later
1004 * sent to the buddy page allocator.
1005 */
1006void __meminit reserve_bootmem_region(unsigned long start, unsigned long end)
1007{
1008 unsigned long start_pfn = PFN_DOWN(start);
1009 unsigned long end_pfn = PFN_UP(end);
1010
1011 for (; start_pfn < end_pfn; start_pfn++) {
1012 if (pfn_valid(start_pfn)) {
1013 struct page *page = pfn_to_page(start_pfn);
1014
1015 init_reserved_page(start_pfn);
1016
1017 /* Avoid false-positive PageTail() */
1018 INIT_LIST_HEAD(&page->lru);
1019
1020 SetPageReserved(page);
1021 }
1022 }
1023}
1024
1025static bool free_pages_prepare(struct page *page, unsigned int order)
1026{
1027 bool compound = PageCompound(page);
1028 int i, bad = 0;
1029
1030 VM_BUG_ON_PAGE(PageTail(page), page);
1031 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1032
1033 trace_mm_page_free(page, order);
1034 kmemcheck_free_shadow(page, order);
1035 kasan_free_pages(page, order);
1036
1037 if (PageAnon(page))
1038 page->mapping = NULL;
1039 bad += free_pages_check(page);
1040 for (i = 1; i < (1 << order); i++) {
1041 if (compound)
1042 bad += free_tail_pages_check(page, page + i);
1043 bad += free_pages_check(page + i);
1044 }
1045 if (bad)
1046 return false;
1047
1048 reset_page_owner(page, order);
1049
1050 if (!PageHighMem(page)) {
1051 debug_check_no_locks_freed(page_address(page),
1052 PAGE_SIZE << order);
1053 debug_check_no_obj_freed(page_address(page),
1054 PAGE_SIZE << order);
1055 }
1056 arch_free_page(page, order);
1057 kernel_poison_pages(page, 1 << order, 0);
1058 kernel_map_pages(page, 1 << order, 0);
1059
1060 return true;
1061}
1062
1063static void __free_pages_ok(struct page *page, unsigned int order)
1064{
1065 unsigned long flags;
1066 int migratetype;
1067 unsigned long pfn = page_to_pfn(page);
1068
1069 if (!free_pages_prepare(page, order))
1070 return;
1071
1072 migratetype = get_pfnblock_migratetype(page, pfn);
1073 local_irq_save(flags);
1074 __count_vm_events(PGFREE, 1 << order);
1075 free_one_page(page_zone(page), page, pfn, order, migratetype);
1076 local_irq_restore(flags);
1077}
1078
1079static void __init __free_pages_boot_core(struct page *page,
1080 unsigned long pfn, unsigned int order)
1081{
1082 unsigned int nr_pages = 1 << order;
1083 struct page *p = page;
1084 unsigned int loop;
1085
1086 prefetchw(p);
1087 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1088 prefetchw(p + 1);
1089 __ClearPageReserved(p);
1090 set_page_count(p, 0);
1091 }
1092 __ClearPageReserved(p);
1093 set_page_count(p, 0);
1094
1095 page_zone(page)->managed_pages += nr_pages;
1096 set_page_refcounted(page);
1097 __free_pages(page, order);
1098}
1099
1100#if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1101 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1102
1103static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1104
1105int __meminit early_pfn_to_nid(unsigned long pfn)
1106{
1107 static DEFINE_SPINLOCK(early_pfn_lock);
1108 int nid;
1109
1110 spin_lock(&early_pfn_lock);
1111 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1112 if (nid < 0)
1113 nid = 0;
1114 spin_unlock(&early_pfn_lock);
1115
1116 return nid;
1117}
1118#endif
1119
1120#ifdef CONFIG_NODES_SPAN_OTHER_NODES
1121static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1122 struct mminit_pfnnid_cache *state)
1123{
1124 int nid;
1125
1126 nid = __early_pfn_to_nid(pfn, state);
1127 if (nid >= 0 && nid != node)
1128 return false;
1129 return true;
1130}
1131
1132/* Only safe to use early in boot when initialisation is single-threaded */
1133static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1134{
1135 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1136}
1137
1138#else
1139
1140static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1141{
1142 return true;
1143}
1144static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1145 struct mminit_pfnnid_cache *state)
1146{
1147 return true;
1148}
1149#endif
1150
1151
1152void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1153 unsigned int order)
1154{
1155 if (early_page_uninitialised(pfn))
1156 return;
1157 return __free_pages_boot_core(page, pfn, order);
1158}
1159
1160/*
1161 * Check that the whole (or subset of) a pageblock given by the interval of
1162 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1163 * with the migration of free compaction scanner. The scanners then need to
1164 * use only pfn_valid_within() check for arches that allow holes within
1165 * pageblocks.
1166 *
1167 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1168 *
1169 * It's possible on some configurations to have a setup like node0 node1 node0
1170 * i.e. it's possible that all pages within a zones range of pages do not
1171 * belong to a single zone. We assume that a border between node0 and node1
1172 * can occur within a single pageblock, but not a node0 node1 node0
1173 * interleaving within a single pageblock. It is therefore sufficient to check
1174 * the first and last page of a pageblock and avoid checking each individual
1175 * page in a pageblock.
1176 */
1177struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1178 unsigned long end_pfn, struct zone *zone)
1179{
1180 struct page *start_page;
1181 struct page *end_page;
1182
1183 /* end_pfn is one past the range we are checking */
1184 end_pfn--;
1185
1186 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1187 return NULL;
1188
1189 start_page = pfn_to_page(start_pfn);
1190
1191 if (page_zone(start_page) != zone)
1192 return NULL;
1193
1194 end_page = pfn_to_page(end_pfn);
1195
1196 /* This gives a shorter code than deriving page_zone(end_page) */
1197 if (page_zone_id(start_page) != page_zone_id(end_page))
1198 return NULL;
1199
1200 return start_page;
1201}
1202
1203void set_zone_contiguous(struct zone *zone)
1204{
1205 unsigned long block_start_pfn = zone->zone_start_pfn;
1206 unsigned long block_end_pfn;
1207
1208 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1209 for (; block_start_pfn < zone_end_pfn(zone);
1210 block_start_pfn = block_end_pfn,
1211 block_end_pfn += pageblock_nr_pages) {
1212
1213 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1214
1215 if (!__pageblock_pfn_to_page(block_start_pfn,
1216 block_end_pfn, zone))
1217 return;
1218 }
1219
1220 /* We confirm that there is no hole */
1221 zone->contiguous = true;
1222}
1223
1224void clear_zone_contiguous(struct zone *zone)
1225{
1226 zone->contiguous = false;
1227}
1228
1229#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1230static void __init deferred_free_range(struct page *page,
1231 unsigned long pfn, int nr_pages)
1232{
1233 int i;
1234
1235 if (!page)
1236 return;
1237
1238 /* Free a large naturally-aligned chunk if possible */
1239 if (nr_pages == MAX_ORDER_NR_PAGES &&
1240 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1241 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1242 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1243 return;
1244 }
1245
1246 for (i = 0; i < nr_pages; i++, page++, pfn++)
1247 __free_pages_boot_core(page, pfn, 0);
1248}
1249
1250/* Completion tracking for deferred_init_memmap() threads */
1251static atomic_t pgdat_init_n_undone __initdata;
1252static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1253
1254static inline void __init pgdat_init_report_one_done(void)
1255{
1256 if (atomic_dec_and_test(&pgdat_init_n_undone))
1257 complete(&pgdat_init_all_done_comp);
1258}
1259
1260/* Initialise remaining memory on a node */
1261static int __init deferred_init_memmap(void *data)
1262{
1263 pg_data_t *pgdat = data;
1264 int nid = pgdat->node_id;
1265 struct mminit_pfnnid_cache nid_init_state = { };
1266 unsigned long start = jiffies;
1267 unsigned long nr_pages = 0;
1268 unsigned long walk_start, walk_end;
1269 int i, zid;
1270 struct zone *zone;
1271 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1272 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1273
1274 if (first_init_pfn == ULONG_MAX) {
1275 pgdat_init_report_one_done();
1276 return 0;
1277 }
1278
1279 /* Bind memory initialisation thread to a local node if possible */
1280 if (!cpumask_empty(cpumask))
1281 set_cpus_allowed_ptr(current, cpumask);
1282
1283 /* Sanity check boundaries */
1284 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1285 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1286 pgdat->first_deferred_pfn = ULONG_MAX;
1287
1288 /* Only the highest zone is deferred so find it */
1289 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1290 zone = pgdat->node_zones + zid;
1291 if (first_init_pfn < zone_end_pfn(zone))
1292 break;
1293 }
1294
1295 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1296 unsigned long pfn, end_pfn;
1297 struct page *page = NULL;
1298 struct page *free_base_page = NULL;
1299 unsigned long free_base_pfn = 0;
1300 int nr_to_free = 0;
1301
1302 end_pfn = min(walk_end, zone_end_pfn(zone));
1303 pfn = first_init_pfn;
1304 if (pfn < walk_start)
1305 pfn = walk_start;
1306 if (pfn < zone->zone_start_pfn)
1307 pfn = zone->zone_start_pfn;
1308
1309 for (; pfn < end_pfn; pfn++) {
1310 if (!pfn_valid_within(pfn))
1311 goto free_range;
1312
1313 /*
1314 * Ensure pfn_valid is checked every
1315 * MAX_ORDER_NR_PAGES for memory holes
1316 */
1317 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1318 if (!pfn_valid(pfn)) {
1319 page = NULL;
1320 goto free_range;
1321 }
1322 }
1323
1324 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1325 page = NULL;
1326 goto free_range;
1327 }
1328
1329 /* Minimise pfn page lookups and scheduler checks */
1330 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1331 page++;
1332 } else {
1333 nr_pages += nr_to_free;
1334 deferred_free_range(free_base_page,
1335 free_base_pfn, nr_to_free);
1336 free_base_page = NULL;
1337 free_base_pfn = nr_to_free = 0;
1338
1339 page = pfn_to_page(pfn);
1340 cond_resched();
1341 }
1342
1343 if (page->flags) {
1344 VM_BUG_ON(page_zone(page) != zone);
1345 goto free_range;
1346 }
1347
1348 __init_single_page(page, pfn, zid, nid);
1349 if (!free_base_page) {
1350 free_base_page = page;
1351 free_base_pfn = pfn;
1352 nr_to_free = 0;
1353 }
1354 nr_to_free++;
1355
1356 /* Where possible, batch up pages for a single free */
1357 continue;
1358free_range:
1359 /* Free the current block of pages to allocator */
1360 nr_pages += nr_to_free;
1361 deferred_free_range(free_base_page, free_base_pfn,
1362 nr_to_free);
1363 free_base_page = NULL;
1364 free_base_pfn = nr_to_free = 0;
1365 }
1366
1367 first_init_pfn = max(end_pfn, first_init_pfn);
1368 }
1369
1370 /* Sanity check that the next zone really is unpopulated */
1371 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1372
1373 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1374 jiffies_to_msecs(jiffies - start));
1375
1376 pgdat_init_report_one_done();
1377 return 0;
1378}
1379#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1380
1381void __init page_alloc_init_late(void)
1382{
1383 struct zone *zone;
1384
1385#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1386 int nid;
1387
1388 /* There will be num_node_state(N_MEMORY) threads */
1389 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1390 for_each_node_state(nid, N_MEMORY) {
1391 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1392 }
1393
1394 /* Block until all are initialised */
1395 wait_for_completion(&pgdat_init_all_done_comp);
1396
1397 /* Reinit limits that are based on free pages after the kernel is up */
1398 files_maxfiles_init();
1399#endif
1400
1401 for_each_populated_zone(zone)
1402 set_zone_contiguous(zone);
1403}
1404
1405#ifdef CONFIG_CMA
1406/* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1407void __init init_cma_reserved_pageblock(struct page *page)
1408{
1409 unsigned i = pageblock_nr_pages;
1410 struct page *p = page;
1411
1412 do {
1413 __ClearPageReserved(p);
1414 set_page_count(p, 0);
1415 } while (++p, --i);
1416
1417 set_pageblock_migratetype(page, MIGRATE_CMA);
1418
1419 if (pageblock_order >= MAX_ORDER) {
1420 i = pageblock_nr_pages;
1421 p = page;
1422 do {
1423 set_page_refcounted(p);
1424 __free_pages(p, MAX_ORDER - 1);
1425 p += MAX_ORDER_NR_PAGES;
1426 } while (i -= MAX_ORDER_NR_PAGES);
1427 } else {
1428 set_page_refcounted(page);
1429 __free_pages(page, pageblock_order);
1430 }
1431
1432 adjust_managed_page_count(page, pageblock_nr_pages);
1433}
1434#endif
1435
1436/*
1437 * The order of subdivision here is critical for the IO subsystem.
1438 * Please do not alter this order without good reasons and regression
1439 * testing. Specifically, as large blocks of memory are subdivided,
1440 * the order in which smaller blocks are delivered depends on the order
1441 * they're subdivided in this function. This is the primary factor
1442 * influencing the order in which pages are delivered to the IO
1443 * subsystem according to empirical testing, and this is also justified
1444 * by considering the behavior of a buddy system containing a single
1445 * large block of memory acted on by a series of small allocations.
1446 * This behavior is a critical factor in sglist merging's success.
1447 *
1448 * -- nyc
1449 */
1450static inline void expand(struct zone *zone, struct page *page,
1451 int low, int high, struct free_area *area,
1452 int migratetype)
1453{
1454 unsigned long size = 1 << high;
1455
1456 while (high > low) {
1457 area--;
1458 high--;
1459 size >>= 1;
1460 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1461
1462 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1463 debug_guardpage_enabled() &&
1464 high < debug_guardpage_minorder()) {
1465 /*
1466 * Mark as guard pages (or page), that will allow to
1467 * merge back to allocator when buddy will be freed.
1468 * Corresponding page table entries will not be touched,
1469 * pages will stay not present in virtual address space
1470 */
1471 set_page_guard(zone, &page[size], high, migratetype);
1472 continue;
1473 }
1474 list_add(&page[size].lru, &area->free_list[migratetype]);
1475 area->nr_free++;
1476 set_page_order(&page[size], high);
1477 }
1478}
1479
1480/*
1481 * This page is about to be returned from the page allocator
1482 */
1483static inline int check_new_page(struct page *page)
1484{
1485 const char *bad_reason = NULL;
1486 unsigned long bad_flags = 0;
1487
1488 if (unlikely(atomic_read(&page->_mapcount) != -1))
1489 bad_reason = "nonzero mapcount";
1490 if (unlikely(page->mapping != NULL))
1491 bad_reason = "non-NULL mapping";
1492 if (unlikely(page_ref_count(page) != 0))
1493 bad_reason = "nonzero _count";
1494 if (unlikely(page->flags & __PG_HWPOISON)) {
1495 bad_reason = "HWPoisoned (hardware-corrupted)";
1496 bad_flags = __PG_HWPOISON;
1497 }
1498 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1499 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1500 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1501 }
1502#ifdef CONFIG_MEMCG
1503 if (unlikely(page->mem_cgroup))
1504 bad_reason = "page still charged to cgroup";
1505#endif
1506 if (unlikely(bad_reason)) {
1507 bad_page(page, bad_reason, bad_flags);
1508 return 1;
1509 }
1510 return 0;
1511}
1512
1513static inline bool free_pages_prezeroed(bool poisoned)
1514{
1515 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1516 page_poisoning_enabled() && poisoned;
1517}
1518
1519static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1520 int alloc_flags)
1521{
1522 int i;
1523 bool poisoned = true;
1524
1525 for (i = 0; i < (1 << order); i++) {
1526 struct page *p = page + i;
1527 if (unlikely(check_new_page(p)))
1528 return 1;
1529 if (poisoned)
1530 poisoned &= page_is_poisoned(p);
1531 }
1532
1533 set_page_private(page, 0);
1534 set_page_refcounted(page);
1535
1536 arch_alloc_page(page, order);
1537 kernel_map_pages(page, 1 << order, 1);
1538 kernel_poison_pages(page, 1 << order, 1);
1539 kasan_alloc_pages(page, order);
1540
1541 if (!free_pages_prezeroed(poisoned) && (gfp_flags & __GFP_ZERO))
1542 for (i = 0; i < (1 << order); i++)
1543 clear_highpage(page + i);
1544
1545 if (order && (gfp_flags & __GFP_COMP))
1546 prep_compound_page(page, order);
1547
1548 set_page_owner(page, order, gfp_flags);
1549
1550 /*
1551 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1552 * allocate the page. The expectation is that the caller is taking
1553 * steps that will free more memory. The caller should avoid the page
1554 * being used for !PFMEMALLOC purposes.
1555 */
1556 if (alloc_flags & ALLOC_NO_WATERMARKS)
1557 set_page_pfmemalloc(page);
1558 else
1559 clear_page_pfmemalloc(page);
1560
1561 return 0;
1562}
1563
1564/*
1565 * Go through the free lists for the given migratetype and remove
1566 * the smallest available page from the freelists
1567 */
1568static inline
1569struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1570 int migratetype)
1571{
1572 unsigned int current_order;
1573 struct free_area *area;
1574 struct page *page;
1575
1576 /* Find a page of the appropriate size in the preferred list */
1577 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1578 area = &(zone->free_area[current_order]);
1579 page = list_first_entry_or_null(&area->free_list[migratetype],
1580 struct page, lru);
1581 if (!page)
1582 continue;
1583 list_del(&page->lru);
1584 rmv_page_order(page);
1585 area->nr_free--;
1586 expand(zone, page, order, current_order, area, migratetype);
1587 set_pcppage_migratetype(page, migratetype);
1588 return page;
1589 }
1590
1591 return NULL;
1592}
1593
1594
1595/*
1596 * This array describes the order lists are fallen back to when
1597 * the free lists for the desirable migrate type are depleted
1598 */
1599static int fallbacks[MIGRATE_TYPES][4] = {
1600 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1601 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1602 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1603#ifdef CONFIG_CMA
1604 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1605#endif
1606#ifdef CONFIG_MEMORY_ISOLATION
1607 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1608#endif
1609};
1610
1611#ifdef CONFIG_CMA
1612static struct page *__rmqueue_cma_fallback(struct zone *zone,
1613 unsigned int order)
1614{
1615 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1616}
1617#else
1618static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1619 unsigned int order) { return NULL; }
1620#endif
1621
1622/*
1623 * Move the free pages in a range to the free lists of the requested type.
1624 * Note that start_page and end_pages are not aligned on a pageblock
1625 * boundary. If alignment is required, use move_freepages_block()
1626 */
1627int move_freepages(struct zone *zone,
1628 struct page *start_page, struct page *end_page,
1629 int migratetype)
1630{
1631 struct page *page;
1632 unsigned int order;
1633 int pages_moved = 0;
1634
1635#ifndef CONFIG_HOLES_IN_ZONE
1636 /*
1637 * page_zone is not safe to call in this context when
1638 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1639 * anyway as we check zone boundaries in move_freepages_block().
1640 * Remove at a later date when no bug reports exist related to
1641 * grouping pages by mobility
1642 */
1643 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1644#endif
1645
1646 for (page = start_page; page <= end_page;) {
1647 /* Make sure we are not inadvertently changing nodes */
1648 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1649
1650 if (!pfn_valid_within(page_to_pfn(page))) {
1651 page++;
1652 continue;
1653 }
1654
1655 if (!PageBuddy(page)) {
1656 page++;
1657 continue;
1658 }
1659
1660 order = page_order(page);
1661 list_move(&page->lru,
1662 &zone->free_area[order].free_list[migratetype]);
1663 page += 1 << order;
1664 pages_moved += 1 << order;
1665 }
1666
1667 return pages_moved;
1668}
1669
1670int move_freepages_block(struct zone *zone, struct page *page,
1671 int migratetype)
1672{
1673 unsigned long start_pfn, end_pfn;
1674 struct page *start_page, *end_page;
1675
1676 start_pfn = page_to_pfn(page);
1677 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1678 start_page = pfn_to_page(start_pfn);
1679 end_page = start_page + pageblock_nr_pages - 1;
1680 end_pfn = start_pfn + pageblock_nr_pages - 1;
1681
1682 /* Do not cross zone boundaries */
1683 if (!zone_spans_pfn(zone, start_pfn))
1684 start_page = page;
1685 if (!zone_spans_pfn(zone, end_pfn))
1686 return 0;
1687
1688 return move_freepages(zone, start_page, end_page, migratetype);
1689}
1690
1691static void change_pageblock_range(struct page *pageblock_page,
1692 int start_order, int migratetype)
1693{
1694 int nr_pageblocks = 1 << (start_order - pageblock_order);
1695
1696 while (nr_pageblocks--) {
1697 set_pageblock_migratetype(pageblock_page, migratetype);
1698 pageblock_page += pageblock_nr_pages;
1699 }
1700}
1701
1702/*
1703 * When we are falling back to another migratetype during allocation, try to
1704 * steal extra free pages from the same pageblocks to satisfy further
1705 * allocations, instead of polluting multiple pageblocks.
1706 *
1707 * If we are stealing a relatively large buddy page, it is likely there will
1708 * be more free pages in the pageblock, so try to steal them all. For
1709 * reclaimable and unmovable allocations, we steal regardless of page size,
1710 * as fragmentation caused by those allocations polluting movable pageblocks
1711 * is worse than movable allocations stealing from unmovable and reclaimable
1712 * pageblocks.
1713 */
1714static bool can_steal_fallback(unsigned int order, int start_mt)
1715{
1716 /*
1717 * Leaving this order check is intended, although there is
1718 * relaxed order check in next check. The reason is that
1719 * we can actually steal whole pageblock if this condition met,
1720 * but, below check doesn't guarantee it and that is just heuristic
1721 * so could be changed anytime.
1722 */
1723 if (order >= pageblock_order)
1724 return true;
1725
1726 if (order >= pageblock_order / 2 ||
1727 start_mt == MIGRATE_RECLAIMABLE ||
1728 start_mt == MIGRATE_UNMOVABLE ||
1729 page_group_by_mobility_disabled)
1730 return true;
1731
1732 return false;
1733}
1734
1735/*
1736 * This function implements actual steal behaviour. If order is large enough,
1737 * we can steal whole pageblock. If not, we first move freepages in this
1738 * pageblock and check whether half of pages are moved or not. If half of
1739 * pages are moved, we can change migratetype of pageblock and permanently
1740 * use it's pages as requested migratetype in the future.
1741 */
1742static void steal_suitable_fallback(struct zone *zone, struct page *page,
1743 int start_type)
1744{
1745 unsigned int current_order = page_order(page);
1746 int pages;
1747
1748 /* Take ownership for orders >= pageblock_order */
1749 if (current_order >= pageblock_order) {
1750 change_pageblock_range(page, current_order, start_type);
1751 return;
1752 }
1753
1754 pages = move_freepages_block(zone, page, start_type);
1755
1756 /* Claim the whole block if over half of it is free */
1757 if (pages >= (1 << (pageblock_order-1)) ||
1758 page_group_by_mobility_disabled)
1759 set_pageblock_migratetype(page, start_type);
1760}
1761
1762/*
1763 * Check whether there is a suitable fallback freepage with requested order.
1764 * If only_stealable is true, this function returns fallback_mt only if
1765 * we can steal other freepages all together. This would help to reduce
1766 * fragmentation due to mixed migratetype pages in one pageblock.
1767 */
1768int find_suitable_fallback(struct free_area *area, unsigned int order,
1769 int migratetype, bool only_stealable, bool *can_steal)
1770{
1771 int i;
1772 int fallback_mt;
1773
1774 if (area->nr_free == 0)
1775 return -1;
1776
1777 *can_steal = false;
1778 for (i = 0;; i++) {
1779 fallback_mt = fallbacks[migratetype][i];
1780 if (fallback_mt == MIGRATE_TYPES)
1781 break;
1782
1783 if (list_empty(&area->free_list[fallback_mt]))
1784 continue;
1785
1786 if (can_steal_fallback(order, migratetype))
1787 *can_steal = true;
1788
1789 if (!only_stealable)
1790 return fallback_mt;
1791
1792 if (*can_steal)
1793 return fallback_mt;
1794 }
1795
1796 return -1;
1797}
1798
1799/*
1800 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1801 * there are no empty page blocks that contain a page with a suitable order
1802 */
1803static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1804 unsigned int alloc_order)
1805{
1806 int mt;
1807 unsigned long max_managed, flags;
1808
1809 /*
1810 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1811 * Check is race-prone but harmless.
1812 */
1813 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1814 if (zone->nr_reserved_highatomic >= max_managed)
1815 return;
1816
1817 spin_lock_irqsave(&zone->lock, flags);
1818
1819 /* Recheck the nr_reserved_highatomic limit under the lock */
1820 if (zone->nr_reserved_highatomic >= max_managed)
1821 goto out_unlock;
1822
1823 /* Yoink! */
1824 mt = get_pageblock_migratetype(page);
1825 if (mt != MIGRATE_HIGHATOMIC &&
1826 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1827 zone->nr_reserved_highatomic += pageblock_nr_pages;
1828 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1829 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1830 }
1831
1832out_unlock:
1833 spin_unlock_irqrestore(&zone->lock, flags);
1834}
1835
1836/*
1837 * Used when an allocation is about to fail under memory pressure. This
1838 * potentially hurts the reliability of high-order allocations when under
1839 * intense memory pressure but failed atomic allocations should be easier
1840 * to recover from than an OOM.
1841 */
1842static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1843{
1844 struct zonelist *zonelist = ac->zonelist;
1845 unsigned long flags;
1846 struct zoneref *z;
1847 struct zone *zone;
1848 struct page *page;
1849 int order;
1850
1851 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1852 ac->nodemask) {
1853 /* Preserve at least one pageblock */
1854 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1855 continue;
1856
1857 spin_lock_irqsave(&zone->lock, flags);
1858 for (order = 0; order < MAX_ORDER; order++) {
1859 struct free_area *area = &(zone->free_area[order]);
1860
1861 page = list_first_entry_or_null(
1862 &area->free_list[MIGRATE_HIGHATOMIC],
1863 struct page, lru);
1864 if (!page)
1865 continue;
1866
1867 /*
1868 * It should never happen but changes to locking could
1869 * inadvertently allow a per-cpu drain to add pages
1870 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1871 * and watch for underflows.
1872 */
1873 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
1874 zone->nr_reserved_highatomic);
1875
1876 /*
1877 * Convert to ac->migratetype and avoid the normal
1878 * pageblock stealing heuristics. Minimally, the caller
1879 * is doing the work and needs the pages. More
1880 * importantly, if the block was always converted to
1881 * MIGRATE_UNMOVABLE or another type then the number
1882 * of pageblocks that cannot be completely freed
1883 * may increase.
1884 */
1885 set_pageblock_migratetype(page, ac->migratetype);
1886 move_freepages_block(zone, page, ac->migratetype);
1887 spin_unlock_irqrestore(&zone->lock, flags);
1888 return;
1889 }
1890 spin_unlock_irqrestore(&zone->lock, flags);
1891 }
1892}
1893
1894/* Remove an element from the buddy allocator from the fallback list */
1895static inline struct page *
1896__rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1897{
1898 struct free_area *area;
1899 unsigned int current_order;
1900 struct page *page;
1901 int fallback_mt;
1902 bool can_steal;
1903
1904 /* Find the largest possible block of pages in the other list */
1905 for (current_order = MAX_ORDER-1;
1906 current_order >= order && current_order <= MAX_ORDER-1;
1907 --current_order) {
1908 area = &(zone->free_area[current_order]);
1909 fallback_mt = find_suitable_fallback(area, current_order,
1910 start_migratetype, false, &can_steal);
1911 if (fallback_mt == -1)
1912 continue;
1913
1914 page = list_first_entry(&area->free_list[fallback_mt],
1915 struct page, lru);
1916 if (can_steal)
1917 steal_suitable_fallback(zone, page, start_migratetype);
1918
1919 /* Remove the page from the freelists */
1920 area->nr_free--;
1921 list_del(&page->lru);
1922 rmv_page_order(page);
1923
1924 expand(zone, page, order, current_order, area,
1925 start_migratetype);
1926 /*
1927 * The pcppage_migratetype may differ from pageblock's
1928 * migratetype depending on the decisions in
1929 * find_suitable_fallback(). This is OK as long as it does not
1930 * differ for MIGRATE_CMA pageblocks. Those can be used as
1931 * fallback only via special __rmqueue_cma_fallback() function
1932 */
1933 set_pcppage_migratetype(page, start_migratetype);
1934
1935 trace_mm_page_alloc_extfrag(page, order, current_order,
1936 start_migratetype, fallback_mt);
1937
1938 return page;
1939 }
1940
1941 return NULL;
1942}
1943
1944/*
1945 * Do the hard work of removing an element from the buddy allocator.
1946 * Call me with the zone->lock already held.
1947 */
1948static struct page *__rmqueue(struct zone *zone, unsigned int order,
1949 int migratetype)
1950{
1951 struct page *page;
1952
1953 page = __rmqueue_smallest(zone, order, migratetype);
1954 if (unlikely(!page)) {
1955 if (migratetype == MIGRATE_MOVABLE)
1956 page = __rmqueue_cma_fallback(zone, order);
1957
1958 if (!page)
1959 page = __rmqueue_fallback(zone, order, migratetype);
1960 }
1961
1962 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1963 return page;
1964}
1965
1966/*
1967 * Obtain a specified number of elements from the buddy allocator, all under
1968 * a single hold of the lock, for efficiency. Add them to the supplied list.
1969 * Returns the number of new pages which were placed at *list.
1970 */
1971static int rmqueue_bulk(struct zone *zone, unsigned int order,
1972 unsigned long count, struct list_head *list,
1973 int migratetype, bool cold)
1974{
1975 int i;
1976
1977 spin_lock(&zone->lock);
1978 for (i = 0; i < count; ++i) {
1979 struct page *page = __rmqueue(zone, order, migratetype);
1980 if (unlikely(page == NULL))
1981 break;
1982
1983 /*
1984 * Split buddy pages returned by expand() are received here
1985 * in physical page order. The page is added to the callers and
1986 * list and the list head then moves forward. From the callers
1987 * perspective, the linked list is ordered by page number in
1988 * some conditions. This is useful for IO devices that can
1989 * merge IO requests if the physical pages are ordered
1990 * properly.
1991 */
1992 if (likely(!cold))
1993 list_add(&page->lru, list);
1994 else
1995 list_add_tail(&page->lru, list);
1996 list = &page->lru;
1997 if (is_migrate_cma(get_pcppage_migratetype(page)))
1998 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1999 -(1 << order));
2000 }
2001 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2002 spin_unlock(&zone->lock);
2003 return i;
2004}
2005
2006#ifdef CONFIG_NUMA
2007/*
2008 * Called from the vmstat counter updater to drain pagesets of this
2009 * currently executing processor on remote nodes after they have
2010 * expired.
2011 *
2012 * Note that this function must be called with the thread pinned to
2013 * a single processor.
2014 */
2015void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2016{
2017 unsigned long flags;
2018 int to_drain, batch;
2019
2020 local_irq_save(flags);
2021 batch = READ_ONCE(pcp->batch);
2022 to_drain = min(pcp->count, batch);
2023 if (to_drain > 0) {
2024 free_pcppages_bulk(zone, to_drain, pcp);
2025 pcp->count -= to_drain;
2026 }
2027 local_irq_restore(flags);
2028}
2029#endif
2030
2031/*
2032 * Drain pcplists of the indicated processor and zone.
2033 *
2034 * The processor must either be the current processor and the
2035 * thread pinned to the current processor or a processor that
2036 * is not online.
2037 */
2038static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2039{
2040 unsigned long flags;
2041 struct per_cpu_pageset *pset;
2042 struct per_cpu_pages *pcp;
2043
2044 local_irq_save(flags);
2045 pset = per_cpu_ptr(zone->pageset, cpu);
2046
2047 pcp = &pset->pcp;
2048 if (pcp->count) {
2049 free_pcppages_bulk(zone, pcp->count, pcp);
2050 pcp->count = 0;
2051 }
2052 local_irq_restore(flags);
2053}
2054
2055/*
2056 * Drain pcplists of all zones on the indicated processor.
2057 *
2058 * The processor must either be the current processor and the
2059 * thread pinned to the current processor or a processor that
2060 * is not online.
2061 */
2062static void drain_pages(unsigned int cpu)
2063{
2064 struct zone *zone;
2065
2066 for_each_populated_zone(zone) {
2067 drain_pages_zone(cpu, zone);
2068 }
2069}
2070
2071/*
2072 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2073 *
2074 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2075 * the single zone's pages.
2076 */
2077void drain_local_pages(struct zone *zone)
2078{
2079 int cpu = smp_processor_id();
2080
2081 if (zone)
2082 drain_pages_zone(cpu, zone);
2083 else
2084 drain_pages(cpu);
2085}
2086
2087/*
2088 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2089 *
2090 * When zone parameter is non-NULL, spill just the single zone's pages.
2091 *
2092 * Note that this code is protected against sending an IPI to an offline
2093 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2094 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2095 * nothing keeps CPUs from showing up after we populated the cpumask and
2096 * before the call to on_each_cpu_mask().
2097 */
2098void drain_all_pages(struct zone *zone)
2099{
2100 int cpu;
2101
2102 /*
2103 * Allocate in the BSS so we wont require allocation in
2104 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2105 */
2106 static cpumask_t cpus_with_pcps;
2107
2108 /*
2109 * We don't care about racing with CPU hotplug event
2110 * as offline notification will cause the notified
2111 * cpu to drain that CPU pcps and on_each_cpu_mask
2112 * disables preemption as part of its processing
2113 */
2114 for_each_online_cpu(cpu) {
2115 struct per_cpu_pageset *pcp;
2116 struct zone *z;
2117 bool has_pcps = false;
2118
2119 if (zone) {
2120 pcp = per_cpu_ptr(zone->pageset, cpu);
2121 if (pcp->pcp.count)
2122 has_pcps = true;
2123 } else {
2124 for_each_populated_zone(z) {
2125 pcp = per_cpu_ptr(z->pageset, cpu);
2126 if (pcp->pcp.count) {
2127 has_pcps = true;
2128 break;
2129 }
2130 }
2131 }
2132
2133 if (has_pcps)
2134 cpumask_set_cpu(cpu, &cpus_with_pcps);
2135 else
2136 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2137 }
2138 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2139 zone, 1);
2140}
2141
2142#ifdef CONFIG_HIBERNATION
2143
2144void mark_free_pages(struct zone *zone)
2145{
2146 unsigned long pfn, max_zone_pfn;
2147 unsigned long flags;
2148 unsigned int order, t;
2149 struct page *page;
2150
2151 if (zone_is_empty(zone))
2152 return;
2153
2154 spin_lock_irqsave(&zone->lock, flags);
2155
2156 max_zone_pfn = zone_end_pfn(zone);
2157 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2158 if (pfn_valid(pfn)) {
2159 page = pfn_to_page(pfn);
2160 if (!swsusp_page_is_forbidden(page))
2161 swsusp_unset_page_free(page);
2162 }
2163
2164 for_each_migratetype_order(order, t) {
2165 list_for_each_entry(page,
2166 &zone->free_area[order].free_list[t], lru) {
2167 unsigned long i;
2168
2169 pfn = page_to_pfn(page);
2170 for (i = 0; i < (1UL << order); i++)
2171 swsusp_set_page_free(pfn_to_page(pfn + i));
2172 }
2173 }
2174 spin_unlock_irqrestore(&zone->lock, flags);
2175}
2176#endif /* CONFIG_PM */
2177
2178/*
2179 * Free a 0-order page
2180 * cold == true ? free a cold page : free a hot page
2181 */
2182void free_hot_cold_page(struct page *page, bool cold)
2183{
2184 struct zone *zone = page_zone(page);
2185 struct per_cpu_pages *pcp;
2186 unsigned long flags;
2187 unsigned long pfn = page_to_pfn(page);
2188 int migratetype;
2189
2190 if (!free_pages_prepare(page, 0))
2191 return;
2192
2193 migratetype = get_pfnblock_migratetype(page, pfn);
2194 set_pcppage_migratetype(page, migratetype);
2195 local_irq_save(flags);
2196 __count_vm_event(PGFREE);
2197
2198 /*
2199 * We only track unmovable, reclaimable and movable on pcp lists.
2200 * Free ISOLATE pages back to the allocator because they are being
2201 * offlined but treat RESERVE as movable pages so we can get those
2202 * areas back if necessary. Otherwise, we may have to free
2203 * excessively into the page allocator
2204 */
2205 if (migratetype >= MIGRATE_PCPTYPES) {
2206 if (unlikely(is_migrate_isolate(migratetype))) {
2207 free_one_page(zone, page, pfn, 0, migratetype);
2208 goto out;
2209 }
2210 migratetype = MIGRATE_MOVABLE;
2211 }
2212
2213 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2214 if (!cold)
2215 list_add(&page->lru, &pcp->lists[migratetype]);
2216 else
2217 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2218 pcp->count++;
2219 if (pcp->count >= pcp->high) {
2220 unsigned long batch = READ_ONCE(pcp->batch);
2221 free_pcppages_bulk(zone, batch, pcp);
2222 pcp->count -= batch;
2223 }
2224
2225out:
2226 local_irq_restore(flags);
2227}
2228
2229/*
2230 * Free a list of 0-order pages
2231 */
2232void free_hot_cold_page_list(struct list_head *list, bool cold)
2233{
2234 struct page *page, *next;
2235
2236 list_for_each_entry_safe(page, next, list, lru) {
2237 trace_mm_page_free_batched(page, cold);
2238 free_hot_cold_page(page, cold);
2239 }
2240}
2241
2242/*
2243 * split_page takes a non-compound higher-order page, and splits it into
2244 * n (1<<order) sub-pages: page[0..n]
2245 * Each sub-page must be freed individually.
2246 *
2247 * Note: this is probably too low level an operation for use in drivers.
2248 * Please consult with lkml before using this in your driver.
2249 */
2250void split_page(struct page *page, unsigned int order)
2251{
2252 int i;
2253 gfp_t gfp_mask;
2254
2255 VM_BUG_ON_PAGE(PageCompound(page), page);
2256 VM_BUG_ON_PAGE(!page_count(page), page);
2257
2258#ifdef CONFIG_KMEMCHECK
2259 /*
2260 * Split shadow pages too, because free(page[0]) would
2261 * otherwise free the whole shadow.
2262 */
2263 if (kmemcheck_page_is_tracked(page))
2264 split_page(virt_to_page(page[0].shadow), order);
2265#endif
2266
2267 gfp_mask = get_page_owner_gfp(page);
2268 set_page_owner(page, 0, gfp_mask);
2269 for (i = 1; i < (1 << order); i++) {
2270 set_page_refcounted(page + i);
2271 set_page_owner(page + i, 0, gfp_mask);
2272 }
2273}
2274EXPORT_SYMBOL_GPL(split_page);
2275
2276int __isolate_free_page(struct page *page, unsigned int order)
2277{
2278 unsigned long watermark;
2279 struct zone *zone;
2280 int mt;
2281
2282 BUG_ON(!PageBuddy(page));
2283
2284 zone = page_zone(page);
2285 mt = get_pageblock_migratetype(page);
2286
2287 if (!is_migrate_isolate(mt)) {
2288 /* Obey watermarks as if the page was being allocated */
2289 watermark = low_wmark_pages(zone) + (1 << order);
2290 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2291 return 0;
2292
2293 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2294 }
2295
2296 /* Remove page from free list */
2297 list_del(&page->lru);
2298 zone->free_area[order].nr_free--;
2299 rmv_page_order(page);
2300
2301 set_page_owner(page, order, __GFP_MOVABLE);
2302
2303 /* Set the pageblock if the isolated page is at least a pageblock */
2304 if (order >= pageblock_order - 1) {
2305 struct page *endpage = page + (1 << order) - 1;
2306 for (; page < endpage; page += pageblock_nr_pages) {
2307 int mt = get_pageblock_migratetype(page);
2308 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2309 set_pageblock_migratetype(page,
2310 MIGRATE_MOVABLE);
2311 }
2312 }
2313
2314
2315 return 1UL << order;
2316}
2317
2318/*
2319 * Similar to split_page except the page is already free. As this is only
2320 * being used for migration, the migratetype of the block also changes.
2321 * As this is called with interrupts disabled, the caller is responsible
2322 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2323 * are enabled.
2324 *
2325 * Note: this is probably too low level an operation for use in drivers.
2326 * Please consult with lkml before using this in your driver.
2327 */
2328int split_free_page(struct page *page)
2329{
2330 unsigned int order;
2331 int nr_pages;
2332
2333 order = page_order(page);
2334
2335 nr_pages = __isolate_free_page(page, order);
2336 if (!nr_pages)
2337 return 0;
2338
2339 /* Split into individual pages */
2340 set_page_refcounted(page);
2341 split_page(page, order);
2342 return nr_pages;
2343}
2344
2345/*
2346 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2347 */
2348static inline
2349struct page *buffered_rmqueue(struct zone *preferred_zone,
2350 struct zone *zone, unsigned int order,
2351 gfp_t gfp_flags, int alloc_flags, int migratetype)
2352{
2353 unsigned long flags;
2354 struct page *page;
2355 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2356
2357 if (likely(order == 0)) {
2358 struct per_cpu_pages *pcp;
2359 struct list_head *list;
2360
2361 local_irq_save(flags);
2362 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2363 list = &pcp->lists[migratetype];
2364 if (list_empty(list)) {
2365 pcp->count += rmqueue_bulk(zone, 0,
2366 pcp->batch, list,
2367 migratetype, cold);
2368 if (unlikely(list_empty(list)))
2369 goto failed;
2370 }
2371
2372 if (cold)
2373 page = list_last_entry(list, struct page, lru);
2374 else
2375 page = list_first_entry(list, struct page, lru);
2376
2377 list_del(&page->lru);
2378 pcp->count--;
2379 } else {
2380 /*
2381 * We most definitely don't want callers attempting to
2382 * allocate greater than order-1 page units with __GFP_NOFAIL.
2383 */
2384 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2385 spin_lock_irqsave(&zone->lock, flags);
2386
2387 page = NULL;
2388 if (alloc_flags & ALLOC_HARDER) {
2389 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2390 if (page)
2391 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2392 }
2393 if (!page)
2394 page = __rmqueue(zone, order, migratetype);
2395 spin_unlock(&zone->lock);
2396 if (!page)
2397 goto failed;
2398 __mod_zone_freepage_state(zone, -(1 << order),
2399 get_pcppage_migratetype(page));
2400 }
2401
2402 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2403 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2404 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2405 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2406
2407 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2408 zone_statistics(preferred_zone, zone, gfp_flags);
2409 local_irq_restore(flags);
2410
2411 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2412 return page;
2413
2414failed:
2415 local_irq_restore(flags);
2416 return NULL;
2417}
2418
2419#ifdef CONFIG_FAIL_PAGE_ALLOC
2420
2421static struct {
2422 struct fault_attr attr;
2423
2424 bool ignore_gfp_highmem;
2425 bool ignore_gfp_reclaim;
2426 u32 min_order;
2427} fail_page_alloc = {
2428 .attr = FAULT_ATTR_INITIALIZER,
2429 .ignore_gfp_reclaim = true,
2430 .ignore_gfp_highmem = true,
2431 .min_order = 1,
2432};
2433
2434static int __init setup_fail_page_alloc(char *str)
2435{
2436 return setup_fault_attr(&fail_page_alloc.attr, str);
2437}
2438__setup("fail_page_alloc=", setup_fail_page_alloc);
2439
2440static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2441{
2442 if (order < fail_page_alloc.min_order)
2443 return false;
2444 if (gfp_mask & __GFP_NOFAIL)
2445 return false;
2446 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2447 return false;
2448 if (fail_page_alloc.ignore_gfp_reclaim &&
2449 (gfp_mask & __GFP_DIRECT_RECLAIM))
2450 return false;
2451
2452 return should_fail(&fail_page_alloc.attr, 1 << order);
2453}
2454
2455#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2456
2457static int __init fail_page_alloc_debugfs(void)
2458{
2459 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2460 struct dentry *dir;
2461
2462 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2463 &fail_page_alloc.attr);
2464 if (IS_ERR(dir))
2465 return PTR_ERR(dir);
2466
2467 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2468 &fail_page_alloc.ignore_gfp_reclaim))
2469 goto fail;
2470 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2471 &fail_page_alloc.ignore_gfp_highmem))
2472 goto fail;
2473 if (!debugfs_create_u32("min-order", mode, dir,
2474 &fail_page_alloc.min_order))
2475 goto fail;
2476
2477 return 0;
2478fail:
2479 debugfs_remove_recursive(dir);
2480
2481 return -ENOMEM;
2482}
2483
2484late_initcall(fail_page_alloc_debugfs);
2485
2486#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2487
2488#else /* CONFIG_FAIL_PAGE_ALLOC */
2489
2490static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2491{
2492 return false;
2493}
2494
2495#endif /* CONFIG_FAIL_PAGE_ALLOC */
2496
2497/*
2498 * Return true if free base pages are above 'mark'. For high-order checks it
2499 * will return true of the order-0 watermark is reached and there is at least
2500 * one free page of a suitable size. Checking now avoids taking the zone lock
2501 * to check in the allocation paths if no pages are free.
2502 */
2503static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2504 unsigned long mark, int classzone_idx, int alloc_flags,
2505 long free_pages)
2506{
2507 long min = mark;
2508 int o;
2509 const int alloc_harder = (alloc_flags & ALLOC_HARDER);
2510
2511 /* free_pages may go negative - that's OK */
2512 free_pages -= (1 << order) - 1;
2513
2514 if (alloc_flags & ALLOC_HIGH)
2515 min -= min / 2;
2516
2517 /*
2518 * If the caller does not have rights to ALLOC_HARDER then subtract
2519 * the high-atomic reserves. This will over-estimate the size of the
2520 * atomic reserve but it avoids a search.
2521 */
2522 if (likely(!alloc_harder))
2523 free_pages -= z->nr_reserved_highatomic;
2524 else
2525 min -= min / 4;
2526
2527#ifdef CONFIG_CMA
2528 /* If allocation can't use CMA areas don't use free CMA pages */
2529 if (!(alloc_flags & ALLOC_CMA))
2530 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2531#endif
2532
2533 /*
2534 * Check watermarks for an order-0 allocation request. If these
2535 * are not met, then a high-order request also cannot go ahead
2536 * even if a suitable page happened to be free.
2537 */
2538 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2539 return false;
2540
2541 /* If this is an order-0 request then the watermark is fine */
2542 if (!order)
2543 return true;
2544
2545 /* For a high-order request, check at least one suitable page is free */
2546 for (o = order; o < MAX_ORDER; o++) {
2547 struct free_area *area = &z->free_area[o];
2548 int mt;
2549
2550 if (!area->nr_free)
2551 continue;
2552
2553 if (alloc_harder)
2554 return true;
2555
2556 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2557 if (!list_empty(&area->free_list[mt]))
2558 return true;
2559 }
2560
2561#ifdef CONFIG_CMA
2562 if ((alloc_flags & ALLOC_CMA) &&
2563 !list_empty(&area->free_list[MIGRATE_CMA])) {
2564 return true;
2565 }
2566#endif
2567 }
2568 return false;
2569}
2570
2571bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2572 int classzone_idx, int alloc_flags)
2573{
2574 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2575 zone_page_state(z, NR_FREE_PAGES));
2576}
2577
2578bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2579 unsigned long mark, int classzone_idx)
2580{
2581 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2582
2583 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2584 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2585
2586 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2587 free_pages);
2588}
2589
2590#ifdef CONFIG_NUMA
2591static bool zone_local(struct zone *local_zone, struct zone *zone)
2592{
2593 return local_zone->node == zone->node;
2594}
2595
2596static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2597{
2598 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2599 RECLAIM_DISTANCE;
2600}
2601#else /* CONFIG_NUMA */
2602static bool zone_local(struct zone *local_zone, struct zone *zone)
2603{
2604 return true;
2605}
2606
2607static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2608{
2609 return true;
2610}
2611#endif /* CONFIG_NUMA */
2612
2613static void reset_alloc_batches(struct zone *preferred_zone)
2614{
2615 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2616
2617 do {
2618 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2619 high_wmark_pages(zone) - low_wmark_pages(zone) -
2620 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2621 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2622 } while (zone++ != preferred_zone);
2623}
2624
2625/*
2626 * get_page_from_freelist goes through the zonelist trying to allocate
2627 * a page.
2628 */
2629static struct page *
2630get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2631 const struct alloc_context *ac)
2632{
2633 struct zonelist *zonelist = ac->zonelist;
2634 struct zoneref *z;
2635 struct page *page = NULL;
2636 struct zone *zone;
2637 int nr_fair_skipped = 0;
2638 bool zonelist_rescan;
2639
2640zonelist_scan:
2641 zonelist_rescan = false;
2642
2643 /*
2644 * Scan zonelist, looking for a zone with enough free.
2645 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2646 */
2647 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2648 ac->nodemask) {
2649 unsigned long mark;
2650
2651 if (cpusets_enabled() &&
2652 (alloc_flags & ALLOC_CPUSET) &&
2653 !cpuset_zone_allowed(zone, gfp_mask))
2654 continue;
2655 /*
2656 * Distribute pages in proportion to the individual
2657 * zone size to ensure fair page aging. The zone a
2658 * page was allocated in should have no effect on the
2659 * time the page has in memory before being reclaimed.
2660 */
2661 if (alloc_flags & ALLOC_FAIR) {
2662 if (!zone_local(ac->preferred_zone, zone))
2663 break;
2664 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2665 nr_fair_skipped++;
2666 continue;
2667 }
2668 }
2669 /*
2670 * When allocating a page cache page for writing, we
2671 * want to get it from a zone that is within its dirty
2672 * limit, such that no single zone holds more than its
2673 * proportional share of globally allowed dirty pages.
2674 * The dirty limits take into account the zone's
2675 * lowmem reserves and high watermark so that kswapd
2676 * should be able to balance it without having to
2677 * write pages from its LRU list.
2678 *
2679 * This may look like it could increase pressure on
2680 * lower zones by failing allocations in higher zones
2681 * before they are full. But the pages that do spill
2682 * over are limited as the lower zones are protected
2683 * by this very same mechanism. It should not become
2684 * a practical burden to them.
2685 *
2686 * XXX: For now, allow allocations to potentially
2687 * exceed the per-zone dirty limit in the slowpath
2688 * (spread_dirty_pages unset) before going into reclaim,
2689 * which is important when on a NUMA setup the allowed
2690 * zones are together not big enough to reach the
2691 * global limit. The proper fix for these situations
2692 * will require awareness of zones in the
2693 * dirty-throttling and the flusher threads.
2694 */
2695 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2696 continue;
2697
2698 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2699 if (!zone_watermark_ok(zone, order, mark,
2700 ac->classzone_idx, alloc_flags)) {
2701 int ret;
2702
2703 /* Checked here to keep the fast path fast */
2704 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2705 if (alloc_flags & ALLOC_NO_WATERMARKS)
2706 goto try_this_zone;
2707
2708 if (zone_reclaim_mode == 0 ||
2709 !zone_allows_reclaim(ac->preferred_zone, zone))
2710 continue;
2711
2712 ret = zone_reclaim(zone, gfp_mask, order);
2713 switch (ret) {
2714 case ZONE_RECLAIM_NOSCAN:
2715 /* did not scan */
2716 continue;
2717 case ZONE_RECLAIM_FULL:
2718 /* scanned but unreclaimable */
2719 continue;
2720 default:
2721 /* did we reclaim enough */
2722 if (zone_watermark_ok(zone, order, mark,
2723 ac->classzone_idx, alloc_flags))
2724 goto try_this_zone;
2725
2726 continue;
2727 }
2728 }
2729
2730try_this_zone:
2731 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2732 gfp_mask, alloc_flags, ac->migratetype);
2733 if (page) {
2734 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2735 goto try_this_zone;
2736
2737 /*
2738 * If this is a high-order atomic allocation then check
2739 * if the pageblock should be reserved for the future
2740 */
2741 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2742 reserve_highatomic_pageblock(page, zone, order);
2743
2744 return page;
2745 }
2746 }
2747
2748 /*
2749 * The first pass makes sure allocations are spread fairly within the
2750 * local node. However, the local node might have free pages left
2751 * after the fairness batches are exhausted, and remote zones haven't
2752 * even been considered yet. Try once more without fairness, and
2753 * include remote zones now, before entering the slowpath and waking
2754 * kswapd: prefer spilling to a remote zone over swapping locally.
2755 */
2756 if (alloc_flags & ALLOC_FAIR) {
2757 alloc_flags &= ~ALLOC_FAIR;
2758 if (nr_fair_skipped) {
2759 zonelist_rescan = true;
2760 reset_alloc_batches(ac->preferred_zone);
2761 }
2762 if (nr_online_nodes > 1)
2763 zonelist_rescan = true;
2764 }
2765
2766 if (zonelist_rescan)
2767 goto zonelist_scan;
2768
2769 return NULL;
2770}
2771
2772/*
2773 * Large machines with many possible nodes should not always dump per-node
2774 * meminfo in irq context.
2775 */
2776static inline bool should_suppress_show_mem(void)
2777{
2778 bool ret = false;
2779
2780#if NODES_SHIFT > 8
2781 ret = in_interrupt();
2782#endif
2783 return ret;
2784}
2785
2786static DEFINE_RATELIMIT_STATE(nopage_rs,
2787 DEFAULT_RATELIMIT_INTERVAL,
2788 DEFAULT_RATELIMIT_BURST);
2789
2790void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
2791{
2792 unsigned int filter = SHOW_MEM_FILTER_NODES;
2793
2794 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2795 debug_guardpage_minorder() > 0)
2796 return;
2797
2798 /*
2799 * This documents exceptions given to allocations in certain
2800 * contexts that are allowed to allocate outside current's set
2801 * of allowed nodes.
2802 */
2803 if (!(gfp_mask & __GFP_NOMEMALLOC))
2804 if (test_thread_flag(TIF_MEMDIE) ||
2805 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2806 filter &= ~SHOW_MEM_FILTER_NODES;
2807 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2808 filter &= ~SHOW_MEM_FILTER_NODES;
2809
2810 if (fmt) {
2811 struct va_format vaf;
2812 va_list args;
2813
2814 va_start(args, fmt);
2815
2816 vaf.fmt = fmt;
2817 vaf.va = &args;
2818
2819 pr_warn("%pV", &vaf);
2820
2821 va_end(args);
2822 }
2823
2824 pr_warn("%s: page allocation failure: order:%u, mode:%#x(%pGg)\n",
2825 current->comm, order, gfp_mask, &gfp_mask);
2826 dump_stack();
2827 if (!should_suppress_show_mem())
2828 show_mem(filter);
2829}
2830
2831static inline struct page *
2832__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2833 const struct alloc_context *ac, unsigned long *did_some_progress)
2834{
2835 struct oom_control oc = {
2836 .zonelist = ac->zonelist,
2837 .nodemask = ac->nodemask,
2838 .gfp_mask = gfp_mask,
2839 .order = order,
2840 };
2841 struct page *page;
2842
2843 *did_some_progress = 0;
2844
2845 /*
2846 * Acquire the oom lock. If that fails, somebody else is
2847 * making progress for us.
2848 */
2849 if (!mutex_trylock(&oom_lock)) {
2850 *did_some_progress = 1;
2851 schedule_timeout_uninterruptible(1);
2852 return NULL;
2853 }
2854
2855 /*
2856 * Go through the zonelist yet one more time, keep very high watermark
2857 * here, this is only to catch a parallel oom killing, we must fail if
2858 * we're still under heavy pressure.
2859 */
2860 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2861 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2862 if (page)
2863 goto out;
2864
2865 if (!(gfp_mask & __GFP_NOFAIL)) {
2866 /* Coredumps can quickly deplete all memory reserves */
2867 if (current->flags & PF_DUMPCORE)
2868 goto out;
2869 /* The OOM killer will not help higher order allocs */
2870 if (order > PAGE_ALLOC_COSTLY_ORDER)
2871 goto out;
2872 /* The OOM killer does not needlessly kill tasks for lowmem */
2873 if (ac->high_zoneidx < ZONE_NORMAL)
2874 goto out;
2875 /* The OOM killer does not compensate for IO-less reclaim */
2876 if (!(gfp_mask & __GFP_FS)) {
2877 /*
2878 * XXX: Page reclaim didn't yield anything,
2879 * and the OOM killer can't be invoked, but
2880 * keep looping as per tradition.
2881 *
2882 * But do not keep looping if oom_killer_disable()
2883 * was already called, for the system is trying to
2884 * enter a quiescent state during suspend.
2885 */
2886 *did_some_progress = !oom_killer_disabled;
2887 goto out;
2888 }
2889 if (pm_suspended_storage())
2890 goto out;
2891 /* The OOM killer may not free memory on a specific node */
2892 if (gfp_mask & __GFP_THISNODE)
2893 goto out;
2894 }
2895 /* Exhausted what can be done so it's blamo time */
2896 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
2897 *did_some_progress = 1;
2898
2899 if (gfp_mask & __GFP_NOFAIL) {
2900 page = get_page_from_freelist(gfp_mask, order,
2901 ALLOC_NO_WATERMARKS|ALLOC_CPUSET, ac);
2902 /*
2903 * fallback to ignore cpuset restriction if our nodes
2904 * are depleted
2905 */
2906 if (!page)
2907 page = get_page_from_freelist(gfp_mask, order,
2908 ALLOC_NO_WATERMARKS, ac);
2909 }
2910 }
2911out:
2912 mutex_unlock(&oom_lock);
2913 return page;
2914}
2915
2916#ifdef CONFIG_COMPACTION
2917/* Try memory compaction for high-order allocations before reclaim */
2918static struct page *
2919__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2920 int alloc_flags, const struct alloc_context *ac,
2921 enum migrate_mode mode, int *contended_compaction,
2922 bool *deferred_compaction)
2923{
2924 unsigned long compact_result;
2925 struct page *page;
2926
2927 if (!order)
2928 return NULL;
2929
2930 current->flags |= PF_MEMALLOC;
2931 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2932 mode, contended_compaction);
2933 current->flags &= ~PF_MEMALLOC;
2934
2935 switch (compact_result) {
2936 case COMPACT_DEFERRED:
2937 *deferred_compaction = true;
2938 /* fall-through */
2939 case COMPACT_SKIPPED:
2940 return NULL;
2941 default:
2942 break;
2943 }
2944
2945 /*
2946 * At least in one zone compaction wasn't deferred or skipped, so let's
2947 * count a compaction stall
2948 */
2949 count_vm_event(COMPACTSTALL);
2950
2951 page = get_page_from_freelist(gfp_mask, order,
2952 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2953
2954 if (page) {
2955 struct zone *zone = page_zone(page);
2956
2957 zone->compact_blockskip_flush = false;
2958 compaction_defer_reset(zone, order, true);
2959 count_vm_event(COMPACTSUCCESS);
2960 return page;
2961 }
2962
2963 /*
2964 * It's bad if compaction run occurs and fails. The most likely reason
2965 * is that pages exist, but not enough to satisfy watermarks.
2966 */
2967 count_vm_event(COMPACTFAIL);
2968
2969 cond_resched();
2970
2971 return NULL;
2972}
2973#else
2974static inline struct page *
2975__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2976 int alloc_flags, const struct alloc_context *ac,
2977 enum migrate_mode mode, int *contended_compaction,
2978 bool *deferred_compaction)
2979{
2980 return NULL;
2981}
2982#endif /* CONFIG_COMPACTION */
2983
2984/* Perform direct synchronous page reclaim */
2985static int
2986__perform_reclaim(gfp_t gfp_mask, unsigned int order,
2987 const struct alloc_context *ac)
2988{
2989 struct reclaim_state reclaim_state;
2990 int progress;
2991
2992 cond_resched();
2993
2994 /* We now go into synchronous reclaim */
2995 cpuset_memory_pressure_bump();
2996 current->flags |= PF_MEMALLOC;
2997 lockdep_set_current_reclaim_state(gfp_mask);
2998 reclaim_state.reclaimed_slab = 0;
2999 current->reclaim_state = &reclaim_state;
3000
3001 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3002 ac->nodemask);
3003
3004 current->reclaim_state = NULL;
3005 lockdep_clear_current_reclaim_state();
3006 current->flags &= ~PF_MEMALLOC;
3007
3008 cond_resched();
3009
3010 return progress;
3011}
3012
3013/* The really slow allocator path where we enter direct reclaim */
3014static inline struct page *
3015__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3016 int alloc_flags, const struct alloc_context *ac,
3017 unsigned long *did_some_progress)
3018{
3019 struct page *page = NULL;
3020 bool drained = false;
3021
3022 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3023 if (unlikely(!(*did_some_progress)))
3024 return NULL;
3025
3026retry:
3027 page = get_page_from_freelist(gfp_mask, order,
3028 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3029
3030 /*
3031 * If an allocation failed after direct reclaim, it could be because
3032 * pages are pinned on the per-cpu lists or in high alloc reserves.
3033 * Shrink them them and try again
3034 */
3035 if (!page && !drained) {
3036 unreserve_highatomic_pageblock(ac);
3037 drain_all_pages(NULL);
3038 drained = true;
3039 goto retry;
3040 }
3041
3042 return page;
3043}
3044
3045static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3046{
3047 struct zoneref *z;
3048 struct zone *zone;
3049
3050 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3051 ac->high_zoneidx, ac->nodemask)
3052 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
3053}
3054
3055static inline int
3056gfp_to_alloc_flags(gfp_t gfp_mask)
3057{
3058 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3059
3060 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3061 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3062
3063 /*
3064 * The caller may dip into page reserves a bit more if the caller
3065 * cannot run direct reclaim, or if the caller has realtime scheduling
3066 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3067 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3068 */
3069 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3070
3071 if (gfp_mask & __GFP_ATOMIC) {
3072 /*
3073 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3074 * if it can't schedule.
3075 */
3076 if (!(gfp_mask & __GFP_NOMEMALLOC))
3077 alloc_flags |= ALLOC_HARDER;
3078 /*
3079 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3080 * comment for __cpuset_node_allowed().
3081 */
3082 alloc_flags &= ~ALLOC_CPUSET;
3083 } else if (unlikely(rt_task(current)) && !in_interrupt())
3084 alloc_flags |= ALLOC_HARDER;
3085
3086 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
3087 if (gfp_mask & __GFP_MEMALLOC)
3088 alloc_flags |= ALLOC_NO_WATERMARKS;
3089 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3090 alloc_flags |= ALLOC_NO_WATERMARKS;
3091 else if (!in_interrupt() &&
3092 ((current->flags & PF_MEMALLOC) ||
3093 unlikely(test_thread_flag(TIF_MEMDIE))))
3094 alloc_flags |= ALLOC_NO_WATERMARKS;
3095 }
3096#ifdef CONFIG_CMA
3097 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3098 alloc_flags |= ALLOC_CMA;
3099#endif
3100 return alloc_flags;
3101}
3102
3103bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3104{
3105 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
3106}
3107
3108static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
3109{
3110 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
3111}
3112
3113static inline struct page *
3114__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3115 struct alloc_context *ac)
3116{
3117 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3118 struct page *page = NULL;
3119 int alloc_flags;
3120 unsigned long pages_reclaimed = 0;
3121 unsigned long did_some_progress;
3122 enum migrate_mode migration_mode = MIGRATE_ASYNC;
3123 bool deferred_compaction = false;
3124 int contended_compaction = COMPACT_CONTENDED_NONE;
3125
3126 /*
3127 * In the slowpath, we sanity check order to avoid ever trying to
3128 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3129 * be using allocators in order of preference for an area that is
3130 * too large.
3131 */
3132 if (order >= MAX_ORDER) {
3133 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3134 return NULL;
3135 }
3136
3137 /*
3138 * We also sanity check to catch abuse of atomic reserves being used by
3139 * callers that are not in atomic context.
3140 */
3141 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3142 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3143 gfp_mask &= ~__GFP_ATOMIC;
3144
3145retry:
3146 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3147 wake_all_kswapds(order, ac);
3148
3149 /*
3150 * OK, we're below the kswapd watermark and have kicked background
3151 * reclaim. Now things get more complex, so set up alloc_flags according
3152 * to how we want to proceed.
3153 */
3154 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3155
3156 /*
3157 * Find the true preferred zone if the allocation is unconstrained by
3158 * cpusets.
3159 */
3160 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3161 struct zoneref *preferred_zoneref;
3162 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3163 ac->high_zoneidx, NULL, &ac->preferred_zone);
3164 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3165 }
3166
3167 /* This is the last chance, in general, before the goto nopage. */
3168 page = get_page_from_freelist(gfp_mask, order,
3169 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3170 if (page)
3171 goto got_pg;
3172
3173 /* Allocate without watermarks if the context allows */
3174 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3175 /*
3176 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3177 * the allocation is high priority and these type of
3178 * allocations are system rather than user orientated
3179 */
3180 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3181 page = get_page_from_freelist(gfp_mask, order,
3182 ALLOC_NO_WATERMARKS, ac);
3183 if (page)
3184 goto got_pg;
3185 }
3186
3187 /* Caller is not willing to reclaim, we can't balance anything */
3188 if (!can_direct_reclaim) {
3189 /*
3190 * All existing users of the __GFP_NOFAIL are blockable, so warn
3191 * of any new users that actually allow this type of allocation
3192 * to fail.
3193 */
3194 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3195 goto nopage;
3196 }
3197
3198 /* Avoid recursion of direct reclaim */
3199 if (current->flags & PF_MEMALLOC) {
3200 /*
3201 * __GFP_NOFAIL request from this context is rather bizarre
3202 * because we cannot reclaim anything and only can loop waiting
3203 * for somebody to do a work for us.
3204 */
3205 if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3206 cond_resched();
3207 goto retry;
3208 }
3209 goto nopage;
3210 }
3211
3212 /* Avoid allocations with no watermarks from looping endlessly */
3213 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3214 goto nopage;
3215
3216 /*
3217 * Try direct compaction. The first pass is asynchronous. Subsequent
3218 * attempts after direct reclaim are synchronous
3219 */
3220 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3221 migration_mode,
3222 &contended_compaction,
3223 &deferred_compaction);
3224 if (page)
3225 goto got_pg;
3226
3227 /* Checks for THP-specific high-order allocations */
3228 if (is_thp_gfp_mask(gfp_mask)) {
3229 /*
3230 * If compaction is deferred for high-order allocations, it is
3231 * because sync compaction recently failed. If this is the case
3232 * and the caller requested a THP allocation, we do not want
3233 * to heavily disrupt the system, so we fail the allocation
3234 * instead of entering direct reclaim.
3235 */
3236 if (deferred_compaction)
3237 goto nopage;
3238
3239 /*
3240 * In all zones where compaction was attempted (and not
3241 * deferred or skipped), lock contention has been detected.
3242 * For THP allocation we do not want to disrupt the others
3243 * so we fallback to base pages instead.
3244 */
3245 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3246 goto nopage;
3247
3248 /*
3249 * If compaction was aborted due to need_resched(), we do not
3250 * want to further increase allocation latency, unless it is
3251 * khugepaged trying to collapse.
3252 */
3253 if (contended_compaction == COMPACT_CONTENDED_SCHED
3254 && !(current->flags & PF_KTHREAD))
3255 goto nopage;
3256 }
3257
3258 /*
3259 * It can become very expensive to allocate transparent hugepages at
3260 * fault, so use asynchronous memory compaction for THP unless it is
3261 * khugepaged trying to collapse.
3262 */
3263 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3264 migration_mode = MIGRATE_SYNC_LIGHT;
3265
3266 /* Try direct reclaim and then allocating */
3267 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3268 &did_some_progress);
3269 if (page)
3270 goto got_pg;
3271
3272 /* Do not loop if specifically requested */
3273 if (gfp_mask & __GFP_NORETRY)
3274 goto noretry;
3275
3276 /* Keep reclaiming pages as long as there is reasonable progress */
3277 pages_reclaimed += did_some_progress;
3278 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3279 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3280 /* Wait for some write requests to complete then retry */
3281 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3282 goto retry;
3283 }
3284
3285 /* Reclaim has failed us, start killing things */
3286 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3287 if (page)
3288 goto got_pg;
3289
3290 /* Retry as long as the OOM killer is making progress */
3291 if (did_some_progress)
3292 goto retry;
3293
3294noretry:
3295 /*
3296 * High-order allocations do not necessarily loop after
3297 * direct reclaim and reclaim/compaction depends on compaction
3298 * being called after reclaim so call directly if necessary
3299 */
3300 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3301 ac, migration_mode,
3302 &contended_compaction,
3303 &deferred_compaction);
3304 if (page)
3305 goto got_pg;
3306nopage:
3307 warn_alloc_failed(gfp_mask, order, NULL);
3308got_pg:
3309 return page;
3310}
3311
3312/*
3313 * This is the 'heart' of the zoned buddy allocator.
3314 */
3315struct page *
3316__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3317 struct zonelist *zonelist, nodemask_t *nodemask)
3318{
3319 struct zoneref *preferred_zoneref;
3320 struct page *page = NULL;
3321 unsigned int cpuset_mems_cookie;
3322 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3323 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3324 struct alloc_context ac = {
3325 .high_zoneidx = gfp_zone(gfp_mask),
3326 .nodemask = nodemask,
3327 .migratetype = gfpflags_to_migratetype(gfp_mask),
3328 };
3329
3330 gfp_mask &= gfp_allowed_mask;
3331
3332 lockdep_trace_alloc(gfp_mask);
3333
3334 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3335
3336 if (should_fail_alloc_page(gfp_mask, order))
3337 return NULL;
3338
3339 /*
3340 * Check the zones suitable for the gfp_mask contain at least one
3341 * valid zone. It's possible to have an empty zonelist as a result
3342 * of __GFP_THISNODE and a memoryless node
3343 */
3344 if (unlikely(!zonelist->_zonerefs->zone))
3345 return NULL;
3346
3347 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3348 alloc_flags |= ALLOC_CMA;
3349
3350retry_cpuset:
3351 cpuset_mems_cookie = read_mems_allowed_begin();
3352
3353 /* We set it here, as __alloc_pages_slowpath might have changed it */
3354 ac.zonelist = zonelist;
3355
3356 /* Dirty zone balancing only done in the fast path */
3357 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3358
3359 /* The preferred zone is used for statistics later */
3360 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3361 ac.nodemask ? : &cpuset_current_mems_allowed,
3362 &ac.preferred_zone);
3363 if (!ac.preferred_zone)
3364 goto out;
3365 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3366
3367 /* First allocation attempt */
3368 alloc_mask = gfp_mask|__GFP_HARDWALL;
3369 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3370 if (unlikely(!page)) {
3371 /*
3372 * Runtime PM, block IO and its error handling path
3373 * can deadlock because I/O on the device might not
3374 * complete.
3375 */
3376 alloc_mask = memalloc_noio_flags(gfp_mask);
3377 ac.spread_dirty_pages = false;
3378
3379 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3380 }
3381
3382 if (kmemcheck_enabled && page)
3383 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3384
3385 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3386
3387out:
3388 /*
3389 * When updating a task's mems_allowed, it is possible to race with
3390 * parallel threads in such a way that an allocation can fail while
3391 * the mask is being updated. If a page allocation is about to fail,
3392 * check if the cpuset changed during allocation and if so, retry.
3393 */
3394 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3395 goto retry_cpuset;
3396
3397 return page;
3398}
3399EXPORT_SYMBOL(__alloc_pages_nodemask);
3400
3401/*
3402 * Common helper functions.
3403 */
3404unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3405{
3406 struct page *page;
3407
3408 /*
3409 * __get_free_pages() returns a 32-bit address, which cannot represent
3410 * a highmem page
3411 */
3412 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3413
3414 page = alloc_pages(gfp_mask, order);
3415 if (!page)
3416 return 0;
3417 return (unsigned long) page_address(page);
3418}
3419EXPORT_SYMBOL(__get_free_pages);
3420
3421unsigned long get_zeroed_page(gfp_t gfp_mask)
3422{
3423 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3424}
3425EXPORT_SYMBOL(get_zeroed_page);
3426
3427void __free_pages(struct page *page, unsigned int order)
3428{
3429 if (put_page_testzero(page)) {
3430 if (order == 0)
3431 free_hot_cold_page(page, false);
3432 else
3433 __free_pages_ok(page, order);
3434 }
3435}
3436
3437EXPORT_SYMBOL(__free_pages);
3438
3439void free_pages(unsigned long addr, unsigned int order)
3440{
3441 if (addr != 0) {
3442 VM_BUG_ON(!virt_addr_valid((void *)addr));
3443 __free_pages(virt_to_page((void *)addr), order);
3444 }
3445}
3446
3447EXPORT_SYMBOL(free_pages);
3448
3449/*
3450 * Page Fragment:
3451 * An arbitrary-length arbitrary-offset area of memory which resides
3452 * within a 0 or higher order page. Multiple fragments within that page
3453 * are individually refcounted, in the page's reference counter.
3454 *
3455 * The page_frag functions below provide a simple allocation framework for
3456 * page fragments. This is used by the network stack and network device
3457 * drivers to provide a backing region of memory for use as either an
3458 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3459 */
3460static struct page *__page_frag_refill(struct page_frag_cache *nc,
3461 gfp_t gfp_mask)
3462{
3463 struct page *page = NULL;
3464 gfp_t gfp = gfp_mask;
3465
3466#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3467 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3468 __GFP_NOMEMALLOC;
3469 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3470 PAGE_FRAG_CACHE_MAX_ORDER);
3471 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3472#endif
3473 if (unlikely(!page))
3474 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3475
3476 nc->va = page ? page_address(page) : NULL;
3477
3478 return page;
3479}
3480
3481void *__alloc_page_frag(struct page_frag_cache *nc,
3482 unsigned int fragsz, gfp_t gfp_mask)
3483{
3484 unsigned int size = PAGE_SIZE;
3485 struct page *page;
3486 int offset;
3487
3488 if (unlikely(!nc->va)) {
3489refill:
3490 page = __page_frag_refill(nc, gfp_mask);
3491 if (!page)
3492 return NULL;
3493
3494#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3495 /* if size can vary use size else just use PAGE_SIZE */
3496 size = nc->size;
3497#endif
3498 /* Even if we own the page, we do not use atomic_set().
3499 * This would break get_page_unless_zero() users.
3500 */
3501 page_ref_add(page, size - 1);
3502
3503 /* reset page count bias and offset to start of new frag */
3504 nc->pfmemalloc = page_is_pfmemalloc(page);
3505 nc->pagecnt_bias = size;
3506 nc->offset = size;
3507 }
3508
3509 offset = nc->offset - fragsz;
3510 if (unlikely(offset < 0)) {
3511 page = virt_to_page(nc->va);
3512
3513 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
3514 goto refill;
3515
3516#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3517 /* if size can vary use size else just use PAGE_SIZE */
3518 size = nc->size;
3519#endif
3520 /* OK, page count is 0, we can safely set it */
3521 set_page_count(page, size);
3522
3523 /* reset page count bias and offset to start of new frag */
3524 nc->pagecnt_bias = size;
3525 offset = size - fragsz;
3526 }
3527
3528 nc->pagecnt_bias--;
3529 nc->offset = offset;
3530
3531 return nc->va + offset;
3532}
3533EXPORT_SYMBOL(__alloc_page_frag);
3534
3535/*
3536 * Frees a page fragment allocated out of either a compound or order 0 page.
3537 */
3538void __free_page_frag(void *addr)
3539{
3540 struct page *page = virt_to_head_page(addr);
3541
3542 if (unlikely(put_page_testzero(page)))
3543 __free_pages_ok(page, compound_order(page));
3544}
3545EXPORT_SYMBOL(__free_page_frag);
3546
3547/*
3548 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3549 * of the current memory cgroup if __GFP_ACCOUNT is set, other than that it is
3550 * equivalent to alloc_pages.
3551 *
3552 * It should be used when the caller would like to use kmalloc, but since the
3553 * allocation is large, it has to fall back to the page allocator.
3554 */
3555struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3556{
3557 struct page *page;
3558
3559 page = alloc_pages(gfp_mask, order);
3560 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3561 __free_pages(page, order);
3562 page = NULL;
3563 }
3564 return page;
3565}
3566
3567struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3568{
3569 struct page *page;
3570
3571 page = alloc_pages_node(nid, gfp_mask, order);
3572 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3573 __free_pages(page, order);
3574 page = NULL;
3575 }
3576 return page;
3577}
3578
3579/*
3580 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3581 * alloc_kmem_pages.
3582 */
3583void __free_kmem_pages(struct page *page, unsigned int order)
3584{
3585 memcg_kmem_uncharge(page, order);
3586 __free_pages(page, order);
3587}
3588
3589void free_kmem_pages(unsigned long addr, unsigned int order)
3590{
3591 if (addr != 0) {
3592 VM_BUG_ON(!virt_addr_valid((void *)addr));
3593 __free_kmem_pages(virt_to_page((void *)addr), order);
3594 }
3595}
3596
3597static void *make_alloc_exact(unsigned long addr, unsigned int order,
3598 size_t size)
3599{
3600 if (addr) {
3601 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3602 unsigned long used = addr + PAGE_ALIGN(size);
3603
3604 split_page(virt_to_page((void *)addr), order);
3605 while (used < alloc_end) {
3606 free_page(used);
3607 used += PAGE_SIZE;
3608 }
3609 }
3610 return (void *)addr;
3611}
3612
3613/**
3614 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3615 * @size: the number of bytes to allocate
3616 * @gfp_mask: GFP flags for the allocation
3617 *
3618 * This function is similar to alloc_pages(), except that it allocates the
3619 * minimum number of pages to satisfy the request. alloc_pages() can only
3620 * allocate memory in power-of-two pages.
3621 *
3622 * This function is also limited by MAX_ORDER.
3623 *
3624 * Memory allocated by this function must be released by free_pages_exact().
3625 */
3626void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3627{
3628 unsigned int order = get_order(size);
3629 unsigned long addr;
3630
3631 addr = __get_free_pages(gfp_mask, order);
3632 return make_alloc_exact(addr, order, size);
3633}
3634EXPORT_SYMBOL(alloc_pages_exact);
3635
3636/**
3637 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3638 * pages on a node.
3639 * @nid: the preferred node ID where memory should be allocated
3640 * @size: the number of bytes to allocate
3641 * @gfp_mask: GFP flags for the allocation
3642 *
3643 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3644 * back.
3645 */
3646void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3647{
3648 unsigned int order = get_order(size);
3649 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3650 if (!p)
3651 return NULL;
3652 return make_alloc_exact((unsigned long)page_address(p), order, size);
3653}
3654
3655/**
3656 * free_pages_exact - release memory allocated via alloc_pages_exact()
3657 * @virt: the value returned by alloc_pages_exact.
3658 * @size: size of allocation, same value as passed to alloc_pages_exact().
3659 *
3660 * Release the memory allocated by a previous call to alloc_pages_exact.
3661 */
3662void free_pages_exact(void *virt, size_t size)
3663{
3664 unsigned long addr = (unsigned long)virt;
3665 unsigned long end = addr + PAGE_ALIGN(size);
3666
3667 while (addr < end) {
3668 free_page(addr);
3669 addr += PAGE_SIZE;
3670 }
3671}
3672EXPORT_SYMBOL(free_pages_exact);
3673
3674/**
3675 * nr_free_zone_pages - count number of pages beyond high watermark
3676 * @offset: The zone index of the highest zone
3677 *
3678 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3679 * high watermark within all zones at or below a given zone index. For each
3680 * zone, the number of pages is calculated as:
3681 * managed_pages - high_pages
3682 */
3683static unsigned long nr_free_zone_pages(int offset)
3684{
3685 struct zoneref *z;
3686 struct zone *zone;
3687
3688 /* Just pick one node, since fallback list is circular */
3689 unsigned long sum = 0;
3690
3691 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3692
3693 for_each_zone_zonelist(zone, z, zonelist, offset) {
3694 unsigned long size = zone->managed_pages;
3695 unsigned long high = high_wmark_pages(zone);
3696 if (size > high)
3697 sum += size - high;
3698 }
3699
3700 return sum;
3701}
3702
3703/**
3704 * nr_free_buffer_pages - count number of pages beyond high watermark
3705 *
3706 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3707 * watermark within ZONE_DMA and ZONE_NORMAL.
3708 */
3709unsigned long nr_free_buffer_pages(void)
3710{
3711 return nr_free_zone_pages(gfp_zone(GFP_USER));
3712}
3713EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3714
3715/**
3716 * nr_free_pagecache_pages - count number of pages beyond high watermark
3717 *
3718 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3719 * high watermark within all zones.
3720 */
3721unsigned long nr_free_pagecache_pages(void)
3722{
3723 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3724}
3725
3726static inline void show_node(struct zone *zone)
3727{
3728 if (IS_ENABLED(CONFIG_NUMA))
3729 printk("Node %d ", zone_to_nid(zone));
3730}
3731
3732long si_mem_available(void)
3733{
3734 long available;
3735 unsigned long pagecache;
3736 unsigned long wmark_low = 0;
3737 unsigned long pages[NR_LRU_LISTS];
3738 struct zone *zone;
3739 int lru;
3740
3741 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
3742 pages[lru] = global_page_state(NR_LRU_BASE + lru);
3743
3744 for_each_zone(zone)
3745 wmark_low += zone->watermark[WMARK_LOW];
3746
3747 /*
3748 * Estimate the amount of memory available for userspace allocations,
3749 * without causing swapping.
3750 */
3751 available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
3752
3753 /*
3754 * Not all the page cache can be freed, otherwise the system will
3755 * start swapping. Assume at least half of the page cache, or the
3756 * low watermark worth of cache, needs to stay.
3757 */
3758 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
3759 pagecache -= min(pagecache / 2, wmark_low);
3760 available += pagecache;
3761
3762 /*
3763 * Part of the reclaimable slab consists of items that are in use,
3764 * and cannot be freed. Cap this estimate at the low watermark.
3765 */
3766 available += global_page_state(NR_SLAB_RECLAIMABLE) -
3767 min(global_page_state(NR_SLAB_RECLAIMABLE) / 2, wmark_low);
3768
3769 if (available < 0)
3770 available = 0;
3771 return available;
3772}
3773EXPORT_SYMBOL_GPL(si_mem_available);
3774
3775void si_meminfo(struct sysinfo *val)
3776{
3777 val->totalram = totalram_pages;
3778 val->sharedram = global_page_state(NR_SHMEM);
3779 val->freeram = global_page_state(NR_FREE_PAGES);
3780 val->bufferram = nr_blockdev_pages();
3781 val->totalhigh = totalhigh_pages;
3782 val->freehigh = nr_free_highpages();
3783 val->mem_unit = PAGE_SIZE;
3784}
3785
3786EXPORT_SYMBOL(si_meminfo);
3787
3788#ifdef CONFIG_NUMA
3789void si_meminfo_node(struct sysinfo *val, int nid)
3790{
3791 int zone_type; /* needs to be signed */
3792 unsigned long managed_pages = 0;
3793 pg_data_t *pgdat = NODE_DATA(nid);
3794
3795 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3796 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3797 val->totalram = managed_pages;
3798 val->sharedram = node_page_state(nid, NR_SHMEM);
3799 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3800#ifdef CONFIG_HIGHMEM
3801 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3802 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3803 NR_FREE_PAGES);
3804#else
3805 val->totalhigh = 0;
3806 val->freehigh = 0;
3807#endif
3808 val->mem_unit = PAGE_SIZE;
3809}
3810#endif
3811
3812/*
3813 * Determine whether the node should be displayed or not, depending on whether
3814 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3815 */
3816bool skip_free_areas_node(unsigned int flags, int nid)
3817{
3818 bool ret = false;
3819 unsigned int cpuset_mems_cookie;
3820
3821 if (!(flags & SHOW_MEM_FILTER_NODES))
3822 goto out;
3823
3824 do {
3825 cpuset_mems_cookie = read_mems_allowed_begin();
3826 ret = !node_isset(nid, cpuset_current_mems_allowed);
3827 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3828out:
3829 return ret;
3830}
3831
3832#define K(x) ((x) << (PAGE_SHIFT-10))
3833
3834static void show_migration_types(unsigned char type)
3835{
3836 static const char types[MIGRATE_TYPES] = {
3837 [MIGRATE_UNMOVABLE] = 'U',
3838 [MIGRATE_MOVABLE] = 'M',
3839 [MIGRATE_RECLAIMABLE] = 'E',
3840 [MIGRATE_HIGHATOMIC] = 'H',
3841#ifdef CONFIG_CMA
3842 [MIGRATE_CMA] = 'C',
3843#endif
3844#ifdef CONFIG_MEMORY_ISOLATION
3845 [MIGRATE_ISOLATE] = 'I',
3846#endif
3847 };
3848 char tmp[MIGRATE_TYPES + 1];
3849 char *p = tmp;
3850 int i;
3851
3852 for (i = 0; i < MIGRATE_TYPES; i++) {
3853 if (type & (1 << i))
3854 *p++ = types[i];
3855 }
3856
3857 *p = '\0';
3858 printk("(%s) ", tmp);
3859}
3860
3861/*
3862 * Show free area list (used inside shift_scroll-lock stuff)
3863 * We also calculate the percentage fragmentation. We do this by counting the
3864 * memory on each free list with the exception of the first item on the list.
3865 *
3866 * Bits in @filter:
3867 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3868 * cpuset.
3869 */
3870void show_free_areas(unsigned int filter)
3871{
3872 unsigned long free_pcp = 0;
3873 int cpu;
3874 struct zone *zone;
3875
3876 for_each_populated_zone(zone) {
3877 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3878 continue;
3879
3880 for_each_online_cpu(cpu)
3881 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3882 }
3883
3884 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3885 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3886 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3887 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3888 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3889 " free:%lu free_pcp:%lu free_cma:%lu\n",
3890 global_page_state(NR_ACTIVE_ANON),
3891 global_page_state(NR_INACTIVE_ANON),
3892 global_page_state(NR_ISOLATED_ANON),
3893 global_page_state(NR_ACTIVE_FILE),
3894 global_page_state(NR_INACTIVE_FILE),
3895 global_page_state(NR_ISOLATED_FILE),
3896 global_page_state(NR_UNEVICTABLE),
3897 global_page_state(NR_FILE_DIRTY),
3898 global_page_state(NR_WRITEBACK),
3899 global_page_state(NR_UNSTABLE_NFS),
3900 global_page_state(NR_SLAB_RECLAIMABLE),
3901 global_page_state(NR_SLAB_UNRECLAIMABLE),
3902 global_page_state(NR_FILE_MAPPED),
3903 global_page_state(NR_SHMEM),
3904 global_page_state(NR_PAGETABLE),
3905 global_page_state(NR_BOUNCE),
3906 global_page_state(NR_FREE_PAGES),
3907 free_pcp,
3908 global_page_state(NR_FREE_CMA_PAGES));
3909
3910 for_each_populated_zone(zone) {
3911 int i;
3912
3913 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3914 continue;
3915
3916 free_pcp = 0;
3917 for_each_online_cpu(cpu)
3918 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3919
3920 show_node(zone);
3921 printk("%s"
3922 " free:%lukB"
3923 " min:%lukB"
3924 " low:%lukB"
3925 " high:%lukB"
3926 " active_anon:%lukB"
3927 " inactive_anon:%lukB"
3928 " active_file:%lukB"
3929 " inactive_file:%lukB"
3930 " unevictable:%lukB"
3931 " isolated(anon):%lukB"
3932 " isolated(file):%lukB"
3933 " present:%lukB"
3934 " managed:%lukB"
3935 " mlocked:%lukB"
3936 " dirty:%lukB"
3937 " writeback:%lukB"
3938 " mapped:%lukB"
3939 " shmem:%lukB"
3940 " slab_reclaimable:%lukB"
3941 " slab_unreclaimable:%lukB"
3942 " kernel_stack:%lukB"
3943 " pagetables:%lukB"
3944 " unstable:%lukB"
3945 " bounce:%lukB"
3946 " free_pcp:%lukB"
3947 " local_pcp:%ukB"
3948 " free_cma:%lukB"
3949 " writeback_tmp:%lukB"
3950 " pages_scanned:%lu"
3951 " all_unreclaimable? %s"
3952 "\n",
3953 zone->name,
3954 K(zone_page_state(zone, NR_FREE_PAGES)),
3955 K(min_wmark_pages(zone)),
3956 K(low_wmark_pages(zone)),
3957 K(high_wmark_pages(zone)),
3958 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3959 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3960 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3961 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3962 K(zone_page_state(zone, NR_UNEVICTABLE)),
3963 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3964 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3965 K(zone->present_pages),
3966 K(zone->managed_pages),
3967 K(zone_page_state(zone, NR_MLOCK)),
3968 K(zone_page_state(zone, NR_FILE_DIRTY)),
3969 K(zone_page_state(zone, NR_WRITEBACK)),
3970 K(zone_page_state(zone, NR_FILE_MAPPED)),
3971 K(zone_page_state(zone, NR_SHMEM)),
3972 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3973 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3974 zone_page_state(zone, NR_KERNEL_STACK) *
3975 THREAD_SIZE / 1024,
3976 K(zone_page_state(zone, NR_PAGETABLE)),
3977 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3978 K(zone_page_state(zone, NR_BOUNCE)),
3979 K(free_pcp),
3980 K(this_cpu_read(zone->pageset->pcp.count)),
3981 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3982 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3983 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3984 (!zone_reclaimable(zone) ? "yes" : "no")
3985 );
3986 printk("lowmem_reserve[]:");
3987 for (i = 0; i < MAX_NR_ZONES; i++)
3988 printk(" %ld", zone->lowmem_reserve[i]);
3989 printk("\n");
3990 }
3991
3992 for_each_populated_zone(zone) {
3993 unsigned int order;
3994 unsigned long nr[MAX_ORDER], flags, total = 0;
3995 unsigned char types[MAX_ORDER];
3996
3997 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3998 continue;
3999 show_node(zone);
4000 printk("%s: ", zone->name);
4001
4002 spin_lock_irqsave(&zone->lock, flags);
4003 for (order = 0; order < MAX_ORDER; order++) {
4004 struct free_area *area = &zone->free_area[order];
4005 int type;
4006
4007 nr[order] = area->nr_free;
4008 total += nr[order] << order;
4009
4010 types[order] = 0;
4011 for (type = 0; type < MIGRATE_TYPES; type++) {
4012 if (!list_empty(&area->free_list[type]))
4013 types[order] |= 1 << type;
4014 }
4015 }
4016 spin_unlock_irqrestore(&zone->lock, flags);
4017 for (order = 0; order < MAX_ORDER; order++) {
4018 printk("%lu*%lukB ", nr[order], K(1UL) << order);
4019 if (nr[order])
4020 show_migration_types(types[order]);
4021 }
4022 printk("= %lukB\n", K(total));
4023 }
4024
4025 hugetlb_show_meminfo();
4026
4027 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
4028
4029 show_swap_cache_info();
4030}
4031
4032static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4033{
4034 zoneref->zone = zone;
4035 zoneref->zone_idx = zone_idx(zone);
4036}
4037
4038/*
4039 * Builds allocation fallback zone lists.
4040 *
4041 * Add all populated zones of a node to the zonelist.
4042 */
4043static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
4044 int nr_zones)
4045{
4046 struct zone *zone;
4047 enum zone_type zone_type = MAX_NR_ZONES;
4048
4049 do {
4050 zone_type--;
4051 zone = pgdat->node_zones + zone_type;
4052 if (populated_zone(zone)) {
4053 zoneref_set_zone(zone,
4054 &zonelist->_zonerefs[nr_zones++]);
4055 check_highest_zone(zone_type);
4056 }
4057 } while (zone_type);
4058
4059 return nr_zones;
4060}
4061
4062
4063/*
4064 * zonelist_order:
4065 * 0 = automatic detection of better ordering.
4066 * 1 = order by ([node] distance, -zonetype)
4067 * 2 = order by (-zonetype, [node] distance)
4068 *
4069 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4070 * the same zonelist. So only NUMA can configure this param.
4071 */
4072#define ZONELIST_ORDER_DEFAULT 0
4073#define ZONELIST_ORDER_NODE 1
4074#define ZONELIST_ORDER_ZONE 2
4075
4076/* zonelist order in the kernel.
4077 * set_zonelist_order() will set this to NODE or ZONE.
4078 */
4079static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4080static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4081
4082
4083#ifdef CONFIG_NUMA
4084/* The value user specified ....changed by config */
4085static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4086/* string for sysctl */
4087#define NUMA_ZONELIST_ORDER_LEN 16
4088char numa_zonelist_order[16] = "default";
4089
4090/*
4091 * interface for configure zonelist ordering.
4092 * command line option "numa_zonelist_order"
4093 * = "[dD]efault - default, automatic configuration.
4094 * = "[nN]ode - order by node locality, then by zone within node
4095 * = "[zZ]one - order by zone, then by locality within zone
4096 */
4097
4098static int __parse_numa_zonelist_order(char *s)
4099{
4100 if (*s == 'd' || *s == 'D') {
4101 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4102 } else if (*s == 'n' || *s == 'N') {
4103 user_zonelist_order = ZONELIST_ORDER_NODE;
4104 } else if (*s == 'z' || *s == 'Z') {
4105 user_zonelist_order = ZONELIST_ORDER_ZONE;
4106 } else {
4107 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s);
4108 return -EINVAL;
4109 }
4110 return 0;
4111}
4112
4113static __init int setup_numa_zonelist_order(char *s)
4114{
4115 int ret;
4116
4117 if (!s)
4118 return 0;
4119
4120 ret = __parse_numa_zonelist_order(s);
4121 if (ret == 0)
4122 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4123
4124 return ret;
4125}
4126early_param("numa_zonelist_order", setup_numa_zonelist_order);
4127
4128/*
4129 * sysctl handler for numa_zonelist_order
4130 */
4131int numa_zonelist_order_handler(struct ctl_table *table, int write,
4132 void __user *buffer, size_t *length,
4133 loff_t *ppos)
4134{
4135 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4136 int ret;
4137 static DEFINE_MUTEX(zl_order_mutex);
4138
4139 mutex_lock(&zl_order_mutex);
4140 if (write) {
4141 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4142 ret = -EINVAL;
4143 goto out;
4144 }
4145 strcpy(saved_string, (char *)table->data);
4146 }
4147 ret = proc_dostring(table, write, buffer, length, ppos);
4148 if (ret)
4149 goto out;
4150 if (write) {
4151 int oldval = user_zonelist_order;
4152
4153 ret = __parse_numa_zonelist_order((char *)table->data);
4154 if (ret) {
4155 /*
4156 * bogus value. restore saved string
4157 */
4158 strncpy((char *)table->data, saved_string,
4159 NUMA_ZONELIST_ORDER_LEN);
4160 user_zonelist_order = oldval;
4161 } else if (oldval != user_zonelist_order) {
4162 mutex_lock(&zonelists_mutex);
4163 build_all_zonelists(NULL, NULL);
4164 mutex_unlock(&zonelists_mutex);
4165 }
4166 }
4167out:
4168 mutex_unlock(&zl_order_mutex);
4169 return ret;
4170}
4171
4172
4173#define MAX_NODE_LOAD (nr_online_nodes)
4174static int node_load[MAX_NUMNODES];
4175
4176/**
4177 * find_next_best_node - find the next node that should appear in a given node's fallback list
4178 * @node: node whose fallback list we're appending
4179 * @used_node_mask: nodemask_t of already used nodes
4180 *
4181 * We use a number of factors to determine which is the next node that should
4182 * appear on a given node's fallback list. The node should not have appeared
4183 * already in @node's fallback list, and it should be the next closest node
4184 * according to the distance array (which contains arbitrary distance values
4185 * from each node to each node in the system), and should also prefer nodes
4186 * with no CPUs, since presumably they'll have very little allocation pressure
4187 * on them otherwise.
4188 * It returns -1 if no node is found.
4189 */
4190static int find_next_best_node(int node, nodemask_t *used_node_mask)
4191{
4192 int n, val;
4193 int min_val = INT_MAX;
4194 int best_node = NUMA_NO_NODE;
4195 const struct cpumask *tmp = cpumask_of_node(0);
4196
4197 /* Use the local node if we haven't already */
4198 if (!node_isset(node, *used_node_mask)) {
4199 node_set(node, *used_node_mask);
4200 return node;
4201 }
4202
4203 for_each_node_state(n, N_MEMORY) {
4204
4205 /* Don't want a node to appear more than once */
4206 if (node_isset(n, *used_node_mask))
4207 continue;
4208
4209 /* Use the distance array to find the distance */
4210 val = node_distance(node, n);
4211
4212 /* Penalize nodes under us ("prefer the next node") */
4213 val += (n < node);
4214
4215 /* Give preference to headless and unused nodes */
4216 tmp = cpumask_of_node(n);
4217 if (!cpumask_empty(tmp))
4218 val += PENALTY_FOR_NODE_WITH_CPUS;
4219
4220 /* Slight preference for less loaded node */
4221 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4222 val += node_load[n];
4223
4224 if (val < min_val) {
4225 min_val = val;
4226 best_node = n;
4227 }
4228 }
4229
4230 if (best_node >= 0)
4231 node_set(best_node, *used_node_mask);
4232
4233 return best_node;
4234}
4235
4236
4237/*
4238 * Build zonelists ordered by node and zones within node.
4239 * This results in maximum locality--normal zone overflows into local
4240 * DMA zone, if any--but risks exhausting DMA zone.
4241 */
4242static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4243{
4244 int j;
4245 struct zonelist *zonelist;
4246
4247 zonelist = &pgdat->node_zonelists[0];
4248 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4249 ;
4250 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4251 zonelist->_zonerefs[j].zone = NULL;
4252 zonelist->_zonerefs[j].zone_idx = 0;
4253}
4254
4255/*
4256 * Build gfp_thisnode zonelists
4257 */
4258static void build_thisnode_zonelists(pg_data_t *pgdat)
4259{
4260 int j;
4261 struct zonelist *zonelist;
4262
4263 zonelist = &pgdat->node_zonelists[1];
4264 j = build_zonelists_node(pgdat, zonelist, 0);
4265 zonelist->_zonerefs[j].zone = NULL;
4266 zonelist->_zonerefs[j].zone_idx = 0;
4267}
4268
4269/*
4270 * Build zonelists ordered by zone and nodes within zones.
4271 * This results in conserving DMA zone[s] until all Normal memory is
4272 * exhausted, but results in overflowing to remote node while memory
4273 * may still exist in local DMA zone.
4274 */
4275static int node_order[MAX_NUMNODES];
4276
4277static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4278{
4279 int pos, j, node;
4280 int zone_type; /* needs to be signed */
4281 struct zone *z;
4282 struct zonelist *zonelist;
4283
4284 zonelist = &pgdat->node_zonelists[0];
4285 pos = 0;
4286 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4287 for (j = 0; j < nr_nodes; j++) {
4288 node = node_order[j];
4289 z = &NODE_DATA(node)->node_zones[zone_type];
4290 if (populated_zone(z)) {
4291 zoneref_set_zone(z,
4292 &zonelist->_zonerefs[pos++]);
4293 check_highest_zone(zone_type);
4294 }
4295 }
4296 }
4297 zonelist->_zonerefs[pos].zone = NULL;
4298 zonelist->_zonerefs[pos].zone_idx = 0;
4299}
4300
4301#if defined(CONFIG_64BIT)
4302/*
4303 * Devices that require DMA32/DMA are relatively rare and do not justify a
4304 * penalty to every machine in case the specialised case applies. Default
4305 * to Node-ordering on 64-bit NUMA machines
4306 */
4307static int default_zonelist_order(void)
4308{
4309 return ZONELIST_ORDER_NODE;
4310}
4311#else
4312/*
4313 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4314 * by the kernel. If processes running on node 0 deplete the low memory zone
4315 * then reclaim will occur more frequency increasing stalls and potentially
4316 * be easier to OOM if a large percentage of the zone is under writeback or
4317 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4318 * Hence, default to zone ordering on 32-bit.
4319 */
4320static int default_zonelist_order(void)
4321{
4322 return ZONELIST_ORDER_ZONE;
4323}
4324#endif /* CONFIG_64BIT */
4325
4326static void set_zonelist_order(void)
4327{
4328 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4329 current_zonelist_order = default_zonelist_order();
4330 else
4331 current_zonelist_order = user_zonelist_order;
4332}
4333
4334static void build_zonelists(pg_data_t *pgdat)
4335{
4336 int i, node, load;
4337 nodemask_t used_mask;
4338 int local_node, prev_node;
4339 struct zonelist *zonelist;
4340 unsigned int order = current_zonelist_order;
4341
4342 /* initialize zonelists */
4343 for (i = 0; i < MAX_ZONELISTS; i++) {
4344 zonelist = pgdat->node_zonelists + i;
4345 zonelist->_zonerefs[0].zone = NULL;
4346 zonelist->_zonerefs[0].zone_idx = 0;
4347 }
4348
4349 /* NUMA-aware ordering of nodes */
4350 local_node = pgdat->node_id;
4351 load = nr_online_nodes;
4352 prev_node = local_node;
4353 nodes_clear(used_mask);
4354
4355 memset(node_order, 0, sizeof(node_order));
4356 i = 0;
4357
4358 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4359 /*
4360 * We don't want to pressure a particular node.
4361 * So adding penalty to the first node in same
4362 * distance group to make it round-robin.
4363 */
4364 if (node_distance(local_node, node) !=
4365 node_distance(local_node, prev_node))
4366 node_load[node] = load;
4367
4368 prev_node = node;
4369 load--;
4370 if (order == ZONELIST_ORDER_NODE)
4371 build_zonelists_in_node_order(pgdat, node);
4372 else
4373 node_order[i++] = node; /* remember order */
4374 }
4375
4376 if (order == ZONELIST_ORDER_ZONE) {
4377 /* calculate node order -- i.e., DMA last! */
4378 build_zonelists_in_zone_order(pgdat, i);
4379 }
4380
4381 build_thisnode_zonelists(pgdat);
4382}
4383
4384#ifdef CONFIG_HAVE_MEMORYLESS_NODES
4385/*
4386 * Return node id of node used for "local" allocations.
4387 * I.e., first node id of first zone in arg node's generic zonelist.
4388 * Used for initializing percpu 'numa_mem', which is used primarily
4389 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4390 */
4391int local_memory_node(int node)
4392{
4393 struct zone *zone;
4394
4395 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4396 gfp_zone(GFP_KERNEL),
4397 NULL,
4398 &zone);
4399 return zone->node;
4400}
4401#endif
4402
4403#else /* CONFIG_NUMA */
4404
4405static void set_zonelist_order(void)
4406{
4407 current_zonelist_order = ZONELIST_ORDER_ZONE;
4408}
4409
4410static void build_zonelists(pg_data_t *pgdat)
4411{
4412 int node, local_node;
4413 enum zone_type j;
4414 struct zonelist *zonelist;
4415
4416 local_node = pgdat->node_id;
4417
4418 zonelist = &pgdat->node_zonelists[0];
4419 j = build_zonelists_node(pgdat, zonelist, 0);
4420
4421 /*
4422 * Now we build the zonelist so that it contains the zones
4423 * of all the other nodes.
4424 * We don't want to pressure a particular node, so when
4425 * building the zones for node N, we make sure that the
4426 * zones coming right after the local ones are those from
4427 * node N+1 (modulo N)
4428 */
4429 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4430 if (!node_online(node))
4431 continue;
4432 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4433 }
4434 for (node = 0; node < local_node; node++) {
4435 if (!node_online(node))
4436 continue;
4437 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4438 }
4439
4440 zonelist->_zonerefs[j].zone = NULL;
4441 zonelist->_zonerefs[j].zone_idx = 0;
4442}
4443
4444#endif /* CONFIG_NUMA */
4445
4446/*
4447 * Boot pageset table. One per cpu which is going to be used for all
4448 * zones and all nodes. The parameters will be set in such a way
4449 * that an item put on a list will immediately be handed over to
4450 * the buddy list. This is safe since pageset manipulation is done
4451 * with interrupts disabled.
4452 *
4453 * The boot_pagesets must be kept even after bootup is complete for
4454 * unused processors and/or zones. They do play a role for bootstrapping
4455 * hotplugged processors.
4456 *
4457 * zoneinfo_show() and maybe other functions do
4458 * not check if the processor is online before following the pageset pointer.
4459 * Other parts of the kernel may not check if the zone is available.
4460 */
4461static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4462static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4463static void setup_zone_pageset(struct zone *zone);
4464
4465/*
4466 * Global mutex to protect against size modification of zonelists
4467 * as well as to serialize pageset setup for the new populated zone.
4468 */
4469DEFINE_MUTEX(zonelists_mutex);
4470
4471/* return values int ....just for stop_machine() */
4472static int __build_all_zonelists(void *data)
4473{
4474 int nid;
4475 int cpu;
4476 pg_data_t *self = data;
4477
4478#ifdef CONFIG_NUMA
4479 memset(node_load, 0, sizeof(node_load));
4480#endif
4481
4482 if (self && !node_online(self->node_id)) {
4483 build_zonelists(self);
4484 }
4485
4486 for_each_online_node(nid) {
4487 pg_data_t *pgdat = NODE_DATA(nid);
4488
4489 build_zonelists(pgdat);
4490 }
4491
4492 /*
4493 * Initialize the boot_pagesets that are going to be used
4494 * for bootstrapping processors. The real pagesets for
4495 * each zone will be allocated later when the per cpu
4496 * allocator is available.
4497 *
4498 * boot_pagesets are used also for bootstrapping offline
4499 * cpus if the system is already booted because the pagesets
4500 * are needed to initialize allocators on a specific cpu too.
4501 * F.e. the percpu allocator needs the page allocator which
4502 * needs the percpu allocator in order to allocate its pagesets
4503 * (a chicken-egg dilemma).
4504 */
4505 for_each_possible_cpu(cpu) {
4506 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4507
4508#ifdef CONFIG_HAVE_MEMORYLESS_NODES
4509 /*
4510 * We now know the "local memory node" for each node--
4511 * i.e., the node of the first zone in the generic zonelist.
4512 * Set up numa_mem percpu variable for on-line cpus. During
4513 * boot, only the boot cpu should be on-line; we'll init the
4514 * secondary cpus' numa_mem as they come on-line. During
4515 * node/memory hotplug, we'll fixup all on-line cpus.
4516 */
4517 if (cpu_online(cpu))
4518 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4519#endif
4520 }
4521
4522 return 0;
4523}
4524
4525static noinline void __init
4526build_all_zonelists_init(void)
4527{
4528 __build_all_zonelists(NULL);
4529 mminit_verify_zonelist();
4530 cpuset_init_current_mems_allowed();
4531}
4532
4533/*
4534 * Called with zonelists_mutex held always
4535 * unless system_state == SYSTEM_BOOTING.
4536 *
4537 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4538 * [we're only called with non-NULL zone through __meminit paths] and
4539 * (2) call of __init annotated helper build_all_zonelists_init
4540 * [protected by SYSTEM_BOOTING].
4541 */
4542void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4543{
4544 set_zonelist_order();
4545
4546 if (system_state == SYSTEM_BOOTING) {
4547 build_all_zonelists_init();
4548 } else {
4549#ifdef CONFIG_MEMORY_HOTPLUG
4550 if (zone)
4551 setup_zone_pageset(zone);
4552#endif
4553 /* we have to stop all cpus to guarantee there is no user
4554 of zonelist */
4555 stop_machine(__build_all_zonelists, pgdat, NULL);
4556 /* cpuset refresh routine should be here */
4557 }
4558 vm_total_pages = nr_free_pagecache_pages();
4559 /*
4560 * Disable grouping by mobility if the number of pages in the
4561 * system is too low to allow the mechanism to work. It would be
4562 * more accurate, but expensive to check per-zone. This check is
4563 * made on memory-hotadd so a system can start with mobility
4564 * disabled and enable it later
4565 */
4566 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4567 page_group_by_mobility_disabled = 1;
4568 else
4569 page_group_by_mobility_disabled = 0;
4570
4571 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
4572 nr_online_nodes,
4573 zonelist_order_name[current_zonelist_order],
4574 page_group_by_mobility_disabled ? "off" : "on",
4575 vm_total_pages);
4576#ifdef CONFIG_NUMA
4577 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4578#endif
4579}
4580
4581/*
4582 * Helper functions to size the waitqueue hash table.
4583 * Essentially these want to choose hash table sizes sufficiently
4584 * large so that collisions trying to wait on pages are rare.
4585 * But in fact, the number of active page waitqueues on typical
4586 * systems is ridiculously low, less than 200. So this is even
4587 * conservative, even though it seems large.
4588 *
4589 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4590 * waitqueues, i.e. the size of the waitq table given the number of pages.
4591 */
4592#define PAGES_PER_WAITQUEUE 256
4593
4594#ifndef CONFIG_MEMORY_HOTPLUG
4595static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4596{
4597 unsigned long size = 1;
4598
4599 pages /= PAGES_PER_WAITQUEUE;
4600
4601 while (size < pages)
4602 size <<= 1;
4603
4604 /*
4605 * Once we have dozens or even hundreds of threads sleeping
4606 * on IO we've got bigger problems than wait queue collision.
4607 * Limit the size of the wait table to a reasonable size.
4608 */
4609 size = min(size, 4096UL);
4610
4611 return max(size, 4UL);
4612}
4613#else
4614/*
4615 * A zone's size might be changed by hot-add, so it is not possible to determine
4616 * a suitable size for its wait_table. So we use the maximum size now.
4617 *
4618 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4619 *
4620 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4621 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4622 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4623 *
4624 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4625 * or more by the traditional way. (See above). It equals:
4626 *
4627 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4628 * ia64(16K page size) : = ( 8G + 4M)byte.
4629 * powerpc (64K page size) : = (32G +16M)byte.
4630 */
4631static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4632{
4633 return 4096UL;
4634}
4635#endif
4636
4637/*
4638 * This is an integer logarithm so that shifts can be used later
4639 * to extract the more random high bits from the multiplicative
4640 * hash function before the remainder is taken.
4641 */
4642static inline unsigned long wait_table_bits(unsigned long size)
4643{
4644 return ffz(~size);
4645}
4646
4647/*
4648 * Initially all pages are reserved - free ones are freed
4649 * up by free_all_bootmem() once the early boot process is
4650 * done. Non-atomic initialization, single-pass.
4651 */
4652void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4653 unsigned long start_pfn, enum memmap_context context)
4654{
4655 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
4656 unsigned long end_pfn = start_pfn + size;
4657 pg_data_t *pgdat = NODE_DATA(nid);
4658 unsigned long pfn;
4659 unsigned long nr_initialised = 0;
4660#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4661 struct memblock_region *r = NULL, *tmp;
4662#endif
4663
4664 if (highest_memmap_pfn < end_pfn - 1)
4665 highest_memmap_pfn = end_pfn - 1;
4666
4667 /*
4668 * Honor reservation requested by the driver for this ZONE_DEVICE
4669 * memory
4670 */
4671 if (altmap && start_pfn == altmap->base_pfn)
4672 start_pfn += altmap->reserve;
4673
4674 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4675 /*
4676 * There can be holes in boot-time mem_map[]s handed to this
4677 * function. They do not exist on hotplugged memory.
4678 */
4679 if (context != MEMMAP_EARLY)
4680 goto not_early;
4681
4682 if (!early_pfn_valid(pfn))
4683 continue;
4684 if (!early_pfn_in_nid(pfn, nid))
4685 continue;
4686 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
4687 break;
4688
4689#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4690 /*
4691 * If not mirrored_kernelcore and ZONE_MOVABLE exists, range
4692 * from zone_movable_pfn[nid] to end of each node should be
4693 * ZONE_MOVABLE not ZONE_NORMAL. skip it.
4694 */
4695 if (!mirrored_kernelcore && zone_movable_pfn[nid])
4696 if (zone == ZONE_NORMAL && pfn >= zone_movable_pfn[nid])
4697 continue;
4698
4699 /*
4700 * Check given memblock attribute by firmware which can affect
4701 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
4702 * mirrored, it's an overlapped memmap init. skip it.
4703 */
4704 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
4705 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
4706 for_each_memblock(memory, tmp)
4707 if (pfn < memblock_region_memory_end_pfn(tmp))
4708 break;
4709 r = tmp;
4710 }
4711 if (pfn >= memblock_region_memory_base_pfn(r) &&
4712 memblock_is_mirror(r)) {
4713 /* already initialized as NORMAL */
4714 pfn = memblock_region_memory_end_pfn(r);
4715 continue;
4716 }
4717 }
4718#endif
4719
4720not_early:
4721 /*
4722 * Mark the block movable so that blocks are reserved for
4723 * movable at startup. This will force kernel allocations
4724 * to reserve their blocks rather than leaking throughout
4725 * the address space during boot when many long-lived
4726 * kernel allocations are made.
4727 *
4728 * bitmap is created for zone's valid pfn range. but memmap
4729 * can be created for invalid pages (for alignment)
4730 * check here not to call set_pageblock_migratetype() against
4731 * pfn out of zone.
4732 */
4733 if (!(pfn & (pageblock_nr_pages - 1))) {
4734 struct page *page = pfn_to_page(pfn);
4735
4736 __init_single_page(page, pfn, zone, nid);
4737 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4738 } else {
4739 __init_single_pfn(pfn, zone, nid);
4740 }
4741 }
4742}
4743
4744static void __meminit zone_init_free_lists(struct zone *zone)
4745{
4746 unsigned int order, t;
4747 for_each_migratetype_order(order, t) {
4748 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4749 zone->free_area[order].nr_free = 0;
4750 }
4751}
4752
4753#ifndef __HAVE_ARCH_MEMMAP_INIT
4754#define memmap_init(size, nid, zone, start_pfn) \
4755 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4756#endif
4757
4758static int zone_batchsize(struct zone *zone)
4759{
4760#ifdef CONFIG_MMU
4761 int batch;
4762
4763 /*
4764 * The per-cpu-pages pools are set to around 1000th of the
4765 * size of the zone. But no more than 1/2 of a meg.
4766 *
4767 * OK, so we don't know how big the cache is. So guess.
4768 */
4769 batch = zone->managed_pages / 1024;
4770 if (batch * PAGE_SIZE > 512 * 1024)
4771 batch = (512 * 1024) / PAGE_SIZE;
4772 batch /= 4; /* We effectively *= 4 below */
4773 if (batch < 1)
4774 batch = 1;
4775
4776 /*
4777 * Clamp the batch to a 2^n - 1 value. Having a power
4778 * of 2 value was found to be more likely to have
4779 * suboptimal cache aliasing properties in some cases.
4780 *
4781 * For example if 2 tasks are alternately allocating
4782 * batches of pages, one task can end up with a lot
4783 * of pages of one half of the possible page colors
4784 * and the other with pages of the other colors.
4785 */
4786 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4787
4788 return batch;
4789
4790#else
4791 /* The deferral and batching of frees should be suppressed under NOMMU
4792 * conditions.
4793 *
4794 * The problem is that NOMMU needs to be able to allocate large chunks
4795 * of contiguous memory as there's no hardware page translation to
4796 * assemble apparent contiguous memory from discontiguous pages.
4797 *
4798 * Queueing large contiguous runs of pages for batching, however,
4799 * causes the pages to actually be freed in smaller chunks. As there
4800 * can be a significant delay between the individual batches being
4801 * recycled, this leads to the once large chunks of space being
4802 * fragmented and becoming unavailable for high-order allocations.
4803 */
4804 return 0;
4805#endif
4806}
4807
4808/*
4809 * pcp->high and pcp->batch values are related and dependent on one another:
4810 * ->batch must never be higher then ->high.
4811 * The following function updates them in a safe manner without read side
4812 * locking.
4813 *
4814 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4815 * those fields changing asynchronously (acording the the above rule).
4816 *
4817 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4818 * outside of boot time (or some other assurance that no concurrent updaters
4819 * exist).
4820 */
4821static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4822 unsigned long batch)
4823{
4824 /* start with a fail safe value for batch */
4825 pcp->batch = 1;
4826 smp_wmb();
4827
4828 /* Update high, then batch, in order */
4829 pcp->high = high;
4830 smp_wmb();
4831
4832 pcp->batch = batch;
4833}
4834
4835/* a companion to pageset_set_high() */
4836static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4837{
4838 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4839}
4840
4841static void pageset_init(struct per_cpu_pageset *p)
4842{
4843 struct per_cpu_pages *pcp;
4844 int migratetype;
4845
4846 memset(p, 0, sizeof(*p));
4847
4848 pcp = &p->pcp;
4849 pcp->count = 0;
4850 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4851 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4852}
4853
4854static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4855{
4856 pageset_init(p);
4857 pageset_set_batch(p, batch);
4858}
4859
4860/*
4861 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4862 * to the value high for the pageset p.
4863 */
4864static void pageset_set_high(struct per_cpu_pageset *p,
4865 unsigned long high)
4866{
4867 unsigned long batch = max(1UL, high / 4);
4868 if ((high / 4) > (PAGE_SHIFT * 8))
4869 batch = PAGE_SHIFT * 8;
4870
4871 pageset_update(&p->pcp, high, batch);
4872}
4873
4874static void pageset_set_high_and_batch(struct zone *zone,
4875 struct per_cpu_pageset *pcp)
4876{
4877 if (percpu_pagelist_fraction)
4878 pageset_set_high(pcp,
4879 (zone->managed_pages /
4880 percpu_pagelist_fraction));
4881 else
4882 pageset_set_batch(pcp, zone_batchsize(zone));
4883}
4884
4885static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4886{
4887 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4888
4889 pageset_init(pcp);
4890 pageset_set_high_and_batch(zone, pcp);
4891}
4892
4893static void __meminit setup_zone_pageset(struct zone *zone)
4894{
4895 int cpu;
4896 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4897 for_each_possible_cpu(cpu)
4898 zone_pageset_init(zone, cpu);
4899}
4900
4901/*
4902 * Allocate per cpu pagesets and initialize them.
4903 * Before this call only boot pagesets were available.
4904 */
4905void __init setup_per_cpu_pageset(void)
4906{
4907 struct zone *zone;
4908
4909 for_each_populated_zone(zone)
4910 setup_zone_pageset(zone);
4911}
4912
4913static noinline __init_refok
4914int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4915{
4916 int i;
4917 size_t alloc_size;
4918
4919 /*
4920 * The per-page waitqueue mechanism uses hashed waitqueues
4921 * per zone.
4922 */
4923 zone->wait_table_hash_nr_entries =
4924 wait_table_hash_nr_entries(zone_size_pages);
4925 zone->wait_table_bits =
4926 wait_table_bits(zone->wait_table_hash_nr_entries);
4927 alloc_size = zone->wait_table_hash_nr_entries
4928 * sizeof(wait_queue_head_t);
4929
4930 if (!slab_is_available()) {
4931 zone->wait_table = (wait_queue_head_t *)
4932 memblock_virt_alloc_node_nopanic(
4933 alloc_size, zone->zone_pgdat->node_id);
4934 } else {
4935 /*
4936 * This case means that a zone whose size was 0 gets new memory
4937 * via memory hot-add.
4938 * But it may be the case that a new node was hot-added. In
4939 * this case vmalloc() will not be able to use this new node's
4940 * memory - this wait_table must be initialized to use this new
4941 * node itself as well.
4942 * To use this new node's memory, further consideration will be
4943 * necessary.
4944 */
4945 zone->wait_table = vmalloc(alloc_size);
4946 }
4947 if (!zone->wait_table)
4948 return -ENOMEM;
4949
4950 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4951 init_waitqueue_head(zone->wait_table + i);
4952
4953 return 0;
4954}
4955
4956static __meminit void zone_pcp_init(struct zone *zone)
4957{
4958 /*
4959 * per cpu subsystem is not up at this point. The following code
4960 * relies on the ability of the linker to provide the
4961 * offset of a (static) per cpu variable into the per cpu area.
4962 */
4963 zone->pageset = &boot_pageset;
4964
4965 if (populated_zone(zone))
4966 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4967 zone->name, zone->present_pages,
4968 zone_batchsize(zone));
4969}
4970
4971int __meminit init_currently_empty_zone(struct zone *zone,
4972 unsigned long zone_start_pfn,
4973 unsigned long size)
4974{
4975 struct pglist_data *pgdat = zone->zone_pgdat;
4976 int ret;
4977 ret = zone_wait_table_init(zone, size);
4978 if (ret)
4979 return ret;
4980 pgdat->nr_zones = zone_idx(zone) + 1;
4981
4982 zone->zone_start_pfn = zone_start_pfn;
4983
4984 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4985 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4986 pgdat->node_id,
4987 (unsigned long)zone_idx(zone),
4988 zone_start_pfn, (zone_start_pfn + size));
4989
4990 zone_init_free_lists(zone);
4991
4992 return 0;
4993}
4994
4995#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4996#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4997
4998/*
4999 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5000 */
5001int __meminit __early_pfn_to_nid(unsigned long pfn,
5002 struct mminit_pfnnid_cache *state)
5003{
5004 unsigned long start_pfn, end_pfn;
5005 int nid;
5006
5007 if (state->last_start <= pfn && pfn < state->last_end)
5008 return state->last_nid;
5009
5010 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5011 if (nid != -1) {
5012 state->last_start = start_pfn;
5013 state->last_end = end_pfn;
5014 state->last_nid = nid;
5015 }
5016
5017 return nid;
5018}
5019#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5020
5021/**
5022 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5023 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5024 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5025 *
5026 * If an architecture guarantees that all ranges registered contain no holes
5027 * and may be freed, this this function may be used instead of calling
5028 * memblock_free_early_nid() manually.
5029 */
5030void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5031{
5032 unsigned long start_pfn, end_pfn;
5033 int i, this_nid;
5034
5035 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5036 start_pfn = min(start_pfn, max_low_pfn);
5037 end_pfn = min(end_pfn, max_low_pfn);
5038
5039 if (start_pfn < end_pfn)
5040 memblock_free_early_nid(PFN_PHYS(start_pfn),
5041 (end_pfn - start_pfn) << PAGE_SHIFT,
5042 this_nid);
5043 }
5044}
5045
5046/**
5047 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5048 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5049 *
5050 * If an architecture guarantees that all ranges registered contain no holes and may
5051 * be freed, this function may be used instead of calling memory_present() manually.
5052 */
5053void __init sparse_memory_present_with_active_regions(int nid)
5054{
5055 unsigned long start_pfn, end_pfn;
5056 int i, this_nid;
5057
5058 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5059 memory_present(this_nid, start_pfn, end_pfn);
5060}
5061
5062/**
5063 * get_pfn_range_for_nid - Return the start and end page frames for a node
5064 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5065 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5066 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5067 *
5068 * It returns the start and end page frame of a node based on information
5069 * provided by memblock_set_node(). If called for a node
5070 * with no available memory, a warning is printed and the start and end
5071 * PFNs will be 0.
5072 */
5073void __meminit get_pfn_range_for_nid(unsigned int nid,
5074 unsigned long *start_pfn, unsigned long *end_pfn)
5075{
5076 unsigned long this_start_pfn, this_end_pfn;
5077 int i;
5078
5079 *start_pfn = -1UL;
5080 *end_pfn = 0;
5081
5082 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5083 *start_pfn = min(*start_pfn, this_start_pfn);
5084 *end_pfn = max(*end_pfn, this_end_pfn);
5085 }
5086
5087 if (*start_pfn == -1UL)
5088 *start_pfn = 0;
5089}
5090
5091/*
5092 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5093 * assumption is made that zones within a node are ordered in monotonic
5094 * increasing memory addresses so that the "highest" populated zone is used
5095 */
5096static void __init find_usable_zone_for_movable(void)
5097{
5098 int zone_index;
5099 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5100 if (zone_index == ZONE_MOVABLE)
5101 continue;
5102
5103 if (arch_zone_highest_possible_pfn[zone_index] >
5104 arch_zone_lowest_possible_pfn[zone_index])
5105 break;
5106 }
5107
5108 VM_BUG_ON(zone_index == -1);
5109 movable_zone = zone_index;
5110}
5111
5112/*
5113 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5114 * because it is sized independent of architecture. Unlike the other zones,
5115 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5116 * in each node depending on the size of each node and how evenly kernelcore
5117 * is distributed. This helper function adjusts the zone ranges
5118 * provided by the architecture for a given node by using the end of the
5119 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5120 * zones within a node are in order of monotonic increases memory addresses
5121 */
5122static void __meminit adjust_zone_range_for_zone_movable(int nid,
5123 unsigned long zone_type,
5124 unsigned long node_start_pfn,
5125 unsigned long node_end_pfn,
5126 unsigned long *zone_start_pfn,
5127 unsigned long *zone_end_pfn)
5128{
5129 /* Only adjust if ZONE_MOVABLE is on this node */
5130 if (zone_movable_pfn[nid]) {
5131 /* Size ZONE_MOVABLE */
5132 if (zone_type == ZONE_MOVABLE) {
5133 *zone_start_pfn = zone_movable_pfn[nid];
5134 *zone_end_pfn = min(node_end_pfn,
5135 arch_zone_highest_possible_pfn[movable_zone]);
5136
5137 /* Check if this whole range is within ZONE_MOVABLE */
5138 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5139 *zone_start_pfn = *zone_end_pfn;
5140 }
5141}
5142
5143/*
5144 * Return the number of pages a zone spans in a node, including holes
5145 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5146 */
5147static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5148 unsigned long zone_type,
5149 unsigned long node_start_pfn,
5150 unsigned long node_end_pfn,
5151 unsigned long *zone_start_pfn,
5152 unsigned long *zone_end_pfn,
5153 unsigned long *ignored)
5154{
5155 /* When hotadd a new node from cpu_up(), the node should be empty */
5156 if (!node_start_pfn && !node_end_pfn)
5157 return 0;
5158
5159 /* Get the start and end of the zone */
5160 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5161 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5162 adjust_zone_range_for_zone_movable(nid, zone_type,
5163 node_start_pfn, node_end_pfn,
5164 zone_start_pfn, zone_end_pfn);
5165
5166 /* Check that this node has pages within the zone's required range */
5167 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5168 return 0;
5169
5170 /* Move the zone boundaries inside the node if necessary */
5171 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5172 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5173
5174 /* Return the spanned pages */
5175 return *zone_end_pfn - *zone_start_pfn;
5176}
5177
5178/*
5179 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5180 * then all holes in the requested range will be accounted for.
5181 */
5182unsigned long __meminit __absent_pages_in_range(int nid,
5183 unsigned long range_start_pfn,
5184 unsigned long range_end_pfn)
5185{
5186 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5187 unsigned long start_pfn, end_pfn;
5188 int i;
5189
5190 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5191 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5192 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5193 nr_absent -= end_pfn - start_pfn;
5194 }
5195 return nr_absent;
5196}
5197
5198/**
5199 * absent_pages_in_range - Return number of page frames in holes within a range
5200 * @start_pfn: The start PFN to start searching for holes
5201 * @end_pfn: The end PFN to stop searching for holes
5202 *
5203 * It returns the number of pages frames in memory holes within a range.
5204 */
5205unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5206 unsigned long end_pfn)
5207{
5208 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5209}
5210
5211/* Return the number of page frames in holes in a zone on a node */
5212static unsigned long __meminit zone_absent_pages_in_node(int nid,
5213 unsigned long zone_type,
5214 unsigned long node_start_pfn,
5215 unsigned long node_end_pfn,
5216 unsigned long *ignored)
5217{
5218 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5219 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5220 unsigned long zone_start_pfn, zone_end_pfn;
5221 unsigned long nr_absent;
5222
5223 /* When hotadd a new node from cpu_up(), the node should be empty */
5224 if (!node_start_pfn && !node_end_pfn)
5225 return 0;
5226
5227 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5228 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5229
5230 adjust_zone_range_for_zone_movable(nid, zone_type,
5231 node_start_pfn, node_end_pfn,
5232 &zone_start_pfn, &zone_end_pfn);
5233 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5234
5235 /*
5236 * ZONE_MOVABLE handling.
5237 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5238 * and vice versa.
5239 */
5240 if (zone_movable_pfn[nid]) {
5241 if (mirrored_kernelcore) {
5242 unsigned long start_pfn, end_pfn;
5243 struct memblock_region *r;
5244
5245 for_each_memblock(memory, r) {
5246 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5247 zone_start_pfn, zone_end_pfn);
5248 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5249 zone_start_pfn, zone_end_pfn);
5250
5251 if (zone_type == ZONE_MOVABLE &&
5252 memblock_is_mirror(r))
5253 nr_absent += end_pfn - start_pfn;
5254
5255 if (zone_type == ZONE_NORMAL &&
5256 !memblock_is_mirror(r))
5257 nr_absent += end_pfn - start_pfn;
5258 }
5259 } else {
5260 if (zone_type == ZONE_NORMAL)
5261 nr_absent += node_end_pfn - zone_movable_pfn[nid];
5262 }
5263 }
5264
5265 return nr_absent;
5266}
5267
5268#else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5269static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5270 unsigned long zone_type,
5271 unsigned long node_start_pfn,
5272 unsigned long node_end_pfn,
5273 unsigned long *zone_start_pfn,
5274 unsigned long *zone_end_pfn,
5275 unsigned long *zones_size)
5276{
5277 unsigned int zone;
5278
5279 *zone_start_pfn = node_start_pfn;
5280 for (zone = 0; zone < zone_type; zone++)
5281 *zone_start_pfn += zones_size[zone];
5282
5283 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5284
5285 return zones_size[zone_type];
5286}
5287
5288static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5289 unsigned long zone_type,
5290 unsigned long node_start_pfn,
5291 unsigned long node_end_pfn,
5292 unsigned long *zholes_size)
5293{
5294 if (!zholes_size)
5295 return 0;
5296
5297 return zholes_size[zone_type];
5298}
5299
5300#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5301
5302static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5303 unsigned long node_start_pfn,
5304 unsigned long node_end_pfn,
5305 unsigned long *zones_size,
5306 unsigned long *zholes_size)
5307{
5308 unsigned long realtotalpages = 0, totalpages = 0;
5309 enum zone_type i;
5310
5311 for (i = 0; i < MAX_NR_ZONES; i++) {
5312 struct zone *zone = pgdat->node_zones + i;
5313 unsigned long zone_start_pfn, zone_end_pfn;
5314 unsigned long size, real_size;
5315
5316 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5317 node_start_pfn,
5318 node_end_pfn,
5319 &zone_start_pfn,
5320 &zone_end_pfn,
5321 zones_size);
5322 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5323 node_start_pfn, node_end_pfn,
5324 zholes_size);
5325 if (size)
5326 zone->zone_start_pfn = zone_start_pfn;
5327 else
5328 zone->zone_start_pfn = 0;
5329 zone->spanned_pages = size;
5330 zone->present_pages = real_size;
5331
5332 totalpages += size;
5333 realtotalpages += real_size;
5334 }
5335
5336 pgdat->node_spanned_pages = totalpages;
5337 pgdat->node_present_pages = realtotalpages;
5338 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5339 realtotalpages);
5340}
5341
5342#ifndef CONFIG_SPARSEMEM
5343/*
5344 * Calculate the size of the zone->blockflags rounded to an unsigned long
5345 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5346 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5347 * round what is now in bits to nearest long in bits, then return it in
5348 * bytes.
5349 */
5350static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5351{
5352 unsigned long usemapsize;
5353
5354 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5355 usemapsize = roundup(zonesize, pageblock_nr_pages);
5356 usemapsize = usemapsize >> pageblock_order;
5357 usemapsize *= NR_PAGEBLOCK_BITS;
5358 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5359
5360 return usemapsize / 8;
5361}
5362
5363static void __init setup_usemap(struct pglist_data *pgdat,
5364 struct zone *zone,
5365 unsigned long zone_start_pfn,
5366 unsigned long zonesize)
5367{
5368 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5369 zone->pageblock_flags = NULL;
5370 if (usemapsize)
5371 zone->pageblock_flags =
5372 memblock_virt_alloc_node_nopanic(usemapsize,
5373 pgdat->node_id);
5374}
5375#else
5376static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5377 unsigned long zone_start_pfn, unsigned long zonesize) {}
5378#endif /* CONFIG_SPARSEMEM */
5379
5380#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5381
5382/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5383void __paginginit set_pageblock_order(void)
5384{
5385 unsigned int order;
5386
5387 /* Check that pageblock_nr_pages has not already been setup */
5388 if (pageblock_order)
5389 return;
5390
5391 if (HPAGE_SHIFT > PAGE_SHIFT)
5392 order = HUGETLB_PAGE_ORDER;
5393 else
5394 order = MAX_ORDER - 1;
5395
5396 /*
5397 * Assume the largest contiguous order of interest is a huge page.
5398 * This value may be variable depending on boot parameters on IA64 and
5399 * powerpc.
5400 */
5401 pageblock_order = order;
5402}
5403#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5404
5405/*
5406 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5407 * is unused as pageblock_order is set at compile-time. See
5408 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5409 * the kernel config
5410 */
5411void __paginginit set_pageblock_order(void)
5412{
5413}
5414
5415#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5416
5417static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5418 unsigned long present_pages)
5419{
5420 unsigned long pages = spanned_pages;
5421
5422 /*
5423 * Provide a more accurate estimation if there are holes within
5424 * the zone and SPARSEMEM is in use. If there are holes within the
5425 * zone, each populated memory region may cost us one or two extra
5426 * memmap pages due to alignment because memmap pages for each
5427 * populated regions may not naturally algined on page boundary.
5428 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5429 */
5430 if (spanned_pages > present_pages + (present_pages >> 4) &&
5431 IS_ENABLED(CONFIG_SPARSEMEM))
5432 pages = present_pages;
5433
5434 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5435}
5436
5437/*
5438 * Set up the zone data structures:
5439 * - mark all pages reserved
5440 * - mark all memory queues empty
5441 * - clear the memory bitmaps
5442 *
5443 * NOTE: pgdat should get zeroed by caller.
5444 */
5445static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5446{
5447 enum zone_type j;
5448 int nid = pgdat->node_id;
5449 int ret;
5450
5451 pgdat_resize_init(pgdat);
5452#ifdef CONFIG_NUMA_BALANCING
5453 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5454 pgdat->numabalancing_migrate_nr_pages = 0;
5455 pgdat->numabalancing_migrate_next_window = jiffies;
5456#endif
5457#ifdef CONFIG_TRANSPARENT_HUGEPAGE
5458 spin_lock_init(&pgdat->split_queue_lock);
5459 INIT_LIST_HEAD(&pgdat->split_queue);
5460 pgdat->split_queue_len = 0;
5461#endif
5462 init_waitqueue_head(&pgdat->kswapd_wait);
5463 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5464#ifdef CONFIG_COMPACTION
5465 init_waitqueue_head(&pgdat->kcompactd_wait);
5466#endif
5467 pgdat_page_ext_init(pgdat);
5468
5469 for (j = 0; j < MAX_NR_ZONES; j++) {
5470 struct zone *zone = pgdat->node_zones + j;
5471 unsigned long size, realsize, freesize, memmap_pages;
5472 unsigned long zone_start_pfn = zone->zone_start_pfn;
5473
5474 size = zone->spanned_pages;
5475 realsize = freesize = zone->present_pages;
5476
5477 /*
5478 * Adjust freesize so that it accounts for how much memory
5479 * is used by this zone for memmap. This affects the watermark
5480 * and per-cpu initialisations
5481 */
5482 memmap_pages = calc_memmap_size(size, realsize);
5483 if (!is_highmem_idx(j)) {
5484 if (freesize >= memmap_pages) {
5485 freesize -= memmap_pages;
5486 if (memmap_pages)
5487 printk(KERN_DEBUG
5488 " %s zone: %lu pages used for memmap\n",
5489 zone_names[j], memmap_pages);
5490 } else
5491 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5492 zone_names[j], memmap_pages, freesize);
5493 }
5494
5495 /* Account for reserved pages */
5496 if (j == 0 && freesize > dma_reserve) {
5497 freesize -= dma_reserve;
5498 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5499 zone_names[0], dma_reserve);
5500 }
5501
5502 if (!is_highmem_idx(j))
5503 nr_kernel_pages += freesize;
5504 /* Charge for highmem memmap if there are enough kernel pages */
5505 else if (nr_kernel_pages > memmap_pages * 2)
5506 nr_kernel_pages -= memmap_pages;
5507 nr_all_pages += freesize;
5508
5509 /*
5510 * Set an approximate value for lowmem here, it will be adjusted
5511 * when the bootmem allocator frees pages into the buddy system.
5512 * And all highmem pages will be managed by the buddy system.
5513 */
5514 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5515#ifdef CONFIG_NUMA
5516 zone->node = nid;
5517 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5518 / 100;
5519 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5520#endif
5521 zone->name = zone_names[j];
5522 spin_lock_init(&zone->lock);
5523 spin_lock_init(&zone->lru_lock);
5524 zone_seqlock_init(zone);
5525 zone->zone_pgdat = pgdat;
5526 zone_pcp_init(zone);
5527
5528 /* For bootup, initialized properly in watermark setup */
5529 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5530
5531 lruvec_init(&zone->lruvec);
5532 if (!size)
5533 continue;
5534
5535 set_pageblock_order();
5536 setup_usemap(pgdat, zone, zone_start_pfn, size);
5537 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5538 BUG_ON(ret);
5539 memmap_init(size, nid, j, zone_start_pfn);
5540 }
5541}
5542
5543static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5544{
5545 unsigned long __maybe_unused start = 0;
5546 unsigned long __maybe_unused offset = 0;
5547
5548 /* Skip empty nodes */
5549 if (!pgdat->node_spanned_pages)
5550 return;
5551
5552#ifdef CONFIG_FLAT_NODE_MEM_MAP
5553 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5554 offset = pgdat->node_start_pfn - start;
5555 /* ia64 gets its own node_mem_map, before this, without bootmem */
5556 if (!pgdat->node_mem_map) {
5557 unsigned long size, end;
5558 struct page *map;
5559
5560 /*
5561 * The zone's endpoints aren't required to be MAX_ORDER
5562 * aligned but the node_mem_map endpoints must be in order
5563 * for the buddy allocator to function correctly.
5564 */
5565 end = pgdat_end_pfn(pgdat);
5566 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5567 size = (end - start) * sizeof(struct page);
5568 map = alloc_remap(pgdat->node_id, size);
5569 if (!map)
5570 map = memblock_virt_alloc_node_nopanic(size,
5571 pgdat->node_id);
5572 pgdat->node_mem_map = map + offset;
5573 }
5574#ifndef CONFIG_NEED_MULTIPLE_NODES
5575 /*
5576 * With no DISCONTIG, the global mem_map is just set as node 0's
5577 */
5578 if (pgdat == NODE_DATA(0)) {
5579 mem_map = NODE_DATA(0)->node_mem_map;
5580#if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5581 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5582 mem_map -= offset;
5583#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5584 }
5585#endif
5586#endif /* CONFIG_FLAT_NODE_MEM_MAP */
5587}
5588
5589void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5590 unsigned long node_start_pfn, unsigned long *zholes_size)
5591{
5592 pg_data_t *pgdat = NODE_DATA(nid);
5593 unsigned long start_pfn = 0;
5594 unsigned long end_pfn = 0;
5595
5596 /* pg_data_t should be reset to zero when it's allocated */
5597 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5598
5599 reset_deferred_meminit(pgdat);
5600 pgdat->node_id = nid;
5601 pgdat->node_start_pfn = node_start_pfn;
5602#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5603 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5604 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5605 (u64)start_pfn << PAGE_SHIFT,
5606 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5607#else
5608 start_pfn = node_start_pfn;
5609#endif
5610 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5611 zones_size, zholes_size);
5612
5613 alloc_node_mem_map(pgdat);
5614#ifdef CONFIG_FLAT_NODE_MEM_MAP
5615 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5616 nid, (unsigned long)pgdat,
5617 (unsigned long)pgdat->node_mem_map);
5618#endif
5619
5620 free_area_init_core(pgdat);
5621}
5622
5623#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5624
5625#if MAX_NUMNODES > 1
5626/*
5627 * Figure out the number of possible node ids.
5628 */
5629void __init setup_nr_node_ids(void)
5630{
5631 unsigned int highest;
5632
5633 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5634 nr_node_ids = highest + 1;
5635}
5636#endif
5637
5638/**
5639 * node_map_pfn_alignment - determine the maximum internode alignment
5640 *
5641 * This function should be called after node map is populated and sorted.
5642 * It calculates the maximum power of two alignment which can distinguish
5643 * all the nodes.
5644 *
5645 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5646 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5647 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5648 * shifted, 1GiB is enough and this function will indicate so.
5649 *
5650 * This is used to test whether pfn -> nid mapping of the chosen memory
5651 * model has fine enough granularity to avoid incorrect mapping for the
5652 * populated node map.
5653 *
5654 * Returns the determined alignment in pfn's. 0 if there is no alignment
5655 * requirement (single node).
5656 */
5657unsigned long __init node_map_pfn_alignment(void)
5658{
5659 unsigned long accl_mask = 0, last_end = 0;
5660 unsigned long start, end, mask;
5661 int last_nid = -1;
5662 int i, nid;
5663
5664 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5665 if (!start || last_nid < 0 || last_nid == nid) {
5666 last_nid = nid;
5667 last_end = end;
5668 continue;
5669 }
5670
5671 /*
5672 * Start with a mask granular enough to pin-point to the
5673 * start pfn and tick off bits one-by-one until it becomes
5674 * too coarse to separate the current node from the last.
5675 */
5676 mask = ~((1 << __ffs(start)) - 1);
5677 while (mask && last_end <= (start & (mask << 1)))
5678 mask <<= 1;
5679
5680 /* accumulate all internode masks */
5681 accl_mask |= mask;
5682 }
5683
5684 /* convert mask to number of pages */
5685 return ~accl_mask + 1;
5686}
5687
5688/* Find the lowest pfn for a node */
5689static unsigned long __init find_min_pfn_for_node(int nid)
5690{
5691 unsigned long min_pfn = ULONG_MAX;
5692 unsigned long start_pfn;
5693 int i;
5694
5695 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5696 min_pfn = min(min_pfn, start_pfn);
5697
5698 if (min_pfn == ULONG_MAX) {
5699 pr_warn("Could not find start_pfn for node %d\n", nid);
5700 return 0;
5701 }
5702
5703 return min_pfn;
5704}
5705
5706/**
5707 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5708 *
5709 * It returns the minimum PFN based on information provided via
5710 * memblock_set_node().
5711 */
5712unsigned long __init find_min_pfn_with_active_regions(void)
5713{
5714 return find_min_pfn_for_node(MAX_NUMNODES);
5715}
5716
5717/*
5718 * early_calculate_totalpages()
5719 * Sum pages in active regions for movable zone.
5720 * Populate N_MEMORY for calculating usable_nodes.
5721 */
5722static unsigned long __init early_calculate_totalpages(void)
5723{
5724 unsigned long totalpages = 0;
5725 unsigned long start_pfn, end_pfn;
5726 int i, nid;
5727
5728 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5729 unsigned long pages = end_pfn - start_pfn;
5730
5731 totalpages += pages;
5732 if (pages)
5733 node_set_state(nid, N_MEMORY);
5734 }
5735 return totalpages;
5736}
5737
5738/*
5739 * Find the PFN the Movable zone begins in each node. Kernel memory
5740 * is spread evenly between nodes as long as the nodes have enough
5741 * memory. When they don't, some nodes will have more kernelcore than
5742 * others
5743 */
5744static void __init find_zone_movable_pfns_for_nodes(void)
5745{
5746 int i, nid;
5747 unsigned long usable_startpfn;
5748 unsigned long kernelcore_node, kernelcore_remaining;
5749 /* save the state before borrow the nodemask */
5750 nodemask_t saved_node_state = node_states[N_MEMORY];
5751 unsigned long totalpages = early_calculate_totalpages();
5752 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5753 struct memblock_region *r;
5754
5755 /* Need to find movable_zone earlier when movable_node is specified. */
5756 find_usable_zone_for_movable();
5757
5758 /*
5759 * If movable_node is specified, ignore kernelcore and movablecore
5760 * options.
5761 */
5762 if (movable_node_is_enabled()) {
5763 for_each_memblock(memory, r) {
5764 if (!memblock_is_hotpluggable(r))
5765 continue;
5766
5767 nid = r->nid;
5768
5769 usable_startpfn = PFN_DOWN(r->base);
5770 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5771 min(usable_startpfn, zone_movable_pfn[nid]) :
5772 usable_startpfn;
5773 }
5774
5775 goto out2;
5776 }
5777
5778 /*
5779 * If kernelcore=mirror is specified, ignore movablecore option
5780 */
5781 if (mirrored_kernelcore) {
5782 bool mem_below_4gb_not_mirrored = false;
5783
5784 for_each_memblock(memory, r) {
5785 if (memblock_is_mirror(r))
5786 continue;
5787
5788 nid = r->nid;
5789
5790 usable_startpfn = memblock_region_memory_base_pfn(r);
5791
5792 if (usable_startpfn < 0x100000) {
5793 mem_below_4gb_not_mirrored = true;
5794 continue;
5795 }
5796
5797 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5798 min(usable_startpfn, zone_movable_pfn[nid]) :
5799 usable_startpfn;
5800 }
5801
5802 if (mem_below_4gb_not_mirrored)
5803 pr_warn("This configuration results in unmirrored kernel memory.");
5804
5805 goto out2;
5806 }
5807
5808 /*
5809 * If movablecore=nn[KMG] was specified, calculate what size of
5810 * kernelcore that corresponds so that memory usable for
5811 * any allocation type is evenly spread. If both kernelcore
5812 * and movablecore are specified, then the value of kernelcore
5813 * will be used for required_kernelcore if it's greater than
5814 * what movablecore would have allowed.
5815 */
5816 if (required_movablecore) {
5817 unsigned long corepages;
5818
5819 /*
5820 * Round-up so that ZONE_MOVABLE is at least as large as what
5821 * was requested by the user
5822 */
5823 required_movablecore =
5824 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5825 required_movablecore = min(totalpages, required_movablecore);
5826 corepages = totalpages - required_movablecore;
5827
5828 required_kernelcore = max(required_kernelcore, corepages);
5829 }
5830
5831 /*
5832 * If kernelcore was not specified or kernelcore size is larger
5833 * than totalpages, there is no ZONE_MOVABLE.
5834 */
5835 if (!required_kernelcore || required_kernelcore >= totalpages)
5836 goto out;
5837
5838 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5839 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5840
5841restart:
5842 /* Spread kernelcore memory as evenly as possible throughout nodes */
5843 kernelcore_node = required_kernelcore / usable_nodes;
5844 for_each_node_state(nid, N_MEMORY) {
5845 unsigned long start_pfn, end_pfn;
5846
5847 /*
5848 * Recalculate kernelcore_node if the division per node
5849 * now exceeds what is necessary to satisfy the requested
5850 * amount of memory for the kernel
5851 */
5852 if (required_kernelcore < kernelcore_node)
5853 kernelcore_node = required_kernelcore / usable_nodes;
5854
5855 /*
5856 * As the map is walked, we track how much memory is usable
5857 * by the kernel using kernelcore_remaining. When it is
5858 * 0, the rest of the node is usable by ZONE_MOVABLE
5859 */
5860 kernelcore_remaining = kernelcore_node;
5861
5862 /* Go through each range of PFNs within this node */
5863 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5864 unsigned long size_pages;
5865
5866 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5867 if (start_pfn >= end_pfn)
5868 continue;
5869
5870 /* Account for what is only usable for kernelcore */
5871 if (start_pfn < usable_startpfn) {
5872 unsigned long kernel_pages;
5873 kernel_pages = min(end_pfn, usable_startpfn)
5874 - start_pfn;
5875
5876 kernelcore_remaining -= min(kernel_pages,
5877 kernelcore_remaining);
5878 required_kernelcore -= min(kernel_pages,
5879 required_kernelcore);
5880
5881 /* Continue if range is now fully accounted */
5882 if (end_pfn <= usable_startpfn) {
5883
5884 /*
5885 * Push zone_movable_pfn to the end so
5886 * that if we have to rebalance
5887 * kernelcore across nodes, we will
5888 * not double account here
5889 */
5890 zone_movable_pfn[nid] = end_pfn;
5891 continue;
5892 }
5893 start_pfn = usable_startpfn;
5894 }
5895
5896 /*
5897 * The usable PFN range for ZONE_MOVABLE is from
5898 * start_pfn->end_pfn. Calculate size_pages as the
5899 * number of pages used as kernelcore
5900 */
5901 size_pages = end_pfn - start_pfn;
5902 if (size_pages > kernelcore_remaining)
5903 size_pages = kernelcore_remaining;
5904 zone_movable_pfn[nid] = start_pfn + size_pages;
5905
5906 /*
5907 * Some kernelcore has been met, update counts and
5908 * break if the kernelcore for this node has been
5909 * satisfied
5910 */
5911 required_kernelcore -= min(required_kernelcore,
5912 size_pages);
5913 kernelcore_remaining -= size_pages;
5914 if (!kernelcore_remaining)
5915 break;
5916 }
5917 }
5918
5919 /*
5920 * If there is still required_kernelcore, we do another pass with one
5921 * less node in the count. This will push zone_movable_pfn[nid] further
5922 * along on the nodes that still have memory until kernelcore is
5923 * satisfied
5924 */
5925 usable_nodes--;
5926 if (usable_nodes && required_kernelcore > usable_nodes)
5927 goto restart;
5928
5929out2:
5930 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5931 for (nid = 0; nid < MAX_NUMNODES; nid++)
5932 zone_movable_pfn[nid] =
5933 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5934
5935out:
5936 /* restore the node_state */
5937 node_states[N_MEMORY] = saved_node_state;
5938}
5939
5940/* Any regular or high memory on that node ? */
5941static void check_for_memory(pg_data_t *pgdat, int nid)
5942{
5943 enum zone_type zone_type;
5944
5945 if (N_MEMORY == N_NORMAL_MEMORY)
5946 return;
5947
5948 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5949 struct zone *zone = &pgdat->node_zones[zone_type];
5950 if (populated_zone(zone)) {
5951 node_set_state(nid, N_HIGH_MEMORY);
5952 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5953 zone_type <= ZONE_NORMAL)
5954 node_set_state(nid, N_NORMAL_MEMORY);
5955 break;
5956 }
5957 }
5958}
5959
5960/**
5961 * free_area_init_nodes - Initialise all pg_data_t and zone data
5962 * @max_zone_pfn: an array of max PFNs for each zone
5963 *
5964 * This will call free_area_init_node() for each active node in the system.
5965 * Using the page ranges provided by memblock_set_node(), the size of each
5966 * zone in each node and their holes is calculated. If the maximum PFN
5967 * between two adjacent zones match, it is assumed that the zone is empty.
5968 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5969 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5970 * starts where the previous one ended. For example, ZONE_DMA32 starts
5971 * at arch_max_dma_pfn.
5972 */
5973void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5974{
5975 unsigned long start_pfn, end_pfn;
5976 int i, nid;
5977
5978 /* Record where the zone boundaries are */
5979 memset(arch_zone_lowest_possible_pfn, 0,
5980 sizeof(arch_zone_lowest_possible_pfn));
5981 memset(arch_zone_highest_possible_pfn, 0,
5982 sizeof(arch_zone_highest_possible_pfn));
5983 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5984 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5985 for (i = 1; i < MAX_NR_ZONES; i++) {
5986 if (i == ZONE_MOVABLE)
5987 continue;
5988 arch_zone_lowest_possible_pfn[i] =
5989 arch_zone_highest_possible_pfn[i-1];
5990 arch_zone_highest_possible_pfn[i] =
5991 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5992 }
5993 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5994 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5995
5996 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5997 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5998 find_zone_movable_pfns_for_nodes();
5999
6000 /* Print out the zone ranges */
6001 pr_info("Zone ranges:\n");
6002 for (i = 0; i < MAX_NR_ZONES; i++) {
6003 if (i == ZONE_MOVABLE)
6004 continue;
6005 pr_info(" %-8s ", zone_names[i]);
6006 if (arch_zone_lowest_possible_pfn[i] ==
6007 arch_zone_highest_possible_pfn[i])
6008 pr_cont("empty\n");
6009 else
6010 pr_cont("[mem %#018Lx-%#018Lx]\n",
6011 (u64)arch_zone_lowest_possible_pfn[i]
6012 << PAGE_SHIFT,
6013 ((u64)arch_zone_highest_possible_pfn[i]
6014 << PAGE_SHIFT) - 1);
6015 }
6016
6017 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6018 pr_info("Movable zone start for each node\n");
6019 for (i = 0; i < MAX_NUMNODES; i++) {
6020 if (zone_movable_pfn[i])
6021 pr_info(" Node %d: %#018Lx\n", i,
6022 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6023 }
6024
6025 /* Print out the early node map */
6026 pr_info("Early memory node ranges\n");
6027 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6028 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6029 (u64)start_pfn << PAGE_SHIFT,
6030 ((u64)end_pfn << PAGE_SHIFT) - 1);
6031
6032 /* Initialise every node */
6033 mminit_verify_pageflags_layout();
6034 setup_nr_node_ids();
6035 for_each_online_node(nid) {
6036 pg_data_t *pgdat = NODE_DATA(nid);
6037 free_area_init_node(nid, NULL,
6038 find_min_pfn_for_node(nid), NULL);
6039
6040 /* Any memory on that node */
6041 if (pgdat->node_present_pages)
6042 node_set_state(nid, N_MEMORY);
6043 check_for_memory(pgdat, nid);
6044 }
6045}
6046
6047static int __init cmdline_parse_core(char *p, unsigned long *core)
6048{
6049 unsigned long long coremem;
6050 if (!p)
6051 return -EINVAL;
6052
6053 coremem = memparse(p, &p);
6054 *core = coremem >> PAGE_SHIFT;
6055
6056 /* Paranoid check that UL is enough for the coremem value */
6057 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6058
6059 return 0;
6060}
6061
6062/*
6063 * kernelcore=size sets the amount of memory for use for allocations that
6064 * cannot be reclaimed or migrated.
6065 */
6066static int __init cmdline_parse_kernelcore(char *p)
6067{
6068 /* parse kernelcore=mirror */
6069 if (parse_option_str(p, "mirror")) {
6070 mirrored_kernelcore = true;
6071 return 0;
6072 }
6073
6074 return cmdline_parse_core(p, &required_kernelcore);
6075}
6076
6077/*
6078 * movablecore=size sets the amount of memory for use for allocations that
6079 * can be reclaimed or migrated.
6080 */
6081static int __init cmdline_parse_movablecore(char *p)
6082{
6083 return cmdline_parse_core(p, &required_movablecore);
6084}
6085
6086early_param("kernelcore", cmdline_parse_kernelcore);
6087early_param("movablecore", cmdline_parse_movablecore);
6088
6089#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6090
6091void adjust_managed_page_count(struct page *page, long count)
6092{
6093 spin_lock(&managed_page_count_lock);
6094 page_zone(page)->managed_pages += count;
6095 totalram_pages += count;
6096#ifdef CONFIG_HIGHMEM
6097 if (PageHighMem(page))
6098 totalhigh_pages += count;
6099#endif
6100 spin_unlock(&managed_page_count_lock);
6101}
6102EXPORT_SYMBOL(adjust_managed_page_count);
6103
6104unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6105{
6106 void *pos;
6107 unsigned long pages = 0;
6108
6109 start = (void *)PAGE_ALIGN((unsigned long)start);
6110 end = (void *)((unsigned long)end & PAGE_MASK);
6111 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6112 if ((unsigned int)poison <= 0xFF)
6113 memset(pos, poison, PAGE_SIZE);
6114 free_reserved_page(virt_to_page(pos));
6115 }
6116
6117 if (pages && s)
6118 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
6119 s, pages << (PAGE_SHIFT - 10), start, end);
6120
6121 return pages;
6122}
6123EXPORT_SYMBOL(free_reserved_area);
6124
6125#ifdef CONFIG_HIGHMEM
6126void free_highmem_page(struct page *page)
6127{
6128 __free_reserved_page(page);
6129 totalram_pages++;
6130 page_zone(page)->managed_pages++;
6131 totalhigh_pages++;
6132}
6133#endif
6134
6135
6136void __init mem_init_print_info(const char *str)
6137{
6138 unsigned long physpages, codesize, datasize, rosize, bss_size;
6139 unsigned long init_code_size, init_data_size;
6140
6141 physpages = get_num_physpages();
6142 codesize = _etext - _stext;
6143 datasize = _edata - _sdata;
6144 rosize = __end_rodata - __start_rodata;
6145 bss_size = __bss_stop - __bss_start;
6146 init_data_size = __init_end - __init_begin;
6147 init_code_size = _einittext - _sinittext;
6148
6149 /*
6150 * Detect special cases and adjust section sizes accordingly:
6151 * 1) .init.* may be embedded into .data sections
6152 * 2) .init.text.* may be out of [__init_begin, __init_end],
6153 * please refer to arch/tile/kernel/vmlinux.lds.S.
6154 * 3) .rodata.* may be embedded into .text or .data sections.
6155 */
6156#define adj_init_size(start, end, size, pos, adj) \
6157 do { \
6158 if (start <= pos && pos < end && size > adj) \
6159 size -= adj; \
6160 } while (0)
6161
6162 adj_init_size(__init_begin, __init_end, init_data_size,
6163 _sinittext, init_code_size);
6164 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6165 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6166 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6167 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6168
6169#undef adj_init_size
6170
6171 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6172#ifdef CONFIG_HIGHMEM
6173 ", %luK highmem"
6174#endif
6175 "%s%s)\n",
6176 nr_free_pages() << (PAGE_SHIFT - 10),
6177 physpages << (PAGE_SHIFT - 10),
6178 codesize >> 10, datasize >> 10, rosize >> 10,
6179 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6180 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6181 totalcma_pages << (PAGE_SHIFT - 10),
6182#ifdef CONFIG_HIGHMEM
6183 totalhigh_pages << (PAGE_SHIFT - 10),
6184#endif
6185 str ? ", " : "", str ? str : "");
6186}
6187
6188/**
6189 * set_dma_reserve - set the specified number of pages reserved in the first zone
6190 * @new_dma_reserve: The number of pages to mark reserved
6191 *
6192 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6193 * In the DMA zone, a significant percentage may be consumed by kernel image
6194 * and other unfreeable allocations which can skew the watermarks badly. This
6195 * function may optionally be used to account for unfreeable pages in the
6196 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6197 * smaller per-cpu batchsize.
6198 */
6199void __init set_dma_reserve(unsigned long new_dma_reserve)
6200{
6201 dma_reserve = new_dma_reserve;
6202}
6203
6204void __init free_area_init(unsigned long *zones_size)
6205{
6206 free_area_init_node(0, zones_size,
6207 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6208}
6209
6210static int page_alloc_cpu_notify(struct notifier_block *self,
6211 unsigned long action, void *hcpu)
6212{
6213 int cpu = (unsigned long)hcpu;
6214
6215 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6216 lru_add_drain_cpu(cpu);
6217 drain_pages(cpu);
6218
6219 /*
6220 * Spill the event counters of the dead processor
6221 * into the current processors event counters.
6222 * This artificially elevates the count of the current
6223 * processor.
6224 */
6225 vm_events_fold_cpu(cpu);
6226
6227 /*
6228 * Zero the differential counters of the dead processor
6229 * so that the vm statistics are consistent.
6230 *
6231 * This is only okay since the processor is dead and cannot
6232 * race with what we are doing.
6233 */
6234 cpu_vm_stats_fold(cpu);
6235 }
6236 return NOTIFY_OK;
6237}
6238
6239void __init page_alloc_init(void)
6240{
6241 hotcpu_notifier(page_alloc_cpu_notify, 0);
6242}
6243
6244/*
6245 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6246 * or min_free_kbytes changes.
6247 */
6248static void calculate_totalreserve_pages(void)
6249{
6250 struct pglist_data *pgdat;
6251 unsigned long reserve_pages = 0;
6252 enum zone_type i, j;
6253
6254 for_each_online_pgdat(pgdat) {
6255 for (i = 0; i < MAX_NR_ZONES; i++) {
6256 struct zone *zone = pgdat->node_zones + i;
6257 long max = 0;
6258
6259 /* Find valid and maximum lowmem_reserve in the zone */
6260 for (j = i; j < MAX_NR_ZONES; j++) {
6261 if (zone->lowmem_reserve[j] > max)
6262 max = zone->lowmem_reserve[j];
6263 }
6264
6265 /* we treat the high watermark as reserved pages. */
6266 max += high_wmark_pages(zone);
6267
6268 if (max > zone->managed_pages)
6269 max = zone->managed_pages;
6270
6271 zone->totalreserve_pages = max;
6272
6273 reserve_pages += max;
6274 }
6275 }
6276 totalreserve_pages = reserve_pages;
6277}
6278
6279/*
6280 * setup_per_zone_lowmem_reserve - called whenever
6281 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6282 * has a correct pages reserved value, so an adequate number of
6283 * pages are left in the zone after a successful __alloc_pages().
6284 */
6285static void setup_per_zone_lowmem_reserve(void)
6286{
6287 struct pglist_data *pgdat;
6288 enum zone_type j, idx;
6289
6290 for_each_online_pgdat(pgdat) {
6291 for (j = 0; j < MAX_NR_ZONES; j++) {
6292 struct zone *zone = pgdat->node_zones + j;
6293 unsigned long managed_pages = zone->managed_pages;
6294
6295 zone->lowmem_reserve[j] = 0;
6296
6297 idx = j;
6298 while (idx) {
6299 struct zone *lower_zone;
6300
6301 idx--;
6302
6303 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6304 sysctl_lowmem_reserve_ratio[idx] = 1;
6305
6306 lower_zone = pgdat->node_zones + idx;
6307 lower_zone->lowmem_reserve[j] = managed_pages /
6308 sysctl_lowmem_reserve_ratio[idx];
6309 managed_pages += lower_zone->managed_pages;
6310 }
6311 }
6312 }
6313
6314 /* update totalreserve_pages */
6315 calculate_totalreserve_pages();
6316}
6317
6318static void __setup_per_zone_wmarks(void)
6319{
6320 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6321 unsigned long lowmem_pages = 0;
6322 struct zone *zone;
6323 unsigned long flags;
6324
6325 /* Calculate total number of !ZONE_HIGHMEM pages */
6326 for_each_zone(zone) {
6327 if (!is_highmem(zone))
6328 lowmem_pages += zone->managed_pages;
6329 }
6330
6331 for_each_zone(zone) {
6332 u64 tmp;
6333
6334 spin_lock_irqsave(&zone->lock, flags);
6335 tmp = (u64)pages_min * zone->managed_pages;
6336 do_div(tmp, lowmem_pages);
6337 if (is_highmem(zone)) {
6338 /*
6339 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6340 * need highmem pages, so cap pages_min to a small
6341 * value here.
6342 *
6343 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6344 * deltas control asynch page reclaim, and so should
6345 * not be capped for highmem.
6346 */
6347 unsigned long min_pages;
6348
6349 min_pages = zone->managed_pages / 1024;
6350 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6351 zone->watermark[WMARK_MIN] = min_pages;
6352 } else {
6353 /*
6354 * If it's a lowmem zone, reserve a number of pages
6355 * proportionate to the zone's size.
6356 */
6357 zone->watermark[WMARK_MIN] = tmp;
6358 }
6359
6360 /*
6361 * Set the kswapd watermarks distance according to the
6362 * scale factor in proportion to available memory, but
6363 * ensure a minimum size on small systems.
6364 */
6365 tmp = max_t(u64, tmp >> 2,
6366 mult_frac(zone->managed_pages,
6367 watermark_scale_factor, 10000));
6368
6369 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6370 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
6371
6372 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6373 high_wmark_pages(zone) - low_wmark_pages(zone) -
6374 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6375
6376 spin_unlock_irqrestore(&zone->lock, flags);
6377 }
6378
6379 /* update totalreserve_pages */
6380 calculate_totalreserve_pages();
6381}
6382
6383/**
6384 * setup_per_zone_wmarks - called when min_free_kbytes changes
6385 * or when memory is hot-{added|removed}
6386 *
6387 * Ensures that the watermark[min,low,high] values for each zone are set
6388 * correctly with respect to min_free_kbytes.
6389 */
6390void setup_per_zone_wmarks(void)
6391{
6392 mutex_lock(&zonelists_mutex);
6393 __setup_per_zone_wmarks();
6394 mutex_unlock(&zonelists_mutex);
6395}
6396
6397/*
6398 * The inactive anon list should be small enough that the VM never has to
6399 * do too much work, but large enough that each inactive page has a chance
6400 * to be referenced again before it is swapped out.
6401 *
6402 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6403 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6404 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6405 * the anonymous pages are kept on the inactive list.
6406 *
6407 * total target max
6408 * memory ratio inactive anon
6409 * -------------------------------------
6410 * 10MB 1 5MB
6411 * 100MB 1 50MB
6412 * 1GB 3 250MB
6413 * 10GB 10 0.9GB
6414 * 100GB 31 3GB
6415 * 1TB 101 10GB
6416 * 10TB 320 32GB
6417 */
6418static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6419{
6420 unsigned int gb, ratio;
6421
6422 /* Zone size in gigabytes */
6423 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6424 if (gb)
6425 ratio = int_sqrt(10 * gb);
6426 else
6427 ratio = 1;
6428
6429 zone->inactive_ratio = ratio;
6430}
6431
6432static void __meminit setup_per_zone_inactive_ratio(void)
6433{
6434 struct zone *zone;
6435
6436 for_each_zone(zone)
6437 calculate_zone_inactive_ratio(zone);
6438}
6439
6440/*
6441 * Initialise min_free_kbytes.
6442 *
6443 * For small machines we want it small (128k min). For large machines
6444 * we want it large (64MB max). But it is not linear, because network
6445 * bandwidth does not increase linearly with machine size. We use
6446 *
6447 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6448 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6449 *
6450 * which yields
6451 *
6452 * 16MB: 512k
6453 * 32MB: 724k
6454 * 64MB: 1024k
6455 * 128MB: 1448k
6456 * 256MB: 2048k
6457 * 512MB: 2896k
6458 * 1024MB: 4096k
6459 * 2048MB: 5792k
6460 * 4096MB: 8192k
6461 * 8192MB: 11584k
6462 * 16384MB: 16384k
6463 */
6464int __meminit init_per_zone_wmark_min(void)
6465{
6466 unsigned long lowmem_kbytes;
6467 int new_min_free_kbytes;
6468
6469 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6470 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6471
6472 if (new_min_free_kbytes > user_min_free_kbytes) {
6473 min_free_kbytes = new_min_free_kbytes;
6474 if (min_free_kbytes < 128)
6475 min_free_kbytes = 128;
6476 if (min_free_kbytes > 65536)
6477 min_free_kbytes = 65536;
6478 } else {
6479 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6480 new_min_free_kbytes, user_min_free_kbytes);
6481 }
6482 setup_per_zone_wmarks();
6483 refresh_zone_stat_thresholds();
6484 setup_per_zone_lowmem_reserve();
6485 setup_per_zone_inactive_ratio();
6486 return 0;
6487}
6488core_initcall(init_per_zone_wmark_min)
6489
6490/*
6491 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6492 * that we can call two helper functions whenever min_free_kbytes
6493 * changes.
6494 */
6495int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6496 void __user *buffer, size_t *length, loff_t *ppos)
6497{
6498 int rc;
6499
6500 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6501 if (rc)
6502 return rc;
6503
6504 if (write) {
6505 user_min_free_kbytes = min_free_kbytes;
6506 setup_per_zone_wmarks();
6507 }
6508 return 0;
6509}
6510
6511int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
6512 void __user *buffer, size_t *length, loff_t *ppos)
6513{
6514 int rc;
6515
6516 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6517 if (rc)
6518 return rc;
6519
6520 if (write)
6521 setup_per_zone_wmarks();
6522
6523 return 0;
6524}
6525
6526#ifdef CONFIG_NUMA
6527int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6528 void __user *buffer, size_t *length, loff_t *ppos)
6529{
6530 struct zone *zone;
6531 int rc;
6532
6533 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6534 if (rc)
6535 return rc;
6536
6537 for_each_zone(zone)
6538 zone->min_unmapped_pages = (zone->managed_pages *
6539 sysctl_min_unmapped_ratio) / 100;
6540 return 0;
6541}
6542
6543int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6544 void __user *buffer, size_t *length, loff_t *ppos)
6545{
6546 struct zone *zone;
6547 int rc;
6548
6549 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6550 if (rc)
6551 return rc;
6552
6553 for_each_zone(zone)
6554 zone->min_slab_pages = (zone->managed_pages *
6555 sysctl_min_slab_ratio) / 100;
6556 return 0;
6557}
6558#endif
6559
6560/*
6561 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6562 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6563 * whenever sysctl_lowmem_reserve_ratio changes.
6564 *
6565 * The reserve ratio obviously has absolutely no relation with the
6566 * minimum watermarks. The lowmem reserve ratio can only make sense
6567 * if in function of the boot time zone sizes.
6568 */
6569int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6570 void __user *buffer, size_t *length, loff_t *ppos)
6571{
6572 proc_dointvec_minmax(table, write, buffer, length, ppos);
6573 setup_per_zone_lowmem_reserve();
6574 return 0;
6575}
6576
6577/*
6578 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6579 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6580 * pagelist can have before it gets flushed back to buddy allocator.
6581 */
6582int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6583 void __user *buffer, size_t *length, loff_t *ppos)
6584{
6585 struct zone *zone;
6586 int old_percpu_pagelist_fraction;
6587 int ret;
6588
6589 mutex_lock(&pcp_batch_high_lock);
6590 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6591
6592 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6593 if (!write || ret < 0)
6594 goto out;
6595
6596 /* Sanity checking to avoid pcp imbalance */
6597 if (percpu_pagelist_fraction &&
6598 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6599 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6600 ret = -EINVAL;
6601 goto out;
6602 }
6603
6604 /* No change? */
6605 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6606 goto out;
6607
6608 for_each_populated_zone(zone) {
6609 unsigned int cpu;
6610
6611 for_each_possible_cpu(cpu)
6612 pageset_set_high_and_batch(zone,
6613 per_cpu_ptr(zone->pageset, cpu));
6614 }
6615out:
6616 mutex_unlock(&pcp_batch_high_lock);
6617 return ret;
6618}
6619
6620#ifdef CONFIG_NUMA
6621int hashdist = HASHDIST_DEFAULT;
6622
6623static int __init set_hashdist(char *str)
6624{
6625 if (!str)
6626 return 0;
6627 hashdist = simple_strtoul(str, &str, 0);
6628 return 1;
6629}
6630__setup("hashdist=", set_hashdist);
6631#endif
6632
6633/*
6634 * allocate a large system hash table from bootmem
6635 * - it is assumed that the hash table must contain an exact power-of-2
6636 * quantity of entries
6637 * - limit is the number of hash buckets, not the total allocation size
6638 */
6639void *__init alloc_large_system_hash(const char *tablename,
6640 unsigned long bucketsize,
6641 unsigned long numentries,
6642 int scale,
6643 int flags,
6644 unsigned int *_hash_shift,
6645 unsigned int *_hash_mask,
6646 unsigned long low_limit,
6647 unsigned long high_limit)
6648{
6649 unsigned long long max = high_limit;
6650 unsigned long log2qty, size;
6651 void *table = NULL;
6652
6653 /* allow the kernel cmdline to have a say */
6654 if (!numentries) {
6655 /* round applicable memory size up to nearest megabyte */
6656 numentries = nr_kernel_pages;
6657
6658 /* It isn't necessary when PAGE_SIZE >= 1MB */
6659 if (PAGE_SHIFT < 20)
6660 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6661
6662 /* limit to 1 bucket per 2^scale bytes of low memory */
6663 if (scale > PAGE_SHIFT)
6664 numentries >>= (scale - PAGE_SHIFT);
6665 else
6666 numentries <<= (PAGE_SHIFT - scale);
6667
6668 /* Make sure we've got at least a 0-order allocation.. */
6669 if (unlikely(flags & HASH_SMALL)) {
6670 /* Makes no sense without HASH_EARLY */
6671 WARN_ON(!(flags & HASH_EARLY));
6672 if (!(numentries >> *_hash_shift)) {
6673 numentries = 1UL << *_hash_shift;
6674 BUG_ON(!numentries);
6675 }
6676 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6677 numentries = PAGE_SIZE / bucketsize;
6678 }
6679 numentries = roundup_pow_of_two(numentries);
6680
6681 /* limit allocation size to 1/16 total memory by default */
6682 if (max == 0) {
6683 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6684 do_div(max, bucketsize);
6685 }
6686 max = min(max, 0x80000000ULL);
6687
6688 if (numentries < low_limit)
6689 numentries = low_limit;
6690 if (numentries > max)
6691 numentries = max;
6692
6693 log2qty = ilog2(numentries);
6694
6695 do {
6696 size = bucketsize << log2qty;
6697 if (flags & HASH_EARLY)
6698 table = memblock_virt_alloc_nopanic(size, 0);
6699 else if (hashdist)
6700 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6701 else {
6702 /*
6703 * If bucketsize is not a power-of-two, we may free
6704 * some pages at the end of hash table which
6705 * alloc_pages_exact() automatically does
6706 */
6707 if (get_order(size) < MAX_ORDER) {
6708 table = alloc_pages_exact(size, GFP_ATOMIC);
6709 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6710 }
6711 }
6712 } while (!table && size > PAGE_SIZE && --log2qty);
6713
6714 if (!table)
6715 panic("Failed to allocate %s hash table\n", tablename);
6716
6717 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
6718 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
6719
6720 if (_hash_shift)
6721 *_hash_shift = log2qty;
6722 if (_hash_mask)
6723 *_hash_mask = (1 << log2qty) - 1;
6724
6725 return table;
6726}
6727
6728/* Return a pointer to the bitmap storing bits affecting a block of pages */
6729static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6730 unsigned long pfn)
6731{
6732#ifdef CONFIG_SPARSEMEM
6733 return __pfn_to_section(pfn)->pageblock_flags;
6734#else
6735 return zone->pageblock_flags;
6736#endif /* CONFIG_SPARSEMEM */
6737}
6738
6739static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6740{
6741#ifdef CONFIG_SPARSEMEM
6742 pfn &= (PAGES_PER_SECTION-1);
6743 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6744#else
6745 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6746 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6747#endif /* CONFIG_SPARSEMEM */
6748}
6749
6750/**
6751 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6752 * @page: The page within the block of interest
6753 * @pfn: The target page frame number
6754 * @end_bitidx: The last bit of interest to retrieve
6755 * @mask: mask of bits that the caller is interested in
6756 *
6757 * Return: pageblock_bits flags
6758 */
6759unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6760 unsigned long end_bitidx,
6761 unsigned long mask)
6762{
6763 struct zone *zone;
6764 unsigned long *bitmap;
6765 unsigned long bitidx, word_bitidx;
6766 unsigned long word;
6767
6768 zone = page_zone(page);
6769 bitmap = get_pageblock_bitmap(zone, pfn);
6770 bitidx = pfn_to_bitidx(zone, pfn);
6771 word_bitidx = bitidx / BITS_PER_LONG;
6772 bitidx &= (BITS_PER_LONG-1);
6773
6774 word = bitmap[word_bitidx];
6775 bitidx += end_bitidx;
6776 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6777}
6778
6779/**
6780 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6781 * @page: The page within the block of interest
6782 * @flags: The flags to set
6783 * @pfn: The target page frame number
6784 * @end_bitidx: The last bit of interest
6785 * @mask: mask of bits that the caller is interested in
6786 */
6787void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6788 unsigned long pfn,
6789 unsigned long end_bitidx,
6790 unsigned long mask)
6791{
6792 struct zone *zone;
6793 unsigned long *bitmap;
6794 unsigned long bitidx, word_bitidx;
6795 unsigned long old_word, word;
6796
6797 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6798
6799 zone = page_zone(page);
6800 bitmap = get_pageblock_bitmap(zone, pfn);
6801 bitidx = pfn_to_bitidx(zone, pfn);
6802 word_bitidx = bitidx / BITS_PER_LONG;
6803 bitidx &= (BITS_PER_LONG-1);
6804
6805 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6806
6807 bitidx += end_bitidx;
6808 mask <<= (BITS_PER_LONG - bitidx - 1);
6809 flags <<= (BITS_PER_LONG - bitidx - 1);
6810
6811 word = READ_ONCE(bitmap[word_bitidx]);
6812 for (;;) {
6813 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6814 if (word == old_word)
6815 break;
6816 word = old_word;
6817 }
6818}
6819
6820/*
6821 * This function checks whether pageblock includes unmovable pages or not.
6822 * If @count is not zero, it is okay to include less @count unmovable pages
6823 *
6824 * PageLRU check without isolation or lru_lock could race so that
6825 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6826 * expect this function should be exact.
6827 */
6828bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6829 bool skip_hwpoisoned_pages)
6830{
6831 unsigned long pfn, iter, found;
6832 int mt;
6833
6834 /*
6835 * For avoiding noise data, lru_add_drain_all() should be called
6836 * If ZONE_MOVABLE, the zone never contains unmovable pages
6837 */
6838 if (zone_idx(zone) == ZONE_MOVABLE)
6839 return false;
6840 mt = get_pageblock_migratetype(page);
6841 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6842 return false;
6843
6844 pfn = page_to_pfn(page);
6845 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6846 unsigned long check = pfn + iter;
6847
6848 if (!pfn_valid_within(check))
6849 continue;
6850
6851 page = pfn_to_page(check);
6852
6853 /*
6854 * Hugepages are not in LRU lists, but they're movable.
6855 * We need not scan over tail pages bacause we don't
6856 * handle each tail page individually in migration.
6857 */
6858 if (PageHuge(page)) {
6859 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6860 continue;
6861 }
6862
6863 /*
6864 * We can't use page_count without pin a page
6865 * because another CPU can free compound page.
6866 * This check already skips compound tails of THP
6867 * because their page->_count is zero at all time.
6868 */
6869 if (!page_ref_count(page)) {
6870 if (PageBuddy(page))
6871 iter += (1 << page_order(page)) - 1;
6872 continue;
6873 }
6874
6875 /*
6876 * The HWPoisoned page may be not in buddy system, and
6877 * page_count() is not 0.
6878 */
6879 if (skip_hwpoisoned_pages && PageHWPoison(page))
6880 continue;
6881
6882 if (!PageLRU(page))
6883 found++;
6884 /*
6885 * If there are RECLAIMABLE pages, we need to check
6886 * it. But now, memory offline itself doesn't call
6887 * shrink_node_slabs() and it still to be fixed.
6888 */
6889 /*
6890 * If the page is not RAM, page_count()should be 0.
6891 * we don't need more check. This is an _used_ not-movable page.
6892 *
6893 * The problematic thing here is PG_reserved pages. PG_reserved
6894 * is set to both of a memory hole page and a _used_ kernel
6895 * page at boot.
6896 */
6897 if (found > count)
6898 return true;
6899 }
6900 return false;
6901}
6902
6903bool is_pageblock_removable_nolock(struct page *page)
6904{
6905 struct zone *zone;
6906 unsigned long pfn;
6907
6908 /*
6909 * We have to be careful here because we are iterating over memory
6910 * sections which are not zone aware so we might end up outside of
6911 * the zone but still within the section.
6912 * We have to take care about the node as well. If the node is offline
6913 * its NODE_DATA will be NULL - see page_zone.
6914 */
6915 if (!node_online(page_to_nid(page)))
6916 return false;
6917
6918 zone = page_zone(page);
6919 pfn = page_to_pfn(page);
6920 if (!zone_spans_pfn(zone, pfn))
6921 return false;
6922
6923 return !has_unmovable_pages(zone, page, 0, true);
6924}
6925
6926#if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
6927
6928static unsigned long pfn_max_align_down(unsigned long pfn)
6929{
6930 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6931 pageblock_nr_pages) - 1);
6932}
6933
6934static unsigned long pfn_max_align_up(unsigned long pfn)
6935{
6936 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6937 pageblock_nr_pages));
6938}
6939
6940/* [start, end) must belong to a single zone. */
6941static int __alloc_contig_migrate_range(struct compact_control *cc,
6942 unsigned long start, unsigned long end)
6943{
6944 /* This function is based on compact_zone() from compaction.c. */
6945 unsigned long nr_reclaimed;
6946 unsigned long pfn = start;
6947 unsigned int tries = 0;
6948 int ret = 0;
6949
6950 migrate_prep();
6951
6952 while (pfn < end || !list_empty(&cc->migratepages)) {
6953 if (fatal_signal_pending(current)) {
6954 ret = -EINTR;
6955 break;
6956 }
6957
6958 if (list_empty(&cc->migratepages)) {
6959 cc->nr_migratepages = 0;
6960 pfn = isolate_migratepages_range(cc, pfn, end);
6961 if (!pfn) {
6962 ret = -EINTR;
6963 break;
6964 }
6965 tries = 0;
6966 } else if (++tries == 5) {
6967 ret = ret < 0 ? ret : -EBUSY;
6968 break;
6969 }
6970
6971 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6972 &cc->migratepages);
6973 cc->nr_migratepages -= nr_reclaimed;
6974
6975 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6976 NULL, 0, cc->mode, MR_CMA);
6977 }
6978 if (ret < 0) {
6979 putback_movable_pages(&cc->migratepages);
6980 return ret;
6981 }
6982 return 0;
6983}
6984
6985/**
6986 * alloc_contig_range() -- tries to allocate given range of pages
6987 * @start: start PFN to allocate
6988 * @end: one-past-the-last PFN to allocate
6989 * @migratetype: migratetype of the underlaying pageblocks (either
6990 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6991 * in range must have the same migratetype and it must
6992 * be either of the two.
6993 *
6994 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6995 * aligned, however it's the caller's responsibility to guarantee that
6996 * we are the only thread that changes migrate type of pageblocks the
6997 * pages fall in.
6998 *
6999 * The PFN range must belong to a single zone.
7000 *
7001 * Returns zero on success or negative error code. On success all
7002 * pages which PFN is in [start, end) are allocated for the caller and
7003 * need to be freed with free_contig_range().
7004 */
7005int alloc_contig_range(unsigned long start, unsigned long end,
7006 unsigned migratetype)
7007{
7008 unsigned long outer_start, outer_end;
7009 unsigned int order;
7010 int ret = 0;
7011
7012 struct compact_control cc = {
7013 .nr_migratepages = 0,
7014 .order = -1,
7015 .zone = page_zone(pfn_to_page(start)),
7016 .mode = MIGRATE_SYNC,
7017 .ignore_skip_hint = true,
7018 };
7019 INIT_LIST_HEAD(&cc.migratepages);
7020
7021 /*
7022 * What we do here is we mark all pageblocks in range as
7023 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7024 * have different sizes, and due to the way page allocator
7025 * work, we align the range to biggest of the two pages so
7026 * that page allocator won't try to merge buddies from
7027 * different pageblocks and change MIGRATE_ISOLATE to some
7028 * other migration type.
7029 *
7030 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7031 * migrate the pages from an unaligned range (ie. pages that
7032 * we are interested in). This will put all the pages in
7033 * range back to page allocator as MIGRATE_ISOLATE.
7034 *
7035 * When this is done, we take the pages in range from page
7036 * allocator removing them from the buddy system. This way
7037 * page allocator will never consider using them.
7038 *
7039 * This lets us mark the pageblocks back as
7040 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7041 * aligned range but not in the unaligned, original range are
7042 * put back to page allocator so that buddy can use them.
7043 */
7044
7045 ret = start_isolate_page_range(pfn_max_align_down(start),
7046 pfn_max_align_up(end), migratetype,
7047 false);
7048 if (ret)
7049 return ret;
7050
7051 /*
7052 * In case of -EBUSY, we'd like to know which page causes problem.
7053 * So, just fall through. We will check it in test_pages_isolated().
7054 */
7055 ret = __alloc_contig_migrate_range(&cc, start, end);
7056 if (ret && ret != -EBUSY)
7057 goto done;
7058
7059 /*
7060 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7061 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7062 * more, all pages in [start, end) are free in page allocator.
7063 * What we are going to do is to allocate all pages from
7064 * [start, end) (that is remove them from page allocator).
7065 *
7066 * The only problem is that pages at the beginning and at the
7067 * end of interesting range may be not aligned with pages that
7068 * page allocator holds, ie. they can be part of higher order
7069 * pages. Because of this, we reserve the bigger range and
7070 * once this is done free the pages we are not interested in.
7071 *
7072 * We don't have to hold zone->lock here because the pages are
7073 * isolated thus they won't get removed from buddy.
7074 */
7075
7076 lru_add_drain_all();
7077 drain_all_pages(cc.zone);
7078
7079 order = 0;
7080 outer_start = start;
7081 while (!PageBuddy(pfn_to_page(outer_start))) {
7082 if (++order >= MAX_ORDER) {
7083 outer_start = start;
7084 break;
7085 }
7086 outer_start &= ~0UL << order;
7087 }
7088
7089 if (outer_start != start) {
7090 order = page_order(pfn_to_page(outer_start));
7091
7092 /*
7093 * outer_start page could be small order buddy page and
7094 * it doesn't include start page. Adjust outer_start
7095 * in this case to report failed page properly
7096 * on tracepoint in test_pages_isolated()
7097 */
7098 if (outer_start + (1UL << order) <= start)
7099 outer_start = start;
7100 }
7101
7102 /* Make sure the range is really isolated. */
7103 if (test_pages_isolated(outer_start, end, false)) {
7104 pr_info("%s: [%lx, %lx) PFNs busy\n",
7105 __func__, outer_start, end);
7106 ret = -EBUSY;
7107 goto done;
7108 }
7109
7110 /* Grab isolated pages from freelists. */
7111 outer_end = isolate_freepages_range(&cc, outer_start, end);
7112 if (!outer_end) {
7113 ret = -EBUSY;
7114 goto done;
7115 }
7116
7117 /* Free head and tail (if any) */
7118 if (start != outer_start)
7119 free_contig_range(outer_start, start - outer_start);
7120 if (end != outer_end)
7121 free_contig_range(end, outer_end - end);
7122
7123done:
7124 undo_isolate_page_range(pfn_max_align_down(start),
7125 pfn_max_align_up(end), migratetype);
7126 return ret;
7127}
7128
7129void free_contig_range(unsigned long pfn, unsigned nr_pages)
7130{
7131 unsigned int count = 0;
7132
7133 for (; nr_pages--; pfn++) {
7134 struct page *page = pfn_to_page(pfn);
7135
7136 count += page_count(page) != 1;
7137 __free_page(page);
7138 }
7139 WARN(count != 0, "%d pages are still in use!\n", count);
7140}
7141#endif
7142
7143#ifdef CONFIG_MEMORY_HOTPLUG
7144/*
7145 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7146 * page high values need to be recalulated.
7147 */
7148void __meminit zone_pcp_update(struct zone *zone)
7149{
7150 unsigned cpu;
7151 mutex_lock(&pcp_batch_high_lock);
7152 for_each_possible_cpu(cpu)
7153 pageset_set_high_and_batch(zone,
7154 per_cpu_ptr(zone->pageset, cpu));
7155 mutex_unlock(&pcp_batch_high_lock);
7156}
7157#endif
7158
7159void zone_pcp_reset(struct zone *zone)
7160{
7161 unsigned long flags;
7162 int cpu;
7163 struct per_cpu_pageset *pset;
7164
7165 /* avoid races with drain_pages() */
7166 local_irq_save(flags);
7167 if (zone->pageset != &boot_pageset) {
7168 for_each_online_cpu(cpu) {
7169 pset = per_cpu_ptr(zone->pageset, cpu);
7170 drain_zonestat(zone, pset);
7171 }
7172 free_percpu(zone->pageset);
7173 zone->pageset = &boot_pageset;
7174 }
7175 local_irq_restore(flags);
7176}
7177
7178#ifdef CONFIG_MEMORY_HOTREMOVE
7179/*
7180 * All pages in the range must be isolated before calling this.
7181 */
7182void
7183__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7184{
7185 struct page *page;
7186 struct zone *zone;
7187 unsigned int order, i;
7188 unsigned long pfn;
7189 unsigned long flags;
7190 /* find the first valid pfn */
7191 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7192 if (pfn_valid(pfn))
7193 break;
7194 if (pfn == end_pfn)
7195 return;
7196 zone = page_zone(pfn_to_page(pfn));
7197 spin_lock_irqsave(&zone->lock, flags);
7198 pfn = start_pfn;
7199 while (pfn < end_pfn) {
7200 if (!pfn_valid(pfn)) {
7201 pfn++;
7202 continue;
7203 }
7204 page = pfn_to_page(pfn);
7205 /*
7206 * The HWPoisoned page may be not in buddy system, and
7207 * page_count() is not 0.
7208 */
7209 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7210 pfn++;
7211 SetPageReserved(page);
7212 continue;
7213 }
7214
7215 BUG_ON(page_count(page));
7216 BUG_ON(!PageBuddy(page));
7217 order = page_order(page);
7218#ifdef CONFIG_DEBUG_VM
7219 pr_info("remove from free list %lx %d %lx\n",
7220 pfn, 1 << order, end_pfn);
7221#endif
7222 list_del(&page->lru);
7223 rmv_page_order(page);
7224 zone->free_area[order].nr_free--;
7225 for (i = 0; i < (1 << order); i++)
7226 SetPageReserved((page+i));
7227 pfn += (1 << order);
7228 }
7229 spin_unlock_irqrestore(&zone->lock, flags);
7230}
7231#endif
7232
7233bool is_free_buddy_page(struct page *page)
7234{
7235 struct zone *zone = page_zone(page);
7236 unsigned long pfn = page_to_pfn(page);
7237 unsigned long flags;
7238 unsigned int order;
7239
7240 spin_lock_irqsave(&zone->lock, flags);
7241 for (order = 0; order < MAX_ORDER; order++) {
7242 struct page *page_head = page - (pfn & ((1 << order) - 1));
7243
7244 if (PageBuddy(page_head) && page_order(page_head) >= order)
7245 break;
7246 }
7247 spin_unlock_irqrestore(&zone->lock, flags);
7248
7249 return order < MAX_ORDER;
7250}
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/page_alloc.c
4 *
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
7 *
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
16 */
17
18#include <linux/stddef.h>
19#include <linux/mm.h>
20#include <linux/highmem.h>
21#include <linux/swap.h>
22#include <linux/interrupt.h>
23#include <linux/pagemap.h>
24#include <linux/jiffies.h>
25#include <linux/memblock.h>
26#include <linux/compiler.h>
27#include <linux/kernel.h>
28#include <linux/kasan.h>
29#include <linux/module.h>
30#include <linux/suspend.h>
31#include <linux/pagevec.h>
32#include <linux/blkdev.h>
33#include <linux/slab.h>
34#include <linux/ratelimit.h>
35#include <linux/oom.h>
36#include <linux/topology.h>
37#include <linux/sysctl.h>
38#include <linux/cpu.h>
39#include <linux/cpuset.h>
40#include <linux/memory_hotplug.h>
41#include <linux/nodemask.h>
42#include <linux/vmalloc.h>
43#include <linux/vmstat.h>
44#include <linux/mempolicy.h>
45#include <linux/memremap.h>
46#include <linux/stop_machine.h>
47#include <linux/random.h>
48#include <linux/sort.h>
49#include <linux/pfn.h>
50#include <linux/backing-dev.h>
51#include <linux/fault-inject.h>
52#include <linux/page-isolation.h>
53#include <linux/debugobjects.h>
54#include <linux/kmemleak.h>
55#include <linux/compaction.h>
56#include <trace/events/kmem.h>
57#include <trace/events/oom.h>
58#include <linux/prefetch.h>
59#include <linux/mm_inline.h>
60#include <linux/mmu_notifier.h>
61#include <linux/migrate.h>
62#include <linux/hugetlb.h>
63#include <linux/sched/rt.h>
64#include <linux/sched/mm.h>
65#include <linux/page_owner.h>
66#include <linux/kthread.h>
67#include <linux/memcontrol.h>
68#include <linux/ftrace.h>
69#include <linux/lockdep.h>
70#include <linux/nmi.h>
71#include <linux/psi.h>
72#include <linux/padata.h>
73#include <linux/khugepaged.h>
74#include <linux/buffer_head.h>
75#include <asm/sections.h>
76#include <asm/tlbflush.h>
77#include <asm/div64.h>
78#include "internal.h"
79#include "shuffle.h"
80#include "page_reporting.h"
81
82/* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
83typedef int __bitwise fpi_t;
84
85/* No special request */
86#define FPI_NONE ((__force fpi_t)0)
87
88/*
89 * Skip free page reporting notification for the (possibly merged) page.
90 * This does not hinder free page reporting from grabbing the page,
91 * reporting it and marking it "reported" - it only skips notifying
92 * the free page reporting infrastructure about a newly freed page. For
93 * example, used when temporarily pulling a page from a freelist and
94 * putting it back unmodified.
95 */
96#define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
97
98/*
99 * Place the (possibly merged) page to the tail of the freelist. Will ignore
100 * page shuffling (relevant code - e.g., memory onlining - is expected to
101 * shuffle the whole zone).
102 *
103 * Note: No code should rely on this flag for correctness - it's purely
104 * to allow for optimizations when handing back either fresh pages
105 * (memory onlining) or untouched pages (page isolation, free page
106 * reporting).
107 */
108#define FPI_TO_TAIL ((__force fpi_t)BIT(1))
109
110/*
111 * Don't poison memory with KASAN (only for the tag-based modes).
112 * During boot, all non-reserved memblock memory is exposed to page_alloc.
113 * Poisoning all that memory lengthens boot time, especially on systems with
114 * large amount of RAM. This flag is used to skip that poisoning.
115 * This is only done for the tag-based KASAN modes, as those are able to
116 * detect memory corruptions with the memory tags assigned by default.
117 * All memory allocated normally after boot gets poisoned as usual.
118 */
119#define FPI_SKIP_KASAN_POISON ((__force fpi_t)BIT(2))
120
121/* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
122static DEFINE_MUTEX(pcp_batch_high_lock);
123#define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
124
125struct pagesets {
126 local_lock_t lock;
127};
128static DEFINE_PER_CPU(struct pagesets, pagesets) = {
129 .lock = INIT_LOCAL_LOCK(lock),
130};
131
132#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
133DEFINE_PER_CPU(int, numa_node);
134EXPORT_PER_CPU_SYMBOL(numa_node);
135#endif
136
137DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
138
139#ifdef CONFIG_HAVE_MEMORYLESS_NODES
140/*
141 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
142 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
143 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
144 * defined in <linux/topology.h>.
145 */
146DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
147EXPORT_PER_CPU_SYMBOL(_numa_mem_);
148#endif
149
150/* work_structs for global per-cpu drains */
151struct pcpu_drain {
152 struct zone *zone;
153 struct work_struct work;
154};
155static DEFINE_MUTEX(pcpu_drain_mutex);
156static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
157
158#ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
159volatile unsigned long latent_entropy __latent_entropy;
160EXPORT_SYMBOL(latent_entropy);
161#endif
162
163/*
164 * Array of node states.
165 */
166nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
167 [N_POSSIBLE] = NODE_MASK_ALL,
168 [N_ONLINE] = { { [0] = 1UL } },
169#ifndef CONFIG_NUMA
170 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
171#ifdef CONFIG_HIGHMEM
172 [N_HIGH_MEMORY] = { { [0] = 1UL } },
173#endif
174 [N_MEMORY] = { { [0] = 1UL } },
175 [N_CPU] = { { [0] = 1UL } },
176#endif /* NUMA */
177};
178EXPORT_SYMBOL(node_states);
179
180atomic_long_t _totalram_pages __read_mostly;
181EXPORT_SYMBOL(_totalram_pages);
182unsigned long totalreserve_pages __read_mostly;
183unsigned long totalcma_pages __read_mostly;
184
185int percpu_pagelist_high_fraction;
186gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
187DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_ALLOC_DEFAULT_ON, init_on_alloc);
188EXPORT_SYMBOL(init_on_alloc);
189
190DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_FREE_DEFAULT_ON, init_on_free);
191EXPORT_SYMBOL(init_on_free);
192
193static bool _init_on_alloc_enabled_early __read_mostly
194 = IS_ENABLED(CONFIG_INIT_ON_ALLOC_DEFAULT_ON);
195static int __init early_init_on_alloc(char *buf)
196{
197
198 return kstrtobool(buf, &_init_on_alloc_enabled_early);
199}
200early_param("init_on_alloc", early_init_on_alloc);
201
202static bool _init_on_free_enabled_early __read_mostly
203 = IS_ENABLED(CONFIG_INIT_ON_FREE_DEFAULT_ON);
204static int __init early_init_on_free(char *buf)
205{
206 return kstrtobool(buf, &_init_on_free_enabled_early);
207}
208early_param("init_on_free", early_init_on_free);
209
210/*
211 * A cached value of the page's pageblock's migratetype, used when the page is
212 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
213 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
214 * Also the migratetype set in the page does not necessarily match the pcplist
215 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
216 * other index - this ensures that it will be put on the correct CMA freelist.
217 */
218static inline int get_pcppage_migratetype(struct page *page)
219{
220 return page->index;
221}
222
223static inline void set_pcppage_migratetype(struct page *page, int migratetype)
224{
225 page->index = migratetype;
226}
227
228#ifdef CONFIG_PM_SLEEP
229/*
230 * The following functions are used by the suspend/hibernate code to temporarily
231 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
232 * while devices are suspended. To avoid races with the suspend/hibernate code,
233 * they should always be called with system_transition_mutex held
234 * (gfp_allowed_mask also should only be modified with system_transition_mutex
235 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
236 * with that modification).
237 */
238
239static gfp_t saved_gfp_mask;
240
241void pm_restore_gfp_mask(void)
242{
243 WARN_ON(!mutex_is_locked(&system_transition_mutex));
244 if (saved_gfp_mask) {
245 gfp_allowed_mask = saved_gfp_mask;
246 saved_gfp_mask = 0;
247 }
248}
249
250void pm_restrict_gfp_mask(void)
251{
252 WARN_ON(!mutex_is_locked(&system_transition_mutex));
253 WARN_ON(saved_gfp_mask);
254 saved_gfp_mask = gfp_allowed_mask;
255 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
256}
257
258bool pm_suspended_storage(void)
259{
260 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
261 return false;
262 return true;
263}
264#endif /* CONFIG_PM_SLEEP */
265
266#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
267unsigned int pageblock_order __read_mostly;
268#endif
269
270static void __free_pages_ok(struct page *page, unsigned int order,
271 fpi_t fpi_flags);
272
273/*
274 * results with 256, 32 in the lowmem_reserve sysctl:
275 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
276 * 1G machine -> (16M dma, 784M normal, 224M high)
277 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
278 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
279 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
280 *
281 * TBD: should special case ZONE_DMA32 machines here - in those we normally
282 * don't need any ZONE_NORMAL reservation
283 */
284int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
285#ifdef CONFIG_ZONE_DMA
286 [ZONE_DMA] = 256,
287#endif
288#ifdef CONFIG_ZONE_DMA32
289 [ZONE_DMA32] = 256,
290#endif
291 [ZONE_NORMAL] = 32,
292#ifdef CONFIG_HIGHMEM
293 [ZONE_HIGHMEM] = 0,
294#endif
295 [ZONE_MOVABLE] = 0,
296};
297
298static char * const zone_names[MAX_NR_ZONES] = {
299#ifdef CONFIG_ZONE_DMA
300 "DMA",
301#endif
302#ifdef CONFIG_ZONE_DMA32
303 "DMA32",
304#endif
305 "Normal",
306#ifdef CONFIG_HIGHMEM
307 "HighMem",
308#endif
309 "Movable",
310#ifdef CONFIG_ZONE_DEVICE
311 "Device",
312#endif
313};
314
315const char * const migratetype_names[MIGRATE_TYPES] = {
316 "Unmovable",
317 "Movable",
318 "Reclaimable",
319 "HighAtomic",
320#ifdef CONFIG_CMA
321 "CMA",
322#endif
323#ifdef CONFIG_MEMORY_ISOLATION
324 "Isolate",
325#endif
326};
327
328compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
329 [NULL_COMPOUND_DTOR] = NULL,
330 [COMPOUND_PAGE_DTOR] = free_compound_page,
331#ifdef CONFIG_HUGETLB_PAGE
332 [HUGETLB_PAGE_DTOR] = free_huge_page,
333#endif
334#ifdef CONFIG_TRANSPARENT_HUGEPAGE
335 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
336#endif
337};
338
339int min_free_kbytes = 1024;
340int user_min_free_kbytes = -1;
341int watermark_boost_factor __read_mostly = 15000;
342int watermark_scale_factor = 10;
343
344static unsigned long nr_kernel_pages __initdata;
345static unsigned long nr_all_pages __initdata;
346static unsigned long dma_reserve __initdata;
347
348static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
349static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
350static unsigned long required_kernelcore __initdata;
351static unsigned long required_kernelcore_percent __initdata;
352static unsigned long required_movablecore __initdata;
353static unsigned long required_movablecore_percent __initdata;
354static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
355static bool mirrored_kernelcore __meminitdata;
356
357/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
358int movable_zone;
359EXPORT_SYMBOL(movable_zone);
360
361#if MAX_NUMNODES > 1
362unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
363unsigned int nr_online_nodes __read_mostly = 1;
364EXPORT_SYMBOL(nr_node_ids);
365EXPORT_SYMBOL(nr_online_nodes);
366#endif
367
368int page_group_by_mobility_disabled __read_mostly;
369
370#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
371/*
372 * During boot we initialize deferred pages on-demand, as needed, but once
373 * page_alloc_init_late() has finished, the deferred pages are all initialized,
374 * and we can permanently disable that path.
375 */
376static DEFINE_STATIC_KEY_TRUE(deferred_pages);
377
378/*
379 * Calling kasan_poison_pages() only after deferred memory initialization
380 * has completed. Poisoning pages during deferred memory init will greatly
381 * lengthen the process and cause problem in large memory systems as the
382 * deferred pages initialization is done with interrupt disabled.
383 *
384 * Assuming that there will be no reference to those newly initialized
385 * pages before they are ever allocated, this should have no effect on
386 * KASAN memory tracking as the poison will be properly inserted at page
387 * allocation time. The only corner case is when pages are allocated by
388 * on-demand allocation and then freed again before the deferred pages
389 * initialization is done, but this is not likely to happen.
390 */
391static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
392{
393 return static_branch_unlikely(&deferred_pages) ||
394 (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
395 (fpi_flags & FPI_SKIP_KASAN_POISON)) ||
396 PageSkipKASanPoison(page);
397}
398
399/* Returns true if the struct page for the pfn is uninitialised */
400static inline bool __meminit early_page_uninitialised(unsigned long pfn)
401{
402 int nid = early_pfn_to_nid(pfn);
403
404 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
405 return true;
406
407 return false;
408}
409
410/*
411 * Returns true when the remaining initialisation should be deferred until
412 * later in the boot cycle when it can be parallelised.
413 */
414static bool __meminit
415defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
416{
417 static unsigned long prev_end_pfn, nr_initialised;
418
419 /*
420 * prev_end_pfn static that contains the end of previous zone
421 * No need to protect because called very early in boot before smp_init.
422 */
423 if (prev_end_pfn != end_pfn) {
424 prev_end_pfn = end_pfn;
425 nr_initialised = 0;
426 }
427
428 /* Always populate low zones for address-constrained allocations */
429 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
430 return false;
431
432 if (NODE_DATA(nid)->first_deferred_pfn != ULONG_MAX)
433 return true;
434 /*
435 * We start only with one section of pages, more pages are added as
436 * needed until the rest of deferred pages are initialized.
437 */
438 nr_initialised++;
439 if ((nr_initialised > PAGES_PER_SECTION) &&
440 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
441 NODE_DATA(nid)->first_deferred_pfn = pfn;
442 return true;
443 }
444 return false;
445}
446#else
447static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
448{
449 return (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
450 (fpi_flags & FPI_SKIP_KASAN_POISON)) ||
451 PageSkipKASanPoison(page);
452}
453
454static inline bool early_page_uninitialised(unsigned long pfn)
455{
456 return false;
457}
458
459static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
460{
461 return false;
462}
463#endif
464
465/* Return a pointer to the bitmap storing bits affecting a block of pages */
466static inline unsigned long *get_pageblock_bitmap(const struct page *page,
467 unsigned long pfn)
468{
469#ifdef CONFIG_SPARSEMEM
470 return section_to_usemap(__pfn_to_section(pfn));
471#else
472 return page_zone(page)->pageblock_flags;
473#endif /* CONFIG_SPARSEMEM */
474}
475
476static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
477{
478#ifdef CONFIG_SPARSEMEM
479 pfn &= (PAGES_PER_SECTION-1);
480#else
481 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
482#endif /* CONFIG_SPARSEMEM */
483 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
484}
485
486static __always_inline
487unsigned long __get_pfnblock_flags_mask(const struct page *page,
488 unsigned long pfn,
489 unsigned long mask)
490{
491 unsigned long *bitmap;
492 unsigned long bitidx, word_bitidx;
493 unsigned long word;
494
495 bitmap = get_pageblock_bitmap(page, pfn);
496 bitidx = pfn_to_bitidx(page, pfn);
497 word_bitidx = bitidx / BITS_PER_LONG;
498 bitidx &= (BITS_PER_LONG-1);
499
500 word = bitmap[word_bitidx];
501 return (word >> bitidx) & mask;
502}
503
504/**
505 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
506 * @page: The page within the block of interest
507 * @pfn: The target page frame number
508 * @mask: mask of bits that the caller is interested in
509 *
510 * Return: pageblock_bits flags
511 */
512unsigned long get_pfnblock_flags_mask(const struct page *page,
513 unsigned long pfn, unsigned long mask)
514{
515 return __get_pfnblock_flags_mask(page, pfn, mask);
516}
517
518static __always_inline int get_pfnblock_migratetype(const struct page *page,
519 unsigned long pfn)
520{
521 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
522}
523
524/**
525 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
526 * @page: The page within the block of interest
527 * @flags: The flags to set
528 * @pfn: The target page frame number
529 * @mask: mask of bits that the caller is interested in
530 */
531void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
532 unsigned long pfn,
533 unsigned long mask)
534{
535 unsigned long *bitmap;
536 unsigned long bitidx, word_bitidx;
537 unsigned long old_word, word;
538
539 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
540 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
541
542 bitmap = get_pageblock_bitmap(page, pfn);
543 bitidx = pfn_to_bitidx(page, pfn);
544 word_bitidx = bitidx / BITS_PER_LONG;
545 bitidx &= (BITS_PER_LONG-1);
546
547 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
548
549 mask <<= bitidx;
550 flags <<= bitidx;
551
552 word = READ_ONCE(bitmap[word_bitidx]);
553 for (;;) {
554 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
555 if (word == old_word)
556 break;
557 word = old_word;
558 }
559}
560
561void set_pageblock_migratetype(struct page *page, int migratetype)
562{
563 if (unlikely(page_group_by_mobility_disabled &&
564 migratetype < MIGRATE_PCPTYPES))
565 migratetype = MIGRATE_UNMOVABLE;
566
567 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
568 page_to_pfn(page), MIGRATETYPE_MASK);
569}
570
571#ifdef CONFIG_DEBUG_VM
572static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
573{
574 int ret = 0;
575 unsigned seq;
576 unsigned long pfn = page_to_pfn(page);
577 unsigned long sp, start_pfn;
578
579 do {
580 seq = zone_span_seqbegin(zone);
581 start_pfn = zone->zone_start_pfn;
582 sp = zone->spanned_pages;
583 if (!zone_spans_pfn(zone, pfn))
584 ret = 1;
585 } while (zone_span_seqretry(zone, seq));
586
587 if (ret)
588 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
589 pfn, zone_to_nid(zone), zone->name,
590 start_pfn, start_pfn + sp);
591
592 return ret;
593}
594
595static int page_is_consistent(struct zone *zone, struct page *page)
596{
597 if (!pfn_valid_within(page_to_pfn(page)))
598 return 0;
599 if (zone != page_zone(page))
600 return 0;
601
602 return 1;
603}
604/*
605 * Temporary debugging check for pages not lying within a given zone.
606 */
607static int __maybe_unused bad_range(struct zone *zone, struct page *page)
608{
609 if (page_outside_zone_boundaries(zone, page))
610 return 1;
611 if (!page_is_consistent(zone, page))
612 return 1;
613
614 return 0;
615}
616#else
617static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
618{
619 return 0;
620}
621#endif
622
623static void bad_page(struct page *page, const char *reason)
624{
625 static unsigned long resume;
626 static unsigned long nr_shown;
627 static unsigned long nr_unshown;
628
629 /*
630 * Allow a burst of 60 reports, then keep quiet for that minute;
631 * or allow a steady drip of one report per second.
632 */
633 if (nr_shown == 60) {
634 if (time_before(jiffies, resume)) {
635 nr_unshown++;
636 goto out;
637 }
638 if (nr_unshown) {
639 pr_alert(
640 "BUG: Bad page state: %lu messages suppressed\n",
641 nr_unshown);
642 nr_unshown = 0;
643 }
644 nr_shown = 0;
645 }
646 if (nr_shown++ == 0)
647 resume = jiffies + 60 * HZ;
648
649 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
650 current->comm, page_to_pfn(page));
651 dump_page(page, reason);
652
653 print_modules();
654 dump_stack();
655out:
656 /* Leave bad fields for debug, except PageBuddy could make trouble */
657 page_mapcount_reset(page); /* remove PageBuddy */
658 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
659}
660
661static inline unsigned int order_to_pindex(int migratetype, int order)
662{
663 int base = order;
664
665#ifdef CONFIG_TRANSPARENT_HUGEPAGE
666 if (order > PAGE_ALLOC_COSTLY_ORDER) {
667 VM_BUG_ON(order != pageblock_order);
668 base = PAGE_ALLOC_COSTLY_ORDER + 1;
669 }
670#else
671 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
672#endif
673
674 return (MIGRATE_PCPTYPES * base) + migratetype;
675}
676
677static inline int pindex_to_order(unsigned int pindex)
678{
679 int order = pindex / MIGRATE_PCPTYPES;
680
681#ifdef CONFIG_TRANSPARENT_HUGEPAGE
682 if (order > PAGE_ALLOC_COSTLY_ORDER) {
683 order = pageblock_order;
684 VM_BUG_ON(order != pageblock_order);
685 }
686#else
687 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
688#endif
689
690 return order;
691}
692
693static inline bool pcp_allowed_order(unsigned int order)
694{
695 if (order <= PAGE_ALLOC_COSTLY_ORDER)
696 return true;
697#ifdef CONFIG_TRANSPARENT_HUGEPAGE
698 if (order == pageblock_order)
699 return true;
700#endif
701 return false;
702}
703
704static inline void free_the_page(struct page *page, unsigned int order)
705{
706 if (pcp_allowed_order(order)) /* Via pcp? */
707 free_unref_page(page, order);
708 else
709 __free_pages_ok(page, order, FPI_NONE);
710}
711
712/*
713 * Higher-order pages are called "compound pages". They are structured thusly:
714 *
715 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
716 *
717 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
718 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
719 *
720 * The first tail page's ->compound_dtor holds the offset in array of compound
721 * page destructors. See compound_page_dtors.
722 *
723 * The first tail page's ->compound_order holds the order of allocation.
724 * This usage means that zero-order pages may not be compound.
725 */
726
727void free_compound_page(struct page *page)
728{
729 mem_cgroup_uncharge(page);
730 free_the_page(page, compound_order(page));
731}
732
733void prep_compound_page(struct page *page, unsigned int order)
734{
735 int i;
736 int nr_pages = 1 << order;
737
738 __SetPageHead(page);
739 for (i = 1; i < nr_pages; i++) {
740 struct page *p = page + i;
741 p->mapping = TAIL_MAPPING;
742 set_compound_head(p, page);
743 }
744
745 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
746 set_compound_order(page, order);
747 atomic_set(compound_mapcount_ptr(page), -1);
748 if (hpage_pincount_available(page))
749 atomic_set(compound_pincount_ptr(page), 0);
750}
751
752#ifdef CONFIG_DEBUG_PAGEALLOC
753unsigned int _debug_guardpage_minorder;
754
755bool _debug_pagealloc_enabled_early __read_mostly
756 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
757EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
758DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
759EXPORT_SYMBOL(_debug_pagealloc_enabled);
760
761DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
762
763static int __init early_debug_pagealloc(char *buf)
764{
765 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
766}
767early_param("debug_pagealloc", early_debug_pagealloc);
768
769static int __init debug_guardpage_minorder_setup(char *buf)
770{
771 unsigned long res;
772
773 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
774 pr_err("Bad debug_guardpage_minorder value\n");
775 return 0;
776 }
777 _debug_guardpage_minorder = res;
778 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
779 return 0;
780}
781early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
782
783static inline bool set_page_guard(struct zone *zone, struct page *page,
784 unsigned int order, int migratetype)
785{
786 if (!debug_guardpage_enabled())
787 return false;
788
789 if (order >= debug_guardpage_minorder())
790 return false;
791
792 __SetPageGuard(page);
793 INIT_LIST_HEAD(&page->lru);
794 set_page_private(page, order);
795 /* Guard pages are not available for any usage */
796 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
797
798 return true;
799}
800
801static inline void clear_page_guard(struct zone *zone, struct page *page,
802 unsigned int order, int migratetype)
803{
804 if (!debug_guardpage_enabled())
805 return;
806
807 __ClearPageGuard(page);
808
809 set_page_private(page, 0);
810 if (!is_migrate_isolate(migratetype))
811 __mod_zone_freepage_state(zone, (1 << order), migratetype);
812}
813#else
814static inline bool set_page_guard(struct zone *zone, struct page *page,
815 unsigned int order, int migratetype) { return false; }
816static inline void clear_page_guard(struct zone *zone, struct page *page,
817 unsigned int order, int migratetype) {}
818#endif
819
820/*
821 * Enable static keys related to various memory debugging and hardening options.
822 * Some override others, and depend on early params that are evaluated in the
823 * order of appearance. So we need to first gather the full picture of what was
824 * enabled, and then make decisions.
825 */
826void init_mem_debugging_and_hardening(void)
827{
828 bool page_poisoning_requested = false;
829
830#ifdef CONFIG_PAGE_POISONING
831 /*
832 * Page poisoning is debug page alloc for some arches. If
833 * either of those options are enabled, enable poisoning.
834 */
835 if (page_poisoning_enabled() ||
836 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
837 debug_pagealloc_enabled())) {
838 static_branch_enable(&_page_poisoning_enabled);
839 page_poisoning_requested = true;
840 }
841#endif
842
843 if ((_init_on_alloc_enabled_early || _init_on_free_enabled_early) &&
844 page_poisoning_requested) {
845 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
846 "will take precedence over init_on_alloc and init_on_free\n");
847 _init_on_alloc_enabled_early = false;
848 _init_on_free_enabled_early = false;
849 }
850
851 if (_init_on_alloc_enabled_early)
852 static_branch_enable(&init_on_alloc);
853 else
854 static_branch_disable(&init_on_alloc);
855
856 if (_init_on_free_enabled_early)
857 static_branch_enable(&init_on_free);
858 else
859 static_branch_disable(&init_on_free);
860
861#ifdef CONFIG_DEBUG_PAGEALLOC
862 if (!debug_pagealloc_enabled())
863 return;
864
865 static_branch_enable(&_debug_pagealloc_enabled);
866
867 if (!debug_guardpage_minorder())
868 return;
869
870 static_branch_enable(&_debug_guardpage_enabled);
871#endif
872}
873
874static inline void set_buddy_order(struct page *page, unsigned int order)
875{
876 set_page_private(page, order);
877 __SetPageBuddy(page);
878}
879
880/*
881 * This function checks whether a page is free && is the buddy
882 * we can coalesce a page and its buddy if
883 * (a) the buddy is not in a hole (check before calling!) &&
884 * (b) the buddy is in the buddy system &&
885 * (c) a page and its buddy have the same order &&
886 * (d) a page and its buddy are in the same zone.
887 *
888 * For recording whether a page is in the buddy system, we set PageBuddy.
889 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
890 *
891 * For recording page's order, we use page_private(page).
892 */
893static inline bool page_is_buddy(struct page *page, struct page *buddy,
894 unsigned int order)
895{
896 if (!page_is_guard(buddy) && !PageBuddy(buddy))
897 return false;
898
899 if (buddy_order(buddy) != order)
900 return false;
901
902 /*
903 * zone check is done late to avoid uselessly calculating
904 * zone/node ids for pages that could never merge.
905 */
906 if (page_zone_id(page) != page_zone_id(buddy))
907 return false;
908
909 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
910
911 return true;
912}
913
914#ifdef CONFIG_COMPACTION
915static inline struct capture_control *task_capc(struct zone *zone)
916{
917 struct capture_control *capc = current->capture_control;
918
919 return unlikely(capc) &&
920 !(current->flags & PF_KTHREAD) &&
921 !capc->page &&
922 capc->cc->zone == zone ? capc : NULL;
923}
924
925static inline bool
926compaction_capture(struct capture_control *capc, struct page *page,
927 int order, int migratetype)
928{
929 if (!capc || order != capc->cc->order)
930 return false;
931
932 /* Do not accidentally pollute CMA or isolated regions*/
933 if (is_migrate_cma(migratetype) ||
934 is_migrate_isolate(migratetype))
935 return false;
936
937 /*
938 * Do not let lower order allocations pollute a movable pageblock.
939 * This might let an unmovable request use a reclaimable pageblock
940 * and vice-versa but no more than normal fallback logic which can
941 * have trouble finding a high-order free page.
942 */
943 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
944 return false;
945
946 capc->page = page;
947 return true;
948}
949
950#else
951static inline struct capture_control *task_capc(struct zone *zone)
952{
953 return NULL;
954}
955
956static inline bool
957compaction_capture(struct capture_control *capc, struct page *page,
958 int order, int migratetype)
959{
960 return false;
961}
962#endif /* CONFIG_COMPACTION */
963
964/* Used for pages not on another list */
965static inline void add_to_free_list(struct page *page, struct zone *zone,
966 unsigned int order, int migratetype)
967{
968 struct free_area *area = &zone->free_area[order];
969
970 list_add(&page->lru, &area->free_list[migratetype]);
971 area->nr_free++;
972}
973
974/* Used for pages not on another list */
975static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
976 unsigned int order, int migratetype)
977{
978 struct free_area *area = &zone->free_area[order];
979
980 list_add_tail(&page->lru, &area->free_list[migratetype]);
981 area->nr_free++;
982}
983
984/*
985 * Used for pages which are on another list. Move the pages to the tail
986 * of the list - so the moved pages won't immediately be considered for
987 * allocation again (e.g., optimization for memory onlining).
988 */
989static inline void move_to_free_list(struct page *page, struct zone *zone,
990 unsigned int order, int migratetype)
991{
992 struct free_area *area = &zone->free_area[order];
993
994 list_move_tail(&page->lru, &area->free_list[migratetype]);
995}
996
997static inline void del_page_from_free_list(struct page *page, struct zone *zone,
998 unsigned int order)
999{
1000 /* clear reported state and update reported page count */
1001 if (page_reported(page))
1002 __ClearPageReported(page);
1003
1004 list_del(&page->lru);
1005 __ClearPageBuddy(page);
1006 set_page_private(page, 0);
1007 zone->free_area[order].nr_free--;
1008}
1009
1010/*
1011 * If this is not the largest possible page, check if the buddy
1012 * of the next-highest order is free. If it is, it's possible
1013 * that pages are being freed that will coalesce soon. In case,
1014 * that is happening, add the free page to the tail of the list
1015 * so it's less likely to be used soon and more likely to be merged
1016 * as a higher order page
1017 */
1018static inline bool
1019buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
1020 struct page *page, unsigned int order)
1021{
1022 struct page *higher_page, *higher_buddy;
1023 unsigned long combined_pfn;
1024
1025 if (order >= MAX_ORDER - 2)
1026 return false;
1027
1028 if (!pfn_valid_within(buddy_pfn))
1029 return false;
1030
1031 combined_pfn = buddy_pfn & pfn;
1032 higher_page = page + (combined_pfn - pfn);
1033 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
1034 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
1035
1036 return pfn_valid_within(buddy_pfn) &&
1037 page_is_buddy(higher_page, higher_buddy, order + 1);
1038}
1039
1040/*
1041 * Freeing function for a buddy system allocator.
1042 *
1043 * The concept of a buddy system is to maintain direct-mapped table
1044 * (containing bit values) for memory blocks of various "orders".
1045 * The bottom level table contains the map for the smallest allocatable
1046 * units of memory (here, pages), and each level above it describes
1047 * pairs of units from the levels below, hence, "buddies".
1048 * At a high level, all that happens here is marking the table entry
1049 * at the bottom level available, and propagating the changes upward
1050 * as necessary, plus some accounting needed to play nicely with other
1051 * parts of the VM system.
1052 * At each level, we keep a list of pages, which are heads of continuous
1053 * free pages of length of (1 << order) and marked with PageBuddy.
1054 * Page's order is recorded in page_private(page) field.
1055 * So when we are allocating or freeing one, we can derive the state of the
1056 * other. That is, if we allocate a small block, and both were
1057 * free, the remainder of the region must be split into blocks.
1058 * If a block is freed, and its buddy is also free, then this
1059 * triggers coalescing into a block of larger size.
1060 *
1061 * -- nyc
1062 */
1063
1064static inline void __free_one_page(struct page *page,
1065 unsigned long pfn,
1066 struct zone *zone, unsigned int order,
1067 int migratetype, fpi_t fpi_flags)
1068{
1069 struct capture_control *capc = task_capc(zone);
1070 unsigned long buddy_pfn;
1071 unsigned long combined_pfn;
1072 unsigned int max_order;
1073 struct page *buddy;
1074 bool to_tail;
1075
1076 max_order = min_t(unsigned int, MAX_ORDER - 1, pageblock_order);
1077
1078 VM_BUG_ON(!zone_is_initialized(zone));
1079 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1080
1081 VM_BUG_ON(migratetype == -1);
1082 if (likely(!is_migrate_isolate(migratetype)))
1083 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1084
1085 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1086 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1087
1088continue_merging:
1089 while (order < max_order) {
1090 if (compaction_capture(capc, page, order, migratetype)) {
1091 __mod_zone_freepage_state(zone, -(1 << order),
1092 migratetype);
1093 return;
1094 }
1095 buddy_pfn = __find_buddy_pfn(pfn, order);
1096 buddy = page + (buddy_pfn - pfn);
1097
1098 if (!pfn_valid_within(buddy_pfn))
1099 goto done_merging;
1100 if (!page_is_buddy(page, buddy, order))
1101 goto done_merging;
1102 /*
1103 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1104 * merge with it and move up one order.
1105 */
1106 if (page_is_guard(buddy))
1107 clear_page_guard(zone, buddy, order, migratetype);
1108 else
1109 del_page_from_free_list(buddy, zone, order);
1110 combined_pfn = buddy_pfn & pfn;
1111 page = page + (combined_pfn - pfn);
1112 pfn = combined_pfn;
1113 order++;
1114 }
1115 if (order < MAX_ORDER - 1) {
1116 /* If we are here, it means order is >= pageblock_order.
1117 * We want to prevent merge between freepages on isolate
1118 * pageblock and normal pageblock. Without this, pageblock
1119 * isolation could cause incorrect freepage or CMA accounting.
1120 *
1121 * We don't want to hit this code for the more frequent
1122 * low-order merging.
1123 */
1124 if (unlikely(has_isolate_pageblock(zone))) {
1125 int buddy_mt;
1126
1127 buddy_pfn = __find_buddy_pfn(pfn, order);
1128 buddy = page + (buddy_pfn - pfn);
1129 buddy_mt = get_pageblock_migratetype(buddy);
1130
1131 if (migratetype != buddy_mt
1132 && (is_migrate_isolate(migratetype) ||
1133 is_migrate_isolate(buddy_mt)))
1134 goto done_merging;
1135 }
1136 max_order = order + 1;
1137 goto continue_merging;
1138 }
1139
1140done_merging:
1141 set_buddy_order(page, order);
1142
1143 if (fpi_flags & FPI_TO_TAIL)
1144 to_tail = true;
1145 else if (is_shuffle_order(order))
1146 to_tail = shuffle_pick_tail();
1147 else
1148 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1149
1150 if (to_tail)
1151 add_to_free_list_tail(page, zone, order, migratetype);
1152 else
1153 add_to_free_list(page, zone, order, migratetype);
1154
1155 /* Notify page reporting subsystem of freed page */
1156 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1157 page_reporting_notify_free(order);
1158}
1159
1160/*
1161 * A bad page could be due to a number of fields. Instead of multiple branches,
1162 * try and check multiple fields with one check. The caller must do a detailed
1163 * check if necessary.
1164 */
1165static inline bool page_expected_state(struct page *page,
1166 unsigned long check_flags)
1167{
1168 if (unlikely(atomic_read(&page->_mapcount) != -1))
1169 return false;
1170
1171 if (unlikely((unsigned long)page->mapping |
1172 page_ref_count(page) |
1173#ifdef CONFIG_MEMCG
1174 page->memcg_data |
1175#endif
1176 (page->flags & check_flags)))
1177 return false;
1178
1179 return true;
1180}
1181
1182static const char *page_bad_reason(struct page *page, unsigned long flags)
1183{
1184 const char *bad_reason = NULL;
1185
1186 if (unlikely(atomic_read(&page->_mapcount) != -1))
1187 bad_reason = "nonzero mapcount";
1188 if (unlikely(page->mapping != NULL))
1189 bad_reason = "non-NULL mapping";
1190 if (unlikely(page_ref_count(page) != 0))
1191 bad_reason = "nonzero _refcount";
1192 if (unlikely(page->flags & flags)) {
1193 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1194 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1195 else
1196 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1197 }
1198#ifdef CONFIG_MEMCG
1199 if (unlikely(page->memcg_data))
1200 bad_reason = "page still charged to cgroup";
1201#endif
1202 return bad_reason;
1203}
1204
1205static void check_free_page_bad(struct page *page)
1206{
1207 bad_page(page,
1208 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1209}
1210
1211static inline int check_free_page(struct page *page)
1212{
1213 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1214 return 0;
1215
1216 /* Something has gone sideways, find it */
1217 check_free_page_bad(page);
1218 return 1;
1219}
1220
1221static int free_tail_pages_check(struct page *head_page, struct page *page)
1222{
1223 int ret = 1;
1224
1225 /*
1226 * We rely page->lru.next never has bit 0 set, unless the page
1227 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1228 */
1229 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1230
1231 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1232 ret = 0;
1233 goto out;
1234 }
1235 switch (page - head_page) {
1236 case 1:
1237 /* the first tail page: ->mapping may be compound_mapcount() */
1238 if (unlikely(compound_mapcount(page))) {
1239 bad_page(page, "nonzero compound_mapcount");
1240 goto out;
1241 }
1242 break;
1243 case 2:
1244 /*
1245 * the second tail page: ->mapping is
1246 * deferred_list.next -- ignore value.
1247 */
1248 break;
1249 default:
1250 if (page->mapping != TAIL_MAPPING) {
1251 bad_page(page, "corrupted mapping in tail page");
1252 goto out;
1253 }
1254 break;
1255 }
1256 if (unlikely(!PageTail(page))) {
1257 bad_page(page, "PageTail not set");
1258 goto out;
1259 }
1260 if (unlikely(compound_head(page) != head_page)) {
1261 bad_page(page, "compound_head not consistent");
1262 goto out;
1263 }
1264 ret = 0;
1265out:
1266 page->mapping = NULL;
1267 clear_compound_head(page);
1268 return ret;
1269}
1270
1271static void kernel_init_free_pages(struct page *page, int numpages, bool zero_tags)
1272{
1273 int i;
1274
1275 if (zero_tags) {
1276 for (i = 0; i < numpages; i++)
1277 tag_clear_highpage(page + i);
1278 return;
1279 }
1280
1281 /* s390's use of memset() could override KASAN redzones. */
1282 kasan_disable_current();
1283 for (i = 0; i < numpages; i++) {
1284 u8 tag = page_kasan_tag(page + i);
1285 page_kasan_tag_reset(page + i);
1286 clear_highpage(page + i);
1287 page_kasan_tag_set(page + i, tag);
1288 }
1289 kasan_enable_current();
1290}
1291
1292static __always_inline bool free_pages_prepare(struct page *page,
1293 unsigned int order, bool check_free, fpi_t fpi_flags)
1294{
1295 int bad = 0;
1296 bool skip_kasan_poison = should_skip_kasan_poison(page, fpi_flags);
1297
1298 VM_BUG_ON_PAGE(PageTail(page), page);
1299
1300 trace_mm_page_free(page, order);
1301
1302 if (unlikely(PageHWPoison(page)) && !order) {
1303 /*
1304 * Do not let hwpoison pages hit pcplists/buddy
1305 * Untie memcg state and reset page's owner
1306 */
1307 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1308 __memcg_kmem_uncharge_page(page, order);
1309 reset_page_owner(page, order);
1310 return false;
1311 }
1312
1313 /*
1314 * Check tail pages before head page information is cleared to
1315 * avoid checking PageCompound for order-0 pages.
1316 */
1317 if (unlikely(order)) {
1318 bool compound = PageCompound(page);
1319 int i;
1320
1321 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1322
1323 if (compound)
1324 ClearPageDoubleMap(page);
1325 for (i = 1; i < (1 << order); i++) {
1326 if (compound)
1327 bad += free_tail_pages_check(page, page + i);
1328 if (unlikely(check_free_page(page + i))) {
1329 bad++;
1330 continue;
1331 }
1332 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1333 }
1334 }
1335 if (PageMappingFlags(page))
1336 page->mapping = NULL;
1337 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1338 __memcg_kmem_uncharge_page(page, order);
1339 if (check_free)
1340 bad += check_free_page(page);
1341 if (bad)
1342 return false;
1343
1344 page_cpupid_reset_last(page);
1345 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1346 reset_page_owner(page, order);
1347
1348 if (!PageHighMem(page)) {
1349 debug_check_no_locks_freed(page_address(page),
1350 PAGE_SIZE << order);
1351 debug_check_no_obj_freed(page_address(page),
1352 PAGE_SIZE << order);
1353 }
1354
1355 kernel_poison_pages(page, 1 << order);
1356
1357 /*
1358 * As memory initialization might be integrated into KASAN,
1359 * kasan_free_pages and kernel_init_free_pages must be
1360 * kept together to avoid discrepancies in behavior.
1361 *
1362 * With hardware tag-based KASAN, memory tags must be set before the
1363 * page becomes unavailable via debug_pagealloc or arch_free_page.
1364 */
1365 if (kasan_has_integrated_init()) {
1366 if (!skip_kasan_poison)
1367 kasan_free_pages(page, order);
1368 } else {
1369 bool init = want_init_on_free();
1370
1371 if (init)
1372 kernel_init_free_pages(page, 1 << order, false);
1373 if (!skip_kasan_poison)
1374 kasan_poison_pages(page, order, init);
1375 }
1376
1377 /*
1378 * arch_free_page() can make the page's contents inaccessible. s390
1379 * does this. So nothing which can access the page's contents should
1380 * happen after this.
1381 */
1382 arch_free_page(page, order);
1383
1384 debug_pagealloc_unmap_pages(page, 1 << order);
1385
1386 return true;
1387}
1388
1389#ifdef CONFIG_DEBUG_VM
1390/*
1391 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1392 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1393 * moved from pcp lists to free lists.
1394 */
1395static bool free_pcp_prepare(struct page *page, unsigned int order)
1396{
1397 return free_pages_prepare(page, order, true, FPI_NONE);
1398}
1399
1400static bool bulkfree_pcp_prepare(struct page *page)
1401{
1402 if (debug_pagealloc_enabled_static())
1403 return check_free_page(page);
1404 else
1405 return false;
1406}
1407#else
1408/*
1409 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1410 * moving from pcp lists to free list in order to reduce overhead. With
1411 * debug_pagealloc enabled, they are checked also immediately when being freed
1412 * to the pcp lists.
1413 */
1414static bool free_pcp_prepare(struct page *page, unsigned int order)
1415{
1416 if (debug_pagealloc_enabled_static())
1417 return free_pages_prepare(page, order, true, FPI_NONE);
1418 else
1419 return free_pages_prepare(page, order, false, FPI_NONE);
1420}
1421
1422static bool bulkfree_pcp_prepare(struct page *page)
1423{
1424 return check_free_page(page);
1425}
1426#endif /* CONFIG_DEBUG_VM */
1427
1428static inline void prefetch_buddy(struct page *page)
1429{
1430 unsigned long pfn = page_to_pfn(page);
1431 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1432 struct page *buddy = page + (buddy_pfn - pfn);
1433
1434 prefetch(buddy);
1435}
1436
1437/*
1438 * Frees a number of pages from the PCP lists
1439 * Assumes all pages on list are in same zone, and of same order.
1440 * count is the number of pages to free.
1441 *
1442 * If the zone was previously in an "all pages pinned" state then look to
1443 * see if this freeing clears that state.
1444 *
1445 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1446 * pinned" detection logic.
1447 */
1448static void free_pcppages_bulk(struct zone *zone, int count,
1449 struct per_cpu_pages *pcp)
1450{
1451 int pindex = 0;
1452 int batch_free = 0;
1453 int nr_freed = 0;
1454 unsigned int order;
1455 int prefetch_nr = READ_ONCE(pcp->batch);
1456 bool isolated_pageblocks;
1457 struct page *page, *tmp;
1458 LIST_HEAD(head);
1459
1460 /*
1461 * Ensure proper count is passed which otherwise would stuck in the
1462 * below while (list_empty(list)) loop.
1463 */
1464 count = min(pcp->count, count);
1465 while (count > 0) {
1466 struct list_head *list;
1467
1468 /*
1469 * Remove pages from lists in a round-robin fashion. A
1470 * batch_free count is maintained that is incremented when an
1471 * empty list is encountered. This is so more pages are freed
1472 * off fuller lists instead of spinning excessively around empty
1473 * lists
1474 */
1475 do {
1476 batch_free++;
1477 if (++pindex == NR_PCP_LISTS)
1478 pindex = 0;
1479 list = &pcp->lists[pindex];
1480 } while (list_empty(list));
1481
1482 /* This is the only non-empty list. Free them all. */
1483 if (batch_free == NR_PCP_LISTS)
1484 batch_free = count;
1485
1486 order = pindex_to_order(pindex);
1487 BUILD_BUG_ON(MAX_ORDER >= (1<<NR_PCP_ORDER_WIDTH));
1488 do {
1489 page = list_last_entry(list, struct page, lru);
1490 /* must delete to avoid corrupting pcp list */
1491 list_del(&page->lru);
1492 nr_freed += 1 << order;
1493 count -= 1 << order;
1494
1495 if (bulkfree_pcp_prepare(page))
1496 continue;
1497
1498 /* Encode order with the migratetype */
1499 page->index <<= NR_PCP_ORDER_WIDTH;
1500 page->index |= order;
1501
1502 list_add_tail(&page->lru, &head);
1503
1504 /*
1505 * We are going to put the page back to the global
1506 * pool, prefetch its buddy to speed up later access
1507 * under zone->lock. It is believed the overhead of
1508 * an additional test and calculating buddy_pfn here
1509 * can be offset by reduced memory latency later. To
1510 * avoid excessive prefetching due to large count, only
1511 * prefetch buddy for the first pcp->batch nr of pages.
1512 */
1513 if (prefetch_nr) {
1514 prefetch_buddy(page);
1515 prefetch_nr--;
1516 }
1517 } while (count > 0 && --batch_free && !list_empty(list));
1518 }
1519 pcp->count -= nr_freed;
1520
1521 /*
1522 * local_lock_irq held so equivalent to spin_lock_irqsave for
1523 * both PREEMPT_RT and non-PREEMPT_RT configurations.
1524 */
1525 spin_lock(&zone->lock);
1526 isolated_pageblocks = has_isolate_pageblock(zone);
1527
1528 /*
1529 * Use safe version since after __free_one_page(),
1530 * page->lru.next will not point to original list.
1531 */
1532 list_for_each_entry_safe(page, tmp, &head, lru) {
1533 int mt = get_pcppage_migratetype(page);
1534
1535 /* mt has been encoded with the order (see above) */
1536 order = mt & NR_PCP_ORDER_MASK;
1537 mt >>= NR_PCP_ORDER_WIDTH;
1538
1539 /* MIGRATE_ISOLATE page should not go to pcplists */
1540 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1541 /* Pageblock could have been isolated meanwhile */
1542 if (unlikely(isolated_pageblocks))
1543 mt = get_pageblock_migratetype(page);
1544
1545 __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1546 trace_mm_page_pcpu_drain(page, order, mt);
1547 }
1548 spin_unlock(&zone->lock);
1549}
1550
1551static void free_one_page(struct zone *zone,
1552 struct page *page, unsigned long pfn,
1553 unsigned int order,
1554 int migratetype, fpi_t fpi_flags)
1555{
1556 unsigned long flags;
1557
1558 spin_lock_irqsave(&zone->lock, flags);
1559 if (unlikely(has_isolate_pageblock(zone) ||
1560 is_migrate_isolate(migratetype))) {
1561 migratetype = get_pfnblock_migratetype(page, pfn);
1562 }
1563 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1564 spin_unlock_irqrestore(&zone->lock, flags);
1565}
1566
1567static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1568 unsigned long zone, int nid)
1569{
1570 mm_zero_struct_page(page);
1571 set_page_links(page, zone, nid, pfn);
1572 init_page_count(page);
1573 page_mapcount_reset(page);
1574 page_cpupid_reset_last(page);
1575 page_kasan_tag_reset(page);
1576
1577 INIT_LIST_HEAD(&page->lru);
1578#ifdef WANT_PAGE_VIRTUAL
1579 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1580 if (!is_highmem_idx(zone))
1581 set_page_address(page, __va(pfn << PAGE_SHIFT));
1582#endif
1583}
1584
1585#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1586static void __meminit init_reserved_page(unsigned long pfn)
1587{
1588 pg_data_t *pgdat;
1589 int nid, zid;
1590
1591 if (!early_page_uninitialised(pfn))
1592 return;
1593
1594 nid = early_pfn_to_nid(pfn);
1595 pgdat = NODE_DATA(nid);
1596
1597 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1598 struct zone *zone = &pgdat->node_zones[zid];
1599
1600 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1601 break;
1602 }
1603 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1604}
1605#else
1606static inline void init_reserved_page(unsigned long pfn)
1607{
1608}
1609#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1610
1611/*
1612 * Initialised pages do not have PageReserved set. This function is
1613 * called for each range allocated by the bootmem allocator and
1614 * marks the pages PageReserved. The remaining valid pages are later
1615 * sent to the buddy page allocator.
1616 */
1617void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1618{
1619 unsigned long start_pfn = PFN_DOWN(start);
1620 unsigned long end_pfn = PFN_UP(end);
1621
1622 for (; start_pfn < end_pfn; start_pfn++) {
1623 if (pfn_valid(start_pfn)) {
1624 struct page *page = pfn_to_page(start_pfn);
1625
1626 init_reserved_page(start_pfn);
1627
1628 /* Avoid false-positive PageTail() */
1629 INIT_LIST_HEAD(&page->lru);
1630
1631 /*
1632 * no need for atomic set_bit because the struct
1633 * page is not visible yet so nobody should
1634 * access it yet.
1635 */
1636 __SetPageReserved(page);
1637 }
1638 }
1639}
1640
1641static void __free_pages_ok(struct page *page, unsigned int order,
1642 fpi_t fpi_flags)
1643{
1644 unsigned long flags;
1645 int migratetype;
1646 unsigned long pfn = page_to_pfn(page);
1647 struct zone *zone = page_zone(page);
1648
1649 if (!free_pages_prepare(page, order, true, fpi_flags))
1650 return;
1651
1652 migratetype = get_pfnblock_migratetype(page, pfn);
1653
1654 spin_lock_irqsave(&zone->lock, flags);
1655 if (unlikely(has_isolate_pageblock(zone) ||
1656 is_migrate_isolate(migratetype))) {
1657 migratetype = get_pfnblock_migratetype(page, pfn);
1658 }
1659 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1660 spin_unlock_irqrestore(&zone->lock, flags);
1661
1662 __count_vm_events(PGFREE, 1 << order);
1663}
1664
1665void __free_pages_core(struct page *page, unsigned int order)
1666{
1667 unsigned int nr_pages = 1 << order;
1668 struct page *p = page;
1669 unsigned int loop;
1670
1671 /*
1672 * When initializing the memmap, __init_single_page() sets the refcount
1673 * of all pages to 1 ("allocated"/"not free"). We have to set the
1674 * refcount of all involved pages to 0.
1675 */
1676 prefetchw(p);
1677 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1678 prefetchw(p + 1);
1679 __ClearPageReserved(p);
1680 set_page_count(p, 0);
1681 }
1682 __ClearPageReserved(p);
1683 set_page_count(p, 0);
1684
1685 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1686
1687 /*
1688 * Bypass PCP and place fresh pages right to the tail, primarily
1689 * relevant for memory onlining.
1690 */
1691 __free_pages_ok(page, order, FPI_TO_TAIL | FPI_SKIP_KASAN_POISON);
1692}
1693
1694#ifdef CONFIG_NUMA
1695
1696/*
1697 * During memory init memblocks map pfns to nids. The search is expensive and
1698 * this caches recent lookups. The implementation of __early_pfn_to_nid
1699 * treats start/end as pfns.
1700 */
1701struct mminit_pfnnid_cache {
1702 unsigned long last_start;
1703 unsigned long last_end;
1704 int last_nid;
1705};
1706
1707static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1708
1709/*
1710 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1711 */
1712static int __meminit __early_pfn_to_nid(unsigned long pfn,
1713 struct mminit_pfnnid_cache *state)
1714{
1715 unsigned long start_pfn, end_pfn;
1716 int nid;
1717
1718 if (state->last_start <= pfn && pfn < state->last_end)
1719 return state->last_nid;
1720
1721 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1722 if (nid != NUMA_NO_NODE) {
1723 state->last_start = start_pfn;
1724 state->last_end = end_pfn;
1725 state->last_nid = nid;
1726 }
1727
1728 return nid;
1729}
1730
1731int __meminit early_pfn_to_nid(unsigned long pfn)
1732{
1733 static DEFINE_SPINLOCK(early_pfn_lock);
1734 int nid;
1735
1736 spin_lock(&early_pfn_lock);
1737 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1738 if (nid < 0)
1739 nid = first_online_node;
1740 spin_unlock(&early_pfn_lock);
1741
1742 return nid;
1743}
1744#endif /* CONFIG_NUMA */
1745
1746void __init memblock_free_pages(struct page *page, unsigned long pfn,
1747 unsigned int order)
1748{
1749 if (early_page_uninitialised(pfn))
1750 return;
1751 __free_pages_core(page, order);
1752}
1753
1754/*
1755 * Check that the whole (or subset of) a pageblock given by the interval of
1756 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1757 * with the migration of free compaction scanner. The scanners then need to
1758 * use only pfn_valid_within() check for arches that allow holes within
1759 * pageblocks.
1760 *
1761 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1762 *
1763 * It's possible on some configurations to have a setup like node0 node1 node0
1764 * i.e. it's possible that all pages within a zones range of pages do not
1765 * belong to a single zone. We assume that a border between node0 and node1
1766 * can occur within a single pageblock, but not a node0 node1 node0
1767 * interleaving within a single pageblock. It is therefore sufficient to check
1768 * the first and last page of a pageblock and avoid checking each individual
1769 * page in a pageblock.
1770 */
1771struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1772 unsigned long end_pfn, struct zone *zone)
1773{
1774 struct page *start_page;
1775 struct page *end_page;
1776
1777 /* end_pfn is one past the range we are checking */
1778 end_pfn--;
1779
1780 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1781 return NULL;
1782
1783 start_page = pfn_to_online_page(start_pfn);
1784 if (!start_page)
1785 return NULL;
1786
1787 if (page_zone(start_page) != zone)
1788 return NULL;
1789
1790 end_page = pfn_to_page(end_pfn);
1791
1792 /* This gives a shorter code than deriving page_zone(end_page) */
1793 if (page_zone_id(start_page) != page_zone_id(end_page))
1794 return NULL;
1795
1796 return start_page;
1797}
1798
1799void set_zone_contiguous(struct zone *zone)
1800{
1801 unsigned long block_start_pfn = zone->zone_start_pfn;
1802 unsigned long block_end_pfn;
1803
1804 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1805 for (; block_start_pfn < zone_end_pfn(zone);
1806 block_start_pfn = block_end_pfn,
1807 block_end_pfn += pageblock_nr_pages) {
1808
1809 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1810
1811 if (!__pageblock_pfn_to_page(block_start_pfn,
1812 block_end_pfn, zone))
1813 return;
1814 cond_resched();
1815 }
1816
1817 /* We confirm that there is no hole */
1818 zone->contiguous = true;
1819}
1820
1821void clear_zone_contiguous(struct zone *zone)
1822{
1823 zone->contiguous = false;
1824}
1825
1826#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1827static void __init deferred_free_range(unsigned long pfn,
1828 unsigned long nr_pages)
1829{
1830 struct page *page;
1831 unsigned long i;
1832
1833 if (!nr_pages)
1834 return;
1835
1836 page = pfn_to_page(pfn);
1837
1838 /* Free a large naturally-aligned chunk if possible */
1839 if (nr_pages == pageblock_nr_pages &&
1840 (pfn & (pageblock_nr_pages - 1)) == 0) {
1841 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1842 __free_pages_core(page, pageblock_order);
1843 return;
1844 }
1845
1846 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1847 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1848 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1849 __free_pages_core(page, 0);
1850 }
1851}
1852
1853/* Completion tracking for deferred_init_memmap() threads */
1854static atomic_t pgdat_init_n_undone __initdata;
1855static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1856
1857static inline void __init pgdat_init_report_one_done(void)
1858{
1859 if (atomic_dec_and_test(&pgdat_init_n_undone))
1860 complete(&pgdat_init_all_done_comp);
1861}
1862
1863/*
1864 * Returns true if page needs to be initialized or freed to buddy allocator.
1865 *
1866 * First we check if pfn is valid on architectures where it is possible to have
1867 * holes within pageblock_nr_pages. On systems where it is not possible, this
1868 * function is optimized out.
1869 *
1870 * Then, we check if a current large page is valid by only checking the validity
1871 * of the head pfn.
1872 */
1873static inline bool __init deferred_pfn_valid(unsigned long pfn)
1874{
1875 if (!pfn_valid_within(pfn))
1876 return false;
1877 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1878 return false;
1879 return true;
1880}
1881
1882/*
1883 * Free pages to buddy allocator. Try to free aligned pages in
1884 * pageblock_nr_pages sizes.
1885 */
1886static void __init deferred_free_pages(unsigned long pfn,
1887 unsigned long end_pfn)
1888{
1889 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1890 unsigned long nr_free = 0;
1891
1892 for (; pfn < end_pfn; pfn++) {
1893 if (!deferred_pfn_valid(pfn)) {
1894 deferred_free_range(pfn - nr_free, nr_free);
1895 nr_free = 0;
1896 } else if (!(pfn & nr_pgmask)) {
1897 deferred_free_range(pfn - nr_free, nr_free);
1898 nr_free = 1;
1899 } else {
1900 nr_free++;
1901 }
1902 }
1903 /* Free the last block of pages to allocator */
1904 deferred_free_range(pfn - nr_free, nr_free);
1905}
1906
1907/*
1908 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1909 * by performing it only once every pageblock_nr_pages.
1910 * Return number of pages initialized.
1911 */
1912static unsigned long __init deferred_init_pages(struct zone *zone,
1913 unsigned long pfn,
1914 unsigned long end_pfn)
1915{
1916 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1917 int nid = zone_to_nid(zone);
1918 unsigned long nr_pages = 0;
1919 int zid = zone_idx(zone);
1920 struct page *page = NULL;
1921
1922 for (; pfn < end_pfn; pfn++) {
1923 if (!deferred_pfn_valid(pfn)) {
1924 page = NULL;
1925 continue;
1926 } else if (!page || !(pfn & nr_pgmask)) {
1927 page = pfn_to_page(pfn);
1928 } else {
1929 page++;
1930 }
1931 __init_single_page(page, pfn, zid, nid);
1932 nr_pages++;
1933 }
1934 return (nr_pages);
1935}
1936
1937/*
1938 * This function is meant to pre-load the iterator for the zone init.
1939 * Specifically it walks through the ranges until we are caught up to the
1940 * first_init_pfn value and exits there. If we never encounter the value we
1941 * return false indicating there are no valid ranges left.
1942 */
1943static bool __init
1944deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1945 unsigned long *spfn, unsigned long *epfn,
1946 unsigned long first_init_pfn)
1947{
1948 u64 j;
1949
1950 /*
1951 * Start out by walking through the ranges in this zone that have
1952 * already been initialized. We don't need to do anything with them
1953 * so we just need to flush them out of the system.
1954 */
1955 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1956 if (*epfn <= first_init_pfn)
1957 continue;
1958 if (*spfn < first_init_pfn)
1959 *spfn = first_init_pfn;
1960 *i = j;
1961 return true;
1962 }
1963
1964 return false;
1965}
1966
1967/*
1968 * Initialize and free pages. We do it in two loops: first we initialize
1969 * struct page, then free to buddy allocator, because while we are
1970 * freeing pages we can access pages that are ahead (computing buddy
1971 * page in __free_one_page()).
1972 *
1973 * In order to try and keep some memory in the cache we have the loop
1974 * broken along max page order boundaries. This way we will not cause
1975 * any issues with the buddy page computation.
1976 */
1977static unsigned long __init
1978deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1979 unsigned long *end_pfn)
1980{
1981 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1982 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1983 unsigned long nr_pages = 0;
1984 u64 j = *i;
1985
1986 /* First we loop through and initialize the page values */
1987 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1988 unsigned long t;
1989
1990 if (mo_pfn <= *start_pfn)
1991 break;
1992
1993 t = min(mo_pfn, *end_pfn);
1994 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1995
1996 if (mo_pfn < *end_pfn) {
1997 *start_pfn = mo_pfn;
1998 break;
1999 }
2000 }
2001
2002 /* Reset values and now loop through freeing pages as needed */
2003 swap(j, *i);
2004
2005 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
2006 unsigned long t;
2007
2008 if (mo_pfn <= spfn)
2009 break;
2010
2011 t = min(mo_pfn, epfn);
2012 deferred_free_pages(spfn, t);
2013
2014 if (mo_pfn <= epfn)
2015 break;
2016 }
2017
2018 return nr_pages;
2019}
2020
2021static void __init
2022deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
2023 void *arg)
2024{
2025 unsigned long spfn, epfn;
2026 struct zone *zone = arg;
2027 u64 i;
2028
2029 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
2030
2031 /*
2032 * Initialize and free pages in MAX_ORDER sized increments so that we
2033 * can avoid introducing any issues with the buddy allocator.
2034 */
2035 while (spfn < end_pfn) {
2036 deferred_init_maxorder(&i, zone, &spfn, &epfn);
2037 cond_resched();
2038 }
2039}
2040
2041/* An arch may override for more concurrency. */
2042__weak int __init
2043deferred_page_init_max_threads(const struct cpumask *node_cpumask)
2044{
2045 return 1;
2046}
2047
2048/* Initialise remaining memory on a node */
2049static int __init deferred_init_memmap(void *data)
2050{
2051 pg_data_t *pgdat = data;
2052 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2053 unsigned long spfn = 0, epfn = 0;
2054 unsigned long first_init_pfn, flags;
2055 unsigned long start = jiffies;
2056 struct zone *zone;
2057 int zid, max_threads;
2058 u64 i;
2059
2060 /* Bind memory initialisation thread to a local node if possible */
2061 if (!cpumask_empty(cpumask))
2062 set_cpus_allowed_ptr(current, cpumask);
2063
2064 pgdat_resize_lock(pgdat, &flags);
2065 first_init_pfn = pgdat->first_deferred_pfn;
2066 if (first_init_pfn == ULONG_MAX) {
2067 pgdat_resize_unlock(pgdat, &flags);
2068 pgdat_init_report_one_done();
2069 return 0;
2070 }
2071
2072 /* Sanity check boundaries */
2073 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
2074 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
2075 pgdat->first_deferred_pfn = ULONG_MAX;
2076
2077 /*
2078 * Once we unlock here, the zone cannot be grown anymore, thus if an
2079 * interrupt thread must allocate this early in boot, zone must be
2080 * pre-grown prior to start of deferred page initialization.
2081 */
2082 pgdat_resize_unlock(pgdat, &flags);
2083
2084 /* Only the highest zone is deferred so find it */
2085 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2086 zone = pgdat->node_zones + zid;
2087 if (first_init_pfn < zone_end_pfn(zone))
2088 break;
2089 }
2090
2091 /* If the zone is empty somebody else may have cleared out the zone */
2092 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2093 first_init_pfn))
2094 goto zone_empty;
2095
2096 max_threads = deferred_page_init_max_threads(cpumask);
2097
2098 while (spfn < epfn) {
2099 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
2100 struct padata_mt_job job = {
2101 .thread_fn = deferred_init_memmap_chunk,
2102 .fn_arg = zone,
2103 .start = spfn,
2104 .size = epfn_align - spfn,
2105 .align = PAGES_PER_SECTION,
2106 .min_chunk = PAGES_PER_SECTION,
2107 .max_threads = max_threads,
2108 };
2109
2110 padata_do_multithreaded(&job);
2111 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2112 epfn_align);
2113 }
2114zone_empty:
2115 /* Sanity check that the next zone really is unpopulated */
2116 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2117
2118 pr_info("node %d deferred pages initialised in %ums\n",
2119 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2120
2121 pgdat_init_report_one_done();
2122 return 0;
2123}
2124
2125/*
2126 * If this zone has deferred pages, try to grow it by initializing enough
2127 * deferred pages to satisfy the allocation specified by order, rounded up to
2128 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2129 * of SECTION_SIZE bytes by initializing struct pages in increments of
2130 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2131 *
2132 * Return true when zone was grown, otherwise return false. We return true even
2133 * when we grow less than requested, to let the caller decide if there are
2134 * enough pages to satisfy the allocation.
2135 *
2136 * Note: We use noinline because this function is needed only during boot, and
2137 * it is called from a __ref function _deferred_grow_zone. This way we are
2138 * making sure that it is not inlined into permanent text section.
2139 */
2140static noinline bool __init
2141deferred_grow_zone(struct zone *zone, unsigned int order)
2142{
2143 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2144 pg_data_t *pgdat = zone->zone_pgdat;
2145 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2146 unsigned long spfn, epfn, flags;
2147 unsigned long nr_pages = 0;
2148 u64 i;
2149
2150 /* Only the last zone may have deferred pages */
2151 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2152 return false;
2153
2154 pgdat_resize_lock(pgdat, &flags);
2155
2156 /*
2157 * If someone grew this zone while we were waiting for spinlock, return
2158 * true, as there might be enough pages already.
2159 */
2160 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2161 pgdat_resize_unlock(pgdat, &flags);
2162 return true;
2163 }
2164
2165 /* If the zone is empty somebody else may have cleared out the zone */
2166 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2167 first_deferred_pfn)) {
2168 pgdat->first_deferred_pfn = ULONG_MAX;
2169 pgdat_resize_unlock(pgdat, &flags);
2170 /* Retry only once. */
2171 return first_deferred_pfn != ULONG_MAX;
2172 }
2173
2174 /*
2175 * Initialize and free pages in MAX_ORDER sized increments so
2176 * that we can avoid introducing any issues with the buddy
2177 * allocator.
2178 */
2179 while (spfn < epfn) {
2180 /* update our first deferred PFN for this section */
2181 first_deferred_pfn = spfn;
2182
2183 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2184 touch_nmi_watchdog();
2185
2186 /* We should only stop along section boundaries */
2187 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2188 continue;
2189
2190 /* If our quota has been met we can stop here */
2191 if (nr_pages >= nr_pages_needed)
2192 break;
2193 }
2194
2195 pgdat->first_deferred_pfn = spfn;
2196 pgdat_resize_unlock(pgdat, &flags);
2197
2198 return nr_pages > 0;
2199}
2200
2201/*
2202 * deferred_grow_zone() is __init, but it is called from
2203 * get_page_from_freelist() during early boot until deferred_pages permanently
2204 * disables this call. This is why we have refdata wrapper to avoid warning,
2205 * and to ensure that the function body gets unloaded.
2206 */
2207static bool __ref
2208_deferred_grow_zone(struct zone *zone, unsigned int order)
2209{
2210 return deferred_grow_zone(zone, order);
2211}
2212
2213#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2214
2215void __init page_alloc_init_late(void)
2216{
2217 struct zone *zone;
2218 int nid;
2219
2220#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2221
2222 /* There will be num_node_state(N_MEMORY) threads */
2223 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2224 for_each_node_state(nid, N_MEMORY) {
2225 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2226 }
2227
2228 /* Block until all are initialised */
2229 wait_for_completion(&pgdat_init_all_done_comp);
2230
2231 /*
2232 * We initialized the rest of the deferred pages. Permanently disable
2233 * on-demand struct page initialization.
2234 */
2235 static_branch_disable(&deferred_pages);
2236
2237 /* Reinit limits that are based on free pages after the kernel is up */
2238 files_maxfiles_init();
2239#endif
2240
2241 buffer_init();
2242
2243 /* Discard memblock private memory */
2244 memblock_discard();
2245
2246 for_each_node_state(nid, N_MEMORY)
2247 shuffle_free_memory(NODE_DATA(nid));
2248
2249 for_each_populated_zone(zone)
2250 set_zone_contiguous(zone);
2251}
2252
2253#ifdef CONFIG_CMA
2254/* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2255void __init init_cma_reserved_pageblock(struct page *page)
2256{
2257 unsigned i = pageblock_nr_pages;
2258 struct page *p = page;
2259
2260 do {
2261 __ClearPageReserved(p);
2262 set_page_count(p, 0);
2263 } while (++p, --i);
2264
2265 set_pageblock_migratetype(page, MIGRATE_CMA);
2266
2267 if (pageblock_order >= MAX_ORDER) {
2268 i = pageblock_nr_pages;
2269 p = page;
2270 do {
2271 set_page_refcounted(p);
2272 __free_pages(p, MAX_ORDER - 1);
2273 p += MAX_ORDER_NR_PAGES;
2274 } while (i -= MAX_ORDER_NR_PAGES);
2275 } else {
2276 set_page_refcounted(page);
2277 __free_pages(page, pageblock_order);
2278 }
2279
2280 adjust_managed_page_count(page, pageblock_nr_pages);
2281 page_zone(page)->cma_pages += pageblock_nr_pages;
2282}
2283#endif
2284
2285/*
2286 * The order of subdivision here is critical for the IO subsystem.
2287 * Please do not alter this order without good reasons and regression
2288 * testing. Specifically, as large blocks of memory are subdivided,
2289 * the order in which smaller blocks are delivered depends on the order
2290 * they're subdivided in this function. This is the primary factor
2291 * influencing the order in which pages are delivered to the IO
2292 * subsystem according to empirical testing, and this is also justified
2293 * by considering the behavior of a buddy system containing a single
2294 * large block of memory acted on by a series of small allocations.
2295 * This behavior is a critical factor in sglist merging's success.
2296 *
2297 * -- nyc
2298 */
2299static inline void expand(struct zone *zone, struct page *page,
2300 int low, int high, int migratetype)
2301{
2302 unsigned long size = 1 << high;
2303
2304 while (high > low) {
2305 high--;
2306 size >>= 1;
2307 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2308
2309 /*
2310 * Mark as guard pages (or page), that will allow to
2311 * merge back to allocator when buddy will be freed.
2312 * Corresponding page table entries will not be touched,
2313 * pages will stay not present in virtual address space
2314 */
2315 if (set_page_guard(zone, &page[size], high, migratetype))
2316 continue;
2317
2318 add_to_free_list(&page[size], zone, high, migratetype);
2319 set_buddy_order(&page[size], high);
2320 }
2321}
2322
2323static void check_new_page_bad(struct page *page)
2324{
2325 if (unlikely(page->flags & __PG_HWPOISON)) {
2326 /* Don't complain about hwpoisoned pages */
2327 page_mapcount_reset(page); /* remove PageBuddy */
2328 return;
2329 }
2330
2331 bad_page(page,
2332 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2333}
2334
2335/*
2336 * This page is about to be returned from the page allocator
2337 */
2338static inline int check_new_page(struct page *page)
2339{
2340 if (likely(page_expected_state(page,
2341 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2342 return 0;
2343
2344 check_new_page_bad(page);
2345 return 1;
2346}
2347
2348#ifdef CONFIG_DEBUG_VM
2349/*
2350 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2351 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2352 * also checked when pcp lists are refilled from the free lists.
2353 */
2354static inline bool check_pcp_refill(struct page *page)
2355{
2356 if (debug_pagealloc_enabled_static())
2357 return check_new_page(page);
2358 else
2359 return false;
2360}
2361
2362static inline bool check_new_pcp(struct page *page)
2363{
2364 return check_new_page(page);
2365}
2366#else
2367/*
2368 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2369 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2370 * enabled, they are also checked when being allocated from the pcp lists.
2371 */
2372static inline bool check_pcp_refill(struct page *page)
2373{
2374 return check_new_page(page);
2375}
2376static inline bool check_new_pcp(struct page *page)
2377{
2378 if (debug_pagealloc_enabled_static())
2379 return check_new_page(page);
2380 else
2381 return false;
2382}
2383#endif /* CONFIG_DEBUG_VM */
2384
2385static bool check_new_pages(struct page *page, unsigned int order)
2386{
2387 int i;
2388 for (i = 0; i < (1 << order); i++) {
2389 struct page *p = page + i;
2390
2391 if (unlikely(check_new_page(p)))
2392 return true;
2393 }
2394
2395 return false;
2396}
2397
2398inline void post_alloc_hook(struct page *page, unsigned int order,
2399 gfp_t gfp_flags)
2400{
2401 set_page_private(page, 0);
2402 set_page_refcounted(page);
2403
2404 arch_alloc_page(page, order);
2405 debug_pagealloc_map_pages(page, 1 << order);
2406
2407 /*
2408 * Page unpoisoning must happen before memory initialization.
2409 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2410 * allocations and the page unpoisoning code will complain.
2411 */
2412 kernel_unpoison_pages(page, 1 << order);
2413
2414 /*
2415 * As memory initialization might be integrated into KASAN,
2416 * kasan_alloc_pages and kernel_init_free_pages must be
2417 * kept together to avoid discrepancies in behavior.
2418 */
2419 if (kasan_has_integrated_init()) {
2420 kasan_alloc_pages(page, order, gfp_flags);
2421 } else {
2422 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags);
2423
2424 kasan_unpoison_pages(page, order, init);
2425 if (init)
2426 kernel_init_free_pages(page, 1 << order,
2427 gfp_flags & __GFP_ZEROTAGS);
2428 }
2429
2430 set_page_owner(page, order, gfp_flags);
2431}
2432
2433static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2434 unsigned int alloc_flags)
2435{
2436 post_alloc_hook(page, order, gfp_flags);
2437
2438 if (order && (gfp_flags & __GFP_COMP))
2439 prep_compound_page(page, order);
2440
2441 /*
2442 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2443 * allocate the page. The expectation is that the caller is taking
2444 * steps that will free more memory. The caller should avoid the page
2445 * being used for !PFMEMALLOC purposes.
2446 */
2447 if (alloc_flags & ALLOC_NO_WATERMARKS)
2448 set_page_pfmemalloc(page);
2449 else
2450 clear_page_pfmemalloc(page);
2451}
2452
2453/*
2454 * Go through the free lists for the given migratetype and remove
2455 * the smallest available page from the freelists
2456 */
2457static __always_inline
2458struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2459 int migratetype)
2460{
2461 unsigned int current_order;
2462 struct free_area *area;
2463 struct page *page;
2464
2465 /* Find a page of the appropriate size in the preferred list */
2466 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2467 area = &(zone->free_area[current_order]);
2468 page = get_page_from_free_area(area, migratetype);
2469 if (!page)
2470 continue;
2471 del_page_from_free_list(page, zone, current_order);
2472 expand(zone, page, order, current_order, migratetype);
2473 set_pcppage_migratetype(page, migratetype);
2474 return page;
2475 }
2476
2477 return NULL;
2478}
2479
2480
2481/*
2482 * This array describes the order lists are fallen back to when
2483 * the free lists for the desirable migrate type are depleted
2484 */
2485static int fallbacks[MIGRATE_TYPES][3] = {
2486 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2487 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2488 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2489#ifdef CONFIG_CMA
2490 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2491#endif
2492#ifdef CONFIG_MEMORY_ISOLATION
2493 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2494#endif
2495};
2496
2497#ifdef CONFIG_CMA
2498static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2499 unsigned int order)
2500{
2501 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2502}
2503#else
2504static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2505 unsigned int order) { return NULL; }
2506#endif
2507
2508/*
2509 * Move the free pages in a range to the freelist tail of the requested type.
2510 * Note that start_page and end_pages are not aligned on a pageblock
2511 * boundary. If alignment is required, use move_freepages_block()
2512 */
2513static int move_freepages(struct zone *zone,
2514 unsigned long start_pfn, unsigned long end_pfn,
2515 int migratetype, int *num_movable)
2516{
2517 struct page *page;
2518 unsigned long pfn;
2519 unsigned int order;
2520 int pages_moved = 0;
2521
2522 for (pfn = start_pfn; pfn <= end_pfn;) {
2523 if (!pfn_valid_within(pfn)) {
2524 pfn++;
2525 continue;
2526 }
2527
2528 page = pfn_to_page(pfn);
2529 if (!PageBuddy(page)) {
2530 /*
2531 * We assume that pages that could be isolated for
2532 * migration are movable. But we don't actually try
2533 * isolating, as that would be expensive.
2534 */
2535 if (num_movable &&
2536 (PageLRU(page) || __PageMovable(page)))
2537 (*num_movable)++;
2538 pfn++;
2539 continue;
2540 }
2541
2542 /* Make sure we are not inadvertently changing nodes */
2543 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2544 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2545
2546 order = buddy_order(page);
2547 move_to_free_list(page, zone, order, migratetype);
2548 pfn += 1 << order;
2549 pages_moved += 1 << order;
2550 }
2551
2552 return pages_moved;
2553}
2554
2555int move_freepages_block(struct zone *zone, struct page *page,
2556 int migratetype, int *num_movable)
2557{
2558 unsigned long start_pfn, end_pfn, pfn;
2559
2560 if (num_movable)
2561 *num_movable = 0;
2562
2563 pfn = page_to_pfn(page);
2564 start_pfn = pfn & ~(pageblock_nr_pages - 1);
2565 end_pfn = start_pfn + pageblock_nr_pages - 1;
2566
2567 /* Do not cross zone boundaries */
2568 if (!zone_spans_pfn(zone, start_pfn))
2569 start_pfn = pfn;
2570 if (!zone_spans_pfn(zone, end_pfn))
2571 return 0;
2572
2573 return move_freepages(zone, start_pfn, end_pfn, migratetype,
2574 num_movable);
2575}
2576
2577static void change_pageblock_range(struct page *pageblock_page,
2578 int start_order, int migratetype)
2579{
2580 int nr_pageblocks = 1 << (start_order - pageblock_order);
2581
2582 while (nr_pageblocks--) {
2583 set_pageblock_migratetype(pageblock_page, migratetype);
2584 pageblock_page += pageblock_nr_pages;
2585 }
2586}
2587
2588/*
2589 * When we are falling back to another migratetype during allocation, try to
2590 * steal extra free pages from the same pageblocks to satisfy further
2591 * allocations, instead of polluting multiple pageblocks.
2592 *
2593 * If we are stealing a relatively large buddy page, it is likely there will
2594 * be more free pages in the pageblock, so try to steal them all. For
2595 * reclaimable and unmovable allocations, we steal regardless of page size,
2596 * as fragmentation caused by those allocations polluting movable pageblocks
2597 * is worse than movable allocations stealing from unmovable and reclaimable
2598 * pageblocks.
2599 */
2600static bool can_steal_fallback(unsigned int order, int start_mt)
2601{
2602 /*
2603 * Leaving this order check is intended, although there is
2604 * relaxed order check in next check. The reason is that
2605 * we can actually steal whole pageblock if this condition met,
2606 * but, below check doesn't guarantee it and that is just heuristic
2607 * so could be changed anytime.
2608 */
2609 if (order >= pageblock_order)
2610 return true;
2611
2612 if (order >= pageblock_order / 2 ||
2613 start_mt == MIGRATE_RECLAIMABLE ||
2614 start_mt == MIGRATE_UNMOVABLE ||
2615 page_group_by_mobility_disabled)
2616 return true;
2617
2618 return false;
2619}
2620
2621static inline bool boost_watermark(struct zone *zone)
2622{
2623 unsigned long max_boost;
2624
2625 if (!watermark_boost_factor)
2626 return false;
2627 /*
2628 * Don't bother in zones that are unlikely to produce results.
2629 * On small machines, including kdump capture kernels running
2630 * in a small area, boosting the watermark can cause an out of
2631 * memory situation immediately.
2632 */
2633 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2634 return false;
2635
2636 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2637 watermark_boost_factor, 10000);
2638
2639 /*
2640 * high watermark may be uninitialised if fragmentation occurs
2641 * very early in boot so do not boost. We do not fall
2642 * through and boost by pageblock_nr_pages as failing
2643 * allocations that early means that reclaim is not going
2644 * to help and it may even be impossible to reclaim the
2645 * boosted watermark resulting in a hang.
2646 */
2647 if (!max_boost)
2648 return false;
2649
2650 max_boost = max(pageblock_nr_pages, max_boost);
2651
2652 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2653 max_boost);
2654
2655 return true;
2656}
2657
2658/*
2659 * This function implements actual steal behaviour. If order is large enough,
2660 * we can steal whole pageblock. If not, we first move freepages in this
2661 * pageblock to our migratetype and determine how many already-allocated pages
2662 * are there in the pageblock with a compatible migratetype. If at least half
2663 * of pages are free or compatible, we can change migratetype of the pageblock
2664 * itself, so pages freed in the future will be put on the correct free list.
2665 */
2666static void steal_suitable_fallback(struct zone *zone, struct page *page,
2667 unsigned int alloc_flags, int start_type, bool whole_block)
2668{
2669 unsigned int current_order = buddy_order(page);
2670 int free_pages, movable_pages, alike_pages;
2671 int old_block_type;
2672
2673 old_block_type = get_pageblock_migratetype(page);
2674
2675 /*
2676 * This can happen due to races and we want to prevent broken
2677 * highatomic accounting.
2678 */
2679 if (is_migrate_highatomic(old_block_type))
2680 goto single_page;
2681
2682 /* Take ownership for orders >= pageblock_order */
2683 if (current_order >= pageblock_order) {
2684 change_pageblock_range(page, current_order, start_type);
2685 goto single_page;
2686 }
2687
2688 /*
2689 * Boost watermarks to increase reclaim pressure to reduce the
2690 * likelihood of future fallbacks. Wake kswapd now as the node
2691 * may be balanced overall and kswapd will not wake naturally.
2692 */
2693 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2694 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2695
2696 /* We are not allowed to try stealing from the whole block */
2697 if (!whole_block)
2698 goto single_page;
2699
2700 free_pages = move_freepages_block(zone, page, start_type,
2701 &movable_pages);
2702 /*
2703 * Determine how many pages are compatible with our allocation.
2704 * For movable allocation, it's the number of movable pages which
2705 * we just obtained. For other types it's a bit more tricky.
2706 */
2707 if (start_type == MIGRATE_MOVABLE) {
2708 alike_pages = movable_pages;
2709 } else {
2710 /*
2711 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2712 * to MOVABLE pageblock, consider all non-movable pages as
2713 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2714 * vice versa, be conservative since we can't distinguish the
2715 * exact migratetype of non-movable pages.
2716 */
2717 if (old_block_type == MIGRATE_MOVABLE)
2718 alike_pages = pageblock_nr_pages
2719 - (free_pages + movable_pages);
2720 else
2721 alike_pages = 0;
2722 }
2723
2724 /* moving whole block can fail due to zone boundary conditions */
2725 if (!free_pages)
2726 goto single_page;
2727
2728 /*
2729 * If a sufficient number of pages in the block are either free or of
2730 * comparable migratability as our allocation, claim the whole block.
2731 */
2732 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2733 page_group_by_mobility_disabled)
2734 set_pageblock_migratetype(page, start_type);
2735
2736 return;
2737
2738single_page:
2739 move_to_free_list(page, zone, current_order, start_type);
2740}
2741
2742/*
2743 * Check whether there is a suitable fallback freepage with requested order.
2744 * If only_stealable is true, this function returns fallback_mt only if
2745 * we can steal other freepages all together. This would help to reduce
2746 * fragmentation due to mixed migratetype pages in one pageblock.
2747 */
2748int find_suitable_fallback(struct free_area *area, unsigned int order,
2749 int migratetype, bool only_stealable, bool *can_steal)
2750{
2751 int i;
2752 int fallback_mt;
2753
2754 if (area->nr_free == 0)
2755 return -1;
2756
2757 *can_steal = false;
2758 for (i = 0;; i++) {
2759 fallback_mt = fallbacks[migratetype][i];
2760 if (fallback_mt == MIGRATE_TYPES)
2761 break;
2762
2763 if (free_area_empty(area, fallback_mt))
2764 continue;
2765
2766 if (can_steal_fallback(order, migratetype))
2767 *can_steal = true;
2768
2769 if (!only_stealable)
2770 return fallback_mt;
2771
2772 if (*can_steal)
2773 return fallback_mt;
2774 }
2775
2776 return -1;
2777}
2778
2779/*
2780 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2781 * there are no empty page blocks that contain a page with a suitable order
2782 */
2783static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2784 unsigned int alloc_order)
2785{
2786 int mt;
2787 unsigned long max_managed, flags;
2788
2789 /*
2790 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2791 * Check is race-prone but harmless.
2792 */
2793 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2794 if (zone->nr_reserved_highatomic >= max_managed)
2795 return;
2796
2797 spin_lock_irqsave(&zone->lock, flags);
2798
2799 /* Recheck the nr_reserved_highatomic limit under the lock */
2800 if (zone->nr_reserved_highatomic >= max_managed)
2801 goto out_unlock;
2802
2803 /* Yoink! */
2804 mt = get_pageblock_migratetype(page);
2805 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2806 && !is_migrate_cma(mt)) {
2807 zone->nr_reserved_highatomic += pageblock_nr_pages;
2808 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2809 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2810 }
2811
2812out_unlock:
2813 spin_unlock_irqrestore(&zone->lock, flags);
2814}
2815
2816/*
2817 * Used when an allocation is about to fail under memory pressure. This
2818 * potentially hurts the reliability of high-order allocations when under
2819 * intense memory pressure but failed atomic allocations should be easier
2820 * to recover from than an OOM.
2821 *
2822 * If @force is true, try to unreserve a pageblock even though highatomic
2823 * pageblock is exhausted.
2824 */
2825static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2826 bool force)
2827{
2828 struct zonelist *zonelist = ac->zonelist;
2829 unsigned long flags;
2830 struct zoneref *z;
2831 struct zone *zone;
2832 struct page *page;
2833 int order;
2834 bool ret;
2835
2836 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2837 ac->nodemask) {
2838 /*
2839 * Preserve at least one pageblock unless memory pressure
2840 * is really high.
2841 */
2842 if (!force && zone->nr_reserved_highatomic <=
2843 pageblock_nr_pages)
2844 continue;
2845
2846 spin_lock_irqsave(&zone->lock, flags);
2847 for (order = 0; order < MAX_ORDER; order++) {
2848 struct free_area *area = &(zone->free_area[order]);
2849
2850 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2851 if (!page)
2852 continue;
2853
2854 /*
2855 * In page freeing path, migratetype change is racy so
2856 * we can counter several free pages in a pageblock
2857 * in this loop although we changed the pageblock type
2858 * from highatomic to ac->migratetype. So we should
2859 * adjust the count once.
2860 */
2861 if (is_migrate_highatomic_page(page)) {
2862 /*
2863 * It should never happen but changes to
2864 * locking could inadvertently allow a per-cpu
2865 * drain to add pages to MIGRATE_HIGHATOMIC
2866 * while unreserving so be safe and watch for
2867 * underflows.
2868 */
2869 zone->nr_reserved_highatomic -= min(
2870 pageblock_nr_pages,
2871 zone->nr_reserved_highatomic);
2872 }
2873
2874 /*
2875 * Convert to ac->migratetype and avoid the normal
2876 * pageblock stealing heuristics. Minimally, the caller
2877 * is doing the work and needs the pages. More
2878 * importantly, if the block was always converted to
2879 * MIGRATE_UNMOVABLE or another type then the number
2880 * of pageblocks that cannot be completely freed
2881 * may increase.
2882 */
2883 set_pageblock_migratetype(page, ac->migratetype);
2884 ret = move_freepages_block(zone, page, ac->migratetype,
2885 NULL);
2886 if (ret) {
2887 spin_unlock_irqrestore(&zone->lock, flags);
2888 return ret;
2889 }
2890 }
2891 spin_unlock_irqrestore(&zone->lock, flags);
2892 }
2893
2894 return false;
2895}
2896
2897/*
2898 * Try finding a free buddy page on the fallback list and put it on the free
2899 * list of requested migratetype, possibly along with other pages from the same
2900 * block, depending on fragmentation avoidance heuristics. Returns true if
2901 * fallback was found so that __rmqueue_smallest() can grab it.
2902 *
2903 * The use of signed ints for order and current_order is a deliberate
2904 * deviation from the rest of this file, to make the for loop
2905 * condition simpler.
2906 */
2907static __always_inline bool
2908__rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2909 unsigned int alloc_flags)
2910{
2911 struct free_area *area;
2912 int current_order;
2913 int min_order = order;
2914 struct page *page;
2915 int fallback_mt;
2916 bool can_steal;
2917
2918 /*
2919 * Do not steal pages from freelists belonging to other pageblocks
2920 * i.e. orders < pageblock_order. If there are no local zones free,
2921 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2922 */
2923 if (alloc_flags & ALLOC_NOFRAGMENT)
2924 min_order = pageblock_order;
2925
2926 /*
2927 * Find the largest available free page in the other list. This roughly
2928 * approximates finding the pageblock with the most free pages, which
2929 * would be too costly to do exactly.
2930 */
2931 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2932 --current_order) {
2933 area = &(zone->free_area[current_order]);
2934 fallback_mt = find_suitable_fallback(area, current_order,
2935 start_migratetype, false, &can_steal);
2936 if (fallback_mt == -1)
2937 continue;
2938
2939 /*
2940 * We cannot steal all free pages from the pageblock and the
2941 * requested migratetype is movable. In that case it's better to
2942 * steal and split the smallest available page instead of the
2943 * largest available page, because even if the next movable
2944 * allocation falls back into a different pageblock than this
2945 * one, it won't cause permanent fragmentation.
2946 */
2947 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2948 && current_order > order)
2949 goto find_smallest;
2950
2951 goto do_steal;
2952 }
2953
2954 return false;
2955
2956find_smallest:
2957 for (current_order = order; current_order < MAX_ORDER;
2958 current_order++) {
2959 area = &(zone->free_area[current_order]);
2960 fallback_mt = find_suitable_fallback(area, current_order,
2961 start_migratetype, false, &can_steal);
2962 if (fallback_mt != -1)
2963 break;
2964 }
2965
2966 /*
2967 * This should not happen - we already found a suitable fallback
2968 * when looking for the largest page.
2969 */
2970 VM_BUG_ON(current_order == MAX_ORDER);
2971
2972do_steal:
2973 page = get_page_from_free_area(area, fallback_mt);
2974
2975 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2976 can_steal);
2977
2978 trace_mm_page_alloc_extfrag(page, order, current_order,
2979 start_migratetype, fallback_mt);
2980
2981 return true;
2982
2983}
2984
2985/*
2986 * Do the hard work of removing an element from the buddy allocator.
2987 * Call me with the zone->lock already held.
2988 */
2989static __always_inline struct page *
2990__rmqueue(struct zone *zone, unsigned int order, int migratetype,
2991 unsigned int alloc_flags)
2992{
2993 struct page *page;
2994
2995 if (IS_ENABLED(CONFIG_CMA)) {
2996 /*
2997 * Balance movable allocations between regular and CMA areas by
2998 * allocating from CMA when over half of the zone's free memory
2999 * is in the CMA area.
3000 */
3001 if (alloc_flags & ALLOC_CMA &&
3002 zone_page_state(zone, NR_FREE_CMA_PAGES) >
3003 zone_page_state(zone, NR_FREE_PAGES) / 2) {
3004 page = __rmqueue_cma_fallback(zone, order);
3005 if (page)
3006 goto out;
3007 }
3008 }
3009retry:
3010 page = __rmqueue_smallest(zone, order, migratetype);
3011 if (unlikely(!page)) {
3012 if (alloc_flags & ALLOC_CMA)
3013 page = __rmqueue_cma_fallback(zone, order);
3014
3015 if (!page && __rmqueue_fallback(zone, order, migratetype,
3016 alloc_flags))
3017 goto retry;
3018 }
3019out:
3020 if (page)
3021 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3022 return page;
3023}
3024
3025/*
3026 * Obtain a specified number of elements from the buddy allocator, all under
3027 * a single hold of the lock, for efficiency. Add them to the supplied list.
3028 * Returns the number of new pages which were placed at *list.
3029 */
3030static int rmqueue_bulk(struct zone *zone, unsigned int order,
3031 unsigned long count, struct list_head *list,
3032 int migratetype, unsigned int alloc_flags)
3033{
3034 int i, allocated = 0;
3035
3036 /*
3037 * local_lock_irq held so equivalent to spin_lock_irqsave for
3038 * both PREEMPT_RT and non-PREEMPT_RT configurations.
3039 */
3040 spin_lock(&zone->lock);
3041 for (i = 0; i < count; ++i) {
3042 struct page *page = __rmqueue(zone, order, migratetype,
3043 alloc_flags);
3044 if (unlikely(page == NULL))
3045 break;
3046
3047 if (unlikely(check_pcp_refill(page)))
3048 continue;
3049
3050 /*
3051 * Split buddy pages returned by expand() are received here in
3052 * physical page order. The page is added to the tail of
3053 * caller's list. From the callers perspective, the linked list
3054 * is ordered by page number under some conditions. This is
3055 * useful for IO devices that can forward direction from the
3056 * head, thus also in the physical page order. This is useful
3057 * for IO devices that can merge IO requests if the physical
3058 * pages are ordered properly.
3059 */
3060 list_add_tail(&page->lru, list);
3061 allocated++;
3062 if (is_migrate_cma(get_pcppage_migratetype(page)))
3063 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
3064 -(1 << order));
3065 }
3066
3067 /*
3068 * i pages were removed from the buddy list even if some leak due
3069 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
3070 * on i. Do not confuse with 'allocated' which is the number of
3071 * pages added to the pcp list.
3072 */
3073 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
3074 spin_unlock(&zone->lock);
3075 return allocated;
3076}
3077
3078#ifdef CONFIG_NUMA
3079/*
3080 * Called from the vmstat counter updater to drain pagesets of this
3081 * currently executing processor on remote nodes after they have
3082 * expired.
3083 *
3084 * Note that this function must be called with the thread pinned to
3085 * a single processor.
3086 */
3087void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
3088{
3089 unsigned long flags;
3090 int to_drain, batch;
3091
3092 local_lock_irqsave(&pagesets.lock, flags);
3093 batch = READ_ONCE(pcp->batch);
3094 to_drain = min(pcp->count, batch);
3095 if (to_drain > 0)
3096 free_pcppages_bulk(zone, to_drain, pcp);
3097 local_unlock_irqrestore(&pagesets.lock, flags);
3098}
3099#endif
3100
3101/*
3102 * Drain pcplists of the indicated processor and zone.
3103 *
3104 * The processor must either be the current processor and the
3105 * thread pinned to the current processor or a processor that
3106 * is not online.
3107 */
3108static void drain_pages_zone(unsigned int cpu, struct zone *zone)
3109{
3110 unsigned long flags;
3111 struct per_cpu_pages *pcp;
3112
3113 local_lock_irqsave(&pagesets.lock, flags);
3114
3115 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3116 if (pcp->count)
3117 free_pcppages_bulk(zone, pcp->count, pcp);
3118
3119 local_unlock_irqrestore(&pagesets.lock, flags);
3120}
3121
3122/*
3123 * Drain pcplists of all zones on the indicated processor.
3124 *
3125 * The processor must either be the current processor and the
3126 * thread pinned to the current processor or a processor that
3127 * is not online.
3128 */
3129static void drain_pages(unsigned int cpu)
3130{
3131 struct zone *zone;
3132
3133 for_each_populated_zone(zone) {
3134 drain_pages_zone(cpu, zone);
3135 }
3136}
3137
3138/*
3139 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3140 *
3141 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3142 * the single zone's pages.
3143 */
3144void drain_local_pages(struct zone *zone)
3145{
3146 int cpu = smp_processor_id();
3147
3148 if (zone)
3149 drain_pages_zone(cpu, zone);
3150 else
3151 drain_pages(cpu);
3152}
3153
3154static void drain_local_pages_wq(struct work_struct *work)
3155{
3156 struct pcpu_drain *drain;
3157
3158 drain = container_of(work, struct pcpu_drain, work);
3159
3160 /*
3161 * drain_all_pages doesn't use proper cpu hotplug protection so
3162 * we can race with cpu offline when the WQ can move this from
3163 * a cpu pinned worker to an unbound one. We can operate on a different
3164 * cpu which is alright but we also have to make sure to not move to
3165 * a different one.
3166 */
3167 preempt_disable();
3168 drain_local_pages(drain->zone);
3169 preempt_enable();
3170}
3171
3172/*
3173 * The implementation of drain_all_pages(), exposing an extra parameter to
3174 * drain on all cpus.
3175 *
3176 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3177 * not empty. The check for non-emptiness can however race with a free to
3178 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3179 * that need the guarantee that every CPU has drained can disable the
3180 * optimizing racy check.
3181 */
3182static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3183{
3184 int cpu;
3185
3186 /*
3187 * Allocate in the BSS so we won't require allocation in
3188 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3189 */
3190 static cpumask_t cpus_with_pcps;
3191
3192 /*
3193 * Make sure nobody triggers this path before mm_percpu_wq is fully
3194 * initialized.
3195 */
3196 if (WARN_ON_ONCE(!mm_percpu_wq))
3197 return;
3198
3199 /*
3200 * Do not drain if one is already in progress unless it's specific to
3201 * a zone. Such callers are primarily CMA and memory hotplug and need
3202 * the drain to be complete when the call returns.
3203 */
3204 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3205 if (!zone)
3206 return;
3207 mutex_lock(&pcpu_drain_mutex);
3208 }
3209
3210 /*
3211 * We don't care about racing with CPU hotplug event
3212 * as offline notification will cause the notified
3213 * cpu to drain that CPU pcps and on_each_cpu_mask
3214 * disables preemption as part of its processing
3215 */
3216 for_each_online_cpu(cpu) {
3217 struct per_cpu_pages *pcp;
3218 struct zone *z;
3219 bool has_pcps = false;
3220
3221 if (force_all_cpus) {
3222 /*
3223 * The pcp.count check is racy, some callers need a
3224 * guarantee that no cpu is missed.
3225 */
3226 has_pcps = true;
3227 } else if (zone) {
3228 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3229 if (pcp->count)
3230 has_pcps = true;
3231 } else {
3232 for_each_populated_zone(z) {
3233 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
3234 if (pcp->count) {
3235 has_pcps = true;
3236 break;
3237 }
3238 }
3239 }
3240
3241 if (has_pcps)
3242 cpumask_set_cpu(cpu, &cpus_with_pcps);
3243 else
3244 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3245 }
3246
3247 for_each_cpu(cpu, &cpus_with_pcps) {
3248 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3249
3250 drain->zone = zone;
3251 INIT_WORK(&drain->work, drain_local_pages_wq);
3252 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3253 }
3254 for_each_cpu(cpu, &cpus_with_pcps)
3255 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3256
3257 mutex_unlock(&pcpu_drain_mutex);
3258}
3259
3260/*
3261 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3262 *
3263 * When zone parameter is non-NULL, spill just the single zone's pages.
3264 *
3265 * Note that this can be extremely slow as the draining happens in a workqueue.
3266 */
3267void drain_all_pages(struct zone *zone)
3268{
3269 __drain_all_pages(zone, false);
3270}
3271
3272#ifdef CONFIG_HIBERNATION
3273
3274/*
3275 * Touch the watchdog for every WD_PAGE_COUNT pages.
3276 */
3277#define WD_PAGE_COUNT (128*1024)
3278
3279void mark_free_pages(struct zone *zone)
3280{
3281 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3282 unsigned long flags;
3283 unsigned int order, t;
3284 struct page *page;
3285
3286 if (zone_is_empty(zone))
3287 return;
3288
3289 spin_lock_irqsave(&zone->lock, flags);
3290
3291 max_zone_pfn = zone_end_pfn(zone);
3292 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3293 if (pfn_valid(pfn)) {
3294 page = pfn_to_page(pfn);
3295
3296 if (!--page_count) {
3297 touch_nmi_watchdog();
3298 page_count = WD_PAGE_COUNT;
3299 }
3300
3301 if (page_zone(page) != zone)
3302 continue;
3303
3304 if (!swsusp_page_is_forbidden(page))
3305 swsusp_unset_page_free(page);
3306 }
3307
3308 for_each_migratetype_order(order, t) {
3309 list_for_each_entry(page,
3310 &zone->free_area[order].free_list[t], lru) {
3311 unsigned long i;
3312
3313 pfn = page_to_pfn(page);
3314 for (i = 0; i < (1UL << order); i++) {
3315 if (!--page_count) {
3316 touch_nmi_watchdog();
3317 page_count = WD_PAGE_COUNT;
3318 }
3319 swsusp_set_page_free(pfn_to_page(pfn + i));
3320 }
3321 }
3322 }
3323 spin_unlock_irqrestore(&zone->lock, flags);
3324}
3325#endif /* CONFIG_PM */
3326
3327static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
3328 unsigned int order)
3329{
3330 int migratetype;
3331
3332 if (!free_pcp_prepare(page, order))
3333 return false;
3334
3335 migratetype = get_pfnblock_migratetype(page, pfn);
3336 set_pcppage_migratetype(page, migratetype);
3337 return true;
3338}
3339
3340static int nr_pcp_free(struct per_cpu_pages *pcp, int high, int batch)
3341{
3342 int min_nr_free, max_nr_free;
3343
3344 /* Check for PCP disabled or boot pageset */
3345 if (unlikely(high < batch))
3346 return 1;
3347
3348 /* Leave at least pcp->batch pages on the list */
3349 min_nr_free = batch;
3350 max_nr_free = high - batch;
3351
3352 /*
3353 * Double the number of pages freed each time there is subsequent
3354 * freeing of pages without any allocation.
3355 */
3356 batch <<= pcp->free_factor;
3357 if (batch < max_nr_free)
3358 pcp->free_factor++;
3359 batch = clamp(batch, min_nr_free, max_nr_free);
3360
3361 return batch;
3362}
3363
3364static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone)
3365{
3366 int high = READ_ONCE(pcp->high);
3367
3368 if (unlikely(!high))
3369 return 0;
3370
3371 if (!test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags))
3372 return high;
3373
3374 /*
3375 * If reclaim is active, limit the number of pages that can be
3376 * stored on pcp lists
3377 */
3378 return min(READ_ONCE(pcp->batch) << 2, high);
3379}
3380
3381static void free_unref_page_commit(struct page *page, unsigned long pfn,
3382 int migratetype, unsigned int order)
3383{
3384 struct zone *zone = page_zone(page);
3385 struct per_cpu_pages *pcp;
3386 int high;
3387 int pindex;
3388
3389 __count_vm_event(PGFREE);
3390 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3391 pindex = order_to_pindex(migratetype, order);
3392 list_add(&page->lru, &pcp->lists[pindex]);
3393 pcp->count += 1 << order;
3394 high = nr_pcp_high(pcp, zone);
3395 if (pcp->count >= high) {
3396 int batch = READ_ONCE(pcp->batch);
3397
3398 free_pcppages_bulk(zone, nr_pcp_free(pcp, high, batch), pcp);
3399 }
3400}
3401
3402/*
3403 * Free a pcp page
3404 */
3405void free_unref_page(struct page *page, unsigned int order)
3406{
3407 unsigned long flags;
3408 unsigned long pfn = page_to_pfn(page);
3409 int migratetype;
3410
3411 if (!free_unref_page_prepare(page, pfn, order))
3412 return;
3413
3414 /*
3415 * We only track unmovable, reclaimable and movable on pcp lists.
3416 * Place ISOLATE pages on the isolated list because they are being
3417 * offlined but treat HIGHATOMIC as movable pages so we can get those
3418 * areas back if necessary. Otherwise, we may have to free
3419 * excessively into the page allocator
3420 */
3421 migratetype = get_pcppage_migratetype(page);
3422 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
3423 if (unlikely(is_migrate_isolate(migratetype))) {
3424 free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
3425 return;
3426 }
3427 migratetype = MIGRATE_MOVABLE;
3428 }
3429
3430 local_lock_irqsave(&pagesets.lock, flags);
3431 free_unref_page_commit(page, pfn, migratetype, order);
3432 local_unlock_irqrestore(&pagesets.lock, flags);
3433}
3434
3435/*
3436 * Free a list of 0-order pages
3437 */
3438void free_unref_page_list(struct list_head *list)
3439{
3440 struct page *page, *next;
3441 unsigned long flags, pfn;
3442 int batch_count = 0;
3443 int migratetype;
3444
3445 /* Prepare pages for freeing */
3446 list_for_each_entry_safe(page, next, list, lru) {
3447 pfn = page_to_pfn(page);
3448 if (!free_unref_page_prepare(page, pfn, 0)) {
3449 list_del(&page->lru);
3450 continue;
3451 }
3452
3453 /*
3454 * Free isolated pages directly to the allocator, see
3455 * comment in free_unref_page.
3456 */
3457 migratetype = get_pcppage_migratetype(page);
3458 if (unlikely(is_migrate_isolate(migratetype))) {
3459 list_del(&page->lru);
3460 free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
3461 continue;
3462 }
3463
3464 set_page_private(page, pfn);
3465 }
3466
3467 local_lock_irqsave(&pagesets.lock, flags);
3468 list_for_each_entry_safe(page, next, list, lru) {
3469 pfn = page_private(page);
3470 set_page_private(page, 0);
3471
3472 /*
3473 * Non-isolated types over MIGRATE_PCPTYPES get added
3474 * to the MIGRATE_MOVABLE pcp list.
3475 */
3476 migratetype = get_pcppage_migratetype(page);
3477 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
3478 migratetype = MIGRATE_MOVABLE;
3479
3480 trace_mm_page_free_batched(page);
3481 free_unref_page_commit(page, pfn, migratetype, 0);
3482
3483 /*
3484 * Guard against excessive IRQ disabled times when we get
3485 * a large list of pages to free.
3486 */
3487 if (++batch_count == SWAP_CLUSTER_MAX) {
3488 local_unlock_irqrestore(&pagesets.lock, flags);
3489 batch_count = 0;
3490 local_lock_irqsave(&pagesets.lock, flags);
3491 }
3492 }
3493 local_unlock_irqrestore(&pagesets.lock, flags);
3494}
3495
3496/*
3497 * split_page takes a non-compound higher-order page, and splits it into
3498 * n (1<<order) sub-pages: page[0..n]
3499 * Each sub-page must be freed individually.
3500 *
3501 * Note: this is probably too low level an operation for use in drivers.
3502 * Please consult with lkml before using this in your driver.
3503 */
3504void split_page(struct page *page, unsigned int order)
3505{
3506 int i;
3507
3508 VM_BUG_ON_PAGE(PageCompound(page), page);
3509 VM_BUG_ON_PAGE(!page_count(page), page);
3510
3511 for (i = 1; i < (1 << order); i++)
3512 set_page_refcounted(page + i);
3513 split_page_owner(page, 1 << order);
3514 split_page_memcg(page, 1 << order);
3515}
3516EXPORT_SYMBOL_GPL(split_page);
3517
3518int __isolate_free_page(struct page *page, unsigned int order)
3519{
3520 unsigned long watermark;
3521 struct zone *zone;
3522 int mt;
3523
3524 BUG_ON(!PageBuddy(page));
3525
3526 zone = page_zone(page);
3527 mt = get_pageblock_migratetype(page);
3528
3529 if (!is_migrate_isolate(mt)) {
3530 /*
3531 * Obey watermarks as if the page was being allocated. We can
3532 * emulate a high-order watermark check with a raised order-0
3533 * watermark, because we already know our high-order page
3534 * exists.
3535 */
3536 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3537 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3538 return 0;
3539
3540 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3541 }
3542
3543 /* Remove page from free list */
3544
3545 del_page_from_free_list(page, zone, order);
3546
3547 /*
3548 * Set the pageblock if the isolated page is at least half of a
3549 * pageblock
3550 */
3551 if (order >= pageblock_order - 1) {
3552 struct page *endpage = page + (1 << order) - 1;
3553 for (; page < endpage; page += pageblock_nr_pages) {
3554 int mt = get_pageblock_migratetype(page);
3555 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3556 && !is_migrate_highatomic(mt))
3557 set_pageblock_migratetype(page,
3558 MIGRATE_MOVABLE);
3559 }
3560 }
3561
3562
3563 return 1UL << order;
3564}
3565
3566/**
3567 * __putback_isolated_page - Return a now-isolated page back where we got it
3568 * @page: Page that was isolated
3569 * @order: Order of the isolated page
3570 * @mt: The page's pageblock's migratetype
3571 *
3572 * This function is meant to return a page pulled from the free lists via
3573 * __isolate_free_page back to the free lists they were pulled from.
3574 */
3575void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3576{
3577 struct zone *zone = page_zone(page);
3578
3579 /* zone lock should be held when this function is called */
3580 lockdep_assert_held(&zone->lock);
3581
3582 /* Return isolated page to tail of freelist. */
3583 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3584 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3585}
3586
3587/*
3588 * Update NUMA hit/miss statistics
3589 *
3590 * Must be called with interrupts disabled.
3591 */
3592static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
3593 long nr_account)
3594{
3595#ifdef CONFIG_NUMA
3596 enum numa_stat_item local_stat = NUMA_LOCAL;
3597
3598 /* skip numa counters update if numa stats is disabled */
3599 if (!static_branch_likely(&vm_numa_stat_key))
3600 return;
3601
3602 if (zone_to_nid(z) != numa_node_id())
3603 local_stat = NUMA_OTHER;
3604
3605 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3606 __count_numa_events(z, NUMA_HIT, nr_account);
3607 else {
3608 __count_numa_events(z, NUMA_MISS, nr_account);
3609 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
3610 }
3611 __count_numa_events(z, local_stat, nr_account);
3612#endif
3613}
3614
3615/* Remove page from the per-cpu list, caller must protect the list */
3616static inline
3617struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
3618 int migratetype,
3619 unsigned int alloc_flags,
3620 struct per_cpu_pages *pcp,
3621 struct list_head *list)
3622{
3623 struct page *page;
3624
3625 do {
3626 if (list_empty(list)) {
3627 int batch = READ_ONCE(pcp->batch);
3628 int alloced;
3629
3630 /*
3631 * Scale batch relative to order if batch implies
3632 * free pages can be stored on the PCP. Batch can
3633 * be 1 for small zones or for boot pagesets which
3634 * should never store free pages as the pages may
3635 * belong to arbitrary zones.
3636 */
3637 if (batch > 1)
3638 batch = max(batch >> order, 2);
3639 alloced = rmqueue_bulk(zone, order,
3640 batch, list,
3641 migratetype, alloc_flags);
3642
3643 pcp->count += alloced << order;
3644 if (unlikely(list_empty(list)))
3645 return NULL;
3646 }
3647
3648 page = list_first_entry(list, struct page, lru);
3649 list_del(&page->lru);
3650 pcp->count -= 1 << order;
3651 } while (check_new_pcp(page));
3652
3653 return page;
3654}
3655
3656/* Lock and remove page from the per-cpu list */
3657static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3658 struct zone *zone, unsigned int order,
3659 gfp_t gfp_flags, int migratetype,
3660 unsigned int alloc_flags)
3661{
3662 struct per_cpu_pages *pcp;
3663 struct list_head *list;
3664 struct page *page;
3665 unsigned long flags;
3666
3667 local_lock_irqsave(&pagesets.lock, flags);
3668
3669 /*
3670 * On allocation, reduce the number of pages that are batch freed.
3671 * See nr_pcp_free() where free_factor is increased for subsequent
3672 * frees.
3673 */
3674 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3675 pcp->free_factor >>= 1;
3676 list = &pcp->lists[order_to_pindex(migratetype, order)];
3677 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
3678 local_unlock_irqrestore(&pagesets.lock, flags);
3679 if (page) {
3680 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3681 zone_statistics(preferred_zone, zone, 1);
3682 }
3683 return page;
3684}
3685
3686/*
3687 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3688 */
3689static inline
3690struct page *rmqueue(struct zone *preferred_zone,
3691 struct zone *zone, unsigned int order,
3692 gfp_t gfp_flags, unsigned int alloc_flags,
3693 int migratetype)
3694{
3695 unsigned long flags;
3696 struct page *page;
3697
3698 if (likely(pcp_allowed_order(order))) {
3699 /*
3700 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3701 * we need to skip it when CMA area isn't allowed.
3702 */
3703 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3704 migratetype != MIGRATE_MOVABLE) {
3705 page = rmqueue_pcplist(preferred_zone, zone, order,
3706 gfp_flags, migratetype, alloc_flags);
3707 goto out;
3708 }
3709 }
3710
3711 /*
3712 * We most definitely don't want callers attempting to
3713 * allocate greater than order-1 page units with __GFP_NOFAIL.
3714 */
3715 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3716 spin_lock_irqsave(&zone->lock, flags);
3717
3718 do {
3719 page = NULL;
3720 /*
3721 * order-0 request can reach here when the pcplist is skipped
3722 * due to non-CMA allocation context. HIGHATOMIC area is
3723 * reserved for high-order atomic allocation, so order-0
3724 * request should skip it.
3725 */
3726 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3727 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3728 if (page)
3729 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3730 }
3731 if (!page)
3732 page = __rmqueue(zone, order, migratetype, alloc_flags);
3733 } while (page && check_new_pages(page, order));
3734 if (!page)
3735 goto failed;
3736
3737 __mod_zone_freepage_state(zone, -(1 << order),
3738 get_pcppage_migratetype(page));
3739 spin_unlock_irqrestore(&zone->lock, flags);
3740
3741 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3742 zone_statistics(preferred_zone, zone, 1);
3743
3744out:
3745 /* Separate test+clear to avoid unnecessary atomics */
3746 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3747 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3748 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3749 }
3750
3751 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3752 return page;
3753
3754failed:
3755 spin_unlock_irqrestore(&zone->lock, flags);
3756 return NULL;
3757}
3758
3759#ifdef CONFIG_FAIL_PAGE_ALLOC
3760
3761static struct {
3762 struct fault_attr attr;
3763
3764 bool ignore_gfp_highmem;
3765 bool ignore_gfp_reclaim;
3766 u32 min_order;
3767} fail_page_alloc = {
3768 .attr = FAULT_ATTR_INITIALIZER,
3769 .ignore_gfp_reclaim = true,
3770 .ignore_gfp_highmem = true,
3771 .min_order = 1,
3772};
3773
3774static int __init setup_fail_page_alloc(char *str)
3775{
3776 return setup_fault_attr(&fail_page_alloc.attr, str);
3777}
3778__setup("fail_page_alloc=", setup_fail_page_alloc);
3779
3780static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3781{
3782 if (order < fail_page_alloc.min_order)
3783 return false;
3784 if (gfp_mask & __GFP_NOFAIL)
3785 return false;
3786 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3787 return false;
3788 if (fail_page_alloc.ignore_gfp_reclaim &&
3789 (gfp_mask & __GFP_DIRECT_RECLAIM))
3790 return false;
3791
3792 return should_fail(&fail_page_alloc.attr, 1 << order);
3793}
3794
3795#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3796
3797static int __init fail_page_alloc_debugfs(void)
3798{
3799 umode_t mode = S_IFREG | 0600;
3800 struct dentry *dir;
3801
3802 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3803 &fail_page_alloc.attr);
3804
3805 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3806 &fail_page_alloc.ignore_gfp_reclaim);
3807 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3808 &fail_page_alloc.ignore_gfp_highmem);
3809 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3810
3811 return 0;
3812}
3813
3814late_initcall(fail_page_alloc_debugfs);
3815
3816#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3817
3818#else /* CONFIG_FAIL_PAGE_ALLOC */
3819
3820static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3821{
3822 return false;
3823}
3824
3825#endif /* CONFIG_FAIL_PAGE_ALLOC */
3826
3827noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3828{
3829 return __should_fail_alloc_page(gfp_mask, order);
3830}
3831ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3832
3833static inline long __zone_watermark_unusable_free(struct zone *z,
3834 unsigned int order, unsigned int alloc_flags)
3835{
3836 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3837 long unusable_free = (1 << order) - 1;
3838
3839 /*
3840 * If the caller does not have rights to ALLOC_HARDER then subtract
3841 * the high-atomic reserves. This will over-estimate the size of the
3842 * atomic reserve but it avoids a search.
3843 */
3844 if (likely(!alloc_harder))
3845 unusable_free += z->nr_reserved_highatomic;
3846
3847#ifdef CONFIG_CMA
3848 /* If allocation can't use CMA areas don't use free CMA pages */
3849 if (!(alloc_flags & ALLOC_CMA))
3850 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3851#endif
3852
3853 return unusable_free;
3854}
3855
3856/*
3857 * Return true if free base pages are above 'mark'. For high-order checks it
3858 * will return true of the order-0 watermark is reached and there is at least
3859 * one free page of a suitable size. Checking now avoids taking the zone lock
3860 * to check in the allocation paths if no pages are free.
3861 */
3862bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3863 int highest_zoneidx, unsigned int alloc_flags,
3864 long free_pages)
3865{
3866 long min = mark;
3867 int o;
3868 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3869
3870 /* free_pages may go negative - that's OK */
3871 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3872
3873 if (alloc_flags & ALLOC_HIGH)
3874 min -= min / 2;
3875
3876 if (unlikely(alloc_harder)) {
3877 /*
3878 * OOM victims can try even harder than normal ALLOC_HARDER
3879 * users on the grounds that it's definitely going to be in
3880 * the exit path shortly and free memory. Any allocation it
3881 * makes during the free path will be small and short-lived.
3882 */
3883 if (alloc_flags & ALLOC_OOM)
3884 min -= min / 2;
3885 else
3886 min -= min / 4;
3887 }
3888
3889 /*
3890 * Check watermarks for an order-0 allocation request. If these
3891 * are not met, then a high-order request also cannot go ahead
3892 * even if a suitable page happened to be free.
3893 */
3894 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3895 return false;
3896
3897 /* If this is an order-0 request then the watermark is fine */
3898 if (!order)
3899 return true;
3900
3901 /* For a high-order request, check at least one suitable page is free */
3902 for (o = order; o < MAX_ORDER; o++) {
3903 struct free_area *area = &z->free_area[o];
3904 int mt;
3905
3906 if (!area->nr_free)
3907 continue;
3908
3909 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3910 if (!free_area_empty(area, mt))
3911 return true;
3912 }
3913
3914#ifdef CONFIG_CMA
3915 if ((alloc_flags & ALLOC_CMA) &&
3916 !free_area_empty(area, MIGRATE_CMA)) {
3917 return true;
3918 }
3919#endif
3920 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3921 return true;
3922 }
3923 return false;
3924}
3925
3926bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3927 int highest_zoneidx, unsigned int alloc_flags)
3928{
3929 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3930 zone_page_state(z, NR_FREE_PAGES));
3931}
3932
3933static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3934 unsigned long mark, int highest_zoneidx,
3935 unsigned int alloc_flags, gfp_t gfp_mask)
3936{
3937 long free_pages;
3938
3939 free_pages = zone_page_state(z, NR_FREE_PAGES);
3940
3941 /*
3942 * Fast check for order-0 only. If this fails then the reserves
3943 * need to be calculated.
3944 */
3945 if (!order) {
3946 long fast_free;
3947
3948 fast_free = free_pages;
3949 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3950 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3951 return true;
3952 }
3953
3954 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3955 free_pages))
3956 return true;
3957 /*
3958 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3959 * when checking the min watermark. The min watermark is the
3960 * point where boosting is ignored so that kswapd is woken up
3961 * when below the low watermark.
3962 */
3963 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3964 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3965 mark = z->_watermark[WMARK_MIN];
3966 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3967 alloc_flags, free_pages);
3968 }
3969
3970 return false;
3971}
3972
3973bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3974 unsigned long mark, int highest_zoneidx)
3975{
3976 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3977
3978 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3979 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3980
3981 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3982 free_pages);
3983}
3984
3985#ifdef CONFIG_NUMA
3986static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3987{
3988 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3989 node_reclaim_distance;
3990}
3991#else /* CONFIG_NUMA */
3992static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3993{
3994 return true;
3995}
3996#endif /* CONFIG_NUMA */
3997
3998/*
3999 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
4000 * fragmentation is subtle. If the preferred zone was HIGHMEM then
4001 * premature use of a lower zone may cause lowmem pressure problems that
4002 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
4003 * probably too small. It only makes sense to spread allocations to avoid
4004 * fragmentation between the Normal and DMA32 zones.
4005 */
4006static inline unsigned int
4007alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
4008{
4009 unsigned int alloc_flags;
4010
4011 /*
4012 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4013 * to save a branch.
4014 */
4015 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
4016
4017#ifdef CONFIG_ZONE_DMA32
4018 if (!zone)
4019 return alloc_flags;
4020
4021 if (zone_idx(zone) != ZONE_NORMAL)
4022 return alloc_flags;
4023
4024 /*
4025 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
4026 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
4027 * on UMA that if Normal is populated then so is DMA32.
4028 */
4029 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
4030 if (nr_online_nodes > 1 && !populated_zone(--zone))
4031 return alloc_flags;
4032
4033 alloc_flags |= ALLOC_NOFRAGMENT;
4034#endif /* CONFIG_ZONE_DMA32 */
4035 return alloc_flags;
4036}
4037
4038/* Must be called after current_gfp_context() which can change gfp_mask */
4039static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
4040 unsigned int alloc_flags)
4041{
4042#ifdef CONFIG_CMA
4043 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4044 alloc_flags |= ALLOC_CMA;
4045#endif
4046 return alloc_flags;
4047}
4048
4049/*
4050 * get_page_from_freelist goes through the zonelist trying to allocate
4051 * a page.
4052 */
4053static struct page *
4054get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
4055 const struct alloc_context *ac)
4056{
4057 struct zoneref *z;
4058 struct zone *zone;
4059 struct pglist_data *last_pgdat_dirty_limit = NULL;
4060 bool no_fallback;
4061
4062retry:
4063 /*
4064 * Scan zonelist, looking for a zone with enough free.
4065 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
4066 */
4067 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
4068 z = ac->preferred_zoneref;
4069 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
4070 ac->nodemask) {
4071 struct page *page;
4072 unsigned long mark;
4073
4074 if (cpusets_enabled() &&
4075 (alloc_flags & ALLOC_CPUSET) &&
4076 !__cpuset_zone_allowed(zone, gfp_mask))
4077 continue;
4078 /*
4079 * When allocating a page cache page for writing, we
4080 * want to get it from a node that is within its dirty
4081 * limit, such that no single node holds more than its
4082 * proportional share of globally allowed dirty pages.
4083 * The dirty limits take into account the node's
4084 * lowmem reserves and high watermark so that kswapd
4085 * should be able to balance it without having to
4086 * write pages from its LRU list.
4087 *
4088 * XXX: For now, allow allocations to potentially
4089 * exceed the per-node dirty limit in the slowpath
4090 * (spread_dirty_pages unset) before going into reclaim,
4091 * which is important when on a NUMA setup the allowed
4092 * nodes are together not big enough to reach the
4093 * global limit. The proper fix for these situations
4094 * will require awareness of nodes in the
4095 * dirty-throttling and the flusher threads.
4096 */
4097 if (ac->spread_dirty_pages) {
4098 if (last_pgdat_dirty_limit == zone->zone_pgdat)
4099 continue;
4100
4101 if (!node_dirty_ok(zone->zone_pgdat)) {
4102 last_pgdat_dirty_limit = zone->zone_pgdat;
4103 continue;
4104 }
4105 }
4106
4107 if (no_fallback && nr_online_nodes > 1 &&
4108 zone != ac->preferred_zoneref->zone) {
4109 int local_nid;
4110
4111 /*
4112 * If moving to a remote node, retry but allow
4113 * fragmenting fallbacks. Locality is more important
4114 * than fragmentation avoidance.
4115 */
4116 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
4117 if (zone_to_nid(zone) != local_nid) {
4118 alloc_flags &= ~ALLOC_NOFRAGMENT;
4119 goto retry;
4120 }
4121 }
4122
4123 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
4124 if (!zone_watermark_fast(zone, order, mark,
4125 ac->highest_zoneidx, alloc_flags,
4126 gfp_mask)) {
4127 int ret;
4128
4129#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4130 /*
4131 * Watermark failed for this zone, but see if we can
4132 * grow this zone if it contains deferred pages.
4133 */
4134 if (static_branch_unlikely(&deferred_pages)) {
4135 if (_deferred_grow_zone(zone, order))
4136 goto try_this_zone;
4137 }
4138#endif
4139 /* Checked here to keep the fast path fast */
4140 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
4141 if (alloc_flags & ALLOC_NO_WATERMARKS)
4142 goto try_this_zone;
4143
4144 if (!node_reclaim_enabled() ||
4145 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
4146 continue;
4147
4148 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
4149 switch (ret) {
4150 case NODE_RECLAIM_NOSCAN:
4151 /* did not scan */
4152 continue;
4153 case NODE_RECLAIM_FULL:
4154 /* scanned but unreclaimable */
4155 continue;
4156 default:
4157 /* did we reclaim enough */
4158 if (zone_watermark_ok(zone, order, mark,
4159 ac->highest_zoneidx, alloc_flags))
4160 goto try_this_zone;
4161
4162 continue;
4163 }
4164 }
4165
4166try_this_zone:
4167 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
4168 gfp_mask, alloc_flags, ac->migratetype);
4169 if (page) {
4170 prep_new_page(page, order, gfp_mask, alloc_flags);
4171
4172 /*
4173 * If this is a high-order atomic allocation then check
4174 * if the pageblock should be reserved for the future
4175 */
4176 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
4177 reserve_highatomic_pageblock(page, zone, order);
4178
4179 return page;
4180 } else {
4181#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4182 /* Try again if zone has deferred pages */
4183 if (static_branch_unlikely(&deferred_pages)) {
4184 if (_deferred_grow_zone(zone, order))
4185 goto try_this_zone;
4186 }
4187#endif
4188 }
4189 }
4190
4191 /*
4192 * It's possible on a UMA machine to get through all zones that are
4193 * fragmented. If avoiding fragmentation, reset and try again.
4194 */
4195 if (no_fallback) {
4196 alloc_flags &= ~ALLOC_NOFRAGMENT;
4197 goto retry;
4198 }
4199
4200 return NULL;
4201}
4202
4203static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4204{
4205 unsigned int filter = SHOW_MEM_FILTER_NODES;
4206
4207 /*
4208 * This documents exceptions given to allocations in certain
4209 * contexts that are allowed to allocate outside current's set
4210 * of allowed nodes.
4211 */
4212 if (!(gfp_mask & __GFP_NOMEMALLOC))
4213 if (tsk_is_oom_victim(current) ||
4214 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4215 filter &= ~SHOW_MEM_FILTER_NODES;
4216 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4217 filter &= ~SHOW_MEM_FILTER_NODES;
4218
4219 show_mem(filter, nodemask);
4220}
4221
4222void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4223{
4224 struct va_format vaf;
4225 va_list args;
4226 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4227
4228 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
4229 return;
4230
4231 va_start(args, fmt);
4232 vaf.fmt = fmt;
4233 vaf.va = &args;
4234 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4235 current->comm, &vaf, gfp_mask, &gfp_mask,
4236 nodemask_pr_args(nodemask));
4237 va_end(args);
4238
4239 cpuset_print_current_mems_allowed();
4240 pr_cont("\n");
4241 dump_stack();
4242 warn_alloc_show_mem(gfp_mask, nodemask);
4243}
4244
4245static inline struct page *
4246__alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4247 unsigned int alloc_flags,
4248 const struct alloc_context *ac)
4249{
4250 struct page *page;
4251
4252 page = get_page_from_freelist(gfp_mask, order,
4253 alloc_flags|ALLOC_CPUSET, ac);
4254 /*
4255 * fallback to ignore cpuset restriction if our nodes
4256 * are depleted
4257 */
4258 if (!page)
4259 page = get_page_from_freelist(gfp_mask, order,
4260 alloc_flags, ac);
4261
4262 return page;
4263}
4264
4265static inline struct page *
4266__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4267 const struct alloc_context *ac, unsigned long *did_some_progress)
4268{
4269 struct oom_control oc = {
4270 .zonelist = ac->zonelist,
4271 .nodemask = ac->nodemask,
4272 .memcg = NULL,
4273 .gfp_mask = gfp_mask,
4274 .order = order,
4275 };
4276 struct page *page;
4277
4278 *did_some_progress = 0;
4279
4280 /*
4281 * Acquire the oom lock. If that fails, somebody else is
4282 * making progress for us.
4283 */
4284 if (!mutex_trylock(&oom_lock)) {
4285 *did_some_progress = 1;
4286 schedule_timeout_uninterruptible(1);
4287 return NULL;
4288 }
4289
4290 /*
4291 * Go through the zonelist yet one more time, keep very high watermark
4292 * here, this is only to catch a parallel oom killing, we must fail if
4293 * we're still under heavy pressure. But make sure that this reclaim
4294 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4295 * allocation which will never fail due to oom_lock already held.
4296 */
4297 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4298 ~__GFP_DIRECT_RECLAIM, order,
4299 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4300 if (page)
4301 goto out;
4302
4303 /* Coredumps can quickly deplete all memory reserves */
4304 if (current->flags & PF_DUMPCORE)
4305 goto out;
4306 /* The OOM killer will not help higher order allocs */
4307 if (order > PAGE_ALLOC_COSTLY_ORDER)
4308 goto out;
4309 /*
4310 * We have already exhausted all our reclaim opportunities without any
4311 * success so it is time to admit defeat. We will skip the OOM killer
4312 * because it is very likely that the caller has a more reasonable
4313 * fallback than shooting a random task.
4314 *
4315 * The OOM killer may not free memory on a specific node.
4316 */
4317 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4318 goto out;
4319 /* The OOM killer does not needlessly kill tasks for lowmem */
4320 if (ac->highest_zoneidx < ZONE_NORMAL)
4321 goto out;
4322 if (pm_suspended_storage())
4323 goto out;
4324 /*
4325 * XXX: GFP_NOFS allocations should rather fail than rely on
4326 * other request to make a forward progress.
4327 * We are in an unfortunate situation where out_of_memory cannot
4328 * do much for this context but let's try it to at least get
4329 * access to memory reserved if the current task is killed (see
4330 * out_of_memory). Once filesystems are ready to handle allocation
4331 * failures more gracefully we should just bail out here.
4332 */
4333
4334 /* Exhausted what can be done so it's blame time */
4335 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4336 *did_some_progress = 1;
4337
4338 /*
4339 * Help non-failing allocations by giving them access to memory
4340 * reserves
4341 */
4342 if (gfp_mask & __GFP_NOFAIL)
4343 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4344 ALLOC_NO_WATERMARKS, ac);
4345 }
4346out:
4347 mutex_unlock(&oom_lock);
4348 return page;
4349}
4350
4351/*
4352 * Maximum number of compaction retries with a progress before OOM
4353 * killer is consider as the only way to move forward.
4354 */
4355#define MAX_COMPACT_RETRIES 16
4356
4357#ifdef CONFIG_COMPACTION
4358/* Try memory compaction for high-order allocations before reclaim */
4359static struct page *
4360__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4361 unsigned int alloc_flags, const struct alloc_context *ac,
4362 enum compact_priority prio, enum compact_result *compact_result)
4363{
4364 struct page *page = NULL;
4365 unsigned long pflags;
4366 unsigned int noreclaim_flag;
4367
4368 if (!order)
4369 return NULL;
4370
4371 psi_memstall_enter(&pflags);
4372 noreclaim_flag = memalloc_noreclaim_save();
4373
4374 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4375 prio, &page);
4376
4377 memalloc_noreclaim_restore(noreclaim_flag);
4378 psi_memstall_leave(&pflags);
4379
4380 if (*compact_result == COMPACT_SKIPPED)
4381 return NULL;
4382 /*
4383 * At least in one zone compaction wasn't deferred or skipped, so let's
4384 * count a compaction stall
4385 */
4386 count_vm_event(COMPACTSTALL);
4387
4388 /* Prep a captured page if available */
4389 if (page)
4390 prep_new_page(page, order, gfp_mask, alloc_flags);
4391
4392 /* Try get a page from the freelist if available */
4393 if (!page)
4394 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4395
4396 if (page) {
4397 struct zone *zone = page_zone(page);
4398
4399 zone->compact_blockskip_flush = false;
4400 compaction_defer_reset(zone, order, true);
4401 count_vm_event(COMPACTSUCCESS);
4402 return page;
4403 }
4404
4405 /*
4406 * It's bad if compaction run occurs and fails. The most likely reason
4407 * is that pages exist, but not enough to satisfy watermarks.
4408 */
4409 count_vm_event(COMPACTFAIL);
4410
4411 cond_resched();
4412
4413 return NULL;
4414}
4415
4416static inline bool
4417should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4418 enum compact_result compact_result,
4419 enum compact_priority *compact_priority,
4420 int *compaction_retries)
4421{
4422 int max_retries = MAX_COMPACT_RETRIES;
4423 int min_priority;
4424 bool ret = false;
4425 int retries = *compaction_retries;
4426 enum compact_priority priority = *compact_priority;
4427
4428 if (!order)
4429 return false;
4430
4431 if (fatal_signal_pending(current))
4432 return false;
4433
4434 if (compaction_made_progress(compact_result))
4435 (*compaction_retries)++;
4436
4437 /*
4438 * compaction considers all the zone as desperately out of memory
4439 * so it doesn't really make much sense to retry except when the
4440 * failure could be caused by insufficient priority
4441 */
4442 if (compaction_failed(compact_result))
4443 goto check_priority;
4444
4445 /*
4446 * compaction was skipped because there are not enough order-0 pages
4447 * to work with, so we retry only if it looks like reclaim can help.
4448 */
4449 if (compaction_needs_reclaim(compact_result)) {
4450 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4451 goto out;
4452 }
4453
4454 /*
4455 * make sure the compaction wasn't deferred or didn't bail out early
4456 * due to locks contention before we declare that we should give up.
4457 * But the next retry should use a higher priority if allowed, so
4458 * we don't just keep bailing out endlessly.
4459 */
4460 if (compaction_withdrawn(compact_result)) {
4461 goto check_priority;
4462 }
4463
4464 /*
4465 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4466 * costly ones because they are de facto nofail and invoke OOM
4467 * killer to move on while costly can fail and users are ready
4468 * to cope with that. 1/4 retries is rather arbitrary but we
4469 * would need much more detailed feedback from compaction to
4470 * make a better decision.
4471 */
4472 if (order > PAGE_ALLOC_COSTLY_ORDER)
4473 max_retries /= 4;
4474 if (*compaction_retries <= max_retries) {
4475 ret = true;
4476 goto out;
4477 }
4478
4479 /*
4480 * Make sure there are attempts at the highest priority if we exhausted
4481 * all retries or failed at the lower priorities.
4482 */
4483check_priority:
4484 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4485 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4486
4487 if (*compact_priority > min_priority) {
4488 (*compact_priority)--;
4489 *compaction_retries = 0;
4490 ret = true;
4491 }
4492out:
4493 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4494 return ret;
4495}
4496#else
4497static inline struct page *
4498__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4499 unsigned int alloc_flags, const struct alloc_context *ac,
4500 enum compact_priority prio, enum compact_result *compact_result)
4501{
4502 *compact_result = COMPACT_SKIPPED;
4503 return NULL;
4504}
4505
4506static inline bool
4507should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4508 enum compact_result compact_result,
4509 enum compact_priority *compact_priority,
4510 int *compaction_retries)
4511{
4512 struct zone *zone;
4513 struct zoneref *z;
4514
4515 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4516 return false;
4517
4518 /*
4519 * There are setups with compaction disabled which would prefer to loop
4520 * inside the allocator rather than hit the oom killer prematurely.
4521 * Let's give them a good hope and keep retrying while the order-0
4522 * watermarks are OK.
4523 */
4524 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4525 ac->highest_zoneidx, ac->nodemask) {
4526 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4527 ac->highest_zoneidx, alloc_flags))
4528 return true;
4529 }
4530 return false;
4531}
4532#endif /* CONFIG_COMPACTION */
4533
4534#ifdef CONFIG_LOCKDEP
4535static struct lockdep_map __fs_reclaim_map =
4536 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4537
4538static bool __need_reclaim(gfp_t gfp_mask)
4539{
4540 /* no reclaim without waiting on it */
4541 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4542 return false;
4543
4544 /* this guy won't enter reclaim */
4545 if (current->flags & PF_MEMALLOC)
4546 return false;
4547
4548 if (gfp_mask & __GFP_NOLOCKDEP)
4549 return false;
4550
4551 return true;
4552}
4553
4554void __fs_reclaim_acquire(void)
4555{
4556 lock_map_acquire(&__fs_reclaim_map);
4557}
4558
4559void __fs_reclaim_release(void)
4560{
4561 lock_map_release(&__fs_reclaim_map);
4562}
4563
4564void fs_reclaim_acquire(gfp_t gfp_mask)
4565{
4566 gfp_mask = current_gfp_context(gfp_mask);
4567
4568 if (__need_reclaim(gfp_mask)) {
4569 if (gfp_mask & __GFP_FS)
4570 __fs_reclaim_acquire();
4571
4572#ifdef CONFIG_MMU_NOTIFIER
4573 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4574 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4575#endif
4576
4577 }
4578}
4579EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4580
4581void fs_reclaim_release(gfp_t gfp_mask)
4582{
4583 gfp_mask = current_gfp_context(gfp_mask);
4584
4585 if (__need_reclaim(gfp_mask)) {
4586 if (gfp_mask & __GFP_FS)
4587 __fs_reclaim_release();
4588 }
4589}
4590EXPORT_SYMBOL_GPL(fs_reclaim_release);
4591#endif
4592
4593/* Perform direct synchronous page reclaim */
4594static unsigned long
4595__perform_reclaim(gfp_t gfp_mask, unsigned int order,
4596 const struct alloc_context *ac)
4597{
4598 unsigned int noreclaim_flag;
4599 unsigned long pflags, progress;
4600
4601 cond_resched();
4602
4603 /* We now go into synchronous reclaim */
4604 cpuset_memory_pressure_bump();
4605 psi_memstall_enter(&pflags);
4606 fs_reclaim_acquire(gfp_mask);
4607 noreclaim_flag = memalloc_noreclaim_save();
4608
4609 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4610 ac->nodemask);
4611
4612 memalloc_noreclaim_restore(noreclaim_flag);
4613 fs_reclaim_release(gfp_mask);
4614 psi_memstall_leave(&pflags);
4615
4616 cond_resched();
4617
4618 return progress;
4619}
4620
4621/* The really slow allocator path where we enter direct reclaim */
4622static inline struct page *
4623__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4624 unsigned int alloc_flags, const struct alloc_context *ac,
4625 unsigned long *did_some_progress)
4626{
4627 struct page *page = NULL;
4628 bool drained = false;
4629
4630 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4631 if (unlikely(!(*did_some_progress)))
4632 return NULL;
4633
4634retry:
4635 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4636
4637 /*
4638 * If an allocation failed after direct reclaim, it could be because
4639 * pages are pinned on the per-cpu lists or in high alloc reserves.
4640 * Shrink them and try again
4641 */
4642 if (!page && !drained) {
4643 unreserve_highatomic_pageblock(ac, false);
4644 drain_all_pages(NULL);
4645 drained = true;
4646 goto retry;
4647 }
4648
4649 return page;
4650}
4651
4652static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4653 const struct alloc_context *ac)
4654{
4655 struct zoneref *z;
4656 struct zone *zone;
4657 pg_data_t *last_pgdat = NULL;
4658 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4659
4660 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4661 ac->nodemask) {
4662 if (last_pgdat != zone->zone_pgdat)
4663 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4664 last_pgdat = zone->zone_pgdat;
4665 }
4666}
4667
4668static inline unsigned int
4669gfp_to_alloc_flags(gfp_t gfp_mask)
4670{
4671 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4672
4673 /*
4674 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4675 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4676 * to save two branches.
4677 */
4678 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4679 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4680
4681 /*
4682 * The caller may dip into page reserves a bit more if the caller
4683 * cannot run direct reclaim, or if the caller has realtime scheduling
4684 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4685 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4686 */
4687 alloc_flags |= (__force int)
4688 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4689
4690 if (gfp_mask & __GFP_ATOMIC) {
4691 /*
4692 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4693 * if it can't schedule.
4694 */
4695 if (!(gfp_mask & __GFP_NOMEMALLOC))
4696 alloc_flags |= ALLOC_HARDER;
4697 /*
4698 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4699 * comment for __cpuset_node_allowed().
4700 */
4701 alloc_flags &= ~ALLOC_CPUSET;
4702 } else if (unlikely(rt_task(current)) && !in_interrupt())
4703 alloc_flags |= ALLOC_HARDER;
4704
4705 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4706
4707 return alloc_flags;
4708}
4709
4710static bool oom_reserves_allowed(struct task_struct *tsk)
4711{
4712 if (!tsk_is_oom_victim(tsk))
4713 return false;
4714
4715 /*
4716 * !MMU doesn't have oom reaper so give access to memory reserves
4717 * only to the thread with TIF_MEMDIE set
4718 */
4719 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4720 return false;
4721
4722 return true;
4723}
4724
4725/*
4726 * Distinguish requests which really need access to full memory
4727 * reserves from oom victims which can live with a portion of it
4728 */
4729static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4730{
4731 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4732 return 0;
4733 if (gfp_mask & __GFP_MEMALLOC)
4734 return ALLOC_NO_WATERMARKS;
4735 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4736 return ALLOC_NO_WATERMARKS;
4737 if (!in_interrupt()) {
4738 if (current->flags & PF_MEMALLOC)
4739 return ALLOC_NO_WATERMARKS;
4740 else if (oom_reserves_allowed(current))
4741 return ALLOC_OOM;
4742 }
4743
4744 return 0;
4745}
4746
4747bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4748{
4749 return !!__gfp_pfmemalloc_flags(gfp_mask);
4750}
4751
4752/*
4753 * Checks whether it makes sense to retry the reclaim to make a forward progress
4754 * for the given allocation request.
4755 *
4756 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4757 * without success, or when we couldn't even meet the watermark if we
4758 * reclaimed all remaining pages on the LRU lists.
4759 *
4760 * Returns true if a retry is viable or false to enter the oom path.
4761 */
4762static inline bool
4763should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4764 struct alloc_context *ac, int alloc_flags,
4765 bool did_some_progress, int *no_progress_loops)
4766{
4767 struct zone *zone;
4768 struct zoneref *z;
4769 bool ret = false;
4770
4771 /*
4772 * Costly allocations might have made a progress but this doesn't mean
4773 * their order will become available due to high fragmentation so
4774 * always increment the no progress counter for them
4775 */
4776 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4777 *no_progress_loops = 0;
4778 else
4779 (*no_progress_loops)++;
4780
4781 /*
4782 * Make sure we converge to OOM if we cannot make any progress
4783 * several times in the row.
4784 */
4785 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4786 /* Before OOM, exhaust highatomic_reserve */
4787 return unreserve_highatomic_pageblock(ac, true);
4788 }
4789
4790 /*
4791 * Keep reclaiming pages while there is a chance this will lead
4792 * somewhere. If none of the target zones can satisfy our allocation
4793 * request even if all reclaimable pages are considered then we are
4794 * screwed and have to go OOM.
4795 */
4796 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4797 ac->highest_zoneidx, ac->nodemask) {
4798 unsigned long available;
4799 unsigned long reclaimable;
4800 unsigned long min_wmark = min_wmark_pages(zone);
4801 bool wmark;
4802
4803 available = reclaimable = zone_reclaimable_pages(zone);
4804 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4805
4806 /*
4807 * Would the allocation succeed if we reclaimed all
4808 * reclaimable pages?
4809 */
4810 wmark = __zone_watermark_ok(zone, order, min_wmark,
4811 ac->highest_zoneidx, alloc_flags, available);
4812 trace_reclaim_retry_zone(z, order, reclaimable,
4813 available, min_wmark, *no_progress_loops, wmark);
4814 if (wmark) {
4815 /*
4816 * If we didn't make any progress and have a lot of
4817 * dirty + writeback pages then we should wait for
4818 * an IO to complete to slow down the reclaim and
4819 * prevent from pre mature OOM
4820 */
4821 if (!did_some_progress) {
4822 unsigned long write_pending;
4823
4824 write_pending = zone_page_state_snapshot(zone,
4825 NR_ZONE_WRITE_PENDING);
4826
4827 if (2 * write_pending > reclaimable) {
4828 congestion_wait(BLK_RW_ASYNC, HZ/10);
4829 return true;
4830 }
4831 }
4832
4833 ret = true;
4834 goto out;
4835 }
4836 }
4837
4838out:
4839 /*
4840 * Memory allocation/reclaim might be called from a WQ context and the
4841 * current implementation of the WQ concurrency control doesn't
4842 * recognize that a particular WQ is congested if the worker thread is
4843 * looping without ever sleeping. Therefore we have to do a short sleep
4844 * here rather than calling cond_resched().
4845 */
4846 if (current->flags & PF_WQ_WORKER)
4847 schedule_timeout_uninterruptible(1);
4848 else
4849 cond_resched();
4850 return ret;
4851}
4852
4853static inline bool
4854check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4855{
4856 /*
4857 * It's possible that cpuset's mems_allowed and the nodemask from
4858 * mempolicy don't intersect. This should be normally dealt with by
4859 * policy_nodemask(), but it's possible to race with cpuset update in
4860 * such a way the check therein was true, and then it became false
4861 * before we got our cpuset_mems_cookie here.
4862 * This assumes that for all allocations, ac->nodemask can come only
4863 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4864 * when it does not intersect with the cpuset restrictions) or the
4865 * caller can deal with a violated nodemask.
4866 */
4867 if (cpusets_enabled() && ac->nodemask &&
4868 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4869 ac->nodemask = NULL;
4870 return true;
4871 }
4872
4873 /*
4874 * When updating a task's mems_allowed or mempolicy nodemask, it is
4875 * possible to race with parallel threads in such a way that our
4876 * allocation can fail while the mask is being updated. If we are about
4877 * to fail, check if the cpuset changed during allocation and if so,
4878 * retry.
4879 */
4880 if (read_mems_allowed_retry(cpuset_mems_cookie))
4881 return true;
4882
4883 return false;
4884}
4885
4886static inline struct page *
4887__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4888 struct alloc_context *ac)
4889{
4890 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4891 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4892 struct page *page = NULL;
4893 unsigned int alloc_flags;
4894 unsigned long did_some_progress;
4895 enum compact_priority compact_priority;
4896 enum compact_result compact_result;
4897 int compaction_retries;
4898 int no_progress_loops;
4899 unsigned int cpuset_mems_cookie;
4900 int reserve_flags;
4901
4902 /*
4903 * We also sanity check to catch abuse of atomic reserves being used by
4904 * callers that are not in atomic context.
4905 */
4906 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4907 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4908 gfp_mask &= ~__GFP_ATOMIC;
4909
4910retry_cpuset:
4911 compaction_retries = 0;
4912 no_progress_loops = 0;
4913 compact_priority = DEF_COMPACT_PRIORITY;
4914 cpuset_mems_cookie = read_mems_allowed_begin();
4915
4916 /*
4917 * The fast path uses conservative alloc_flags to succeed only until
4918 * kswapd needs to be woken up, and to avoid the cost of setting up
4919 * alloc_flags precisely. So we do that now.
4920 */
4921 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4922
4923 /*
4924 * We need to recalculate the starting point for the zonelist iterator
4925 * because we might have used different nodemask in the fast path, or
4926 * there was a cpuset modification and we are retrying - otherwise we
4927 * could end up iterating over non-eligible zones endlessly.
4928 */
4929 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4930 ac->highest_zoneidx, ac->nodemask);
4931 if (!ac->preferred_zoneref->zone)
4932 goto nopage;
4933
4934 if (alloc_flags & ALLOC_KSWAPD)
4935 wake_all_kswapds(order, gfp_mask, ac);
4936
4937 /*
4938 * The adjusted alloc_flags might result in immediate success, so try
4939 * that first
4940 */
4941 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4942 if (page)
4943 goto got_pg;
4944
4945 /*
4946 * For costly allocations, try direct compaction first, as it's likely
4947 * that we have enough base pages and don't need to reclaim. For non-
4948 * movable high-order allocations, do that as well, as compaction will
4949 * try prevent permanent fragmentation by migrating from blocks of the
4950 * same migratetype.
4951 * Don't try this for allocations that are allowed to ignore
4952 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4953 */
4954 if (can_direct_reclaim &&
4955 (costly_order ||
4956 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4957 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4958 page = __alloc_pages_direct_compact(gfp_mask, order,
4959 alloc_flags, ac,
4960 INIT_COMPACT_PRIORITY,
4961 &compact_result);
4962 if (page)
4963 goto got_pg;
4964
4965 /*
4966 * Checks for costly allocations with __GFP_NORETRY, which
4967 * includes some THP page fault allocations
4968 */
4969 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4970 /*
4971 * If allocating entire pageblock(s) and compaction
4972 * failed because all zones are below low watermarks
4973 * or is prohibited because it recently failed at this
4974 * order, fail immediately unless the allocator has
4975 * requested compaction and reclaim retry.
4976 *
4977 * Reclaim is
4978 * - potentially very expensive because zones are far
4979 * below their low watermarks or this is part of very
4980 * bursty high order allocations,
4981 * - not guaranteed to help because isolate_freepages()
4982 * may not iterate over freed pages as part of its
4983 * linear scan, and
4984 * - unlikely to make entire pageblocks free on its
4985 * own.
4986 */
4987 if (compact_result == COMPACT_SKIPPED ||
4988 compact_result == COMPACT_DEFERRED)
4989 goto nopage;
4990
4991 /*
4992 * Looks like reclaim/compaction is worth trying, but
4993 * sync compaction could be very expensive, so keep
4994 * using async compaction.
4995 */
4996 compact_priority = INIT_COMPACT_PRIORITY;
4997 }
4998 }
4999
5000retry:
5001 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
5002 if (alloc_flags & ALLOC_KSWAPD)
5003 wake_all_kswapds(order, gfp_mask, ac);
5004
5005 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
5006 if (reserve_flags)
5007 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags);
5008
5009 /*
5010 * Reset the nodemask and zonelist iterators if memory policies can be
5011 * ignored. These allocations are high priority and system rather than
5012 * user oriented.
5013 */
5014 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
5015 ac->nodemask = NULL;
5016 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5017 ac->highest_zoneidx, ac->nodemask);
5018 }
5019
5020 /* Attempt with potentially adjusted zonelist and alloc_flags */
5021 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
5022 if (page)
5023 goto got_pg;
5024
5025 /* Caller is not willing to reclaim, we can't balance anything */
5026 if (!can_direct_reclaim)
5027 goto nopage;
5028
5029 /* Avoid recursion of direct reclaim */
5030 if (current->flags & PF_MEMALLOC)
5031 goto nopage;
5032
5033 /* Try direct reclaim and then allocating */
5034 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
5035 &did_some_progress);
5036 if (page)
5037 goto got_pg;
5038
5039 /* Try direct compaction and then allocating */
5040 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
5041 compact_priority, &compact_result);
5042 if (page)
5043 goto got_pg;
5044
5045 /* Do not loop if specifically requested */
5046 if (gfp_mask & __GFP_NORETRY)
5047 goto nopage;
5048
5049 /*
5050 * Do not retry costly high order allocations unless they are
5051 * __GFP_RETRY_MAYFAIL
5052 */
5053 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
5054 goto nopage;
5055
5056 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
5057 did_some_progress > 0, &no_progress_loops))
5058 goto retry;
5059
5060 /*
5061 * It doesn't make any sense to retry for the compaction if the order-0
5062 * reclaim is not able to make any progress because the current
5063 * implementation of the compaction depends on the sufficient amount
5064 * of free memory (see __compaction_suitable)
5065 */
5066 if (did_some_progress > 0 &&
5067 should_compact_retry(ac, order, alloc_flags,
5068 compact_result, &compact_priority,
5069 &compaction_retries))
5070 goto retry;
5071
5072
5073 /* Deal with possible cpuset update races before we start OOM killing */
5074 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5075 goto retry_cpuset;
5076
5077 /* Reclaim has failed us, start killing things */
5078 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
5079 if (page)
5080 goto got_pg;
5081
5082 /* Avoid allocations with no watermarks from looping endlessly */
5083 if (tsk_is_oom_victim(current) &&
5084 (alloc_flags & ALLOC_OOM ||
5085 (gfp_mask & __GFP_NOMEMALLOC)))
5086 goto nopage;
5087
5088 /* Retry as long as the OOM killer is making progress */
5089 if (did_some_progress) {
5090 no_progress_loops = 0;
5091 goto retry;
5092 }
5093
5094nopage:
5095 /* Deal with possible cpuset update races before we fail */
5096 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5097 goto retry_cpuset;
5098
5099 /*
5100 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
5101 * we always retry
5102 */
5103 if (gfp_mask & __GFP_NOFAIL) {
5104 /*
5105 * All existing users of the __GFP_NOFAIL are blockable, so warn
5106 * of any new users that actually require GFP_NOWAIT
5107 */
5108 if (WARN_ON_ONCE(!can_direct_reclaim))
5109 goto fail;
5110
5111 /*
5112 * PF_MEMALLOC request from this context is rather bizarre
5113 * because we cannot reclaim anything and only can loop waiting
5114 * for somebody to do a work for us
5115 */
5116 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
5117
5118 /*
5119 * non failing costly orders are a hard requirement which we
5120 * are not prepared for much so let's warn about these users
5121 * so that we can identify them and convert them to something
5122 * else.
5123 */
5124 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
5125
5126 /*
5127 * Help non-failing allocations by giving them access to memory
5128 * reserves but do not use ALLOC_NO_WATERMARKS because this
5129 * could deplete whole memory reserves which would just make
5130 * the situation worse
5131 */
5132 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
5133 if (page)
5134 goto got_pg;
5135
5136 cond_resched();
5137 goto retry;
5138 }
5139fail:
5140 warn_alloc(gfp_mask, ac->nodemask,
5141 "page allocation failure: order:%u", order);
5142got_pg:
5143 return page;
5144}
5145
5146static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
5147 int preferred_nid, nodemask_t *nodemask,
5148 struct alloc_context *ac, gfp_t *alloc_gfp,
5149 unsigned int *alloc_flags)
5150{
5151 ac->highest_zoneidx = gfp_zone(gfp_mask);
5152 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
5153 ac->nodemask = nodemask;
5154 ac->migratetype = gfp_migratetype(gfp_mask);
5155
5156 if (cpusets_enabled()) {
5157 *alloc_gfp |= __GFP_HARDWALL;
5158 /*
5159 * When we are in the interrupt context, it is irrelevant
5160 * to the current task context. It means that any node ok.
5161 */
5162 if (!in_interrupt() && !ac->nodemask)
5163 ac->nodemask = &cpuset_current_mems_allowed;
5164 else
5165 *alloc_flags |= ALLOC_CPUSET;
5166 }
5167
5168 fs_reclaim_acquire(gfp_mask);
5169 fs_reclaim_release(gfp_mask);
5170
5171 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
5172
5173 if (should_fail_alloc_page(gfp_mask, order))
5174 return false;
5175
5176 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
5177
5178 /* Dirty zone balancing only done in the fast path */
5179 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5180
5181 /*
5182 * The preferred zone is used for statistics but crucially it is
5183 * also used as the starting point for the zonelist iterator. It
5184 * may get reset for allocations that ignore memory policies.
5185 */
5186 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5187 ac->highest_zoneidx, ac->nodemask);
5188
5189 return true;
5190}
5191
5192/*
5193 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
5194 * @gfp: GFP flags for the allocation
5195 * @preferred_nid: The preferred NUMA node ID to allocate from
5196 * @nodemask: Set of nodes to allocate from, may be NULL
5197 * @nr_pages: The number of pages desired on the list or array
5198 * @page_list: Optional list to store the allocated pages
5199 * @page_array: Optional array to store the pages
5200 *
5201 * This is a batched version of the page allocator that attempts to
5202 * allocate nr_pages quickly. Pages are added to page_list if page_list
5203 * is not NULL, otherwise it is assumed that the page_array is valid.
5204 *
5205 * For lists, nr_pages is the number of pages that should be allocated.
5206 *
5207 * For arrays, only NULL elements are populated with pages and nr_pages
5208 * is the maximum number of pages that will be stored in the array.
5209 *
5210 * Returns the number of pages on the list or array.
5211 */
5212unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
5213 nodemask_t *nodemask, int nr_pages,
5214 struct list_head *page_list,
5215 struct page **page_array)
5216{
5217 struct page *page;
5218 unsigned long flags;
5219 struct zone *zone;
5220 struct zoneref *z;
5221 struct per_cpu_pages *pcp;
5222 struct list_head *pcp_list;
5223 struct alloc_context ac;
5224 gfp_t alloc_gfp;
5225 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5226 int nr_populated = 0, nr_account = 0;
5227
5228 /*
5229 * Skip populated array elements to determine if any pages need
5230 * to be allocated before disabling IRQs.
5231 */
5232 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
5233 nr_populated++;
5234
5235 /* No pages requested? */
5236 if (unlikely(nr_pages <= 0))
5237 goto out;
5238
5239 /* Already populated array? */
5240 if (unlikely(page_array && nr_pages - nr_populated == 0))
5241 goto out;
5242
5243 /* Use the single page allocator for one page. */
5244 if (nr_pages - nr_populated == 1)
5245 goto failed;
5246
5247#ifdef CONFIG_PAGE_OWNER
5248 /*
5249 * PAGE_OWNER may recurse into the allocator to allocate space to
5250 * save the stack with pagesets.lock held. Releasing/reacquiring
5251 * removes much of the performance benefit of bulk allocation so
5252 * force the caller to allocate one page at a time as it'll have
5253 * similar performance to added complexity to the bulk allocator.
5254 */
5255 if (static_branch_unlikely(&page_owner_inited))
5256 goto failed;
5257#endif
5258
5259 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5260 gfp &= gfp_allowed_mask;
5261 alloc_gfp = gfp;
5262 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
5263 goto out;
5264 gfp = alloc_gfp;
5265
5266 /* Find an allowed local zone that meets the low watermark. */
5267 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
5268 unsigned long mark;
5269
5270 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5271 !__cpuset_zone_allowed(zone, gfp)) {
5272 continue;
5273 }
5274
5275 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
5276 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
5277 goto failed;
5278 }
5279
5280 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
5281 if (zone_watermark_fast(zone, 0, mark,
5282 zonelist_zone_idx(ac.preferred_zoneref),
5283 alloc_flags, gfp)) {
5284 break;
5285 }
5286 }
5287
5288 /*
5289 * If there are no allowed local zones that meets the watermarks then
5290 * try to allocate a single page and reclaim if necessary.
5291 */
5292 if (unlikely(!zone))
5293 goto failed;
5294
5295 /* Attempt the batch allocation */
5296 local_lock_irqsave(&pagesets.lock, flags);
5297 pcp = this_cpu_ptr(zone->per_cpu_pageset);
5298 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
5299
5300 while (nr_populated < nr_pages) {
5301
5302 /* Skip existing pages */
5303 if (page_array && page_array[nr_populated]) {
5304 nr_populated++;
5305 continue;
5306 }
5307
5308 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
5309 pcp, pcp_list);
5310 if (unlikely(!page)) {
5311 /* Try and get at least one page */
5312 if (!nr_populated)
5313 goto failed_irq;
5314 break;
5315 }
5316 nr_account++;
5317
5318 prep_new_page(page, 0, gfp, 0);
5319 if (page_list)
5320 list_add(&page->lru, page_list);
5321 else
5322 page_array[nr_populated] = page;
5323 nr_populated++;
5324 }
5325
5326 local_unlock_irqrestore(&pagesets.lock, flags);
5327
5328 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
5329 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
5330
5331out:
5332 return nr_populated;
5333
5334failed_irq:
5335 local_unlock_irqrestore(&pagesets.lock, flags);
5336
5337failed:
5338 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
5339 if (page) {
5340 if (page_list)
5341 list_add(&page->lru, page_list);
5342 else
5343 page_array[nr_populated] = page;
5344 nr_populated++;
5345 }
5346
5347 goto out;
5348}
5349EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
5350
5351/*
5352 * This is the 'heart' of the zoned buddy allocator.
5353 */
5354struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
5355 nodemask_t *nodemask)
5356{
5357 struct page *page;
5358 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5359 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5360 struct alloc_context ac = { };
5361
5362 /*
5363 * There are several places where we assume that the order value is sane
5364 * so bail out early if the request is out of bound.
5365 */
5366 if (unlikely(order >= MAX_ORDER)) {
5367 WARN_ON_ONCE(!(gfp & __GFP_NOWARN));
5368 return NULL;
5369 }
5370
5371 gfp &= gfp_allowed_mask;
5372 /*
5373 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5374 * resp. GFP_NOIO which has to be inherited for all allocation requests
5375 * from a particular context which has been marked by
5376 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5377 * movable zones are not used during allocation.
5378 */
5379 gfp = current_gfp_context(gfp);
5380 alloc_gfp = gfp;
5381 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5382 &alloc_gfp, &alloc_flags))
5383 return NULL;
5384
5385 /*
5386 * Forbid the first pass from falling back to types that fragment
5387 * memory until all local zones are considered.
5388 */
5389 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5390
5391 /* First allocation attempt */
5392 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5393 if (likely(page))
5394 goto out;
5395
5396 alloc_gfp = gfp;
5397 ac.spread_dirty_pages = false;
5398
5399 /*
5400 * Restore the original nodemask if it was potentially replaced with
5401 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5402 */
5403 ac.nodemask = nodemask;
5404
5405 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5406
5407out:
5408 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5409 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5410 __free_pages(page, order);
5411 page = NULL;
5412 }
5413
5414 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5415
5416 return page;
5417}
5418EXPORT_SYMBOL(__alloc_pages);
5419
5420/*
5421 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5422 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5423 * you need to access high mem.
5424 */
5425unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5426{
5427 struct page *page;
5428
5429 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5430 if (!page)
5431 return 0;
5432 return (unsigned long) page_address(page);
5433}
5434EXPORT_SYMBOL(__get_free_pages);
5435
5436unsigned long get_zeroed_page(gfp_t gfp_mask)
5437{
5438 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5439}
5440EXPORT_SYMBOL(get_zeroed_page);
5441
5442/**
5443 * __free_pages - Free pages allocated with alloc_pages().
5444 * @page: The page pointer returned from alloc_pages().
5445 * @order: The order of the allocation.
5446 *
5447 * This function can free multi-page allocations that are not compound
5448 * pages. It does not check that the @order passed in matches that of
5449 * the allocation, so it is easy to leak memory. Freeing more memory
5450 * than was allocated will probably emit a warning.
5451 *
5452 * If the last reference to this page is speculative, it will be released
5453 * by put_page() which only frees the first page of a non-compound
5454 * allocation. To prevent the remaining pages from being leaked, we free
5455 * the subsequent pages here. If you want to use the page's reference
5456 * count to decide when to free the allocation, you should allocate a
5457 * compound page, and use put_page() instead of __free_pages().
5458 *
5459 * Context: May be called in interrupt context or while holding a normal
5460 * spinlock, but not in NMI context or while holding a raw spinlock.
5461 */
5462void __free_pages(struct page *page, unsigned int order)
5463{
5464 if (put_page_testzero(page))
5465 free_the_page(page, order);
5466 else if (!PageHead(page))
5467 while (order-- > 0)
5468 free_the_page(page + (1 << order), order);
5469}
5470EXPORT_SYMBOL(__free_pages);
5471
5472void free_pages(unsigned long addr, unsigned int order)
5473{
5474 if (addr != 0) {
5475 VM_BUG_ON(!virt_addr_valid((void *)addr));
5476 __free_pages(virt_to_page((void *)addr), order);
5477 }
5478}
5479
5480EXPORT_SYMBOL(free_pages);
5481
5482/*
5483 * Page Fragment:
5484 * An arbitrary-length arbitrary-offset area of memory which resides
5485 * within a 0 or higher order page. Multiple fragments within that page
5486 * are individually refcounted, in the page's reference counter.
5487 *
5488 * The page_frag functions below provide a simple allocation framework for
5489 * page fragments. This is used by the network stack and network device
5490 * drivers to provide a backing region of memory for use as either an
5491 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5492 */
5493static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5494 gfp_t gfp_mask)
5495{
5496 struct page *page = NULL;
5497 gfp_t gfp = gfp_mask;
5498
5499#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5500 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5501 __GFP_NOMEMALLOC;
5502 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5503 PAGE_FRAG_CACHE_MAX_ORDER);
5504 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5505#endif
5506 if (unlikely(!page))
5507 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5508
5509 nc->va = page ? page_address(page) : NULL;
5510
5511 return page;
5512}
5513
5514void __page_frag_cache_drain(struct page *page, unsigned int count)
5515{
5516 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5517
5518 if (page_ref_sub_and_test(page, count))
5519 free_the_page(page, compound_order(page));
5520}
5521EXPORT_SYMBOL(__page_frag_cache_drain);
5522
5523void *page_frag_alloc_align(struct page_frag_cache *nc,
5524 unsigned int fragsz, gfp_t gfp_mask,
5525 unsigned int align_mask)
5526{
5527 unsigned int size = PAGE_SIZE;
5528 struct page *page;
5529 int offset;
5530
5531 if (unlikely(!nc->va)) {
5532refill:
5533 page = __page_frag_cache_refill(nc, gfp_mask);
5534 if (!page)
5535 return NULL;
5536
5537#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5538 /* if size can vary use size else just use PAGE_SIZE */
5539 size = nc->size;
5540#endif
5541 /* Even if we own the page, we do not use atomic_set().
5542 * This would break get_page_unless_zero() users.
5543 */
5544 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5545
5546 /* reset page count bias and offset to start of new frag */
5547 nc->pfmemalloc = page_is_pfmemalloc(page);
5548 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5549 nc->offset = size;
5550 }
5551
5552 offset = nc->offset - fragsz;
5553 if (unlikely(offset < 0)) {
5554 page = virt_to_page(nc->va);
5555
5556 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5557 goto refill;
5558
5559 if (unlikely(nc->pfmemalloc)) {
5560 free_the_page(page, compound_order(page));
5561 goto refill;
5562 }
5563
5564#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5565 /* if size can vary use size else just use PAGE_SIZE */
5566 size = nc->size;
5567#endif
5568 /* OK, page count is 0, we can safely set it */
5569 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5570
5571 /* reset page count bias and offset to start of new frag */
5572 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5573 offset = size - fragsz;
5574 }
5575
5576 nc->pagecnt_bias--;
5577 offset &= align_mask;
5578 nc->offset = offset;
5579
5580 return nc->va + offset;
5581}
5582EXPORT_SYMBOL(page_frag_alloc_align);
5583
5584/*
5585 * Frees a page fragment allocated out of either a compound or order 0 page.
5586 */
5587void page_frag_free(void *addr)
5588{
5589 struct page *page = virt_to_head_page(addr);
5590
5591 if (unlikely(put_page_testzero(page)))
5592 free_the_page(page, compound_order(page));
5593}
5594EXPORT_SYMBOL(page_frag_free);
5595
5596static void *make_alloc_exact(unsigned long addr, unsigned int order,
5597 size_t size)
5598{
5599 if (addr) {
5600 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5601 unsigned long used = addr + PAGE_ALIGN(size);
5602
5603 split_page(virt_to_page((void *)addr), order);
5604 while (used < alloc_end) {
5605 free_page(used);
5606 used += PAGE_SIZE;
5607 }
5608 }
5609 return (void *)addr;
5610}
5611
5612/**
5613 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5614 * @size: the number of bytes to allocate
5615 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5616 *
5617 * This function is similar to alloc_pages(), except that it allocates the
5618 * minimum number of pages to satisfy the request. alloc_pages() can only
5619 * allocate memory in power-of-two pages.
5620 *
5621 * This function is also limited by MAX_ORDER.
5622 *
5623 * Memory allocated by this function must be released by free_pages_exact().
5624 *
5625 * Return: pointer to the allocated area or %NULL in case of error.
5626 */
5627void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5628{
5629 unsigned int order = get_order(size);
5630 unsigned long addr;
5631
5632 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5633 gfp_mask &= ~__GFP_COMP;
5634
5635 addr = __get_free_pages(gfp_mask, order);
5636 return make_alloc_exact(addr, order, size);
5637}
5638EXPORT_SYMBOL(alloc_pages_exact);
5639
5640/**
5641 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5642 * pages on a node.
5643 * @nid: the preferred node ID where memory should be allocated
5644 * @size: the number of bytes to allocate
5645 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5646 *
5647 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5648 * back.
5649 *
5650 * Return: pointer to the allocated area or %NULL in case of error.
5651 */
5652void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5653{
5654 unsigned int order = get_order(size);
5655 struct page *p;
5656
5657 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5658 gfp_mask &= ~__GFP_COMP;
5659
5660 p = alloc_pages_node(nid, gfp_mask, order);
5661 if (!p)
5662 return NULL;
5663 return make_alloc_exact((unsigned long)page_address(p), order, size);
5664}
5665
5666/**
5667 * free_pages_exact - release memory allocated via alloc_pages_exact()
5668 * @virt: the value returned by alloc_pages_exact.
5669 * @size: size of allocation, same value as passed to alloc_pages_exact().
5670 *
5671 * Release the memory allocated by a previous call to alloc_pages_exact.
5672 */
5673void free_pages_exact(void *virt, size_t size)
5674{
5675 unsigned long addr = (unsigned long)virt;
5676 unsigned long end = addr + PAGE_ALIGN(size);
5677
5678 while (addr < end) {
5679 free_page(addr);
5680 addr += PAGE_SIZE;
5681 }
5682}
5683EXPORT_SYMBOL(free_pages_exact);
5684
5685/**
5686 * nr_free_zone_pages - count number of pages beyond high watermark
5687 * @offset: The zone index of the highest zone
5688 *
5689 * nr_free_zone_pages() counts the number of pages which are beyond the
5690 * high watermark within all zones at or below a given zone index. For each
5691 * zone, the number of pages is calculated as:
5692 *
5693 * nr_free_zone_pages = managed_pages - high_pages
5694 *
5695 * Return: number of pages beyond high watermark.
5696 */
5697static unsigned long nr_free_zone_pages(int offset)
5698{
5699 struct zoneref *z;
5700 struct zone *zone;
5701
5702 /* Just pick one node, since fallback list is circular */
5703 unsigned long sum = 0;
5704
5705 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5706
5707 for_each_zone_zonelist(zone, z, zonelist, offset) {
5708 unsigned long size = zone_managed_pages(zone);
5709 unsigned long high = high_wmark_pages(zone);
5710 if (size > high)
5711 sum += size - high;
5712 }
5713
5714 return sum;
5715}
5716
5717/**
5718 * nr_free_buffer_pages - count number of pages beyond high watermark
5719 *
5720 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5721 * watermark within ZONE_DMA and ZONE_NORMAL.
5722 *
5723 * Return: number of pages beyond high watermark within ZONE_DMA and
5724 * ZONE_NORMAL.
5725 */
5726unsigned long nr_free_buffer_pages(void)
5727{
5728 return nr_free_zone_pages(gfp_zone(GFP_USER));
5729}
5730EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5731
5732static inline void show_node(struct zone *zone)
5733{
5734 if (IS_ENABLED(CONFIG_NUMA))
5735 printk("Node %d ", zone_to_nid(zone));
5736}
5737
5738long si_mem_available(void)
5739{
5740 long available;
5741 unsigned long pagecache;
5742 unsigned long wmark_low = 0;
5743 unsigned long pages[NR_LRU_LISTS];
5744 unsigned long reclaimable;
5745 struct zone *zone;
5746 int lru;
5747
5748 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5749 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5750
5751 for_each_zone(zone)
5752 wmark_low += low_wmark_pages(zone);
5753
5754 /*
5755 * Estimate the amount of memory available for userspace allocations,
5756 * without causing swapping.
5757 */
5758 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5759
5760 /*
5761 * Not all the page cache can be freed, otherwise the system will
5762 * start swapping. Assume at least half of the page cache, or the
5763 * low watermark worth of cache, needs to stay.
5764 */
5765 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5766 pagecache -= min(pagecache / 2, wmark_low);
5767 available += pagecache;
5768
5769 /*
5770 * Part of the reclaimable slab and other kernel memory consists of
5771 * items that are in use, and cannot be freed. Cap this estimate at the
5772 * low watermark.
5773 */
5774 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5775 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5776 available += reclaimable - min(reclaimable / 2, wmark_low);
5777
5778 if (available < 0)
5779 available = 0;
5780 return available;
5781}
5782EXPORT_SYMBOL_GPL(si_mem_available);
5783
5784void si_meminfo(struct sysinfo *val)
5785{
5786 val->totalram = totalram_pages();
5787 val->sharedram = global_node_page_state(NR_SHMEM);
5788 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5789 val->bufferram = nr_blockdev_pages();
5790 val->totalhigh = totalhigh_pages();
5791 val->freehigh = nr_free_highpages();
5792 val->mem_unit = PAGE_SIZE;
5793}
5794
5795EXPORT_SYMBOL(si_meminfo);
5796
5797#ifdef CONFIG_NUMA
5798void si_meminfo_node(struct sysinfo *val, int nid)
5799{
5800 int zone_type; /* needs to be signed */
5801 unsigned long managed_pages = 0;
5802 unsigned long managed_highpages = 0;
5803 unsigned long free_highpages = 0;
5804 pg_data_t *pgdat = NODE_DATA(nid);
5805
5806 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5807 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5808 val->totalram = managed_pages;
5809 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5810 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5811#ifdef CONFIG_HIGHMEM
5812 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5813 struct zone *zone = &pgdat->node_zones[zone_type];
5814
5815 if (is_highmem(zone)) {
5816 managed_highpages += zone_managed_pages(zone);
5817 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5818 }
5819 }
5820 val->totalhigh = managed_highpages;
5821 val->freehigh = free_highpages;
5822#else
5823 val->totalhigh = managed_highpages;
5824 val->freehigh = free_highpages;
5825#endif
5826 val->mem_unit = PAGE_SIZE;
5827}
5828#endif
5829
5830/*
5831 * Determine whether the node should be displayed or not, depending on whether
5832 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5833 */
5834static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5835{
5836 if (!(flags & SHOW_MEM_FILTER_NODES))
5837 return false;
5838
5839 /*
5840 * no node mask - aka implicit memory numa policy. Do not bother with
5841 * the synchronization - read_mems_allowed_begin - because we do not
5842 * have to be precise here.
5843 */
5844 if (!nodemask)
5845 nodemask = &cpuset_current_mems_allowed;
5846
5847 return !node_isset(nid, *nodemask);
5848}
5849
5850#define K(x) ((x) << (PAGE_SHIFT-10))
5851
5852static void show_migration_types(unsigned char type)
5853{
5854 static const char types[MIGRATE_TYPES] = {
5855 [MIGRATE_UNMOVABLE] = 'U',
5856 [MIGRATE_MOVABLE] = 'M',
5857 [MIGRATE_RECLAIMABLE] = 'E',
5858 [MIGRATE_HIGHATOMIC] = 'H',
5859#ifdef CONFIG_CMA
5860 [MIGRATE_CMA] = 'C',
5861#endif
5862#ifdef CONFIG_MEMORY_ISOLATION
5863 [MIGRATE_ISOLATE] = 'I',
5864#endif
5865 };
5866 char tmp[MIGRATE_TYPES + 1];
5867 char *p = tmp;
5868 int i;
5869
5870 for (i = 0; i < MIGRATE_TYPES; i++) {
5871 if (type & (1 << i))
5872 *p++ = types[i];
5873 }
5874
5875 *p = '\0';
5876 printk(KERN_CONT "(%s) ", tmp);
5877}
5878
5879/*
5880 * Show free area list (used inside shift_scroll-lock stuff)
5881 * We also calculate the percentage fragmentation. We do this by counting the
5882 * memory on each free list with the exception of the first item on the list.
5883 *
5884 * Bits in @filter:
5885 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5886 * cpuset.
5887 */
5888void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5889{
5890 unsigned long free_pcp = 0;
5891 int cpu;
5892 struct zone *zone;
5893 pg_data_t *pgdat;
5894
5895 for_each_populated_zone(zone) {
5896 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5897 continue;
5898
5899 for_each_online_cpu(cpu)
5900 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5901 }
5902
5903 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5904 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5905 " unevictable:%lu dirty:%lu writeback:%lu\n"
5906 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5907 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5908 " free:%lu free_pcp:%lu free_cma:%lu\n",
5909 global_node_page_state(NR_ACTIVE_ANON),
5910 global_node_page_state(NR_INACTIVE_ANON),
5911 global_node_page_state(NR_ISOLATED_ANON),
5912 global_node_page_state(NR_ACTIVE_FILE),
5913 global_node_page_state(NR_INACTIVE_FILE),
5914 global_node_page_state(NR_ISOLATED_FILE),
5915 global_node_page_state(NR_UNEVICTABLE),
5916 global_node_page_state(NR_FILE_DIRTY),
5917 global_node_page_state(NR_WRITEBACK),
5918 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5919 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5920 global_node_page_state(NR_FILE_MAPPED),
5921 global_node_page_state(NR_SHMEM),
5922 global_node_page_state(NR_PAGETABLE),
5923 global_zone_page_state(NR_BOUNCE),
5924 global_zone_page_state(NR_FREE_PAGES),
5925 free_pcp,
5926 global_zone_page_state(NR_FREE_CMA_PAGES));
5927
5928 for_each_online_pgdat(pgdat) {
5929 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5930 continue;
5931
5932 printk("Node %d"
5933 " active_anon:%lukB"
5934 " inactive_anon:%lukB"
5935 " active_file:%lukB"
5936 " inactive_file:%lukB"
5937 " unevictable:%lukB"
5938 " isolated(anon):%lukB"
5939 " isolated(file):%lukB"
5940 " mapped:%lukB"
5941 " dirty:%lukB"
5942 " writeback:%lukB"
5943 " shmem:%lukB"
5944#ifdef CONFIG_TRANSPARENT_HUGEPAGE
5945 " shmem_thp: %lukB"
5946 " shmem_pmdmapped: %lukB"
5947 " anon_thp: %lukB"
5948#endif
5949 " writeback_tmp:%lukB"
5950 " kernel_stack:%lukB"
5951#ifdef CONFIG_SHADOW_CALL_STACK
5952 " shadow_call_stack:%lukB"
5953#endif
5954 " pagetables:%lukB"
5955 " all_unreclaimable? %s"
5956 "\n",
5957 pgdat->node_id,
5958 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5959 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5960 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5961 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5962 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5963 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5964 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5965 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5966 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5967 K(node_page_state(pgdat, NR_WRITEBACK)),
5968 K(node_page_state(pgdat, NR_SHMEM)),
5969#ifdef CONFIG_TRANSPARENT_HUGEPAGE
5970 K(node_page_state(pgdat, NR_SHMEM_THPS)),
5971 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
5972 K(node_page_state(pgdat, NR_ANON_THPS)),
5973#endif
5974 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5975 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5976#ifdef CONFIG_SHADOW_CALL_STACK
5977 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5978#endif
5979 K(node_page_state(pgdat, NR_PAGETABLE)),
5980 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5981 "yes" : "no");
5982 }
5983
5984 for_each_populated_zone(zone) {
5985 int i;
5986
5987 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5988 continue;
5989
5990 free_pcp = 0;
5991 for_each_online_cpu(cpu)
5992 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5993
5994 show_node(zone);
5995 printk(KERN_CONT
5996 "%s"
5997 " free:%lukB"
5998 " min:%lukB"
5999 " low:%lukB"
6000 " high:%lukB"
6001 " reserved_highatomic:%luKB"
6002 " active_anon:%lukB"
6003 " inactive_anon:%lukB"
6004 " active_file:%lukB"
6005 " inactive_file:%lukB"
6006 " unevictable:%lukB"
6007 " writepending:%lukB"
6008 " present:%lukB"
6009 " managed:%lukB"
6010 " mlocked:%lukB"
6011 " bounce:%lukB"
6012 " free_pcp:%lukB"
6013 " local_pcp:%ukB"
6014 " free_cma:%lukB"
6015 "\n",
6016 zone->name,
6017 K(zone_page_state(zone, NR_FREE_PAGES)),
6018 K(min_wmark_pages(zone)),
6019 K(low_wmark_pages(zone)),
6020 K(high_wmark_pages(zone)),
6021 K(zone->nr_reserved_highatomic),
6022 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
6023 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
6024 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
6025 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
6026 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
6027 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
6028 K(zone->present_pages),
6029 K(zone_managed_pages(zone)),
6030 K(zone_page_state(zone, NR_MLOCK)),
6031 K(zone_page_state(zone, NR_BOUNCE)),
6032 K(free_pcp),
6033 K(this_cpu_read(zone->per_cpu_pageset->count)),
6034 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
6035 printk("lowmem_reserve[]:");
6036 for (i = 0; i < MAX_NR_ZONES; i++)
6037 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
6038 printk(KERN_CONT "\n");
6039 }
6040
6041 for_each_populated_zone(zone) {
6042 unsigned int order;
6043 unsigned long nr[MAX_ORDER], flags, total = 0;
6044 unsigned char types[MAX_ORDER];
6045
6046 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6047 continue;
6048 show_node(zone);
6049 printk(KERN_CONT "%s: ", zone->name);
6050
6051 spin_lock_irqsave(&zone->lock, flags);
6052 for (order = 0; order < MAX_ORDER; order++) {
6053 struct free_area *area = &zone->free_area[order];
6054 int type;
6055
6056 nr[order] = area->nr_free;
6057 total += nr[order] << order;
6058
6059 types[order] = 0;
6060 for (type = 0; type < MIGRATE_TYPES; type++) {
6061 if (!free_area_empty(area, type))
6062 types[order] |= 1 << type;
6063 }
6064 }
6065 spin_unlock_irqrestore(&zone->lock, flags);
6066 for (order = 0; order < MAX_ORDER; order++) {
6067 printk(KERN_CONT "%lu*%lukB ",
6068 nr[order], K(1UL) << order);
6069 if (nr[order])
6070 show_migration_types(types[order]);
6071 }
6072 printk(KERN_CONT "= %lukB\n", K(total));
6073 }
6074
6075 hugetlb_show_meminfo();
6076
6077 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
6078
6079 show_swap_cache_info();
6080}
6081
6082static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
6083{
6084 zoneref->zone = zone;
6085 zoneref->zone_idx = zone_idx(zone);
6086}
6087
6088/*
6089 * Builds allocation fallback zone lists.
6090 *
6091 * Add all populated zones of a node to the zonelist.
6092 */
6093static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
6094{
6095 struct zone *zone;
6096 enum zone_type zone_type = MAX_NR_ZONES;
6097 int nr_zones = 0;
6098
6099 do {
6100 zone_type--;
6101 zone = pgdat->node_zones + zone_type;
6102 if (managed_zone(zone)) {
6103 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
6104 check_highest_zone(zone_type);
6105 }
6106 } while (zone_type);
6107
6108 return nr_zones;
6109}
6110
6111#ifdef CONFIG_NUMA
6112
6113static int __parse_numa_zonelist_order(char *s)
6114{
6115 /*
6116 * We used to support different zonelists modes but they turned
6117 * out to be just not useful. Let's keep the warning in place
6118 * if somebody still use the cmd line parameter so that we do
6119 * not fail it silently
6120 */
6121 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
6122 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
6123 return -EINVAL;
6124 }
6125 return 0;
6126}
6127
6128char numa_zonelist_order[] = "Node";
6129
6130/*
6131 * sysctl handler for numa_zonelist_order
6132 */
6133int numa_zonelist_order_handler(struct ctl_table *table, int write,
6134 void *buffer, size_t *length, loff_t *ppos)
6135{
6136 if (write)
6137 return __parse_numa_zonelist_order(buffer);
6138 return proc_dostring(table, write, buffer, length, ppos);
6139}
6140
6141
6142#define MAX_NODE_LOAD (nr_online_nodes)
6143static int node_load[MAX_NUMNODES];
6144
6145/**
6146 * find_next_best_node - find the next node that should appear in a given node's fallback list
6147 * @node: node whose fallback list we're appending
6148 * @used_node_mask: nodemask_t of already used nodes
6149 *
6150 * We use a number of factors to determine which is the next node that should
6151 * appear on a given node's fallback list. The node should not have appeared
6152 * already in @node's fallback list, and it should be the next closest node
6153 * according to the distance array (which contains arbitrary distance values
6154 * from each node to each node in the system), and should also prefer nodes
6155 * with no CPUs, since presumably they'll have very little allocation pressure
6156 * on them otherwise.
6157 *
6158 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
6159 */
6160static int find_next_best_node(int node, nodemask_t *used_node_mask)
6161{
6162 int n, val;
6163 int min_val = INT_MAX;
6164 int best_node = NUMA_NO_NODE;
6165
6166 /* Use the local node if we haven't already */
6167 if (!node_isset(node, *used_node_mask)) {
6168 node_set(node, *used_node_mask);
6169 return node;
6170 }
6171
6172 for_each_node_state(n, N_MEMORY) {
6173
6174 /* Don't want a node to appear more than once */
6175 if (node_isset(n, *used_node_mask))
6176 continue;
6177
6178 /* Use the distance array to find the distance */
6179 val = node_distance(node, n);
6180
6181 /* Penalize nodes under us ("prefer the next node") */
6182 val += (n < node);
6183
6184 /* Give preference to headless and unused nodes */
6185 if (!cpumask_empty(cpumask_of_node(n)))
6186 val += PENALTY_FOR_NODE_WITH_CPUS;
6187
6188 /* Slight preference for less loaded node */
6189 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
6190 val += node_load[n];
6191
6192 if (val < min_val) {
6193 min_val = val;
6194 best_node = n;
6195 }
6196 }
6197
6198 if (best_node >= 0)
6199 node_set(best_node, *used_node_mask);
6200
6201 return best_node;
6202}
6203
6204
6205/*
6206 * Build zonelists ordered by node and zones within node.
6207 * This results in maximum locality--normal zone overflows into local
6208 * DMA zone, if any--but risks exhausting DMA zone.
6209 */
6210static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
6211 unsigned nr_nodes)
6212{
6213 struct zoneref *zonerefs;
6214 int i;
6215
6216 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6217
6218 for (i = 0; i < nr_nodes; i++) {
6219 int nr_zones;
6220
6221 pg_data_t *node = NODE_DATA(node_order[i]);
6222
6223 nr_zones = build_zonerefs_node(node, zonerefs);
6224 zonerefs += nr_zones;
6225 }
6226 zonerefs->zone = NULL;
6227 zonerefs->zone_idx = 0;
6228}
6229
6230/*
6231 * Build gfp_thisnode zonelists
6232 */
6233static void build_thisnode_zonelists(pg_data_t *pgdat)
6234{
6235 struct zoneref *zonerefs;
6236 int nr_zones;
6237
6238 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
6239 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6240 zonerefs += nr_zones;
6241 zonerefs->zone = NULL;
6242 zonerefs->zone_idx = 0;
6243}
6244
6245/*
6246 * Build zonelists ordered by zone and nodes within zones.
6247 * This results in conserving DMA zone[s] until all Normal memory is
6248 * exhausted, but results in overflowing to remote node while memory
6249 * may still exist in local DMA zone.
6250 */
6251
6252static void build_zonelists(pg_data_t *pgdat)
6253{
6254 static int node_order[MAX_NUMNODES];
6255 int node, load, nr_nodes = 0;
6256 nodemask_t used_mask = NODE_MASK_NONE;
6257 int local_node, prev_node;
6258
6259 /* NUMA-aware ordering of nodes */
6260 local_node = pgdat->node_id;
6261 load = nr_online_nodes;
6262 prev_node = local_node;
6263
6264 memset(node_order, 0, sizeof(node_order));
6265 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
6266 /*
6267 * We don't want to pressure a particular node.
6268 * So adding penalty to the first node in same
6269 * distance group to make it round-robin.
6270 */
6271 if (node_distance(local_node, node) !=
6272 node_distance(local_node, prev_node))
6273 node_load[node] = load;
6274
6275 node_order[nr_nodes++] = node;
6276 prev_node = node;
6277 load--;
6278 }
6279
6280 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
6281 build_thisnode_zonelists(pgdat);
6282}
6283
6284#ifdef CONFIG_HAVE_MEMORYLESS_NODES
6285/*
6286 * Return node id of node used for "local" allocations.
6287 * I.e., first node id of first zone in arg node's generic zonelist.
6288 * Used for initializing percpu 'numa_mem', which is used primarily
6289 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
6290 */
6291int local_memory_node(int node)
6292{
6293 struct zoneref *z;
6294
6295 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
6296 gfp_zone(GFP_KERNEL),
6297 NULL);
6298 return zone_to_nid(z->zone);
6299}
6300#endif
6301
6302static void setup_min_unmapped_ratio(void);
6303static void setup_min_slab_ratio(void);
6304#else /* CONFIG_NUMA */
6305
6306static void build_zonelists(pg_data_t *pgdat)
6307{
6308 int node, local_node;
6309 struct zoneref *zonerefs;
6310 int nr_zones;
6311
6312 local_node = pgdat->node_id;
6313
6314 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6315 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6316 zonerefs += nr_zones;
6317
6318 /*
6319 * Now we build the zonelist so that it contains the zones
6320 * of all the other nodes.
6321 * We don't want to pressure a particular node, so when
6322 * building the zones for node N, we make sure that the
6323 * zones coming right after the local ones are those from
6324 * node N+1 (modulo N)
6325 */
6326 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
6327 if (!node_online(node))
6328 continue;
6329 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6330 zonerefs += nr_zones;
6331 }
6332 for (node = 0; node < local_node; node++) {
6333 if (!node_online(node))
6334 continue;
6335 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6336 zonerefs += nr_zones;
6337 }
6338
6339 zonerefs->zone = NULL;
6340 zonerefs->zone_idx = 0;
6341}
6342
6343#endif /* CONFIG_NUMA */
6344
6345/*
6346 * Boot pageset table. One per cpu which is going to be used for all
6347 * zones and all nodes. The parameters will be set in such a way
6348 * that an item put on a list will immediately be handed over to
6349 * the buddy list. This is safe since pageset manipulation is done
6350 * with interrupts disabled.
6351 *
6352 * The boot_pagesets must be kept even after bootup is complete for
6353 * unused processors and/or zones. They do play a role for bootstrapping
6354 * hotplugged processors.
6355 *
6356 * zoneinfo_show() and maybe other functions do
6357 * not check if the processor is online before following the pageset pointer.
6358 * Other parts of the kernel may not check if the zone is available.
6359 */
6360static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
6361/* These effectively disable the pcplists in the boot pageset completely */
6362#define BOOT_PAGESET_HIGH 0
6363#define BOOT_PAGESET_BATCH 1
6364static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
6365static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
6366static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6367
6368static void __build_all_zonelists(void *data)
6369{
6370 int nid;
6371 int __maybe_unused cpu;
6372 pg_data_t *self = data;
6373 static DEFINE_SPINLOCK(lock);
6374
6375 spin_lock(&lock);
6376
6377#ifdef CONFIG_NUMA
6378 memset(node_load, 0, sizeof(node_load));
6379#endif
6380
6381 /*
6382 * This node is hotadded and no memory is yet present. So just
6383 * building zonelists is fine - no need to touch other nodes.
6384 */
6385 if (self && !node_online(self->node_id)) {
6386 build_zonelists(self);
6387 } else {
6388 for_each_online_node(nid) {
6389 pg_data_t *pgdat = NODE_DATA(nid);
6390
6391 build_zonelists(pgdat);
6392 }
6393
6394#ifdef CONFIG_HAVE_MEMORYLESS_NODES
6395 /*
6396 * We now know the "local memory node" for each node--
6397 * i.e., the node of the first zone in the generic zonelist.
6398 * Set up numa_mem percpu variable for on-line cpus. During
6399 * boot, only the boot cpu should be on-line; we'll init the
6400 * secondary cpus' numa_mem as they come on-line. During
6401 * node/memory hotplug, we'll fixup all on-line cpus.
6402 */
6403 for_each_online_cpu(cpu)
6404 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6405#endif
6406 }
6407
6408 spin_unlock(&lock);
6409}
6410
6411static noinline void __init
6412build_all_zonelists_init(void)
6413{
6414 int cpu;
6415
6416 __build_all_zonelists(NULL);
6417
6418 /*
6419 * Initialize the boot_pagesets that are going to be used
6420 * for bootstrapping processors. The real pagesets for
6421 * each zone will be allocated later when the per cpu
6422 * allocator is available.
6423 *
6424 * boot_pagesets are used also for bootstrapping offline
6425 * cpus if the system is already booted because the pagesets
6426 * are needed to initialize allocators on a specific cpu too.
6427 * F.e. the percpu allocator needs the page allocator which
6428 * needs the percpu allocator in order to allocate its pagesets
6429 * (a chicken-egg dilemma).
6430 */
6431 for_each_possible_cpu(cpu)
6432 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
6433
6434 mminit_verify_zonelist();
6435 cpuset_init_current_mems_allowed();
6436}
6437
6438/*
6439 * unless system_state == SYSTEM_BOOTING.
6440 *
6441 * __ref due to call of __init annotated helper build_all_zonelists_init
6442 * [protected by SYSTEM_BOOTING].
6443 */
6444void __ref build_all_zonelists(pg_data_t *pgdat)
6445{
6446 unsigned long vm_total_pages;
6447
6448 if (system_state == SYSTEM_BOOTING) {
6449 build_all_zonelists_init();
6450 } else {
6451 __build_all_zonelists(pgdat);
6452 /* cpuset refresh routine should be here */
6453 }
6454 /* Get the number of free pages beyond high watermark in all zones. */
6455 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6456 /*
6457 * Disable grouping by mobility if the number of pages in the
6458 * system is too low to allow the mechanism to work. It would be
6459 * more accurate, but expensive to check per-zone. This check is
6460 * made on memory-hotadd so a system can start with mobility
6461 * disabled and enable it later
6462 */
6463 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6464 page_group_by_mobility_disabled = 1;
6465 else
6466 page_group_by_mobility_disabled = 0;
6467
6468 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6469 nr_online_nodes,
6470 page_group_by_mobility_disabled ? "off" : "on",
6471 vm_total_pages);
6472#ifdef CONFIG_NUMA
6473 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6474#endif
6475}
6476
6477/* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6478static bool __meminit
6479overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6480{
6481 static struct memblock_region *r;
6482
6483 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6484 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6485 for_each_mem_region(r) {
6486 if (*pfn < memblock_region_memory_end_pfn(r))
6487 break;
6488 }
6489 }
6490 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6491 memblock_is_mirror(r)) {
6492 *pfn = memblock_region_memory_end_pfn(r);
6493 return true;
6494 }
6495 }
6496 return false;
6497}
6498
6499/*
6500 * Initially all pages are reserved - free ones are freed
6501 * up by memblock_free_all() once the early boot process is
6502 * done. Non-atomic initialization, single-pass.
6503 *
6504 * All aligned pageblocks are initialized to the specified migratetype
6505 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6506 * zone stats (e.g., nr_isolate_pageblock) are touched.
6507 */
6508void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6509 unsigned long start_pfn, unsigned long zone_end_pfn,
6510 enum meminit_context context,
6511 struct vmem_altmap *altmap, int migratetype)
6512{
6513 unsigned long pfn, end_pfn = start_pfn + size;
6514 struct page *page;
6515
6516 if (highest_memmap_pfn < end_pfn - 1)
6517 highest_memmap_pfn = end_pfn - 1;
6518
6519#ifdef CONFIG_ZONE_DEVICE
6520 /*
6521 * Honor reservation requested by the driver for this ZONE_DEVICE
6522 * memory. We limit the total number of pages to initialize to just
6523 * those that might contain the memory mapping. We will defer the
6524 * ZONE_DEVICE page initialization until after we have released
6525 * the hotplug lock.
6526 */
6527 if (zone == ZONE_DEVICE) {
6528 if (!altmap)
6529 return;
6530
6531 if (start_pfn == altmap->base_pfn)
6532 start_pfn += altmap->reserve;
6533 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6534 }
6535#endif
6536
6537 for (pfn = start_pfn; pfn < end_pfn; ) {
6538 /*
6539 * There can be holes in boot-time mem_map[]s handed to this
6540 * function. They do not exist on hotplugged memory.
6541 */
6542 if (context == MEMINIT_EARLY) {
6543 if (overlap_memmap_init(zone, &pfn))
6544 continue;
6545 if (defer_init(nid, pfn, zone_end_pfn))
6546 break;
6547 }
6548
6549 page = pfn_to_page(pfn);
6550 __init_single_page(page, pfn, zone, nid);
6551 if (context == MEMINIT_HOTPLUG)
6552 __SetPageReserved(page);
6553
6554 /*
6555 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6556 * such that unmovable allocations won't be scattered all
6557 * over the place during system boot.
6558 */
6559 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6560 set_pageblock_migratetype(page, migratetype);
6561 cond_resched();
6562 }
6563 pfn++;
6564 }
6565}
6566
6567#ifdef CONFIG_ZONE_DEVICE
6568void __ref memmap_init_zone_device(struct zone *zone,
6569 unsigned long start_pfn,
6570 unsigned long nr_pages,
6571 struct dev_pagemap *pgmap)
6572{
6573 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6574 struct pglist_data *pgdat = zone->zone_pgdat;
6575 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6576 unsigned long zone_idx = zone_idx(zone);
6577 unsigned long start = jiffies;
6578 int nid = pgdat->node_id;
6579
6580 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6581 return;
6582
6583 /*
6584 * The call to memmap_init should have already taken care
6585 * of the pages reserved for the memmap, so we can just jump to
6586 * the end of that region and start processing the device pages.
6587 */
6588 if (altmap) {
6589 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6590 nr_pages = end_pfn - start_pfn;
6591 }
6592
6593 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6594 struct page *page = pfn_to_page(pfn);
6595
6596 __init_single_page(page, pfn, zone_idx, nid);
6597
6598 /*
6599 * Mark page reserved as it will need to wait for onlining
6600 * phase for it to be fully associated with a zone.
6601 *
6602 * We can use the non-atomic __set_bit operation for setting
6603 * the flag as we are still initializing the pages.
6604 */
6605 __SetPageReserved(page);
6606
6607 /*
6608 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6609 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6610 * ever freed or placed on a driver-private list.
6611 */
6612 page->pgmap = pgmap;
6613 page->zone_device_data = NULL;
6614
6615 /*
6616 * Mark the block movable so that blocks are reserved for
6617 * movable at startup. This will force kernel allocations
6618 * to reserve their blocks rather than leaking throughout
6619 * the address space during boot when many long-lived
6620 * kernel allocations are made.
6621 *
6622 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6623 * because this is done early in section_activate()
6624 */
6625 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6626 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6627 cond_resched();
6628 }
6629 }
6630
6631 pr_info("%s initialised %lu pages in %ums\n", __func__,
6632 nr_pages, jiffies_to_msecs(jiffies - start));
6633}
6634
6635#endif
6636static void __meminit zone_init_free_lists(struct zone *zone)
6637{
6638 unsigned int order, t;
6639 for_each_migratetype_order(order, t) {
6640 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6641 zone->free_area[order].nr_free = 0;
6642 }
6643}
6644
6645#if !defined(CONFIG_FLATMEM)
6646/*
6647 * Only struct pages that correspond to ranges defined by memblock.memory
6648 * are zeroed and initialized by going through __init_single_page() during
6649 * memmap_init_zone_range().
6650 *
6651 * But, there could be struct pages that correspond to holes in
6652 * memblock.memory. This can happen because of the following reasons:
6653 * - physical memory bank size is not necessarily the exact multiple of the
6654 * arbitrary section size
6655 * - early reserved memory may not be listed in memblock.memory
6656 * - memory layouts defined with memmap= kernel parameter may not align
6657 * nicely with memmap sections
6658 *
6659 * Explicitly initialize those struct pages so that:
6660 * - PG_Reserved is set
6661 * - zone and node links point to zone and node that span the page if the
6662 * hole is in the middle of a zone
6663 * - zone and node links point to adjacent zone/node if the hole falls on
6664 * the zone boundary; the pages in such holes will be prepended to the
6665 * zone/node above the hole except for the trailing pages in the last
6666 * section that will be appended to the zone/node below.
6667 */
6668static void __init init_unavailable_range(unsigned long spfn,
6669 unsigned long epfn,
6670 int zone, int node)
6671{
6672 unsigned long pfn;
6673 u64 pgcnt = 0;
6674
6675 for (pfn = spfn; pfn < epfn; pfn++) {
6676 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6677 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6678 + pageblock_nr_pages - 1;
6679 continue;
6680 }
6681 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6682 __SetPageReserved(pfn_to_page(pfn));
6683 pgcnt++;
6684 }
6685
6686 if (pgcnt)
6687 pr_info("On node %d, zone %s: %lld pages in unavailable ranges",
6688 node, zone_names[zone], pgcnt);
6689}
6690#else
6691static inline void init_unavailable_range(unsigned long spfn,
6692 unsigned long epfn,
6693 int zone, int node)
6694{
6695}
6696#endif
6697
6698static void __init memmap_init_zone_range(struct zone *zone,
6699 unsigned long start_pfn,
6700 unsigned long end_pfn,
6701 unsigned long *hole_pfn)
6702{
6703 unsigned long zone_start_pfn = zone->zone_start_pfn;
6704 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6705 int nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6706
6707 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6708 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
6709
6710 if (start_pfn >= end_pfn)
6711 return;
6712
6713 memmap_init_range(end_pfn - start_pfn, nid, zone_id, start_pfn,
6714 zone_end_pfn, MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6715
6716 if (*hole_pfn < start_pfn)
6717 init_unavailable_range(*hole_pfn, start_pfn, zone_id, nid);
6718
6719 *hole_pfn = end_pfn;
6720}
6721
6722static void __init memmap_init(void)
6723{
6724 unsigned long start_pfn, end_pfn;
6725 unsigned long hole_pfn = 0;
6726 int i, j, zone_id, nid;
6727
6728 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6729 struct pglist_data *node = NODE_DATA(nid);
6730
6731 for (j = 0; j < MAX_NR_ZONES; j++) {
6732 struct zone *zone = node->node_zones + j;
6733
6734 if (!populated_zone(zone))
6735 continue;
6736
6737 memmap_init_zone_range(zone, start_pfn, end_pfn,
6738 &hole_pfn);
6739 zone_id = j;
6740 }
6741 }
6742
6743#ifdef CONFIG_SPARSEMEM
6744 /*
6745 * Initialize the memory map for hole in the range [memory_end,
6746 * section_end].
6747 * Append the pages in this hole to the highest zone in the last
6748 * node.
6749 * The call to init_unavailable_range() is outside the ifdef to
6750 * silence the compiler warining about zone_id set but not used;
6751 * for FLATMEM it is a nop anyway
6752 */
6753 end_pfn = round_up(end_pfn, PAGES_PER_SECTION);
6754 if (hole_pfn < end_pfn)
6755#endif
6756 init_unavailable_range(hole_pfn, end_pfn, zone_id, nid);
6757}
6758
6759static int zone_batchsize(struct zone *zone)
6760{
6761#ifdef CONFIG_MMU
6762 int batch;
6763
6764 /*
6765 * The number of pages to batch allocate is either ~0.1%
6766 * of the zone or 1MB, whichever is smaller. The batch
6767 * size is striking a balance between allocation latency
6768 * and zone lock contention.
6769 */
6770 batch = min(zone_managed_pages(zone) >> 10, (1024 * 1024) / PAGE_SIZE);
6771 batch /= 4; /* We effectively *= 4 below */
6772 if (batch < 1)
6773 batch = 1;
6774
6775 /*
6776 * Clamp the batch to a 2^n - 1 value. Having a power
6777 * of 2 value was found to be more likely to have
6778 * suboptimal cache aliasing properties in some cases.
6779 *
6780 * For example if 2 tasks are alternately allocating
6781 * batches of pages, one task can end up with a lot
6782 * of pages of one half of the possible page colors
6783 * and the other with pages of the other colors.
6784 */
6785 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6786
6787 return batch;
6788
6789#else
6790 /* The deferral and batching of frees should be suppressed under NOMMU
6791 * conditions.
6792 *
6793 * The problem is that NOMMU needs to be able to allocate large chunks
6794 * of contiguous memory as there's no hardware page translation to
6795 * assemble apparent contiguous memory from discontiguous pages.
6796 *
6797 * Queueing large contiguous runs of pages for batching, however,
6798 * causes the pages to actually be freed in smaller chunks. As there
6799 * can be a significant delay between the individual batches being
6800 * recycled, this leads to the once large chunks of space being
6801 * fragmented and becoming unavailable for high-order allocations.
6802 */
6803 return 0;
6804#endif
6805}
6806
6807static int zone_highsize(struct zone *zone, int batch, int cpu_online)
6808{
6809#ifdef CONFIG_MMU
6810 int high;
6811 int nr_split_cpus;
6812 unsigned long total_pages;
6813
6814 if (!percpu_pagelist_high_fraction) {
6815 /*
6816 * By default, the high value of the pcp is based on the zone
6817 * low watermark so that if they are full then background
6818 * reclaim will not be started prematurely.
6819 */
6820 total_pages = low_wmark_pages(zone);
6821 } else {
6822 /*
6823 * If percpu_pagelist_high_fraction is configured, the high
6824 * value is based on a fraction of the managed pages in the
6825 * zone.
6826 */
6827 total_pages = zone_managed_pages(zone) / percpu_pagelist_high_fraction;
6828 }
6829
6830 /*
6831 * Split the high value across all online CPUs local to the zone. Note
6832 * that early in boot that CPUs may not be online yet and that during
6833 * CPU hotplug that the cpumask is not yet updated when a CPU is being
6834 * onlined. For memory nodes that have no CPUs, split pcp->high across
6835 * all online CPUs to mitigate the risk that reclaim is triggered
6836 * prematurely due to pages stored on pcp lists.
6837 */
6838 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
6839 if (!nr_split_cpus)
6840 nr_split_cpus = num_online_cpus();
6841 high = total_pages / nr_split_cpus;
6842
6843 /*
6844 * Ensure high is at least batch*4. The multiple is based on the
6845 * historical relationship between high and batch.
6846 */
6847 high = max(high, batch << 2);
6848
6849 return high;
6850#else
6851 return 0;
6852#endif
6853}
6854
6855/*
6856 * pcp->high and pcp->batch values are related and generally batch is lower
6857 * than high. They are also related to pcp->count such that count is lower
6858 * than high, and as soon as it reaches high, the pcplist is flushed.
6859 *
6860 * However, guaranteeing these relations at all times would require e.g. write
6861 * barriers here but also careful usage of read barriers at the read side, and
6862 * thus be prone to error and bad for performance. Thus the update only prevents
6863 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6864 * can cope with those fields changing asynchronously, and fully trust only the
6865 * pcp->count field on the local CPU with interrupts disabled.
6866 *
6867 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6868 * outside of boot time (or some other assurance that no concurrent updaters
6869 * exist).
6870 */
6871static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6872 unsigned long batch)
6873{
6874 WRITE_ONCE(pcp->batch, batch);
6875 WRITE_ONCE(pcp->high, high);
6876}
6877
6878static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
6879{
6880 int pindex;
6881
6882 memset(pcp, 0, sizeof(*pcp));
6883 memset(pzstats, 0, sizeof(*pzstats));
6884
6885 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
6886 INIT_LIST_HEAD(&pcp->lists[pindex]);
6887
6888 /*
6889 * Set batch and high values safe for a boot pageset. A true percpu
6890 * pageset's initialization will update them subsequently. Here we don't
6891 * need to be as careful as pageset_update() as nobody can access the
6892 * pageset yet.
6893 */
6894 pcp->high = BOOT_PAGESET_HIGH;
6895 pcp->batch = BOOT_PAGESET_BATCH;
6896 pcp->free_factor = 0;
6897}
6898
6899static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
6900 unsigned long batch)
6901{
6902 struct per_cpu_pages *pcp;
6903 int cpu;
6904
6905 for_each_possible_cpu(cpu) {
6906 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6907 pageset_update(pcp, high, batch);
6908 }
6909}
6910
6911/*
6912 * Calculate and set new high and batch values for all per-cpu pagesets of a
6913 * zone based on the zone's size.
6914 */
6915static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
6916{
6917 int new_high, new_batch;
6918
6919 new_batch = max(1, zone_batchsize(zone));
6920 new_high = zone_highsize(zone, new_batch, cpu_online);
6921
6922 if (zone->pageset_high == new_high &&
6923 zone->pageset_batch == new_batch)
6924 return;
6925
6926 zone->pageset_high = new_high;
6927 zone->pageset_batch = new_batch;
6928
6929 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
6930}
6931
6932void __meminit setup_zone_pageset(struct zone *zone)
6933{
6934 int cpu;
6935
6936 /* Size may be 0 on !SMP && !NUMA */
6937 if (sizeof(struct per_cpu_zonestat) > 0)
6938 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
6939
6940 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
6941 for_each_possible_cpu(cpu) {
6942 struct per_cpu_pages *pcp;
6943 struct per_cpu_zonestat *pzstats;
6944
6945 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6946 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
6947 per_cpu_pages_init(pcp, pzstats);
6948 }
6949
6950 zone_set_pageset_high_and_batch(zone, 0);
6951}
6952
6953/*
6954 * Allocate per cpu pagesets and initialize them.
6955 * Before this call only boot pagesets were available.
6956 */
6957void __init setup_per_cpu_pageset(void)
6958{
6959 struct pglist_data *pgdat;
6960 struct zone *zone;
6961 int __maybe_unused cpu;
6962
6963 for_each_populated_zone(zone)
6964 setup_zone_pageset(zone);
6965
6966#ifdef CONFIG_NUMA
6967 /*
6968 * Unpopulated zones continue using the boot pagesets.
6969 * The numa stats for these pagesets need to be reset.
6970 * Otherwise, they will end up skewing the stats of
6971 * the nodes these zones are associated with.
6972 */
6973 for_each_possible_cpu(cpu) {
6974 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
6975 memset(pzstats->vm_numa_event, 0,
6976 sizeof(pzstats->vm_numa_event));
6977 }
6978#endif
6979
6980 for_each_online_pgdat(pgdat)
6981 pgdat->per_cpu_nodestats =
6982 alloc_percpu(struct per_cpu_nodestat);
6983}
6984
6985static __meminit void zone_pcp_init(struct zone *zone)
6986{
6987 /*
6988 * per cpu subsystem is not up at this point. The following code
6989 * relies on the ability of the linker to provide the
6990 * offset of a (static) per cpu variable into the per cpu area.
6991 */
6992 zone->per_cpu_pageset = &boot_pageset;
6993 zone->per_cpu_zonestats = &boot_zonestats;
6994 zone->pageset_high = BOOT_PAGESET_HIGH;
6995 zone->pageset_batch = BOOT_PAGESET_BATCH;
6996
6997 if (populated_zone(zone))
6998 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
6999 zone->present_pages, zone_batchsize(zone));
7000}
7001
7002void __meminit init_currently_empty_zone(struct zone *zone,
7003 unsigned long zone_start_pfn,
7004 unsigned long size)
7005{
7006 struct pglist_data *pgdat = zone->zone_pgdat;
7007 int zone_idx = zone_idx(zone) + 1;
7008
7009 if (zone_idx > pgdat->nr_zones)
7010 pgdat->nr_zones = zone_idx;
7011
7012 zone->zone_start_pfn = zone_start_pfn;
7013
7014 mminit_dprintk(MMINIT_TRACE, "memmap_init",
7015 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
7016 pgdat->node_id,
7017 (unsigned long)zone_idx(zone),
7018 zone_start_pfn, (zone_start_pfn + size));
7019
7020 zone_init_free_lists(zone);
7021 zone->initialized = 1;
7022}
7023
7024/**
7025 * get_pfn_range_for_nid - Return the start and end page frames for a node
7026 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
7027 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
7028 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
7029 *
7030 * It returns the start and end page frame of a node based on information
7031 * provided by memblock_set_node(). If called for a node
7032 * with no available memory, a warning is printed and the start and end
7033 * PFNs will be 0.
7034 */
7035void __init get_pfn_range_for_nid(unsigned int nid,
7036 unsigned long *start_pfn, unsigned long *end_pfn)
7037{
7038 unsigned long this_start_pfn, this_end_pfn;
7039 int i;
7040
7041 *start_pfn = -1UL;
7042 *end_pfn = 0;
7043
7044 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
7045 *start_pfn = min(*start_pfn, this_start_pfn);
7046 *end_pfn = max(*end_pfn, this_end_pfn);
7047 }
7048
7049 if (*start_pfn == -1UL)
7050 *start_pfn = 0;
7051}
7052
7053/*
7054 * This finds a zone that can be used for ZONE_MOVABLE pages. The
7055 * assumption is made that zones within a node are ordered in monotonic
7056 * increasing memory addresses so that the "highest" populated zone is used
7057 */
7058static void __init find_usable_zone_for_movable(void)
7059{
7060 int zone_index;
7061 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
7062 if (zone_index == ZONE_MOVABLE)
7063 continue;
7064
7065 if (arch_zone_highest_possible_pfn[zone_index] >
7066 arch_zone_lowest_possible_pfn[zone_index])
7067 break;
7068 }
7069
7070 VM_BUG_ON(zone_index == -1);
7071 movable_zone = zone_index;
7072}
7073
7074/*
7075 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
7076 * because it is sized independent of architecture. Unlike the other zones,
7077 * the starting point for ZONE_MOVABLE is not fixed. It may be different
7078 * in each node depending on the size of each node and how evenly kernelcore
7079 * is distributed. This helper function adjusts the zone ranges
7080 * provided by the architecture for a given node by using the end of the
7081 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
7082 * zones within a node are in order of monotonic increases memory addresses
7083 */
7084static void __init adjust_zone_range_for_zone_movable(int nid,
7085 unsigned long zone_type,
7086 unsigned long node_start_pfn,
7087 unsigned long node_end_pfn,
7088 unsigned long *zone_start_pfn,
7089 unsigned long *zone_end_pfn)
7090{
7091 /* Only adjust if ZONE_MOVABLE is on this node */
7092 if (zone_movable_pfn[nid]) {
7093 /* Size ZONE_MOVABLE */
7094 if (zone_type == ZONE_MOVABLE) {
7095 *zone_start_pfn = zone_movable_pfn[nid];
7096 *zone_end_pfn = min(node_end_pfn,
7097 arch_zone_highest_possible_pfn[movable_zone]);
7098
7099 /* Adjust for ZONE_MOVABLE starting within this range */
7100 } else if (!mirrored_kernelcore &&
7101 *zone_start_pfn < zone_movable_pfn[nid] &&
7102 *zone_end_pfn > zone_movable_pfn[nid]) {
7103 *zone_end_pfn = zone_movable_pfn[nid];
7104
7105 /* Check if this whole range is within ZONE_MOVABLE */
7106 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
7107 *zone_start_pfn = *zone_end_pfn;
7108 }
7109}
7110
7111/*
7112 * Return the number of pages a zone spans in a node, including holes
7113 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
7114 */
7115static unsigned long __init zone_spanned_pages_in_node(int nid,
7116 unsigned long zone_type,
7117 unsigned long node_start_pfn,
7118 unsigned long node_end_pfn,
7119 unsigned long *zone_start_pfn,
7120 unsigned long *zone_end_pfn)
7121{
7122 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7123 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7124 /* When hotadd a new node from cpu_up(), the node should be empty */
7125 if (!node_start_pfn && !node_end_pfn)
7126 return 0;
7127
7128 /* Get the start and end of the zone */
7129 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7130 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7131 adjust_zone_range_for_zone_movable(nid, zone_type,
7132 node_start_pfn, node_end_pfn,
7133 zone_start_pfn, zone_end_pfn);
7134
7135 /* Check that this node has pages within the zone's required range */
7136 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
7137 return 0;
7138
7139 /* Move the zone boundaries inside the node if necessary */
7140 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
7141 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
7142
7143 /* Return the spanned pages */
7144 return *zone_end_pfn - *zone_start_pfn;
7145}
7146
7147/*
7148 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
7149 * then all holes in the requested range will be accounted for.
7150 */
7151unsigned long __init __absent_pages_in_range(int nid,
7152 unsigned long range_start_pfn,
7153 unsigned long range_end_pfn)
7154{
7155 unsigned long nr_absent = range_end_pfn - range_start_pfn;
7156 unsigned long start_pfn, end_pfn;
7157 int i;
7158
7159 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7160 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
7161 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
7162 nr_absent -= end_pfn - start_pfn;
7163 }
7164 return nr_absent;
7165}
7166
7167/**
7168 * absent_pages_in_range - Return number of page frames in holes within a range
7169 * @start_pfn: The start PFN to start searching for holes
7170 * @end_pfn: The end PFN to stop searching for holes
7171 *
7172 * Return: the number of pages frames in memory holes within a range.
7173 */
7174unsigned long __init absent_pages_in_range(unsigned long start_pfn,
7175 unsigned long end_pfn)
7176{
7177 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
7178}
7179
7180/* Return the number of page frames in holes in a zone on a node */
7181static unsigned long __init zone_absent_pages_in_node(int nid,
7182 unsigned long zone_type,
7183 unsigned long node_start_pfn,
7184 unsigned long node_end_pfn)
7185{
7186 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7187 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7188 unsigned long zone_start_pfn, zone_end_pfn;
7189 unsigned long nr_absent;
7190
7191 /* When hotadd a new node from cpu_up(), the node should be empty */
7192 if (!node_start_pfn && !node_end_pfn)
7193 return 0;
7194
7195 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7196 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7197
7198 adjust_zone_range_for_zone_movable(nid, zone_type,
7199 node_start_pfn, node_end_pfn,
7200 &zone_start_pfn, &zone_end_pfn);
7201 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
7202
7203 /*
7204 * ZONE_MOVABLE handling.
7205 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
7206 * and vice versa.
7207 */
7208 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
7209 unsigned long start_pfn, end_pfn;
7210 struct memblock_region *r;
7211
7212 for_each_mem_region(r) {
7213 start_pfn = clamp(memblock_region_memory_base_pfn(r),
7214 zone_start_pfn, zone_end_pfn);
7215 end_pfn = clamp(memblock_region_memory_end_pfn(r),
7216 zone_start_pfn, zone_end_pfn);
7217
7218 if (zone_type == ZONE_MOVABLE &&
7219 memblock_is_mirror(r))
7220 nr_absent += end_pfn - start_pfn;
7221
7222 if (zone_type == ZONE_NORMAL &&
7223 !memblock_is_mirror(r))
7224 nr_absent += end_pfn - start_pfn;
7225 }
7226 }
7227
7228 return nr_absent;
7229}
7230
7231static void __init calculate_node_totalpages(struct pglist_data *pgdat,
7232 unsigned long node_start_pfn,
7233 unsigned long node_end_pfn)
7234{
7235 unsigned long realtotalpages = 0, totalpages = 0;
7236 enum zone_type i;
7237
7238 for (i = 0; i < MAX_NR_ZONES; i++) {
7239 struct zone *zone = pgdat->node_zones + i;
7240 unsigned long zone_start_pfn, zone_end_pfn;
7241 unsigned long spanned, absent;
7242 unsigned long size, real_size;
7243
7244 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
7245 node_start_pfn,
7246 node_end_pfn,
7247 &zone_start_pfn,
7248 &zone_end_pfn);
7249 absent = zone_absent_pages_in_node(pgdat->node_id, i,
7250 node_start_pfn,
7251 node_end_pfn);
7252
7253 size = spanned;
7254 real_size = size - absent;
7255
7256 if (size)
7257 zone->zone_start_pfn = zone_start_pfn;
7258 else
7259 zone->zone_start_pfn = 0;
7260 zone->spanned_pages = size;
7261 zone->present_pages = real_size;
7262
7263 totalpages += size;
7264 realtotalpages += real_size;
7265 }
7266
7267 pgdat->node_spanned_pages = totalpages;
7268 pgdat->node_present_pages = realtotalpages;
7269 pr_debug("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
7270}
7271
7272#ifndef CONFIG_SPARSEMEM
7273/*
7274 * Calculate the size of the zone->blockflags rounded to an unsigned long
7275 * Start by making sure zonesize is a multiple of pageblock_order by rounding
7276 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
7277 * round what is now in bits to nearest long in bits, then return it in
7278 * bytes.
7279 */
7280static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
7281{
7282 unsigned long usemapsize;
7283
7284 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
7285 usemapsize = roundup(zonesize, pageblock_nr_pages);
7286 usemapsize = usemapsize >> pageblock_order;
7287 usemapsize *= NR_PAGEBLOCK_BITS;
7288 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
7289
7290 return usemapsize / 8;
7291}
7292
7293static void __ref setup_usemap(struct zone *zone)
7294{
7295 unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
7296 zone->spanned_pages);
7297 zone->pageblock_flags = NULL;
7298 if (usemapsize) {
7299 zone->pageblock_flags =
7300 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
7301 zone_to_nid(zone));
7302 if (!zone->pageblock_flags)
7303 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
7304 usemapsize, zone->name, zone_to_nid(zone));
7305 }
7306}
7307#else
7308static inline void setup_usemap(struct zone *zone) {}
7309#endif /* CONFIG_SPARSEMEM */
7310
7311#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
7312
7313/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
7314void __init set_pageblock_order(void)
7315{
7316 unsigned int order;
7317
7318 /* Check that pageblock_nr_pages has not already been setup */
7319 if (pageblock_order)
7320 return;
7321
7322 if (HPAGE_SHIFT > PAGE_SHIFT)
7323 order = HUGETLB_PAGE_ORDER;
7324 else
7325 order = MAX_ORDER - 1;
7326
7327 /*
7328 * Assume the largest contiguous order of interest is a huge page.
7329 * This value may be variable depending on boot parameters on IA64 and
7330 * powerpc.
7331 */
7332 pageblock_order = order;
7333}
7334#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7335
7336/*
7337 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
7338 * is unused as pageblock_order is set at compile-time. See
7339 * include/linux/pageblock-flags.h for the values of pageblock_order based on
7340 * the kernel config
7341 */
7342void __init set_pageblock_order(void)
7343{
7344}
7345
7346#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7347
7348static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
7349 unsigned long present_pages)
7350{
7351 unsigned long pages = spanned_pages;
7352
7353 /*
7354 * Provide a more accurate estimation if there are holes within
7355 * the zone and SPARSEMEM is in use. If there are holes within the
7356 * zone, each populated memory region may cost us one or two extra
7357 * memmap pages due to alignment because memmap pages for each
7358 * populated regions may not be naturally aligned on page boundary.
7359 * So the (present_pages >> 4) heuristic is a tradeoff for that.
7360 */
7361 if (spanned_pages > present_pages + (present_pages >> 4) &&
7362 IS_ENABLED(CONFIG_SPARSEMEM))
7363 pages = present_pages;
7364
7365 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
7366}
7367
7368#ifdef CONFIG_TRANSPARENT_HUGEPAGE
7369static void pgdat_init_split_queue(struct pglist_data *pgdat)
7370{
7371 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
7372
7373 spin_lock_init(&ds_queue->split_queue_lock);
7374 INIT_LIST_HEAD(&ds_queue->split_queue);
7375 ds_queue->split_queue_len = 0;
7376}
7377#else
7378static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
7379#endif
7380
7381#ifdef CONFIG_COMPACTION
7382static void pgdat_init_kcompactd(struct pglist_data *pgdat)
7383{
7384 init_waitqueue_head(&pgdat->kcompactd_wait);
7385}
7386#else
7387static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
7388#endif
7389
7390static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
7391{
7392 pgdat_resize_init(pgdat);
7393
7394 pgdat_init_split_queue(pgdat);
7395 pgdat_init_kcompactd(pgdat);
7396
7397 init_waitqueue_head(&pgdat->kswapd_wait);
7398 init_waitqueue_head(&pgdat->pfmemalloc_wait);
7399
7400 pgdat_page_ext_init(pgdat);
7401 lruvec_init(&pgdat->__lruvec);
7402}
7403
7404static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7405 unsigned long remaining_pages)
7406{
7407 atomic_long_set(&zone->managed_pages, remaining_pages);
7408 zone_set_nid(zone, nid);
7409 zone->name = zone_names[idx];
7410 zone->zone_pgdat = NODE_DATA(nid);
7411 spin_lock_init(&zone->lock);
7412 zone_seqlock_init(zone);
7413 zone_pcp_init(zone);
7414}
7415
7416/*
7417 * Set up the zone data structures
7418 * - init pgdat internals
7419 * - init all zones belonging to this node
7420 *
7421 * NOTE: this function is only called during memory hotplug
7422 */
7423#ifdef CONFIG_MEMORY_HOTPLUG
7424void __ref free_area_init_core_hotplug(int nid)
7425{
7426 enum zone_type z;
7427 pg_data_t *pgdat = NODE_DATA(nid);
7428
7429 pgdat_init_internals(pgdat);
7430 for (z = 0; z < MAX_NR_ZONES; z++)
7431 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7432}
7433#endif
7434
7435/*
7436 * Set up the zone data structures:
7437 * - mark all pages reserved
7438 * - mark all memory queues empty
7439 * - clear the memory bitmaps
7440 *
7441 * NOTE: pgdat should get zeroed by caller.
7442 * NOTE: this function is only called during early init.
7443 */
7444static void __init free_area_init_core(struct pglist_data *pgdat)
7445{
7446 enum zone_type j;
7447 int nid = pgdat->node_id;
7448
7449 pgdat_init_internals(pgdat);
7450 pgdat->per_cpu_nodestats = &boot_nodestats;
7451
7452 for (j = 0; j < MAX_NR_ZONES; j++) {
7453 struct zone *zone = pgdat->node_zones + j;
7454 unsigned long size, freesize, memmap_pages;
7455
7456 size = zone->spanned_pages;
7457 freesize = zone->present_pages;
7458
7459 /*
7460 * Adjust freesize so that it accounts for how much memory
7461 * is used by this zone for memmap. This affects the watermark
7462 * and per-cpu initialisations
7463 */
7464 memmap_pages = calc_memmap_size(size, freesize);
7465 if (!is_highmem_idx(j)) {
7466 if (freesize >= memmap_pages) {
7467 freesize -= memmap_pages;
7468 if (memmap_pages)
7469 pr_debug(" %s zone: %lu pages used for memmap\n",
7470 zone_names[j], memmap_pages);
7471 } else
7472 pr_warn(" %s zone: %lu memmap pages exceeds freesize %lu\n",
7473 zone_names[j], memmap_pages, freesize);
7474 }
7475
7476 /* Account for reserved pages */
7477 if (j == 0 && freesize > dma_reserve) {
7478 freesize -= dma_reserve;
7479 pr_debug(" %s zone: %lu pages reserved\n", zone_names[0], dma_reserve);
7480 }
7481
7482 if (!is_highmem_idx(j))
7483 nr_kernel_pages += freesize;
7484 /* Charge for highmem memmap if there are enough kernel pages */
7485 else if (nr_kernel_pages > memmap_pages * 2)
7486 nr_kernel_pages -= memmap_pages;
7487 nr_all_pages += freesize;
7488
7489 /*
7490 * Set an approximate value for lowmem here, it will be adjusted
7491 * when the bootmem allocator frees pages into the buddy system.
7492 * And all highmem pages will be managed by the buddy system.
7493 */
7494 zone_init_internals(zone, j, nid, freesize);
7495
7496 if (!size)
7497 continue;
7498
7499 set_pageblock_order();
7500 setup_usemap(zone);
7501 init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7502 }
7503}
7504
7505#ifdef CONFIG_FLATMEM
7506static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
7507{
7508 unsigned long __maybe_unused start = 0;
7509 unsigned long __maybe_unused offset = 0;
7510
7511 /* Skip empty nodes */
7512 if (!pgdat->node_spanned_pages)
7513 return;
7514
7515 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7516 offset = pgdat->node_start_pfn - start;
7517 /* ia64 gets its own node_mem_map, before this, without bootmem */
7518 if (!pgdat->node_mem_map) {
7519 unsigned long size, end;
7520 struct page *map;
7521
7522 /*
7523 * The zone's endpoints aren't required to be MAX_ORDER
7524 * aligned but the node_mem_map endpoints must be in order
7525 * for the buddy allocator to function correctly.
7526 */
7527 end = pgdat_end_pfn(pgdat);
7528 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7529 size = (end - start) * sizeof(struct page);
7530 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
7531 pgdat->node_id);
7532 if (!map)
7533 panic("Failed to allocate %ld bytes for node %d memory map\n",
7534 size, pgdat->node_id);
7535 pgdat->node_mem_map = map + offset;
7536 }
7537 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7538 __func__, pgdat->node_id, (unsigned long)pgdat,
7539 (unsigned long)pgdat->node_mem_map);
7540#ifndef CONFIG_NUMA
7541 /*
7542 * With no DISCONTIG, the global mem_map is just set as node 0's
7543 */
7544 if (pgdat == NODE_DATA(0)) {
7545 mem_map = NODE_DATA(0)->node_mem_map;
7546 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7547 mem_map -= offset;
7548 }
7549#endif
7550}
7551#else
7552static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
7553#endif /* CONFIG_FLATMEM */
7554
7555#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7556static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7557{
7558 pgdat->first_deferred_pfn = ULONG_MAX;
7559}
7560#else
7561static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7562#endif
7563
7564static void __init free_area_init_node(int nid)
7565{
7566 pg_data_t *pgdat = NODE_DATA(nid);
7567 unsigned long start_pfn = 0;
7568 unsigned long end_pfn = 0;
7569
7570 /* pg_data_t should be reset to zero when it's allocated */
7571 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7572
7573 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7574
7575 pgdat->node_id = nid;
7576 pgdat->node_start_pfn = start_pfn;
7577 pgdat->per_cpu_nodestats = NULL;
7578
7579 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7580 (u64)start_pfn << PAGE_SHIFT,
7581 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7582 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7583
7584 alloc_node_mem_map(pgdat);
7585 pgdat_set_deferred_range(pgdat);
7586
7587 free_area_init_core(pgdat);
7588}
7589
7590void __init free_area_init_memoryless_node(int nid)
7591{
7592 free_area_init_node(nid);
7593}
7594
7595#if MAX_NUMNODES > 1
7596/*
7597 * Figure out the number of possible node ids.
7598 */
7599void __init setup_nr_node_ids(void)
7600{
7601 unsigned int highest;
7602
7603 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7604 nr_node_ids = highest + 1;
7605}
7606#endif
7607
7608/**
7609 * node_map_pfn_alignment - determine the maximum internode alignment
7610 *
7611 * This function should be called after node map is populated and sorted.
7612 * It calculates the maximum power of two alignment which can distinguish
7613 * all the nodes.
7614 *
7615 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7616 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7617 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7618 * shifted, 1GiB is enough and this function will indicate so.
7619 *
7620 * This is used to test whether pfn -> nid mapping of the chosen memory
7621 * model has fine enough granularity to avoid incorrect mapping for the
7622 * populated node map.
7623 *
7624 * Return: the determined alignment in pfn's. 0 if there is no alignment
7625 * requirement (single node).
7626 */
7627unsigned long __init node_map_pfn_alignment(void)
7628{
7629 unsigned long accl_mask = 0, last_end = 0;
7630 unsigned long start, end, mask;
7631 int last_nid = NUMA_NO_NODE;
7632 int i, nid;
7633
7634 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7635 if (!start || last_nid < 0 || last_nid == nid) {
7636 last_nid = nid;
7637 last_end = end;
7638 continue;
7639 }
7640
7641 /*
7642 * Start with a mask granular enough to pin-point to the
7643 * start pfn and tick off bits one-by-one until it becomes
7644 * too coarse to separate the current node from the last.
7645 */
7646 mask = ~((1 << __ffs(start)) - 1);
7647 while (mask && last_end <= (start & (mask << 1)))
7648 mask <<= 1;
7649
7650 /* accumulate all internode masks */
7651 accl_mask |= mask;
7652 }
7653
7654 /* convert mask to number of pages */
7655 return ~accl_mask + 1;
7656}
7657
7658/**
7659 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7660 *
7661 * Return: the minimum PFN based on information provided via
7662 * memblock_set_node().
7663 */
7664unsigned long __init find_min_pfn_with_active_regions(void)
7665{
7666 return PHYS_PFN(memblock_start_of_DRAM());
7667}
7668
7669/*
7670 * early_calculate_totalpages()
7671 * Sum pages in active regions for movable zone.
7672 * Populate N_MEMORY for calculating usable_nodes.
7673 */
7674static unsigned long __init early_calculate_totalpages(void)
7675{
7676 unsigned long totalpages = 0;
7677 unsigned long start_pfn, end_pfn;
7678 int i, nid;
7679
7680 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7681 unsigned long pages = end_pfn - start_pfn;
7682
7683 totalpages += pages;
7684 if (pages)
7685 node_set_state(nid, N_MEMORY);
7686 }
7687 return totalpages;
7688}
7689
7690/*
7691 * Find the PFN the Movable zone begins in each node. Kernel memory
7692 * is spread evenly between nodes as long as the nodes have enough
7693 * memory. When they don't, some nodes will have more kernelcore than
7694 * others
7695 */
7696static void __init find_zone_movable_pfns_for_nodes(void)
7697{
7698 int i, nid;
7699 unsigned long usable_startpfn;
7700 unsigned long kernelcore_node, kernelcore_remaining;
7701 /* save the state before borrow the nodemask */
7702 nodemask_t saved_node_state = node_states[N_MEMORY];
7703 unsigned long totalpages = early_calculate_totalpages();
7704 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7705 struct memblock_region *r;
7706
7707 /* Need to find movable_zone earlier when movable_node is specified. */
7708 find_usable_zone_for_movable();
7709
7710 /*
7711 * If movable_node is specified, ignore kernelcore and movablecore
7712 * options.
7713 */
7714 if (movable_node_is_enabled()) {
7715 for_each_mem_region(r) {
7716 if (!memblock_is_hotpluggable(r))
7717 continue;
7718
7719 nid = memblock_get_region_node(r);
7720
7721 usable_startpfn = PFN_DOWN(r->base);
7722 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7723 min(usable_startpfn, zone_movable_pfn[nid]) :
7724 usable_startpfn;
7725 }
7726
7727 goto out2;
7728 }
7729
7730 /*
7731 * If kernelcore=mirror is specified, ignore movablecore option
7732 */
7733 if (mirrored_kernelcore) {
7734 bool mem_below_4gb_not_mirrored = false;
7735
7736 for_each_mem_region(r) {
7737 if (memblock_is_mirror(r))
7738 continue;
7739
7740 nid = memblock_get_region_node(r);
7741
7742 usable_startpfn = memblock_region_memory_base_pfn(r);
7743
7744 if (usable_startpfn < 0x100000) {
7745 mem_below_4gb_not_mirrored = true;
7746 continue;
7747 }
7748
7749 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7750 min(usable_startpfn, zone_movable_pfn[nid]) :
7751 usable_startpfn;
7752 }
7753
7754 if (mem_below_4gb_not_mirrored)
7755 pr_warn("This configuration results in unmirrored kernel memory.\n");
7756
7757 goto out2;
7758 }
7759
7760 /*
7761 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7762 * amount of necessary memory.
7763 */
7764 if (required_kernelcore_percent)
7765 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7766 10000UL;
7767 if (required_movablecore_percent)
7768 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7769 10000UL;
7770
7771 /*
7772 * If movablecore= was specified, calculate what size of
7773 * kernelcore that corresponds so that memory usable for
7774 * any allocation type is evenly spread. If both kernelcore
7775 * and movablecore are specified, then the value of kernelcore
7776 * will be used for required_kernelcore if it's greater than
7777 * what movablecore would have allowed.
7778 */
7779 if (required_movablecore) {
7780 unsigned long corepages;
7781
7782 /*
7783 * Round-up so that ZONE_MOVABLE is at least as large as what
7784 * was requested by the user
7785 */
7786 required_movablecore =
7787 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7788 required_movablecore = min(totalpages, required_movablecore);
7789 corepages = totalpages - required_movablecore;
7790
7791 required_kernelcore = max(required_kernelcore, corepages);
7792 }
7793
7794 /*
7795 * If kernelcore was not specified or kernelcore size is larger
7796 * than totalpages, there is no ZONE_MOVABLE.
7797 */
7798 if (!required_kernelcore || required_kernelcore >= totalpages)
7799 goto out;
7800
7801 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7802 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7803
7804restart:
7805 /* Spread kernelcore memory as evenly as possible throughout nodes */
7806 kernelcore_node = required_kernelcore / usable_nodes;
7807 for_each_node_state(nid, N_MEMORY) {
7808 unsigned long start_pfn, end_pfn;
7809
7810 /*
7811 * Recalculate kernelcore_node if the division per node
7812 * now exceeds what is necessary to satisfy the requested
7813 * amount of memory for the kernel
7814 */
7815 if (required_kernelcore < kernelcore_node)
7816 kernelcore_node = required_kernelcore / usable_nodes;
7817
7818 /*
7819 * As the map is walked, we track how much memory is usable
7820 * by the kernel using kernelcore_remaining. When it is
7821 * 0, the rest of the node is usable by ZONE_MOVABLE
7822 */
7823 kernelcore_remaining = kernelcore_node;
7824
7825 /* Go through each range of PFNs within this node */
7826 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7827 unsigned long size_pages;
7828
7829 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7830 if (start_pfn >= end_pfn)
7831 continue;
7832
7833 /* Account for what is only usable for kernelcore */
7834 if (start_pfn < usable_startpfn) {
7835 unsigned long kernel_pages;
7836 kernel_pages = min(end_pfn, usable_startpfn)
7837 - start_pfn;
7838
7839 kernelcore_remaining -= min(kernel_pages,
7840 kernelcore_remaining);
7841 required_kernelcore -= min(kernel_pages,
7842 required_kernelcore);
7843
7844 /* Continue if range is now fully accounted */
7845 if (end_pfn <= usable_startpfn) {
7846
7847 /*
7848 * Push zone_movable_pfn to the end so
7849 * that if we have to rebalance
7850 * kernelcore across nodes, we will
7851 * not double account here
7852 */
7853 zone_movable_pfn[nid] = end_pfn;
7854 continue;
7855 }
7856 start_pfn = usable_startpfn;
7857 }
7858
7859 /*
7860 * The usable PFN range for ZONE_MOVABLE is from
7861 * start_pfn->end_pfn. Calculate size_pages as the
7862 * number of pages used as kernelcore
7863 */
7864 size_pages = end_pfn - start_pfn;
7865 if (size_pages > kernelcore_remaining)
7866 size_pages = kernelcore_remaining;
7867 zone_movable_pfn[nid] = start_pfn + size_pages;
7868
7869 /*
7870 * Some kernelcore has been met, update counts and
7871 * break if the kernelcore for this node has been
7872 * satisfied
7873 */
7874 required_kernelcore -= min(required_kernelcore,
7875 size_pages);
7876 kernelcore_remaining -= size_pages;
7877 if (!kernelcore_remaining)
7878 break;
7879 }
7880 }
7881
7882 /*
7883 * If there is still required_kernelcore, we do another pass with one
7884 * less node in the count. This will push zone_movable_pfn[nid] further
7885 * along on the nodes that still have memory until kernelcore is
7886 * satisfied
7887 */
7888 usable_nodes--;
7889 if (usable_nodes && required_kernelcore > usable_nodes)
7890 goto restart;
7891
7892out2:
7893 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7894 for (nid = 0; nid < MAX_NUMNODES; nid++)
7895 zone_movable_pfn[nid] =
7896 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7897
7898out:
7899 /* restore the node_state */
7900 node_states[N_MEMORY] = saved_node_state;
7901}
7902
7903/* Any regular or high memory on that node ? */
7904static void check_for_memory(pg_data_t *pgdat, int nid)
7905{
7906 enum zone_type zone_type;
7907
7908 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7909 struct zone *zone = &pgdat->node_zones[zone_type];
7910 if (populated_zone(zone)) {
7911 if (IS_ENABLED(CONFIG_HIGHMEM))
7912 node_set_state(nid, N_HIGH_MEMORY);
7913 if (zone_type <= ZONE_NORMAL)
7914 node_set_state(nid, N_NORMAL_MEMORY);
7915 break;
7916 }
7917 }
7918}
7919
7920/*
7921 * Some architectures, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7922 * such cases we allow max_zone_pfn sorted in the descending order
7923 */
7924bool __weak arch_has_descending_max_zone_pfns(void)
7925{
7926 return false;
7927}
7928
7929/**
7930 * free_area_init - Initialise all pg_data_t and zone data
7931 * @max_zone_pfn: an array of max PFNs for each zone
7932 *
7933 * This will call free_area_init_node() for each active node in the system.
7934 * Using the page ranges provided by memblock_set_node(), the size of each
7935 * zone in each node and their holes is calculated. If the maximum PFN
7936 * between two adjacent zones match, it is assumed that the zone is empty.
7937 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7938 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7939 * starts where the previous one ended. For example, ZONE_DMA32 starts
7940 * at arch_max_dma_pfn.
7941 */
7942void __init free_area_init(unsigned long *max_zone_pfn)
7943{
7944 unsigned long start_pfn, end_pfn;
7945 int i, nid, zone;
7946 bool descending;
7947
7948 /* Record where the zone boundaries are */
7949 memset(arch_zone_lowest_possible_pfn, 0,
7950 sizeof(arch_zone_lowest_possible_pfn));
7951 memset(arch_zone_highest_possible_pfn, 0,
7952 sizeof(arch_zone_highest_possible_pfn));
7953
7954 start_pfn = find_min_pfn_with_active_regions();
7955 descending = arch_has_descending_max_zone_pfns();
7956
7957 for (i = 0; i < MAX_NR_ZONES; i++) {
7958 if (descending)
7959 zone = MAX_NR_ZONES - i - 1;
7960 else
7961 zone = i;
7962
7963 if (zone == ZONE_MOVABLE)
7964 continue;
7965
7966 end_pfn = max(max_zone_pfn[zone], start_pfn);
7967 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7968 arch_zone_highest_possible_pfn[zone] = end_pfn;
7969
7970 start_pfn = end_pfn;
7971 }
7972
7973 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7974 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7975 find_zone_movable_pfns_for_nodes();
7976
7977 /* Print out the zone ranges */
7978 pr_info("Zone ranges:\n");
7979 for (i = 0; i < MAX_NR_ZONES; i++) {
7980 if (i == ZONE_MOVABLE)
7981 continue;
7982 pr_info(" %-8s ", zone_names[i]);
7983 if (arch_zone_lowest_possible_pfn[i] ==
7984 arch_zone_highest_possible_pfn[i])
7985 pr_cont("empty\n");
7986 else
7987 pr_cont("[mem %#018Lx-%#018Lx]\n",
7988 (u64)arch_zone_lowest_possible_pfn[i]
7989 << PAGE_SHIFT,
7990 ((u64)arch_zone_highest_possible_pfn[i]
7991 << PAGE_SHIFT) - 1);
7992 }
7993
7994 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7995 pr_info("Movable zone start for each node\n");
7996 for (i = 0; i < MAX_NUMNODES; i++) {
7997 if (zone_movable_pfn[i])
7998 pr_info(" Node %d: %#018Lx\n", i,
7999 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
8000 }
8001
8002 /*
8003 * Print out the early node map, and initialize the
8004 * subsection-map relative to active online memory ranges to
8005 * enable future "sub-section" extensions of the memory map.
8006 */
8007 pr_info("Early memory node ranges\n");
8008 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
8009 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
8010 (u64)start_pfn << PAGE_SHIFT,
8011 ((u64)end_pfn << PAGE_SHIFT) - 1);
8012 subsection_map_init(start_pfn, end_pfn - start_pfn);
8013 }
8014
8015 /* Initialise every node */
8016 mminit_verify_pageflags_layout();
8017 setup_nr_node_ids();
8018 for_each_online_node(nid) {
8019 pg_data_t *pgdat = NODE_DATA(nid);
8020 free_area_init_node(nid);
8021
8022 /* Any memory on that node */
8023 if (pgdat->node_present_pages)
8024 node_set_state(nid, N_MEMORY);
8025 check_for_memory(pgdat, nid);
8026 }
8027
8028 memmap_init();
8029}
8030
8031static int __init cmdline_parse_core(char *p, unsigned long *core,
8032 unsigned long *percent)
8033{
8034 unsigned long long coremem;
8035 char *endptr;
8036
8037 if (!p)
8038 return -EINVAL;
8039
8040 /* Value may be a percentage of total memory, otherwise bytes */
8041 coremem = simple_strtoull(p, &endptr, 0);
8042 if (*endptr == '%') {
8043 /* Paranoid check for percent values greater than 100 */
8044 WARN_ON(coremem > 100);
8045
8046 *percent = coremem;
8047 } else {
8048 coremem = memparse(p, &p);
8049 /* Paranoid check that UL is enough for the coremem value */
8050 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
8051
8052 *core = coremem >> PAGE_SHIFT;
8053 *percent = 0UL;
8054 }
8055 return 0;
8056}
8057
8058/*
8059 * kernelcore=size sets the amount of memory for use for allocations that
8060 * cannot be reclaimed or migrated.
8061 */
8062static int __init cmdline_parse_kernelcore(char *p)
8063{
8064 /* parse kernelcore=mirror */
8065 if (parse_option_str(p, "mirror")) {
8066 mirrored_kernelcore = true;
8067 return 0;
8068 }
8069
8070 return cmdline_parse_core(p, &required_kernelcore,
8071 &required_kernelcore_percent);
8072}
8073
8074/*
8075 * movablecore=size sets the amount of memory for use for allocations that
8076 * can be reclaimed or migrated.
8077 */
8078static int __init cmdline_parse_movablecore(char *p)
8079{
8080 return cmdline_parse_core(p, &required_movablecore,
8081 &required_movablecore_percent);
8082}
8083
8084early_param("kernelcore", cmdline_parse_kernelcore);
8085early_param("movablecore", cmdline_parse_movablecore);
8086
8087void adjust_managed_page_count(struct page *page, long count)
8088{
8089 atomic_long_add(count, &page_zone(page)->managed_pages);
8090 totalram_pages_add(count);
8091#ifdef CONFIG_HIGHMEM
8092 if (PageHighMem(page))
8093 totalhigh_pages_add(count);
8094#endif
8095}
8096EXPORT_SYMBOL(adjust_managed_page_count);
8097
8098unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
8099{
8100 void *pos;
8101 unsigned long pages = 0;
8102
8103 start = (void *)PAGE_ALIGN((unsigned long)start);
8104 end = (void *)((unsigned long)end & PAGE_MASK);
8105 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
8106 struct page *page = virt_to_page(pos);
8107 void *direct_map_addr;
8108
8109 /*
8110 * 'direct_map_addr' might be different from 'pos'
8111 * because some architectures' virt_to_page()
8112 * work with aliases. Getting the direct map
8113 * address ensures that we get a _writeable_
8114 * alias for the memset().
8115 */
8116 direct_map_addr = page_address(page);
8117 /*
8118 * Perform a kasan-unchecked memset() since this memory
8119 * has not been initialized.
8120 */
8121 direct_map_addr = kasan_reset_tag(direct_map_addr);
8122 if ((unsigned int)poison <= 0xFF)
8123 memset(direct_map_addr, poison, PAGE_SIZE);
8124
8125 free_reserved_page(page);
8126 }
8127
8128 if (pages && s)
8129 pr_info("Freeing %s memory: %ldK\n",
8130 s, pages << (PAGE_SHIFT - 10));
8131
8132 return pages;
8133}
8134
8135void __init mem_init_print_info(void)
8136{
8137 unsigned long physpages, codesize, datasize, rosize, bss_size;
8138 unsigned long init_code_size, init_data_size;
8139
8140 physpages = get_num_physpages();
8141 codesize = _etext - _stext;
8142 datasize = _edata - _sdata;
8143 rosize = __end_rodata - __start_rodata;
8144 bss_size = __bss_stop - __bss_start;
8145 init_data_size = __init_end - __init_begin;
8146 init_code_size = _einittext - _sinittext;
8147
8148 /*
8149 * Detect special cases and adjust section sizes accordingly:
8150 * 1) .init.* may be embedded into .data sections
8151 * 2) .init.text.* may be out of [__init_begin, __init_end],
8152 * please refer to arch/tile/kernel/vmlinux.lds.S.
8153 * 3) .rodata.* may be embedded into .text or .data sections.
8154 */
8155#define adj_init_size(start, end, size, pos, adj) \
8156 do { \
8157 if (start <= pos && pos < end && size > adj) \
8158 size -= adj; \
8159 } while (0)
8160
8161 adj_init_size(__init_begin, __init_end, init_data_size,
8162 _sinittext, init_code_size);
8163 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
8164 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
8165 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
8166 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
8167
8168#undef adj_init_size
8169
8170 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
8171#ifdef CONFIG_HIGHMEM
8172 ", %luK highmem"
8173#endif
8174 ")\n",
8175 nr_free_pages() << (PAGE_SHIFT - 10),
8176 physpages << (PAGE_SHIFT - 10),
8177 codesize >> 10, datasize >> 10, rosize >> 10,
8178 (init_data_size + init_code_size) >> 10, bss_size >> 10,
8179 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
8180 totalcma_pages << (PAGE_SHIFT - 10)
8181#ifdef CONFIG_HIGHMEM
8182 , totalhigh_pages() << (PAGE_SHIFT - 10)
8183#endif
8184 );
8185}
8186
8187/**
8188 * set_dma_reserve - set the specified number of pages reserved in the first zone
8189 * @new_dma_reserve: The number of pages to mark reserved
8190 *
8191 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
8192 * In the DMA zone, a significant percentage may be consumed by kernel image
8193 * and other unfreeable allocations which can skew the watermarks badly. This
8194 * function may optionally be used to account for unfreeable pages in the
8195 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
8196 * smaller per-cpu batchsize.
8197 */
8198void __init set_dma_reserve(unsigned long new_dma_reserve)
8199{
8200 dma_reserve = new_dma_reserve;
8201}
8202
8203static int page_alloc_cpu_dead(unsigned int cpu)
8204{
8205 struct zone *zone;
8206
8207 lru_add_drain_cpu(cpu);
8208 drain_pages(cpu);
8209
8210 /*
8211 * Spill the event counters of the dead processor
8212 * into the current processors event counters.
8213 * This artificially elevates the count of the current
8214 * processor.
8215 */
8216 vm_events_fold_cpu(cpu);
8217
8218 /*
8219 * Zero the differential counters of the dead processor
8220 * so that the vm statistics are consistent.
8221 *
8222 * This is only okay since the processor is dead and cannot
8223 * race with what we are doing.
8224 */
8225 cpu_vm_stats_fold(cpu);
8226
8227 for_each_populated_zone(zone)
8228 zone_pcp_update(zone, 0);
8229
8230 return 0;
8231}
8232
8233static int page_alloc_cpu_online(unsigned int cpu)
8234{
8235 struct zone *zone;
8236
8237 for_each_populated_zone(zone)
8238 zone_pcp_update(zone, 1);
8239 return 0;
8240}
8241
8242#ifdef CONFIG_NUMA
8243int hashdist = HASHDIST_DEFAULT;
8244
8245static int __init set_hashdist(char *str)
8246{
8247 if (!str)
8248 return 0;
8249 hashdist = simple_strtoul(str, &str, 0);
8250 return 1;
8251}
8252__setup("hashdist=", set_hashdist);
8253#endif
8254
8255void __init page_alloc_init(void)
8256{
8257 int ret;
8258
8259#ifdef CONFIG_NUMA
8260 if (num_node_state(N_MEMORY) == 1)
8261 hashdist = 0;
8262#endif
8263
8264 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
8265 "mm/page_alloc:pcp",
8266 page_alloc_cpu_online,
8267 page_alloc_cpu_dead);
8268 WARN_ON(ret < 0);
8269}
8270
8271/*
8272 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
8273 * or min_free_kbytes changes.
8274 */
8275static void calculate_totalreserve_pages(void)
8276{
8277 struct pglist_data *pgdat;
8278 unsigned long reserve_pages = 0;
8279 enum zone_type i, j;
8280
8281 for_each_online_pgdat(pgdat) {
8282
8283 pgdat->totalreserve_pages = 0;
8284
8285 for (i = 0; i < MAX_NR_ZONES; i++) {
8286 struct zone *zone = pgdat->node_zones + i;
8287 long max = 0;
8288 unsigned long managed_pages = zone_managed_pages(zone);
8289
8290 /* Find valid and maximum lowmem_reserve in the zone */
8291 for (j = i; j < MAX_NR_ZONES; j++) {
8292 if (zone->lowmem_reserve[j] > max)
8293 max = zone->lowmem_reserve[j];
8294 }
8295
8296 /* we treat the high watermark as reserved pages. */
8297 max += high_wmark_pages(zone);
8298
8299 if (max > managed_pages)
8300 max = managed_pages;
8301
8302 pgdat->totalreserve_pages += max;
8303
8304 reserve_pages += max;
8305 }
8306 }
8307 totalreserve_pages = reserve_pages;
8308}
8309
8310/*
8311 * setup_per_zone_lowmem_reserve - called whenever
8312 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
8313 * has a correct pages reserved value, so an adequate number of
8314 * pages are left in the zone after a successful __alloc_pages().
8315 */
8316static void setup_per_zone_lowmem_reserve(void)
8317{
8318 struct pglist_data *pgdat;
8319 enum zone_type i, j;
8320
8321 for_each_online_pgdat(pgdat) {
8322 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
8323 struct zone *zone = &pgdat->node_zones[i];
8324 int ratio = sysctl_lowmem_reserve_ratio[i];
8325 bool clear = !ratio || !zone_managed_pages(zone);
8326 unsigned long managed_pages = 0;
8327
8328 for (j = i + 1; j < MAX_NR_ZONES; j++) {
8329 struct zone *upper_zone = &pgdat->node_zones[j];
8330
8331 managed_pages += zone_managed_pages(upper_zone);
8332
8333 if (clear)
8334 zone->lowmem_reserve[j] = 0;
8335 else
8336 zone->lowmem_reserve[j] = managed_pages / ratio;
8337 }
8338 }
8339 }
8340
8341 /* update totalreserve_pages */
8342 calculate_totalreserve_pages();
8343}
8344
8345static void __setup_per_zone_wmarks(void)
8346{
8347 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
8348 unsigned long lowmem_pages = 0;
8349 struct zone *zone;
8350 unsigned long flags;
8351
8352 /* Calculate total number of !ZONE_HIGHMEM pages */
8353 for_each_zone(zone) {
8354 if (!is_highmem(zone))
8355 lowmem_pages += zone_managed_pages(zone);
8356 }
8357
8358 for_each_zone(zone) {
8359 u64 tmp;
8360
8361 spin_lock_irqsave(&zone->lock, flags);
8362 tmp = (u64)pages_min * zone_managed_pages(zone);
8363 do_div(tmp, lowmem_pages);
8364 if (is_highmem(zone)) {
8365 /*
8366 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
8367 * need highmem pages, so cap pages_min to a small
8368 * value here.
8369 *
8370 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
8371 * deltas control async page reclaim, and so should
8372 * not be capped for highmem.
8373 */
8374 unsigned long min_pages;
8375
8376 min_pages = zone_managed_pages(zone) / 1024;
8377 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
8378 zone->_watermark[WMARK_MIN] = min_pages;
8379 } else {
8380 /*
8381 * If it's a lowmem zone, reserve a number of pages
8382 * proportionate to the zone's size.
8383 */
8384 zone->_watermark[WMARK_MIN] = tmp;
8385 }
8386
8387 /*
8388 * Set the kswapd watermarks distance according to the
8389 * scale factor in proportion to available memory, but
8390 * ensure a minimum size on small systems.
8391 */
8392 tmp = max_t(u64, tmp >> 2,
8393 mult_frac(zone_managed_pages(zone),
8394 watermark_scale_factor, 10000));
8395
8396 zone->watermark_boost = 0;
8397 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
8398 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
8399
8400 spin_unlock_irqrestore(&zone->lock, flags);
8401 }
8402
8403 /* update totalreserve_pages */
8404 calculate_totalreserve_pages();
8405}
8406
8407/**
8408 * setup_per_zone_wmarks - called when min_free_kbytes changes
8409 * or when memory is hot-{added|removed}
8410 *
8411 * Ensures that the watermark[min,low,high] values for each zone are set
8412 * correctly with respect to min_free_kbytes.
8413 */
8414void setup_per_zone_wmarks(void)
8415{
8416 struct zone *zone;
8417 static DEFINE_SPINLOCK(lock);
8418
8419 spin_lock(&lock);
8420 __setup_per_zone_wmarks();
8421 spin_unlock(&lock);
8422
8423 /*
8424 * The watermark size have changed so update the pcpu batch
8425 * and high limits or the limits may be inappropriate.
8426 */
8427 for_each_zone(zone)
8428 zone_pcp_update(zone, 0);
8429}
8430
8431/*
8432 * Initialise min_free_kbytes.
8433 *
8434 * For small machines we want it small (128k min). For large machines
8435 * we want it large (256MB max). But it is not linear, because network
8436 * bandwidth does not increase linearly with machine size. We use
8437 *
8438 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8439 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8440 *
8441 * which yields
8442 *
8443 * 16MB: 512k
8444 * 32MB: 724k
8445 * 64MB: 1024k
8446 * 128MB: 1448k
8447 * 256MB: 2048k
8448 * 512MB: 2896k
8449 * 1024MB: 4096k
8450 * 2048MB: 5792k
8451 * 4096MB: 8192k
8452 * 8192MB: 11584k
8453 * 16384MB: 16384k
8454 */
8455int __meminit init_per_zone_wmark_min(void)
8456{
8457 unsigned long lowmem_kbytes;
8458 int new_min_free_kbytes;
8459
8460 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8461 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8462
8463 if (new_min_free_kbytes > user_min_free_kbytes) {
8464 min_free_kbytes = new_min_free_kbytes;
8465 if (min_free_kbytes < 128)
8466 min_free_kbytes = 128;
8467 if (min_free_kbytes > 262144)
8468 min_free_kbytes = 262144;
8469 } else {
8470 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8471 new_min_free_kbytes, user_min_free_kbytes);
8472 }
8473 setup_per_zone_wmarks();
8474 refresh_zone_stat_thresholds();
8475 setup_per_zone_lowmem_reserve();
8476
8477#ifdef CONFIG_NUMA
8478 setup_min_unmapped_ratio();
8479 setup_min_slab_ratio();
8480#endif
8481
8482 khugepaged_min_free_kbytes_update();
8483
8484 return 0;
8485}
8486postcore_initcall(init_per_zone_wmark_min)
8487
8488/*
8489 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8490 * that we can call two helper functions whenever min_free_kbytes
8491 * changes.
8492 */
8493int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8494 void *buffer, size_t *length, loff_t *ppos)
8495{
8496 int rc;
8497
8498 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8499 if (rc)
8500 return rc;
8501
8502 if (write) {
8503 user_min_free_kbytes = min_free_kbytes;
8504 setup_per_zone_wmarks();
8505 }
8506 return 0;
8507}
8508
8509int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8510 void *buffer, size_t *length, loff_t *ppos)
8511{
8512 int rc;
8513
8514 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8515 if (rc)
8516 return rc;
8517
8518 if (write)
8519 setup_per_zone_wmarks();
8520
8521 return 0;
8522}
8523
8524#ifdef CONFIG_NUMA
8525static void setup_min_unmapped_ratio(void)
8526{
8527 pg_data_t *pgdat;
8528 struct zone *zone;
8529
8530 for_each_online_pgdat(pgdat)
8531 pgdat->min_unmapped_pages = 0;
8532
8533 for_each_zone(zone)
8534 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8535 sysctl_min_unmapped_ratio) / 100;
8536}
8537
8538
8539int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8540 void *buffer, size_t *length, loff_t *ppos)
8541{
8542 int rc;
8543
8544 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8545 if (rc)
8546 return rc;
8547
8548 setup_min_unmapped_ratio();
8549
8550 return 0;
8551}
8552
8553static void setup_min_slab_ratio(void)
8554{
8555 pg_data_t *pgdat;
8556 struct zone *zone;
8557
8558 for_each_online_pgdat(pgdat)
8559 pgdat->min_slab_pages = 0;
8560
8561 for_each_zone(zone)
8562 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8563 sysctl_min_slab_ratio) / 100;
8564}
8565
8566int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8567 void *buffer, size_t *length, loff_t *ppos)
8568{
8569 int rc;
8570
8571 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8572 if (rc)
8573 return rc;
8574
8575 setup_min_slab_ratio();
8576
8577 return 0;
8578}
8579#endif
8580
8581/*
8582 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8583 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8584 * whenever sysctl_lowmem_reserve_ratio changes.
8585 *
8586 * The reserve ratio obviously has absolutely no relation with the
8587 * minimum watermarks. The lowmem reserve ratio can only make sense
8588 * if in function of the boot time zone sizes.
8589 */
8590int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8591 void *buffer, size_t *length, loff_t *ppos)
8592{
8593 int i;
8594
8595 proc_dointvec_minmax(table, write, buffer, length, ppos);
8596
8597 for (i = 0; i < MAX_NR_ZONES; i++) {
8598 if (sysctl_lowmem_reserve_ratio[i] < 1)
8599 sysctl_lowmem_reserve_ratio[i] = 0;
8600 }
8601
8602 setup_per_zone_lowmem_reserve();
8603 return 0;
8604}
8605
8606/*
8607 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
8608 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8609 * pagelist can have before it gets flushed back to buddy allocator.
8610 */
8611int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
8612 int write, void *buffer, size_t *length, loff_t *ppos)
8613{
8614 struct zone *zone;
8615 int old_percpu_pagelist_high_fraction;
8616 int ret;
8617
8618 mutex_lock(&pcp_batch_high_lock);
8619 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
8620
8621 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8622 if (!write || ret < 0)
8623 goto out;
8624
8625 /* Sanity checking to avoid pcp imbalance */
8626 if (percpu_pagelist_high_fraction &&
8627 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
8628 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
8629 ret = -EINVAL;
8630 goto out;
8631 }
8632
8633 /* No change? */
8634 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
8635 goto out;
8636
8637 for_each_populated_zone(zone)
8638 zone_set_pageset_high_and_batch(zone, 0);
8639out:
8640 mutex_unlock(&pcp_batch_high_lock);
8641 return ret;
8642}
8643
8644#ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8645/*
8646 * Returns the number of pages that arch has reserved but
8647 * is not known to alloc_large_system_hash().
8648 */
8649static unsigned long __init arch_reserved_kernel_pages(void)
8650{
8651 return 0;
8652}
8653#endif
8654
8655/*
8656 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8657 * machines. As memory size is increased the scale is also increased but at
8658 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8659 * quadruples the scale is increased by one, which means the size of hash table
8660 * only doubles, instead of quadrupling as well.
8661 * Because 32-bit systems cannot have large physical memory, where this scaling
8662 * makes sense, it is disabled on such platforms.
8663 */
8664#if __BITS_PER_LONG > 32
8665#define ADAPT_SCALE_BASE (64ul << 30)
8666#define ADAPT_SCALE_SHIFT 2
8667#define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8668#endif
8669
8670/*
8671 * allocate a large system hash table from bootmem
8672 * - it is assumed that the hash table must contain an exact power-of-2
8673 * quantity of entries
8674 * - limit is the number of hash buckets, not the total allocation size
8675 */
8676void *__init alloc_large_system_hash(const char *tablename,
8677 unsigned long bucketsize,
8678 unsigned long numentries,
8679 int scale,
8680 int flags,
8681 unsigned int *_hash_shift,
8682 unsigned int *_hash_mask,
8683 unsigned long low_limit,
8684 unsigned long high_limit)
8685{
8686 unsigned long long max = high_limit;
8687 unsigned long log2qty, size;
8688 void *table = NULL;
8689 gfp_t gfp_flags;
8690 bool virt;
8691 bool huge;
8692
8693 /* allow the kernel cmdline to have a say */
8694 if (!numentries) {
8695 /* round applicable memory size up to nearest megabyte */
8696 numentries = nr_kernel_pages;
8697 numentries -= arch_reserved_kernel_pages();
8698
8699 /* It isn't necessary when PAGE_SIZE >= 1MB */
8700 if (PAGE_SHIFT < 20)
8701 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8702
8703#if __BITS_PER_LONG > 32
8704 if (!high_limit) {
8705 unsigned long adapt;
8706
8707 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8708 adapt <<= ADAPT_SCALE_SHIFT)
8709 scale++;
8710 }
8711#endif
8712
8713 /* limit to 1 bucket per 2^scale bytes of low memory */
8714 if (scale > PAGE_SHIFT)
8715 numentries >>= (scale - PAGE_SHIFT);
8716 else
8717 numentries <<= (PAGE_SHIFT - scale);
8718
8719 /* Make sure we've got at least a 0-order allocation.. */
8720 if (unlikely(flags & HASH_SMALL)) {
8721 /* Makes no sense without HASH_EARLY */
8722 WARN_ON(!(flags & HASH_EARLY));
8723 if (!(numentries >> *_hash_shift)) {
8724 numentries = 1UL << *_hash_shift;
8725 BUG_ON(!numentries);
8726 }
8727 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8728 numentries = PAGE_SIZE / bucketsize;
8729 }
8730 numentries = roundup_pow_of_two(numentries);
8731
8732 /* limit allocation size to 1/16 total memory by default */
8733 if (max == 0) {
8734 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8735 do_div(max, bucketsize);
8736 }
8737 max = min(max, 0x80000000ULL);
8738
8739 if (numentries < low_limit)
8740 numentries = low_limit;
8741 if (numentries > max)
8742 numentries = max;
8743
8744 log2qty = ilog2(numentries);
8745
8746 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8747 do {
8748 virt = false;
8749 size = bucketsize << log2qty;
8750 if (flags & HASH_EARLY) {
8751 if (flags & HASH_ZERO)
8752 table = memblock_alloc(size, SMP_CACHE_BYTES);
8753 else
8754 table = memblock_alloc_raw(size,
8755 SMP_CACHE_BYTES);
8756 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8757 table = __vmalloc(size, gfp_flags);
8758 virt = true;
8759 huge = is_vm_area_hugepages(table);
8760 } else {
8761 /*
8762 * If bucketsize is not a power-of-two, we may free
8763 * some pages at the end of hash table which
8764 * alloc_pages_exact() automatically does
8765 */
8766 table = alloc_pages_exact(size, gfp_flags);
8767 kmemleak_alloc(table, size, 1, gfp_flags);
8768 }
8769 } while (!table && size > PAGE_SIZE && --log2qty);
8770
8771 if (!table)
8772 panic("Failed to allocate %s hash table\n", tablename);
8773
8774 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8775 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8776 virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
8777
8778 if (_hash_shift)
8779 *_hash_shift = log2qty;
8780 if (_hash_mask)
8781 *_hash_mask = (1 << log2qty) - 1;
8782
8783 return table;
8784}
8785
8786/*
8787 * This function checks whether pageblock includes unmovable pages or not.
8788 *
8789 * PageLRU check without isolation or lru_lock could race so that
8790 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8791 * check without lock_page also may miss some movable non-lru pages at
8792 * race condition. So you can't expect this function should be exact.
8793 *
8794 * Returns a page without holding a reference. If the caller wants to
8795 * dereference that page (e.g., dumping), it has to make sure that it
8796 * cannot get removed (e.g., via memory unplug) concurrently.
8797 *
8798 */
8799struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8800 int migratetype, int flags)
8801{
8802 unsigned long iter = 0;
8803 unsigned long pfn = page_to_pfn(page);
8804 unsigned long offset = pfn % pageblock_nr_pages;
8805
8806 if (is_migrate_cma_page(page)) {
8807 /*
8808 * CMA allocations (alloc_contig_range) really need to mark
8809 * isolate CMA pageblocks even when they are not movable in fact
8810 * so consider them movable here.
8811 */
8812 if (is_migrate_cma(migratetype))
8813 return NULL;
8814
8815 return page;
8816 }
8817
8818 for (; iter < pageblock_nr_pages - offset; iter++) {
8819 if (!pfn_valid_within(pfn + iter))
8820 continue;
8821
8822 page = pfn_to_page(pfn + iter);
8823
8824 /*
8825 * Both, bootmem allocations and memory holes are marked
8826 * PG_reserved and are unmovable. We can even have unmovable
8827 * allocations inside ZONE_MOVABLE, for example when
8828 * specifying "movablecore".
8829 */
8830 if (PageReserved(page))
8831 return page;
8832
8833 /*
8834 * If the zone is movable and we have ruled out all reserved
8835 * pages then it should be reasonably safe to assume the rest
8836 * is movable.
8837 */
8838 if (zone_idx(zone) == ZONE_MOVABLE)
8839 continue;
8840
8841 /*
8842 * Hugepages are not in LRU lists, but they're movable.
8843 * THPs are on the LRU, but need to be counted as #small pages.
8844 * We need not scan over tail pages because we don't
8845 * handle each tail page individually in migration.
8846 */
8847 if (PageHuge(page) || PageTransCompound(page)) {
8848 struct page *head = compound_head(page);
8849 unsigned int skip_pages;
8850
8851 if (PageHuge(page)) {
8852 if (!hugepage_migration_supported(page_hstate(head)))
8853 return page;
8854 } else if (!PageLRU(head) && !__PageMovable(head)) {
8855 return page;
8856 }
8857
8858 skip_pages = compound_nr(head) - (page - head);
8859 iter += skip_pages - 1;
8860 continue;
8861 }
8862
8863 /*
8864 * We can't use page_count without pin a page
8865 * because another CPU can free compound page.
8866 * This check already skips compound tails of THP
8867 * because their page->_refcount is zero at all time.
8868 */
8869 if (!page_ref_count(page)) {
8870 if (PageBuddy(page))
8871 iter += (1 << buddy_order(page)) - 1;
8872 continue;
8873 }
8874
8875 /*
8876 * The HWPoisoned page may be not in buddy system, and
8877 * page_count() is not 0.
8878 */
8879 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8880 continue;
8881
8882 /*
8883 * We treat all PageOffline() pages as movable when offlining
8884 * to give drivers a chance to decrement their reference count
8885 * in MEM_GOING_OFFLINE in order to indicate that these pages
8886 * can be offlined as there are no direct references anymore.
8887 * For actually unmovable PageOffline() where the driver does
8888 * not support this, we will fail later when trying to actually
8889 * move these pages that still have a reference count > 0.
8890 * (false negatives in this function only)
8891 */
8892 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8893 continue;
8894
8895 if (__PageMovable(page) || PageLRU(page))
8896 continue;
8897
8898 /*
8899 * If there are RECLAIMABLE pages, we need to check
8900 * it. But now, memory offline itself doesn't call
8901 * shrink_node_slabs() and it still to be fixed.
8902 */
8903 return page;
8904 }
8905 return NULL;
8906}
8907
8908#ifdef CONFIG_CONTIG_ALLOC
8909static unsigned long pfn_max_align_down(unsigned long pfn)
8910{
8911 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8912 pageblock_nr_pages) - 1);
8913}
8914
8915static unsigned long pfn_max_align_up(unsigned long pfn)
8916{
8917 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8918 pageblock_nr_pages));
8919}
8920
8921#if defined(CONFIG_DYNAMIC_DEBUG) || \
8922 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
8923/* Usage: See admin-guide/dynamic-debug-howto.rst */
8924static void alloc_contig_dump_pages(struct list_head *page_list)
8925{
8926 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
8927
8928 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
8929 struct page *page;
8930
8931 dump_stack();
8932 list_for_each_entry(page, page_list, lru)
8933 dump_page(page, "migration failure");
8934 }
8935}
8936#else
8937static inline void alloc_contig_dump_pages(struct list_head *page_list)
8938{
8939}
8940#endif
8941
8942/* [start, end) must belong to a single zone. */
8943static int __alloc_contig_migrate_range(struct compact_control *cc,
8944 unsigned long start, unsigned long end)
8945{
8946 /* This function is based on compact_zone() from compaction.c. */
8947 unsigned int nr_reclaimed;
8948 unsigned long pfn = start;
8949 unsigned int tries = 0;
8950 int ret = 0;
8951 struct migration_target_control mtc = {
8952 .nid = zone_to_nid(cc->zone),
8953 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8954 };
8955
8956 lru_cache_disable();
8957
8958 while (pfn < end || !list_empty(&cc->migratepages)) {
8959 if (fatal_signal_pending(current)) {
8960 ret = -EINTR;
8961 break;
8962 }
8963
8964 if (list_empty(&cc->migratepages)) {
8965 cc->nr_migratepages = 0;
8966 ret = isolate_migratepages_range(cc, pfn, end);
8967 if (ret && ret != -EAGAIN)
8968 break;
8969 pfn = cc->migrate_pfn;
8970 tries = 0;
8971 } else if (++tries == 5) {
8972 ret = -EBUSY;
8973 break;
8974 }
8975
8976 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8977 &cc->migratepages);
8978 cc->nr_migratepages -= nr_reclaimed;
8979
8980 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
8981 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE);
8982
8983 /*
8984 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
8985 * to retry again over this error, so do the same here.
8986 */
8987 if (ret == -ENOMEM)
8988 break;
8989 }
8990
8991 lru_cache_enable();
8992 if (ret < 0) {
8993 if (ret == -EBUSY)
8994 alloc_contig_dump_pages(&cc->migratepages);
8995 putback_movable_pages(&cc->migratepages);
8996 return ret;
8997 }
8998 return 0;
8999}
9000
9001/**
9002 * alloc_contig_range() -- tries to allocate given range of pages
9003 * @start: start PFN to allocate
9004 * @end: one-past-the-last PFN to allocate
9005 * @migratetype: migratetype of the underlying pageblocks (either
9006 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
9007 * in range must have the same migratetype and it must
9008 * be either of the two.
9009 * @gfp_mask: GFP mask to use during compaction
9010 *
9011 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
9012 * aligned. The PFN range must belong to a single zone.
9013 *
9014 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
9015 * pageblocks in the range. Once isolated, the pageblocks should not
9016 * be modified by others.
9017 *
9018 * Return: zero on success or negative error code. On success all
9019 * pages which PFN is in [start, end) are allocated for the caller and
9020 * need to be freed with free_contig_range().
9021 */
9022int alloc_contig_range(unsigned long start, unsigned long end,
9023 unsigned migratetype, gfp_t gfp_mask)
9024{
9025 unsigned long outer_start, outer_end;
9026 unsigned int order;
9027 int ret = 0;
9028
9029 struct compact_control cc = {
9030 .nr_migratepages = 0,
9031 .order = -1,
9032 .zone = page_zone(pfn_to_page(start)),
9033 .mode = MIGRATE_SYNC,
9034 .ignore_skip_hint = true,
9035 .no_set_skip_hint = true,
9036 .gfp_mask = current_gfp_context(gfp_mask),
9037 .alloc_contig = true,
9038 };
9039 INIT_LIST_HEAD(&cc.migratepages);
9040
9041 /*
9042 * What we do here is we mark all pageblocks in range as
9043 * MIGRATE_ISOLATE. Because pageblock and max order pages may
9044 * have different sizes, and due to the way page allocator
9045 * work, we align the range to biggest of the two pages so
9046 * that page allocator won't try to merge buddies from
9047 * different pageblocks and change MIGRATE_ISOLATE to some
9048 * other migration type.
9049 *
9050 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
9051 * migrate the pages from an unaligned range (ie. pages that
9052 * we are interested in). This will put all the pages in
9053 * range back to page allocator as MIGRATE_ISOLATE.
9054 *
9055 * When this is done, we take the pages in range from page
9056 * allocator removing them from the buddy system. This way
9057 * page allocator will never consider using them.
9058 *
9059 * This lets us mark the pageblocks back as
9060 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
9061 * aligned range but not in the unaligned, original range are
9062 * put back to page allocator so that buddy can use them.
9063 */
9064
9065 ret = start_isolate_page_range(pfn_max_align_down(start),
9066 pfn_max_align_up(end), migratetype, 0);
9067 if (ret)
9068 return ret;
9069
9070 drain_all_pages(cc.zone);
9071
9072 /*
9073 * In case of -EBUSY, we'd like to know which page causes problem.
9074 * So, just fall through. test_pages_isolated() has a tracepoint
9075 * which will report the busy page.
9076 *
9077 * It is possible that busy pages could become available before
9078 * the call to test_pages_isolated, and the range will actually be
9079 * allocated. So, if we fall through be sure to clear ret so that
9080 * -EBUSY is not accidentally used or returned to caller.
9081 */
9082 ret = __alloc_contig_migrate_range(&cc, start, end);
9083 if (ret && ret != -EBUSY)
9084 goto done;
9085 ret = 0;
9086
9087 /*
9088 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
9089 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
9090 * more, all pages in [start, end) are free in page allocator.
9091 * What we are going to do is to allocate all pages from
9092 * [start, end) (that is remove them from page allocator).
9093 *
9094 * The only problem is that pages at the beginning and at the
9095 * end of interesting range may be not aligned with pages that
9096 * page allocator holds, ie. they can be part of higher order
9097 * pages. Because of this, we reserve the bigger range and
9098 * once this is done free the pages we are not interested in.
9099 *
9100 * We don't have to hold zone->lock here because the pages are
9101 * isolated thus they won't get removed from buddy.
9102 */
9103
9104 order = 0;
9105 outer_start = start;
9106 while (!PageBuddy(pfn_to_page(outer_start))) {
9107 if (++order >= MAX_ORDER) {
9108 outer_start = start;
9109 break;
9110 }
9111 outer_start &= ~0UL << order;
9112 }
9113
9114 if (outer_start != start) {
9115 order = buddy_order(pfn_to_page(outer_start));
9116
9117 /*
9118 * outer_start page could be small order buddy page and
9119 * it doesn't include start page. Adjust outer_start
9120 * in this case to report failed page properly
9121 * on tracepoint in test_pages_isolated()
9122 */
9123 if (outer_start + (1UL << order) <= start)
9124 outer_start = start;
9125 }
9126
9127 /* Make sure the range is really isolated. */
9128 if (test_pages_isolated(outer_start, end, 0)) {
9129 ret = -EBUSY;
9130 goto done;
9131 }
9132
9133 /* Grab isolated pages from freelists. */
9134 outer_end = isolate_freepages_range(&cc, outer_start, end);
9135 if (!outer_end) {
9136 ret = -EBUSY;
9137 goto done;
9138 }
9139
9140 /* Free head and tail (if any) */
9141 if (start != outer_start)
9142 free_contig_range(outer_start, start - outer_start);
9143 if (end != outer_end)
9144 free_contig_range(end, outer_end - end);
9145
9146done:
9147 undo_isolate_page_range(pfn_max_align_down(start),
9148 pfn_max_align_up(end), migratetype);
9149 return ret;
9150}
9151EXPORT_SYMBOL(alloc_contig_range);
9152
9153static int __alloc_contig_pages(unsigned long start_pfn,
9154 unsigned long nr_pages, gfp_t gfp_mask)
9155{
9156 unsigned long end_pfn = start_pfn + nr_pages;
9157
9158 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
9159 gfp_mask);
9160}
9161
9162static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
9163 unsigned long nr_pages)
9164{
9165 unsigned long i, end_pfn = start_pfn + nr_pages;
9166 struct page *page;
9167
9168 for (i = start_pfn; i < end_pfn; i++) {
9169 page = pfn_to_online_page(i);
9170 if (!page)
9171 return false;
9172
9173 if (page_zone(page) != z)
9174 return false;
9175
9176 if (PageReserved(page))
9177 return false;
9178 }
9179 return true;
9180}
9181
9182static bool zone_spans_last_pfn(const struct zone *zone,
9183 unsigned long start_pfn, unsigned long nr_pages)
9184{
9185 unsigned long last_pfn = start_pfn + nr_pages - 1;
9186
9187 return zone_spans_pfn(zone, last_pfn);
9188}
9189
9190/**
9191 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
9192 * @nr_pages: Number of contiguous pages to allocate
9193 * @gfp_mask: GFP mask to limit search and used during compaction
9194 * @nid: Target node
9195 * @nodemask: Mask for other possible nodes
9196 *
9197 * This routine is a wrapper around alloc_contig_range(). It scans over zones
9198 * on an applicable zonelist to find a contiguous pfn range which can then be
9199 * tried for allocation with alloc_contig_range(). This routine is intended
9200 * for allocation requests which can not be fulfilled with the buddy allocator.
9201 *
9202 * The allocated memory is always aligned to a page boundary. If nr_pages is a
9203 * power of two then the alignment is guaranteed to be to the given nr_pages
9204 * (e.g. 1GB request would be aligned to 1GB).
9205 *
9206 * Allocated pages can be freed with free_contig_range() or by manually calling
9207 * __free_page() on each allocated page.
9208 *
9209 * Return: pointer to contiguous pages on success, or NULL if not successful.
9210 */
9211struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
9212 int nid, nodemask_t *nodemask)
9213{
9214 unsigned long ret, pfn, flags;
9215 struct zonelist *zonelist;
9216 struct zone *zone;
9217 struct zoneref *z;
9218
9219 zonelist = node_zonelist(nid, gfp_mask);
9220 for_each_zone_zonelist_nodemask(zone, z, zonelist,
9221 gfp_zone(gfp_mask), nodemask) {
9222 spin_lock_irqsave(&zone->lock, flags);
9223
9224 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
9225 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
9226 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
9227 /*
9228 * We release the zone lock here because
9229 * alloc_contig_range() will also lock the zone
9230 * at some point. If there's an allocation
9231 * spinning on this lock, it may win the race
9232 * and cause alloc_contig_range() to fail...
9233 */
9234 spin_unlock_irqrestore(&zone->lock, flags);
9235 ret = __alloc_contig_pages(pfn, nr_pages,
9236 gfp_mask);
9237 if (!ret)
9238 return pfn_to_page(pfn);
9239 spin_lock_irqsave(&zone->lock, flags);
9240 }
9241 pfn += nr_pages;
9242 }
9243 spin_unlock_irqrestore(&zone->lock, flags);
9244 }
9245 return NULL;
9246}
9247#endif /* CONFIG_CONTIG_ALLOC */
9248
9249void free_contig_range(unsigned long pfn, unsigned long nr_pages)
9250{
9251 unsigned long count = 0;
9252
9253 for (; nr_pages--; pfn++) {
9254 struct page *page = pfn_to_page(pfn);
9255
9256 count += page_count(page) != 1;
9257 __free_page(page);
9258 }
9259 WARN(count != 0, "%lu pages are still in use!\n", count);
9260}
9261EXPORT_SYMBOL(free_contig_range);
9262
9263/*
9264 * The zone indicated has a new number of managed_pages; batch sizes and percpu
9265 * page high values need to be recalculated.
9266 */
9267void zone_pcp_update(struct zone *zone, int cpu_online)
9268{
9269 mutex_lock(&pcp_batch_high_lock);
9270 zone_set_pageset_high_and_batch(zone, cpu_online);
9271 mutex_unlock(&pcp_batch_high_lock);
9272}
9273
9274/*
9275 * Effectively disable pcplists for the zone by setting the high limit to 0
9276 * and draining all cpus. A concurrent page freeing on another CPU that's about
9277 * to put the page on pcplist will either finish before the drain and the page
9278 * will be drained, or observe the new high limit and skip the pcplist.
9279 *
9280 * Must be paired with a call to zone_pcp_enable().
9281 */
9282void zone_pcp_disable(struct zone *zone)
9283{
9284 mutex_lock(&pcp_batch_high_lock);
9285 __zone_set_pageset_high_and_batch(zone, 0, 1);
9286 __drain_all_pages(zone, true);
9287}
9288
9289void zone_pcp_enable(struct zone *zone)
9290{
9291 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
9292 mutex_unlock(&pcp_batch_high_lock);
9293}
9294
9295void zone_pcp_reset(struct zone *zone)
9296{
9297 int cpu;
9298 struct per_cpu_zonestat *pzstats;
9299
9300 if (zone->per_cpu_pageset != &boot_pageset) {
9301 for_each_online_cpu(cpu) {
9302 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
9303 drain_zonestat(zone, pzstats);
9304 }
9305 free_percpu(zone->per_cpu_pageset);
9306 free_percpu(zone->per_cpu_zonestats);
9307 zone->per_cpu_pageset = &boot_pageset;
9308 zone->per_cpu_zonestats = &boot_zonestats;
9309 }
9310}
9311
9312#ifdef CONFIG_MEMORY_HOTREMOVE
9313/*
9314 * All pages in the range must be in a single zone, must not contain holes,
9315 * must span full sections, and must be isolated before calling this function.
9316 */
9317void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
9318{
9319 unsigned long pfn = start_pfn;
9320 struct page *page;
9321 struct zone *zone;
9322 unsigned int order;
9323 unsigned long flags;
9324
9325 offline_mem_sections(pfn, end_pfn);
9326 zone = page_zone(pfn_to_page(pfn));
9327 spin_lock_irqsave(&zone->lock, flags);
9328 while (pfn < end_pfn) {
9329 page = pfn_to_page(pfn);
9330 /*
9331 * The HWPoisoned page may be not in buddy system, and
9332 * page_count() is not 0.
9333 */
9334 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
9335 pfn++;
9336 continue;
9337 }
9338 /*
9339 * At this point all remaining PageOffline() pages have a
9340 * reference count of 0 and can simply be skipped.
9341 */
9342 if (PageOffline(page)) {
9343 BUG_ON(page_count(page));
9344 BUG_ON(PageBuddy(page));
9345 pfn++;
9346 continue;
9347 }
9348
9349 BUG_ON(page_count(page));
9350 BUG_ON(!PageBuddy(page));
9351 order = buddy_order(page);
9352 del_page_from_free_list(page, zone, order);
9353 pfn += (1 << order);
9354 }
9355 spin_unlock_irqrestore(&zone->lock, flags);
9356}
9357#endif
9358
9359bool is_free_buddy_page(struct page *page)
9360{
9361 struct zone *zone = page_zone(page);
9362 unsigned long pfn = page_to_pfn(page);
9363 unsigned long flags;
9364 unsigned int order;
9365
9366 spin_lock_irqsave(&zone->lock, flags);
9367 for (order = 0; order < MAX_ORDER; order++) {
9368 struct page *page_head = page - (pfn & ((1 << order) - 1));
9369
9370 if (PageBuddy(page_head) && buddy_order(page_head) >= order)
9371 break;
9372 }
9373 spin_unlock_irqrestore(&zone->lock, flags);
9374
9375 return order < MAX_ORDER;
9376}
9377
9378#ifdef CONFIG_MEMORY_FAILURE
9379/*
9380 * Break down a higher-order page in sub-pages, and keep our target out of
9381 * buddy allocator.
9382 */
9383static void break_down_buddy_pages(struct zone *zone, struct page *page,
9384 struct page *target, int low, int high,
9385 int migratetype)
9386{
9387 unsigned long size = 1 << high;
9388 struct page *current_buddy, *next_page;
9389
9390 while (high > low) {
9391 high--;
9392 size >>= 1;
9393
9394 if (target >= &page[size]) {
9395 next_page = page + size;
9396 current_buddy = page;
9397 } else {
9398 next_page = page;
9399 current_buddy = page + size;
9400 }
9401
9402 if (set_page_guard(zone, current_buddy, high, migratetype))
9403 continue;
9404
9405 if (current_buddy != target) {
9406 add_to_free_list(current_buddy, zone, high, migratetype);
9407 set_buddy_order(current_buddy, high);
9408 page = next_page;
9409 }
9410 }
9411}
9412
9413/*
9414 * Take a page that will be marked as poisoned off the buddy allocator.
9415 */
9416bool take_page_off_buddy(struct page *page)
9417{
9418 struct zone *zone = page_zone(page);
9419 unsigned long pfn = page_to_pfn(page);
9420 unsigned long flags;
9421 unsigned int order;
9422 bool ret = false;
9423
9424 spin_lock_irqsave(&zone->lock, flags);
9425 for (order = 0; order < MAX_ORDER; order++) {
9426 struct page *page_head = page - (pfn & ((1 << order) - 1));
9427 int page_order = buddy_order(page_head);
9428
9429 if (PageBuddy(page_head) && page_order >= order) {
9430 unsigned long pfn_head = page_to_pfn(page_head);
9431 int migratetype = get_pfnblock_migratetype(page_head,
9432 pfn_head);
9433
9434 del_page_from_free_list(page_head, zone, page_order);
9435 break_down_buddy_pages(zone, page_head, page, 0,
9436 page_order, migratetype);
9437 if (!is_migrate_isolate(migratetype))
9438 __mod_zone_freepage_state(zone, -1, migratetype);
9439 ret = true;
9440 break;
9441 }
9442 if (page_count(page_head) > 0)
9443 break;
9444 }
9445 spin_unlock_irqrestore(&zone->lock, flags);
9446 return ret;
9447}
9448#endif