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/migrate.h>
61#include <linux/hugetlb.h>
62#include <linux/sched/rt.h>
63#include <linux/sched/mm.h>
64#include <linux/page_owner.h>
65#include <linux/kthread.h>
66#include <linux/memcontrol.h>
67#include <linux/ftrace.h>
68#include <linux/lockdep.h>
69#include <linux/nmi.h>
70#include <linux/psi.h>
71#include <linux/padata.h>
72#include <linux/khugepaged.h>
73
74#include <asm/sections.h>
75#include <asm/tlbflush.h>
76#include <asm/div64.h>
77#include "internal.h"
78#include "shuffle.h"
79#include "page_reporting.h"
80
81/* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
82static DEFINE_MUTEX(pcp_batch_high_lock);
83#define MIN_PERCPU_PAGELIST_FRACTION (8)
84
85#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
86DEFINE_PER_CPU(int, numa_node);
87EXPORT_PER_CPU_SYMBOL(numa_node);
88#endif
89
90DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
91
92#ifdef CONFIG_HAVE_MEMORYLESS_NODES
93/*
94 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
95 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
96 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
97 * defined in <linux/topology.h>.
98 */
99DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
100EXPORT_PER_CPU_SYMBOL(_numa_mem_);
101#endif
102
103/* work_structs for global per-cpu drains */
104struct pcpu_drain {
105 struct zone *zone;
106 struct work_struct work;
107};
108static DEFINE_MUTEX(pcpu_drain_mutex);
109static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
110
111#ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
112volatile unsigned long latent_entropy __latent_entropy;
113EXPORT_SYMBOL(latent_entropy);
114#endif
115
116/*
117 * Array of node states.
118 */
119nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
120 [N_POSSIBLE] = NODE_MASK_ALL,
121 [N_ONLINE] = { { [0] = 1UL } },
122#ifndef CONFIG_NUMA
123 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
124#ifdef CONFIG_HIGHMEM
125 [N_HIGH_MEMORY] = { { [0] = 1UL } },
126#endif
127 [N_MEMORY] = { { [0] = 1UL } },
128 [N_CPU] = { { [0] = 1UL } },
129#endif /* NUMA */
130};
131EXPORT_SYMBOL(node_states);
132
133atomic_long_t _totalram_pages __read_mostly;
134EXPORT_SYMBOL(_totalram_pages);
135unsigned long totalreserve_pages __read_mostly;
136unsigned long totalcma_pages __read_mostly;
137
138int percpu_pagelist_fraction;
139gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
140#ifdef CONFIG_INIT_ON_ALLOC_DEFAULT_ON
141DEFINE_STATIC_KEY_TRUE(init_on_alloc);
142#else
143DEFINE_STATIC_KEY_FALSE(init_on_alloc);
144#endif
145EXPORT_SYMBOL(init_on_alloc);
146
147#ifdef CONFIG_INIT_ON_FREE_DEFAULT_ON
148DEFINE_STATIC_KEY_TRUE(init_on_free);
149#else
150DEFINE_STATIC_KEY_FALSE(init_on_free);
151#endif
152EXPORT_SYMBOL(init_on_free);
153
154static int __init early_init_on_alloc(char *buf)
155{
156 int ret;
157 bool bool_result;
158
159 if (!buf)
160 return -EINVAL;
161 ret = kstrtobool(buf, &bool_result);
162 if (bool_result && page_poisoning_enabled())
163 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_alloc\n");
164 if (bool_result)
165 static_branch_enable(&init_on_alloc);
166 else
167 static_branch_disable(&init_on_alloc);
168 return ret;
169}
170early_param("init_on_alloc", early_init_on_alloc);
171
172static int __init early_init_on_free(char *buf)
173{
174 int ret;
175 bool bool_result;
176
177 if (!buf)
178 return -EINVAL;
179 ret = kstrtobool(buf, &bool_result);
180 if (bool_result && page_poisoning_enabled())
181 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_free\n");
182 if (bool_result)
183 static_branch_enable(&init_on_free);
184 else
185 static_branch_disable(&init_on_free);
186 return ret;
187}
188early_param("init_on_free", early_init_on_free);
189
190/*
191 * A cached value of the page's pageblock's migratetype, used when the page is
192 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
193 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
194 * Also the migratetype set in the page does not necessarily match the pcplist
195 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
196 * other index - this ensures that it will be put on the correct CMA freelist.
197 */
198static inline int get_pcppage_migratetype(struct page *page)
199{
200 return page->index;
201}
202
203static inline void set_pcppage_migratetype(struct page *page, int migratetype)
204{
205 page->index = migratetype;
206}
207
208#ifdef CONFIG_PM_SLEEP
209/*
210 * The following functions are used by the suspend/hibernate code to temporarily
211 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
212 * while devices are suspended. To avoid races with the suspend/hibernate code,
213 * they should always be called with system_transition_mutex held
214 * (gfp_allowed_mask also should only be modified with system_transition_mutex
215 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
216 * with that modification).
217 */
218
219static gfp_t saved_gfp_mask;
220
221void pm_restore_gfp_mask(void)
222{
223 WARN_ON(!mutex_is_locked(&system_transition_mutex));
224 if (saved_gfp_mask) {
225 gfp_allowed_mask = saved_gfp_mask;
226 saved_gfp_mask = 0;
227 }
228}
229
230void pm_restrict_gfp_mask(void)
231{
232 WARN_ON(!mutex_is_locked(&system_transition_mutex));
233 WARN_ON(saved_gfp_mask);
234 saved_gfp_mask = gfp_allowed_mask;
235 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
236}
237
238bool pm_suspended_storage(void)
239{
240 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
241 return false;
242 return true;
243}
244#endif /* CONFIG_PM_SLEEP */
245
246#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
247unsigned int pageblock_order __read_mostly;
248#endif
249
250static void __free_pages_ok(struct page *page, unsigned int order);
251
252/*
253 * results with 256, 32 in the lowmem_reserve sysctl:
254 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
255 * 1G machine -> (16M dma, 784M normal, 224M high)
256 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
257 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
258 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
259 *
260 * TBD: should special case ZONE_DMA32 machines here - in those we normally
261 * don't need any ZONE_NORMAL reservation
262 */
263int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
264#ifdef CONFIG_ZONE_DMA
265 [ZONE_DMA] = 256,
266#endif
267#ifdef CONFIG_ZONE_DMA32
268 [ZONE_DMA32] = 256,
269#endif
270 [ZONE_NORMAL] = 32,
271#ifdef CONFIG_HIGHMEM
272 [ZONE_HIGHMEM] = 0,
273#endif
274 [ZONE_MOVABLE] = 0,
275};
276
277static char * const zone_names[MAX_NR_ZONES] = {
278#ifdef CONFIG_ZONE_DMA
279 "DMA",
280#endif
281#ifdef CONFIG_ZONE_DMA32
282 "DMA32",
283#endif
284 "Normal",
285#ifdef CONFIG_HIGHMEM
286 "HighMem",
287#endif
288 "Movable",
289#ifdef CONFIG_ZONE_DEVICE
290 "Device",
291#endif
292};
293
294const char * const migratetype_names[MIGRATE_TYPES] = {
295 "Unmovable",
296 "Movable",
297 "Reclaimable",
298 "HighAtomic",
299#ifdef CONFIG_CMA
300 "CMA",
301#endif
302#ifdef CONFIG_MEMORY_ISOLATION
303 "Isolate",
304#endif
305};
306
307compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
308 [NULL_COMPOUND_DTOR] = NULL,
309 [COMPOUND_PAGE_DTOR] = free_compound_page,
310#ifdef CONFIG_HUGETLB_PAGE
311 [HUGETLB_PAGE_DTOR] = free_huge_page,
312#endif
313#ifdef CONFIG_TRANSPARENT_HUGEPAGE
314 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
315#endif
316};
317
318int min_free_kbytes = 1024;
319int user_min_free_kbytes = -1;
320#ifdef CONFIG_DISCONTIGMEM
321/*
322 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
323 * are not on separate NUMA nodes. Functionally this works but with
324 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
325 * quite small. By default, do not boost watermarks on discontigmem as in
326 * many cases very high-order allocations like THP are likely to be
327 * unsupported and the premature reclaim offsets the advantage of long-term
328 * fragmentation avoidance.
329 */
330int watermark_boost_factor __read_mostly;
331#else
332int watermark_boost_factor __read_mostly = 15000;
333#endif
334int watermark_scale_factor = 10;
335
336static unsigned long nr_kernel_pages __initdata;
337static unsigned long nr_all_pages __initdata;
338static unsigned long dma_reserve __initdata;
339
340static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
341static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
342static unsigned long required_kernelcore __initdata;
343static unsigned long required_kernelcore_percent __initdata;
344static unsigned long required_movablecore __initdata;
345static unsigned long required_movablecore_percent __initdata;
346static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
347static bool mirrored_kernelcore __meminitdata;
348
349/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
350int movable_zone;
351EXPORT_SYMBOL(movable_zone);
352
353#if MAX_NUMNODES > 1
354unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
355unsigned int nr_online_nodes __read_mostly = 1;
356EXPORT_SYMBOL(nr_node_ids);
357EXPORT_SYMBOL(nr_online_nodes);
358#endif
359
360int page_group_by_mobility_disabled __read_mostly;
361
362#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
363/*
364 * During boot we initialize deferred pages on-demand, as needed, but once
365 * page_alloc_init_late() has finished, the deferred pages are all initialized,
366 * and we can permanently disable that path.
367 */
368static DEFINE_STATIC_KEY_TRUE(deferred_pages);
369
370/*
371 * Calling kasan_free_pages() only after deferred memory initialization
372 * has completed. Poisoning pages during deferred memory init will greatly
373 * lengthen the process and cause problem in large memory systems as the
374 * deferred pages initialization is done with interrupt disabled.
375 *
376 * Assuming that there will be no reference to those newly initialized
377 * pages before they are ever allocated, this should have no effect on
378 * KASAN memory tracking as the poison will be properly inserted at page
379 * allocation time. The only corner case is when pages are allocated by
380 * on-demand allocation and then freed again before the deferred pages
381 * initialization is done, but this is not likely to happen.
382 */
383static inline void kasan_free_nondeferred_pages(struct page *page, int order)
384{
385 if (!static_branch_unlikely(&deferred_pages))
386 kasan_free_pages(page, order);
387}
388
389/* Returns true if the struct page for the pfn is uninitialised */
390static inline bool __meminit early_page_uninitialised(unsigned long pfn)
391{
392 int nid = early_pfn_to_nid(pfn);
393
394 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
395 return true;
396
397 return false;
398}
399
400/*
401 * Returns true when the remaining initialisation should be deferred until
402 * later in the boot cycle when it can be parallelised.
403 */
404static bool __meminit
405defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
406{
407 static unsigned long prev_end_pfn, nr_initialised;
408
409 /*
410 * prev_end_pfn static that contains the end of previous zone
411 * No need to protect because called very early in boot before smp_init.
412 */
413 if (prev_end_pfn != end_pfn) {
414 prev_end_pfn = end_pfn;
415 nr_initialised = 0;
416 }
417
418 /* Always populate low zones for address-constrained allocations */
419 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
420 return false;
421
422 /*
423 * We start only with one section of pages, more pages are added as
424 * needed until the rest of deferred pages are initialized.
425 */
426 nr_initialised++;
427 if ((nr_initialised > PAGES_PER_SECTION) &&
428 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
429 NODE_DATA(nid)->first_deferred_pfn = pfn;
430 return true;
431 }
432 return false;
433}
434#else
435#define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
436
437static inline bool early_page_uninitialised(unsigned long pfn)
438{
439 return false;
440}
441
442static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
443{
444 return false;
445}
446#endif
447
448/* Return a pointer to the bitmap storing bits affecting a block of pages */
449static inline unsigned long *get_pageblock_bitmap(struct page *page,
450 unsigned long pfn)
451{
452#ifdef CONFIG_SPARSEMEM
453 return section_to_usemap(__pfn_to_section(pfn));
454#else
455 return page_zone(page)->pageblock_flags;
456#endif /* CONFIG_SPARSEMEM */
457}
458
459static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
460{
461#ifdef CONFIG_SPARSEMEM
462 pfn &= (PAGES_PER_SECTION-1);
463#else
464 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
465#endif /* CONFIG_SPARSEMEM */
466 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
467}
468
469/**
470 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
471 * @page: The page within the block of interest
472 * @pfn: The target page frame number
473 * @mask: mask of bits that the caller is interested in
474 *
475 * Return: pageblock_bits flags
476 */
477static __always_inline
478unsigned long __get_pfnblock_flags_mask(struct page *page,
479 unsigned long pfn,
480 unsigned long mask)
481{
482 unsigned long *bitmap;
483 unsigned long bitidx, word_bitidx;
484 unsigned long word;
485
486 bitmap = get_pageblock_bitmap(page, pfn);
487 bitidx = pfn_to_bitidx(page, pfn);
488 word_bitidx = bitidx / BITS_PER_LONG;
489 bitidx &= (BITS_PER_LONG-1);
490
491 word = bitmap[word_bitidx];
492 return (word >> bitidx) & mask;
493}
494
495unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
496 unsigned long mask)
497{
498 return __get_pfnblock_flags_mask(page, pfn, mask);
499}
500
501static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
502{
503 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
504}
505
506/**
507 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
508 * @page: The page within the block of interest
509 * @flags: The flags to set
510 * @pfn: The target page frame number
511 * @mask: mask of bits that the caller is interested in
512 */
513void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
514 unsigned long pfn,
515 unsigned long mask)
516{
517 unsigned long *bitmap;
518 unsigned long bitidx, word_bitidx;
519 unsigned long old_word, word;
520
521 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
522 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
523
524 bitmap = get_pageblock_bitmap(page, pfn);
525 bitidx = pfn_to_bitidx(page, pfn);
526 word_bitidx = bitidx / BITS_PER_LONG;
527 bitidx &= (BITS_PER_LONG-1);
528
529 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
530
531 mask <<= bitidx;
532 flags <<= bitidx;
533
534 word = READ_ONCE(bitmap[word_bitidx]);
535 for (;;) {
536 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
537 if (word == old_word)
538 break;
539 word = old_word;
540 }
541}
542
543void set_pageblock_migratetype(struct page *page, int migratetype)
544{
545 if (unlikely(page_group_by_mobility_disabled &&
546 migratetype < MIGRATE_PCPTYPES))
547 migratetype = MIGRATE_UNMOVABLE;
548
549 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
550 page_to_pfn(page), MIGRATETYPE_MASK);
551}
552
553#ifdef CONFIG_DEBUG_VM
554static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
555{
556 int ret = 0;
557 unsigned seq;
558 unsigned long pfn = page_to_pfn(page);
559 unsigned long sp, start_pfn;
560
561 do {
562 seq = zone_span_seqbegin(zone);
563 start_pfn = zone->zone_start_pfn;
564 sp = zone->spanned_pages;
565 if (!zone_spans_pfn(zone, pfn))
566 ret = 1;
567 } while (zone_span_seqretry(zone, seq));
568
569 if (ret)
570 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
571 pfn, zone_to_nid(zone), zone->name,
572 start_pfn, start_pfn + sp);
573
574 return ret;
575}
576
577static int page_is_consistent(struct zone *zone, struct page *page)
578{
579 if (!pfn_valid_within(page_to_pfn(page)))
580 return 0;
581 if (zone != page_zone(page))
582 return 0;
583
584 return 1;
585}
586/*
587 * Temporary debugging check for pages not lying within a given zone.
588 */
589static int __maybe_unused bad_range(struct zone *zone, struct page *page)
590{
591 if (page_outside_zone_boundaries(zone, page))
592 return 1;
593 if (!page_is_consistent(zone, page))
594 return 1;
595
596 return 0;
597}
598#else
599static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
600{
601 return 0;
602}
603#endif
604
605static void bad_page(struct page *page, const char *reason)
606{
607 static unsigned long resume;
608 static unsigned long nr_shown;
609 static unsigned long nr_unshown;
610
611 /*
612 * Allow a burst of 60 reports, then keep quiet for that minute;
613 * or allow a steady drip of one report per second.
614 */
615 if (nr_shown == 60) {
616 if (time_before(jiffies, resume)) {
617 nr_unshown++;
618 goto out;
619 }
620 if (nr_unshown) {
621 pr_alert(
622 "BUG: Bad page state: %lu messages suppressed\n",
623 nr_unshown);
624 nr_unshown = 0;
625 }
626 nr_shown = 0;
627 }
628 if (nr_shown++ == 0)
629 resume = jiffies + 60 * HZ;
630
631 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
632 current->comm, page_to_pfn(page));
633 __dump_page(page, reason);
634 dump_page_owner(page);
635
636 print_modules();
637 dump_stack();
638out:
639 /* Leave bad fields for debug, except PageBuddy could make trouble */
640 page_mapcount_reset(page); /* remove PageBuddy */
641 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
642}
643
644/*
645 * Higher-order pages are called "compound pages". They are structured thusly:
646 *
647 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
648 *
649 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
650 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
651 *
652 * The first tail page's ->compound_dtor holds the offset in array of compound
653 * page destructors. See compound_page_dtors.
654 *
655 * The first tail page's ->compound_order holds the order of allocation.
656 * This usage means that zero-order pages may not be compound.
657 */
658
659void free_compound_page(struct page *page)
660{
661 mem_cgroup_uncharge(page);
662 __free_pages_ok(page, compound_order(page));
663}
664
665void prep_compound_page(struct page *page, unsigned int order)
666{
667 int i;
668 int nr_pages = 1 << order;
669
670 __SetPageHead(page);
671 for (i = 1; i < nr_pages; i++) {
672 struct page *p = page + i;
673 set_page_count(p, 0);
674 p->mapping = TAIL_MAPPING;
675 set_compound_head(p, page);
676 }
677
678 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
679 set_compound_order(page, order);
680 atomic_set(compound_mapcount_ptr(page), -1);
681 if (hpage_pincount_available(page))
682 atomic_set(compound_pincount_ptr(page), 0);
683}
684
685#ifdef CONFIG_DEBUG_PAGEALLOC
686unsigned int _debug_guardpage_minorder;
687
688bool _debug_pagealloc_enabled_early __read_mostly
689 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
690EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
691DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
692EXPORT_SYMBOL(_debug_pagealloc_enabled);
693
694DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
695
696static int __init early_debug_pagealloc(char *buf)
697{
698 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
699}
700early_param("debug_pagealloc", early_debug_pagealloc);
701
702void init_debug_pagealloc(void)
703{
704 if (!debug_pagealloc_enabled())
705 return;
706
707 static_branch_enable(&_debug_pagealloc_enabled);
708
709 if (!debug_guardpage_minorder())
710 return;
711
712 static_branch_enable(&_debug_guardpage_enabled);
713}
714
715static int __init debug_guardpage_minorder_setup(char *buf)
716{
717 unsigned long res;
718
719 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
720 pr_err("Bad debug_guardpage_minorder value\n");
721 return 0;
722 }
723 _debug_guardpage_minorder = res;
724 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
725 return 0;
726}
727early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
728
729static inline bool set_page_guard(struct zone *zone, struct page *page,
730 unsigned int order, int migratetype)
731{
732 if (!debug_guardpage_enabled())
733 return false;
734
735 if (order >= debug_guardpage_minorder())
736 return false;
737
738 __SetPageGuard(page);
739 INIT_LIST_HEAD(&page->lru);
740 set_page_private(page, order);
741 /* Guard pages are not available for any usage */
742 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
743
744 return true;
745}
746
747static inline void clear_page_guard(struct zone *zone, struct page *page,
748 unsigned int order, int migratetype)
749{
750 if (!debug_guardpage_enabled())
751 return;
752
753 __ClearPageGuard(page);
754
755 set_page_private(page, 0);
756 if (!is_migrate_isolate(migratetype))
757 __mod_zone_freepage_state(zone, (1 << order), migratetype);
758}
759#else
760static inline bool set_page_guard(struct zone *zone, struct page *page,
761 unsigned int order, int migratetype) { return false; }
762static inline void clear_page_guard(struct zone *zone, struct page *page,
763 unsigned int order, int migratetype) {}
764#endif
765
766static inline void set_page_order(struct page *page, unsigned int order)
767{
768 set_page_private(page, order);
769 __SetPageBuddy(page);
770}
771
772/*
773 * This function checks whether a page is free && is the buddy
774 * we can coalesce a page and its buddy if
775 * (a) the buddy is not in a hole (check before calling!) &&
776 * (b) the buddy is in the buddy system &&
777 * (c) a page and its buddy have the same order &&
778 * (d) a page and its buddy are in the same zone.
779 *
780 * For recording whether a page is in the buddy system, we set PageBuddy.
781 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
782 *
783 * For recording page's order, we use page_private(page).
784 */
785static inline bool page_is_buddy(struct page *page, struct page *buddy,
786 unsigned int order)
787{
788 if (!page_is_guard(buddy) && !PageBuddy(buddy))
789 return false;
790
791 if (page_order(buddy) != order)
792 return false;
793
794 /*
795 * zone check is done late to avoid uselessly calculating
796 * zone/node ids for pages that could never merge.
797 */
798 if (page_zone_id(page) != page_zone_id(buddy))
799 return false;
800
801 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
802
803 return true;
804}
805
806#ifdef CONFIG_COMPACTION
807static inline struct capture_control *task_capc(struct zone *zone)
808{
809 struct capture_control *capc = current->capture_control;
810
811 return unlikely(capc) &&
812 !(current->flags & PF_KTHREAD) &&
813 !capc->page &&
814 capc->cc->zone == zone ? capc : NULL;
815}
816
817static inline bool
818compaction_capture(struct capture_control *capc, struct page *page,
819 int order, int migratetype)
820{
821 if (!capc || order != capc->cc->order)
822 return false;
823
824 /* Do not accidentally pollute CMA or isolated regions*/
825 if (is_migrate_cma(migratetype) ||
826 is_migrate_isolate(migratetype))
827 return false;
828
829 /*
830 * Do not let lower order allocations polluate a movable pageblock.
831 * This might let an unmovable request use a reclaimable pageblock
832 * and vice-versa but no more than normal fallback logic which can
833 * have trouble finding a high-order free page.
834 */
835 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
836 return false;
837
838 capc->page = page;
839 return true;
840}
841
842#else
843static inline struct capture_control *task_capc(struct zone *zone)
844{
845 return NULL;
846}
847
848static inline bool
849compaction_capture(struct capture_control *capc, struct page *page,
850 int order, int migratetype)
851{
852 return false;
853}
854#endif /* CONFIG_COMPACTION */
855
856/* Used for pages not on another list */
857static inline void add_to_free_list(struct page *page, struct zone *zone,
858 unsigned int order, int migratetype)
859{
860 struct free_area *area = &zone->free_area[order];
861
862 list_add(&page->lru, &area->free_list[migratetype]);
863 area->nr_free++;
864}
865
866/* Used for pages not on another list */
867static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
868 unsigned int order, int migratetype)
869{
870 struct free_area *area = &zone->free_area[order];
871
872 list_add_tail(&page->lru, &area->free_list[migratetype]);
873 area->nr_free++;
874}
875
876/* Used for pages which are on another list */
877static inline void move_to_free_list(struct page *page, struct zone *zone,
878 unsigned int order, int migratetype)
879{
880 struct free_area *area = &zone->free_area[order];
881
882 list_move(&page->lru, &area->free_list[migratetype]);
883}
884
885static inline void del_page_from_free_list(struct page *page, struct zone *zone,
886 unsigned int order)
887{
888 /* clear reported state and update reported page count */
889 if (page_reported(page))
890 __ClearPageReported(page);
891
892 list_del(&page->lru);
893 __ClearPageBuddy(page);
894 set_page_private(page, 0);
895 zone->free_area[order].nr_free--;
896}
897
898/*
899 * If this is not the largest possible page, check if the buddy
900 * of the next-highest order is free. If it is, it's possible
901 * that pages are being freed that will coalesce soon. In case,
902 * that is happening, add the free page to the tail of the list
903 * so it's less likely to be used soon and more likely to be merged
904 * as a higher order page
905 */
906static inline bool
907buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
908 struct page *page, unsigned int order)
909{
910 struct page *higher_page, *higher_buddy;
911 unsigned long combined_pfn;
912
913 if (order >= MAX_ORDER - 2)
914 return false;
915
916 if (!pfn_valid_within(buddy_pfn))
917 return false;
918
919 combined_pfn = buddy_pfn & pfn;
920 higher_page = page + (combined_pfn - pfn);
921 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
922 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
923
924 return pfn_valid_within(buddy_pfn) &&
925 page_is_buddy(higher_page, higher_buddy, order + 1);
926}
927
928/*
929 * Freeing function for a buddy system allocator.
930 *
931 * The concept of a buddy system is to maintain direct-mapped table
932 * (containing bit values) for memory blocks of various "orders".
933 * The bottom level table contains the map for the smallest allocatable
934 * units of memory (here, pages), and each level above it describes
935 * pairs of units from the levels below, hence, "buddies".
936 * At a high level, all that happens here is marking the table entry
937 * at the bottom level available, and propagating the changes upward
938 * as necessary, plus some accounting needed to play nicely with other
939 * parts of the VM system.
940 * At each level, we keep a list of pages, which are heads of continuous
941 * free pages of length of (1 << order) and marked with PageBuddy.
942 * Page's order is recorded in page_private(page) field.
943 * So when we are allocating or freeing one, we can derive the state of the
944 * other. That is, if we allocate a small block, and both were
945 * free, the remainder of the region must be split into blocks.
946 * If a block is freed, and its buddy is also free, then this
947 * triggers coalescing into a block of larger size.
948 *
949 * -- nyc
950 */
951
952static inline void __free_one_page(struct page *page,
953 unsigned long pfn,
954 struct zone *zone, unsigned int order,
955 int migratetype, bool report)
956{
957 struct capture_control *capc = task_capc(zone);
958 unsigned long buddy_pfn;
959 unsigned long combined_pfn;
960 unsigned int max_order;
961 struct page *buddy;
962 bool to_tail;
963
964 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
965
966 VM_BUG_ON(!zone_is_initialized(zone));
967 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
968
969 VM_BUG_ON(migratetype == -1);
970 if (likely(!is_migrate_isolate(migratetype)))
971 __mod_zone_freepage_state(zone, 1 << order, migratetype);
972
973 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
974 VM_BUG_ON_PAGE(bad_range(zone, page), page);
975
976continue_merging:
977 while (order < max_order - 1) {
978 if (compaction_capture(capc, page, order, migratetype)) {
979 __mod_zone_freepage_state(zone, -(1 << order),
980 migratetype);
981 return;
982 }
983 buddy_pfn = __find_buddy_pfn(pfn, order);
984 buddy = page + (buddy_pfn - pfn);
985
986 if (!pfn_valid_within(buddy_pfn))
987 goto done_merging;
988 if (!page_is_buddy(page, buddy, order))
989 goto done_merging;
990 /*
991 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
992 * merge with it and move up one order.
993 */
994 if (page_is_guard(buddy))
995 clear_page_guard(zone, buddy, order, migratetype);
996 else
997 del_page_from_free_list(buddy, zone, order);
998 combined_pfn = buddy_pfn & pfn;
999 page = page + (combined_pfn - pfn);
1000 pfn = combined_pfn;
1001 order++;
1002 }
1003 if (max_order < MAX_ORDER) {
1004 /* If we are here, it means order is >= pageblock_order.
1005 * We want to prevent merge between freepages on isolate
1006 * pageblock and normal pageblock. Without this, pageblock
1007 * isolation could cause incorrect freepage or CMA accounting.
1008 *
1009 * We don't want to hit this code for the more frequent
1010 * low-order merging.
1011 */
1012 if (unlikely(has_isolate_pageblock(zone))) {
1013 int buddy_mt;
1014
1015 buddy_pfn = __find_buddy_pfn(pfn, order);
1016 buddy = page + (buddy_pfn - pfn);
1017 buddy_mt = get_pageblock_migratetype(buddy);
1018
1019 if (migratetype != buddy_mt
1020 && (is_migrate_isolate(migratetype) ||
1021 is_migrate_isolate(buddy_mt)))
1022 goto done_merging;
1023 }
1024 max_order++;
1025 goto continue_merging;
1026 }
1027
1028done_merging:
1029 set_page_order(page, order);
1030
1031 if (is_shuffle_order(order))
1032 to_tail = shuffle_pick_tail();
1033 else
1034 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1035
1036 if (to_tail)
1037 add_to_free_list_tail(page, zone, order, migratetype);
1038 else
1039 add_to_free_list(page, zone, order, migratetype);
1040
1041 /* Notify page reporting subsystem of freed page */
1042 if (report)
1043 page_reporting_notify_free(order);
1044}
1045
1046/*
1047 * A bad page could be due to a number of fields. Instead of multiple branches,
1048 * try and check multiple fields with one check. The caller must do a detailed
1049 * check if necessary.
1050 */
1051static inline bool page_expected_state(struct page *page,
1052 unsigned long check_flags)
1053{
1054 if (unlikely(atomic_read(&page->_mapcount) != -1))
1055 return false;
1056
1057 if (unlikely((unsigned long)page->mapping |
1058 page_ref_count(page) |
1059#ifdef CONFIG_MEMCG
1060 (unsigned long)page->mem_cgroup |
1061#endif
1062 (page->flags & check_flags)))
1063 return false;
1064
1065 return true;
1066}
1067
1068static const char *page_bad_reason(struct page *page, unsigned long flags)
1069{
1070 const char *bad_reason = NULL;
1071
1072 if (unlikely(atomic_read(&page->_mapcount) != -1))
1073 bad_reason = "nonzero mapcount";
1074 if (unlikely(page->mapping != NULL))
1075 bad_reason = "non-NULL mapping";
1076 if (unlikely(page_ref_count(page) != 0))
1077 bad_reason = "nonzero _refcount";
1078 if (unlikely(page->flags & flags)) {
1079 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1080 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1081 else
1082 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1083 }
1084#ifdef CONFIG_MEMCG
1085 if (unlikely(page->mem_cgroup))
1086 bad_reason = "page still charged to cgroup";
1087#endif
1088 return bad_reason;
1089}
1090
1091static void check_free_page_bad(struct page *page)
1092{
1093 bad_page(page,
1094 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1095}
1096
1097static inline int check_free_page(struct page *page)
1098{
1099 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1100 return 0;
1101
1102 /* Something has gone sideways, find it */
1103 check_free_page_bad(page);
1104 return 1;
1105}
1106
1107static int free_tail_pages_check(struct page *head_page, struct page *page)
1108{
1109 int ret = 1;
1110
1111 /*
1112 * We rely page->lru.next never has bit 0 set, unless the page
1113 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1114 */
1115 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1116
1117 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1118 ret = 0;
1119 goto out;
1120 }
1121 switch (page - head_page) {
1122 case 1:
1123 /* the first tail page: ->mapping may be compound_mapcount() */
1124 if (unlikely(compound_mapcount(page))) {
1125 bad_page(page, "nonzero compound_mapcount");
1126 goto out;
1127 }
1128 break;
1129 case 2:
1130 /*
1131 * the second tail page: ->mapping is
1132 * deferred_list.next -- ignore value.
1133 */
1134 break;
1135 default:
1136 if (page->mapping != TAIL_MAPPING) {
1137 bad_page(page, "corrupted mapping in tail page");
1138 goto out;
1139 }
1140 break;
1141 }
1142 if (unlikely(!PageTail(page))) {
1143 bad_page(page, "PageTail not set");
1144 goto out;
1145 }
1146 if (unlikely(compound_head(page) != head_page)) {
1147 bad_page(page, "compound_head not consistent");
1148 goto out;
1149 }
1150 ret = 0;
1151out:
1152 page->mapping = NULL;
1153 clear_compound_head(page);
1154 return ret;
1155}
1156
1157static void kernel_init_free_pages(struct page *page, int numpages)
1158{
1159 int i;
1160
1161 /* s390's use of memset() could override KASAN redzones. */
1162 kasan_disable_current();
1163 for (i = 0; i < numpages; i++)
1164 clear_highpage(page + i);
1165 kasan_enable_current();
1166}
1167
1168static __always_inline bool free_pages_prepare(struct page *page,
1169 unsigned int order, bool check_free)
1170{
1171 int bad = 0;
1172
1173 VM_BUG_ON_PAGE(PageTail(page), page);
1174
1175 trace_mm_page_free(page, order);
1176
1177 /*
1178 * Check tail pages before head page information is cleared to
1179 * avoid checking PageCompound for order-0 pages.
1180 */
1181 if (unlikely(order)) {
1182 bool compound = PageCompound(page);
1183 int i;
1184
1185 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1186
1187 if (compound)
1188 ClearPageDoubleMap(page);
1189 for (i = 1; i < (1 << order); i++) {
1190 if (compound)
1191 bad += free_tail_pages_check(page, page + i);
1192 if (unlikely(check_free_page(page + i))) {
1193 bad++;
1194 continue;
1195 }
1196 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1197 }
1198 }
1199 if (PageMappingFlags(page))
1200 page->mapping = NULL;
1201 if (memcg_kmem_enabled() && PageKmemcg(page))
1202 __memcg_kmem_uncharge_page(page, order);
1203 if (check_free)
1204 bad += check_free_page(page);
1205 if (bad)
1206 return false;
1207
1208 page_cpupid_reset_last(page);
1209 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1210 reset_page_owner(page, order);
1211
1212 if (!PageHighMem(page)) {
1213 debug_check_no_locks_freed(page_address(page),
1214 PAGE_SIZE << order);
1215 debug_check_no_obj_freed(page_address(page),
1216 PAGE_SIZE << order);
1217 }
1218 if (want_init_on_free())
1219 kernel_init_free_pages(page, 1 << order);
1220
1221 kernel_poison_pages(page, 1 << order, 0);
1222 /*
1223 * arch_free_page() can make the page's contents inaccessible. s390
1224 * does this. So nothing which can access the page's contents should
1225 * happen after this.
1226 */
1227 arch_free_page(page, order);
1228
1229 if (debug_pagealloc_enabled_static())
1230 kernel_map_pages(page, 1 << order, 0);
1231
1232 kasan_free_nondeferred_pages(page, order);
1233
1234 return true;
1235}
1236
1237#ifdef CONFIG_DEBUG_VM
1238/*
1239 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1240 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1241 * moved from pcp lists to free lists.
1242 */
1243static bool free_pcp_prepare(struct page *page)
1244{
1245 return free_pages_prepare(page, 0, true);
1246}
1247
1248static bool bulkfree_pcp_prepare(struct page *page)
1249{
1250 if (debug_pagealloc_enabled_static())
1251 return check_free_page(page);
1252 else
1253 return false;
1254}
1255#else
1256/*
1257 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1258 * moving from pcp lists to free list in order to reduce overhead. With
1259 * debug_pagealloc enabled, they are checked also immediately when being freed
1260 * to the pcp lists.
1261 */
1262static bool free_pcp_prepare(struct page *page)
1263{
1264 if (debug_pagealloc_enabled_static())
1265 return free_pages_prepare(page, 0, true);
1266 else
1267 return free_pages_prepare(page, 0, false);
1268}
1269
1270static bool bulkfree_pcp_prepare(struct page *page)
1271{
1272 return check_free_page(page);
1273}
1274#endif /* CONFIG_DEBUG_VM */
1275
1276static inline void prefetch_buddy(struct page *page)
1277{
1278 unsigned long pfn = page_to_pfn(page);
1279 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1280 struct page *buddy = page + (buddy_pfn - pfn);
1281
1282 prefetch(buddy);
1283}
1284
1285/*
1286 * Frees a number of pages from the PCP lists
1287 * Assumes all pages on list are in same zone, and of same order.
1288 * count is the number of pages to free.
1289 *
1290 * If the zone was previously in an "all pages pinned" state then look to
1291 * see if this freeing clears that state.
1292 *
1293 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1294 * pinned" detection logic.
1295 */
1296static void free_pcppages_bulk(struct zone *zone, int count,
1297 struct per_cpu_pages *pcp)
1298{
1299 int migratetype = 0;
1300 int batch_free = 0;
1301 int prefetch_nr = 0;
1302 bool isolated_pageblocks;
1303 struct page *page, *tmp;
1304 LIST_HEAD(head);
1305
1306 /*
1307 * Ensure proper count is passed which otherwise would stuck in the
1308 * below while (list_empty(list)) loop.
1309 */
1310 count = min(pcp->count, count);
1311 while (count) {
1312 struct list_head *list;
1313
1314 /*
1315 * Remove pages from lists in a round-robin fashion. A
1316 * batch_free count is maintained that is incremented when an
1317 * empty list is encountered. This is so more pages are freed
1318 * off fuller lists instead of spinning excessively around empty
1319 * lists
1320 */
1321 do {
1322 batch_free++;
1323 if (++migratetype == MIGRATE_PCPTYPES)
1324 migratetype = 0;
1325 list = &pcp->lists[migratetype];
1326 } while (list_empty(list));
1327
1328 /* This is the only non-empty list. Free them all. */
1329 if (batch_free == MIGRATE_PCPTYPES)
1330 batch_free = count;
1331
1332 do {
1333 page = list_last_entry(list, struct page, lru);
1334 /* must delete to avoid corrupting pcp list */
1335 list_del(&page->lru);
1336 pcp->count--;
1337
1338 if (bulkfree_pcp_prepare(page))
1339 continue;
1340
1341 list_add_tail(&page->lru, &head);
1342
1343 /*
1344 * We are going to put the page back to the global
1345 * pool, prefetch its buddy to speed up later access
1346 * under zone->lock. It is believed the overhead of
1347 * an additional test and calculating buddy_pfn here
1348 * can be offset by reduced memory latency later. To
1349 * avoid excessive prefetching due to large count, only
1350 * prefetch buddy for the first pcp->batch nr of pages.
1351 */
1352 if (prefetch_nr++ < pcp->batch)
1353 prefetch_buddy(page);
1354 } while (--count && --batch_free && !list_empty(list));
1355 }
1356
1357 spin_lock(&zone->lock);
1358 isolated_pageblocks = has_isolate_pageblock(zone);
1359
1360 /*
1361 * Use safe version since after __free_one_page(),
1362 * page->lru.next will not point to original list.
1363 */
1364 list_for_each_entry_safe(page, tmp, &head, lru) {
1365 int mt = get_pcppage_migratetype(page);
1366 /* MIGRATE_ISOLATE page should not go to pcplists */
1367 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1368 /* Pageblock could have been isolated meanwhile */
1369 if (unlikely(isolated_pageblocks))
1370 mt = get_pageblock_migratetype(page);
1371
1372 __free_one_page(page, page_to_pfn(page), zone, 0, mt, true);
1373 trace_mm_page_pcpu_drain(page, 0, mt);
1374 }
1375 spin_unlock(&zone->lock);
1376}
1377
1378static void free_one_page(struct zone *zone,
1379 struct page *page, unsigned long pfn,
1380 unsigned int order,
1381 int migratetype)
1382{
1383 spin_lock(&zone->lock);
1384 if (unlikely(has_isolate_pageblock(zone) ||
1385 is_migrate_isolate(migratetype))) {
1386 migratetype = get_pfnblock_migratetype(page, pfn);
1387 }
1388 __free_one_page(page, pfn, zone, order, migratetype, true);
1389 spin_unlock(&zone->lock);
1390}
1391
1392static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1393 unsigned long zone, int nid)
1394{
1395 mm_zero_struct_page(page);
1396 set_page_links(page, zone, nid, pfn);
1397 init_page_count(page);
1398 page_mapcount_reset(page);
1399 page_cpupid_reset_last(page);
1400 page_kasan_tag_reset(page);
1401
1402 INIT_LIST_HEAD(&page->lru);
1403#ifdef WANT_PAGE_VIRTUAL
1404 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1405 if (!is_highmem_idx(zone))
1406 set_page_address(page, __va(pfn << PAGE_SHIFT));
1407#endif
1408}
1409
1410#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1411static void __meminit init_reserved_page(unsigned long pfn)
1412{
1413 pg_data_t *pgdat;
1414 int nid, zid;
1415
1416 if (!early_page_uninitialised(pfn))
1417 return;
1418
1419 nid = early_pfn_to_nid(pfn);
1420 pgdat = NODE_DATA(nid);
1421
1422 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1423 struct zone *zone = &pgdat->node_zones[zid];
1424
1425 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1426 break;
1427 }
1428 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1429}
1430#else
1431static inline void init_reserved_page(unsigned long pfn)
1432{
1433}
1434#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1435
1436/*
1437 * Initialised pages do not have PageReserved set. This function is
1438 * called for each range allocated by the bootmem allocator and
1439 * marks the pages PageReserved. The remaining valid pages are later
1440 * sent to the buddy page allocator.
1441 */
1442void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1443{
1444 unsigned long start_pfn = PFN_DOWN(start);
1445 unsigned long end_pfn = PFN_UP(end);
1446
1447 for (; start_pfn < end_pfn; start_pfn++) {
1448 if (pfn_valid(start_pfn)) {
1449 struct page *page = pfn_to_page(start_pfn);
1450
1451 init_reserved_page(start_pfn);
1452
1453 /* Avoid false-positive PageTail() */
1454 INIT_LIST_HEAD(&page->lru);
1455
1456 /*
1457 * no need for atomic set_bit because the struct
1458 * page is not visible yet so nobody should
1459 * access it yet.
1460 */
1461 __SetPageReserved(page);
1462 }
1463 }
1464}
1465
1466static void __free_pages_ok(struct page *page, unsigned int order)
1467{
1468 unsigned long flags;
1469 int migratetype;
1470 unsigned long pfn = page_to_pfn(page);
1471
1472 if (!free_pages_prepare(page, order, true))
1473 return;
1474
1475 migratetype = get_pfnblock_migratetype(page, pfn);
1476 local_irq_save(flags);
1477 __count_vm_events(PGFREE, 1 << order);
1478 free_one_page(page_zone(page), page, pfn, order, migratetype);
1479 local_irq_restore(flags);
1480}
1481
1482void __free_pages_core(struct page *page, unsigned int order)
1483{
1484 unsigned int nr_pages = 1 << order;
1485 struct page *p = page;
1486 unsigned int loop;
1487
1488 prefetchw(p);
1489 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1490 prefetchw(p + 1);
1491 __ClearPageReserved(p);
1492 set_page_count(p, 0);
1493 }
1494 __ClearPageReserved(p);
1495 set_page_count(p, 0);
1496
1497 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1498 set_page_refcounted(page);
1499 __free_pages(page, order);
1500}
1501
1502#ifdef CONFIG_NEED_MULTIPLE_NODES
1503
1504static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1505
1506#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
1507
1508/*
1509 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1510 */
1511int __meminit __early_pfn_to_nid(unsigned long pfn,
1512 struct mminit_pfnnid_cache *state)
1513{
1514 unsigned long start_pfn, end_pfn;
1515 int nid;
1516
1517 if (state->last_start <= pfn && pfn < state->last_end)
1518 return state->last_nid;
1519
1520 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1521 if (nid != NUMA_NO_NODE) {
1522 state->last_start = start_pfn;
1523 state->last_end = end_pfn;
1524 state->last_nid = nid;
1525 }
1526
1527 return nid;
1528}
1529#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
1530
1531int __meminit early_pfn_to_nid(unsigned long pfn)
1532{
1533 static DEFINE_SPINLOCK(early_pfn_lock);
1534 int nid;
1535
1536 spin_lock(&early_pfn_lock);
1537 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1538 if (nid < 0)
1539 nid = first_online_node;
1540 spin_unlock(&early_pfn_lock);
1541
1542 return nid;
1543}
1544#endif /* CONFIG_NEED_MULTIPLE_NODES */
1545
1546void __init memblock_free_pages(struct page *page, unsigned long pfn,
1547 unsigned int order)
1548{
1549 if (early_page_uninitialised(pfn))
1550 return;
1551 __free_pages_core(page, order);
1552}
1553
1554/*
1555 * Check that the whole (or subset of) a pageblock given by the interval of
1556 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1557 * with the migration of free compaction scanner. The scanners then need to
1558 * use only pfn_valid_within() check for arches that allow holes within
1559 * pageblocks.
1560 *
1561 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1562 *
1563 * It's possible on some configurations to have a setup like node0 node1 node0
1564 * i.e. it's possible that all pages within a zones range of pages do not
1565 * belong to a single zone. We assume that a border between node0 and node1
1566 * can occur within a single pageblock, but not a node0 node1 node0
1567 * interleaving within a single pageblock. It is therefore sufficient to check
1568 * the first and last page of a pageblock and avoid checking each individual
1569 * page in a pageblock.
1570 */
1571struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1572 unsigned long end_pfn, struct zone *zone)
1573{
1574 struct page *start_page;
1575 struct page *end_page;
1576
1577 /* end_pfn is one past the range we are checking */
1578 end_pfn--;
1579
1580 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1581 return NULL;
1582
1583 start_page = pfn_to_online_page(start_pfn);
1584 if (!start_page)
1585 return NULL;
1586
1587 if (page_zone(start_page) != zone)
1588 return NULL;
1589
1590 end_page = pfn_to_page(end_pfn);
1591
1592 /* This gives a shorter code than deriving page_zone(end_page) */
1593 if (page_zone_id(start_page) != page_zone_id(end_page))
1594 return NULL;
1595
1596 return start_page;
1597}
1598
1599void set_zone_contiguous(struct zone *zone)
1600{
1601 unsigned long block_start_pfn = zone->zone_start_pfn;
1602 unsigned long block_end_pfn;
1603
1604 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1605 for (; block_start_pfn < zone_end_pfn(zone);
1606 block_start_pfn = block_end_pfn,
1607 block_end_pfn += pageblock_nr_pages) {
1608
1609 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1610
1611 if (!__pageblock_pfn_to_page(block_start_pfn,
1612 block_end_pfn, zone))
1613 return;
1614 cond_resched();
1615 }
1616
1617 /* We confirm that there is no hole */
1618 zone->contiguous = true;
1619}
1620
1621void clear_zone_contiguous(struct zone *zone)
1622{
1623 zone->contiguous = false;
1624}
1625
1626#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1627static void __init deferred_free_range(unsigned long pfn,
1628 unsigned long nr_pages)
1629{
1630 struct page *page;
1631 unsigned long i;
1632
1633 if (!nr_pages)
1634 return;
1635
1636 page = pfn_to_page(pfn);
1637
1638 /* Free a large naturally-aligned chunk if possible */
1639 if (nr_pages == pageblock_nr_pages &&
1640 (pfn & (pageblock_nr_pages - 1)) == 0) {
1641 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1642 __free_pages_core(page, pageblock_order);
1643 return;
1644 }
1645
1646 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1647 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1648 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1649 __free_pages_core(page, 0);
1650 }
1651}
1652
1653/* Completion tracking for deferred_init_memmap() threads */
1654static atomic_t pgdat_init_n_undone __initdata;
1655static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1656
1657static inline void __init pgdat_init_report_one_done(void)
1658{
1659 if (atomic_dec_and_test(&pgdat_init_n_undone))
1660 complete(&pgdat_init_all_done_comp);
1661}
1662
1663/*
1664 * Returns true if page needs to be initialized or freed to buddy allocator.
1665 *
1666 * First we check if pfn is valid on architectures where it is possible to have
1667 * holes within pageblock_nr_pages. On systems where it is not possible, this
1668 * function is optimized out.
1669 *
1670 * Then, we check if a current large page is valid by only checking the validity
1671 * of the head pfn.
1672 */
1673static inline bool __init deferred_pfn_valid(unsigned long pfn)
1674{
1675 if (!pfn_valid_within(pfn))
1676 return false;
1677 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1678 return false;
1679 return true;
1680}
1681
1682/*
1683 * Free pages to buddy allocator. Try to free aligned pages in
1684 * pageblock_nr_pages sizes.
1685 */
1686static void __init deferred_free_pages(unsigned long pfn,
1687 unsigned long end_pfn)
1688{
1689 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1690 unsigned long nr_free = 0;
1691
1692 for (; pfn < end_pfn; pfn++) {
1693 if (!deferred_pfn_valid(pfn)) {
1694 deferred_free_range(pfn - nr_free, nr_free);
1695 nr_free = 0;
1696 } else if (!(pfn & nr_pgmask)) {
1697 deferred_free_range(pfn - nr_free, nr_free);
1698 nr_free = 1;
1699 } else {
1700 nr_free++;
1701 }
1702 }
1703 /* Free the last block of pages to allocator */
1704 deferred_free_range(pfn - nr_free, nr_free);
1705}
1706
1707/*
1708 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1709 * by performing it only once every pageblock_nr_pages.
1710 * Return number of pages initialized.
1711 */
1712static unsigned long __init deferred_init_pages(struct zone *zone,
1713 unsigned long pfn,
1714 unsigned long end_pfn)
1715{
1716 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1717 int nid = zone_to_nid(zone);
1718 unsigned long nr_pages = 0;
1719 int zid = zone_idx(zone);
1720 struct page *page = NULL;
1721
1722 for (; pfn < end_pfn; pfn++) {
1723 if (!deferred_pfn_valid(pfn)) {
1724 page = NULL;
1725 continue;
1726 } else if (!page || !(pfn & nr_pgmask)) {
1727 page = pfn_to_page(pfn);
1728 } else {
1729 page++;
1730 }
1731 __init_single_page(page, pfn, zid, nid);
1732 nr_pages++;
1733 }
1734 return (nr_pages);
1735}
1736
1737/*
1738 * This function is meant to pre-load the iterator for the zone init.
1739 * Specifically it walks through the ranges until we are caught up to the
1740 * first_init_pfn value and exits there. If we never encounter the value we
1741 * return false indicating there are no valid ranges left.
1742 */
1743static bool __init
1744deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1745 unsigned long *spfn, unsigned long *epfn,
1746 unsigned long first_init_pfn)
1747{
1748 u64 j;
1749
1750 /*
1751 * Start out by walking through the ranges in this zone that have
1752 * already been initialized. We don't need to do anything with them
1753 * so we just need to flush them out of the system.
1754 */
1755 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1756 if (*epfn <= first_init_pfn)
1757 continue;
1758 if (*spfn < first_init_pfn)
1759 *spfn = first_init_pfn;
1760 *i = j;
1761 return true;
1762 }
1763
1764 return false;
1765}
1766
1767/*
1768 * Initialize and free pages. We do it in two loops: first we initialize
1769 * struct page, then free to buddy allocator, because while we are
1770 * freeing pages we can access pages that are ahead (computing buddy
1771 * page in __free_one_page()).
1772 *
1773 * In order to try and keep some memory in the cache we have the loop
1774 * broken along max page order boundaries. This way we will not cause
1775 * any issues with the buddy page computation.
1776 */
1777static unsigned long __init
1778deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1779 unsigned long *end_pfn)
1780{
1781 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1782 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1783 unsigned long nr_pages = 0;
1784 u64 j = *i;
1785
1786 /* First we loop through and initialize the page values */
1787 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1788 unsigned long t;
1789
1790 if (mo_pfn <= *start_pfn)
1791 break;
1792
1793 t = min(mo_pfn, *end_pfn);
1794 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1795
1796 if (mo_pfn < *end_pfn) {
1797 *start_pfn = mo_pfn;
1798 break;
1799 }
1800 }
1801
1802 /* Reset values and now loop through freeing pages as needed */
1803 swap(j, *i);
1804
1805 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1806 unsigned long t;
1807
1808 if (mo_pfn <= spfn)
1809 break;
1810
1811 t = min(mo_pfn, epfn);
1812 deferred_free_pages(spfn, t);
1813
1814 if (mo_pfn <= epfn)
1815 break;
1816 }
1817
1818 return nr_pages;
1819}
1820
1821static void __init
1822deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1823 void *arg)
1824{
1825 unsigned long spfn, epfn;
1826 struct zone *zone = arg;
1827 u64 i;
1828
1829 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1830
1831 /*
1832 * Initialize and free pages in MAX_ORDER sized increments so that we
1833 * can avoid introducing any issues with the buddy allocator.
1834 */
1835 while (spfn < end_pfn) {
1836 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1837 cond_resched();
1838 }
1839}
1840
1841/* An arch may override for more concurrency. */
1842__weak int __init
1843deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1844{
1845 return 1;
1846}
1847
1848/* Initialise remaining memory on a node */
1849static int __init deferred_init_memmap(void *data)
1850{
1851 pg_data_t *pgdat = data;
1852 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1853 unsigned long spfn = 0, epfn = 0;
1854 unsigned long first_init_pfn, flags;
1855 unsigned long start = jiffies;
1856 struct zone *zone;
1857 int zid, max_threads;
1858 u64 i;
1859
1860 /* Bind memory initialisation thread to a local node if possible */
1861 if (!cpumask_empty(cpumask))
1862 set_cpus_allowed_ptr(current, cpumask);
1863
1864 pgdat_resize_lock(pgdat, &flags);
1865 first_init_pfn = pgdat->first_deferred_pfn;
1866 if (first_init_pfn == ULONG_MAX) {
1867 pgdat_resize_unlock(pgdat, &flags);
1868 pgdat_init_report_one_done();
1869 return 0;
1870 }
1871
1872 /* Sanity check boundaries */
1873 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1874 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1875 pgdat->first_deferred_pfn = ULONG_MAX;
1876
1877 /*
1878 * Once we unlock here, the zone cannot be grown anymore, thus if an
1879 * interrupt thread must allocate this early in boot, zone must be
1880 * pre-grown prior to start of deferred page initialization.
1881 */
1882 pgdat_resize_unlock(pgdat, &flags);
1883
1884 /* Only the highest zone is deferred so find it */
1885 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1886 zone = pgdat->node_zones + zid;
1887 if (first_init_pfn < zone_end_pfn(zone))
1888 break;
1889 }
1890
1891 /* If the zone is empty somebody else may have cleared out the zone */
1892 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1893 first_init_pfn))
1894 goto zone_empty;
1895
1896 max_threads = deferred_page_init_max_threads(cpumask);
1897
1898 while (spfn < epfn) {
1899 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
1900 struct padata_mt_job job = {
1901 .thread_fn = deferred_init_memmap_chunk,
1902 .fn_arg = zone,
1903 .start = spfn,
1904 .size = epfn_align - spfn,
1905 .align = PAGES_PER_SECTION,
1906 .min_chunk = PAGES_PER_SECTION,
1907 .max_threads = max_threads,
1908 };
1909
1910 padata_do_multithreaded(&job);
1911 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1912 epfn_align);
1913 }
1914zone_empty:
1915 /* Sanity check that the next zone really is unpopulated */
1916 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1917
1918 pr_info("node %d deferred pages initialised in %ums\n",
1919 pgdat->node_id, jiffies_to_msecs(jiffies - start));
1920
1921 pgdat_init_report_one_done();
1922 return 0;
1923}
1924
1925/*
1926 * If this zone has deferred pages, try to grow it by initializing enough
1927 * deferred pages to satisfy the allocation specified by order, rounded up to
1928 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1929 * of SECTION_SIZE bytes by initializing struct pages in increments of
1930 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1931 *
1932 * Return true when zone was grown, otherwise return false. We return true even
1933 * when we grow less than requested, to let the caller decide if there are
1934 * enough pages to satisfy the allocation.
1935 *
1936 * Note: We use noinline because this function is needed only during boot, and
1937 * it is called from a __ref function _deferred_grow_zone. This way we are
1938 * making sure that it is not inlined into permanent text section.
1939 */
1940static noinline bool __init
1941deferred_grow_zone(struct zone *zone, unsigned int order)
1942{
1943 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1944 pg_data_t *pgdat = zone->zone_pgdat;
1945 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1946 unsigned long spfn, epfn, flags;
1947 unsigned long nr_pages = 0;
1948 u64 i;
1949
1950 /* Only the last zone may have deferred pages */
1951 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1952 return false;
1953
1954 pgdat_resize_lock(pgdat, &flags);
1955
1956 /*
1957 * If someone grew this zone while we were waiting for spinlock, return
1958 * true, as there might be enough pages already.
1959 */
1960 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1961 pgdat_resize_unlock(pgdat, &flags);
1962 return true;
1963 }
1964
1965 /* If the zone is empty somebody else may have cleared out the zone */
1966 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1967 first_deferred_pfn)) {
1968 pgdat->first_deferred_pfn = ULONG_MAX;
1969 pgdat_resize_unlock(pgdat, &flags);
1970 /* Retry only once. */
1971 return first_deferred_pfn != ULONG_MAX;
1972 }
1973
1974 /*
1975 * Initialize and free pages in MAX_ORDER sized increments so
1976 * that we can avoid introducing any issues with the buddy
1977 * allocator.
1978 */
1979 while (spfn < epfn) {
1980 /* update our first deferred PFN for this section */
1981 first_deferred_pfn = spfn;
1982
1983 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1984 touch_nmi_watchdog();
1985
1986 /* We should only stop along section boundaries */
1987 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
1988 continue;
1989
1990 /* If our quota has been met we can stop here */
1991 if (nr_pages >= nr_pages_needed)
1992 break;
1993 }
1994
1995 pgdat->first_deferred_pfn = spfn;
1996 pgdat_resize_unlock(pgdat, &flags);
1997
1998 return nr_pages > 0;
1999}
2000
2001/*
2002 * deferred_grow_zone() is __init, but it is called from
2003 * get_page_from_freelist() during early boot until deferred_pages permanently
2004 * disables this call. This is why we have refdata wrapper to avoid warning,
2005 * and to ensure that the function body gets unloaded.
2006 */
2007static bool __ref
2008_deferred_grow_zone(struct zone *zone, unsigned int order)
2009{
2010 return deferred_grow_zone(zone, order);
2011}
2012
2013#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2014
2015void __init page_alloc_init_late(void)
2016{
2017 struct zone *zone;
2018 int nid;
2019
2020#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2021
2022 /* There will be num_node_state(N_MEMORY) threads */
2023 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2024 for_each_node_state(nid, N_MEMORY) {
2025 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2026 }
2027
2028 /* Block until all are initialised */
2029 wait_for_completion(&pgdat_init_all_done_comp);
2030
2031 /*
2032 * The number of managed pages has changed due to the initialisation
2033 * so the pcpu batch and high limits needs to be updated or the limits
2034 * will be artificially small.
2035 */
2036 for_each_populated_zone(zone)
2037 zone_pcp_update(zone);
2038
2039 /*
2040 * We initialized the rest of the deferred pages. Permanently disable
2041 * on-demand struct page initialization.
2042 */
2043 static_branch_disable(&deferred_pages);
2044
2045 /* Reinit limits that are based on free pages after the kernel is up */
2046 files_maxfiles_init();
2047#endif
2048
2049 /* Discard memblock private memory */
2050 memblock_discard();
2051
2052 for_each_node_state(nid, N_MEMORY)
2053 shuffle_free_memory(NODE_DATA(nid));
2054
2055 for_each_populated_zone(zone)
2056 set_zone_contiguous(zone);
2057}
2058
2059#ifdef CONFIG_CMA
2060/* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2061void __init init_cma_reserved_pageblock(struct page *page)
2062{
2063 unsigned i = pageblock_nr_pages;
2064 struct page *p = page;
2065
2066 do {
2067 __ClearPageReserved(p);
2068 set_page_count(p, 0);
2069 } while (++p, --i);
2070
2071 set_pageblock_migratetype(page, MIGRATE_CMA);
2072
2073 if (pageblock_order >= MAX_ORDER) {
2074 i = pageblock_nr_pages;
2075 p = page;
2076 do {
2077 set_page_refcounted(p);
2078 __free_pages(p, MAX_ORDER - 1);
2079 p += MAX_ORDER_NR_PAGES;
2080 } while (i -= MAX_ORDER_NR_PAGES);
2081 } else {
2082 set_page_refcounted(page);
2083 __free_pages(page, pageblock_order);
2084 }
2085
2086 adjust_managed_page_count(page, pageblock_nr_pages);
2087}
2088#endif
2089
2090/*
2091 * The order of subdivision here is critical for the IO subsystem.
2092 * Please do not alter this order without good reasons and regression
2093 * testing. Specifically, as large blocks of memory are subdivided,
2094 * the order in which smaller blocks are delivered depends on the order
2095 * they're subdivided in this function. This is the primary factor
2096 * influencing the order in which pages are delivered to the IO
2097 * subsystem according to empirical testing, and this is also justified
2098 * by considering the behavior of a buddy system containing a single
2099 * large block of memory acted on by a series of small allocations.
2100 * This behavior is a critical factor in sglist merging's success.
2101 *
2102 * -- nyc
2103 */
2104static inline void expand(struct zone *zone, struct page *page,
2105 int low, int high, int migratetype)
2106{
2107 unsigned long size = 1 << high;
2108
2109 while (high > low) {
2110 high--;
2111 size >>= 1;
2112 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2113
2114 /*
2115 * Mark as guard pages (or page), that will allow to
2116 * merge back to allocator when buddy will be freed.
2117 * Corresponding page table entries will not be touched,
2118 * pages will stay not present in virtual address space
2119 */
2120 if (set_page_guard(zone, &page[size], high, migratetype))
2121 continue;
2122
2123 add_to_free_list(&page[size], zone, high, migratetype);
2124 set_page_order(&page[size], high);
2125 }
2126}
2127
2128static void check_new_page_bad(struct page *page)
2129{
2130 if (unlikely(page->flags & __PG_HWPOISON)) {
2131 /* Don't complain about hwpoisoned pages */
2132 page_mapcount_reset(page); /* remove PageBuddy */
2133 return;
2134 }
2135
2136 bad_page(page,
2137 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2138}
2139
2140/*
2141 * This page is about to be returned from the page allocator
2142 */
2143static inline int check_new_page(struct page *page)
2144{
2145 if (likely(page_expected_state(page,
2146 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2147 return 0;
2148
2149 check_new_page_bad(page);
2150 return 1;
2151}
2152
2153static inline bool free_pages_prezeroed(void)
2154{
2155 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2156 page_poisoning_enabled()) || want_init_on_free();
2157}
2158
2159#ifdef CONFIG_DEBUG_VM
2160/*
2161 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2162 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2163 * also checked when pcp lists are refilled from the free lists.
2164 */
2165static inline bool check_pcp_refill(struct page *page)
2166{
2167 if (debug_pagealloc_enabled_static())
2168 return check_new_page(page);
2169 else
2170 return false;
2171}
2172
2173static inline bool check_new_pcp(struct page *page)
2174{
2175 return check_new_page(page);
2176}
2177#else
2178/*
2179 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2180 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2181 * enabled, they are also checked when being allocated from the pcp lists.
2182 */
2183static inline bool check_pcp_refill(struct page *page)
2184{
2185 return check_new_page(page);
2186}
2187static inline bool check_new_pcp(struct page *page)
2188{
2189 if (debug_pagealloc_enabled_static())
2190 return check_new_page(page);
2191 else
2192 return false;
2193}
2194#endif /* CONFIG_DEBUG_VM */
2195
2196static bool check_new_pages(struct page *page, unsigned int order)
2197{
2198 int i;
2199 for (i = 0; i < (1 << order); i++) {
2200 struct page *p = page + i;
2201
2202 if (unlikely(check_new_page(p)))
2203 return true;
2204 }
2205
2206 return false;
2207}
2208
2209inline void post_alloc_hook(struct page *page, unsigned int order,
2210 gfp_t gfp_flags)
2211{
2212 set_page_private(page, 0);
2213 set_page_refcounted(page);
2214
2215 arch_alloc_page(page, order);
2216 if (debug_pagealloc_enabled_static())
2217 kernel_map_pages(page, 1 << order, 1);
2218 kasan_alloc_pages(page, order);
2219 kernel_poison_pages(page, 1 << order, 1);
2220 set_page_owner(page, order, gfp_flags);
2221}
2222
2223static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2224 unsigned int alloc_flags)
2225{
2226 post_alloc_hook(page, order, gfp_flags);
2227
2228 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2229 kernel_init_free_pages(page, 1 << order);
2230
2231 if (order && (gfp_flags & __GFP_COMP))
2232 prep_compound_page(page, order);
2233
2234 /*
2235 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2236 * allocate the page. The expectation is that the caller is taking
2237 * steps that will free more memory. The caller should avoid the page
2238 * being used for !PFMEMALLOC purposes.
2239 */
2240 if (alloc_flags & ALLOC_NO_WATERMARKS)
2241 set_page_pfmemalloc(page);
2242 else
2243 clear_page_pfmemalloc(page);
2244}
2245
2246/*
2247 * Go through the free lists for the given migratetype and remove
2248 * the smallest available page from the freelists
2249 */
2250static __always_inline
2251struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2252 int migratetype)
2253{
2254 unsigned int current_order;
2255 struct free_area *area;
2256 struct page *page;
2257
2258 /* Find a page of the appropriate size in the preferred list */
2259 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2260 area = &(zone->free_area[current_order]);
2261 page = get_page_from_free_area(area, migratetype);
2262 if (!page)
2263 continue;
2264 del_page_from_free_list(page, zone, current_order);
2265 expand(zone, page, order, current_order, migratetype);
2266 set_pcppage_migratetype(page, migratetype);
2267 return page;
2268 }
2269
2270 return NULL;
2271}
2272
2273
2274/*
2275 * This array describes the order lists are fallen back to when
2276 * the free lists for the desirable migrate type are depleted
2277 */
2278static int fallbacks[MIGRATE_TYPES][3] = {
2279 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2280 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2281 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2282#ifdef CONFIG_CMA
2283 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2284#endif
2285#ifdef CONFIG_MEMORY_ISOLATION
2286 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2287#endif
2288};
2289
2290#ifdef CONFIG_CMA
2291static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2292 unsigned int order)
2293{
2294 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2295}
2296#else
2297static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2298 unsigned int order) { return NULL; }
2299#endif
2300
2301/*
2302 * Move the free pages in a range to the free lists of the requested type.
2303 * Note that start_page and end_pages are not aligned on a pageblock
2304 * boundary. If alignment is required, use move_freepages_block()
2305 */
2306static int move_freepages(struct zone *zone,
2307 struct page *start_page, struct page *end_page,
2308 int migratetype, int *num_movable)
2309{
2310 struct page *page;
2311 unsigned int order;
2312 int pages_moved = 0;
2313
2314 for (page = start_page; page <= end_page;) {
2315 if (!pfn_valid_within(page_to_pfn(page))) {
2316 page++;
2317 continue;
2318 }
2319
2320 if (!PageBuddy(page)) {
2321 /*
2322 * We assume that pages that could be isolated for
2323 * migration are movable. But we don't actually try
2324 * isolating, as that would be expensive.
2325 */
2326 if (num_movable &&
2327 (PageLRU(page) || __PageMovable(page)))
2328 (*num_movable)++;
2329
2330 page++;
2331 continue;
2332 }
2333
2334 /* Make sure we are not inadvertently changing nodes */
2335 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2336 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2337
2338 order = page_order(page);
2339 move_to_free_list(page, zone, order, migratetype);
2340 page += 1 << order;
2341 pages_moved += 1 << order;
2342 }
2343
2344 return pages_moved;
2345}
2346
2347int move_freepages_block(struct zone *zone, struct page *page,
2348 int migratetype, int *num_movable)
2349{
2350 unsigned long start_pfn, end_pfn;
2351 struct page *start_page, *end_page;
2352
2353 if (num_movable)
2354 *num_movable = 0;
2355
2356 start_pfn = page_to_pfn(page);
2357 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2358 start_page = pfn_to_page(start_pfn);
2359 end_page = start_page + pageblock_nr_pages - 1;
2360 end_pfn = start_pfn + pageblock_nr_pages - 1;
2361
2362 /* Do not cross zone boundaries */
2363 if (!zone_spans_pfn(zone, start_pfn))
2364 start_page = page;
2365 if (!zone_spans_pfn(zone, end_pfn))
2366 return 0;
2367
2368 return move_freepages(zone, start_page, end_page, migratetype,
2369 num_movable);
2370}
2371
2372static void change_pageblock_range(struct page *pageblock_page,
2373 int start_order, int migratetype)
2374{
2375 int nr_pageblocks = 1 << (start_order - pageblock_order);
2376
2377 while (nr_pageblocks--) {
2378 set_pageblock_migratetype(pageblock_page, migratetype);
2379 pageblock_page += pageblock_nr_pages;
2380 }
2381}
2382
2383/*
2384 * When we are falling back to another migratetype during allocation, try to
2385 * steal extra free pages from the same pageblocks to satisfy further
2386 * allocations, instead of polluting multiple pageblocks.
2387 *
2388 * If we are stealing a relatively large buddy page, it is likely there will
2389 * be more free pages in the pageblock, so try to steal them all. For
2390 * reclaimable and unmovable allocations, we steal regardless of page size,
2391 * as fragmentation caused by those allocations polluting movable pageblocks
2392 * is worse than movable allocations stealing from unmovable and reclaimable
2393 * pageblocks.
2394 */
2395static bool can_steal_fallback(unsigned int order, int start_mt)
2396{
2397 /*
2398 * Leaving this order check is intended, although there is
2399 * relaxed order check in next check. The reason is that
2400 * we can actually steal whole pageblock if this condition met,
2401 * but, below check doesn't guarantee it and that is just heuristic
2402 * so could be changed anytime.
2403 */
2404 if (order >= pageblock_order)
2405 return true;
2406
2407 if (order >= pageblock_order / 2 ||
2408 start_mt == MIGRATE_RECLAIMABLE ||
2409 start_mt == MIGRATE_UNMOVABLE ||
2410 page_group_by_mobility_disabled)
2411 return true;
2412
2413 return false;
2414}
2415
2416static inline void boost_watermark(struct zone *zone)
2417{
2418 unsigned long max_boost;
2419
2420 if (!watermark_boost_factor)
2421 return;
2422 /*
2423 * Don't bother in zones that are unlikely to produce results.
2424 * On small machines, including kdump capture kernels running
2425 * in a small area, boosting the watermark can cause an out of
2426 * memory situation immediately.
2427 */
2428 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2429 return;
2430
2431 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2432 watermark_boost_factor, 10000);
2433
2434 /*
2435 * high watermark may be uninitialised if fragmentation occurs
2436 * very early in boot so do not boost. We do not fall
2437 * through and boost by pageblock_nr_pages as failing
2438 * allocations that early means that reclaim is not going
2439 * to help and it may even be impossible to reclaim the
2440 * boosted watermark resulting in a hang.
2441 */
2442 if (!max_boost)
2443 return;
2444
2445 max_boost = max(pageblock_nr_pages, max_boost);
2446
2447 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2448 max_boost);
2449}
2450
2451/*
2452 * This function implements actual steal behaviour. If order is large enough,
2453 * we can steal whole pageblock. If not, we first move freepages in this
2454 * pageblock to our migratetype and determine how many already-allocated pages
2455 * are there in the pageblock with a compatible migratetype. If at least half
2456 * of pages are free or compatible, we can change migratetype of the pageblock
2457 * itself, so pages freed in the future will be put on the correct free list.
2458 */
2459static void steal_suitable_fallback(struct zone *zone, struct page *page,
2460 unsigned int alloc_flags, int start_type, bool whole_block)
2461{
2462 unsigned int current_order = page_order(page);
2463 int free_pages, movable_pages, alike_pages;
2464 int old_block_type;
2465
2466 old_block_type = get_pageblock_migratetype(page);
2467
2468 /*
2469 * This can happen due to races and we want to prevent broken
2470 * highatomic accounting.
2471 */
2472 if (is_migrate_highatomic(old_block_type))
2473 goto single_page;
2474
2475 /* Take ownership for orders >= pageblock_order */
2476 if (current_order >= pageblock_order) {
2477 change_pageblock_range(page, current_order, start_type);
2478 goto single_page;
2479 }
2480
2481 /*
2482 * Boost watermarks to increase reclaim pressure to reduce the
2483 * likelihood of future fallbacks. Wake kswapd now as the node
2484 * may be balanced overall and kswapd will not wake naturally.
2485 */
2486 boost_watermark(zone);
2487 if (alloc_flags & ALLOC_KSWAPD)
2488 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2489
2490 /* We are not allowed to try stealing from the whole block */
2491 if (!whole_block)
2492 goto single_page;
2493
2494 free_pages = move_freepages_block(zone, page, start_type,
2495 &movable_pages);
2496 /*
2497 * Determine how many pages are compatible with our allocation.
2498 * For movable allocation, it's the number of movable pages which
2499 * we just obtained. For other types it's a bit more tricky.
2500 */
2501 if (start_type == MIGRATE_MOVABLE) {
2502 alike_pages = movable_pages;
2503 } else {
2504 /*
2505 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2506 * to MOVABLE pageblock, consider all non-movable pages as
2507 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2508 * vice versa, be conservative since we can't distinguish the
2509 * exact migratetype of non-movable pages.
2510 */
2511 if (old_block_type == MIGRATE_MOVABLE)
2512 alike_pages = pageblock_nr_pages
2513 - (free_pages + movable_pages);
2514 else
2515 alike_pages = 0;
2516 }
2517
2518 /* moving whole block can fail due to zone boundary conditions */
2519 if (!free_pages)
2520 goto single_page;
2521
2522 /*
2523 * If a sufficient number of pages in the block are either free or of
2524 * comparable migratability as our allocation, claim the whole block.
2525 */
2526 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2527 page_group_by_mobility_disabled)
2528 set_pageblock_migratetype(page, start_type);
2529
2530 return;
2531
2532single_page:
2533 move_to_free_list(page, zone, current_order, start_type);
2534}
2535
2536/*
2537 * Check whether there is a suitable fallback freepage with requested order.
2538 * If only_stealable is true, this function returns fallback_mt only if
2539 * we can steal other freepages all together. This would help to reduce
2540 * fragmentation due to mixed migratetype pages in one pageblock.
2541 */
2542int find_suitable_fallback(struct free_area *area, unsigned int order,
2543 int migratetype, bool only_stealable, bool *can_steal)
2544{
2545 int i;
2546 int fallback_mt;
2547
2548 if (area->nr_free == 0)
2549 return -1;
2550
2551 *can_steal = false;
2552 for (i = 0;; i++) {
2553 fallback_mt = fallbacks[migratetype][i];
2554 if (fallback_mt == MIGRATE_TYPES)
2555 break;
2556
2557 if (free_area_empty(area, fallback_mt))
2558 continue;
2559
2560 if (can_steal_fallback(order, migratetype))
2561 *can_steal = true;
2562
2563 if (!only_stealable)
2564 return fallback_mt;
2565
2566 if (*can_steal)
2567 return fallback_mt;
2568 }
2569
2570 return -1;
2571}
2572
2573/*
2574 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2575 * there are no empty page blocks that contain a page with a suitable order
2576 */
2577static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2578 unsigned int alloc_order)
2579{
2580 int mt;
2581 unsigned long max_managed, flags;
2582
2583 /*
2584 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2585 * Check is race-prone but harmless.
2586 */
2587 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2588 if (zone->nr_reserved_highatomic >= max_managed)
2589 return;
2590
2591 spin_lock_irqsave(&zone->lock, flags);
2592
2593 /* Recheck the nr_reserved_highatomic limit under the lock */
2594 if (zone->nr_reserved_highatomic >= max_managed)
2595 goto out_unlock;
2596
2597 /* Yoink! */
2598 mt = get_pageblock_migratetype(page);
2599 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2600 && !is_migrate_cma(mt)) {
2601 zone->nr_reserved_highatomic += pageblock_nr_pages;
2602 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2603 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2604 }
2605
2606out_unlock:
2607 spin_unlock_irqrestore(&zone->lock, flags);
2608}
2609
2610/*
2611 * Used when an allocation is about to fail under memory pressure. This
2612 * potentially hurts the reliability of high-order allocations when under
2613 * intense memory pressure but failed atomic allocations should be easier
2614 * to recover from than an OOM.
2615 *
2616 * If @force is true, try to unreserve a pageblock even though highatomic
2617 * pageblock is exhausted.
2618 */
2619static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2620 bool force)
2621{
2622 struct zonelist *zonelist = ac->zonelist;
2623 unsigned long flags;
2624 struct zoneref *z;
2625 struct zone *zone;
2626 struct page *page;
2627 int order;
2628 bool ret;
2629
2630 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2631 ac->nodemask) {
2632 /*
2633 * Preserve at least one pageblock unless memory pressure
2634 * is really high.
2635 */
2636 if (!force && zone->nr_reserved_highatomic <=
2637 pageblock_nr_pages)
2638 continue;
2639
2640 spin_lock_irqsave(&zone->lock, flags);
2641 for (order = 0; order < MAX_ORDER; order++) {
2642 struct free_area *area = &(zone->free_area[order]);
2643
2644 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2645 if (!page)
2646 continue;
2647
2648 /*
2649 * In page freeing path, migratetype change is racy so
2650 * we can counter several free pages in a pageblock
2651 * in this loop althoug we changed the pageblock type
2652 * from highatomic to ac->migratetype. So we should
2653 * adjust the count once.
2654 */
2655 if (is_migrate_highatomic_page(page)) {
2656 /*
2657 * It should never happen but changes to
2658 * locking could inadvertently allow a per-cpu
2659 * drain to add pages to MIGRATE_HIGHATOMIC
2660 * while unreserving so be safe and watch for
2661 * underflows.
2662 */
2663 zone->nr_reserved_highatomic -= min(
2664 pageblock_nr_pages,
2665 zone->nr_reserved_highatomic);
2666 }
2667
2668 /*
2669 * Convert to ac->migratetype and avoid the normal
2670 * pageblock stealing heuristics. Minimally, the caller
2671 * is doing the work and needs the pages. More
2672 * importantly, if the block was always converted to
2673 * MIGRATE_UNMOVABLE or another type then the number
2674 * of pageblocks that cannot be completely freed
2675 * may increase.
2676 */
2677 set_pageblock_migratetype(page, ac->migratetype);
2678 ret = move_freepages_block(zone, page, ac->migratetype,
2679 NULL);
2680 if (ret) {
2681 spin_unlock_irqrestore(&zone->lock, flags);
2682 return ret;
2683 }
2684 }
2685 spin_unlock_irqrestore(&zone->lock, flags);
2686 }
2687
2688 return false;
2689}
2690
2691/*
2692 * Try finding a free buddy page on the fallback list and put it on the free
2693 * list of requested migratetype, possibly along with other pages from the same
2694 * block, depending on fragmentation avoidance heuristics. Returns true if
2695 * fallback was found so that __rmqueue_smallest() can grab it.
2696 *
2697 * The use of signed ints for order and current_order is a deliberate
2698 * deviation from the rest of this file, to make the for loop
2699 * condition simpler.
2700 */
2701static __always_inline bool
2702__rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2703 unsigned int alloc_flags)
2704{
2705 struct free_area *area;
2706 int current_order;
2707 int min_order = order;
2708 struct page *page;
2709 int fallback_mt;
2710 bool can_steal;
2711
2712 /*
2713 * Do not steal pages from freelists belonging to other pageblocks
2714 * i.e. orders < pageblock_order. If there are no local zones free,
2715 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2716 */
2717 if (alloc_flags & ALLOC_NOFRAGMENT)
2718 min_order = pageblock_order;
2719
2720 /*
2721 * Find the largest available free page in the other list. This roughly
2722 * approximates finding the pageblock with the most free pages, which
2723 * would be too costly to do exactly.
2724 */
2725 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2726 --current_order) {
2727 area = &(zone->free_area[current_order]);
2728 fallback_mt = find_suitable_fallback(area, current_order,
2729 start_migratetype, false, &can_steal);
2730 if (fallback_mt == -1)
2731 continue;
2732
2733 /*
2734 * We cannot steal all free pages from the pageblock and the
2735 * requested migratetype is movable. In that case it's better to
2736 * steal and split the smallest available page instead of the
2737 * largest available page, because even if the next movable
2738 * allocation falls back into a different pageblock than this
2739 * one, it won't cause permanent fragmentation.
2740 */
2741 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2742 && current_order > order)
2743 goto find_smallest;
2744
2745 goto do_steal;
2746 }
2747
2748 return false;
2749
2750find_smallest:
2751 for (current_order = order; current_order < MAX_ORDER;
2752 current_order++) {
2753 area = &(zone->free_area[current_order]);
2754 fallback_mt = find_suitable_fallback(area, current_order,
2755 start_migratetype, false, &can_steal);
2756 if (fallback_mt != -1)
2757 break;
2758 }
2759
2760 /*
2761 * This should not happen - we already found a suitable fallback
2762 * when looking for the largest page.
2763 */
2764 VM_BUG_ON(current_order == MAX_ORDER);
2765
2766do_steal:
2767 page = get_page_from_free_area(area, fallback_mt);
2768
2769 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2770 can_steal);
2771
2772 trace_mm_page_alloc_extfrag(page, order, current_order,
2773 start_migratetype, fallback_mt);
2774
2775 return true;
2776
2777}
2778
2779/*
2780 * Do the hard work of removing an element from the buddy allocator.
2781 * Call me with the zone->lock already held.
2782 */
2783static __always_inline struct page *
2784__rmqueue(struct zone *zone, unsigned int order, int migratetype,
2785 unsigned int alloc_flags)
2786{
2787 struct page *page;
2788
2789#ifdef CONFIG_CMA
2790 /*
2791 * Balance movable allocations between regular and CMA areas by
2792 * allocating from CMA when over half of the zone's free memory
2793 * is in the CMA area.
2794 */
2795 if (alloc_flags & ALLOC_CMA &&
2796 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2797 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2798 page = __rmqueue_cma_fallback(zone, order);
2799 if (page)
2800 return page;
2801 }
2802#endif
2803retry:
2804 page = __rmqueue_smallest(zone, order, migratetype);
2805 if (unlikely(!page)) {
2806 if (alloc_flags & ALLOC_CMA)
2807 page = __rmqueue_cma_fallback(zone, order);
2808
2809 if (!page && __rmqueue_fallback(zone, order, migratetype,
2810 alloc_flags))
2811 goto retry;
2812 }
2813
2814 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2815 return page;
2816}
2817
2818/*
2819 * Obtain a specified number of elements from the buddy allocator, all under
2820 * a single hold of the lock, for efficiency. Add them to the supplied list.
2821 * Returns the number of new pages which were placed at *list.
2822 */
2823static int rmqueue_bulk(struct zone *zone, unsigned int order,
2824 unsigned long count, struct list_head *list,
2825 int migratetype, unsigned int alloc_flags)
2826{
2827 int i, alloced = 0;
2828
2829 spin_lock(&zone->lock);
2830 for (i = 0; i < count; ++i) {
2831 struct page *page = __rmqueue(zone, order, migratetype,
2832 alloc_flags);
2833 if (unlikely(page == NULL))
2834 break;
2835
2836 if (unlikely(check_pcp_refill(page)))
2837 continue;
2838
2839 /*
2840 * Split buddy pages returned by expand() are received here in
2841 * physical page order. The page is added to the tail of
2842 * caller's list. From the callers perspective, the linked list
2843 * is ordered by page number under some conditions. This is
2844 * useful for IO devices that can forward direction from the
2845 * head, thus also in the physical page order. This is useful
2846 * for IO devices that can merge IO requests if the physical
2847 * pages are ordered properly.
2848 */
2849 list_add_tail(&page->lru, list);
2850 alloced++;
2851 if (is_migrate_cma(get_pcppage_migratetype(page)))
2852 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2853 -(1 << order));
2854 }
2855
2856 /*
2857 * i pages were removed from the buddy list even if some leak due
2858 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2859 * on i. Do not confuse with 'alloced' which is the number of
2860 * pages added to the pcp list.
2861 */
2862 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2863 spin_unlock(&zone->lock);
2864 return alloced;
2865}
2866
2867#ifdef CONFIG_NUMA
2868/*
2869 * Called from the vmstat counter updater to drain pagesets of this
2870 * currently executing processor on remote nodes after they have
2871 * expired.
2872 *
2873 * Note that this function must be called with the thread pinned to
2874 * a single processor.
2875 */
2876void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2877{
2878 unsigned long flags;
2879 int to_drain, batch;
2880
2881 local_irq_save(flags);
2882 batch = READ_ONCE(pcp->batch);
2883 to_drain = min(pcp->count, batch);
2884 if (to_drain > 0)
2885 free_pcppages_bulk(zone, to_drain, pcp);
2886 local_irq_restore(flags);
2887}
2888#endif
2889
2890/*
2891 * Drain pcplists of the indicated processor and zone.
2892 *
2893 * The processor must either be the current processor and the
2894 * thread pinned to the current processor or a processor that
2895 * is not online.
2896 */
2897static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2898{
2899 unsigned long flags;
2900 struct per_cpu_pageset *pset;
2901 struct per_cpu_pages *pcp;
2902
2903 local_irq_save(flags);
2904 pset = per_cpu_ptr(zone->pageset, cpu);
2905
2906 pcp = &pset->pcp;
2907 if (pcp->count)
2908 free_pcppages_bulk(zone, pcp->count, pcp);
2909 local_irq_restore(flags);
2910}
2911
2912/*
2913 * Drain pcplists of all zones on the indicated processor.
2914 *
2915 * The processor must either be the current processor and the
2916 * thread pinned to the current processor or a processor that
2917 * is not online.
2918 */
2919static void drain_pages(unsigned int cpu)
2920{
2921 struct zone *zone;
2922
2923 for_each_populated_zone(zone) {
2924 drain_pages_zone(cpu, zone);
2925 }
2926}
2927
2928/*
2929 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2930 *
2931 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2932 * the single zone's pages.
2933 */
2934void drain_local_pages(struct zone *zone)
2935{
2936 int cpu = smp_processor_id();
2937
2938 if (zone)
2939 drain_pages_zone(cpu, zone);
2940 else
2941 drain_pages(cpu);
2942}
2943
2944static void drain_local_pages_wq(struct work_struct *work)
2945{
2946 struct pcpu_drain *drain;
2947
2948 drain = container_of(work, struct pcpu_drain, work);
2949
2950 /*
2951 * drain_all_pages doesn't use proper cpu hotplug protection so
2952 * we can race with cpu offline when the WQ can move this from
2953 * a cpu pinned worker to an unbound one. We can operate on a different
2954 * cpu which is allright but we also have to make sure to not move to
2955 * a different one.
2956 */
2957 preempt_disable();
2958 drain_local_pages(drain->zone);
2959 preempt_enable();
2960}
2961
2962/*
2963 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2964 *
2965 * When zone parameter is non-NULL, spill just the single zone's pages.
2966 *
2967 * Note that this can be extremely slow as the draining happens in a workqueue.
2968 */
2969void drain_all_pages(struct zone *zone)
2970{
2971 int cpu;
2972
2973 /*
2974 * Allocate in the BSS so we wont require allocation in
2975 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2976 */
2977 static cpumask_t cpus_with_pcps;
2978
2979 /*
2980 * Make sure nobody triggers this path before mm_percpu_wq is fully
2981 * initialized.
2982 */
2983 if (WARN_ON_ONCE(!mm_percpu_wq))
2984 return;
2985
2986 /*
2987 * Do not drain if one is already in progress unless it's specific to
2988 * a zone. Such callers are primarily CMA and memory hotplug and need
2989 * the drain to be complete when the call returns.
2990 */
2991 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2992 if (!zone)
2993 return;
2994 mutex_lock(&pcpu_drain_mutex);
2995 }
2996
2997 /*
2998 * We don't care about racing with CPU hotplug event
2999 * as offline notification will cause the notified
3000 * cpu to drain that CPU pcps and on_each_cpu_mask
3001 * disables preemption as part of its processing
3002 */
3003 for_each_online_cpu(cpu) {
3004 struct per_cpu_pageset *pcp;
3005 struct zone *z;
3006 bool has_pcps = false;
3007
3008 if (zone) {
3009 pcp = per_cpu_ptr(zone->pageset, cpu);
3010 if (pcp->pcp.count)
3011 has_pcps = true;
3012 } else {
3013 for_each_populated_zone(z) {
3014 pcp = per_cpu_ptr(z->pageset, cpu);
3015 if (pcp->pcp.count) {
3016 has_pcps = true;
3017 break;
3018 }
3019 }
3020 }
3021
3022 if (has_pcps)
3023 cpumask_set_cpu(cpu, &cpus_with_pcps);
3024 else
3025 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3026 }
3027
3028 for_each_cpu(cpu, &cpus_with_pcps) {
3029 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3030
3031 drain->zone = zone;
3032 INIT_WORK(&drain->work, drain_local_pages_wq);
3033 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3034 }
3035 for_each_cpu(cpu, &cpus_with_pcps)
3036 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3037
3038 mutex_unlock(&pcpu_drain_mutex);
3039}
3040
3041#ifdef CONFIG_HIBERNATION
3042
3043/*
3044 * Touch the watchdog for every WD_PAGE_COUNT pages.
3045 */
3046#define WD_PAGE_COUNT (128*1024)
3047
3048void mark_free_pages(struct zone *zone)
3049{
3050 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3051 unsigned long flags;
3052 unsigned int order, t;
3053 struct page *page;
3054
3055 if (zone_is_empty(zone))
3056 return;
3057
3058 spin_lock_irqsave(&zone->lock, flags);
3059
3060 max_zone_pfn = zone_end_pfn(zone);
3061 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3062 if (pfn_valid(pfn)) {
3063 page = pfn_to_page(pfn);
3064
3065 if (!--page_count) {
3066 touch_nmi_watchdog();
3067 page_count = WD_PAGE_COUNT;
3068 }
3069
3070 if (page_zone(page) != zone)
3071 continue;
3072
3073 if (!swsusp_page_is_forbidden(page))
3074 swsusp_unset_page_free(page);
3075 }
3076
3077 for_each_migratetype_order(order, t) {
3078 list_for_each_entry(page,
3079 &zone->free_area[order].free_list[t], lru) {
3080 unsigned long i;
3081
3082 pfn = page_to_pfn(page);
3083 for (i = 0; i < (1UL << order); i++) {
3084 if (!--page_count) {
3085 touch_nmi_watchdog();
3086 page_count = WD_PAGE_COUNT;
3087 }
3088 swsusp_set_page_free(pfn_to_page(pfn + i));
3089 }
3090 }
3091 }
3092 spin_unlock_irqrestore(&zone->lock, flags);
3093}
3094#endif /* CONFIG_PM */
3095
3096static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3097{
3098 int migratetype;
3099
3100 if (!free_pcp_prepare(page))
3101 return false;
3102
3103 migratetype = get_pfnblock_migratetype(page, pfn);
3104 set_pcppage_migratetype(page, migratetype);
3105 return true;
3106}
3107
3108static void free_unref_page_commit(struct page *page, unsigned long pfn)
3109{
3110 struct zone *zone = page_zone(page);
3111 struct per_cpu_pages *pcp;
3112 int migratetype;
3113
3114 migratetype = get_pcppage_migratetype(page);
3115 __count_vm_event(PGFREE);
3116
3117 /*
3118 * We only track unmovable, reclaimable and movable on pcp lists.
3119 * Free ISOLATE pages back to the allocator because they are being
3120 * offlined but treat HIGHATOMIC as movable pages so we can get those
3121 * areas back if necessary. Otherwise, we may have to free
3122 * excessively into the page allocator
3123 */
3124 if (migratetype >= MIGRATE_PCPTYPES) {
3125 if (unlikely(is_migrate_isolate(migratetype))) {
3126 free_one_page(zone, page, pfn, 0, migratetype);
3127 return;
3128 }
3129 migratetype = MIGRATE_MOVABLE;
3130 }
3131
3132 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3133 list_add(&page->lru, &pcp->lists[migratetype]);
3134 pcp->count++;
3135 if (pcp->count >= pcp->high) {
3136 unsigned long batch = READ_ONCE(pcp->batch);
3137 free_pcppages_bulk(zone, batch, pcp);
3138 }
3139}
3140
3141/*
3142 * Free a 0-order page
3143 */
3144void free_unref_page(struct page *page)
3145{
3146 unsigned long flags;
3147 unsigned long pfn = page_to_pfn(page);
3148
3149 if (!free_unref_page_prepare(page, pfn))
3150 return;
3151
3152 local_irq_save(flags);
3153 free_unref_page_commit(page, pfn);
3154 local_irq_restore(flags);
3155}
3156
3157/*
3158 * Free a list of 0-order pages
3159 */
3160void free_unref_page_list(struct list_head *list)
3161{
3162 struct page *page, *next;
3163 unsigned long flags, pfn;
3164 int batch_count = 0;
3165
3166 /* Prepare pages for freeing */
3167 list_for_each_entry_safe(page, next, list, lru) {
3168 pfn = page_to_pfn(page);
3169 if (!free_unref_page_prepare(page, pfn))
3170 list_del(&page->lru);
3171 set_page_private(page, pfn);
3172 }
3173
3174 local_irq_save(flags);
3175 list_for_each_entry_safe(page, next, list, lru) {
3176 unsigned long pfn = page_private(page);
3177
3178 set_page_private(page, 0);
3179 trace_mm_page_free_batched(page);
3180 free_unref_page_commit(page, pfn);
3181
3182 /*
3183 * Guard against excessive IRQ disabled times when we get
3184 * a large list of pages to free.
3185 */
3186 if (++batch_count == SWAP_CLUSTER_MAX) {
3187 local_irq_restore(flags);
3188 batch_count = 0;
3189 local_irq_save(flags);
3190 }
3191 }
3192 local_irq_restore(flags);
3193}
3194
3195/*
3196 * split_page takes a non-compound higher-order page, and splits it into
3197 * n (1<<order) sub-pages: page[0..n]
3198 * Each sub-page must be freed individually.
3199 *
3200 * Note: this is probably too low level an operation for use in drivers.
3201 * Please consult with lkml before using this in your driver.
3202 */
3203void split_page(struct page *page, unsigned int order)
3204{
3205 int i;
3206
3207 VM_BUG_ON_PAGE(PageCompound(page), page);
3208 VM_BUG_ON_PAGE(!page_count(page), page);
3209
3210 for (i = 1; i < (1 << order); i++)
3211 set_page_refcounted(page + i);
3212 split_page_owner(page, order);
3213}
3214EXPORT_SYMBOL_GPL(split_page);
3215
3216int __isolate_free_page(struct page *page, unsigned int order)
3217{
3218 unsigned long watermark;
3219 struct zone *zone;
3220 int mt;
3221
3222 BUG_ON(!PageBuddy(page));
3223
3224 zone = page_zone(page);
3225 mt = get_pageblock_migratetype(page);
3226
3227 if (!is_migrate_isolate(mt)) {
3228 /*
3229 * Obey watermarks as if the page was being allocated. We can
3230 * emulate a high-order watermark check with a raised order-0
3231 * watermark, because we already know our high-order page
3232 * exists.
3233 */
3234 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3235 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3236 return 0;
3237
3238 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3239 }
3240
3241 /* Remove page from free list */
3242
3243 del_page_from_free_list(page, zone, order);
3244
3245 /*
3246 * Set the pageblock if the isolated page is at least half of a
3247 * pageblock
3248 */
3249 if (order >= pageblock_order - 1) {
3250 struct page *endpage = page + (1 << order) - 1;
3251 for (; page < endpage; page += pageblock_nr_pages) {
3252 int mt = get_pageblock_migratetype(page);
3253 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3254 && !is_migrate_highatomic(mt))
3255 set_pageblock_migratetype(page,
3256 MIGRATE_MOVABLE);
3257 }
3258 }
3259
3260
3261 return 1UL << order;
3262}
3263
3264/**
3265 * __putback_isolated_page - Return a now-isolated page back where we got it
3266 * @page: Page that was isolated
3267 * @order: Order of the isolated page
3268 * @mt: The page's pageblock's migratetype
3269 *
3270 * This function is meant to return a page pulled from the free lists via
3271 * __isolate_free_page back to the free lists they were pulled from.
3272 */
3273void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3274{
3275 struct zone *zone = page_zone(page);
3276
3277 /* zone lock should be held when this function is called */
3278 lockdep_assert_held(&zone->lock);
3279
3280 /* Return isolated page to tail of freelist. */
3281 __free_one_page(page, page_to_pfn(page), zone, order, mt, false);
3282}
3283
3284/*
3285 * Update NUMA hit/miss statistics
3286 *
3287 * Must be called with interrupts disabled.
3288 */
3289static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3290{
3291#ifdef CONFIG_NUMA
3292 enum numa_stat_item local_stat = NUMA_LOCAL;
3293
3294 /* skip numa counters update if numa stats is disabled */
3295 if (!static_branch_likely(&vm_numa_stat_key))
3296 return;
3297
3298 if (zone_to_nid(z) != numa_node_id())
3299 local_stat = NUMA_OTHER;
3300
3301 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3302 __inc_numa_state(z, NUMA_HIT);
3303 else {
3304 __inc_numa_state(z, NUMA_MISS);
3305 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3306 }
3307 __inc_numa_state(z, local_stat);
3308#endif
3309}
3310
3311/* Remove page from the per-cpu list, caller must protect the list */
3312static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3313 unsigned int alloc_flags,
3314 struct per_cpu_pages *pcp,
3315 struct list_head *list)
3316{
3317 struct page *page;
3318
3319 do {
3320 if (list_empty(list)) {
3321 pcp->count += rmqueue_bulk(zone, 0,
3322 pcp->batch, list,
3323 migratetype, alloc_flags);
3324 if (unlikely(list_empty(list)))
3325 return NULL;
3326 }
3327
3328 page = list_first_entry(list, struct page, lru);
3329 list_del(&page->lru);
3330 pcp->count--;
3331 } while (check_new_pcp(page));
3332
3333 return page;
3334}
3335
3336/* Lock and remove page from the per-cpu list */
3337static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3338 struct zone *zone, gfp_t gfp_flags,
3339 int migratetype, unsigned int alloc_flags)
3340{
3341 struct per_cpu_pages *pcp;
3342 struct list_head *list;
3343 struct page *page;
3344 unsigned long flags;
3345
3346 local_irq_save(flags);
3347 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3348 list = &pcp->lists[migratetype];
3349 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3350 if (page) {
3351 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3352 zone_statistics(preferred_zone, zone);
3353 }
3354 local_irq_restore(flags);
3355 return page;
3356}
3357
3358/*
3359 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3360 */
3361static inline
3362struct page *rmqueue(struct zone *preferred_zone,
3363 struct zone *zone, unsigned int order,
3364 gfp_t gfp_flags, unsigned int alloc_flags,
3365 int migratetype)
3366{
3367 unsigned long flags;
3368 struct page *page;
3369
3370 if (likely(order == 0)) {
3371 /*
3372 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3373 * we need to skip it when CMA area isn't allowed.
3374 */
3375 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3376 migratetype != MIGRATE_MOVABLE) {
3377 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3378 migratetype, alloc_flags);
3379 goto out;
3380 }
3381 }
3382
3383 /*
3384 * We most definitely don't want callers attempting to
3385 * allocate greater than order-1 page units with __GFP_NOFAIL.
3386 */
3387 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3388 spin_lock_irqsave(&zone->lock, flags);
3389
3390 do {
3391 page = NULL;
3392 /*
3393 * order-0 request can reach here when the pcplist is skipped
3394 * due to non-CMA allocation context. HIGHATOMIC area is
3395 * reserved for high-order atomic allocation, so order-0
3396 * request should skip it.
3397 */
3398 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3399 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3400 if (page)
3401 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3402 }
3403 if (!page)
3404 page = __rmqueue(zone, order, migratetype, alloc_flags);
3405 } while (page && check_new_pages(page, order));
3406 spin_unlock(&zone->lock);
3407 if (!page)
3408 goto failed;
3409 __mod_zone_freepage_state(zone, -(1 << order),
3410 get_pcppage_migratetype(page));
3411
3412 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3413 zone_statistics(preferred_zone, zone);
3414 local_irq_restore(flags);
3415
3416out:
3417 /* Separate test+clear to avoid unnecessary atomics */
3418 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3419 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3420 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3421 }
3422
3423 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3424 return page;
3425
3426failed:
3427 local_irq_restore(flags);
3428 return NULL;
3429}
3430
3431#ifdef CONFIG_FAIL_PAGE_ALLOC
3432
3433static struct {
3434 struct fault_attr attr;
3435
3436 bool ignore_gfp_highmem;
3437 bool ignore_gfp_reclaim;
3438 u32 min_order;
3439} fail_page_alloc = {
3440 .attr = FAULT_ATTR_INITIALIZER,
3441 .ignore_gfp_reclaim = true,
3442 .ignore_gfp_highmem = true,
3443 .min_order = 1,
3444};
3445
3446static int __init setup_fail_page_alloc(char *str)
3447{
3448 return setup_fault_attr(&fail_page_alloc.attr, str);
3449}
3450__setup("fail_page_alloc=", setup_fail_page_alloc);
3451
3452static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3453{
3454 if (order < fail_page_alloc.min_order)
3455 return false;
3456 if (gfp_mask & __GFP_NOFAIL)
3457 return false;
3458 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3459 return false;
3460 if (fail_page_alloc.ignore_gfp_reclaim &&
3461 (gfp_mask & __GFP_DIRECT_RECLAIM))
3462 return false;
3463
3464 return should_fail(&fail_page_alloc.attr, 1 << order);
3465}
3466
3467#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3468
3469static int __init fail_page_alloc_debugfs(void)
3470{
3471 umode_t mode = S_IFREG | 0600;
3472 struct dentry *dir;
3473
3474 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3475 &fail_page_alloc.attr);
3476
3477 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3478 &fail_page_alloc.ignore_gfp_reclaim);
3479 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3480 &fail_page_alloc.ignore_gfp_highmem);
3481 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3482
3483 return 0;
3484}
3485
3486late_initcall(fail_page_alloc_debugfs);
3487
3488#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3489
3490#else /* CONFIG_FAIL_PAGE_ALLOC */
3491
3492static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3493{
3494 return false;
3495}
3496
3497#endif /* CONFIG_FAIL_PAGE_ALLOC */
3498
3499static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3500{
3501 return __should_fail_alloc_page(gfp_mask, order);
3502}
3503ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3504
3505static inline long __zone_watermark_unusable_free(struct zone *z,
3506 unsigned int order, unsigned int alloc_flags)
3507{
3508 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3509 long unusable_free = (1 << order) - 1;
3510
3511 /*
3512 * If the caller does not have rights to ALLOC_HARDER then subtract
3513 * the high-atomic reserves. This will over-estimate the size of the
3514 * atomic reserve but it avoids a search.
3515 */
3516 if (likely(!alloc_harder))
3517 unusable_free += z->nr_reserved_highatomic;
3518
3519#ifdef CONFIG_CMA
3520 /* If allocation can't use CMA areas don't use free CMA pages */
3521 if (!(alloc_flags & ALLOC_CMA))
3522 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3523#endif
3524
3525 return unusable_free;
3526}
3527
3528/*
3529 * Return true if free base pages are above 'mark'. For high-order checks it
3530 * will return true of the order-0 watermark is reached and there is at least
3531 * one free page of a suitable size. Checking now avoids taking the zone lock
3532 * to check in the allocation paths if no pages are free.
3533 */
3534bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3535 int highest_zoneidx, unsigned int alloc_flags,
3536 long free_pages)
3537{
3538 long min = mark;
3539 int o;
3540 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3541
3542 /* free_pages may go negative - that's OK */
3543 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3544
3545 if (alloc_flags & ALLOC_HIGH)
3546 min -= min / 2;
3547
3548 if (unlikely(alloc_harder)) {
3549 /*
3550 * OOM victims can try even harder than normal ALLOC_HARDER
3551 * users on the grounds that it's definitely going to be in
3552 * the exit path shortly and free memory. Any allocation it
3553 * makes during the free path will be small and short-lived.
3554 */
3555 if (alloc_flags & ALLOC_OOM)
3556 min -= min / 2;
3557 else
3558 min -= min / 4;
3559 }
3560
3561 /*
3562 * Check watermarks for an order-0 allocation request. If these
3563 * are not met, then a high-order request also cannot go ahead
3564 * even if a suitable page happened to be free.
3565 */
3566 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3567 return false;
3568
3569 /* If this is an order-0 request then the watermark is fine */
3570 if (!order)
3571 return true;
3572
3573 /* For a high-order request, check at least one suitable page is free */
3574 for (o = order; o < MAX_ORDER; o++) {
3575 struct free_area *area = &z->free_area[o];
3576 int mt;
3577
3578 if (!area->nr_free)
3579 continue;
3580
3581 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3582 if (!free_area_empty(area, mt))
3583 return true;
3584 }
3585
3586#ifdef CONFIG_CMA
3587 if ((alloc_flags & ALLOC_CMA) &&
3588 !free_area_empty(area, MIGRATE_CMA)) {
3589 return true;
3590 }
3591#endif
3592 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3593 return true;
3594 }
3595 return false;
3596}
3597
3598bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3599 int highest_zoneidx, unsigned int alloc_flags)
3600{
3601 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3602 zone_page_state(z, NR_FREE_PAGES));
3603}
3604
3605static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3606 unsigned long mark, int highest_zoneidx,
3607 unsigned int alloc_flags, gfp_t gfp_mask)
3608{
3609 long free_pages;
3610
3611 free_pages = zone_page_state(z, NR_FREE_PAGES);
3612
3613 /*
3614 * Fast check for order-0 only. If this fails then the reserves
3615 * need to be calculated.
3616 */
3617 if (!order) {
3618 long fast_free;
3619
3620 fast_free = free_pages;
3621 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3622 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3623 return true;
3624 }
3625
3626 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3627 free_pages))
3628 return true;
3629 /*
3630 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3631 * when checking the min watermark. The min watermark is the
3632 * point where boosting is ignored so that kswapd is woken up
3633 * when below the low watermark.
3634 */
3635 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3636 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3637 mark = z->_watermark[WMARK_MIN];
3638 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3639 alloc_flags, free_pages);
3640 }
3641
3642 return false;
3643}
3644
3645bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3646 unsigned long mark, int highest_zoneidx)
3647{
3648 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3649
3650 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3651 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3652
3653 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3654 free_pages);
3655}
3656
3657#ifdef CONFIG_NUMA
3658static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3659{
3660 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3661 node_reclaim_distance;
3662}
3663#else /* CONFIG_NUMA */
3664static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3665{
3666 return true;
3667}
3668#endif /* CONFIG_NUMA */
3669
3670/*
3671 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3672 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3673 * premature use of a lower zone may cause lowmem pressure problems that
3674 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3675 * probably too small. It only makes sense to spread allocations to avoid
3676 * fragmentation between the Normal and DMA32 zones.
3677 */
3678static inline unsigned int
3679alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3680{
3681 unsigned int alloc_flags;
3682
3683 /*
3684 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3685 * to save a branch.
3686 */
3687 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3688
3689#ifdef CONFIG_ZONE_DMA32
3690 if (!zone)
3691 return alloc_flags;
3692
3693 if (zone_idx(zone) != ZONE_NORMAL)
3694 return alloc_flags;
3695
3696 /*
3697 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3698 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3699 * on UMA that if Normal is populated then so is DMA32.
3700 */
3701 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3702 if (nr_online_nodes > 1 && !populated_zone(--zone))
3703 return alloc_flags;
3704
3705 alloc_flags |= ALLOC_NOFRAGMENT;
3706#endif /* CONFIG_ZONE_DMA32 */
3707 return alloc_flags;
3708}
3709
3710static inline unsigned int current_alloc_flags(gfp_t gfp_mask,
3711 unsigned int alloc_flags)
3712{
3713#ifdef CONFIG_CMA
3714 unsigned int pflags = current->flags;
3715
3716 if (!(pflags & PF_MEMALLOC_NOCMA) &&
3717 gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3718 alloc_flags |= ALLOC_CMA;
3719
3720#endif
3721 return alloc_flags;
3722}
3723
3724/*
3725 * get_page_from_freelist goes through the zonelist trying to allocate
3726 * a page.
3727 */
3728static struct page *
3729get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3730 const struct alloc_context *ac)
3731{
3732 struct zoneref *z;
3733 struct zone *zone;
3734 struct pglist_data *last_pgdat_dirty_limit = NULL;
3735 bool no_fallback;
3736
3737retry:
3738 /*
3739 * Scan zonelist, looking for a zone with enough free.
3740 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3741 */
3742 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3743 z = ac->preferred_zoneref;
3744 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist,
3745 ac->highest_zoneidx, ac->nodemask) {
3746 struct page *page;
3747 unsigned long mark;
3748
3749 if (cpusets_enabled() &&
3750 (alloc_flags & ALLOC_CPUSET) &&
3751 !__cpuset_zone_allowed(zone, gfp_mask))
3752 continue;
3753 /*
3754 * When allocating a page cache page for writing, we
3755 * want to get it from a node that is within its dirty
3756 * limit, such that no single node holds more than its
3757 * proportional share of globally allowed dirty pages.
3758 * The dirty limits take into account the node's
3759 * lowmem reserves and high watermark so that kswapd
3760 * should be able to balance it without having to
3761 * write pages from its LRU list.
3762 *
3763 * XXX: For now, allow allocations to potentially
3764 * exceed the per-node dirty limit in the slowpath
3765 * (spread_dirty_pages unset) before going into reclaim,
3766 * which is important when on a NUMA setup the allowed
3767 * nodes are together not big enough to reach the
3768 * global limit. The proper fix for these situations
3769 * will require awareness of nodes in the
3770 * dirty-throttling and the flusher threads.
3771 */
3772 if (ac->spread_dirty_pages) {
3773 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3774 continue;
3775
3776 if (!node_dirty_ok(zone->zone_pgdat)) {
3777 last_pgdat_dirty_limit = zone->zone_pgdat;
3778 continue;
3779 }
3780 }
3781
3782 if (no_fallback && nr_online_nodes > 1 &&
3783 zone != ac->preferred_zoneref->zone) {
3784 int local_nid;
3785
3786 /*
3787 * If moving to a remote node, retry but allow
3788 * fragmenting fallbacks. Locality is more important
3789 * than fragmentation avoidance.
3790 */
3791 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3792 if (zone_to_nid(zone) != local_nid) {
3793 alloc_flags &= ~ALLOC_NOFRAGMENT;
3794 goto retry;
3795 }
3796 }
3797
3798 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3799 if (!zone_watermark_fast(zone, order, mark,
3800 ac->highest_zoneidx, alloc_flags,
3801 gfp_mask)) {
3802 int ret;
3803
3804#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3805 /*
3806 * Watermark failed for this zone, but see if we can
3807 * grow this zone if it contains deferred pages.
3808 */
3809 if (static_branch_unlikely(&deferred_pages)) {
3810 if (_deferred_grow_zone(zone, order))
3811 goto try_this_zone;
3812 }
3813#endif
3814 /* Checked here to keep the fast path fast */
3815 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3816 if (alloc_flags & ALLOC_NO_WATERMARKS)
3817 goto try_this_zone;
3818
3819 if (node_reclaim_mode == 0 ||
3820 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3821 continue;
3822
3823 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3824 switch (ret) {
3825 case NODE_RECLAIM_NOSCAN:
3826 /* did not scan */
3827 continue;
3828 case NODE_RECLAIM_FULL:
3829 /* scanned but unreclaimable */
3830 continue;
3831 default:
3832 /* did we reclaim enough */
3833 if (zone_watermark_ok(zone, order, mark,
3834 ac->highest_zoneidx, alloc_flags))
3835 goto try_this_zone;
3836
3837 continue;
3838 }
3839 }
3840
3841try_this_zone:
3842 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3843 gfp_mask, alloc_flags, ac->migratetype);
3844 if (page) {
3845 prep_new_page(page, order, gfp_mask, alloc_flags);
3846
3847 /*
3848 * If this is a high-order atomic allocation then check
3849 * if the pageblock should be reserved for the future
3850 */
3851 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3852 reserve_highatomic_pageblock(page, zone, order);
3853
3854 return page;
3855 } else {
3856#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3857 /* Try again if zone has deferred pages */
3858 if (static_branch_unlikely(&deferred_pages)) {
3859 if (_deferred_grow_zone(zone, order))
3860 goto try_this_zone;
3861 }
3862#endif
3863 }
3864 }
3865
3866 /*
3867 * It's possible on a UMA machine to get through all zones that are
3868 * fragmented. If avoiding fragmentation, reset and try again.
3869 */
3870 if (no_fallback) {
3871 alloc_flags &= ~ALLOC_NOFRAGMENT;
3872 goto retry;
3873 }
3874
3875 return NULL;
3876}
3877
3878static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3879{
3880 unsigned int filter = SHOW_MEM_FILTER_NODES;
3881
3882 /*
3883 * This documents exceptions given to allocations in certain
3884 * contexts that are allowed to allocate outside current's set
3885 * of allowed nodes.
3886 */
3887 if (!(gfp_mask & __GFP_NOMEMALLOC))
3888 if (tsk_is_oom_victim(current) ||
3889 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3890 filter &= ~SHOW_MEM_FILTER_NODES;
3891 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3892 filter &= ~SHOW_MEM_FILTER_NODES;
3893
3894 show_mem(filter, nodemask);
3895}
3896
3897void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3898{
3899 struct va_format vaf;
3900 va_list args;
3901 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3902
3903 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3904 return;
3905
3906 va_start(args, fmt);
3907 vaf.fmt = fmt;
3908 vaf.va = &args;
3909 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3910 current->comm, &vaf, gfp_mask, &gfp_mask,
3911 nodemask_pr_args(nodemask));
3912 va_end(args);
3913
3914 cpuset_print_current_mems_allowed();
3915 pr_cont("\n");
3916 dump_stack();
3917 warn_alloc_show_mem(gfp_mask, nodemask);
3918}
3919
3920static inline struct page *
3921__alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3922 unsigned int alloc_flags,
3923 const struct alloc_context *ac)
3924{
3925 struct page *page;
3926
3927 page = get_page_from_freelist(gfp_mask, order,
3928 alloc_flags|ALLOC_CPUSET, ac);
3929 /*
3930 * fallback to ignore cpuset restriction if our nodes
3931 * are depleted
3932 */
3933 if (!page)
3934 page = get_page_from_freelist(gfp_mask, order,
3935 alloc_flags, ac);
3936
3937 return page;
3938}
3939
3940static inline struct page *
3941__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3942 const struct alloc_context *ac, unsigned long *did_some_progress)
3943{
3944 struct oom_control oc = {
3945 .zonelist = ac->zonelist,
3946 .nodemask = ac->nodemask,
3947 .memcg = NULL,
3948 .gfp_mask = gfp_mask,
3949 .order = order,
3950 };
3951 struct page *page;
3952
3953 *did_some_progress = 0;
3954
3955 /*
3956 * Acquire the oom lock. If that fails, somebody else is
3957 * making progress for us.
3958 */
3959 if (!mutex_trylock(&oom_lock)) {
3960 *did_some_progress = 1;
3961 schedule_timeout_uninterruptible(1);
3962 return NULL;
3963 }
3964
3965 /*
3966 * Go through the zonelist yet one more time, keep very high watermark
3967 * here, this is only to catch a parallel oom killing, we must fail if
3968 * we're still under heavy pressure. But make sure that this reclaim
3969 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3970 * allocation which will never fail due to oom_lock already held.
3971 */
3972 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3973 ~__GFP_DIRECT_RECLAIM, order,
3974 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3975 if (page)
3976 goto out;
3977
3978 /* Coredumps can quickly deplete all memory reserves */
3979 if (current->flags & PF_DUMPCORE)
3980 goto out;
3981 /* The OOM killer will not help higher order allocs */
3982 if (order > PAGE_ALLOC_COSTLY_ORDER)
3983 goto out;
3984 /*
3985 * We have already exhausted all our reclaim opportunities without any
3986 * success so it is time to admit defeat. We will skip the OOM killer
3987 * because it is very likely that the caller has a more reasonable
3988 * fallback than shooting a random task.
3989 */
3990 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3991 goto out;
3992 /* The OOM killer does not needlessly kill tasks for lowmem */
3993 if (ac->highest_zoneidx < ZONE_NORMAL)
3994 goto out;
3995 if (pm_suspended_storage())
3996 goto out;
3997 /*
3998 * XXX: GFP_NOFS allocations should rather fail than rely on
3999 * other request to make a forward progress.
4000 * We are in an unfortunate situation where out_of_memory cannot
4001 * do much for this context but let's try it to at least get
4002 * access to memory reserved if the current task is killed (see
4003 * out_of_memory). Once filesystems are ready to handle allocation
4004 * failures more gracefully we should just bail out here.
4005 */
4006
4007 /* The OOM killer may not free memory on a specific node */
4008 if (gfp_mask & __GFP_THISNODE)
4009 goto out;
4010
4011 /* Exhausted what can be done so it's blame time */
4012 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4013 *did_some_progress = 1;
4014
4015 /*
4016 * Help non-failing allocations by giving them access to memory
4017 * reserves
4018 */
4019 if (gfp_mask & __GFP_NOFAIL)
4020 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4021 ALLOC_NO_WATERMARKS, ac);
4022 }
4023out:
4024 mutex_unlock(&oom_lock);
4025 return page;
4026}
4027
4028/*
4029 * Maximum number of compaction retries wit a progress before OOM
4030 * killer is consider as the only way to move forward.
4031 */
4032#define MAX_COMPACT_RETRIES 16
4033
4034#ifdef CONFIG_COMPACTION
4035/* Try memory compaction for high-order allocations before reclaim */
4036static struct page *
4037__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4038 unsigned int alloc_flags, const struct alloc_context *ac,
4039 enum compact_priority prio, enum compact_result *compact_result)
4040{
4041 struct page *page = NULL;
4042 unsigned long pflags;
4043 unsigned int noreclaim_flag;
4044
4045 if (!order)
4046 return NULL;
4047
4048 psi_memstall_enter(&pflags);
4049 noreclaim_flag = memalloc_noreclaim_save();
4050
4051 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4052 prio, &page);
4053
4054 memalloc_noreclaim_restore(noreclaim_flag);
4055 psi_memstall_leave(&pflags);
4056
4057 /*
4058 * At least in one zone compaction wasn't deferred or skipped, so let's
4059 * count a compaction stall
4060 */
4061 count_vm_event(COMPACTSTALL);
4062
4063 /* Prep a captured page if available */
4064 if (page)
4065 prep_new_page(page, order, gfp_mask, alloc_flags);
4066
4067 /* Try get a page from the freelist if available */
4068 if (!page)
4069 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4070
4071 if (page) {
4072 struct zone *zone = page_zone(page);
4073
4074 zone->compact_blockskip_flush = false;
4075 compaction_defer_reset(zone, order, true);
4076 count_vm_event(COMPACTSUCCESS);
4077 return page;
4078 }
4079
4080 /*
4081 * It's bad if compaction run occurs and fails. The most likely reason
4082 * is that pages exist, but not enough to satisfy watermarks.
4083 */
4084 count_vm_event(COMPACTFAIL);
4085
4086 cond_resched();
4087
4088 return NULL;
4089}
4090
4091static inline bool
4092should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4093 enum compact_result compact_result,
4094 enum compact_priority *compact_priority,
4095 int *compaction_retries)
4096{
4097 int max_retries = MAX_COMPACT_RETRIES;
4098 int min_priority;
4099 bool ret = false;
4100 int retries = *compaction_retries;
4101 enum compact_priority priority = *compact_priority;
4102
4103 if (!order)
4104 return false;
4105
4106 if (compaction_made_progress(compact_result))
4107 (*compaction_retries)++;
4108
4109 /*
4110 * compaction considers all the zone as desperately out of memory
4111 * so it doesn't really make much sense to retry except when the
4112 * failure could be caused by insufficient priority
4113 */
4114 if (compaction_failed(compact_result))
4115 goto check_priority;
4116
4117 /*
4118 * compaction was skipped because there are not enough order-0 pages
4119 * to work with, so we retry only if it looks like reclaim can help.
4120 */
4121 if (compaction_needs_reclaim(compact_result)) {
4122 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4123 goto out;
4124 }
4125
4126 /*
4127 * make sure the compaction wasn't deferred or didn't bail out early
4128 * due to locks contention before we declare that we should give up.
4129 * But the next retry should use a higher priority if allowed, so
4130 * we don't just keep bailing out endlessly.
4131 */
4132 if (compaction_withdrawn(compact_result)) {
4133 goto check_priority;
4134 }
4135
4136 /*
4137 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4138 * costly ones because they are de facto nofail and invoke OOM
4139 * killer to move on while costly can fail and users are ready
4140 * to cope with that. 1/4 retries is rather arbitrary but we
4141 * would need much more detailed feedback from compaction to
4142 * make a better decision.
4143 */
4144 if (order > PAGE_ALLOC_COSTLY_ORDER)
4145 max_retries /= 4;
4146 if (*compaction_retries <= max_retries) {
4147 ret = true;
4148 goto out;
4149 }
4150
4151 /*
4152 * Make sure there are attempts at the highest priority if we exhausted
4153 * all retries or failed at the lower priorities.
4154 */
4155check_priority:
4156 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4157 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4158
4159 if (*compact_priority > min_priority) {
4160 (*compact_priority)--;
4161 *compaction_retries = 0;
4162 ret = true;
4163 }
4164out:
4165 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4166 return ret;
4167}
4168#else
4169static inline struct page *
4170__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4171 unsigned int alloc_flags, const struct alloc_context *ac,
4172 enum compact_priority prio, enum compact_result *compact_result)
4173{
4174 *compact_result = COMPACT_SKIPPED;
4175 return NULL;
4176}
4177
4178static inline bool
4179should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4180 enum compact_result compact_result,
4181 enum compact_priority *compact_priority,
4182 int *compaction_retries)
4183{
4184 struct zone *zone;
4185 struct zoneref *z;
4186
4187 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4188 return false;
4189
4190 /*
4191 * There are setups with compaction disabled which would prefer to loop
4192 * inside the allocator rather than hit the oom killer prematurely.
4193 * Let's give them a good hope and keep retrying while the order-0
4194 * watermarks are OK.
4195 */
4196 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4197 ac->highest_zoneidx, ac->nodemask) {
4198 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4199 ac->highest_zoneidx, alloc_flags))
4200 return true;
4201 }
4202 return false;
4203}
4204#endif /* CONFIG_COMPACTION */
4205
4206#ifdef CONFIG_LOCKDEP
4207static struct lockdep_map __fs_reclaim_map =
4208 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4209
4210static bool __need_fs_reclaim(gfp_t gfp_mask)
4211{
4212 gfp_mask = current_gfp_context(gfp_mask);
4213
4214 /* no reclaim without waiting on it */
4215 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4216 return false;
4217
4218 /* this guy won't enter reclaim */
4219 if (current->flags & PF_MEMALLOC)
4220 return false;
4221
4222 /* We're only interested __GFP_FS allocations for now */
4223 if (!(gfp_mask & __GFP_FS))
4224 return false;
4225
4226 if (gfp_mask & __GFP_NOLOCKDEP)
4227 return false;
4228
4229 return true;
4230}
4231
4232void __fs_reclaim_acquire(void)
4233{
4234 lock_map_acquire(&__fs_reclaim_map);
4235}
4236
4237void __fs_reclaim_release(void)
4238{
4239 lock_map_release(&__fs_reclaim_map);
4240}
4241
4242void fs_reclaim_acquire(gfp_t gfp_mask)
4243{
4244 if (__need_fs_reclaim(gfp_mask))
4245 __fs_reclaim_acquire();
4246}
4247EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4248
4249void fs_reclaim_release(gfp_t gfp_mask)
4250{
4251 if (__need_fs_reclaim(gfp_mask))
4252 __fs_reclaim_release();
4253}
4254EXPORT_SYMBOL_GPL(fs_reclaim_release);
4255#endif
4256
4257/* Perform direct synchronous page reclaim */
4258static int
4259__perform_reclaim(gfp_t gfp_mask, unsigned int order,
4260 const struct alloc_context *ac)
4261{
4262 int progress;
4263 unsigned int noreclaim_flag;
4264 unsigned long pflags;
4265
4266 cond_resched();
4267
4268 /* We now go into synchronous reclaim */
4269 cpuset_memory_pressure_bump();
4270 psi_memstall_enter(&pflags);
4271 fs_reclaim_acquire(gfp_mask);
4272 noreclaim_flag = memalloc_noreclaim_save();
4273
4274 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4275 ac->nodemask);
4276
4277 memalloc_noreclaim_restore(noreclaim_flag);
4278 fs_reclaim_release(gfp_mask);
4279 psi_memstall_leave(&pflags);
4280
4281 cond_resched();
4282
4283 return progress;
4284}
4285
4286/* The really slow allocator path where we enter direct reclaim */
4287static inline struct page *
4288__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4289 unsigned int alloc_flags, const struct alloc_context *ac,
4290 unsigned long *did_some_progress)
4291{
4292 struct page *page = NULL;
4293 bool drained = false;
4294
4295 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4296 if (unlikely(!(*did_some_progress)))
4297 return NULL;
4298
4299retry:
4300 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4301
4302 /*
4303 * If an allocation failed after direct reclaim, it could be because
4304 * pages are pinned on the per-cpu lists or in high alloc reserves.
4305 * Shrink them and try again
4306 */
4307 if (!page && !drained) {
4308 unreserve_highatomic_pageblock(ac, false);
4309 drain_all_pages(NULL);
4310 drained = true;
4311 goto retry;
4312 }
4313
4314 return page;
4315}
4316
4317static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4318 const struct alloc_context *ac)
4319{
4320 struct zoneref *z;
4321 struct zone *zone;
4322 pg_data_t *last_pgdat = NULL;
4323 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4324
4325 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4326 ac->nodemask) {
4327 if (last_pgdat != zone->zone_pgdat)
4328 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4329 last_pgdat = zone->zone_pgdat;
4330 }
4331}
4332
4333static inline unsigned int
4334gfp_to_alloc_flags(gfp_t gfp_mask)
4335{
4336 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4337
4338 /*
4339 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4340 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4341 * to save two branches.
4342 */
4343 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4344 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4345
4346 /*
4347 * The caller may dip into page reserves a bit more if the caller
4348 * cannot run direct reclaim, or if the caller has realtime scheduling
4349 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4350 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4351 */
4352 alloc_flags |= (__force int)
4353 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4354
4355 if (gfp_mask & __GFP_ATOMIC) {
4356 /*
4357 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4358 * if it can't schedule.
4359 */
4360 if (!(gfp_mask & __GFP_NOMEMALLOC))
4361 alloc_flags |= ALLOC_HARDER;
4362 /*
4363 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4364 * comment for __cpuset_node_allowed().
4365 */
4366 alloc_flags &= ~ALLOC_CPUSET;
4367 } else if (unlikely(rt_task(current)) && !in_interrupt())
4368 alloc_flags |= ALLOC_HARDER;
4369
4370 alloc_flags = current_alloc_flags(gfp_mask, alloc_flags);
4371
4372 return alloc_flags;
4373}
4374
4375static bool oom_reserves_allowed(struct task_struct *tsk)
4376{
4377 if (!tsk_is_oom_victim(tsk))
4378 return false;
4379
4380 /*
4381 * !MMU doesn't have oom reaper so give access to memory reserves
4382 * only to the thread with TIF_MEMDIE set
4383 */
4384 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4385 return false;
4386
4387 return true;
4388}
4389
4390/*
4391 * Distinguish requests which really need access to full memory
4392 * reserves from oom victims which can live with a portion of it
4393 */
4394static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4395{
4396 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4397 return 0;
4398 if (gfp_mask & __GFP_MEMALLOC)
4399 return ALLOC_NO_WATERMARKS;
4400 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4401 return ALLOC_NO_WATERMARKS;
4402 if (!in_interrupt()) {
4403 if (current->flags & PF_MEMALLOC)
4404 return ALLOC_NO_WATERMARKS;
4405 else if (oom_reserves_allowed(current))
4406 return ALLOC_OOM;
4407 }
4408
4409 return 0;
4410}
4411
4412bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4413{
4414 return !!__gfp_pfmemalloc_flags(gfp_mask);
4415}
4416
4417/*
4418 * Checks whether it makes sense to retry the reclaim to make a forward progress
4419 * for the given allocation request.
4420 *
4421 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4422 * without success, or when we couldn't even meet the watermark if we
4423 * reclaimed all remaining pages on the LRU lists.
4424 *
4425 * Returns true if a retry is viable or false to enter the oom path.
4426 */
4427static inline bool
4428should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4429 struct alloc_context *ac, int alloc_flags,
4430 bool did_some_progress, int *no_progress_loops)
4431{
4432 struct zone *zone;
4433 struct zoneref *z;
4434 bool ret = false;
4435
4436 /*
4437 * Costly allocations might have made a progress but this doesn't mean
4438 * their order will become available due to high fragmentation so
4439 * always increment the no progress counter for them
4440 */
4441 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4442 *no_progress_loops = 0;
4443 else
4444 (*no_progress_loops)++;
4445
4446 /*
4447 * Make sure we converge to OOM if we cannot make any progress
4448 * several times in the row.
4449 */
4450 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4451 /* Before OOM, exhaust highatomic_reserve */
4452 return unreserve_highatomic_pageblock(ac, true);
4453 }
4454
4455 /*
4456 * Keep reclaiming pages while there is a chance this will lead
4457 * somewhere. If none of the target zones can satisfy our allocation
4458 * request even if all reclaimable pages are considered then we are
4459 * screwed and have to go OOM.
4460 */
4461 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4462 ac->highest_zoneidx, ac->nodemask) {
4463 unsigned long available;
4464 unsigned long reclaimable;
4465 unsigned long min_wmark = min_wmark_pages(zone);
4466 bool wmark;
4467
4468 available = reclaimable = zone_reclaimable_pages(zone);
4469 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4470
4471 /*
4472 * Would the allocation succeed if we reclaimed all
4473 * reclaimable pages?
4474 */
4475 wmark = __zone_watermark_ok(zone, order, min_wmark,
4476 ac->highest_zoneidx, alloc_flags, available);
4477 trace_reclaim_retry_zone(z, order, reclaimable,
4478 available, min_wmark, *no_progress_loops, wmark);
4479 if (wmark) {
4480 /*
4481 * If we didn't make any progress and have a lot of
4482 * dirty + writeback pages then we should wait for
4483 * an IO to complete to slow down the reclaim and
4484 * prevent from pre mature OOM
4485 */
4486 if (!did_some_progress) {
4487 unsigned long write_pending;
4488
4489 write_pending = zone_page_state_snapshot(zone,
4490 NR_ZONE_WRITE_PENDING);
4491
4492 if (2 * write_pending > reclaimable) {
4493 congestion_wait(BLK_RW_ASYNC, HZ/10);
4494 return true;
4495 }
4496 }
4497
4498 ret = true;
4499 goto out;
4500 }
4501 }
4502
4503out:
4504 /*
4505 * Memory allocation/reclaim might be called from a WQ context and the
4506 * current implementation of the WQ concurrency control doesn't
4507 * recognize that a particular WQ is congested if the worker thread is
4508 * looping without ever sleeping. Therefore we have to do a short sleep
4509 * here rather than calling cond_resched().
4510 */
4511 if (current->flags & PF_WQ_WORKER)
4512 schedule_timeout_uninterruptible(1);
4513 else
4514 cond_resched();
4515 return ret;
4516}
4517
4518static inline bool
4519check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4520{
4521 /*
4522 * It's possible that cpuset's mems_allowed and the nodemask from
4523 * mempolicy don't intersect. This should be normally dealt with by
4524 * policy_nodemask(), but it's possible to race with cpuset update in
4525 * such a way the check therein was true, and then it became false
4526 * before we got our cpuset_mems_cookie here.
4527 * This assumes that for all allocations, ac->nodemask can come only
4528 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4529 * when it does not intersect with the cpuset restrictions) or the
4530 * caller can deal with a violated nodemask.
4531 */
4532 if (cpusets_enabled() && ac->nodemask &&
4533 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4534 ac->nodemask = NULL;
4535 return true;
4536 }
4537
4538 /*
4539 * When updating a task's mems_allowed or mempolicy nodemask, it is
4540 * possible to race with parallel threads in such a way that our
4541 * allocation can fail while the mask is being updated. If we are about
4542 * to fail, check if the cpuset changed during allocation and if so,
4543 * retry.
4544 */
4545 if (read_mems_allowed_retry(cpuset_mems_cookie))
4546 return true;
4547
4548 return false;
4549}
4550
4551static inline struct page *
4552__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4553 struct alloc_context *ac)
4554{
4555 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4556 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4557 struct page *page = NULL;
4558 unsigned int alloc_flags;
4559 unsigned long did_some_progress;
4560 enum compact_priority compact_priority;
4561 enum compact_result compact_result;
4562 int compaction_retries;
4563 int no_progress_loops;
4564 unsigned int cpuset_mems_cookie;
4565 int reserve_flags;
4566
4567 /*
4568 * We also sanity check to catch abuse of atomic reserves being used by
4569 * callers that are not in atomic context.
4570 */
4571 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4572 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4573 gfp_mask &= ~__GFP_ATOMIC;
4574
4575retry_cpuset:
4576 compaction_retries = 0;
4577 no_progress_loops = 0;
4578 compact_priority = DEF_COMPACT_PRIORITY;
4579 cpuset_mems_cookie = read_mems_allowed_begin();
4580
4581 /*
4582 * The fast path uses conservative alloc_flags to succeed only until
4583 * kswapd needs to be woken up, and to avoid the cost of setting up
4584 * alloc_flags precisely. So we do that now.
4585 */
4586 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4587
4588 /*
4589 * We need to recalculate the starting point for the zonelist iterator
4590 * because we might have used different nodemask in the fast path, or
4591 * there was a cpuset modification and we are retrying - otherwise we
4592 * could end up iterating over non-eligible zones endlessly.
4593 */
4594 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4595 ac->highest_zoneidx, ac->nodemask);
4596 if (!ac->preferred_zoneref->zone)
4597 goto nopage;
4598
4599 if (alloc_flags & ALLOC_KSWAPD)
4600 wake_all_kswapds(order, gfp_mask, ac);
4601
4602 /*
4603 * The adjusted alloc_flags might result in immediate success, so try
4604 * that first
4605 */
4606 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4607 if (page)
4608 goto got_pg;
4609
4610 /*
4611 * For costly allocations, try direct compaction first, as it's likely
4612 * that we have enough base pages and don't need to reclaim. For non-
4613 * movable high-order allocations, do that as well, as compaction will
4614 * try prevent permanent fragmentation by migrating from blocks of the
4615 * same migratetype.
4616 * Don't try this for allocations that are allowed to ignore
4617 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4618 */
4619 if (can_direct_reclaim &&
4620 (costly_order ||
4621 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4622 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4623 page = __alloc_pages_direct_compact(gfp_mask, order,
4624 alloc_flags, ac,
4625 INIT_COMPACT_PRIORITY,
4626 &compact_result);
4627 if (page)
4628 goto got_pg;
4629
4630 /*
4631 * Checks for costly allocations with __GFP_NORETRY, which
4632 * includes some THP page fault allocations
4633 */
4634 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4635 /*
4636 * If allocating entire pageblock(s) and compaction
4637 * failed because all zones are below low watermarks
4638 * or is prohibited because it recently failed at this
4639 * order, fail immediately unless the allocator has
4640 * requested compaction and reclaim retry.
4641 *
4642 * Reclaim is
4643 * - potentially very expensive because zones are far
4644 * below their low watermarks or this is part of very
4645 * bursty high order allocations,
4646 * - not guaranteed to help because isolate_freepages()
4647 * may not iterate over freed pages as part of its
4648 * linear scan, and
4649 * - unlikely to make entire pageblocks free on its
4650 * own.
4651 */
4652 if (compact_result == COMPACT_SKIPPED ||
4653 compact_result == COMPACT_DEFERRED)
4654 goto nopage;
4655
4656 /*
4657 * Looks like reclaim/compaction is worth trying, but
4658 * sync compaction could be very expensive, so keep
4659 * using async compaction.
4660 */
4661 compact_priority = INIT_COMPACT_PRIORITY;
4662 }
4663 }
4664
4665retry:
4666 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4667 if (alloc_flags & ALLOC_KSWAPD)
4668 wake_all_kswapds(order, gfp_mask, ac);
4669
4670 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4671 if (reserve_flags)
4672 alloc_flags = current_alloc_flags(gfp_mask, reserve_flags);
4673
4674 /*
4675 * Reset the nodemask and zonelist iterators if memory policies can be
4676 * ignored. These allocations are high priority and system rather than
4677 * user oriented.
4678 */
4679 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4680 ac->nodemask = NULL;
4681 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4682 ac->highest_zoneidx, ac->nodemask);
4683 }
4684
4685 /* Attempt with potentially adjusted zonelist and alloc_flags */
4686 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4687 if (page)
4688 goto got_pg;
4689
4690 /* Caller is not willing to reclaim, we can't balance anything */
4691 if (!can_direct_reclaim)
4692 goto nopage;
4693
4694 /* Avoid recursion of direct reclaim */
4695 if (current->flags & PF_MEMALLOC)
4696 goto nopage;
4697
4698 /* Try direct reclaim and then allocating */
4699 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4700 &did_some_progress);
4701 if (page)
4702 goto got_pg;
4703
4704 /* Try direct compaction and then allocating */
4705 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4706 compact_priority, &compact_result);
4707 if (page)
4708 goto got_pg;
4709
4710 /* Do not loop if specifically requested */
4711 if (gfp_mask & __GFP_NORETRY)
4712 goto nopage;
4713
4714 /*
4715 * Do not retry costly high order allocations unless they are
4716 * __GFP_RETRY_MAYFAIL
4717 */
4718 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4719 goto nopage;
4720
4721 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4722 did_some_progress > 0, &no_progress_loops))
4723 goto retry;
4724
4725 /*
4726 * It doesn't make any sense to retry for the compaction if the order-0
4727 * reclaim is not able to make any progress because the current
4728 * implementation of the compaction depends on the sufficient amount
4729 * of free memory (see __compaction_suitable)
4730 */
4731 if (did_some_progress > 0 &&
4732 should_compact_retry(ac, order, alloc_flags,
4733 compact_result, &compact_priority,
4734 &compaction_retries))
4735 goto retry;
4736
4737
4738 /* Deal with possible cpuset update races before we start OOM killing */
4739 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4740 goto retry_cpuset;
4741
4742 /* Reclaim has failed us, start killing things */
4743 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4744 if (page)
4745 goto got_pg;
4746
4747 /* Avoid allocations with no watermarks from looping endlessly */
4748 if (tsk_is_oom_victim(current) &&
4749 (alloc_flags & ALLOC_OOM ||
4750 (gfp_mask & __GFP_NOMEMALLOC)))
4751 goto nopage;
4752
4753 /* Retry as long as the OOM killer is making progress */
4754 if (did_some_progress) {
4755 no_progress_loops = 0;
4756 goto retry;
4757 }
4758
4759nopage:
4760 /* Deal with possible cpuset update races before we fail */
4761 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4762 goto retry_cpuset;
4763
4764 /*
4765 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4766 * we always retry
4767 */
4768 if (gfp_mask & __GFP_NOFAIL) {
4769 /*
4770 * All existing users of the __GFP_NOFAIL are blockable, so warn
4771 * of any new users that actually require GFP_NOWAIT
4772 */
4773 if (WARN_ON_ONCE(!can_direct_reclaim))
4774 goto fail;
4775
4776 /*
4777 * PF_MEMALLOC request from this context is rather bizarre
4778 * because we cannot reclaim anything and only can loop waiting
4779 * for somebody to do a work for us
4780 */
4781 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4782
4783 /*
4784 * non failing costly orders are a hard requirement which we
4785 * are not prepared for much so let's warn about these users
4786 * so that we can identify them and convert them to something
4787 * else.
4788 */
4789 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4790
4791 /*
4792 * Help non-failing allocations by giving them access to memory
4793 * reserves but do not use ALLOC_NO_WATERMARKS because this
4794 * could deplete whole memory reserves which would just make
4795 * the situation worse
4796 */
4797 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4798 if (page)
4799 goto got_pg;
4800
4801 cond_resched();
4802 goto retry;
4803 }
4804fail:
4805 warn_alloc(gfp_mask, ac->nodemask,
4806 "page allocation failure: order:%u", order);
4807got_pg:
4808 return page;
4809}
4810
4811static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4812 int preferred_nid, nodemask_t *nodemask,
4813 struct alloc_context *ac, gfp_t *alloc_mask,
4814 unsigned int *alloc_flags)
4815{
4816 ac->highest_zoneidx = gfp_zone(gfp_mask);
4817 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4818 ac->nodemask = nodemask;
4819 ac->migratetype = gfp_migratetype(gfp_mask);
4820
4821 if (cpusets_enabled()) {
4822 *alloc_mask |= __GFP_HARDWALL;
4823 /*
4824 * When we are in the interrupt context, it is irrelevant
4825 * to the current task context. It means that any node ok.
4826 */
4827 if (!in_interrupt() && !ac->nodemask)
4828 ac->nodemask = &cpuset_current_mems_allowed;
4829 else
4830 *alloc_flags |= ALLOC_CPUSET;
4831 }
4832
4833 fs_reclaim_acquire(gfp_mask);
4834 fs_reclaim_release(gfp_mask);
4835
4836 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4837
4838 if (should_fail_alloc_page(gfp_mask, order))
4839 return false;
4840
4841 *alloc_flags = current_alloc_flags(gfp_mask, *alloc_flags);
4842
4843 return true;
4844}
4845
4846/* Determine whether to spread dirty pages and what the first usable zone */
4847static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4848{
4849 /* Dirty zone balancing only done in the fast path */
4850 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4851
4852 /*
4853 * The preferred zone is used for statistics but crucially it is
4854 * also used as the starting point for the zonelist iterator. It
4855 * may get reset for allocations that ignore memory policies.
4856 */
4857 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4858 ac->highest_zoneidx, ac->nodemask);
4859}
4860
4861/*
4862 * This is the 'heart' of the zoned buddy allocator.
4863 */
4864struct page *
4865__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4866 nodemask_t *nodemask)
4867{
4868 struct page *page;
4869 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4870 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4871 struct alloc_context ac = { };
4872
4873 /*
4874 * There are several places where we assume that the order value is sane
4875 * so bail out early if the request is out of bound.
4876 */
4877 if (unlikely(order >= MAX_ORDER)) {
4878 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4879 return NULL;
4880 }
4881
4882 gfp_mask &= gfp_allowed_mask;
4883 alloc_mask = gfp_mask;
4884 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4885 return NULL;
4886
4887 finalise_ac(gfp_mask, &ac);
4888
4889 /*
4890 * Forbid the first pass from falling back to types that fragment
4891 * memory until all local zones are considered.
4892 */
4893 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4894
4895 /* First allocation attempt */
4896 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4897 if (likely(page))
4898 goto out;
4899
4900 /*
4901 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4902 * resp. GFP_NOIO which has to be inherited for all allocation requests
4903 * from a particular context which has been marked by
4904 * memalloc_no{fs,io}_{save,restore}.
4905 */
4906 alloc_mask = current_gfp_context(gfp_mask);
4907 ac.spread_dirty_pages = false;
4908
4909 /*
4910 * Restore the original nodemask if it was potentially replaced with
4911 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4912 */
4913 ac.nodemask = nodemask;
4914
4915 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4916
4917out:
4918 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4919 unlikely(__memcg_kmem_charge_page(page, gfp_mask, order) != 0)) {
4920 __free_pages(page, order);
4921 page = NULL;
4922 }
4923
4924 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4925
4926 return page;
4927}
4928EXPORT_SYMBOL(__alloc_pages_nodemask);
4929
4930/*
4931 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4932 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4933 * you need to access high mem.
4934 */
4935unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4936{
4937 struct page *page;
4938
4939 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4940 if (!page)
4941 return 0;
4942 return (unsigned long) page_address(page);
4943}
4944EXPORT_SYMBOL(__get_free_pages);
4945
4946unsigned long get_zeroed_page(gfp_t gfp_mask)
4947{
4948 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4949}
4950EXPORT_SYMBOL(get_zeroed_page);
4951
4952static inline void free_the_page(struct page *page, unsigned int order)
4953{
4954 if (order == 0) /* Via pcp? */
4955 free_unref_page(page);
4956 else
4957 __free_pages_ok(page, order);
4958}
4959
4960void __free_pages(struct page *page, unsigned int order)
4961{
4962 if (put_page_testzero(page))
4963 free_the_page(page, order);
4964}
4965EXPORT_SYMBOL(__free_pages);
4966
4967void free_pages(unsigned long addr, unsigned int order)
4968{
4969 if (addr != 0) {
4970 VM_BUG_ON(!virt_addr_valid((void *)addr));
4971 __free_pages(virt_to_page((void *)addr), order);
4972 }
4973}
4974
4975EXPORT_SYMBOL(free_pages);
4976
4977/*
4978 * Page Fragment:
4979 * An arbitrary-length arbitrary-offset area of memory which resides
4980 * within a 0 or higher order page. Multiple fragments within that page
4981 * are individually refcounted, in the page's reference counter.
4982 *
4983 * The page_frag functions below provide a simple allocation framework for
4984 * page fragments. This is used by the network stack and network device
4985 * drivers to provide a backing region of memory for use as either an
4986 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4987 */
4988static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4989 gfp_t gfp_mask)
4990{
4991 struct page *page = NULL;
4992 gfp_t gfp = gfp_mask;
4993
4994#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4995 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4996 __GFP_NOMEMALLOC;
4997 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4998 PAGE_FRAG_CACHE_MAX_ORDER);
4999 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5000#endif
5001 if (unlikely(!page))
5002 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5003
5004 nc->va = page ? page_address(page) : NULL;
5005
5006 return page;
5007}
5008
5009void __page_frag_cache_drain(struct page *page, unsigned int count)
5010{
5011 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5012
5013 if (page_ref_sub_and_test(page, count))
5014 free_the_page(page, compound_order(page));
5015}
5016EXPORT_SYMBOL(__page_frag_cache_drain);
5017
5018void *page_frag_alloc(struct page_frag_cache *nc,
5019 unsigned int fragsz, gfp_t gfp_mask)
5020{
5021 unsigned int size = PAGE_SIZE;
5022 struct page *page;
5023 int offset;
5024
5025 if (unlikely(!nc->va)) {
5026refill:
5027 page = __page_frag_cache_refill(nc, gfp_mask);
5028 if (!page)
5029 return NULL;
5030
5031#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5032 /* if size can vary use size else just use PAGE_SIZE */
5033 size = nc->size;
5034#endif
5035 /* Even if we own the page, we do not use atomic_set().
5036 * This would break get_page_unless_zero() users.
5037 */
5038 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5039
5040 /* reset page count bias and offset to start of new frag */
5041 nc->pfmemalloc = page_is_pfmemalloc(page);
5042 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5043 nc->offset = size;
5044 }
5045
5046 offset = nc->offset - fragsz;
5047 if (unlikely(offset < 0)) {
5048 page = virt_to_page(nc->va);
5049
5050 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5051 goto refill;
5052
5053#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5054 /* if size can vary use size else just use PAGE_SIZE */
5055 size = nc->size;
5056#endif
5057 /* OK, page count is 0, we can safely set it */
5058 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5059
5060 /* reset page count bias and offset to start of new frag */
5061 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5062 offset = size - fragsz;
5063 }
5064
5065 nc->pagecnt_bias--;
5066 nc->offset = offset;
5067
5068 return nc->va + offset;
5069}
5070EXPORT_SYMBOL(page_frag_alloc);
5071
5072/*
5073 * Frees a page fragment allocated out of either a compound or order 0 page.
5074 */
5075void page_frag_free(void *addr)
5076{
5077 struct page *page = virt_to_head_page(addr);
5078
5079 if (unlikely(put_page_testzero(page)))
5080 free_the_page(page, compound_order(page));
5081}
5082EXPORT_SYMBOL(page_frag_free);
5083
5084static void *make_alloc_exact(unsigned long addr, unsigned int order,
5085 size_t size)
5086{
5087 if (addr) {
5088 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5089 unsigned long used = addr + PAGE_ALIGN(size);
5090
5091 split_page(virt_to_page((void *)addr), order);
5092 while (used < alloc_end) {
5093 free_page(used);
5094 used += PAGE_SIZE;
5095 }
5096 }
5097 return (void *)addr;
5098}
5099
5100/**
5101 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5102 * @size: the number of bytes to allocate
5103 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5104 *
5105 * This function is similar to alloc_pages(), except that it allocates the
5106 * minimum number of pages to satisfy the request. alloc_pages() can only
5107 * allocate memory in power-of-two pages.
5108 *
5109 * This function is also limited by MAX_ORDER.
5110 *
5111 * Memory allocated by this function must be released by free_pages_exact().
5112 *
5113 * Return: pointer to the allocated area or %NULL in case of error.
5114 */
5115void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5116{
5117 unsigned int order = get_order(size);
5118 unsigned long addr;
5119
5120 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5121 gfp_mask &= ~__GFP_COMP;
5122
5123 addr = __get_free_pages(gfp_mask, order);
5124 return make_alloc_exact(addr, order, size);
5125}
5126EXPORT_SYMBOL(alloc_pages_exact);
5127
5128/**
5129 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5130 * pages on a node.
5131 * @nid: the preferred node ID where memory should be allocated
5132 * @size: the number of bytes to allocate
5133 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5134 *
5135 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5136 * back.
5137 *
5138 * Return: pointer to the allocated area or %NULL in case of error.
5139 */
5140void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5141{
5142 unsigned int order = get_order(size);
5143 struct page *p;
5144
5145 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5146 gfp_mask &= ~__GFP_COMP;
5147
5148 p = alloc_pages_node(nid, gfp_mask, order);
5149 if (!p)
5150 return NULL;
5151 return make_alloc_exact((unsigned long)page_address(p), order, size);
5152}
5153
5154/**
5155 * free_pages_exact - release memory allocated via alloc_pages_exact()
5156 * @virt: the value returned by alloc_pages_exact.
5157 * @size: size of allocation, same value as passed to alloc_pages_exact().
5158 *
5159 * Release the memory allocated by a previous call to alloc_pages_exact.
5160 */
5161void free_pages_exact(void *virt, size_t size)
5162{
5163 unsigned long addr = (unsigned long)virt;
5164 unsigned long end = addr + PAGE_ALIGN(size);
5165
5166 while (addr < end) {
5167 free_page(addr);
5168 addr += PAGE_SIZE;
5169 }
5170}
5171EXPORT_SYMBOL(free_pages_exact);
5172
5173/**
5174 * nr_free_zone_pages - count number of pages beyond high watermark
5175 * @offset: The zone index of the highest zone
5176 *
5177 * nr_free_zone_pages() counts the number of pages which are beyond the
5178 * high watermark within all zones at or below a given zone index. For each
5179 * zone, the number of pages is calculated as:
5180 *
5181 * nr_free_zone_pages = managed_pages - high_pages
5182 *
5183 * Return: number of pages beyond high watermark.
5184 */
5185static unsigned long nr_free_zone_pages(int offset)
5186{
5187 struct zoneref *z;
5188 struct zone *zone;
5189
5190 /* Just pick one node, since fallback list is circular */
5191 unsigned long sum = 0;
5192
5193 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5194
5195 for_each_zone_zonelist(zone, z, zonelist, offset) {
5196 unsigned long size = zone_managed_pages(zone);
5197 unsigned long high = high_wmark_pages(zone);
5198 if (size > high)
5199 sum += size - high;
5200 }
5201
5202 return sum;
5203}
5204
5205/**
5206 * nr_free_buffer_pages - count number of pages beyond high watermark
5207 *
5208 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5209 * watermark within ZONE_DMA and ZONE_NORMAL.
5210 *
5211 * Return: number of pages beyond high watermark within ZONE_DMA and
5212 * ZONE_NORMAL.
5213 */
5214unsigned long nr_free_buffer_pages(void)
5215{
5216 return nr_free_zone_pages(gfp_zone(GFP_USER));
5217}
5218EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5219
5220static inline void show_node(struct zone *zone)
5221{
5222 if (IS_ENABLED(CONFIG_NUMA))
5223 printk("Node %d ", zone_to_nid(zone));
5224}
5225
5226long si_mem_available(void)
5227{
5228 long available;
5229 unsigned long pagecache;
5230 unsigned long wmark_low = 0;
5231 unsigned long pages[NR_LRU_LISTS];
5232 unsigned long reclaimable;
5233 struct zone *zone;
5234 int lru;
5235
5236 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5237 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5238
5239 for_each_zone(zone)
5240 wmark_low += low_wmark_pages(zone);
5241
5242 /*
5243 * Estimate the amount of memory available for userspace allocations,
5244 * without causing swapping.
5245 */
5246 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5247
5248 /*
5249 * Not all the page cache can be freed, otherwise the system will
5250 * start swapping. Assume at least half of the page cache, or the
5251 * low watermark worth of cache, needs to stay.
5252 */
5253 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5254 pagecache -= min(pagecache / 2, wmark_low);
5255 available += pagecache;
5256
5257 /*
5258 * Part of the reclaimable slab and other kernel memory consists of
5259 * items that are in use, and cannot be freed. Cap this estimate at the
5260 * low watermark.
5261 */
5262 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5263 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5264 available += reclaimable - min(reclaimable / 2, wmark_low);
5265
5266 if (available < 0)
5267 available = 0;
5268 return available;
5269}
5270EXPORT_SYMBOL_GPL(si_mem_available);
5271
5272void si_meminfo(struct sysinfo *val)
5273{
5274 val->totalram = totalram_pages();
5275 val->sharedram = global_node_page_state(NR_SHMEM);
5276 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5277 val->bufferram = nr_blockdev_pages();
5278 val->totalhigh = totalhigh_pages();
5279 val->freehigh = nr_free_highpages();
5280 val->mem_unit = PAGE_SIZE;
5281}
5282
5283EXPORT_SYMBOL(si_meminfo);
5284
5285#ifdef CONFIG_NUMA
5286void si_meminfo_node(struct sysinfo *val, int nid)
5287{
5288 int zone_type; /* needs to be signed */
5289 unsigned long managed_pages = 0;
5290 unsigned long managed_highpages = 0;
5291 unsigned long free_highpages = 0;
5292 pg_data_t *pgdat = NODE_DATA(nid);
5293
5294 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5295 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5296 val->totalram = managed_pages;
5297 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5298 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5299#ifdef CONFIG_HIGHMEM
5300 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5301 struct zone *zone = &pgdat->node_zones[zone_type];
5302
5303 if (is_highmem(zone)) {
5304 managed_highpages += zone_managed_pages(zone);
5305 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5306 }
5307 }
5308 val->totalhigh = managed_highpages;
5309 val->freehigh = free_highpages;
5310#else
5311 val->totalhigh = managed_highpages;
5312 val->freehigh = free_highpages;
5313#endif
5314 val->mem_unit = PAGE_SIZE;
5315}
5316#endif
5317
5318/*
5319 * Determine whether the node should be displayed or not, depending on whether
5320 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5321 */
5322static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5323{
5324 if (!(flags & SHOW_MEM_FILTER_NODES))
5325 return false;
5326
5327 /*
5328 * no node mask - aka implicit memory numa policy. Do not bother with
5329 * the synchronization - read_mems_allowed_begin - because we do not
5330 * have to be precise here.
5331 */
5332 if (!nodemask)
5333 nodemask = &cpuset_current_mems_allowed;
5334
5335 return !node_isset(nid, *nodemask);
5336}
5337
5338#define K(x) ((x) << (PAGE_SHIFT-10))
5339
5340static void show_migration_types(unsigned char type)
5341{
5342 static const char types[MIGRATE_TYPES] = {
5343 [MIGRATE_UNMOVABLE] = 'U',
5344 [MIGRATE_MOVABLE] = 'M',
5345 [MIGRATE_RECLAIMABLE] = 'E',
5346 [MIGRATE_HIGHATOMIC] = 'H',
5347#ifdef CONFIG_CMA
5348 [MIGRATE_CMA] = 'C',
5349#endif
5350#ifdef CONFIG_MEMORY_ISOLATION
5351 [MIGRATE_ISOLATE] = 'I',
5352#endif
5353 };
5354 char tmp[MIGRATE_TYPES + 1];
5355 char *p = tmp;
5356 int i;
5357
5358 for (i = 0; i < MIGRATE_TYPES; i++) {
5359 if (type & (1 << i))
5360 *p++ = types[i];
5361 }
5362
5363 *p = '\0';
5364 printk(KERN_CONT "(%s) ", tmp);
5365}
5366
5367/*
5368 * Show free area list (used inside shift_scroll-lock stuff)
5369 * We also calculate the percentage fragmentation. We do this by counting the
5370 * memory on each free list with the exception of the first item on the list.
5371 *
5372 * Bits in @filter:
5373 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5374 * cpuset.
5375 */
5376void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5377{
5378 unsigned long free_pcp = 0;
5379 int cpu;
5380 struct zone *zone;
5381 pg_data_t *pgdat;
5382
5383 for_each_populated_zone(zone) {
5384 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5385 continue;
5386
5387 for_each_online_cpu(cpu)
5388 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5389 }
5390
5391 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5392 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5393 " unevictable:%lu dirty:%lu writeback:%lu\n"
5394 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5395 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5396 " free:%lu free_pcp:%lu free_cma:%lu\n",
5397 global_node_page_state(NR_ACTIVE_ANON),
5398 global_node_page_state(NR_INACTIVE_ANON),
5399 global_node_page_state(NR_ISOLATED_ANON),
5400 global_node_page_state(NR_ACTIVE_FILE),
5401 global_node_page_state(NR_INACTIVE_FILE),
5402 global_node_page_state(NR_ISOLATED_FILE),
5403 global_node_page_state(NR_UNEVICTABLE),
5404 global_node_page_state(NR_FILE_DIRTY),
5405 global_node_page_state(NR_WRITEBACK),
5406 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5407 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5408 global_node_page_state(NR_FILE_MAPPED),
5409 global_node_page_state(NR_SHMEM),
5410 global_zone_page_state(NR_PAGETABLE),
5411 global_zone_page_state(NR_BOUNCE),
5412 global_zone_page_state(NR_FREE_PAGES),
5413 free_pcp,
5414 global_zone_page_state(NR_FREE_CMA_PAGES));
5415
5416 for_each_online_pgdat(pgdat) {
5417 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5418 continue;
5419
5420 printk("Node %d"
5421 " active_anon:%lukB"
5422 " inactive_anon:%lukB"
5423 " active_file:%lukB"
5424 " inactive_file:%lukB"
5425 " unevictable:%lukB"
5426 " isolated(anon):%lukB"
5427 " isolated(file):%lukB"
5428 " mapped:%lukB"
5429 " dirty:%lukB"
5430 " writeback:%lukB"
5431 " shmem:%lukB"
5432#ifdef CONFIG_TRANSPARENT_HUGEPAGE
5433 " shmem_thp: %lukB"
5434 " shmem_pmdmapped: %lukB"
5435 " anon_thp: %lukB"
5436#endif
5437 " writeback_tmp:%lukB"
5438 " kernel_stack:%lukB"
5439#ifdef CONFIG_SHADOW_CALL_STACK
5440 " shadow_call_stack:%lukB"
5441#endif
5442 " all_unreclaimable? %s"
5443 "\n",
5444 pgdat->node_id,
5445 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5446 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5447 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5448 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5449 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5450 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5451 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5452 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5453 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5454 K(node_page_state(pgdat, NR_WRITEBACK)),
5455 K(node_page_state(pgdat, NR_SHMEM)),
5456#ifdef CONFIG_TRANSPARENT_HUGEPAGE
5457 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5458 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5459 * HPAGE_PMD_NR),
5460 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5461#endif
5462 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5463 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5464#ifdef CONFIG_SHADOW_CALL_STACK
5465 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5466#endif
5467 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5468 "yes" : "no");
5469 }
5470
5471 for_each_populated_zone(zone) {
5472 int i;
5473
5474 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5475 continue;
5476
5477 free_pcp = 0;
5478 for_each_online_cpu(cpu)
5479 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5480
5481 show_node(zone);
5482 printk(KERN_CONT
5483 "%s"
5484 " free:%lukB"
5485 " min:%lukB"
5486 " low:%lukB"
5487 " high:%lukB"
5488 " reserved_highatomic:%luKB"
5489 " active_anon:%lukB"
5490 " inactive_anon:%lukB"
5491 " active_file:%lukB"
5492 " inactive_file:%lukB"
5493 " unevictable:%lukB"
5494 " writepending:%lukB"
5495 " present:%lukB"
5496 " managed:%lukB"
5497 " mlocked:%lukB"
5498 " pagetables:%lukB"
5499 " bounce:%lukB"
5500 " free_pcp:%lukB"
5501 " local_pcp:%ukB"
5502 " free_cma:%lukB"
5503 "\n",
5504 zone->name,
5505 K(zone_page_state(zone, NR_FREE_PAGES)),
5506 K(min_wmark_pages(zone)),
5507 K(low_wmark_pages(zone)),
5508 K(high_wmark_pages(zone)),
5509 K(zone->nr_reserved_highatomic),
5510 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5511 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5512 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5513 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5514 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5515 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5516 K(zone->present_pages),
5517 K(zone_managed_pages(zone)),
5518 K(zone_page_state(zone, NR_MLOCK)),
5519 K(zone_page_state(zone, NR_PAGETABLE)),
5520 K(zone_page_state(zone, NR_BOUNCE)),
5521 K(free_pcp),
5522 K(this_cpu_read(zone->pageset->pcp.count)),
5523 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5524 printk("lowmem_reserve[]:");
5525 for (i = 0; i < MAX_NR_ZONES; i++)
5526 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5527 printk(KERN_CONT "\n");
5528 }
5529
5530 for_each_populated_zone(zone) {
5531 unsigned int order;
5532 unsigned long nr[MAX_ORDER], flags, total = 0;
5533 unsigned char types[MAX_ORDER];
5534
5535 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5536 continue;
5537 show_node(zone);
5538 printk(KERN_CONT "%s: ", zone->name);
5539
5540 spin_lock_irqsave(&zone->lock, flags);
5541 for (order = 0; order < MAX_ORDER; order++) {
5542 struct free_area *area = &zone->free_area[order];
5543 int type;
5544
5545 nr[order] = area->nr_free;
5546 total += nr[order] << order;
5547
5548 types[order] = 0;
5549 for (type = 0; type < MIGRATE_TYPES; type++) {
5550 if (!free_area_empty(area, type))
5551 types[order] |= 1 << type;
5552 }
5553 }
5554 spin_unlock_irqrestore(&zone->lock, flags);
5555 for (order = 0; order < MAX_ORDER; order++) {
5556 printk(KERN_CONT "%lu*%lukB ",
5557 nr[order], K(1UL) << order);
5558 if (nr[order])
5559 show_migration_types(types[order]);
5560 }
5561 printk(KERN_CONT "= %lukB\n", K(total));
5562 }
5563
5564 hugetlb_show_meminfo();
5565
5566 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5567
5568 show_swap_cache_info();
5569}
5570
5571static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5572{
5573 zoneref->zone = zone;
5574 zoneref->zone_idx = zone_idx(zone);
5575}
5576
5577/*
5578 * Builds allocation fallback zone lists.
5579 *
5580 * Add all populated zones of a node to the zonelist.
5581 */
5582static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5583{
5584 struct zone *zone;
5585 enum zone_type zone_type = MAX_NR_ZONES;
5586 int nr_zones = 0;
5587
5588 do {
5589 zone_type--;
5590 zone = pgdat->node_zones + zone_type;
5591 if (managed_zone(zone)) {
5592 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5593 check_highest_zone(zone_type);
5594 }
5595 } while (zone_type);
5596
5597 return nr_zones;
5598}
5599
5600#ifdef CONFIG_NUMA
5601
5602static int __parse_numa_zonelist_order(char *s)
5603{
5604 /*
5605 * We used to support different zonlists modes but they turned
5606 * out to be just not useful. Let's keep the warning in place
5607 * if somebody still use the cmd line parameter so that we do
5608 * not fail it silently
5609 */
5610 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5611 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5612 return -EINVAL;
5613 }
5614 return 0;
5615}
5616
5617char numa_zonelist_order[] = "Node";
5618
5619/*
5620 * sysctl handler for numa_zonelist_order
5621 */
5622int numa_zonelist_order_handler(struct ctl_table *table, int write,
5623 void *buffer, size_t *length, loff_t *ppos)
5624{
5625 if (write)
5626 return __parse_numa_zonelist_order(buffer);
5627 return proc_dostring(table, write, buffer, length, ppos);
5628}
5629
5630
5631#define MAX_NODE_LOAD (nr_online_nodes)
5632static int node_load[MAX_NUMNODES];
5633
5634/**
5635 * find_next_best_node - find the next node that should appear in a given node's fallback list
5636 * @node: node whose fallback list we're appending
5637 * @used_node_mask: nodemask_t of already used nodes
5638 *
5639 * We use a number of factors to determine which is the next node that should
5640 * appear on a given node's fallback list. The node should not have appeared
5641 * already in @node's fallback list, and it should be the next closest node
5642 * according to the distance array (which contains arbitrary distance values
5643 * from each node to each node in the system), and should also prefer nodes
5644 * with no CPUs, since presumably they'll have very little allocation pressure
5645 * on them otherwise.
5646 *
5647 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5648 */
5649static int find_next_best_node(int node, nodemask_t *used_node_mask)
5650{
5651 int n, val;
5652 int min_val = INT_MAX;
5653 int best_node = NUMA_NO_NODE;
5654 const struct cpumask *tmp = cpumask_of_node(0);
5655
5656 /* Use the local node if we haven't already */
5657 if (!node_isset(node, *used_node_mask)) {
5658 node_set(node, *used_node_mask);
5659 return node;
5660 }
5661
5662 for_each_node_state(n, N_MEMORY) {
5663
5664 /* Don't want a node to appear more than once */
5665 if (node_isset(n, *used_node_mask))
5666 continue;
5667
5668 /* Use the distance array to find the distance */
5669 val = node_distance(node, n);
5670
5671 /* Penalize nodes under us ("prefer the next node") */
5672 val += (n < node);
5673
5674 /* Give preference to headless and unused nodes */
5675 tmp = cpumask_of_node(n);
5676 if (!cpumask_empty(tmp))
5677 val += PENALTY_FOR_NODE_WITH_CPUS;
5678
5679 /* Slight preference for less loaded node */
5680 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5681 val += node_load[n];
5682
5683 if (val < min_val) {
5684 min_val = val;
5685 best_node = n;
5686 }
5687 }
5688
5689 if (best_node >= 0)
5690 node_set(best_node, *used_node_mask);
5691
5692 return best_node;
5693}
5694
5695
5696/*
5697 * Build zonelists ordered by node and zones within node.
5698 * This results in maximum locality--normal zone overflows into local
5699 * DMA zone, if any--but risks exhausting DMA zone.
5700 */
5701static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5702 unsigned nr_nodes)
5703{
5704 struct zoneref *zonerefs;
5705 int i;
5706
5707 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5708
5709 for (i = 0; i < nr_nodes; i++) {
5710 int nr_zones;
5711
5712 pg_data_t *node = NODE_DATA(node_order[i]);
5713
5714 nr_zones = build_zonerefs_node(node, zonerefs);
5715 zonerefs += nr_zones;
5716 }
5717 zonerefs->zone = NULL;
5718 zonerefs->zone_idx = 0;
5719}
5720
5721/*
5722 * Build gfp_thisnode zonelists
5723 */
5724static void build_thisnode_zonelists(pg_data_t *pgdat)
5725{
5726 struct zoneref *zonerefs;
5727 int nr_zones;
5728
5729 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5730 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5731 zonerefs += nr_zones;
5732 zonerefs->zone = NULL;
5733 zonerefs->zone_idx = 0;
5734}
5735
5736/*
5737 * Build zonelists ordered by zone and nodes within zones.
5738 * This results in conserving DMA zone[s] until all Normal memory is
5739 * exhausted, but results in overflowing to remote node while memory
5740 * may still exist in local DMA zone.
5741 */
5742
5743static void build_zonelists(pg_data_t *pgdat)
5744{
5745 static int node_order[MAX_NUMNODES];
5746 int node, load, nr_nodes = 0;
5747 nodemask_t used_mask = NODE_MASK_NONE;
5748 int local_node, prev_node;
5749
5750 /* NUMA-aware ordering of nodes */
5751 local_node = pgdat->node_id;
5752 load = nr_online_nodes;
5753 prev_node = local_node;
5754
5755 memset(node_order, 0, sizeof(node_order));
5756 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5757 /*
5758 * We don't want to pressure a particular node.
5759 * So adding penalty to the first node in same
5760 * distance group to make it round-robin.
5761 */
5762 if (node_distance(local_node, node) !=
5763 node_distance(local_node, prev_node))
5764 node_load[node] = load;
5765
5766 node_order[nr_nodes++] = node;
5767 prev_node = node;
5768 load--;
5769 }
5770
5771 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5772 build_thisnode_zonelists(pgdat);
5773}
5774
5775#ifdef CONFIG_HAVE_MEMORYLESS_NODES
5776/*
5777 * Return node id of node used for "local" allocations.
5778 * I.e., first node id of first zone in arg node's generic zonelist.
5779 * Used for initializing percpu 'numa_mem', which is used primarily
5780 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5781 */
5782int local_memory_node(int node)
5783{
5784 struct zoneref *z;
5785
5786 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5787 gfp_zone(GFP_KERNEL),
5788 NULL);
5789 return zone_to_nid(z->zone);
5790}
5791#endif
5792
5793static void setup_min_unmapped_ratio(void);
5794static void setup_min_slab_ratio(void);
5795#else /* CONFIG_NUMA */
5796
5797static void build_zonelists(pg_data_t *pgdat)
5798{
5799 int node, local_node;
5800 struct zoneref *zonerefs;
5801 int nr_zones;
5802
5803 local_node = pgdat->node_id;
5804
5805 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5806 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5807 zonerefs += nr_zones;
5808
5809 /*
5810 * Now we build the zonelist so that it contains the zones
5811 * of all the other nodes.
5812 * We don't want to pressure a particular node, so when
5813 * building the zones for node N, we make sure that the
5814 * zones coming right after the local ones are those from
5815 * node N+1 (modulo N)
5816 */
5817 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5818 if (!node_online(node))
5819 continue;
5820 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5821 zonerefs += nr_zones;
5822 }
5823 for (node = 0; node < local_node; node++) {
5824 if (!node_online(node))
5825 continue;
5826 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5827 zonerefs += nr_zones;
5828 }
5829
5830 zonerefs->zone = NULL;
5831 zonerefs->zone_idx = 0;
5832}
5833
5834#endif /* CONFIG_NUMA */
5835
5836/*
5837 * Boot pageset table. One per cpu which is going to be used for all
5838 * zones and all nodes. The parameters will be set in such a way
5839 * that an item put on a list will immediately be handed over to
5840 * the buddy list. This is safe since pageset manipulation is done
5841 * with interrupts disabled.
5842 *
5843 * The boot_pagesets must be kept even after bootup is complete for
5844 * unused processors and/or zones. They do play a role for bootstrapping
5845 * hotplugged processors.
5846 *
5847 * zoneinfo_show() and maybe other functions do
5848 * not check if the processor is online before following the pageset pointer.
5849 * Other parts of the kernel may not check if the zone is available.
5850 */
5851static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5852static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5853static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5854
5855static void __build_all_zonelists(void *data)
5856{
5857 int nid;
5858 int __maybe_unused cpu;
5859 pg_data_t *self = data;
5860 static DEFINE_SPINLOCK(lock);
5861
5862 spin_lock(&lock);
5863
5864#ifdef CONFIG_NUMA
5865 memset(node_load, 0, sizeof(node_load));
5866#endif
5867
5868 /*
5869 * This node is hotadded and no memory is yet present. So just
5870 * building zonelists is fine - no need to touch other nodes.
5871 */
5872 if (self && !node_online(self->node_id)) {
5873 build_zonelists(self);
5874 } else {
5875 for_each_online_node(nid) {
5876 pg_data_t *pgdat = NODE_DATA(nid);
5877
5878 build_zonelists(pgdat);
5879 }
5880
5881#ifdef CONFIG_HAVE_MEMORYLESS_NODES
5882 /*
5883 * We now know the "local memory node" for each node--
5884 * i.e., the node of the first zone in the generic zonelist.
5885 * Set up numa_mem percpu variable for on-line cpus. During
5886 * boot, only the boot cpu should be on-line; we'll init the
5887 * secondary cpus' numa_mem as they come on-line. During
5888 * node/memory hotplug, we'll fixup all on-line cpus.
5889 */
5890 for_each_online_cpu(cpu)
5891 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5892#endif
5893 }
5894
5895 spin_unlock(&lock);
5896}
5897
5898static noinline void __init
5899build_all_zonelists_init(void)
5900{
5901 int cpu;
5902
5903 __build_all_zonelists(NULL);
5904
5905 /*
5906 * Initialize the boot_pagesets that are going to be used
5907 * for bootstrapping processors. The real pagesets for
5908 * each zone will be allocated later when the per cpu
5909 * allocator is available.
5910 *
5911 * boot_pagesets are used also for bootstrapping offline
5912 * cpus if the system is already booted because the pagesets
5913 * are needed to initialize allocators on a specific cpu too.
5914 * F.e. the percpu allocator needs the page allocator which
5915 * needs the percpu allocator in order to allocate its pagesets
5916 * (a chicken-egg dilemma).
5917 */
5918 for_each_possible_cpu(cpu)
5919 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5920
5921 mminit_verify_zonelist();
5922 cpuset_init_current_mems_allowed();
5923}
5924
5925/*
5926 * unless system_state == SYSTEM_BOOTING.
5927 *
5928 * __ref due to call of __init annotated helper build_all_zonelists_init
5929 * [protected by SYSTEM_BOOTING].
5930 */
5931void __ref build_all_zonelists(pg_data_t *pgdat)
5932{
5933 unsigned long vm_total_pages;
5934
5935 if (system_state == SYSTEM_BOOTING) {
5936 build_all_zonelists_init();
5937 } else {
5938 __build_all_zonelists(pgdat);
5939 /* cpuset refresh routine should be here */
5940 }
5941 /* Get the number of free pages beyond high watermark in all zones. */
5942 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5943 /*
5944 * Disable grouping by mobility if the number of pages in the
5945 * system is too low to allow the mechanism to work. It would be
5946 * more accurate, but expensive to check per-zone. This check is
5947 * made on memory-hotadd so a system can start with mobility
5948 * disabled and enable it later
5949 */
5950 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5951 page_group_by_mobility_disabled = 1;
5952 else
5953 page_group_by_mobility_disabled = 0;
5954
5955 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5956 nr_online_nodes,
5957 page_group_by_mobility_disabled ? "off" : "on",
5958 vm_total_pages);
5959#ifdef CONFIG_NUMA
5960 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5961#endif
5962}
5963
5964/* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5965static bool __meminit
5966overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5967{
5968 static struct memblock_region *r;
5969
5970 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5971 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5972 for_each_memblock(memory, r) {
5973 if (*pfn < memblock_region_memory_end_pfn(r))
5974 break;
5975 }
5976 }
5977 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5978 memblock_is_mirror(r)) {
5979 *pfn = memblock_region_memory_end_pfn(r);
5980 return true;
5981 }
5982 }
5983 return false;
5984}
5985
5986/*
5987 * Initially all pages are reserved - free ones are freed
5988 * up by memblock_free_all() once the early boot process is
5989 * done. Non-atomic initialization, single-pass.
5990 */
5991void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5992 unsigned long start_pfn, enum meminit_context context,
5993 struct vmem_altmap *altmap)
5994{
5995 unsigned long pfn, end_pfn = start_pfn + size;
5996 struct page *page;
5997
5998 if (highest_memmap_pfn < end_pfn - 1)
5999 highest_memmap_pfn = end_pfn - 1;
6000
6001#ifdef CONFIG_ZONE_DEVICE
6002 /*
6003 * Honor reservation requested by the driver for this ZONE_DEVICE
6004 * memory. We limit the total number of pages to initialize to just
6005 * those that might contain the memory mapping. We will defer the
6006 * ZONE_DEVICE page initialization until after we have released
6007 * the hotplug lock.
6008 */
6009 if (zone == ZONE_DEVICE) {
6010 if (!altmap)
6011 return;
6012
6013 if (start_pfn == altmap->base_pfn)
6014 start_pfn += altmap->reserve;
6015 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6016 }
6017#endif
6018
6019 for (pfn = start_pfn; pfn < end_pfn; ) {
6020 /*
6021 * There can be holes in boot-time mem_map[]s handed to this
6022 * function. They do not exist on hotplugged memory.
6023 */
6024 if (context == MEMINIT_EARLY) {
6025 if (overlap_memmap_init(zone, &pfn))
6026 continue;
6027 if (defer_init(nid, pfn, end_pfn))
6028 break;
6029 }
6030
6031 page = pfn_to_page(pfn);
6032 __init_single_page(page, pfn, zone, nid);
6033 if (context == MEMINIT_HOTPLUG)
6034 __SetPageReserved(page);
6035
6036 /*
6037 * Mark the block movable so that blocks are reserved for
6038 * movable at startup. This will force kernel allocations
6039 * to reserve their blocks rather than leaking throughout
6040 * the address space during boot when many long-lived
6041 * kernel allocations are made.
6042 *
6043 * bitmap is created for zone's valid pfn range. but memmap
6044 * can be created for invalid pages (for alignment)
6045 * check here not to call set_pageblock_migratetype() against
6046 * pfn out of zone.
6047 */
6048 if (!(pfn & (pageblock_nr_pages - 1))) {
6049 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6050 cond_resched();
6051 }
6052 pfn++;
6053 }
6054}
6055
6056#ifdef CONFIG_ZONE_DEVICE
6057void __ref memmap_init_zone_device(struct zone *zone,
6058 unsigned long start_pfn,
6059 unsigned long nr_pages,
6060 struct dev_pagemap *pgmap)
6061{
6062 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6063 struct pglist_data *pgdat = zone->zone_pgdat;
6064 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6065 unsigned long zone_idx = zone_idx(zone);
6066 unsigned long start = jiffies;
6067 int nid = pgdat->node_id;
6068
6069 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6070 return;
6071
6072 /*
6073 * The call to memmap_init_zone should have already taken care
6074 * of the pages reserved for the memmap, so we can just jump to
6075 * the end of that region and start processing the device pages.
6076 */
6077 if (altmap) {
6078 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6079 nr_pages = end_pfn - start_pfn;
6080 }
6081
6082 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6083 struct page *page = pfn_to_page(pfn);
6084
6085 __init_single_page(page, pfn, zone_idx, nid);
6086
6087 /*
6088 * Mark page reserved as it will need to wait for onlining
6089 * phase for it to be fully associated with a zone.
6090 *
6091 * We can use the non-atomic __set_bit operation for setting
6092 * the flag as we are still initializing the pages.
6093 */
6094 __SetPageReserved(page);
6095
6096 /*
6097 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6098 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6099 * ever freed or placed on a driver-private list.
6100 */
6101 page->pgmap = pgmap;
6102 page->zone_device_data = NULL;
6103
6104 /*
6105 * Mark the block movable so that blocks are reserved for
6106 * movable at startup. This will force kernel allocations
6107 * to reserve their blocks rather than leaking throughout
6108 * the address space during boot when many long-lived
6109 * kernel allocations are made.
6110 *
6111 * bitmap is created for zone's valid pfn range. but memmap
6112 * can be created for invalid pages (for alignment)
6113 * check here not to call set_pageblock_migratetype() against
6114 * pfn out of zone.
6115 *
6116 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6117 * because this is done early in section_activate()
6118 */
6119 if (!(pfn & (pageblock_nr_pages - 1))) {
6120 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6121 cond_resched();
6122 }
6123 }
6124
6125 pr_info("%s initialised %lu pages in %ums\n", __func__,
6126 nr_pages, jiffies_to_msecs(jiffies - start));
6127}
6128
6129#endif
6130static void __meminit zone_init_free_lists(struct zone *zone)
6131{
6132 unsigned int order, t;
6133 for_each_migratetype_order(order, t) {
6134 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6135 zone->free_area[order].nr_free = 0;
6136 }
6137}
6138
6139void __meminit __weak memmap_init(unsigned long size, int nid,
6140 unsigned long zone,
6141 unsigned long range_start_pfn)
6142{
6143 unsigned long start_pfn, end_pfn;
6144 unsigned long range_end_pfn = range_start_pfn + size;
6145 int i;
6146
6147 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6148 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6149 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6150
6151 if (end_pfn > start_pfn) {
6152 size = end_pfn - start_pfn;
6153 memmap_init_zone(size, nid, zone, start_pfn,
6154 MEMINIT_EARLY, NULL);
6155 }
6156 }
6157}
6158
6159static int zone_batchsize(struct zone *zone)
6160{
6161#ifdef CONFIG_MMU
6162 int batch;
6163
6164 /*
6165 * The per-cpu-pages pools are set to around 1000th of the
6166 * size of the zone.
6167 */
6168 batch = zone_managed_pages(zone) / 1024;
6169 /* But no more than a meg. */
6170 if (batch * PAGE_SIZE > 1024 * 1024)
6171 batch = (1024 * 1024) / PAGE_SIZE;
6172 batch /= 4; /* We effectively *= 4 below */
6173 if (batch < 1)
6174 batch = 1;
6175
6176 /*
6177 * Clamp the batch to a 2^n - 1 value. Having a power
6178 * of 2 value was found to be more likely to have
6179 * suboptimal cache aliasing properties in some cases.
6180 *
6181 * For example if 2 tasks are alternately allocating
6182 * batches of pages, one task can end up with a lot
6183 * of pages of one half of the possible page colors
6184 * and the other with pages of the other colors.
6185 */
6186 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6187
6188 return batch;
6189
6190#else
6191 /* The deferral and batching of frees should be suppressed under NOMMU
6192 * conditions.
6193 *
6194 * The problem is that NOMMU needs to be able to allocate large chunks
6195 * of contiguous memory as there's no hardware page translation to
6196 * assemble apparent contiguous memory from discontiguous pages.
6197 *
6198 * Queueing large contiguous runs of pages for batching, however,
6199 * causes the pages to actually be freed in smaller chunks. As there
6200 * can be a significant delay between the individual batches being
6201 * recycled, this leads to the once large chunks of space being
6202 * fragmented and becoming unavailable for high-order allocations.
6203 */
6204 return 0;
6205#endif
6206}
6207
6208/*
6209 * pcp->high and pcp->batch values are related and dependent on one another:
6210 * ->batch must never be higher then ->high.
6211 * The following function updates them in a safe manner without read side
6212 * locking.
6213 *
6214 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6215 * those fields changing asynchronously (acording to the above rule).
6216 *
6217 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6218 * outside of boot time (or some other assurance that no concurrent updaters
6219 * exist).
6220 */
6221static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6222 unsigned long batch)
6223{
6224 /* start with a fail safe value for batch */
6225 pcp->batch = 1;
6226 smp_wmb();
6227
6228 /* Update high, then batch, in order */
6229 pcp->high = high;
6230 smp_wmb();
6231
6232 pcp->batch = batch;
6233}
6234
6235/* a companion to pageset_set_high() */
6236static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6237{
6238 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6239}
6240
6241static void pageset_init(struct per_cpu_pageset *p)
6242{
6243 struct per_cpu_pages *pcp;
6244 int migratetype;
6245
6246 memset(p, 0, sizeof(*p));
6247
6248 pcp = &p->pcp;
6249 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6250 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6251}
6252
6253static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6254{
6255 pageset_init(p);
6256 pageset_set_batch(p, batch);
6257}
6258
6259/*
6260 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6261 * to the value high for the pageset p.
6262 */
6263static void pageset_set_high(struct per_cpu_pageset *p,
6264 unsigned long high)
6265{
6266 unsigned long batch = max(1UL, high / 4);
6267 if ((high / 4) > (PAGE_SHIFT * 8))
6268 batch = PAGE_SHIFT * 8;
6269
6270 pageset_update(&p->pcp, high, batch);
6271}
6272
6273static void pageset_set_high_and_batch(struct zone *zone,
6274 struct per_cpu_pageset *pcp)
6275{
6276 if (percpu_pagelist_fraction)
6277 pageset_set_high(pcp,
6278 (zone_managed_pages(zone) /
6279 percpu_pagelist_fraction));
6280 else
6281 pageset_set_batch(pcp, zone_batchsize(zone));
6282}
6283
6284static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6285{
6286 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6287
6288 pageset_init(pcp);
6289 pageset_set_high_and_batch(zone, pcp);
6290}
6291
6292void __meminit setup_zone_pageset(struct zone *zone)
6293{
6294 int cpu;
6295 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6296 for_each_possible_cpu(cpu)
6297 zone_pageset_init(zone, cpu);
6298}
6299
6300/*
6301 * Allocate per cpu pagesets and initialize them.
6302 * Before this call only boot pagesets were available.
6303 */
6304void __init setup_per_cpu_pageset(void)
6305{
6306 struct pglist_data *pgdat;
6307 struct zone *zone;
6308 int __maybe_unused cpu;
6309
6310 for_each_populated_zone(zone)
6311 setup_zone_pageset(zone);
6312
6313#ifdef CONFIG_NUMA
6314 /*
6315 * Unpopulated zones continue using the boot pagesets.
6316 * The numa stats for these pagesets need to be reset.
6317 * Otherwise, they will end up skewing the stats of
6318 * the nodes these zones are associated with.
6319 */
6320 for_each_possible_cpu(cpu) {
6321 struct per_cpu_pageset *pcp = &per_cpu(boot_pageset, cpu);
6322 memset(pcp->vm_numa_stat_diff, 0,
6323 sizeof(pcp->vm_numa_stat_diff));
6324 }
6325#endif
6326
6327 for_each_online_pgdat(pgdat)
6328 pgdat->per_cpu_nodestats =
6329 alloc_percpu(struct per_cpu_nodestat);
6330}
6331
6332static __meminit void zone_pcp_init(struct zone *zone)
6333{
6334 /*
6335 * per cpu subsystem is not up at this point. The following code
6336 * relies on the ability of the linker to provide the
6337 * offset of a (static) per cpu variable into the per cpu area.
6338 */
6339 zone->pageset = &boot_pageset;
6340
6341 if (populated_zone(zone))
6342 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6343 zone->name, zone->present_pages,
6344 zone_batchsize(zone));
6345}
6346
6347void __meminit init_currently_empty_zone(struct zone *zone,
6348 unsigned long zone_start_pfn,
6349 unsigned long size)
6350{
6351 struct pglist_data *pgdat = zone->zone_pgdat;
6352 int zone_idx = zone_idx(zone) + 1;
6353
6354 if (zone_idx > pgdat->nr_zones)
6355 pgdat->nr_zones = zone_idx;
6356
6357 zone->zone_start_pfn = zone_start_pfn;
6358
6359 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6360 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6361 pgdat->node_id,
6362 (unsigned long)zone_idx(zone),
6363 zone_start_pfn, (zone_start_pfn + size));
6364
6365 zone_init_free_lists(zone);
6366 zone->initialized = 1;
6367}
6368
6369/**
6370 * get_pfn_range_for_nid - Return the start and end page frames for a node
6371 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6372 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6373 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6374 *
6375 * It returns the start and end page frame of a node based on information
6376 * provided by memblock_set_node(). If called for a node
6377 * with no available memory, a warning is printed and the start and end
6378 * PFNs will be 0.
6379 */
6380void __init get_pfn_range_for_nid(unsigned int nid,
6381 unsigned long *start_pfn, unsigned long *end_pfn)
6382{
6383 unsigned long this_start_pfn, this_end_pfn;
6384 int i;
6385
6386 *start_pfn = -1UL;
6387 *end_pfn = 0;
6388
6389 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6390 *start_pfn = min(*start_pfn, this_start_pfn);
6391 *end_pfn = max(*end_pfn, this_end_pfn);
6392 }
6393
6394 if (*start_pfn == -1UL)
6395 *start_pfn = 0;
6396}
6397
6398/*
6399 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6400 * assumption is made that zones within a node are ordered in monotonic
6401 * increasing memory addresses so that the "highest" populated zone is used
6402 */
6403static void __init find_usable_zone_for_movable(void)
6404{
6405 int zone_index;
6406 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6407 if (zone_index == ZONE_MOVABLE)
6408 continue;
6409
6410 if (arch_zone_highest_possible_pfn[zone_index] >
6411 arch_zone_lowest_possible_pfn[zone_index])
6412 break;
6413 }
6414
6415 VM_BUG_ON(zone_index == -1);
6416 movable_zone = zone_index;
6417}
6418
6419/*
6420 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6421 * because it is sized independent of architecture. Unlike the other zones,
6422 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6423 * in each node depending on the size of each node and how evenly kernelcore
6424 * is distributed. This helper function adjusts the zone ranges
6425 * provided by the architecture for a given node by using the end of the
6426 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6427 * zones within a node are in order of monotonic increases memory addresses
6428 */
6429static void __init adjust_zone_range_for_zone_movable(int nid,
6430 unsigned long zone_type,
6431 unsigned long node_start_pfn,
6432 unsigned long node_end_pfn,
6433 unsigned long *zone_start_pfn,
6434 unsigned long *zone_end_pfn)
6435{
6436 /* Only adjust if ZONE_MOVABLE is on this node */
6437 if (zone_movable_pfn[nid]) {
6438 /* Size ZONE_MOVABLE */
6439 if (zone_type == ZONE_MOVABLE) {
6440 *zone_start_pfn = zone_movable_pfn[nid];
6441 *zone_end_pfn = min(node_end_pfn,
6442 arch_zone_highest_possible_pfn[movable_zone]);
6443
6444 /* Adjust for ZONE_MOVABLE starting within this range */
6445 } else if (!mirrored_kernelcore &&
6446 *zone_start_pfn < zone_movable_pfn[nid] &&
6447 *zone_end_pfn > zone_movable_pfn[nid]) {
6448 *zone_end_pfn = zone_movable_pfn[nid];
6449
6450 /* Check if this whole range is within ZONE_MOVABLE */
6451 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6452 *zone_start_pfn = *zone_end_pfn;
6453 }
6454}
6455
6456/*
6457 * Return the number of pages a zone spans in a node, including holes
6458 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6459 */
6460static unsigned long __init zone_spanned_pages_in_node(int nid,
6461 unsigned long zone_type,
6462 unsigned long node_start_pfn,
6463 unsigned long node_end_pfn,
6464 unsigned long *zone_start_pfn,
6465 unsigned long *zone_end_pfn)
6466{
6467 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6468 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6469 /* When hotadd a new node from cpu_up(), the node should be empty */
6470 if (!node_start_pfn && !node_end_pfn)
6471 return 0;
6472
6473 /* Get the start and end of the zone */
6474 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6475 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6476 adjust_zone_range_for_zone_movable(nid, zone_type,
6477 node_start_pfn, node_end_pfn,
6478 zone_start_pfn, zone_end_pfn);
6479
6480 /* Check that this node has pages within the zone's required range */
6481 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6482 return 0;
6483
6484 /* Move the zone boundaries inside the node if necessary */
6485 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6486 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6487
6488 /* Return the spanned pages */
6489 return *zone_end_pfn - *zone_start_pfn;
6490}
6491
6492/*
6493 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6494 * then all holes in the requested range will be accounted for.
6495 */
6496unsigned long __init __absent_pages_in_range(int nid,
6497 unsigned long range_start_pfn,
6498 unsigned long range_end_pfn)
6499{
6500 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6501 unsigned long start_pfn, end_pfn;
6502 int i;
6503
6504 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6505 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6506 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6507 nr_absent -= end_pfn - start_pfn;
6508 }
6509 return nr_absent;
6510}
6511
6512/**
6513 * absent_pages_in_range - Return number of page frames in holes within a range
6514 * @start_pfn: The start PFN to start searching for holes
6515 * @end_pfn: The end PFN to stop searching for holes
6516 *
6517 * Return: the number of pages frames in memory holes within a range.
6518 */
6519unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6520 unsigned long end_pfn)
6521{
6522 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6523}
6524
6525/* Return the number of page frames in holes in a zone on a node */
6526static unsigned long __init zone_absent_pages_in_node(int nid,
6527 unsigned long zone_type,
6528 unsigned long node_start_pfn,
6529 unsigned long node_end_pfn)
6530{
6531 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6532 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6533 unsigned long zone_start_pfn, zone_end_pfn;
6534 unsigned long nr_absent;
6535
6536 /* When hotadd a new node from cpu_up(), the node should be empty */
6537 if (!node_start_pfn && !node_end_pfn)
6538 return 0;
6539
6540 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6541 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6542
6543 adjust_zone_range_for_zone_movable(nid, zone_type,
6544 node_start_pfn, node_end_pfn,
6545 &zone_start_pfn, &zone_end_pfn);
6546 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6547
6548 /*
6549 * ZONE_MOVABLE handling.
6550 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6551 * and vice versa.
6552 */
6553 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6554 unsigned long start_pfn, end_pfn;
6555 struct memblock_region *r;
6556
6557 for_each_memblock(memory, r) {
6558 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6559 zone_start_pfn, zone_end_pfn);
6560 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6561 zone_start_pfn, zone_end_pfn);
6562
6563 if (zone_type == ZONE_MOVABLE &&
6564 memblock_is_mirror(r))
6565 nr_absent += end_pfn - start_pfn;
6566
6567 if (zone_type == ZONE_NORMAL &&
6568 !memblock_is_mirror(r))
6569 nr_absent += end_pfn - start_pfn;
6570 }
6571 }
6572
6573 return nr_absent;
6574}
6575
6576static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6577 unsigned long node_start_pfn,
6578 unsigned long node_end_pfn)
6579{
6580 unsigned long realtotalpages = 0, totalpages = 0;
6581 enum zone_type i;
6582
6583 for (i = 0; i < MAX_NR_ZONES; i++) {
6584 struct zone *zone = pgdat->node_zones + i;
6585 unsigned long zone_start_pfn, zone_end_pfn;
6586 unsigned long spanned, absent;
6587 unsigned long size, real_size;
6588
6589 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
6590 node_start_pfn,
6591 node_end_pfn,
6592 &zone_start_pfn,
6593 &zone_end_pfn);
6594 absent = zone_absent_pages_in_node(pgdat->node_id, i,
6595 node_start_pfn,
6596 node_end_pfn);
6597
6598 size = spanned;
6599 real_size = size - absent;
6600
6601 if (size)
6602 zone->zone_start_pfn = zone_start_pfn;
6603 else
6604 zone->zone_start_pfn = 0;
6605 zone->spanned_pages = size;
6606 zone->present_pages = real_size;
6607
6608 totalpages += size;
6609 realtotalpages += real_size;
6610 }
6611
6612 pgdat->node_spanned_pages = totalpages;
6613 pgdat->node_present_pages = realtotalpages;
6614 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6615 realtotalpages);
6616}
6617
6618#ifndef CONFIG_SPARSEMEM
6619/*
6620 * Calculate the size of the zone->blockflags rounded to an unsigned long
6621 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6622 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6623 * round what is now in bits to nearest long in bits, then return it in
6624 * bytes.
6625 */
6626static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6627{
6628 unsigned long usemapsize;
6629
6630 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6631 usemapsize = roundup(zonesize, pageblock_nr_pages);
6632 usemapsize = usemapsize >> pageblock_order;
6633 usemapsize *= NR_PAGEBLOCK_BITS;
6634 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6635
6636 return usemapsize / 8;
6637}
6638
6639static void __ref setup_usemap(struct pglist_data *pgdat,
6640 struct zone *zone,
6641 unsigned long zone_start_pfn,
6642 unsigned long zonesize)
6643{
6644 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6645 zone->pageblock_flags = NULL;
6646 if (usemapsize) {
6647 zone->pageblock_flags =
6648 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6649 pgdat->node_id);
6650 if (!zone->pageblock_flags)
6651 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6652 usemapsize, zone->name, pgdat->node_id);
6653 }
6654}
6655#else
6656static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6657 unsigned long zone_start_pfn, unsigned long zonesize) {}
6658#endif /* CONFIG_SPARSEMEM */
6659
6660#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6661
6662/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6663void __init set_pageblock_order(void)
6664{
6665 unsigned int order;
6666
6667 /* Check that pageblock_nr_pages has not already been setup */
6668 if (pageblock_order)
6669 return;
6670
6671 if (HPAGE_SHIFT > PAGE_SHIFT)
6672 order = HUGETLB_PAGE_ORDER;
6673 else
6674 order = MAX_ORDER - 1;
6675
6676 /*
6677 * Assume the largest contiguous order of interest is a huge page.
6678 * This value may be variable depending on boot parameters on IA64 and
6679 * powerpc.
6680 */
6681 pageblock_order = order;
6682}
6683#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6684
6685/*
6686 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6687 * is unused as pageblock_order is set at compile-time. See
6688 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6689 * the kernel config
6690 */
6691void __init set_pageblock_order(void)
6692{
6693}
6694
6695#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6696
6697static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6698 unsigned long present_pages)
6699{
6700 unsigned long pages = spanned_pages;
6701
6702 /*
6703 * Provide a more accurate estimation if there are holes within
6704 * the zone and SPARSEMEM is in use. If there are holes within the
6705 * zone, each populated memory region may cost us one or two extra
6706 * memmap pages due to alignment because memmap pages for each
6707 * populated regions may not be naturally aligned on page boundary.
6708 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6709 */
6710 if (spanned_pages > present_pages + (present_pages >> 4) &&
6711 IS_ENABLED(CONFIG_SPARSEMEM))
6712 pages = present_pages;
6713
6714 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6715}
6716
6717#ifdef CONFIG_TRANSPARENT_HUGEPAGE
6718static void pgdat_init_split_queue(struct pglist_data *pgdat)
6719{
6720 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6721
6722 spin_lock_init(&ds_queue->split_queue_lock);
6723 INIT_LIST_HEAD(&ds_queue->split_queue);
6724 ds_queue->split_queue_len = 0;
6725}
6726#else
6727static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6728#endif
6729
6730#ifdef CONFIG_COMPACTION
6731static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6732{
6733 init_waitqueue_head(&pgdat->kcompactd_wait);
6734}
6735#else
6736static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6737#endif
6738
6739static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6740{
6741 pgdat_resize_init(pgdat);
6742
6743 pgdat_init_split_queue(pgdat);
6744 pgdat_init_kcompactd(pgdat);
6745
6746 init_waitqueue_head(&pgdat->kswapd_wait);
6747 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6748
6749 pgdat_page_ext_init(pgdat);
6750 spin_lock_init(&pgdat->lru_lock);
6751 lruvec_init(&pgdat->__lruvec);
6752}
6753
6754static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6755 unsigned long remaining_pages)
6756{
6757 atomic_long_set(&zone->managed_pages, remaining_pages);
6758 zone_set_nid(zone, nid);
6759 zone->name = zone_names[idx];
6760 zone->zone_pgdat = NODE_DATA(nid);
6761 spin_lock_init(&zone->lock);
6762 zone_seqlock_init(zone);
6763 zone_pcp_init(zone);
6764}
6765
6766/*
6767 * Set up the zone data structures
6768 * - init pgdat internals
6769 * - init all zones belonging to this node
6770 *
6771 * NOTE: this function is only called during memory hotplug
6772 */
6773#ifdef CONFIG_MEMORY_HOTPLUG
6774void __ref free_area_init_core_hotplug(int nid)
6775{
6776 enum zone_type z;
6777 pg_data_t *pgdat = NODE_DATA(nid);
6778
6779 pgdat_init_internals(pgdat);
6780 for (z = 0; z < MAX_NR_ZONES; z++)
6781 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6782}
6783#endif
6784
6785/*
6786 * Set up the zone data structures:
6787 * - mark all pages reserved
6788 * - mark all memory queues empty
6789 * - clear the memory bitmaps
6790 *
6791 * NOTE: pgdat should get zeroed by caller.
6792 * NOTE: this function is only called during early init.
6793 */
6794static void __init free_area_init_core(struct pglist_data *pgdat)
6795{
6796 enum zone_type j;
6797 int nid = pgdat->node_id;
6798
6799 pgdat_init_internals(pgdat);
6800 pgdat->per_cpu_nodestats = &boot_nodestats;
6801
6802 for (j = 0; j < MAX_NR_ZONES; j++) {
6803 struct zone *zone = pgdat->node_zones + j;
6804 unsigned long size, freesize, memmap_pages;
6805 unsigned long zone_start_pfn = zone->zone_start_pfn;
6806
6807 size = zone->spanned_pages;
6808 freesize = zone->present_pages;
6809
6810 /*
6811 * Adjust freesize so that it accounts for how much memory
6812 * is used by this zone for memmap. This affects the watermark
6813 * and per-cpu initialisations
6814 */
6815 memmap_pages = calc_memmap_size(size, freesize);
6816 if (!is_highmem_idx(j)) {
6817 if (freesize >= memmap_pages) {
6818 freesize -= memmap_pages;
6819 if (memmap_pages)
6820 printk(KERN_DEBUG
6821 " %s zone: %lu pages used for memmap\n",
6822 zone_names[j], memmap_pages);
6823 } else
6824 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6825 zone_names[j], memmap_pages, freesize);
6826 }
6827
6828 /* Account for reserved pages */
6829 if (j == 0 && freesize > dma_reserve) {
6830 freesize -= dma_reserve;
6831 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6832 zone_names[0], dma_reserve);
6833 }
6834
6835 if (!is_highmem_idx(j))
6836 nr_kernel_pages += freesize;
6837 /* Charge for highmem memmap if there are enough kernel pages */
6838 else if (nr_kernel_pages > memmap_pages * 2)
6839 nr_kernel_pages -= memmap_pages;
6840 nr_all_pages += freesize;
6841
6842 /*
6843 * Set an approximate value for lowmem here, it will be adjusted
6844 * when the bootmem allocator frees pages into the buddy system.
6845 * And all highmem pages will be managed by the buddy system.
6846 */
6847 zone_init_internals(zone, j, nid, freesize);
6848
6849 if (!size)
6850 continue;
6851
6852 set_pageblock_order();
6853 setup_usemap(pgdat, zone, zone_start_pfn, size);
6854 init_currently_empty_zone(zone, zone_start_pfn, size);
6855 memmap_init(size, nid, j, zone_start_pfn);
6856 }
6857}
6858
6859#ifdef CONFIG_FLAT_NODE_MEM_MAP
6860static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6861{
6862 unsigned long __maybe_unused start = 0;
6863 unsigned long __maybe_unused offset = 0;
6864
6865 /* Skip empty nodes */
6866 if (!pgdat->node_spanned_pages)
6867 return;
6868
6869 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6870 offset = pgdat->node_start_pfn - start;
6871 /* ia64 gets its own node_mem_map, before this, without bootmem */
6872 if (!pgdat->node_mem_map) {
6873 unsigned long size, end;
6874 struct page *map;
6875
6876 /*
6877 * The zone's endpoints aren't required to be MAX_ORDER
6878 * aligned but the node_mem_map endpoints must be in order
6879 * for the buddy allocator to function correctly.
6880 */
6881 end = pgdat_end_pfn(pgdat);
6882 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6883 size = (end - start) * sizeof(struct page);
6884 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6885 pgdat->node_id);
6886 if (!map)
6887 panic("Failed to allocate %ld bytes for node %d memory map\n",
6888 size, pgdat->node_id);
6889 pgdat->node_mem_map = map + offset;
6890 }
6891 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6892 __func__, pgdat->node_id, (unsigned long)pgdat,
6893 (unsigned long)pgdat->node_mem_map);
6894#ifndef CONFIG_NEED_MULTIPLE_NODES
6895 /*
6896 * With no DISCONTIG, the global mem_map is just set as node 0's
6897 */
6898 if (pgdat == NODE_DATA(0)) {
6899 mem_map = NODE_DATA(0)->node_mem_map;
6900 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6901 mem_map -= offset;
6902 }
6903#endif
6904}
6905#else
6906static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6907#endif /* CONFIG_FLAT_NODE_MEM_MAP */
6908
6909#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6910static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6911{
6912 pgdat->first_deferred_pfn = ULONG_MAX;
6913}
6914#else
6915static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6916#endif
6917
6918static void __init free_area_init_node(int nid)
6919{
6920 pg_data_t *pgdat = NODE_DATA(nid);
6921 unsigned long start_pfn = 0;
6922 unsigned long end_pfn = 0;
6923
6924 /* pg_data_t should be reset to zero when it's allocated */
6925 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
6926
6927 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6928
6929 pgdat->node_id = nid;
6930 pgdat->node_start_pfn = start_pfn;
6931 pgdat->per_cpu_nodestats = NULL;
6932
6933 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6934 (u64)start_pfn << PAGE_SHIFT,
6935 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6936 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
6937
6938 alloc_node_mem_map(pgdat);
6939 pgdat_set_deferred_range(pgdat);
6940
6941 free_area_init_core(pgdat);
6942}
6943
6944void __init free_area_init_memoryless_node(int nid)
6945{
6946 free_area_init_node(nid);
6947}
6948
6949#if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6950/*
6951 * Initialize all valid struct pages in the range [spfn, epfn) and mark them
6952 * PageReserved(). Return the number of struct pages that were initialized.
6953 */
6954static u64 __init init_unavailable_range(unsigned long spfn, unsigned long epfn)
6955{
6956 unsigned long pfn;
6957 u64 pgcnt = 0;
6958
6959 for (pfn = spfn; pfn < epfn; pfn++) {
6960 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6961 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6962 + pageblock_nr_pages - 1;
6963 continue;
6964 }
6965 /*
6966 * Use a fake node/zone (0) for now. Some of these pages
6967 * (in memblock.reserved but not in memblock.memory) will
6968 * get re-initialized via reserve_bootmem_region() later.
6969 */
6970 __init_single_page(pfn_to_page(pfn), pfn, 0, 0);
6971 __SetPageReserved(pfn_to_page(pfn));
6972 pgcnt++;
6973 }
6974
6975 return pgcnt;
6976}
6977
6978/*
6979 * Only struct pages that are backed by physical memory are zeroed and
6980 * initialized by going through __init_single_page(). But, there are some
6981 * struct pages which are reserved in memblock allocator and their fields
6982 * may be accessed (for example page_to_pfn() on some configuration accesses
6983 * flags). We must explicitly initialize those struct pages.
6984 *
6985 * This function also addresses a similar issue where struct pages are left
6986 * uninitialized because the physical address range is not covered by
6987 * memblock.memory or memblock.reserved. That could happen when memblock
6988 * layout is manually configured via memmap=, or when the highest physical
6989 * address (max_pfn) does not end on a section boundary.
6990 */
6991static void __init init_unavailable_mem(void)
6992{
6993 phys_addr_t start, end;
6994 u64 i, pgcnt;
6995 phys_addr_t next = 0;
6996
6997 /*
6998 * Loop through unavailable ranges not covered by memblock.memory.
6999 */
7000 pgcnt = 0;
7001 for_each_mem_range(i, &memblock.memory, NULL,
7002 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
7003 if (next < start)
7004 pgcnt += init_unavailable_range(PFN_DOWN(next),
7005 PFN_UP(start));
7006 next = end;
7007 }
7008
7009 /*
7010 * Early sections always have a fully populated memmap for the whole
7011 * section - see pfn_valid(). If the last section has holes at the
7012 * end and that section is marked "online", the memmap will be
7013 * considered initialized. Make sure that memmap has a well defined
7014 * state.
7015 */
7016 pgcnt += init_unavailable_range(PFN_DOWN(next),
7017 round_up(max_pfn, PAGES_PER_SECTION));
7018
7019 /*
7020 * Struct pages that do not have backing memory. This could be because
7021 * firmware is using some of this memory, or for some other reasons.
7022 */
7023 if (pgcnt)
7024 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
7025}
7026#else
7027static inline void __init init_unavailable_mem(void)
7028{
7029}
7030#endif /* !CONFIG_FLAT_NODE_MEM_MAP */
7031
7032#if MAX_NUMNODES > 1
7033/*
7034 * Figure out the number of possible node ids.
7035 */
7036void __init setup_nr_node_ids(void)
7037{
7038 unsigned int highest;
7039
7040 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7041 nr_node_ids = highest + 1;
7042}
7043#endif
7044
7045/**
7046 * node_map_pfn_alignment - determine the maximum internode alignment
7047 *
7048 * This function should be called after node map is populated and sorted.
7049 * It calculates the maximum power of two alignment which can distinguish
7050 * all the nodes.
7051 *
7052 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7053 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7054 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7055 * shifted, 1GiB is enough and this function will indicate so.
7056 *
7057 * This is used to test whether pfn -> nid mapping of the chosen memory
7058 * model has fine enough granularity to avoid incorrect mapping for the
7059 * populated node map.
7060 *
7061 * Return: the determined alignment in pfn's. 0 if there is no alignment
7062 * requirement (single node).
7063 */
7064unsigned long __init node_map_pfn_alignment(void)
7065{
7066 unsigned long accl_mask = 0, last_end = 0;
7067 unsigned long start, end, mask;
7068 int last_nid = NUMA_NO_NODE;
7069 int i, nid;
7070
7071 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7072 if (!start || last_nid < 0 || last_nid == nid) {
7073 last_nid = nid;
7074 last_end = end;
7075 continue;
7076 }
7077
7078 /*
7079 * Start with a mask granular enough to pin-point to the
7080 * start pfn and tick off bits one-by-one until it becomes
7081 * too coarse to separate the current node from the last.
7082 */
7083 mask = ~((1 << __ffs(start)) - 1);
7084 while (mask && last_end <= (start & (mask << 1)))
7085 mask <<= 1;
7086
7087 /* accumulate all internode masks */
7088 accl_mask |= mask;
7089 }
7090
7091 /* convert mask to number of pages */
7092 return ~accl_mask + 1;
7093}
7094
7095/**
7096 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7097 *
7098 * Return: the minimum PFN based on information provided via
7099 * memblock_set_node().
7100 */
7101unsigned long __init find_min_pfn_with_active_regions(void)
7102{
7103 return PHYS_PFN(memblock_start_of_DRAM());
7104}
7105
7106/*
7107 * early_calculate_totalpages()
7108 * Sum pages in active regions for movable zone.
7109 * Populate N_MEMORY for calculating usable_nodes.
7110 */
7111static unsigned long __init early_calculate_totalpages(void)
7112{
7113 unsigned long totalpages = 0;
7114 unsigned long start_pfn, end_pfn;
7115 int i, nid;
7116
7117 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7118 unsigned long pages = end_pfn - start_pfn;
7119
7120 totalpages += pages;
7121 if (pages)
7122 node_set_state(nid, N_MEMORY);
7123 }
7124 return totalpages;
7125}
7126
7127/*
7128 * Find the PFN the Movable zone begins in each node. Kernel memory
7129 * is spread evenly between nodes as long as the nodes have enough
7130 * memory. When they don't, some nodes will have more kernelcore than
7131 * others
7132 */
7133static void __init find_zone_movable_pfns_for_nodes(void)
7134{
7135 int i, nid;
7136 unsigned long usable_startpfn;
7137 unsigned long kernelcore_node, kernelcore_remaining;
7138 /* save the state before borrow the nodemask */
7139 nodemask_t saved_node_state = node_states[N_MEMORY];
7140 unsigned long totalpages = early_calculate_totalpages();
7141 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7142 struct memblock_region *r;
7143
7144 /* Need to find movable_zone earlier when movable_node is specified. */
7145 find_usable_zone_for_movable();
7146
7147 /*
7148 * If movable_node is specified, ignore kernelcore and movablecore
7149 * options.
7150 */
7151 if (movable_node_is_enabled()) {
7152 for_each_memblock(memory, r) {
7153 if (!memblock_is_hotpluggable(r))
7154 continue;
7155
7156 nid = memblock_get_region_node(r);
7157
7158 usable_startpfn = PFN_DOWN(r->base);
7159 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7160 min(usable_startpfn, zone_movable_pfn[nid]) :
7161 usable_startpfn;
7162 }
7163
7164 goto out2;
7165 }
7166
7167 /*
7168 * If kernelcore=mirror is specified, ignore movablecore option
7169 */
7170 if (mirrored_kernelcore) {
7171 bool mem_below_4gb_not_mirrored = false;
7172
7173 for_each_memblock(memory, r) {
7174 if (memblock_is_mirror(r))
7175 continue;
7176
7177 nid = memblock_get_region_node(r);
7178
7179 usable_startpfn = memblock_region_memory_base_pfn(r);
7180
7181 if (usable_startpfn < 0x100000) {
7182 mem_below_4gb_not_mirrored = true;
7183 continue;
7184 }
7185
7186 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7187 min(usable_startpfn, zone_movable_pfn[nid]) :
7188 usable_startpfn;
7189 }
7190
7191 if (mem_below_4gb_not_mirrored)
7192 pr_warn("This configuration results in unmirrored kernel memory.\n");
7193
7194 goto out2;
7195 }
7196
7197 /*
7198 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7199 * amount of necessary memory.
7200 */
7201 if (required_kernelcore_percent)
7202 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7203 10000UL;
7204 if (required_movablecore_percent)
7205 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7206 10000UL;
7207
7208 /*
7209 * If movablecore= was specified, calculate what size of
7210 * kernelcore that corresponds so that memory usable for
7211 * any allocation type is evenly spread. If both kernelcore
7212 * and movablecore are specified, then the value of kernelcore
7213 * will be used for required_kernelcore if it's greater than
7214 * what movablecore would have allowed.
7215 */
7216 if (required_movablecore) {
7217 unsigned long corepages;
7218
7219 /*
7220 * Round-up so that ZONE_MOVABLE is at least as large as what
7221 * was requested by the user
7222 */
7223 required_movablecore =
7224 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7225 required_movablecore = min(totalpages, required_movablecore);
7226 corepages = totalpages - required_movablecore;
7227
7228 required_kernelcore = max(required_kernelcore, corepages);
7229 }
7230
7231 /*
7232 * If kernelcore was not specified or kernelcore size is larger
7233 * than totalpages, there is no ZONE_MOVABLE.
7234 */
7235 if (!required_kernelcore || required_kernelcore >= totalpages)
7236 goto out;
7237
7238 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7239 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7240
7241restart:
7242 /* Spread kernelcore memory as evenly as possible throughout nodes */
7243 kernelcore_node = required_kernelcore / usable_nodes;
7244 for_each_node_state(nid, N_MEMORY) {
7245 unsigned long start_pfn, end_pfn;
7246
7247 /*
7248 * Recalculate kernelcore_node if the division per node
7249 * now exceeds what is necessary to satisfy the requested
7250 * amount of memory for the kernel
7251 */
7252 if (required_kernelcore < kernelcore_node)
7253 kernelcore_node = required_kernelcore / usable_nodes;
7254
7255 /*
7256 * As the map is walked, we track how much memory is usable
7257 * by the kernel using kernelcore_remaining. When it is
7258 * 0, the rest of the node is usable by ZONE_MOVABLE
7259 */
7260 kernelcore_remaining = kernelcore_node;
7261
7262 /* Go through each range of PFNs within this node */
7263 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7264 unsigned long size_pages;
7265
7266 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7267 if (start_pfn >= end_pfn)
7268 continue;
7269
7270 /* Account for what is only usable for kernelcore */
7271 if (start_pfn < usable_startpfn) {
7272 unsigned long kernel_pages;
7273 kernel_pages = min(end_pfn, usable_startpfn)
7274 - start_pfn;
7275
7276 kernelcore_remaining -= min(kernel_pages,
7277 kernelcore_remaining);
7278 required_kernelcore -= min(kernel_pages,
7279 required_kernelcore);
7280
7281 /* Continue if range is now fully accounted */
7282 if (end_pfn <= usable_startpfn) {
7283
7284 /*
7285 * Push zone_movable_pfn to the end so
7286 * that if we have to rebalance
7287 * kernelcore across nodes, we will
7288 * not double account here
7289 */
7290 zone_movable_pfn[nid] = end_pfn;
7291 continue;
7292 }
7293 start_pfn = usable_startpfn;
7294 }
7295
7296 /*
7297 * The usable PFN range for ZONE_MOVABLE is from
7298 * start_pfn->end_pfn. Calculate size_pages as the
7299 * number of pages used as kernelcore
7300 */
7301 size_pages = end_pfn - start_pfn;
7302 if (size_pages > kernelcore_remaining)
7303 size_pages = kernelcore_remaining;
7304 zone_movable_pfn[nid] = start_pfn + size_pages;
7305
7306 /*
7307 * Some kernelcore has been met, update counts and
7308 * break if the kernelcore for this node has been
7309 * satisfied
7310 */
7311 required_kernelcore -= min(required_kernelcore,
7312 size_pages);
7313 kernelcore_remaining -= size_pages;
7314 if (!kernelcore_remaining)
7315 break;
7316 }
7317 }
7318
7319 /*
7320 * If there is still required_kernelcore, we do another pass with one
7321 * less node in the count. This will push zone_movable_pfn[nid] further
7322 * along on the nodes that still have memory until kernelcore is
7323 * satisfied
7324 */
7325 usable_nodes--;
7326 if (usable_nodes && required_kernelcore > usable_nodes)
7327 goto restart;
7328
7329out2:
7330 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7331 for (nid = 0; nid < MAX_NUMNODES; nid++)
7332 zone_movable_pfn[nid] =
7333 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7334
7335out:
7336 /* restore the node_state */
7337 node_states[N_MEMORY] = saved_node_state;
7338}
7339
7340/* Any regular or high memory on that node ? */
7341static void check_for_memory(pg_data_t *pgdat, int nid)
7342{
7343 enum zone_type zone_type;
7344
7345 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7346 struct zone *zone = &pgdat->node_zones[zone_type];
7347 if (populated_zone(zone)) {
7348 if (IS_ENABLED(CONFIG_HIGHMEM))
7349 node_set_state(nid, N_HIGH_MEMORY);
7350 if (zone_type <= ZONE_NORMAL)
7351 node_set_state(nid, N_NORMAL_MEMORY);
7352 break;
7353 }
7354 }
7355}
7356
7357/*
7358 * Some architecturs, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7359 * such cases we allow max_zone_pfn sorted in the descending order
7360 */
7361bool __weak arch_has_descending_max_zone_pfns(void)
7362{
7363 return false;
7364}
7365
7366/**
7367 * free_area_init - Initialise all pg_data_t and zone data
7368 * @max_zone_pfn: an array of max PFNs for each zone
7369 *
7370 * This will call free_area_init_node() for each active node in the system.
7371 * Using the page ranges provided by memblock_set_node(), the size of each
7372 * zone in each node and their holes is calculated. If the maximum PFN
7373 * between two adjacent zones match, it is assumed that the zone is empty.
7374 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7375 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7376 * starts where the previous one ended. For example, ZONE_DMA32 starts
7377 * at arch_max_dma_pfn.
7378 */
7379void __init free_area_init(unsigned long *max_zone_pfn)
7380{
7381 unsigned long start_pfn, end_pfn;
7382 int i, nid, zone;
7383 bool descending;
7384
7385 /* Record where the zone boundaries are */
7386 memset(arch_zone_lowest_possible_pfn, 0,
7387 sizeof(arch_zone_lowest_possible_pfn));
7388 memset(arch_zone_highest_possible_pfn, 0,
7389 sizeof(arch_zone_highest_possible_pfn));
7390
7391 start_pfn = find_min_pfn_with_active_regions();
7392 descending = arch_has_descending_max_zone_pfns();
7393
7394 for (i = 0; i < MAX_NR_ZONES; i++) {
7395 if (descending)
7396 zone = MAX_NR_ZONES - i - 1;
7397 else
7398 zone = i;
7399
7400 if (zone == ZONE_MOVABLE)
7401 continue;
7402
7403 end_pfn = max(max_zone_pfn[zone], start_pfn);
7404 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7405 arch_zone_highest_possible_pfn[zone] = end_pfn;
7406
7407 start_pfn = end_pfn;
7408 }
7409
7410 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7411 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7412 find_zone_movable_pfns_for_nodes();
7413
7414 /* Print out the zone ranges */
7415 pr_info("Zone ranges:\n");
7416 for (i = 0; i < MAX_NR_ZONES; i++) {
7417 if (i == ZONE_MOVABLE)
7418 continue;
7419 pr_info(" %-8s ", zone_names[i]);
7420 if (arch_zone_lowest_possible_pfn[i] ==
7421 arch_zone_highest_possible_pfn[i])
7422 pr_cont("empty\n");
7423 else
7424 pr_cont("[mem %#018Lx-%#018Lx]\n",
7425 (u64)arch_zone_lowest_possible_pfn[i]
7426 << PAGE_SHIFT,
7427 ((u64)arch_zone_highest_possible_pfn[i]
7428 << PAGE_SHIFT) - 1);
7429 }
7430
7431 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7432 pr_info("Movable zone start for each node\n");
7433 for (i = 0; i < MAX_NUMNODES; i++) {
7434 if (zone_movable_pfn[i])
7435 pr_info(" Node %d: %#018Lx\n", i,
7436 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7437 }
7438
7439 /*
7440 * Print out the early node map, and initialize the
7441 * subsection-map relative to active online memory ranges to
7442 * enable future "sub-section" extensions of the memory map.
7443 */
7444 pr_info("Early memory node ranges\n");
7445 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7446 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7447 (u64)start_pfn << PAGE_SHIFT,
7448 ((u64)end_pfn << PAGE_SHIFT) - 1);
7449 subsection_map_init(start_pfn, end_pfn - start_pfn);
7450 }
7451
7452 /* Initialise every node */
7453 mminit_verify_pageflags_layout();
7454 setup_nr_node_ids();
7455 init_unavailable_mem();
7456 for_each_online_node(nid) {
7457 pg_data_t *pgdat = NODE_DATA(nid);
7458 free_area_init_node(nid);
7459
7460 /* Any memory on that node */
7461 if (pgdat->node_present_pages)
7462 node_set_state(nid, N_MEMORY);
7463 check_for_memory(pgdat, nid);
7464 }
7465}
7466
7467static int __init cmdline_parse_core(char *p, unsigned long *core,
7468 unsigned long *percent)
7469{
7470 unsigned long long coremem;
7471 char *endptr;
7472
7473 if (!p)
7474 return -EINVAL;
7475
7476 /* Value may be a percentage of total memory, otherwise bytes */
7477 coremem = simple_strtoull(p, &endptr, 0);
7478 if (*endptr == '%') {
7479 /* Paranoid check for percent values greater than 100 */
7480 WARN_ON(coremem > 100);
7481
7482 *percent = coremem;
7483 } else {
7484 coremem = memparse(p, &p);
7485 /* Paranoid check that UL is enough for the coremem value */
7486 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7487
7488 *core = coremem >> PAGE_SHIFT;
7489 *percent = 0UL;
7490 }
7491 return 0;
7492}
7493
7494/*
7495 * kernelcore=size sets the amount of memory for use for allocations that
7496 * cannot be reclaimed or migrated.
7497 */
7498static int __init cmdline_parse_kernelcore(char *p)
7499{
7500 /* parse kernelcore=mirror */
7501 if (parse_option_str(p, "mirror")) {
7502 mirrored_kernelcore = true;
7503 return 0;
7504 }
7505
7506 return cmdline_parse_core(p, &required_kernelcore,
7507 &required_kernelcore_percent);
7508}
7509
7510/*
7511 * movablecore=size sets the amount of memory for use for allocations that
7512 * can be reclaimed or migrated.
7513 */
7514static int __init cmdline_parse_movablecore(char *p)
7515{
7516 return cmdline_parse_core(p, &required_movablecore,
7517 &required_movablecore_percent);
7518}
7519
7520early_param("kernelcore", cmdline_parse_kernelcore);
7521early_param("movablecore", cmdline_parse_movablecore);
7522
7523void adjust_managed_page_count(struct page *page, long count)
7524{
7525 atomic_long_add(count, &page_zone(page)->managed_pages);
7526 totalram_pages_add(count);
7527#ifdef CONFIG_HIGHMEM
7528 if (PageHighMem(page))
7529 totalhigh_pages_add(count);
7530#endif
7531}
7532EXPORT_SYMBOL(adjust_managed_page_count);
7533
7534unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7535{
7536 void *pos;
7537 unsigned long pages = 0;
7538
7539 start = (void *)PAGE_ALIGN((unsigned long)start);
7540 end = (void *)((unsigned long)end & PAGE_MASK);
7541 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7542 struct page *page = virt_to_page(pos);
7543 void *direct_map_addr;
7544
7545 /*
7546 * 'direct_map_addr' might be different from 'pos'
7547 * because some architectures' virt_to_page()
7548 * work with aliases. Getting the direct map
7549 * address ensures that we get a _writeable_
7550 * alias for the memset().
7551 */
7552 direct_map_addr = page_address(page);
7553 if ((unsigned int)poison <= 0xFF)
7554 memset(direct_map_addr, poison, PAGE_SIZE);
7555
7556 free_reserved_page(page);
7557 }
7558
7559 if (pages && s)
7560 pr_info("Freeing %s memory: %ldK\n",
7561 s, pages << (PAGE_SHIFT - 10));
7562
7563 return pages;
7564}
7565
7566#ifdef CONFIG_HIGHMEM
7567void free_highmem_page(struct page *page)
7568{
7569 __free_reserved_page(page);
7570 totalram_pages_inc();
7571 atomic_long_inc(&page_zone(page)->managed_pages);
7572 totalhigh_pages_inc();
7573}
7574#endif
7575
7576
7577void __init mem_init_print_info(const char *str)
7578{
7579 unsigned long physpages, codesize, datasize, rosize, bss_size;
7580 unsigned long init_code_size, init_data_size;
7581
7582 physpages = get_num_physpages();
7583 codesize = _etext - _stext;
7584 datasize = _edata - _sdata;
7585 rosize = __end_rodata - __start_rodata;
7586 bss_size = __bss_stop - __bss_start;
7587 init_data_size = __init_end - __init_begin;
7588 init_code_size = _einittext - _sinittext;
7589
7590 /*
7591 * Detect special cases and adjust section sizes accordingly:
7592 * 1) .init.* may be embedded into .data sections
7593 * 2) .init.text.* may be out of [__init_begin, __init_end],
7594 * please refer to arch/tile/kernel/vmlinux.lds.S.
7595 * 3) .rodata.* may be embedded into .text or .data sections.
7596 */
7597#define adj_init_size(start, end, size, pos, adj) \
7598 do { \
7599 if (start <= pos && pos < end && size > adj) \
7600 size -= adj; \
7601 } while (0)
7602
7603 adj_init_size(__init_begin, __init_end, init_data_size,
7604 _sinittext, init_code_size);
7605 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7606 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7607 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7608 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7609
7610#undef adj_init_size
7611
7612 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7613#ifdef CONFIG_HIGHMEM
7614 ", %luK highmem"
7615#endif
7616 "%s%s)\n",
7617 nr_free_pages() << (PAGE_SHIFT - 10),
7618 physpages << (PAGE_SHIFT - 10),
7619 codesize >> 10, datasize >> 10, rosize >> 10,
7620 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7621 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7622 totalcma_pages << (PAGE_SHIFT - 10),
7623#ifdef CONFIG_HIGHMEM
7624 totalhigh_pages() << (PAGE_SHIFT - 10),
7625#endif
7626 str ? ", " : "", str ? str : "");
7627}
7628
7629/**
7630 * set_dma_reserve - set the specified number of pages reserved in the first zone
7631 * @new_dma_reserve: The number of pages to mark reserved
7632 *
7633 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7634 * In the DMA zone, a significant percentage may be consumed by kernel image
7635 * and other unfreeable allocations which can skew the watermarks badly. This
7636 * function may optionally be used to account for unfreeable pages in the
7637 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7638 * smaller per-cpu batchsize.
7639 */
7640void __init set_dma_reserve(unsigned long new_dma_reserve)
7641{
7642 dma_reserve = new_dma_reserve;
7643}
7644
7645static int page_alloc_cpu_dead(unsigned int cpu)
7646{
7647
7648 lru_add_drain_cpu(cpu);
7649 drain_pages(cpu);
7650
7651 /*
7652 * Spill the event counters of the dead processor
7653 * into the current processors event counters.
7654 * This artificially elevates the count of the current
7655 * processor.
7656 */
7657 vm_events_fold_cpu(cpu);
7658
7659 /*
7660 * Zero the differential counters of the dead processor
7661 * so that the vm statistics are consistent.
7662 *
7663 * This is only okay since the processor is dead and cannot
7664 * race with what we are doing.
7665 */
7666 cpu_vm_stats_fold(cpu);
7667 return 0;
7668}
7669
7670#ifdef CONFIG_NUMA
7671int hashdist = HASHDIST_DEFAULT;
7672
7673static int __init set_hashdist(char *str)
7674{
7675 if (!str)
7676 return 0;
7677 hashdist = simple_strtoul(str, &str, 0);
7678 return 1;
7679}
7680__setup("hashdist=", set_hashdist);
7681#endif
7682
7683void __init page_alloc_init(void)
7684{
7685 int ret;
7686
7687#ifdef CONFIG_NUMA
7688 if (num_node_state(N_MEMORY) == 1)
7689 hashdist = 0;
7690#endif
7691
7692 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7693 "mm/page_alloc:dead", NULL,
7694 page_alloc_cpu_dead);
7695 WARN_ON(ret < 0);
7696}
7697
7698/*
7699 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7700 * or min_free_kbytes changes.
7701 */
7702static void calculate_totalreserve_pages(void)
7703{
7704 struct pglist_data *pgdat;
7705 unsigned long reserve_pages = 0;
7706 enum zone_type i, j;
7707
7708 for_each_online_pgdat(pgdat) {
7709
7710 pgdat->totalreserve_pages = 0;
7711
7712 for (i = 0; i < MAX_NR_ZONES; i++) {
7713 struct zone *zone = pgdat->node_zones + i;
7714 long max = 0;
7715 unsigned long managed_pages = zone_managed_pages(zone);
7716
7717 /* Find valid and maximum lowmem_reserve in the zone */
7718 for (j = i; j < MAX_NR_ZONES; j++) {
7719 if (zone->lowmem_reserve[j] > max)
7720 max = zone->lowmem_reserve[j];
7721 }
7722
7723 /* we treat the high watermark as reserved pages. */
7724 max += high_wmark_pages(zone);
7725
7726 if (max > managed_pages)
7727 max = managed_pages;
7728
7729 pgdat->totalreserve_pages += max;
7730
7731 reserve_pages += max;
7732 }
7733 }
7734 totalreserve_pages = reserve_pages;
7735}
7736
7737/*
7738 * setup_per_zone_lowmem_reserve - called whenever
7739 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7740 * has a correct pages reserved value, so an adequate number of
7741 * pages are left in the zone after a successful __alloc_pages().
7742 */
7743static void setup_per_zone_lowmem_reserve(void)
7744{
7745 struct pglist_data *pgdat;
7746 enum zone_type j, idx;
7747
7748 for_each_online_pgdat(pgdat) {
7749 for (j = 0; j < MAX_NR_ZONES; j++) {
7750 struct zone *zone = pgdat->node_zones + j;
7751 unsigned long managed_pages = zone_managed_pages(zone);
7752
7753 zone->lowmem_reserve[j] = 0;
7754
7755 idx = j;
7756 while (idx) {
7757 struct zone *lower_zone;
7758
7759 idx--;
7760 lower_zone = pgdat->node_zones + idx;
7761
7762 if (!sysctl_lowmem_reserve_ratio[idx] ||
7763 !zone_managed_pages(lower_zone)) {
7764 lower_zone->lowmem_reserve[j] = 0;
7765 continue;
7766 } else {
7767 lower_zone->lowmem_reserve[j] =
7768 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7769 }
7770 managed_pages += zone_managed_pages(lower_zone);
7771 }
7772 }
7773 }
7774
7775 /* update totalreserve_pages */
7776 calculate_totalreserve_pages();
7777}
7778
7779static void __setup_per_zone_wmarks(void)
7780{
7781 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7782 unsigned long lowmem_pages = 0;
7783 struct zone *zone;
7784 unsigned long flags;
7785
7786 /* Calculate total number of !ZONE_HIGHMEM pages */
7787 for_each_zone(zone) {
7788 if (!is_highmem(zone))
7789 lowmem_pages += zone_managed_pages(zone);
7790 }
7791
7792 for_each_zone(zone) {
7793 u64 tmp;
7794
7795 spin_lock_irqsave(&zone->lock, flags);
7796 tmp = (u64)pages_min * zone_managed_pages(zone);
7797 do_div(tmp, lowmem_pages);
7798 if (is_highmem(zone)) {
7799 /*
7800 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7801 * need highmem pages, so cap pages_min to a small
7802 * value here.
7803 *
7804 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7805 * deltas control async page reclaim, and so should
7806 * not be capped for highmem.
7807 */
7808 unsigned long min_pages;
7809
7810 min_pages = zone_managed_pages(zone) / 1024;
7811 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7812 zone->_watermark[WMARK_MIN] = min_pages;
7813 } else {
7814 /*
7815 * If it's a lowmem zone, reserve a number of pages
7816 * proportionate to the zone's size.
7817 */
7818 zone->_watermark[WMARK_MIN] = tmp;
7819 }
7820
7821 /*
7822 * Set the kswapd watermarks distance according to the
7823 * scale factor in proportion to available memory, but
7824 * ensure a minimum size on small systems.
7825 */
7826 tmp = max_t(u64, tmp >> 2,
7827 mult_frac(zone_managed_pages(zone),
7828 watermark_scale_factor, 10000));
7829
7830 zone->watermark_boost = 0;
7831 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7832 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7833
7834 spin_unlock_irqrestore(&zone->lock, flags);
7835 }
7836
7837 /* update totalreserve_pages */
7838 calculate_totalreserve_pages();
7839}
7840
7841/**
7842 * setup_per_zone_wmarks - called when min_free_kbytes changes
7843 * or when memory is hot-{added|removed}
7844 *
7845 * Ensures that the watermark[min,low,high] values for each zone are set
7846 * correctly with respect to min_free_kbytes.
7847 */
7848void setup_per_zone_wmarks(void)
7849{
7850 static DEFINE_SPINLOCK(lock);
7851
7852 spin_lock(&lock);
7853 __setup_per_zone_wmarks();
7854 spin_unlock(&lock);
7855}
7856
7857/*
7858 * Initialise min_free_kbytes.
7859 *
7860 * For small machines we want it small (128k min). For large machines
7861 * we want it large (256MB max). But it is not linear, because network
7862 * bandwidth does not increase linearly with machine size. We use
7863 *
7864 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7865 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7866 *
7867 * which yields
7868 *
7869 * 16MB: 512k
7870 * 32MB: 724k
7871 * 64MB: 1024k
7872 * 128MB: 1448k
7873 * 256MB: 2048k
7874 * 512MB: 2896k
7875 * 1024MB: 4096k
7876 * 2048MB: 5792k
7877 * 4096MB: 8192k
7878 * 8192MB: 11584k
7879 * 16384MB: 16384k
7880 */
7881int __meminit init_per_zone_wmark_min(void)
7882{
7883 unsigned long lowmem_kbytes;
7884 int new_min_free_kbytes;
7885
7886 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7887 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7888
7889 if (new_min_free_kbytes > user_min_free_kbytes) {
7890 min_free_kbytes = new_min_free_kbytes;
7891 if (min_free_kbytes < 128)
7892 min_free_kbytes = 128;
7893 if (min_free_kbytes > 262144)
7894 min_free_kbytes = 262144;
7895 } else {
7896 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7897 new_min_free_kbytes, user_min_free_kbytes);
7898 }
7899 setup_per_zone_wmarks();
7900 refresh_zone_stat_thresholds();
7901 setup_per_zone_lowmem_reserve();
7902
7903#ifdef CONFIG_NUMA
7904 setup_min_unmapped_ratio();
7905 setup_min_slab_ratio();
7906#endif
7907
7908 khugepaged_min_free_kbytes_update();
7909
7910 return 0;
7911}
7912postcore_initcall(init_per_zone_wmark_min)
7913
7914/*
7915 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7916 * that we can call two helper functions whenever min_free_kbytes
7917 * changes.
7918 */
7919int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7920 void *buffer, size_t *length, loff_t *ppos)
7921{
7922 int rc;
7923
7924 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7925 if (rc)
7926 return rc;
7927
7928 if (write) {
7929 user_min_free_kbytes = min_free_kbytes;
7930 setup_per_zone_wmarks();
7931 }
7932 return 0;
7933}
7934
7935int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7936 void *buffer, size_t *length, loff_t *ppos)
7937{
7938 int rc;
7939
7940 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7941 if (rc)
7942 return rc;
7943
7944 if (write)
7945 setup_per_zone_wmarks();
7946
7947 return 0;
7948}
7949
7950#ifdef CONFIG_NUMA
7951static void setup_min_unmapped_ratio(void)
7952{
7953 pg_data_t *pgdat;
7954 struct zone *zone;
7955
7956 for_each_online_pgdat(pgdat)
7957 pgdat->min_unmapped_pages = 0;
7958
7959 for_each_zone(zone)
7960 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7961 sysctl_min_unmapped_ratio) / 100;
7962}
7963
7964
7965int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7966 void *buffer, size_t *length, loff_t *ppos)
7967{
7968 int rc;
7969
7970 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7971 if (rc)
7972 return rc;
7973
7974 setup_min_unmapped_ratio();
7975
7976 return 0;
7977}
7978
7979static void setup_min_slab_ratio(void)
7980{
7981 pg_data_t *pgdat;
7982 struct zone *zone;
7983
7984 for_each_online_pgdat(pgdat)
7985 pgdat->min_slab_pages = 0;
7986
7987 for_each_zone(zone)
7988 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7989 sysctl_min_slab_ratio) / 100;
7990}
7991
7992int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7993 void *buffer, size_t *length, loff_t *ppos)
7994{
7995 int rc;
7996
7997 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7998 if (rc)
7999 return rc;
8000
8001 setup_min_slab_ratio();
8002
8003 return 0;
8004}
8005#endif
8006
8007/*
8008 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8009 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8010 * whenever sysctl_lowmem_reserve_ratio changes.
8011 *
8012 * The reserve ratio obviously has absolutely no relation with the
8013 * minimum watermarks. The lowmem reserve ratio can only make sense
8014 * if in function of the boot time zone sizes.
8015 */
8016int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8017 void *buffer, size_t *length, loff_t *ppos)
8018{
8019 int i;
8020
8021 proc_dointvec_minmax(table, write, buffer, length, ppos);
8022
8023 for (i = 0; i < MAX_NR_ZONES; i++) {
8024 if (sysctl_lowmem_reserve_ratio[i] < 1)
8025 sysctl_lowmem_reserve_ratio[i] = 0;
8026 }
8027
8028 setup_per_zone_lowmem_reserve();
8029 return 0;
8030}
8031
8032static void __zone_pcp_update(struct zone *zone)
8033{
8034 unsigned int cpu;
8035
8036 for_each_possible_cpu(cpu)
8037 pageset_set_high_and_batch(zone,
8038 per_cpu_ptr(zone->pageset, cpu));
8039}
8040
8041/*
8042 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8043 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8044 * pagelist can have before it gets flushed back to buddy allocator.
8045 */
8046int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8047 void *buffer, size_t *length, loff_t *ppos)
8048{
8049 struct zone *zone;
8050 int old_percpu_pagelist_fraction;
8051 int ret;
8052
8053 mutex_lock(&pcp_batch_high_lock);
8054 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8055
8056 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8057 if (!write || ret < 0)
8058 goto out;
8059
8060 /* Sanity checking to avoid pcp imbalance */
8061 if (percpu_pagelist_fraction &&
8062 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8063 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8064 ret = -EINVAL;
8065 goto out;
8066 }
8067
8068 /* No change? */
8069 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8070 goto out;
8071
8072 for_each_populated_zone(zone)
8073 __zone_pcp_update(zone);
8074out:
8075 mutex_unlock(&pcp_batch_high_lock);
8076 return ret;
8077}
8078
8079#ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8080/*
8081 * Returns the number of pages that arch has reserved but
8082 * is not known to alloc_large_system_hash().
8083 */
8084static unsigned long __init arch_reserved_kernel_pages(void)
8085{
8086 return 0;
8087}
8088#endif
8089
8090/*
8091 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8092 * machines. As memory size is increased the scale is also increased but at
8093 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8094 * quadruples the scale is increased by one, which means the size of hash table
8095 * only doubles, instead of quadrupling as well.
8096 * Because 32-bit systems cannot have large physical memory, where this scaling
8097 * makes sense, it is disabled on such platforms.
8098 */
8099#if __BITS_PER_LONG > 32
8100#define ADAPT_SCALE_BASE (64ul << 30)
8101#define ADAPT_SCALE_SHIFT 2
8102#define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8103#endif
8104
8105/*
8106 * allocate a large system hash table from bootmem
8107 * - it is assumed that the hash table must contain an exact power-of-2
8108 * quantity of entries
8109 * - limit is the number of hash buckets, not the total allocation size
8110 */
8111void *__init alloc_large_system_hash(const char *tablename,
8112 unsigned long bucketsize,
8113 unsigned long numentries,
8114 int scale,
8115 int flags,
8116 unsigned int *_hash_shift,
8117 unsigned int *_hash_mask,
8118 unsigned long low_limit,
8119 unsigned long high_limit)
8120{
8121 unsigned long long max = high_limit;
8122 unsigned long log2qty, size;
8123 void *table = NULL;
8124 gfp_t gfp_flags;
8125 bool virt;
8126
8127 /* allow the kernel cmdline to have a say */
8128 if (!numentries) {
8129 /* round applicable memory size up to nearest megabyte */
8130 numentries = nr_kernel_pages;
8131 numentries -= arch_reserved_kernel_pages();
8132
8133 /* It isn't necessary when PAGE_SIZE >= 1MB */
8134 if (PAGE_SHIFT < 20)
8135 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8136
8137#if __BITS_PER_LONG > 32
8138 if (!high_limit) {
8139 unsigned long adapt;
8140
8141 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8142 adapt <<= ADAPT_SCALE_SHIFT)
8143 scale++;
8144 }
8145#endif
8146
8147 /* limit to 1 bucket per 2^scale bytes of low memory */
8148 if (scale > PAGE_SHIFT)
8149 numentries >>= (scale - PAGE_SHIFT);
8150 else
8151 numentries <<= (PAGE_SHIFT - scale);
8152
8153 /* Make sure we've got at least a 0-order allocation.. */
8154 if (unlikely(flags & HASH_SMALL)) {
8155 /* Makes no sense without HASH_EARLY */
8156 WARN_ON(!(flags & HASH_EARLY));
8157 if (!(numentries >> *_hash_shift)) {
8158 numentries = 1UL << *_hash_shift;
8159 BUG_ON(!numentries);
8160 }
8161 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8162 numentries = PAGE_SIZE / bucketsize;
8163 }
8164 numentries = roundup_pow_of_two(numentries);
8165
8166 /* limit allocation size to 1/16 total memory by default */
8167 if (max == 0) {
8168 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8169 do_div(max, bucketsize);
8170 }
8171 max = min(max, 0x80000000ULL);
8172
8173 if (numentries < low_limit)
8174 numentries = low_limit;
8175 if (numentries > max)
8176 numentries = max;
8177
8178 log2qty = ilog2(numentries);
8179
8180 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8181 do {
8182 virt = false;
8183 size = bucketsize << log2qty;
8184 if (flags & HASH_EARLY) {
8185 if (flags & HASH_ZERO)
8186 table = memblock_alloc(size, SMP_CACHE_BYTES);
8187 else
8188 table = memblock_alloc_raw(size,
8189 SMP_CACHE_BYTES);
8190 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8191 table = __vmalloc(size, gfp_flags);
8192 virt = true;
8193 } else {
8194 /*
8195 * If bucketsize is not a power-of-two, we may free
8196 * some pages at the end of hash table which
8197 * alloc_pages_exact() automatically does
8198 */
8199 table = alloc_pages_exact(size, gfp_flags);
8200 kmemleak_alloc(table, size, 1, gfp_flags);
8201 }
8202 } while (!table && size > PAGE_SIZE && --log2qty);
8203
8204 if (!table)
8205 panic("Failed to allocate %s hash table\n", tablename);
8206
8207 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8208 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8209 virt ? "vmalloc" : "linear");
8210
8211 if (_hash_shift)
8212 *_hash_shift = log2qty;
8213 if (_hash_mask)
8214 *_hash_mask = (1 << log2qty) - 1;
8215
8216 return table;
8217}
8218
8219/*
8220 * This function checks whether pageblock includes unmovable pages or not.
8221 *
8222 * PageLRU check without isolation or lru_lock could race so that
8223 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8224 * check without lock_page also may miss some movable non-lru pages at
8225 * race condition. So you can't expect this function should be exact.
8226 *
8227 * Returns a page without holding a reference. If the caller wants to
8228 * dereference that page (e.g., dumping), it has to make sure that it
8229 * cannot get removed (e.g., via memory unplug) concurrently.
8230 *
8231 */
8232struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8233 int migratetype, int flags)
8234{
8235 unsigned long iter = 0;
8236 unsigned long pfn = page_to_pfn(page);
8237
8238 /*
8239 * TODO we could make this much more efficient by not checking every
8240 * page in the range if we know all of them are in MOVABLE_ZONE and
8241 * that the movable zone guarantees that pages are migratable but
8242 * the later is not the case right now unfortunatelly. E.g. movablecore
8243 * can still lead to having bootmem allocations in zone_movable.
8244 */
8245
8246 if (is_migrate_cma_page(page)) {
8247 /*
8248 * CMA allocations (alloc_contig_range) really need to mark
8249 * isolate CMA pageblocks even when they are not movable in fact
8250 * so consider them movable here.
8251 */
8252 if (is_migrate_cma(migratetype))
8253 return NULL;
8254
8255 return page;
8256 }
8257
8258 for (; iter < pageblock_nr_pages; iter++) {
8259 if (!pfn_valid_within(pfn + iter))
8260 continue;
8261
8262 page = pfn_to_page(pfn + iter);
8263
8264 if (PageReserved(page))
8265 return page;
8266
8267 /*
8268 * If the zone is movable and we have ruled out all reserved
8269 * pages then it should be reasonably safe to assume the rest
8270 * is movable.
8271 */
8272 if (zone_idx(zone) == ZONE_MOVABLE)
8273 continue;
8274
8275 /*
8276 * Hugepages are not in LRU lists, but they're movable.
8277 * THPs are on the LRU, but need to be counted as #small pages.
8278 * We need not scan over tail pages because we don't
8279 * handle each tail page individually in migration.
8280 */
8281 if (PageHuge(page) || PageTransCompound(page)) {
8282 struct page *head = compound_head(page);
8283 unsigned int skip_pages;
8284
8285 if (PageHuge(page)) {
8286 if (!hugepage_migration_supported(page_hstate(head)))
8287 return page;
8288 } else if (!PageLRU(head) && !__PageMovable(head)) {
8289 return page;
8290 }
8291
8292 skip_pages = compound_nr(head) - (page - head);
8293 iter += skip_pages - 1;
8294 continue;
8295 }
8296
8297 /*
8298 * We can't use page_count without pin a page
8299 * because another CPU can free compound page.
8300 * This check already skips compound tails of THP
8301 * because their page->_refcount is zero at all time.
8302 */
8303 if (!page_ref_count(page)) {
8304 if (PageBuddy(page))
8305 iter += (1 << page_order(page)) - 1;
8306 continue;
8307 }
8308
8309 /*
8310 * The HWPoisoned page may be not in buddy system, and
8311 * page_count() is not 0.
8312 */
8313 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8314 continue;
8315
8316 /*
8317 * We treat all PageOffline() pages as movable when offlining
8318 * to give drivers a chance to decrement their reference count
8319 * in MEM_GOING_OFFLINE in order to indicate that these pages
8320 * can be offlined as there are no direct references anymore.
8321 * For actually unmovable PageOffline() where the driver does
8322 * not support this, we will fail later when trying to actually
8323 * move these pages that still have a reference count > 0.
8324 * (false negatives in this function only)
8325 */
8326 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8327 continue;
8328
8329 if (__PageMovable(page) || PageLRU(page))
8330 continue;
8331
8332 /*
8333 * If there are RECLAIMABLE pages, we need to check
8334 * it. But now, memory offline itself doesn't call
8335 * shrink_node_slabs() and it still to be fixed.
8336 */
8337 /*
8338 * If the page is not RAM, page_count()should be 0.
8339 * we don't need more check. This is an _used_ not-movable page.
8340 *
8341 * The problematic thing here is PG_reserved pages. PG_reserved
8342 * is set to both of a memory hole page and a _used_ kernel
8343 * page at boot.
8344 */
8345 return page;
8346 }
8347 return NULL;
8348}
8349
8350#ifdef CONFIG_CONTIG_ALLOC
8351static unsigned long pfn_max_align_down(unsigned long pfn)
8352{
8353 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8354 pageblock_nr_pages) - 1);
8355}
8356
8357static unsigned long pfn_max_align_up(unsigned long pfn)
8358{
8359 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8360 pageblock_nr_pages));
8361}
8362
8363/* [start, end) must belong to a single zone. */
8364static int __alloc_contig_migrate_range(struct compact_control *cc,
8365 unsigned long start, unsigned long end)
8366{
8367 /* This function is based on compact_zone() from compaction.c. */
8368 unsigned int nr_reclaimed;
8369 unsigned long pfn = start;
8370 unsigned int tries = 0;
8371 int ret = 0;
8372 struct migration_target_control mtc = {
8373 .nid = zone_to_nid(cc->zone),
8374 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8375 };
8376
8377 migrate_prep();
8378
8379 while (pfn < end || !list_empty(&cc->migratepages)) {
8380 if (fatal_signal_pending(current)) {
8381 ret = -EINTR;
8382 break;
8383 }
8384
8385 if (list_empty(&cc->migratepages)) {
8386 cc->nr_migratepages = 0;
8387 pfn = isolate_migratepages_range(cc, pfn, end);
8388 if (!pfn) {
8389 ret = -EINTR;
8390 break;
8391 }
8392 tries = 0;
8393 } else if (++tries == 5) {
8394 ret = ret < 0 ? ret : -EBUSY;
8395 break;
8396 }
8397
8398 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8399 &cc->migratepages);
8400 cc->nr_migratepages -= nr_reclaimed;
8401
8402 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
8403 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE);
8404 }
8405 if (ret < 0) {
8406 putback_movable_pages(&cc->migratepages);
8407 return ret;
8408 }
8409 return 0;
8410}
8411
8412/**
8413 * alloc_contig_range() -- tries to allocate given range of pages
8414 * @start: start PFN to allocate
8415 * @end: one-past-the-last PFN to allocate
8416 * @migratetype: migratetype of the underlaying pageblocks (either
8417 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8418 * in range must have the same migratetype and it must
8419 * be either of the two.
8420 * @gfp_mask: GFP mask to use during compaction
8421 *
8422 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8423 * aligned. The PFN range must belong to a single zone.
8424 *
8425 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8426 * pageblocks in the range. Once isolated, the pageblocks should not
8427 * be modified by others.
8428 *
8429 * Return: zero on success or negative error code. On success all
8430 * pages which PFN is in [start, end) are allocated for the caller and
8431 * need to be freed with free_contig_range().
8432 */
8433int alloc_contig_range(unsigned long start, unsigned long end,
8434 unsigned migratetype, gfp_t gfp_mask)
8435{
8436 unsigned long outer_start, outer_end;
8437 unsigned int order;
8438 int ret = 0;
8439
8440 struct compact_control cc = {
8441 .nr_migratepages = 0,
8442 .order = -1,
8443 .zone = page_zone(pfn_to_page(start)),
8444 .mode = MIGRATE_SYNC,
8445 .ignore_skip_hint = true,
8446 .no_set_skip_hint = true,
8447 .gfp_mask = current_gfp_context(gfp_mask),
8448 .alloc_contig = true,
8449 };
8450 INIT_LIST_HEAD(&cc.migratepages);
8451
8452 /*
8453 * What we do here is we mark all pageblocks in range as
8454 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8455 * have different sizes, and due to the way page allocator
8456 * work, we align the range to biggest of the two pages so
8457 * that page allocator won't try to merge buddies from
8458 * different pageblocks and change MIGRATE_ISOLATE to some
8459 * other migration type.
8460 *
8461 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8462 * migrate the pages from an unaligned range (ie. pages that
8463 * we are interested in). This will put all the pages in
8464 * range back to page allocator as MIGRATE_ISOLATE.
8465 *
8466 * When this is done, we take the pages in range from page
8467 * allocator removing them from the buddy system. This way
8468 * page allocator will never consider using them.
8469 *
8470 * This lets us mark the pageblocks back as
8471 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8472 * aligned range but not in the unaligned, original range are
8473 * put back to page allocator so that buddy can use them.
8474 */
8475
8476 ret = start_isolate_page_range(pfn_max_align_down(start),
8477 pfn_max_align_up(end), migratetype, 0);
8478 if (ret < 0)
8479 return ret;
8480
8481 /*
8482 * In case of -EBUSY, we'd like to know which page causes problem.
8483 * So, just fall through. test_pages_isolated() has a tracepoint
8484 * which will report the busy page.
8485 *
8486 * It is possible that busy pages could become available before
8487 * the call to test_pages_isolated, and the range will actually be
8488 * allocated. So, if we fall through be sure to clear ret so that
8489 * -EBUSY is not accidentally used or returned to caller.
8490 */
8491 ret = __alloc_contig_migrate_range(&cc, start, end);
8492 if (ret && ret != -EBUSY)
8493 goto done;
8494 ret =0;
8495
8496 /*
8497 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8498 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8499 * more, all pages in [start, end) are free in page allocator.
8500 * What we are going to do is to allocate all pages from
8501 * [start, end) (that is remove them from page allocator).
8502 *
8503 * The only problem is that pages at the beginning and at the
8504 * end of interesting range may be not aligned with pages that
8505 * page allocator holds, ie. they can be part of higher order
8506 * pages. Because of this, we reserve the bigger range and
8507 * once this is done free the pages we are not interested in.
8508 *
8509 * We don't have to hold zone->lock here because the pages are
8510 * isolated thus they won't get removed from buddy.
8511 */
8512
8513 lru_add_drain_all();
8514
8515 order = 0;
8516 outer_start = start;
8517 while (!PageBuddy(pfn_to_page(outer_start))) {
8518 if (++order >= MAX_ORDER) {
8519 outer_start = start;
8520 break;
8521 }
8522 outer_start &= ~0UL << order;
8523 }
8524
8525 if (outer_start != start) {
8526 order = page_order(pfn_to_page(outer_start));
8527
8528 /*
8529 * outer_start page could be small order buddy page and
8530 * it doesn't include start page. Adjust outer_start
8531 * in this case to report failed page properly
8532 * on tracepoint in test_pages_isolated()
8533 */
8534 if (outer_start + (1UL << order) <= start)
8535 outer_start = start;
8536 }
8537
8538 /* Make sure the range is really isolated. */
8539 if (test_pages_isolated(outer_start, end, 0)) {
8540 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8541 __func__, outer_start, end);
8542 ret = -EBUSY;
8543 goto done;
8544 }
8545
8546 /* Grab isolated pages from freelists. */
8547 outer_end = isolate_freepages_range(&cc, outer_start, end);
8548 if (!outer_end) {
8549 ret = -EBUSY;
8550 goto done;
8551 }
8552
8553 /* Free head and tail (if any) */
8554 if (start != outer_start)
8555 free_contig_range(outer_start, start - outer_start);
8556 if (end != outer_end)
8557 free_contig_range(end, outer_end - end);
8558
8559done:
8560 undo_isolate_page_range(pfn_max_align_down(start),
8561 pfn_max_align_up(end), migratetype);
8562 return ret;
8563}
8564EXPORT_SYMBOL(alloc_contig_range);
8565
8566static int __alloc_contig_pages(unsigned long start_pfn,
8567 unsigned long nr_pages, gfp_t gfp_mask)
8568{
8569 unsigned long end_pfn = start_pfn + nr_pages;
8570
8571 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8572 gfp_mask);
8573}
8574
8575static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8576 unsigned long nr_pages)
8577{
8578 unsigned long i, end_pfn = start_pfn + nr_pages;
8579 struct page *page;
8580
8581 for (i = start_pfn; i < end_pfn; i++) {
8582 page = pfn_to_online_page(i);
8583 if (!page)
8584 return false;
8585
8586 if (page_zone(page) != z)
8587 return false;
8588
8589 if (PageReserved(page))
8590 return false;
8591
8592 if (page_count(page) > 0)
8593 return false;
8594
8595 if (PageHuge(page))
8596 return false;
8597 }
8598 return true;
8599}
8600
8601static bool zone_spans_last_pfn(const struct zone *zone,
8602 unsigned long start_pfn, unsigned long nr_pages)
8603{
8604 unsigned long last_pfn = start_pfn + nr_pages - 1;
8605
8606 return zone_spans_pfn(zone, last_pfn);
8607}
8608
8609/**
8610 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8611 * @nr_pages: Number of contiguous pages to allocate
8612 * @gfp_mask: GFP mask to limit search and used during compaction
8613 * @nid: Target node
8614 * @nodemask: Mask for other possible nodes
8615 *
8616 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8617 * on an applicable zonelist to find a contiguous pfn range which can then be
8618 * tried for allocation with alloc_contig_range(). This routine is intended
8619 * for allocation requests which can not be fulfilled with the buddy allocator.
8620 *
8621 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8622 * power of two then the alignment is guaranteed to be to the given nr_pages
8623 * (e.g. 1GB request would be aligned to 1GB).
8624 *
8625 * Allocated pages can be freed with free_contig_range() or by manually calling
8626 * __free_page() on each allocated page.
8627 *
8628 * Return: pointer to contiguous pages on success, or NULL if not successful.
8629 */
8630struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8631 int nid, nodemask_t *nodemask)
8632{
8633 unsigned long ret, pfn, flags;
8634 struct zonelist *zonelist;
8635 struct zone *zone;
8636 struct zoneref *z;
8637
8638 zonelist = node_zonelist(nid, gfp_mask);
8639 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8640 gfp_zone(gfp_mask), nodemask) {
8641 spin_lock_irqsave(&zone->lock, flags);
8642
8643 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8644 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8645 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8646 /*
8647 * We release the zone lock here because
8648 * alloc_contig_range() will also lock the zone
8649 * at some point. If there's an allocation
8650 * spinning on this lock, it may win the race
8651 * and cause alloc_contig_range() to fail...
8652 */
8653 spin_unlock_irqrestore(&zone->lock, flags);
8654 ret = __alloc_contig_pages(pfn, nr_pages,
8655 gfp_mask);
8656 if (!ret)
8657 return pfn_to_page(pfn);
8658 spin_lock_irqsave(&zone->lock, flags);
8659 }
8660 pfn += nr_pages;
8661 }
8662 spin_unlock_irqrestore(&zone->lock, flags);
8663 }
8664 return NULL;
8665}
8666#endif /* CONFIG_CONTIG_ALLOC */
8667
8668void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8669{
8670 unsigned int count = 0;
8671
8672 for (; nr_pages--; pfn++) {
8673 struct page *page = pfn_to_page(pfn);
8674
8675 count += page_count(page) != 1;
8676 __free_page(page);
8677 }
8678 WARN(count != 0, "%d pages are still in use!\n", count);
8679}
8680EXPORT_SYMBOL(free_contig_range);
8681
8682/*
8683 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8684 * page high values need to be recalulated.
8685 */
8686void __meminit zone_pcp_update(struct zone *zone)
8687{
8688 mutex_lock(&pcp_batch_high_lock);
8689 __zone_pcp_update(zone);
8690 mutex_unlock(&pcp_batch_high_lock);
8691}
8692
8693void zone_pcp_reset(struct zone *zone)
8694{
8695 unsigned long flags;
8696 int cpu;
8697 struct per_cpu_pageset *pset;
8698
8699 /* avoid races with drain_pages() */
8700 local_irq_save(flags);
8701 if (zone->pageset != &boot_pageset) {
8702 for_each_online_cpu(cpu) {
8703 pset = per_cpu_ptr(zone->pageset, cpu);
8704 drain_zonestat(zone, pset);
8705 }
8706 free_percpu(zone->pageset);
8707 zone->pageset = &boot_pageset;
8708 }
8709 local_irq_restore(flags);
8710}
8711
8712#ifdef CONFIG_MEMORY_HOTREMOVE
8713/*
8714 * All pages in the range must be in a single zone and isolated
8715 * before calling this.
8716 */
8717unsigned long
8718__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8719{
8720 struct page *page;
8721 struct zone *zone;
8722 unsigned int order;
8723 unsigned long pfn;
8724 unsigned long flags;
8725 unsigned long offlined_pages = 0;
8726
8727 /* find the first valid pfn */
8728 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8729 if (pfn_valid(pfn))
8730 break;
8731 if (pfn == end_pfn)
8732 return offlined_pages;
8733
8734 offline_mem_sections(pfn, end_pfn);
8735 zone = page_zone(pfn_to_page(pfn));
8736 spin_lock_irqsave(&zone->lock, flags);
8737 pfn = start_pfn;
8738 while (pfn < end_pfn) {
8739 if (!pfn_valid(pfn)) {
8740 pfn++;
8741 continue;
8742 }
8743 page = pfn_to_page(pfn);
8744 /*
8745 * The HWPoisoned page may be not in buddy system, and
8746 * page_count() is not 0.
8747 */
8748 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8749 pfn++;
8750 offlined_pages++;
8751 continue;
8752 }
8753 /*
8754 * At this point all remaining PageOffline() pages have a
8755 * reference count of 0 and can simply be skipped.
8756 */
8757 if (PageOffline(page)) {
8758 BUG_ON(page_count(page));
8759 BUG_ON(PageBuddy(page));
8760 pfn++;
8761 offlined_pages++;
8762 continue;
8763 }
8764
8765 BUG_ON(page_count(page));
8766 BUG_ON(!PageBuddy(page));
8767 order = page_order(page);
8768 offlined_pages += 1 << order;
8769 del_page_from_free_list(page, zone, order);
8770 pfn += (1 << order);
8771 }
8772 spin_unlock_irqrestore(&zone->lock, flags);
8773
8774 return offlined_pages;
8775}
8776#endif
8777
8778bool is_free_buddy_page(struct page *page)
8779{
8780 struct zone *zone = page_zone(page);
8781 unsigned long pfn = page_to_pfn(page);
8782 unsigned long flags;
8783 unsigned int order;
8784
8785 spin_lock_irqsave(&zone->lock, flags);
8786 for (order = 0; order < MAX_ORDER; order++) {
8787 struct page *page_head = page - (pfn & ((1 << order) - 1));
8788
8789 if (PageBuddy(page_head) && page_order(page_head) >= order)
8790 break;
8791 }
8792 spin_unlock_irqrestore(&zone->lock, flags);
8793
8794 return order < MAX_ORDER;
8795}
8796
8797#ifdef CONFIG_MEMORY_FAILURE
8798/*
8799 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8800 * test is performed under the zone lock to prevent a race against page
8801 * allocation.
8802 */
8803bool set_hwpoison_free_buddy_page(struct page *page)
8804{
8805 struct zone *zone = page_zone(page);
8806 unsigned long pfn = page_to_pfn(page);
8807 unsigned long flags;
8808 unsigned int order;
8809 bool hwpoisoned = false;
8810
8811 spin_lock_irqsave(&zone->lock, flags);
8812 for (order = 0; order < MAX_ORDER; order++) {
8813 struct page *page_head = page - (pfn & ((1 << order) - 1));
8814
8815 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8816 if (!TestSetPageHWPoison(page))
8817 hwpoisoned = true;
8818 break;
8819 }
8820 }
8821 spin_unlock_irqrestore(&zone->lock, flags);
8822
8823 return hwpoisoned;
8824}
8825#endif