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/interrupt.h>
22#include <linux/jiffies.h>
23#include <linux/compiler.h>
24#include <linux/kernel.h>
25#include <linux/kasan.h>
26#include <linux/kmsan.h>
27#include <linux/module.h>
28#include <linux/suspend.h>
29#include <linux/ratelimit.h>
30#include <linux/oom.h>
31#include <linux/topology.h>
32#include <linux/sysctl.h>
33#include <linux/cpu.h>
34#include <linux/cpuset.h>
35#include <linux/memory_hotplug.h>
36#include <linux/nodemask.h>
37#include <linux/vmstat.h>
38#include <linux/fault-inject.h>
39#include <linux/compaction.h>
40#include <trace/events/kmem.h>
41#include <trace/events/oom.h>
42#include <linux/prefetch.h>
43#include <linux/mm_inline.h>
44#include <linux/mmu_notifier.h>
45#include <linux/migrate.h>
46#include <linux/sched/mm.h>
47#include <linux/page_owner.h>
48#include <linux/page_table_check.h>
49#include <linux/memcontrol.h>
50#include <linux/ftrace.h>
51#include <linux/lockdep.h>
52#include <linux/psi.h>
53#include <linux/khugepaged.h>
54#include <linux/delayacct.h>
55#include <linux/cacheinfo.h>
56#include <asm/div64.h>
57#include "internal.h"
58#include "shuffle.h"
59#include "page_reporting.h"
60
61/* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
62typedef int __bitwise fpi_t;
63
64/* No special request */
65#define FPI_NONE ((__force fpi_t)0)
66
67/*
68 * Skip free page reporting notification for the (possibly merged) page.
69 * This does not hinder free page reporting from grabbing the page,
70 * reporting it and marking it "reported" - it only skips notifying
71 * the free page reporting infrastructure about a newly freed page. For
72 * example, used when temporarily pulling a page from a freelist and
73 * putting it back unmodified.
74 */
75#define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
76
77/*
78 * Place the (possibly merged) page to the tail of the freelist. Will ignore
79 * page shuffling (relevant code - e.g., memory onlining - is expected to
80 * shuffle the whole zone).
81 *
82 * Note: No code should rely on this flag for correctness - it's purely
83 * to allow for optimizations when handing back either fresh pages
84 * (memory onlining) or untouched pages (page isolation, free page
85 * reporting).
86 */
87#define FPI_TO_TAIL ((__force fpi_t)BIT(1))
88
89/* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
90static DEFINE_MUTEX(pcp_batch_high_lock);
91#define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
92
93#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
94/*
95 * On SMP, spin_trylock is sufficient protection.
96 * On PREEMPT_RT, spin_trylock is equivalent on both SMP and UP.
97 */
98#define pcp_trylock_prepare(flags) do { } while (0)
99#define pcp_trylock_finish(flag) do { } while (0)
100#else
101
102/* UP spin_trylock always succeeds so disable IRQs to prevent re-entrancy. */
103#define pcp_trylock_prepare(flags) local_irq_save(flags)
104#define pcp_trylock_finish(flags) local_irq_restore(flags)
105#endif
106
107/*
108 * Locking a pcp requires a PCP lookup followed by a spinlock. To avoid
109 * a migration causing the wrong PCP to be locked and remote memory being
110 * potentially allocated, pin the task to the CPU for the lookup+lock.
111 * preempt_disable is used on !RT because it is faster than migrate_disable.
112 * migrate_disable is used on RT because otherwise RT spinlock usage is
113 * interfered with and a high priority task cannot preempt the allocator.
114 */
115#ifndef CONFIG_PREEMPT_RT
116#define pcpu_task_pin() preempt_disable()
117#define pcpu_task_unpin() preempt_enable()
118#else
119#define pcpu_task_pin() migrate_disable()
120#define pcpu_task_unpin() migrate_enable()
121#endif
122
123/*
124 * Generic helper to lookup and a per-cpu variable with an embedded spinlock.
125 * Return value should be used with equivalent unlock helper.
126 */
127#define pcpu_spin_lock(type, member, ptr) \
128({ \
129 type *_ret; \
130 pcpu_task_pin(); \
131 _ret = this_cpu_ptr(ptr); \
132 spin_lock(&_ret->member); \
133 _ret; \
134})
135
136#define pcpu_spin_trylock(type, member, ptr) \
137({ \
138 type *_ret; \
139 pcpu_task_pin(); \
140 _ret = this_cpu_ptr(ptr); \
141 if (!spin_trylock(&_ret->member)) { \
142 pcpu_task_unpin(); \
143 _ret = NULL; \
144 } \
145 _ret; \
146})
147
148#define pcpu_spin_unlock(member, ptr) \
149({ \
150 spin_unlock(&ptr->member); \
151 pcpu_task_unpin(); \
152})
153
154/* struct per_cpu_pages specific helpers. */
155#define pcp_spin_lock(ptr) \
156 pcpu_spin_lock(struct per_cpu_pages, lock, ptr)
157
158#define pcp_spin_trylock(ptr) \
159 pcpu_spin_trylock(struct per_cpu_pages, lock, ptr)
160
161#define pcp_spin_unlock(ptr) \
162 pcpu_spin_unlock(lock, ptr)
163
164#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
165DEFINE_PER_CPU(int, numa_node);
166EXPORT_PER_CPU_SYMBOL(numa_node);
167#endif
168
169DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
170
171#ifdef CONFIG_HAVE_MEMORYLESS_NODES
172/*
173 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
174 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
175 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
176 * defined in <linux/topology.h>.
177 */
178DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
179EXPORT_PER_CPU_SYMBOL(_numa_mem_);
180#endif
181
182static DEFINE_MUTEX(pcpu_drain_mutex);
183
184#ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
185volatile unsigned long latent_entropy __latent_entropy;
186EXPORT_SYMBOL(latent_entropy);
187#endif
188
189/*
190 * Array of node states.
191 */
192nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
193 [N_POSSIBLE] = NODE_MASK_ALL,
194 [N_ONLINE] = { { [0] = 1UL } },
195#ifndef CONFIG_NUMA
196 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
197#ifdef CONFIG_HIGHMEM
198 [N_HIGH_MEMORY] = { { [0] = 1UL } },
199#endif
200 [N_MEMORY] = { { [0] = 1UL } },
201 [N_CPU] = { { [0] = 1UL } },
202#endif /* NUMA */
203};
204EXPORT_SYMBOL(node_states);
205
206gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
207
208/*
209 * A cached value of the page's pageblock's migratetype, used when the page is
210 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
211 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
212 * Also the migratetype set in the page does not necessarily match the pcplist
213 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
214 * other index - this ensures that it will be put on the correct CMA freelist.
215 */
216static inline int get_pcppage_migratetype(struct page *page)
217{
218 return page->index;
219}
220
221static inline void set_pcppage_migratetype(struct page *page, int migratetype)
222{
223 page->index = migratetype;
224}
225
226#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
227unsigned int pageblock_order __read_mostly;
228#endif
229
230static void __free_pages_ok(struct page *page, unsigned int order,
231 fpi_t fpi_flags);
232
233/*
234 * results with 256, 32 in the lowmem_reserve sysctl:
235 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
236 * 1G machine -> (16M dma, 784M normal, 224M high)
237 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
238 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
239 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
240 *
241 * TBD: should special case ZONE_DMA32 machines here - in those we normally
242 * don't need any ZONE_NORMAL reservation
243 */
244static int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
245#ifdef CONFIG_ZONE_DMA
246 [ZONE_DMA] = 256,
247#endif
248#ifdef CONFIG_ZONE_DMA32
249 [ZONE_DMA32] = 256,
250#endif
251 [ZONE_NORMAL] = 32,
252#ifdef CONFIG_HIGHMEM
253 [ZONE_HIGHMEM] = 0,
254#endif
255 [ZONE_MOVABLE] = 0,
256};
257
258char * const zone_names[MAX_NR_ZONES] = {
259#ifdef CONFIG_ZONE_DMA
260 "DMA",
261#endif
262#ifdef CONFIG_ZONE_DMA32
263 "DMA32",
264#endif
265 "Normal",
266#ifdef CONFIG_HIGHMEM
267 "HighMem",
268#endif
269 "Movable",
270#ifdef CONFIG_ZONE_DEVICE
271 "Device",
272#endif
273};
274
275const char * const migratetype_names[MIGRATE_TYPES] = {
276 "Unmovable",
277 "Movable",
278 "Reclaimable",
279 "HighAtomic",
280#ifdef CONFIG_CMA
281 "CMA",
282#endif
283#ifdef CONFIG_MEMORY_ISOLATION
284 "Isolate",
285#endif
286};
287
288int min_free_kbytes = 1024;
289int user_min_free_kbytes = -1;
290static int watermark_boost_factor __read_mostly = 15000;
291static int watermark_scale_factor = 10;
292
293/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
294int movable_zone;
295EXPORT_SYMBOL(movable_zone);
296
297#if MAX_NUMNODES > 1
298unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
299unsigned int nr_online_nodes __read_mostly = 1;
300EXPORT_SYMBOL(nr_node_ids);
301EXPORT_SYMBOL(nr_online_nodes);
302#endif
303
304static bool page_contains_unaccepted(struct page *page, unsigned int order);
305static void accept_page(struct page *page, unsigned int order);
306static bool try_to_accept_memory(struct zone *zone, unsigned int order);
307static inline bool has_unaccepted_memory(void);
308static bool __free_unaccepted(struct page *page);
309
310int page_group_by_mobility_disabled __read_mostly;
311
312#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
313/*
314 * During boot we initialize deferred pages on-demand, as needed, but once
315 * page_alloc_init_late() has finished, the deferred pages are all initialized,
316 * and we can permanently disable that path.
317 */
318DEFINE_STATIC_KEY_TRUE(deferred_pages);
319
320static inline bool deferred_pages_enabled(void)
321{
322 return static_branch_unlikely(&deferred_pages);
323}
324
325/*
326 * deferred_grow_zone() is __init, but it is called from
327 * get_page_from_freelist() during early boot until deferred_pages permanently
328 * disables this call. This is why we have refdata wrapper to avoid warning,
329 * and to ensure that the function body gets unloaded.
330 */
331static bool __ref
332_deferred_grow_zone(struct zone *zone, unsigned int order)
333{
334 return deferred_grow_zone(zone, order);
335}
336#else
337static inline bool deferred_pages_enabled(void)
338{
339 return false;
340}
341#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
342
343/* Return a pointer to the bitmap storing bits affecting a block of pages */
344static inline unsigned long *get_pageblock_bitmap(const struct page *page,
345 unsigned long pfn)
346{
347#ifdef CONFIG_SPARSEMEM
348 return section_to_usemap(__pfn_to_section(pfn));
349#else
350 return page_zone(page)->pageblock_flags;
351#endif /* CONFIG_SPARSEMEM */
352}
353
354static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
355{
356#ifdef CONFIG_SPARSEMEM
357 pfn &= (PAGES_PER_SECTION-1);
358#else
359 pfn = pfn - pageblock_start_pfn(page_zone(page)->zone_start_pfn);
360#endif /* CONFIG_SPARSEMEM */
361 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
362}
363
364/**
365 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
366 * @page: The page within the block of interest
367 * @pfn: The target page frame number
368 * @mask: mask of bits that the caller is interested in
369 *
370 * Return: pageblock_bits flags
371 */
372unsigned long get_pfnblock_flags_mask(const struct page *page,
373 unsigned long pfn, unsigned long mask)
374{
375 unsigned long *bitmap;
376 unsigned long bitidx, word_bitidx;
377 unsigned long word;
378
379 bitmap = get_pageblock_bitmap(page, pfn);
380 bitidx = pfn_to_bitidx(page, pfn);
381 word_bitidx = bitidx / BITS_PER_LONG;
382 bitidx &= (BITS_PER_LONG-1);
383 /*
384 * This races, without locks, with set_pfnblock_flags_mask(). Ensure
385 * a consistent read of the memory array, so that results, even though
386 * racy, are not corrupted.
387 */
388 word = READ_ONCE(bitmap[word_bitidx]);
389 return (word >> bitidx) & mask;
390}
391
392static __always_inline int get_pfnblock_migratetype(const struct page *page,
393 unsigned long pfn)
394{
395 return get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
396}
397
398/**
399 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
400 * @page: The page within the block of interest
401 * @flags: The flags to set
402 * @pfn: The target page frame number
403 * @mask: mask of bits that the caller is interested in
404 */
405void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
406 unsigned long pfn,
407 unsigned long mask)
408{
409 unsigned long *bitmap;
410 unsigned long bitidx, word_bitidx;
411 unsigned long word;
412
413 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
414 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
415
416 bitmap = get_pageblock_bitmap(page, pfn);
417 bitidx = pfn_to_bitidx(page, pfn);
418 word_bitidx = bitidx / BITS_PER_LONG;
419 bitidx &= (BITS_PER_LONG-1);
420
421 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
422
423 mask <<= bitidx;
424 flags <<= bitidx;
425
426 word = READ_ONCE(bitmap[word_bitidx]);
427 do {
428 } while (!try_cmpxchg(&bitmap[word_bitidx], &word, (word & ~mask) | flags));
429}
430
431void set_pageblock_migratetype(struct page *page, int migratetype)
432{
433 if (unlikely(page_group_by_mobility_disabled &&
434 migratetype < MIGRATE_PCPTYPES))
435 migratetype = MIGRATE_UNMOVABLE;
436
437 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
438 page_to_pfn(page), MIGRATETYPE_MASK);
439}
440
441#ifdef CONFIG_DEBUG_VM
442static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
443{
444 int ret;
445 unsigned seq;
446 unsigned long pfn = page_to_pfn(page);
447 unsigned long sp, start_pfn;
448
449 do {
450 seq = zone_span_seqbegin(zone);
451 start_pfn = zone->zone_start_pfn;
452 sp = zone->spanned_pages;
453 ret = !zone_spans_pfn(zone, pfn);
454 } while (zone_span_seqretry(zone, seq));
455
456 if (ret)
457 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
458 pfn, zone_to_nid(zone), zone->name,
459 start_pfn, start_pfn + sp);
460
461 return ret;
462}
463
464/*
465 * Temporary debugging check for pages not lying within a given zone.
466 */
467static int __maybe_unused bad_range(struct zone *zone, struct page *page)
468{
469 if (page_outside_zone_boundaries(zone, page))
470 return 1;
471 if (zone != page_zone(page))
472 return 1;
473
474 return 0;
475}
476#else
477static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
478{
479 return 0;
480}
481#endif
482
483static void bad_page(struct page *page, const char *reason)
484{
485 static unsigned long resume;
486 static unsigned long nr_shown;
487 static unsigned long nr_unshown;
488
489 /*
490 * Allow a burst of 60 reports, then keep quiet for that minute;
491 * or allow a steady drip of one report per second.
492 */
493 if (nr_shown == 60) {
494 if (time_before(jiffies, resume)) {
495 nr_unshown++;
496 goto out;
497 }
498 if (nr_unshown) {
499 pr_alert(
500 "BUG: Bad page state: %lu messages suppressed\n",
501 nr_unshown);
502 nr_unshown = 0;
503 }
504 nr_shown = 0;
505 }
506 if (nr_shown++ == 0)
507 resume = jiffies + 60 * HZ;
508
509 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
510 current->comm, page_to_pfn(page));
511 dump_page(page, reason);
512
513 print_modules();
514 dump_stack();
515out:
516 /* Leave bad fields for debug, except PageBuddy could make trouble */
517 page_mapcount_reset(page); /* remove PageBuddy */
518 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
519}
520
521static inline unsigned int order_to_pindex(int migratetype, int order)
522{
523#ifdef CONFIG_TRANSPARENT_HUGEPAGE
524 if (order > PAGE_ALLOC_COSTLY_ORDER) {
525 VM_BUG_ON(order != pageblock_order);
526 return NR_LOWORDER_PCP_LISTS;
527 }
528#else
529 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
530#endif
531
532 return (MIGRATE_PCPTYPES * order) + migratetype;
533}
534
535static inline int pindex_to_order(unsigned int pindex)
536{
537 int order = pindex / MIGRATE_PCPTYPES;
538
539#ifdef CONFIG_TRANSPARENT_HUGEPAGE
540 if (pindex == NR_LOWORDER_PCP_LISTS)
541 order = pageblock_order;
542#else
543 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
544#endif
545
546 return order;
547}
548
549static inline bool pcp_allowed_order(unsigned int order)
550{
551 if (order <= PAGE_ALLOC_COSTLY_ORDER)
552 return true;
553#ifdef CONFIG_TRANSPARENT_HUGEPAGE
554 if (order == pageblock_order)
555 return true;
556#endif
557 return false;
558}
559
560static inline void free_the_page(struct page *page, unsigned int order)
561{
562 if (pcp_allowed_order(order)) /* Via pcp? */
563 free_unref_page(page, order);
564 else
565 __free_pages_ok(page, order, FPI_NONE);
566}
567
568/*
569 * Higher-order pages are called "compound pages". They are structured thusly:
570 *
571 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
572 *
573 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
574 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
575 *
576 * The first tail page's ->compound_order holds the order of allocation.
577 * This usage means that zero-order pages may not be compound.
578 */
579
580void prep_compound_page(struct page *page, unsigned int order)
581{
582 int i;
583 int nr_pages = 1 << order;
584
585 __SetPageHead(page);
586 for (i = 1; i < nr_pages; i++)
587 prep_compound_tail(page, i);
588
589 prep_compound_head(page, order);
590}
591
592void destroy_large_folio(struct folio *folio)
593{
594 if (folio_test_hugetlb(folio)) {
595 free_huge_folio(folio);
596 return;
597 }
598
599 if (folio_test_large_rmappable(folio))
600 folio_undo_large_rmappable(folio);
601
602 mem_cgroup_uncharge(folio);
603 free_the_page(&folio->page, folio_order(folio));
604}
605
606static inline void set_buddy_order(struct page *page, unsigned int order)
607{
608 set_page_private(page, order);
609 __SetPageBuddy(page);
610}
611
612#ifdef CONFIG_COMPACTION
613static inline struct capture_control *task_capc(struct zone *zone)
614{
615 struct capture_control *capc = current->capture_control;
616
617 return unlikely(capc) &&
618 !(current->flags & PF_KTHREAD) &&
619 !capc->page &&
620 capc->cc->zone == zone ? capc : NULL;
621}
622
623static inline bool
624compaction_capture(struct capture_control *capc, struct page *page,
625 int order, int migratetype)
626{
627 if (!capc || order != capc->cc->order)
628 return false;
629
630 /* Do not accidentally pollute CMA or isolated regions*/
631 if (is_migrate_cma(migratetype) ||
632 is_migrate_isolate(migratetype))
633 return false;
634
635 /*
636 * Do not let lower order allocations pollute a movable pageblock.
637 * This might let an unmovable request use a reclaimable pageblock
638 * and vice-versa but no more than normal fallback logic which can
639 * have trouble finding a high-order free page.
640 */
641 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
642 return false;
643
644 capc->page = page;
645 return true;
646}
647
648#else
649static inline struct capture_control *task_capc(struct zone *zone)
650{
651 return NULL;
652}
653
654static inline bool
655compaction_capture(struct capture_control *capc, struct page *page,
656 int order, int migratetype)
657{
658 return false;
659}
660#endif /* CONFIG_COMPACTION */
661
662/* Used for pages not on another list */
663static inline void add_to_free_list(struct page *page, struct zone *zone,
664 unsigned int order, int migratetype)
665{
666 struct free_area *area = &zone->free_area[order];
667
668 list_add(&page->buddy_list, &area->free_list[migratetype]);
669 area->nr_free++;
670}
671
672/* Used for pages not on another list */
673static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
674 unsigned int order, int migratetype)
675{
676 struct free_area *area = &zone->free_area[order];
677
678 list_add_tail(&page->buddy_list, &area->free_list[migratetype]);
679 area->nr_free++;
680}
681
682/*
683 * Used for pages which are on another list. Move the pages to the tail
684 * of the list - so the moved pages won't immediately be considered for
685 * allocation again (e.g., optimization for memory onlining).
686 */
687static inline void move_to_free_list(struct page *page, struct zone *zone,
688 unsigned int order, int migratetype)
689{
690 struct free_area *area = &zone->free_area[order];
691
692 list_move_tail(&page->buddy_list, &area->free_list[migratetype]);
693}
694
695static inline void del_page_from_free_list(struct page *page, struct zone *zone,
696 unsigned int order)
697{
698 /* clear reported state and update reported page count */
699 if (page_reported(page))
700 __ClearPageReported(page);
701
702 list_del(&page->buddy_list);
703 __ClearPageBuddy(page);
704 set_page_private(page, 0);
705 zone->free_area[order].nr_free--;
706}
707
708static inline struct page *get_page_from_free_area(struct free_area *area,
709 int migratetype)
710{
711 return list_first_entry_or_null(&area->free_list[migratetype],
712 struct page, buddy_list);
713}
714
715/*
716 * If this is not the largest possible page, check if the buddy
717 * of the next-highest order is free. If it is, it's possible
718 * that pages are being freed that will coalesce soon. In case,
719 * that is happening, add the free page to the tail of the list
720 * so it's less likely to be used soon and more likely to be merged
721 * as a higher order page
722 */
723static inline bool
724buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
725 struct page *page, unsigned int order)
726{
727 unsigned long higher_page_pfn;
728 struct page *higher_page;
729
730 if (order >= MAX_PAGE_ORDER - 1)
731 return false;
732
733 higher_page_pfn = buddy_pfn & pfn;
734 higher_page = page + (higher_page_pfn - pfn);
735
736 return find_buddy_page_pfn(higher_page, higher_page_pfn, order + 1,
737 NULL) != NULL;
738}
739
740/*
741 * Freeing function for a buddy system allocator.
742 *
743 * The concept of a buddy system is to maintain direct-mapped table
744 * (containing bit values) for memory blocks of various "orders".
745 * The bottom level table contains the map for the smallest allocatable
746 * units of memory (here, pages), and each level above it describes
747 * pairs of units from the levels below, hence, "buddies".
748 * At a high level, all that happens here is marking the table entry
749 * at the bottom level available, and propagating the changes upward
750 * as necessary, plus some accounting needed to play nicely with other
751 * parts of the VM system.
752 * At each level, we keep a list of pages, which are heads of continuous
753 * free pages of length of (1 << order) and marked with PageBuddy.
754 * Page's order is recorded in page_private(page) field.
755 * So when we are allocating or freeing one, we can derive the state of the
756 * other. That is, if we allocate a small block, and both were
757 * free, the remainder of the region must be split into blocks.
758 * If a block is freed, and its buddy is also free, then this
759 * triggers coalescing into a block of larger size.
760 *
761 * -- nyc
762 */
763
764static inline void __free_one_page(struct page *page,
765 unsigned long pfn,
766 struct zone *zone, unsigned int order,
767 int migratetype, fpi_t fpi_flags)
768{
769 struct capture_control *capc = task_capc(zone);
770 unsigned long buddy_pfn = 0;
771 unsigned long combined_pfn;
772 struct page *buddy;
773 bool to_tail;
774
775 VM_BUG_ON(!zone_is_initialized(zone));
776 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
777
778 VM_BUG_ON(migratetype == -1);
779 if (likely(!is_migrate_isolate(migratetype)))
780 __mod_zone_freepage_state(zone, 1 << order, migratetype);
781
782 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
783 VM_BUG_ON_PAGE(bad_range(zone, page), page);
784
785 while (order < MAX_PAGE_ORDER) {
786 if (compaction_capture(capc, page, order, migratetype)) {
787 __mod_zone_freepage_state(zone, -(1 << order),
788 migratetype);
789 return;
790 }
791
792 buddy = find_buddy_page_pfn(page, pfn, order, &buddy_pfn);
793 if (!buddy)
794 goto done_merging;
795
796 if (unlikely(order >= pageblock_order)) {
797 /*
798 * We want to prevent merge between freepages on pageblock
799 * without fallbacks and normal pageblock. Without this,
800 * pageblock isolation could cause incorrect freepage or CMA
801 * accounting or HIGHATOMIC accounting.
802 */
803 int buddy_mt = get_pfnblock_migratetype(buddy, buddy_pfn);
804
805 if (migratetype != buddy_mt
806 && (!migratetype_is_mergeable(migratetype) ||
807 !migratetype_is_mergeable(buddy_mt)))
808 goto done_merging;
809 }
810
811 /*
812 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
813 * merge with it and move up one order.
814 */
815 if (page_is_guard(buddy))
816 clear_page_guard(zone, buddy, order, migratetype);
817 else
818 del_page_from_free_list(buddy, zone, order);
819 combined_pfn = buddy_pfn & pfn;
820 page = page + (combined_pfn - pfn);
821 pfn = combined_pfn;
822 order++;
823 }
824
825done_merging:
826 set_buddy_order(page, order);
827
828 if (fpi_flags & FPI_TO_TAIL)
829 to_tail = true;
830 else if (is_shuffle_order(order))
831 to_tail = shuffle_pick_tail();
832 else
833 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
834
835 if (to_tail)
836 add_to_free_list_tail(page, zone, order, migratetype);
837 else
838 add_to_free_list(page, zone, order, migratetype);
839
840 /* Notify page reporting subsystem of freed page */
841 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
842 page_reporting_notify_free(order);
843}
844
845/**
846 * split_free_page() -- split a free page at split_pfn_offset
847 * @free_page: the original free page
848 * @order: the order of the page
849 * @split_pfn_offset: split offset within the page
850 *
851 * Return -ENOENT if the free page is changed, otherwise 0
852 *
853 * It is used when the free page crosses two pageblocks with different migratetypes
854 * at split_pfn_offset within the page. The split free page will be put into
855 * separate migratetype lists afterwards. Otherwise, the function achieves
856 * nothing.
857 */
858int split_free_page(struct page *free_page,
859 unsigned int order, unsigned long split_pfn_offset)
860{
861 struct zone *zone = page_zone(free_page);
862 unsigned long free_page_pfn = page_to_pfn(free_page);
863 unsigned long pfn;
864 unsigned long flags;
865 int free_page_order;
866 int mt;
867 int ret = 0;
868
869 if (split_pfn_offset == 0)
870 return ret;
871
872 spin_lock_irqsave(&zone->lock, flags);
873
874 if (!PageBuddy(free_page) || buddy_order(free_page) != order) {
875 ret = -ENOENT;
876 goto out;
877 }
878
879 mt = get_pfnblock_migratetype(free_page, free_page_pfn);
880 if (likely(!is_migrate_isolate(mt)))
881 __mod_zone_freepage_state(zone, -(1UL << order), mt);
882
883 del_page_from_free_list(free_page, zone, order);
884 for (pfn = free_page_pfn;
885 pfn < free_page_pfn + (1UL << order);) {
886 int mt = get_pfnblock_migratetype(pfn_to_page(pfn), pfn);
887
888 free_page_order = min_t(unsigned int,
889 pfn ? __ffs(pfn) : order,
890 __fls(split_pfn_offset));
891 __free_one_page(pfn_to_page(pfn), pfn, zone, free_page_order,
892 mt, FPI_NONE);
893 pfn += 1UL << free_page_order;
894 split_pfn_offset -= (1UL << free_page_order);
895 /* we have done the first part, now switch to second part */
896 if (split_pfn_offset == 0)
897 split_pfn_offset = (1UL << order) - (pfn - free_page_pfn);
898 }
899out:
900 spin_unlock_irqrestore(&zone->lock, flags);
901 return ret;
902}
903/*
904 * A bad page could be due to a number of fields. Instead of multiple branches,
905 * try and check multiple fields with one check. The caller must do a detailed
906 * check if necessary.
907 */
908static inline bool page_expected_state(struct page *page,
909 unsigned long check_flags)
910{
911 if (unlikely(atomic_read(&page->_mapcount) != -1))
912 return false;
913
914 if (unlikely((unsigned long)page->mapping |
915 page_ref_count(page) |
916#ifdef CONFIG_MEMCG
917 page->memcg_data |
918#endif
919#ifdef CONFIG_PAGE_POOL
920 ((page->pp_magic & ~0x3UL) == PP_SIGNATURE) |
921#endif
922 (page->flags & check_flags)))
923 return false;
924
925 return true;
926}
927
928static const char *page_bad_reason(struct page *page, unsigned long flags)
929{
930 const char *bad_reason = NULL;
931
932 if (unlikely(atomic_read(&page->_mapcount) != -1))
933 bad_reason = "nonzero mapcount";
934 if (unlikely(page->mapping != NULL))
935 bad_reason = "non-NULL mapping";
936 if (unlikely(page_ref_count(page) != 0))
937 bad_reason = "nonzero _refcount";
938 if (unlikely(page->flags & flags)) {
939 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
940 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
941 else
942 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
943 }
944#ifdef CONFIG_MEMCG
945 if (unlikely(page->memcg_data))
946 bad_reason = "page still charged to cgroup";
947#endif
948#ifdef CONFIG_PAGE_POOL
949 if (unlikely((page->pp_magic & ~0x3UL) == PP_SIGNATURE))
950 bad_reason = "page_pool leak";
951#endif
952 return bad_reason;
953}
954
955static void free_page_is_bad_report(struct page *page)
956{
957 bad_page(page,
958 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
959}
960
961static inline bool free_page_is_bad(struct page *page)
962{
963 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
964 return false;
965
966 /* Something has gone sideways, find it */
967 free_page_is_bad_report(page);
968 return true;
969}
970
971static inline bool is_check_pages_enabled(void)
972{
973 return static_branch_unlikely(&check_pages_enabled);
974}
975
976static int free_tail_page_prepare(struct page *head_page, struct page *page)
977{
978 struct folio *folio = (struct folio *)head_page;
979 int ret = 1;
980
981 /*
982 * We rely page->lru.next never has bit 0 set, unless the page
983 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
984 */
985 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
986
987 if (!is_check_pages_enabled()) {
988 ret = 0;
989 goto out;
990 }
991 switch (page - head_page) {
992 case 1:
993 /* the first tail page: these may be in place of ->mapping */
994 if (unlikely(folio_entire_mapcount(folio))) {
995 bad_page(page, "nonzero entire_mapcount");
996 goto out;
997 }
998 if (unlikely(atomic_read(&folio->_nr_pages_mapped))) {
999 bad_page(page, "nonzero nr_pages_mapped");
1000 goto out;
1001 }
1002 if (unlikely(atomic_read(&folio->_pincount))) {
1003 bad_page(page, "nonzero pincount");
1004 goto out;
1005 }
1006 break;
1007 case 2:
1008 /*
1009 * the second tail page: ->mapping is
1010 * deferred_list.next -- ignore value.
1011 */
1012 break;
1013 default:
1014 if (page->mapping != TAIL_MAPPING) {
1015 bad_page(page, "corrupted mapping in tail page");
1016 goto out;
1017 }
1018 break;
1019 }
1020 if (unlikely(!PageTail(page))) {
1021 bad_page(page, "PageTail not set");
1022 goto out;
1023 }
1024 if (unlikely(compound_head(page) != head_page)) {
1025 bad_page(page, "compound_head not consistent");
1026 goto out;
1027 }
1028 ret = 0;
1029out:
1030 page->mapping = NULL;
1031 clear_compound_head(page);
1032 return ret;
1033}
1034
1035/*
1036 * Skip KASAN memory poisoning when either:
1037 *
1038 * 1. For generic KASAN: deferred memory initialization has not yet completed.
1039 * Tag-based KASAN modes skip pages freed via deferred memory initialization
1040 * using page tags instead (see below).
1041 * 2. For tag-based KASAN modes: the page has a match-all KASAN tag, indicating
1042 * that error detection is disabled for accesses via the page address.
1043 *
1044 * Pages will have match-all tags in the following circumstances:
1045 *
1046 * 1. Pages are being initialized for the first time, including during deferred
1047 * memory init; see the call to page_kasan_tag_reset in __init_single_page.
1048 * 2. The allocation was not unpoisoned due to __GFP_SKIP_KASAN, with the
1049 * exception of pages unpoisoned by kasan_unpoison_vmalloc.
1050 * 3. The allocation was excluded from being checked due to sampling,
1051 * see the call to kasan_unpoison_pages.
1052 *
1053 * Poisoning pages during deferred memory init will greatly lengthen the
1054 * process and cause problem in large memory systems as the deferred pages
1055 * initialization is done with interrupt disabled.
1056 *
1057 * Assuming that there will be no reference to those newly initialized
1058 * pages before they are ever allocated, this should have no effect on
1059 * KASAN memory tracking as the poison will be properly inserted at page
1060 * allocation time. The only corner case is when pages are allocated by
1061 * on-demand allocation and then freed again before the deferred pages
1062 * initialization is done, but this is not likely to happen.
1063 */
1064static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
1065{
1066 if (IS_ENABLED(CONFIG_KASAN_GENERIC))
1067 return deferred_pages_enabled();
1068
1069 return page_kasan_tag(page) == KASAN_TAG_KERNEL;
1070}
1071
1072static void kernel_init_pages(struct page *page, int numpages)
1073{
1074 int i;
1075
1076 /* s390's use of memset() could override KASAN redzones. */
1077 kasan_disable_current();
1078 for (i = 0; i < numpages; i++)
1079 clear_highpage_kasan_tagged(page + i);
1080 kasan_enable_current();
1081}
1082
1083static __always_inline bool free_pages_prepare(struct page *page,
1084 unsigned int order, fpi_t fpi_flags)
1085{
1086 int bad = 0;
1087 bool skip_kasan_poison = should_skip_kasan_poison(page, fpi_flags);
1088 bool init = want_init_on_free();
1089 bool compound = PageCompound(page);
1090
1091 VM_BUG_ON_PAGE(PageTail(page), page);
1092
1093 trace_mm_page_free(page, order);
1094 kmsan_free_page(page, order);
1095
1096 if (memcg_kmem_online() && PageMemcgKmem(page))
1097 __memcg_kmem_uncharge_page(page, order);
1098
1099 if (unlikely(PageHWPoison(page)) && !order) {
1100 /* Do not let hwpoison pages hit pcplists/buddy */
1101 reset_page_owner(page, order);
1102 page_table_check_free(page, order);
1103 return false;
1104 }
1105
1106 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1107
1108 /*
1109 * Check tail pages before head page information is cleared to
1110 * avoid checking PageCompound for order-0 pages.
1111 */
1112 if (unlikely(order)) {
1113 int i;
1114
1115 if (compound)
1116 page[1].flags &= ~PAGE_FLAGS_SECOND;
1117 for (i = 1; i < (1 << order); i++) {
1118 if (compound)
1119 bad += free_tail_page_prepare(page, page + i);
1120 if (is_check_pages_enabled()) {
1121 if (free_page_is_bad(page + i)) {
1122 bad++;
1123 continue;
1124 }
1125 }
1126 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1127 }
1128 }
1129 if (PageMappingFlags(page))
1130 page->mapping = NULL;
1131 if (is_check_pages_enabled()) {
1132 if (free_page_is_bad(page))
1133 bad++;
1134 if (bad)
1135 return false;
1136 }
1137
1138 page_cpupid_reset_last(page);
1139 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1140 reset_page_owner(page, order);
1141 page_table_check_free(page, order);
1142
1143 if (!PageHighMem(page)) {
1144 debug_check_no_locks_freed(page_address(page),
1145 PAGE_SIZE << order);
1146 debug_check_no_obj_freed(page_address(page),
1147 PAGE_SIZE << order);
1148 }
1149
1150 kernel_poison_pages(page, 1 << order);
1151
1152 /*
1153 * As memory initialization might be integrated into KASAN,
1154 * KASAN poisoning and memory initialization code must be
1155 * kept together to avoid discrepancies in behavior.
1156 *
1157 * With hardware tag-based KASAN, memory tags must be set before the
1158 * page becomes unavailable via debug_pagealloc or arch_free_page.
1159 */
1160 if (!skip_kasan_poison) {
1161 kasan_poison_pages(page, order, init);
1162
1163 /* Memory is already initialized if KASAN did it internally. */
1164 if (kasan_has_integrated_init())
1165 init = false;
1166 }
1167 if (init)
1168 kernel_init_pages(page, 1 << order);
1169
1170 /*
1171 * arch_free_page() can make the page's contents inaccessible. s390
1172 * does this. So nothing which can access the page's contents should
1173 * happen after this.
1174 */
1175 arch_free_page(page, order);
1176
1177 debug_pagealloc_unmap_pages(page, 1 << order);
1178
1179 return true;
1180}
1181
1182/*
1183 * Frees a number of pages from the PCP lists
1184 * Assumes all pages on list are in same zone.
1185 * count is the number of pages to free.
1186 */
1187static void free_pcppages_bulk(struct zone *zone, int count,
1188 struct per_cpu_pages *pcp,
1189 int pindex)
1190{
1191 unsigned long flags;
1192 unsigned int order;
1193 bool isolated_pageblocks;
1194 struct page *page;
1195
1196 /*
1197 * Ensure proper count is passed which otherwise would stuck in the
1198 * below while (list_empty(list)) loop.
1199 */
1200 count = min(pcp->count, count);
1201
1202 /* Ensure requested pindex is drained first. */
1203 pindex = pindex - 1;
1204
1205 spin_lock_irqsave(&zone->lock, flags);
1206 isolated_pageblocks = has_isolate_pageblock(zone);
1207
1208 while (count > 0) {
1209 struct list_head *list;
1210 int nr_pages;
1211
1212 /* Remove pages from lists in a round-robin fashion. */
1213 do {
1214 if (++pindex > NR_PCP_LISTS - 1)
1215 pindex = 0;
1216 list = &pcp->lists[pindex];
1217 } while (list_empty(list));
1218
1219 order = pindex_to_order(pindex);
1220 nr_pages = 1 << order;
1221 do {
1222 int mt;
1223
1224 page = list_last_entry(list, struct page, pcp_list);
1225 mt = get_pcppage_migratetype(page);
1226
1227 /* must delete to avoid corrupting pcp list */
1228 list_del(&page->pcp_list);
1229 count -= nr_pages;
1230 pcp->count -= nr_pages;
1231
1232 /* MIGRATE_ISOLATE page should not go to pcplists */
1233 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1234 /* Pageblock could have been isolated meanwhile */
1235 if (unlikely(isolated_pageblocks))
1236 mt = get_pageblock_migratetype(page);
1237
1238 __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1239 trace_mm_page_pcpu_drain(page, order, mt);
1240 } while (count > 0 && !list_empty(list));
1241 }
1242
1243 spin_unlock_irqrestore(&zone->lock, flags);
1244}
1245
1246static void free_one_page(struct zone *zone,
1247 struct page *page, unsigned long pfn,
1248 unsigned int order,
1249 int migratetype, fpi_t fpi_flags)
1250{
1251 unsigned long flags;
1252
1253 spin_lock_irqsave(&zone->lock, flags);
1254 if (unlikely(has_isolate_pageblock(zone) ||
1255 is_migrate_isolate(migratetype))) {
1256 migratetype = get_pfnblock_migratetype(page, pfn);
1257 }
1258 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1259 spin_unlock_irqrestore(&zone->lock, flags);
1260}
1261
1262static void __free_pages_ok(struct page *page, unsigned int order,
1263 fpi_t fpi_flags)
1264{
1265 int migratetype;
1266 unsigned long pfn = page_to_pfn(page);
1267 struct zone *zone = page_zone(page);
1268
1269 if (!free_pages_prepare(page, order, fpi_flags))
1270 return;
1271
1272 /*
1273 * Calling get_pfnblock_migratetype() without spin_lock_irqsave() here
1274 * is used to avoid calling get_pfnblock_migratetype() under the lock.
1275 * This will reduce the lock holding time.
1276 */
1277 migratetype = get_pfnblock_migratetype(page, pfn);
1278
1279 free_one_page(zone, page, pfn, order, migratetype, fpi_flags);
1280
1281 __count_vm_events(PGFREE, 1 << order);
1282}
1283
1284void __free_pages_core(struct page *page, unsigned int order)
1285{
1286 unsigned int nr_pages = 1 << order;
1287 struct page *p = page;
1288 unsigned int loop;
1289
1290 /*
1291 * When initializing the memmap, __init_single_page() sets the refcount
1292 * of all pages to 1 ("allocated"/"not free"). We have to set the
1293 * refcount of all involved pages to 0.
1294 */
1295 prefetchw(p);
1296 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1297 prefetchw(p + 1);
1298 __ClearPageReserved(p);
1299 set_page_count(p, 0);
1300 }
1301 __ClearPageReserved(p);
1302 set_page_count(p, 0);
1303
1304 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1305
1306 if (page_contains_unaccepted(page, order)) {
1307 if (order == MAX_PAGE_ORDER && __free_unaccepted(page))
1308 return;
1309
1310 accept_page(page, order);
1311 }
1312
1313 /*
1314 * Bypass PCP and place fresh pages right to the tail, primarily
1315 * relevant for memory onlining.
1316 */
1317 __free_pages_ok(page, order, FPI_TO_TAIL);
1318}
1319
1320/*
1321 * Check that the whole (or subset of) a pageblock given by the interval of
1322 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1323 * with the migration of free compaction scanner.
1324 *
1325 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1326 *
1327 * It's possible on some configurations to have a setup like node0 node1 node0
1328 * i.e. it's possible that all pages within a zones range of pages do not
1329 * belong to a single zone. We assume that a border between node0 and node1
1330 * can occur within a single pageblock, but not a node0 node1 node0
1331 * interleaving within a single pageblock. It is therefore sufficient to check
1332 * the first and last page of a pageblock and avoid checking each individual
1333 * page in a pageblock.
1334 *
1335 * Note: the function may return non-NULL struct page even for a page block
1336 * which contains a memory hole (i.e. there is no physical memory for a subset
1337 * of the pfn range). For example, if the pageblock order is MAX_PAGE_ORDER, which
1338 * will fall into 2 sub-sections, and the end pfn of the pageblock may be hole
1339 * even though the start pfn is online and valid. This should be safe most of
1340 * the time because struct pages are still initialized via init_unavailable_range()
1341 * and pfn walkers shouldn't touch any physical memory range for which they do
1342 * not recognize any specific metadata in struct pages.
1343 */
1344struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1345 unsigned long end_pfn, struct zone *zone)
1346{
1347 struct page *start_page;
1348 struct page *end_page;
1349
1350 /* end_pfn is one past the range we are checking */
1351 end_pfn--;
1352
1353 if (!pfn_valid(end_pfn))
1354 return NULL;
1355
1356 start_page = pfn_to_online_page(start_pfn);
1357 if (!start_page)
1358 return NULL;
1359
1360 if (page_zone(start_page) != zone)
1361 return NULL;
1362
1363 end_page = pfn_to_page(end_pfn);
1364
1365 /* This gives a shorter code than deriving page_zone(end_page) */
1366 if (page_zone_id(start_page) != page_zone_id(end_page))
1367 return NULL;
1368
1369 return start_page;
1370}
1371
1372/*
1373 * The order of subdivision here is critical for the IO subsystem.
1374 * Please do not alter this order without good reasons and regression
1375 * testing. Specifically, as large blocks of memory are subdivided,
1376 * the order in which smaller blocks are delivered depends on the order
1377 * they're subdivided in this function. This is the primary factor
1378 * influencing the order in which pages are delivered to the IO
1379 * subsystem according to empirical testing, and this is also justified
1380 * by considering the behavior of a buddy system containing a single
1381 * large block of memory acted on by a series of small allocations.
1382 * This behavior is a critical factor in sglist merging's success.
1383 *
1384 * -- nyc
1385 */
1386static inline void expand(struct zone *zone, struct page *page,
1387 int low, int high, int migratetype)
1388{
1389 unsigned long size = 1 << high;
1390
1391 while (high > low) {
1392 high--;
1393 size >>= 1;
1394 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1395
1396 /*
1397 * Mark as guard pages (or page), that will allow to
1398 * merge back to allocator when buddy will be freed.
1399 * Corresponding page table entries will not be touched,
1400 * pages will stay not present in virtual address space
1401 */
1402 if (set_page_guard(zone, &page[size], high, migratetype))
1403 continue;
1404
1405 add_to_free_list(&page[size], zone, high, migratetype);
1406 set_buddy_order(&page[size], high);
1407 }
1408}
1409
1410static void check_new_page_bad(struct page *page)
1411{
1412 if (unlikely(page->flags & __PG_HWPOISON)) {
1413 /* Don't complain about hwpoisoned pages */
1414 page_mapcount_reset(page); /* remove PageBuddy */
1415 return;
1416 }
1417
1418 bad_page(page,
1419 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
1420}
1421
1422/*
1423 * This page is about to be returned from the page allocator
1424 */
1425static int check_new_page(struct page *page)
1426{
1427 if (likely(page_expected_state(page,
1428 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1429 return 0;
1430
1431 check_new_page_bad(page);
1432 return 1;
1433}
1434
1435static inline bool check_new_pages(struct page *page, unsigned int order)
1436{
1437 if (is_check_pages_enabled()) {
1438 for (int i = 0; i < (1 << order); i++) {
1439 struct page *p = page + i;
1440
1441 if (check_new_page(p))
1442 return true;
1443 }
1444 }
1445
1446 return false;
1447}
1448
1449static inline bool should_skip_kasan_unpoison(gfp_t flags)
1450{
1451 /* Don't skip if a software KASAN mode is enabled. */
1452 if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
1453 IS_ENABLED(CONFIG_KASAN_SW_TAGS))
1454 return false;
1455
1456 /* Skip, if hardware tag-based KASAN is not enabled. */
1457 if (!kasan_hw_tags_enabled())
1458 return true;
1459
1460 /*
1461 * With hardware tag-based KASAN enabled, skip if this has been
1462 * requested via __GFP_SKIP_KASAN.
1463 */
1464 return flags & __GFP_SKIP_KASAN;
1465}
1466
1467static inline bool should_skip_init(gfp_t flags)
1468{
1469 /* Don't skip, if hardware tag-based KASAN is not enabled. */
1470 if (!kasan_hw_tags_enabled())
1471 return false;
1472
1473 /* For hardware tag-based KASAN, skip if requested. */
1474 return (flags & __GFP_SKIP_ZERO);
1475}
1476
1477inline void post_alloc_hook(struct page *page, unsigned int order,
1478 gfp_t gfp_flags)
1479{
1480 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) &&
1481 !should_skip_init(gfp_flags);
1482 bool zero_tags = init && (gfp_flags & __GFP_ZEROTAGS);
1483 int i;
1484
1485 set_page_private(page, 0);
1486 set_page_refcounted(page);
1487
1488 arch_alloc_page(page, order);
1489 debug_pagealloc_map_pages(page, 1 << order);
1490
1491 /*
1492 * Page unpoisoning must happen before memory initialization.
1493 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
1494 * allocations and the page unpoisoning code will complain.
1495 */
1496 kernel_unpoison_pages(page, 1 << order);
1497
1498 /*
1499 * As memory initialization might be integrated into KASAN,
1500 * KASAN unpoisoning and memory initializion code must be
1501 * kept together to avoid discrepancies in behavior.
1502 */
1503
1504 /*
1505 * If memory tags should be zeroed
1506 * (which happens only when memory should be initialized as well).
1507 */
1508 if (zero_tags) {
1509 /* Initialize both memory and memory tags. */
1510 for (i = 0; i != 1 << order; ++i)
1511 tag_clear_highpage(page + i);
1512
1513 /* Take note that memory was initialized by the loop above. */
1514 init = false;
1515 }
1516 if (!should_skip_kasan_unpoison(gfp_flags) &&
1517 kasan_unpoison_pages(page, order, init)) {
1518 /* Take note that memory was initialized by KASAN. */
1519 if (kasan_has_integrated_init())
1520 init = false;
1521 } else {
1522 /*
1523 * If memory tags have not been set by KASAN, reset the page
1524 * tags to ensure page_address() dereferencing does not fault.
1525 */
1526 for (i = 0; i != 1 << order; ++i)
1527 page_kasan_tag_reset(page + i);
1528 }
1529 /* If memory is still not initialized, initialize it now. */
1530 if (init)
1531 kernel_init_pages(page, 1 << order);
1532
1533 set_page_owner(page, order, gfp_flags);
1534 page_table_check_alloc(page, order);
1535}
1536
1537static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1538 unsigned int alloc_flags)
1539{
1540 post_alloc_hook(page, order, gfp_flags);
1541
1542 if (order && (gfp_flags & __GFP_COMP))
1543 prep_compound_page(page, order);
1544
1545 /*
1546 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1547 * allocate the page. The expectation is that the caller is taking
1548 * steps that will free more memory. The caller should avoid the page
1549 * being used for !PFMEMALLOC purposes.
1550 */
1551 if (alloc_flags & ALLOC_NO_WATERMARKS)
1552 set_page_pfmemalloc(page);
1553 else
1554 clear_page_pfmemalloc(page);
1555}
1556
1557/*
1558 * Go through the free lists for the given migratetype and remove
1559 * the smallest available page from the freelists
1560 */
1561static __always_inline
1562struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1563 int migratetype)
1564{
1565 unsigned int current_order;
1566 struct free_area *area;
1567 struct page *page;
1568
1569 /* Find a page of the appropriate size in the preferred list */
1570 for (current_order = order; current_order < NR_PAGE_ORDERS; ++current_order) {
1571 area = &(zone->free_area[current_order]);
1572 page = get_page_from_free_area(area, migratetype);
1573 if (!page)
1574 continue;
1575 del_page_from_free_list(page, zone, current_order);
1576 expand(zone, page, order, current_order, migratetype);
1577 set_pcppage_migratetype(page, migratetype);
1578 trace_mm_page_alloc_zone_locked(page, order, migratetype,
1579 pcp_allowed_order(order) &&
1580 migratetype < MIGRATE_PCPTYPES);
1581 return page;
1582 }
1583
1584 return NULL;
1585}
1586
1587
1588/*
1589 * This array describes the order lists are fallen back to when
1590 * the free lists for the desirable migrate type are depleted
1591 *
1592 * The other migratetypes do not have fallbacks.
1593 */
1594static int fallbacks[MIGRATE_TYPES][MIGRATE_PCPTYPES - 1] = {
1595 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE },
1596 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE },
1597 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE },
1598};
1599
1600#ifdef CONFIG_CMA
1601static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1602 unsigned int order)
1603{
1604 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1605}
1606#else
1607static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1608 unsigned int order) { return NULL; }
1609#endif
1610
1611/*
1612 * Move the free pages in a range to the freelist tail of the requested type.
1613 * Note that start_page and end_pages are not aligned on a pageblock
1614 * boundary. If alignment is required, use move_freepages_block()
1615 */
1616static int move_freepages(struct zone *zone,
1617 unsigned long start_pfn, unsigned long end_pfn,
1618 int migratetype, int *num_movable)
1619{
1620 struct page *page;
1621 unsigned long pfn;
1622 unsigned int order;
1623 int pages_moved = 0;
1624
1625 for (pfn = start_pfn; pfn <= end_pfn;) {
1626 page = pfn_to_page(pfn);
1627 if (!PageBuddy(page)) {
1628 /*
1629 * We assume that pages that could be isolated for
1630 * migration are movable. But we don't actually try
1631 * isolating, as that would be expensive.
1632 */
1633 if (num_movable &&
1634 (PageLRU(page) || __PageMovable(page)))
1635 (*num_movable)++;
1636 pfn++;
1637 continue;
1638 }
1639
1640 /* Make sure we are not inadvertently changing nodes */
1641 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1642 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
1643
1644 order = buddy_order(page);
1645 move_to_free_list(page, zone, order, migratetype);
1646 pfn += 1 << order;
1647 pages_moved += 1 << order;
1648 }
1649
1650 return pages_moved;
1651}
1652
1653int move_freepages_block(struct zone *zone, struct page *page,
1654 int migratetype, int *num_movable)
1655{
1656 unsigned long start_pfn, end_pfn, pfn;
1657
1658 if (num_movable)
1659 *num_movable = 0;
1660
1661 pfn = page_to_pfn(page);
1662 start_pfn = pageblock_start_pfn(pfn);
1663 end_pfn = pageblock_end_pfn(pfn) - 1;
1664
1665 /* Do not cross zone boundaries */
1666 if (!zone_spans_pfn(zone, start_pfn))
1667 start_pfn = pfn;
1668 if (!zone_spans_pfn(zone, end_pfn))
1669 return 0;
1670
1671 return move_freepages(zone, start_pfn, end_pfn, migratetype,
1672 num_movable);
1673}
1674
1675static void change_pageblock_range(struct page *pageblock_page,
1676 int start_order, int migratetype)
1677{
1678 int nr_pageblocks = 1 << (start_order - pageblock_order);
1679
1680 while (nr_pageblocks--) {
1681 set_pageblock_migratetype(pageblock_page, migratetype);
1682 pageblock_page += pageblock_nr_pages;
1683 }
1684}
1685
1686/*
1687 * When we are falling back to another migratetype during allocation, try to
1688 * steal extra free pages from the same pageblocks to satisfy further
1689 * allocations, instead of polluting multiple pageblocks.
1690 *
1691 * If we are stealing a relatively large buddy page, it is likely there will
1692 * be more free pages in the pageblock, so try to steal them all. For
1693 * reclaimable and unmovable allocations, we steal regardless of page size,
1694 * as fragmentation caused by those allocations polluting movable pageblocks
1695 * is worse than movable allocations stealing from unmovable and reclaimable
1696 * pageblocks.
1697 */
1698static bool can_steal_fallback(unsigned int order, int start_mt)
1699{
1700 /*
1701 * Leaving this order check is intended, although there is
1702 * relaxed order check in next check. The reason is that
1703 * we can actually steal whole pageblock if this condition met,
1704 * but, below check doesn't guarantee it and that is just heuristic
1705 * so could be changed anytime.
1706 */
1707 if (order >= pageblock_order)
1708 return true;
1709
1710 if (order >= pageblock_order / 2 ||
1711 start_mt == MIGRATE_RECLAIMABLE ||
1712 start_mt == MIGRATE_UNMOVABLE ||
1713 page_group_by_mobility_disabled)
1714 return true;
1715
1716 return false;
1717}
1718
1719static inline bool boost_watermark(struct zone *zone)
1720{
1721 unsigned long max_boost;
1722
1723 if (!watermark_boost_factor)
1724 return false;
1725 /*
1726 * Don't bother in zones that are unlikely to produce results.
1727 * On small machines, including kdump capture kernels running
1728 * in a small area, boosting the watermark can cause an out of
1729 * memory situation immediately.
1730 */
1731 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
1732 return false;
1733
1734 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
1735 watermark_boost_factor, 10000);
1736
1737 /*
1738 * high watermark may be uninitialised if fragmentation occurs
1739 * very early in boot so do not boost. We do not fall
1740 * through and boost by pageblock_nr_pages as failing
1741 * allocations that early means that reclaim is not going
1742 * to help and it may even be impossible to reclaim the
1743 * boosted watermark resulting in a hang.
1744 */
1745 if (!max_boost)
1746 return false;
1747
1748 max_boost = max(pageblock_nr_pages, max_boost);
1749
1750 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
1751 max_boost);
1752
1753 return true;
1754}
1755
1756/*
1757 * This function implements actual steal behaviour. If order is large enough,
1758 * we can steal whole pageblock. If not, we first move freepages in this
1759 * pageblock to our migratetype and determine how many already-allocated pages
1760 * are there in the pageblock with a compatible migratetype. If at least half
1761 * of pages are free or compatible, we can change migratetype of the pageblock
1762 * itself, so pages freed in the future will be put on the correct free list.
1763 */
1764static void steal_suitable_fallback(struct zone *zone, struct page *page,
1765 unsigned int alloc_flags, int start_type, bool whole_block)
1766{
1767 unsigned int current_order = buddy_order(page);
1768 int free_pages, movable_pages, alike_pages;
1769 int old_block_type;
1770
1771 old_block_type = get_pageblock_migratetype(page);
1772
1773 /*
1774 * This can happen due to races and we want to prevent broken
1775 * highatomic accounting.
1776 */
1777 if (is_migrate_highatomic(old_block_type))
1778 goto single_page;
1779
1780 /* Take ownership for orders >= pageblock_order */
1781 if (current_order >= pageblock_order) {
1782 change_pageblock_range(page, current_order, start_type);
1783 goto single_page;
1784 }
1785
1786 /*
1787 * Boost watermarks to increase reclaim pressure to reduce the
1788 * likelihood of future fallbacks. Wake kswapd now as the node
1789 * may be balanced overall and kswapd will not wake naturally.
1790 */
1791 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
1792 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
1793
1794 /* We are not allowed to try stealing from the whole block */
1795 if (!whole_block)
1796 goto single_page;
1797
1798 free_pages = move_freepages_block(zone, page, start_type,
1799 &movable_pages);
1800 /* moving whole block can fail due to zone boundary conditions */
1801 if (!free_pages)
1802 goto single_page;
1803
1804 /*
1805 * Determine how many pages are compatible with our allocation.
1806 * For movable allocation, it's the number of movable pages which
1807 * we just obtained. For other types it's a bit more tricky.
1808 */
1809 if (start_type == MIGRATE_MOVABLE) {
1810 alike_pages = movable_pages;
1811 } else {
1812 /*
1813 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
1814 * to MOVABLE pageblock, consider all non-movable pages as
1815 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
1816 * vice versa, be conservative since we can't distinguish the
1817 * exact migratetype of non-movable pages.
1818 */
1819 if (old_block_type == MIGRATE_MOVABLE)
1820 alike_pages = pageblock_nr_pages
1821 - (free_pages + movable_pages);
1822 else
1823 alike_pages = 0;
1824 }
1825 /*
1826 * If a sufficient number of pages in the block are either free or of
1827 * compatible migratability as our allocation, claim the whole block.
1828 */
1829 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
1830 page_group_by_mobility_disabled)
1831 set_pageblock_migratetype(page, start_type);
1832
1833 return;
1834
1835single_page:
1836 move_to_free_list(page, zone, current_order, start_type);
1837}
1838
1839/*
1840 * Check whether there is a suitable fallback freepage with requested order.
1841 * If only_stealable is true, this function returns fallback_mt only if
1842 * we can steal other freepages all together. This would help to reduce
1843 * fragmentation due to mixed migratetype pages in one pageblock.
1844 */
1845int find_suitable_fallback(struct free_area *area, unsigned int order,
1846 int migratetype, bool only_stealable, bool *can_steal)
1847{
1848 int i;
1849 int fallback_mt;
1850
1851 if (area->nr_free == 0)
1852 return -1;
1853
1854 *can_steal = false;
1855 for (i = 0; i < MIGRATE_PCPTYPES - 1 ; i++) {
1856 fallback_mt = fallbacks[migratetype][i];
1857 if (free_area_empty(area, fallback_mt))
1858 continue;
1859
1860 if (can_steal_fallback(order, migratetype))
1861 *can_steal = true;
1862
1863 if (!only_stealable)
1864 return fallback_mt;
1865
1866 if (*can_steal)
1867 return fallback_mt;
1868 }
1869
1870 return -1;
1871}
1872
1873/*
1874 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1875 * there are no empty page blocks that contain a page with a suitable order
1876 */
1877static void reserve_highatomic_pageblock(struct page *page, struct zone *zone)
1878{
1879 int mt;
1880 unsigned long max_managed, flags;
1881
1882 /*
1883 * The number reserved as: minimum is 1 pageblock, maximum is
1884 * roughly 1% of a zone. But if 1% of a zone falls below a
1885 * pageblock size, then don't reserve any pageblocks.
1886 * Check is race-prone but harmless.
1887 */
1888 if ((zone_managed_pages(zone) / 100) < pageblock_nr_pages)
1889 return;
1890 max_managed = ALIGN((zone_managed_pages(zone) / 100), pageblock_nr_pages);
1891 if (zone->nr_reserved_highatomic >= max_managed)
1892 return;
1893
1894 spin_lock_irqsave(&zone->lock, flags);
1895
1896 /* Recheck the nr_reserved_highatomic limit under the lock */
1897 if (zone->nr_reserved_highatomic >= max_managed)
1898 goto out_unlock;
1899
1900 /* Yoink! */
1901 mt = get_pageblock_migratetype(page);
1902 /* Only reserve normal pageblocks (i.e., they can merge with others) */
1903 if (migratetype_is_mergeable(mt)) {
1904 zone->nr_reserved_highatomic += pageblock_nr_pages;
1905 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1906 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
1907 }
1908
1909out_unlock:
1910 spin_unlock_irqrestore(&zone->lock, flags);
1911}
1912
1913/*
1914 * Used when an allocation is about to fail under memory pressure. This
1915 * potentially hurts the reliability of high-order allocations when under
1916 * intense memory pressure but failed atomic allocations should be easier
1917 * to recover from than an OOM.
1918 *
1919 * If @force is true, try to unreserve a pageblock even though highatomic
1920 * pageblock is exhausted.
1921 */
1922static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
1923 bool force)
1924{
1925 struct zonelist *zonelist = ac->zonelist;
1926 unsigned long flags;
1927 struct zoneref *z;
1928 struct zone *zone;
1929 struct page *page;
1930 int order;
1931 bool ret;
1932
1933 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
1934 ac->nodemask) {
1935 /*
1936 * Preserve at least one pageblock unless memory pressure
1937 * is really high.
1938 */
1939 if (!force && zone->nr_reserved_highatomic <=
1940 pageblock_nr_pages)
1941 continue;
1942
1943 spin_lock_irqsave(&zone->lock, flags);
1944 for (order = 0; order < NR_PAGE_ORDERS; order++) {
1945 struct free_area *area = &(zone->free_area[order]);
1946
1947 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
1948 if (!page)
1949 continue;
1950
1951 /*
1952 * In page freeing path, migratetype change is racy so
1953 * we can counter several free pages in a pageblock
1954 * in this loop although we changed the pageblock type
1955 * from highatomic to ac->migratetype. So we should
1956 * adjust the count once.
1957 */
1958 if (is_migrate_highatomic_page(page)) {
1959 /*
1960 * It should never happen but changes to
1961 * locking could inadvertently allow a per-cpu
1962 * drain to add pages to MIGRATE_HIGHATOMIC
1963 * while unreserving so be safe and watch for
1964 * underflows.
1965 */
1966 zone->nr_reserved_highatomic -= min(
1967 pageblock_nr_pages,
1968 zone->nr_reserved_highatomic);
1969 }
1970
1971 /*
1972 * Convert to ac->migratetype and avoid the normal
1973 * pageblock stealing heuristics. Minimally, the caller
1974 * is doing the work and needs the pages. More
1975 * importantly, if the block was always converted to
1976 * MIGRATE_UNMOVABLE or another type then the number
1977 * of pageblocks that cannot be completely freed
1978 * may increase.
1979 */
1980 set_pageblock_migratetype(page, ac->migratetype);
1981 ret = move_freepages_block(zone, page, ac->migratetype,
1982 NULL);
1983 if (ret) {
1984 spin_unlock_irqrestore(&zone->lock, flags);
1985 return ret;
1986 }
1987 }
1988 spin_unlock_irqrestore(&zone->lock, flags);
1989 }
1990
1991 return false;
1992}
1993
1994/*
1995 * Try finding a free buddy page on the fallback list and put it on the free
1996 * list of requested migratetype, possibly along with other pages from the same
1997 * block, depending on fragmentation avoidance heuristics. Returns true if
1998 * fallback was found so that __rmqueue_smallest() can grab it.
1999 *
2000 * The use of signed ints for order and current_order is a deliberate
2001 * deviation from the rest of this file, to make the for loop
2002 * condition simpler.
2003 */
2004static __always_inline bool
2005__rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2006 unsigned int alloc_flags)
2007{
2008 struct free_area *area;
2009 int current_order;
2010 int min_order = order;
2011 struct page *page;
2012 int fallback_mt;
2013 bool can_steal;
2014
2015 /*
2016 * Do not steal pages from freelists belonging to other pageblocks
2017 * i.e. orders < pageblock_order. If there are no local zones free,
2018 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2019 */
2020 if (order < pageblock_order && alloc_flags & ALLOC_NOFRAGMENT)
2021 min_order = pageblock_order;
2022
2023 /*
2024 * Find the largest available free page in the other list. This roughly
2025 * approximates finding the pageblock with the most free pages, which
2026 * would be too costly to do exactly.
2027 */
2028 for (current_order = MAX_PAGE_ORDER; current_order >= min_order;
2029 --current_order) {
2030 area = &(zone->free_area[current_order]);
2031 fallback_mt = find_suitable_fallback(area, current_order,
2032 start_migratetype, false, &can_steal);
2033 if (fallback_mt == -1)
2034 continue;
2035
2036 /*
2037 * We cannot steal all free pages from the pageblock and the
2038 * requested migratetype is movable. In that case it's better to
2039 * steal and split the smallest available page instead of the
2040 * largest available page, because even if the next movable
2041 * allocation falls back into a different pageblock than this
2042 * one, it won't cause permanent fragmentation.
2043 */
2044 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2045 && current_order > order)
2046 goto find_smallest;
2047
2048 goto do_steal;
2049 }
2050
2051 return false;
2052
2053find_smallest:
2054 for (current_order = order; current_order < NR_PAGE_ORDERS; current_order++) {
2055 area = &(zone->free_area[current_order]);
2056 fallback_mt = find_suitable_fallback(area, current_order,
2057 start_migratetype, false, &can_steal);
2058 if (fallback_mt != -1)
2059 break;
2060 }
2061
2062 /*
2063 * This should not happen - we already found a suitable fallback
2064 * when looking for the largest page.
2065 */
2066 VM_BUG_ON(current_order > MAX_PAGE_ORDER);
2067
2068do_steal:
2069 page = get_page_from_free_area(area, fallback_mt);
2070
2071 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2072 can_steal);
2073
2074 trace_mm_page_alloc_extfrag(page, order, current_order,
2075 start_migratetype, fallback_mt);
2076
2077 return true;
2078
2079}
2080
2081/*
2082 * Do the hard work of removing an element from the buddy allocator.
2083 * Call me with the zone->lock already held.
2084 */
2085static __always_inline struct page *
2086__rmqueue(struct zone *zone, unsigned int order, int migratetype,
2087 unsigned int alloc_flags)
2088{
2089 struct page *page;
2090
2091 if (IS_ENABLED(CONFIG_CMA)) {
2092 /*
2093 * Balance movable allocations between regular and CMA areas by
2094 * allocating from CMA when over half of the zone's free memory
2095 * is in the CMA area.
2096 */
2097 if (alloc_flags & ALLOC_CMA &&
2098 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2099 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2100 page = __rmqueue_cma_fallback(zone, order);
2101 if (page)
2102 return page;
2103 }
2104 }
2105retry:
2106 page = __rmqueue_smallest(zone, order, migratetype);
2107 if (unlikely(!page)) {
2108 if (alloc_flags & ALLOC_CMA)
2109 page = __rmqueue_cma_fallback(zone, order);
2110
2111 if (!page && __rmqueue_fallback(zone, order, migratetype,
2112 alloc_flags))
2113 goto retry;
2114 }
2115 return page;
2116}
2117
2118/*
2119 * Obtain a specified number of elements from the buddy allocator, all under
2120 * a single hold of the lock, for efficiency. Add them to the supplied list.
2121 * Returns the number of new pages which were placed at *list.
2122 */
2123static int rmqueue_bulk(struct zone *zone, unsigned int order,
2124 unsigned long count, struct list_head *list,
2125 int migratetype, unsigned int alloc_flags)
2126{
2127 unsigned long flags;
2128 int i;
2129
2130 spin_lock_irqsave(&zone->lock, flags);
2131 for (i = 0; i < count; ++i) {
2132 struct page *page = __rmqueue(zone, order, migratetype,
2133 alloc_flags);
2134 if (unlikely(page == NULL))
2135 break;
2136
2137 /*
2138 * Split buddy pages returned by expand() are received here in
2139 * physical page order. The page is added to the tail of
2140 * caller's list. From the callers perspective, the linked list
2141 * is ordered by page number under some conditions. This is
2142 * useful for IO devices that can forward direction from the
2143 * head, thus also in the physical page order. This is useful
2144 * for IO devices that can merge IO requests if the physical
2145 * pages are ordered properly.
2146 */
2147 list_add_tail(&page->pcp_list, list);
2148 if (is_migrate_cma(get_pcppage_migratetype(page)))
2149 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2150 -(1 << order));
2151 }
2152
2153 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2154 spin_unlock_irqrestore(&zone->lock, flags);
2155
2156 return i;
2157}
2158
2159/*
2160 * Called from the vmstat counter updater to decay the PCP high.
2161 * Return whether there are addition works to do.
2162 */
2163int decay_pcp_high(struct zone *zone, struct per_cpu_pages *pcp)
2164{
2165 int high_min, to_drain, batch;
2166 int todo = 0;
2167
2168 high_min = READ_ONCE(pcp->high_min);
2169 batch = READ_ONCE(pcp->batch);
2170 /*
2171 * Decrease pcp->high periodically to try to free possible
2172 * idle PCP pages. And, avoid to free too many pages to
2173 * control latency. This caps pcp->high decrement too.
2174 */
2175 if (pcp->high > high_min) {
2176 pcp->high = max3(pcp->count - (batch << CONFIG_PCP_BATCH_SCALE_MAX),
2177 pcp->high - (pcp->high >> 3), high_min);
2178 if (pcp->high > high_min)
2179 todo++;
2180 }
2181
2182 to_drain = pcp->count - pcp->high;
2183 if (to_drain > 0) {
2184 spin_lock(&pcp->lock);
2185 free_pcppages_bulk(zone, to_drain, pcp, 0);
2186 spin_unlock(&pcp->lock);
2187 todo++;
2188 }
2189
2190 return todo;
2191}
2192
2193#ifdef CONFIG_NUMA
2194/*
2195 * Called from the vmstat counter updater to drain pagesets of this
2196 * currently executing processor on remote nodes after they have
2197 * expired.
2198 */
2199void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2200{
2201 int to_drain, batch;
2202
2203 batch = READ_ONCE(pcp->batch);
2204 to_drain = min(pcp->count, batch);
2205 if (to_drain > 0) {
2206 spin_lock(&pcp->lock);
2207 free_pcppages_bulk(zone, to_drain, pcp, 0);
2208 spin_unlock(&pcp->lock);
2209 }
2210}
2211#endif
2212
2213/*
2214 * Drain pcplists of the indicated processor and zone.
2215 */
2216static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2217{
2218 struct per_cpu_pages *pcp;
2219
2220 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2221 if (pcp->count) {
2222 spin_lock(&pcp->lock);
2223 free_pcppages_bulk(zone, pcp->count, pcp, 0);
2224 spin_unlock(&pcp->lock);
2225 }
2226}
2227
2228/*
2229 * Drain pcplists of all zones on the indicated processor.
2230 */
2231static void drain_pages(unsigned int cpu)
2232{
2233 struct zone *zone;
2234
2235 for_each_populated_zone(zone) {
2236 drain_pages_zone(cpu, zone);
2237 }
2238}
2239
2240/*
2241 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2242 */
2243void drain_local_pages(struct zone *zone)
2244{
2245 int cpu = smp_processor_id();
2246
2247 if (zone)
2248 drain_pages_zone(cpu, zone);
2249 else
2250 drain_pages(cpu);
2251}
2252
2253/*
2254 * The implementation of drain_all_pages(), exposing an extra parameter to
2255 * drain on all cpus.
2256 *
2257 * drain_all_pages() is optimized to only execute on cpus where pcplists are
2258 * not empty. The check for non-emptiness can however race with a free to
2259 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
2260 * that need the guarantee that every CPU has drained can disable the
2261 * optimizing racy check.
2262 */
2263static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
2264{
2265 int cpu;
2266
2267 /*
2268 * Allocate in the BSS so we won't require allocation in
2269 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2270 */
2271 static cpumask_t cpus_with_pcps;
2272
2273 /*
2274 * Do not drain if one is already in progress unless it's specific to
2275 * a zone. Such callers are primarily CMA and memory hotplug and need
2276 * the drain to be complete when the call returns.
2277 */
2278 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2279 if (!zone)
2280 return;
2281 mutex_lock(&pcpu_drain_mutex);
2282 }
2283
2284 /*
2285 * We don't care about racing with CPU hotplug event
2286 * as offline notification will cause the notified
2287 * cpu to drain that CPU pcps and on_each_cpu_mask
2288 * disables preemption as part of its processing
2289 */
2290 for_each_online_cpu(cpu) {
2291 struct per_cpu_pages *pcp;
2292 struct zone *z;
2293 bool has_pcps = false;
2294
2295 if (force_all_cpus) {
2296 /*
2297 * The pcp.count check is racy, some callers need a
2298 * guarantee that no cpu is missed.
2299 */
2300 has_pcps = true;
2301 } else if (zone) {
2302 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2303 if (pcp->count)
2304 has_pcps = true;
2305 } else {
2306 for_each_populated_zone(z) {
2307 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
2308 if (pcp->count) {
2309 has_pcps = true;
2310 break;
2311 }
2312 }
2313 }
2314
2315 if (has_pcps)
2316 cpumask_set_cpu(cpu, &cpus_with_pcps);
2317 else
2318 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2319 }
2320
2321 for_each_cpu(cpu, &cpus_with_pcps) {
2322 if (zone)
2323 drain_pages_zone(cpu, zone);
2324 else
2325 drain_pages(cpu);
2326 }
2327
2328 mutex_unlock(&pcpu_drain_mutex);
2329}
2330
2331/*
2332 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2333 *
2334 * When zone parameter is non-NULL, spill just the single zone's pages.
2335 */
2336void drain_all_pages(struct zone *zone)
2337{
2338 __drain_all_pages(zone, false);
2339}
2340
2341static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
2342 unsigned int order)
2343{
2344 int migratetype;
2345
2346 if (!free_pages_prepare(page, order, FPI_NONE))
2347 return false;
2348
2349 migratetype = get_pfnblock_migratetype(page, pfn);
2350 set_pcppage_migratetype(page, migratetype);
2351 return true;
2352}
2353
2354static int nr_pcp_free(struct per_cpu_pages *pcp, int batch, int high, bool free_high)
2355{
2356 int min_nr_free, max_nr_free;
2357
2358 /* Free as much as possible if batch freeing high-order pages. */
2359 if (unlikely(free_high))
2360 return min(pcp->count, batch << CONFIG_PCP_BATCH_SCALE_MAX);
2361
2362 /* Check for PCP disabled or boot pageset */
2363 if (unlikely(high < batch))
2364 return 1;
2365
2366 /* Leave at least pcp->batch pages on the list */
2367 min_nr_free = batch;
2368 max_nr_free = high - batch;
2369
2370 /*
2371 * Increase the batch number to the number of the consecutive
2372 * freed pages to reduce zone lock contention.
2373 */
2374 batch = clamp_t(int, pcp->free_count, min_nr_free, max_nr_free);
2375
2376 return batch;
2377}
2378
2379static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
2380 int batch, bool free_high)
2381{
2382 int high, high_min, high_max;
2383
2384 high_min = READ_ONCE(pcp->high_min);
2385 high_max = READ_ONCE(pcp->high_max);
2386 high = pcp->high = clamp(pcp->high, high_min, high_max);
2387
2388 if (unlikely(!high))
2389 return 0;
2390
2391 if (unlikely(free_high)) {
2392 pcp->high = max(high - (batch << CONFIG_PCP_BATCH_SCALE_MAX),
2393 high_min);
2394 return 0;
2395 }
2396
2397 /*
2398 * If reclaim is active, limit the number of pages that can be
2399 * stored on pcp lists
2400 */
2401 if (test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags)) {
2402 int free_count = max_t(int, pcp->free_count, batch);
2403
2404 pcp->high = max(high - free_count, high_min);
2405 return min(batch << 2, pcp->high);
2406 }
2407
2408 if (high_min == high_max)
2409 return high;
2410
2411 if (test_bit(ZONE_BELOW_HIGH, &zone->flags)) {
2412 int free_count = max_t(int, pcp->free_count, batch);
2413
2414 pcp->high = max(high - free_count, high_min);
2415 high = max(pcp->count, high_min);
2416 } else if (pcp->count >= high) {
2417 int need_high = pcp->free_count + batch;
2418
2419 /* pcp->high should be large enough to hold batch freed pages */
2420 if (pcp->high < need_high)
2421 pcp->high = clamp(need_high, high_min, high_max);
2422 }
2423
2424 return high;
2425}
2426
2427static void free_unref_page_commit(struct zone *zone, struct per_cpu_pages *pcp,
2428 struct page *page, int migratetype,
2429 unsigned int order)
2430{
2431 int high, batch;
2432 int pindex;
2433 bool free_high = false;
2434
2435 /*
2436 * On freeing, reduce the number of pages that are batch allocated.
2437 * See nr_pcp_alloc() where alloc_factor is increased for subsequent
2438 * allocations.
2439 */
2440 pcp->alloc_factor >>= 1;
2441 __count_vm_events(PGFREE, 1 << order);
2442 pindex = order_to_pindex(migratetype, order);
2443 list_add(&page->pcp_list, &pcp->lists[pindex]);
2444 pcp->count += 1 << order;
2445
2446 batch = READ_ONCE(pcp->batch);
2447 /*
2448 * As high-order pages other than THP's stored on PCP can contribute
2449 * to fragmentation, limit the number stored when PCP is heavily
2450 * freeing without allocation. The remainder after bulk freeing
2451 * stops will be drained from vmstat refresh context.
2452 */
2453 if (order && order <= PAGE_ALLOC_COSTLY_ORDER) {
2454 free_high = (pcp->free_count >= batch &&
2455 (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) &&
2456 (!(pcp->flags & PCPF_FREE_HIGH_BATCH) ||
2457 pcp->count >= READ_ONCE(batch)));
2458 pcp->flags |= PCPF_PREV_FREE_HIGH_ORDER;
2459 } else if (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) {
2460 pcp->flags &= ~PCPF_PREV_FREE_HIGH_ORDER;
2461 }
2462 if (pcp->free_count < (batch << CONFIG_PCP_BATCH_SCALE_MAX))
2463 pcp->free_count += (1 << order);
2464 high = nr_pcp_high(pcp, zone, batch, free_high);
2465 if (pcp->count >= high) {
2466 free_pcppages_bulk(zone, nr_pcp_free(pcp, batch, high, free_high),
2467 pcp, pindex);
2468 if (test_bit(ZONE_BELOW_HIGH, &zone->flags) &&
2469 zone_watermark_ok(zone, 0, high_wmark_pages(zone),
2470 ZONE_MOVABLE, 0))
2471 clear_bit(ZONE_BELOW_HIGH, &zone->flags);
2472 }
2473}
2474
2475/*
2476 * Free a pcp page
2477 */
2478void free_unref_page(struct page *page, unsigned int order)
2479{
2480 unsigned long __maybe_unused UP_flags;
2481 struct per_cpu_pages *pcp;
2482 struct zone *zone;
2483 unsigned long pfn = page_to_pfn(page);
2484 int migratetype, pcpmigratetype;
2485
2486 if (!free_unref_page_prepare(page, pfn, order))
2487 return;
2488
2489 /*
2490 * We only track unmovable, reclaimable and movable on pcp lists.
2491 * Place ISOLATE pages on the isolated list because they are being
2492 * offlined but treat HIGHATOMIC and CMA as movable pages so we can
2493 * get those areas back if necessary. Otherwise, we may have to free
2494 * excessively into the page allocator
2495 */
2496 migratetype = pcpmigratetype = get_pcppage_migratetype(page);
2497 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
2498 if (unlikely(is_migrate_isolate(migratetype))) {
2499 free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
2500 return;
2501 }
2502 pcpmigratetype = MIGRATE_MOVABLE;
2503 }
2504
2505 zone = page_zone(page);
2506 pcp_trylock_prepare(UP_flags);
2507 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2508 if (pcp) {
2509 free_unref_page_commit(zone, pcp, page, pcpmigratetype, order);
2510 pcp_spin_unlock(pcp);
2511 } else {
2512 free_one_page(zone, page, pfn, order, migratetype, FPI_NONE);
2513 }
2514 pcp_trylock_finish(UP_flags);
2515}
2516
2517/*
2518 * Free a list of 0-order pages
2519 */
2520void free_unref_page_list(struct list_head *list)
2521{
2522 unsigned long __maybe_unused UP_flags;
2523 struct page *page, *next;
2524 struct per_cpu_pages *pcp = NULL;
2525 struct zone *locked_zone = NULL;
2526 int batch_count = 0;
2527 int migratetype;
2528
2529 /* Prepare pages for freeing */
2530 list_for_each_entry_safe(page, next, list, lru) {
2531 unsigned long pfn = page_to_pfn(page);
2532 if (!free_unref_page_prepare(page, pfn, 0)) {
2533 list_del(&page->lru);
2534 continue;
2535 }
2536
2537 /*
2538 * Free isolated pages directly to the allocator, see
2539 * comment in free_unref_page.
2540 */
2541 migratetype = get_pcppage_migratetype(page);
2542 if (unlikely(is_migrate_isolate(migratetype))) {
2543 list_del(&page->lru);
2544 free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
2545 continue;
2546 }
2547 }
2548
2549 list_for_each_entry_safe(page, next, list, lru) {
2550 struct zone *zone = page_zone(page);
2551
2552 list_del(&page->lru);
2553 migratetype = get_pcppage_migratetype(page);
2554
2555 /*
2556 * Either different zone requiring a different pcp lock or
2557 * excessive lock hold times when freeing a large list of
2558 * pages.
2559 */
2560 if (zone != locked_zone || batch_count == SWAP_CLUSTER_MAX) {
2561 if (pcp) {
2562 pcp_spin_unlock(pcp);
2563 pcp_trylock_finish(UP_flags);
2564 }
2565
2566 batch_count = 0;
2567
2568 /*
2569 * trylock is necessary as pages may be getting freed
2570 * from IRQ or SoftIRQ context after an IO completion.
2571 */
2572 pcp_trylock_prepare(UP_flags);
2573 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2574 if (unlikely(!pcp)) {
2575 pcp_trylock_finish(UP_flags);
2576 free_one_page(zone, page, page_to_pfn(page),
2577 0, migratetype, FPI_NONE);
2578 locked_zone = NULL;
2579 continue;
2580 }
2581 locked_zone = zone;
2582 }
2583
2584 /*
2585 * Non-isolated types over MIGRATE_PCPTYPES get added
2586 * to the MIGRATE_MOVABLE pcp list.
2587 */
2588 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
2589 migratetype = MIGRATE_MOVABLE;
2590
2591 trace_mm_page_free_batched(page);
2592 free_unref_page_commit(zone, pcp, page, migratetype, 0);
2593 batch_count++;
2594 }
2595
2596 if (pcp) {
2597 pcp_spin_unlock(pcp);
2598 pcp_trylock_finish(UP_flags);
2599 }
2600}
2601
2602/*
2603 * split_page takes a non-compound higher-order page, and splits it into
2604 * n (1<<order) sub-pages: page[0..n]
2605 * Each sub-page must be freed individually.
2606 *
2607 * Note: this is probably too low level an operation for use in drivers.
2608 * Please consult with lkml before using this in your driver.
2609 */
2610void split_page(struct page *page, unsigned int order)
2611{
2612 int i;
2613
2614 VM_BUG_ON_PAGE(PageCompound(page), page);
2615 VM_BUG_ON_PAGE(!page_count(page), page);
2616
2617 for (i = 1; i < (1 << order); i++)
2618 set_page_refcounted(page + i);
2619 split_page_owner(page, 1 << order);
2620 split_page_memcg(page, 1 << order);
2621}
2622EXPORT_SYMBOL_GPL(split_page);
2623
2624int __isolate_free_page(struct page *page, unsigned int order)
2625{
2626 struct zone *zone = page_zone(page);
2627 int mt = get_pageblock_migratetype(page);
2628
2629 if (!is_migrate_isolate(mt)) {
2630 unsigned long watermark;
2631 /*
2632 * Obey watermarks as if the page was being allocated. We can
2633 * emulate a high-order watermark check with a raised order-0
2634 * watermark, because we already know our high-order page
2635 * exists.
2636 */
2637 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
2638 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2639 return 0;
2640
2641 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2642 }
2643
2644 del_page_from_free_list(page, zone, order);
2645
2646 /*
2647 * Set the pageblock if the isolated page is at least half of a
2648 * pageblock
2649 */
2650 if (order >= pageblock_order - 1) {
2651 struct page *endpage = page + (1 << order) - 1;
2652 for (; page < endpage; page += pageblock_nr_pages) {
2653 int mt = get_pageblock_migratetype(page);
2654 /*
2655 * Only change normal pageblocks (i.e., they can merge
2656 * with others)
2657 */
2658 if (migratetype_is_mergeable(mt))
2659 set_pageblock_migratetype(page,
2660 MIGRATE_MOVABLE);
2661 }
2662 }
2663
2664 return 1UL << order;
2665}
2666
2667/**
2668 * __putback_isolated_page - Return a now-isolated page back where we got it
2669 * @page: Page that was isolated
2670 * @order: Order of the isolated page
2671 * @mt: The page's pageblock's migratetype
2672 *
2673 * This function is meant to return a page pulled from the free lists via
2674 * __isolate_free_page back to the free lists they were pulled from.
2675 */
2676void __putback_isolated_page(struct page *page, unsigned int order, int mt)
2677{
2678 struct zone *zone = page_zone(page);
2679
2680 /* zone lock should be held when this function is called */
2681 lockdep_assert_held(&zone->lock);
2682
2683 /* Return isolated page to tail of freelist. */
2684 __free_one_page(page, page_to_pfn(page), zone, order, mt,
2685 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
2686}
2687
2688/*
2689 * Update NUMA hit/miss statistics
2690 */
2691static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
2692 long nr_account)
2693{
2694#ifdef CONFIG_NUMA
2695 enum numa_stat_item local_stat = NUMA_LOCAL;
2696
2697 /* skip numa counters update if numa stats is disabled */
2698 if (!static_branch_likely(&vm_numa_stat_key))
2699 return;
2700
2701 if (zone_to_nid(z) != numa_node_id())
2702 local_stat = NUMA_OTHER;
2703
2704 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
2705 __count_numa_events(z, NUMA_HIT, nr_account);
2706 else {
2707 __count_numa_events(z, NUMA_MISS, nr_account);
2708 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
2709 }
2710 __count_numa_events(z, local_stat, nr_account);
2711#endif
2712}
2713
2714static __always_inline
2715struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone,
2716 unsigned int order, unsigned int alloc_flags,
2717 int migratetype)
2718{
2719 struct page *page;
2720 unsigned long flags;
2721
2722 do {
2723 page = NULL;
2724 spin_lock_irqsave(&zone->lock, flags);
2725 if (alloc_flags & ALLOC_HIGHATOMIC)
2726 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2727 if (!page) {
2728 page = __rmqueue(zone, order, migratetype, alloc_flags);
2729
2730 /*
2731 * If the allocation fails, allow OOM handling access
2732 * to HIGHATOMIC reserves as failing now is worse than
2733 * failing a high-order atomic allocation in the
2734 * future.
2735 */
2736 if (!page && (alloc_flags & ALLOC_OOM))
2737 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2738
2739 if (!page) {
2740 spin_unlock_irqrestore(&zone->lock, flags);
2741 return NULL;
2742 }
2743 }
2744 __mod_zone_freepage_state(zone, -(1 << order),
2745 get_pcppage_migratetype(page));
2746 spin_unlock_irqrestore(&zone->lock, flags);
2747 } while (check_new_pages(page, order));
2748
2749 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2750 zone_statistics(preferred_zone, zone, 1);
2751
2752 return page;
2753}
2754
2755static int nr_pcp_alloc(struct per_cpu_pages *pcp, struct zone *zone, int order)
2756{
2757 int high, base_batch, batch, max_nr_alloc;
2758 int high_max, high_min;
2759
2760 base_batch = READ_ONCE(pcp->batch);
2761 high_min = READ_ONCE(pcp->high_min);
2762 high_max = READ_ONCE(pcp->high_max);
2763 high = pcp->high = clamp(pcp->high, high_min, high_max);
2764
2765 /* Check for PCP disabled or boot pageset */
2766 if (unlikely(high < base_batch))
2767 return 1;
2768
2769 if (order)
2770 batch = base_batch;
2771 else
2772 batch = (base_batch << pcp->alloc_factor);
2773
2774 /*
2775 * If we had larger pcp->high, we could avoid to allocate from
2776 * zone.
2777 */
2778 if (high_min != high_max && !test_bit(ZONE_BELOW_HIGH, &zone->flags))
2779 high = pcp->high = min(high + batch, high_max);
2780
2781 if (!order) {
2782 max_nr_alloc = max(high - pcp->count - base_batch, base_batch);
2783 /*
2784 * Double the number of pages allocated each time there is
2785 * subsequent allocation of order-0 pages without any freeing.
2786 */
2787 if (batch <= max_nr_alloc &&
2788 pcp->alloc_factor < CONFIG_PCP_BATCH_SCALE_MAX)
2789 pcp->alloc_factor++;
2790 batch = min(batch, max_nr_alloc);
2791 }
2792
2793 /*
2794 * Scale batch relative to order if batch implies free pages
2795 * can be stored on the PCP. Batch can be 1 for small zones or
2796 * for boot pagesets which should never store free pages as
2797 * the pages may belong to arbitrary zones.
2798 */
2799 if (batch > 1)
2800 batch = max(batch >> order, 2);
2801
2802 return batch;
2803}
2804
2805/* Remove page from the per-cpu list, caller must protect the list */
2806static inline
2807struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
2808 int migratetype,
2809 unsigned int alloc_flags,
2810 struct per_cpu_pages *pcp,
2811 struct list_head *list)
2812{
2813 struct page *page;
2814
2815 do {
2816 if (list_empty(list)) {
2817 int batch = nr_pcp_alloc(pcp, zone, order);
2818 int alloced;
2819
2820 alloced = rmqueue_bulk(zone, order,
2821 batch, list,
2822 migratetype, alloc_flags);
2823
2824 pcp->count += alloced << order;
2825 if (unlikely(list_empty(list)))
2826 return NULL;
2827 }
2828
2829 page = list_first_entry(list, struct page, pcp_list);
2830 list_del(&page->pcp_list);
2831 pcp->count -= 1 << order;
2832 } while (check_new_pages(page, order));
2833
2834 return page;
2835}
2836
2837/* Lock and remove page from the per-cpu list */
2838static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2839 struct zone *zone, unsigned int order,
2840 int migratetype, unsigned int alloc_flags)
2841{
2842 struct per_cpu_pages *pcp;
2843 struct list_head *list;
2844 struct page *page;
2845 unsigned long __maybe_unused UP_flags;
2846
2847 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
2848 pcp_trylock_prepare(UP_flags);
2849 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2850 if (!pcp) {
2851 pcp_trylock_finish(UP_flags);
2852 return NULL;
2853 }
2854
2855 /*
2856 * On allocation, reduce the number of pages that are batch freed.
2857 * See nr_pcp_free() where free_factor is increased for subsequent
2858 * frees.
2859 */
2860 pcp->free_count >>= 1;
2861 list = &pcp->lists[order_to_pindex(migratetype, order)];
2862 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
2863 pcp_spin_unlock(pcp);
2864 pcp_trylock_finish(UP_flags);
2865 if (page) {
2866 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2867 zone_statistics(preferred_zone, zone, 1);
2868 }
2869 return page;
2870}
2871
2872/*
2873 * Allocate a page from the given zone.
2874 * Use pcplists for THP or "cheap" high-order allocations.
2875 */
2876
2877/*
2878 * Do not instrument rmqueue() with KMSAN. This function may call
2879 * __msan_poison_alloca() through a call to set_pfnblock_flags_mask().
2880 * If __msan_poison_alloca() attempts to allocate pages for the stack depot, it
2881 * may call rmqueue() again, which will result in a deadlock.
2882 */
2883__no_sanitize_memory
2884static inline
2885struct page *rmqueue(struct zone *preferred_zone,
2886 struct zone *zone, unsigned int order,
2887 gfp_t gfp_flags, unsigned int alloc_flags,
2888 int migratetype)
2889{
2890 struct page *page;
2891
2892 /*
2893 * We most definitely don't want callers attempting to
2894 * allocate greater than order-1 page units with __GFP_NOFAIL.
2895 */
2896 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2897
2898 if (likely(pcp_allowed_order(order))) {
2899 page = rmqueue_pcplist(preferred_zone, zone, order,
2900 migratetype, alloc_flags);
2901 if (likely(page))
2902 goto out;
2903 }
2904
2905 page = rmqueue_buddy(preferred_zone, zone, order, alloc_flags,
2906 migratetype);
2907
2908out:
2909 /* Separate test+clear to avoid unnecessary atomics */
2910 if ((alloc_flags & ALLOC_KSWAPD) &&
2911 unlikely(test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags))) {
2912 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2913 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
2914 }
2915
2916 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
2917 return page;
2918}
2919
2920noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2921{
2922 return __should_fail_alloc_page(gfp_mask, order);
2923}
2924ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
2925
2926static inline long __zone_watermark_unusable_free(struct zone *z,
2927 unsigned int order, unsigned int alloc_flags)
2928{
2929 long unusable_free = (1 << order) - 1;
2930
2931 /*
2932 * If the caller does not have rights to reserves below the min
2933 * watermark then subtract the high-atomic reserves. This will
2934 * over-estimate the size of the atomic reserve but it avoids a search.
2935 */
2936 if (likely(!(alloc_flags & ALLOC_RESERVES)))
2937 unusable_free += z->nr_reserved_highatomic;
2938
2939#ifdef CONFIG_CMA
2940 /* If allocation can't use CMA areas don't use free CMA pages */
2941 if (!(alloc_flags & ALLOC_CMA))
2942 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
2943#endif
2944#ifdef CONFIG_UNACCEPTED_MEMORY
2945 unusable_free += zone_page_state(z, NR_UNACCEPTED);
2946#endif
2947
2948 return unusable_free;
2949}
2950
2951/*
2952 * Return true if free base pages are above 'mark'. For high-order checks it
2953 * will return true of the order-0 watermark is reached and there is at least
2954 * one free page of a suitable size. Checking now avoids taking the zone lock
2955 * to check in the allocation paths if no pages are free.
2956 */
2957bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2958 int highest_zoneidx, unsigned int alloc_flags,
2959 long free_pages)
2960{
2961 long min = mark;
2962 int o;
2963
2964 /* free_pages may go negative - that's OK */
2965 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
2966
2967 if (unlikely(alloc_flags & ALLOC_RESERVES)) {
2968 /*
2969 * __GFP_HIGH allows access to 50% of the min reserve as well
2970 * as OOM.
2971 */
2972 if (alloc_flags & ALLOC_MIN_RESERVE) {
2973 min -= min / 2;
2974
2975 /*
2976 * Non-blocking allocations (e.g. GFP_ATOMIC) can
2977 * access more reserves than just __GFP_HIGH. Other
2978 * non-blocking allocations requests such as GFP_NOWAIT
2979 * or (GFP_KERNEL & ~__GFP_DIRECT_RECLAIM) do not get
2980 * access to the min reserve.
2981 */
2982 if (alloc_flags & ALLOC_NON_BLOCK)
2983 min -= min / 4;
2984 }
2985
2986 /*
2987 * OOM victims can try even harder than the normal reserve
2988 * users on the grounds that it's definitely going to be in
2989 * the exit path shortly and free memory. Any allocation it
2990 * makes during the free path will be small and short-lived.
2991 */
2992 if (alloc_flags & ALLOC_OOM)
2993 min -= min / 2;
2994 }
2995
2996 /*
2997 * Check watermarks for an order-0 allocation request. If these
2998 * are not met, then a high-order request also cannot go ahead
2999 * even if a suitable page happened to be free.
3000 */
3001 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3002 return false;
3003
3004 /* If this is an order-0 request then the watermark is fine */
3005 if (!order)
3006 return true;
3007
3008 /* For a high-order request, check at least one suitable page is free */
3009 for (o = order; o < NR_PAGE_ORDERS; o++) {
3010 struct free_area *area = &z->free_area[o];
3011 int mt;
3012
3013 if (!area->nr_free)
3014 continue;
3015
3016 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3017 if (!free_area_empty(area, mt))
3018 return true;
3019 }
3020
3021#ifdef CONFIG_CMA
3022 if ((alloc_flags & ALLOC_CMA) &&
3023 !free_area_empty(area, MIGRATE_CMA)) {
3024 return true;
3025 }
3026#endif
3027 if ((alloc_flags & (ALLOC_HIGHATOMIC|ALLOC_OOM)) &&
3028 !free_area_empty(area, MIGRATE_HIGHATOMIC)) {
3029 return true;
3030 }
3031 }
3032 return false;
3033}
3034
3035bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3036 int highest_zoneidx, unsigned int alloc_flags)
3037{
3038 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3039 zone_page_state(z, NR_FREE_PAGES));
3040}
3041
3042static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3043 unsigned long mark, int highest_zoneidx,
3044 unsigned int alloc_flags, gfp_t gfp_mask)
3045{
3046 long free_pages;
3047
3048 free_pages = zone_page_state(z, NR_FREE_PAGES);
3049
3050 /*
3051 * Fast check for order-0 only. If this fails then the reserves
3052 * need to be calculated.
3053 */
3054 if (!order) {
3055 long usable_free;
3056 long reserved;
3057
3058 usable_free = free_pages;
3059 reserved = __zone_watermark_unusable_free(z, 0, alloc_flags);
3060
3061 /* reserved may over estimate high-atomic reserves. */
3062 usable_free -= min(usable_free, reserved);
3063 if (usable_free > mark + z->lowmem_reserve[highest_zoneidx])
3064 return true;
3065 }
3066
3067 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3068 free_pages))
3069 return true;
3070
3071 /*
3072 * Ignore watermark boosting for __GFP_HIGH order-0 allocations
3073 * when checking the min watermark. The min watermark is the
3074 * point where boosting is ignored so that kswapd is woken up
3075 * when below the low watermark.
3076 */
3077 if (unlikely(!order && (alloc_flags & ALLOC_MIN_RESERVE) && z->watermark_boost
3078 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3079 mark = z->_watermark[WMARK_MIN];
3080 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3081 alloc_flags, free_pages);
3082 }
3083
3084 return false;
3085}
3086
3087bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3088 unsigned long mark, int highest_zoneidx)
3089{
3090 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3091
3092 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3093 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3094
3095 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3096 free_pages);
3097}
3098
3099#ifdef CONFIG_NUMA
3100int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
3101
3102static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3103{
3104 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3105 node_reclaim_distance;
3106}
3107#else /* CONFIG_NUMA */
3108static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3109{
3110 return true;
3111}
3112#endif /* CONFIG_NUMA */
3113
3114/*
3115 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3116 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3117 * premature use of a lower zone may cause lowmem pressure problems that
3118 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3119 * probably too small. It only makes sense to spread allocations to avoid
3120 * fragmentation between the Normal and DMA32 zones.
3121 */
3122static inline unsigned int
3123alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3124{
3125 unsigned int alloc_flags;
3126
3127 /*
3128 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3129 * to save a branch.
3130 */
3131 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3132
3133#ifdef CONFIG_ZONE_DMA32
3134 if (!zone)
3135 return alloc_flags;
3136
3137 if (zone_idx(zone) != ZONE_NORMAL)
3138 return alloc_flags;
3139
3140 /*
3141 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3142 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3143 * on UMA that if Normal is populated then so is DMA32.
3144 */
3145 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3146 if (nr_online_nodes > 1 && !populated_zone(--zone))
3147 return alloc_flags;
3148
3149 alloc_flags |= ALLOC_NOFRAGMENT;
3150#endif /* CONFIG_ZONE_DMA32 */
3151 return alloc_flags;
3152}
3153
3154/* Must be called after current_gfp_context() which can change gfp_mask */
3155static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
3156 unsigned int alloc_flags)
3157{
3158#ifdef CONFIG_CMA
3159 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3160 alloc_flags |= ALLOC_CMA;
3161#endif
3162 return alloc_flags;
3163}
3164
3165/*
3166 * get_page_from_freelist goes through the zonelist trying to allocate
3167 * a page.
3168 */
3169static struct page *
3170get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3171 const struct alloc_context *ac)
3172{
3173 struct zoneref *z;
3174 struct zone *zone;
3175 struct pglist_data *last_pgdat = NULL;
3176 bool last_pgdat_dirty_ok = false;
3177 bool no_fallback;
3178
3179retry:
3180 /*
3181 * Scan zonelist, looking for a zone with enough free.
3182 * See also cpuset_node_allowed() comment in kernel/cgroup/cpuset.c.
3183 */
3184 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3185 z = ac->preferred_zoneref;
3186 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3187 ac->nodemask) {
3188 struct page *page;
3189 unsigned long mark;
3190
3191 if (cpusets_enabled() &&
3192 (alloc_flags & ALLOC_CPUSET) &&
3193 !__cpuset_zone_allowed(zone, gfp_mask))
3194 continue;
3195 /*
3196 * When allocating a page cache page for writing, we
3197 * want to get it from a node that is within its dirty
3198 * limit, such that no single node holds more than its
3199 * proportional share of globally allowed dirty pages.
3200 * The dirty limits take into account the node's
3201 * lowmem reserves and high watermark so that kswapd
3202 * should be able to balance it without having to
3203 * write pages from its LRU list.
3204 *
3205 * XXX: For now, allow allocations to potentially
3206 * exceed the per-node dirty limit in the slowpath
3207 * (spread_dirty_pages unset) before going into reclaim,
3208 * which is important when on a NUMA setup the allowed
3209 * nodes are together not big enough to reach the
3210 * global limit. The proper fix for these situations
3211 * will require awareness of nodes in the
3212 * dirty-throttling and the flusher threads.
3213 */
3214 if (ac->spread_dirty_pages) {
3215 if (last_pgdat != zone->zone_pgdat) {
3216 last_pgdat = zone->zone_pgdat;
3217 last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat);
3218 }
3219
3220 if (!last_pgdat_dirty_ok)
3221 continue;
3222 }
3223
3224 if (no_fallback && nr_online_nodes > 1 &&
3225 zone != ac->preferred_zoneref->zone) {
3226 int local_nid;
3227
3228 /*
3229 * If moving to a remote node, retry but allow
3230 * fragmenting fallbacks. Locality is more important
3231 * than fragmentation avoidance.
3232 */
3233 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3234 if (zone_to_nid(zone) != local_nid) {
3235 alloc_flags &= ~ALLOC_NOFRAGMENT;
3236 goto retry;
3237 }
3238 }
3239
3240 /*
3241 * Detect whether the number of free pages is below high
3242 * watermark. If so, we will decrease pcp->high and free
3243 * PCP pages in free path to reduce the possibility of
3244 * premature page reclaiming. Detection is done here to
3245 * avoid to do that in hotter free path.
3246 */
3247 if (test_bit(ZONE_BELOW_HIGH, &zone->flags))
3248 goto check_alloc_wmark;
3249
3250 mark = high_wmark_pages(zone);
3251 if (zone_watermark_fast(zone, order, mark,
3252 ac->highest_zoneidx, alloc_flags,
3253 gfp_mask))
3254 goto try_this_zone;
3255 else
3256 set_bit(ZONE_BELOW_HIGH, &zone->flags);
3257
3258check_alloc_wmark:
3259 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3260 if (!zone_watermark_fast(zone, order, mark,
3261 ac->highest_zoneidx, alloc_flags,
3262 gfp_mask)) {
3263 int ret;
3264
3265 if (has_unaccepted_memory()) {
3266 if (try_to_accept_memory(zone, order))
3267 goto try_this_zone;
3268 }
3269
3270#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3271 /*
3272 * Watermark failed for this zone, but see if we can
3273 * grow this zone if it contains deferred pages.
3274 */
3275 if (deferred_pages_enabled()) {
3276 if (_deferred_grow_zone(zone, order))
3277 goto try_this_zone;
3278 }
3279#endif
3280 /* Checked here to keep the fast path fast */
3281 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3282 if (alloc_flags & ALLOC_NO_WATERMARKS)
3283 goto try_this_zone;
3284
3285 if (!node_reclaim_enabled() ||
3286 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3287 continue;
3288
3289 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3290 switch (ret) {
3291 case NODE_RECLAIM_NOSCAN:
3292 /* did not scan */
3293 continue;
3294 case NODE_RECLAIM_FULL:
3295 /* scanned but unreclaimable */
3296 continue;
3297 default:
3298 /* did we reclaim enough */
3299 if (zone_watermark_ok(zone, order, mark,
3300 ac->highest_zoneidx, alloc_flags))
3301 goto try_this_zone;
3302
3303 continue;
3304 }
3305 }
3306
3307try_this_zone:
3308 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3309 gfp_mask, alloc_flags, ac->migratetype);
3310 if (page) {
3311 prep_new_page(page, order, gfp_mask, alloc_flags);
3312
3313 /*
3314 * If this is a high-order atomic allocation then check
3315 * if the pageblock should be reserved for the future
3316 */
3317 if (unlikely(alloc_flags & ALLOC_HIGHATOMIC))
3318 reserve_highatomic_pageblock(page, zone);
3319
3320 return page;
3321 } else {
3322 if (has_unaccepted_memory()) {
3323 if (try_to_accept_memory(zone, order))
3324 goto try_this_zone;
3325 }
3326
3327#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3328 /* Try again if zone has deferred pages */
3329 if (deferred_pages_enabled()) {
3330 if (_deferred_grow_zone(zone, order))
3331 goto try_this_zone;
3332 }
3333#endif
3334 }
3335 }
3336
3337 /*
3338 * It's possible on a UMA machine to get through all zones that are
3339 * fragmented. If avoiding fragmentation, reset and try again.
3340 */
3341 if (no_fallback) {
3342 alloc_flags &= ~ALLOC_NOFRAGMENT;
3343 goto retry;
3344 }
3345
3346 return NULL;
3347}
3348
3349static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3350{
3351 unsigned int filter = SHOW_MEM_FILTER_NODES;
3352
3353 /*
3354 * This documents exceptions given to allocations in certain
3355 * contexts that are allowed to allocate outside current's set
3356 * of allowed nodes.
3357 */
3358 if (!(gfp_mask & __GFP_NOMEMALLOC))
3359 if (tsk_is_oom_victim(current) ||
3360 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3361 filter &= ~SHOW_MEM_FILTER_NODES;
3362 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3363 filter &= ~SHOW_MEM_FILTER_NODES;
3364
3365 __show_mem(filter, nodemask, gfp_zone(gfp_mask));
3366}
3367
3368void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3369{
3370 struct va_format vaf;
3371 va_list args;
3372 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3373
3374 if ((gfp_mask & __GFP_NOWARN) ||
3375 !__ratelimit(&nopage_rs) ||
3376 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
3377 return;
3378
3379 va_start(args, fmt);
3380 vaf.fmt = fmt;
3381 vaf.va = &args;
3382 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3383 current->comm, &vaf, gfp_mask, &gfp_mask,
3384 nodemask_pr_args(nodemask));
3385 va_end(args);
3386
3387 cpuset_print_current_mems_allowed();
3388 pr_cont("\n");
3389 dump_stack();
3390 warn_alloc_show_mem(gfp_mask, nodemask);
3391}
3392
3393static inline struct page *
3394__alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3395 unsigned int alloc_flags,
3396 const struct alloc_context *ac)
3397{
3398 struct page *page;
3399
3400 page = get_page_from_freelist(gfp_mask, order,
3401 alloc_flags|ALLOC_CPUSET, ac);
3402 /*
3403 * fallback to ignore cpuset restriction if our nodes
3404 * are depleted
3405 */
3406 if (!page)
3407 page = get_page_from_freelist(gfp_mask, order,
3408 alloc_flags, ac);
3409
3410 return page;
3411}
3412
3413static inline struct page *
3414__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3415 const struct alloc_context *ac, unsigned long *did_some_progress)
3416{
3417 struct oom_control oc = {
3418 .zonelist = ac->zonelist,
3419 .nodemask = ac->nodemask,
3420 .memcg = NULL,
3421 .gfp_mask = gfp_mask,
3422 .order = order,
3423 };
3424 struct page *page;
3425
3426 *did_some_progress = 0;
3427
3428 /*
3429 * Acquire the oom lock. If that fails, somebody else is
3430 * making progress for us.
3431 */
3432 if (!mutex_trylock(&oom_lock)) {
3433 *did_some_progress = 1;
3434 schedule_timeout_uninterruptible(1);
3435 return NULL;
3436 }
3437
3438 /*
3439 * Go through the zonelist yet one more time, keep very high watermark
3440 * here, this is only to catch a parallel oom killing, we must fail if
3441 * we're still under heavy pressure. But make sure that this reclaim
3442 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3443 * allocation which will never fail due to oom_lock already held.
3444 */
3445 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3446 ~__GFP_DIRECT_RECLAIM, order,
3447 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3448 if (page)
3449 goto out;
3450
3451 /* Coredumps can quickly deplete all memory reserves */
3452 if (current->flags & PF_DUMPCORE)
3453 goto out;
3454 /* The OOM killer will not help higher order allocs */
3455 if (order > PAGE_ALLOC_COSTLY_ORDER)
3456 goto out;
3457 /*
3458 * We have already exhausted all our reclaim opportunities without any
3459 * success so it is time to admit defeat. We will skip the OOM killer
3460 * because it is very likely that the caller has a more reasonable
3461 * fallback than shooting a random task.
3462 *
3463 * The OOM killer may not free memory on a specific node.
3464 */
3465 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
3466 goto out;
3467 /* The OOM killer does not needlessly kill tasks for lowmem */
3468 if (ac->highest_zoneidx < ZONE_NORMAL)
3469 goto out;
3470 if (pm_suspended_storage())
3471 goto out;
3472 /*
3473 * XXX: GFP_NOFS allocations should rather fail than rely on
3474 * other request to make a forward progress.
3475 * We are in an unfortunate situation where out_of_memory cannot
3476 * do much for this context but let's try it to at least get
3477 * access to memory reserved if the current task is killed (see
3478 * out_of_memory). Once filesystems are ready to handle allocation
3479 * failures more gracefully we should just bail out here.
3480 */
3481
3482 /* Exhausted what can be done so it's blame time */
3483 if (out_of_memory(&oc) ||
3484 WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) {
3485 *did_some_progress = 1;
3486
3487 /*
3488 * Help non-failing allocations by giving them access to memory
3489 * reserves
3490 */
3491 if (gfp_mask & __GFP_NOFAIL)
3492 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3493 ALLOC_NO_WATERMARKS, ac);
3494 }
3495out:
3496 mutex_unlock(&oom_lock);
3497 return page;
3498}
3499
3500/*
3501 * Maximum number of compaction retries with a progress before OOM
3502 * killer is consider as the only way to move forward.
3503 */
3504#define MAX_COMPACT_RETRIES 16
3505
3506#ifdef CONFIG_COMPACTION
3507/* Try memory compaction for high-order allocations before reclaim */
3508static struct page *
3509__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3510 unsigned int alloc_flags, const struct alloc_context *ac,
3511 enum compact_priority prio, enum compact_result *compact_result)
3512{
3513 struct page *page = NULL;
3514 unsigned long pflags;
3515 unsigned int noreclaim_flag;
3516
3517 if (!order)
3518 return NULL;
3519
3520 psi_memstall_enter(&pflags);
3521 delayacct_compact_start();
3522 noreclaim_flag = memalloc_noreclaim_save();
3523
3524 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3525 prio, &page);
3526
3527 memalloc_noreclaim_restore(noreclaim_flag);
3528 psi_memstall_leave(&pflags);
3529 delayacct_compact_end();
3530
3531 if (*compact_result == COMPACT_SKIPPED)
3532 return NULL;
3533 /*
3534 * At least in one zone compaction wasn't deferred or skipped, so let's
3535 * count a compaction stall
3536 */
3537 count_vm_event(COMPACTSTALL);
3538
3539 /* Prep a captured page if available */
3540 if (page)
3541 prep_new_page(page, order, gfp_mask, alloc_flags);
3542
3543 /* Try get a page from the freelist if available */
3544 if (!page)
3545 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3546
3547 if (page) {
3548 struct zone *zone = page_zone(page);
3549
3550 zone->compact_blockskip_flush = false;
3551 compaction_defer_reset(zone, order, true);
3552 count_vm_event(COMPACTSUCCESS);
3553 return page;
3554 }
3555
3556 /*
3557 * It's bad if compaction run occurs and fails. The most likely reason
3558 * is that pages exist, but not enough to satisfy watermarks.
3559 */
3560 count_vm_event(COMPACTFAIL);
3561
3562 cond_resched();
3563
3564 return NULL;
3565}
3566
3567static inline bool
3568should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3569 enum compact_result compact_result,
3570 enum compact_priority *compact_priority,
3571 int *compaction_retries)
3572{
3573 int max_retries = MAX_COMPACT_RETRIES;
3574 int min_priority;
3575 bool ret = false;
3576 int retries = *compaction_retries;
3577 enum compact_priority priority = *compact_priority;
3578
3579 if (!order)
3580 return false;
3581
3582 if (fatal_signal_pending(current))
3583 return false;
3584
3585 /*
3586 * Compaction was skipped due to a lack of free order-0
3587 * migration targets. Continue if reclaim can help.
3588 */
3589 if (compact_result == COMPACT_SKIPPED) {
3590 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3591 goto out;
3592 }
3593
3594 /*
3595 * Compaction managed to coalesce some page blocks, but the
3596 * allocation failed presumably due to a race. Retry some.
3597 */
3598 if (compact_result == COMPACT_SUCCESS) {
3599 /*
3600 * !costly requests are much more important than
3601 * __GFP_RETRY_MAYFAIL costly ones because they are de
3602 * facto nofail and invoke OOM killer to move on while
3603 * costly can fail and users are ready to cope with
3604 * that. 1/4 retries is rather arbitrary but we would
3605 * need much more detailed feedback from compaction to
3606 * make a better decision.
3607 */
3608 if (order > PAGE_ALLOC_COSTLY_ORDER)
3609 max_retries /= 4;
3610
3611 if (++(*compaction_retries) <= max_retries) {
3612 ret = true;
3613 goto out;
3614 }
3615 }
3616
3617 /*
3618 * Compaction failed. Retry with increasing priority.
3619 */
3620 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3621 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3622
3623 if (*compact_priority > min_priority) {
3624 (*compact_priority)--;
3625 *compaction_retries = 0;
3626 ret = true;
3627 }
3628out:
3629 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3630 return ret;
3631}
3632#else
3633static inline struct page *
3634__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3635 unsigned int alloc_flags, const struct alloc_context *ac,
3636 enum compact_priority prio, enum compact_result *compact_result)
3637{
3638 *compact_result = COMPACT_SKIPPED;
3639 return NULL;
3640}
3641
3642static inline bool
3643should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3644 enum compact_result compact_result,
3645 enum compact_priority *compact_priority,
3646 int *compaction_retries)
3647{
3648 struct zone *zone;
3649 struct zoneref *z;
3650
3651 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3652 return false;
3653
3654 /*
3655 * There are setups with compaction disabled which would prefer to loop
3656 * inside the allocator rather than hit the oom killer prematurely.
3657 * Let's give them a good hope and keep retrying while the order-0
3658 * watermarks are OK.
3659 */
3660 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3661 ac->highest_zoneidx, ac->nodemask) {
3662 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3663 ac->highest_zoneidx, alloc_flags))
3664 return true;
3665 }
3666 return false;
3667}
3668#endif /* CONFIG_COMPACTION */
3669
3670#ifdef CONFIG_LOCKDEP
3671static struct lockdep_map __fs_reclaim_map =
3672 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3673
3674static bool __need_reclaim(gfp_t gfp_mask)
3675{
3676 /* no reclaim without waiting on it */
3677 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3678 return false;
3679
3680 /* this guy won't enter reclaim */
3681 if (current->flags & PF_MEMALLOC)
3682 return false;
3683
3684 if (gfp_mask & __GFP_NOLOCKDEP)
3685 return false;
3686
3687 return true;
3688}
3689
3690void __fs_reclaim_acquire(unsigned long ip)
3691{
3692 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
3693}
3694
3695void __fs_reclaim_release(unsigned long ip)
3696{
3697 lock_release(&__fs_reclaim_map, ip);
3698}
3699
3700void fs_reclaim_acquire(gfp_t gfp_mask)
3701{
3702 gfp_mask = current_gfp_context(gfp_mask);
3703
3704 if (__need_reclaim(gfp_mask)) {
3705 if (gfp_mask & __GFP_FS)
3706 __fs_reclaim_acquire(_RET_IP_);
3707
3708#ifdef CONFIG_MMU_NOTIFIER
3709 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
3710 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
3711#endif
3712
3713 }
3714}
3715EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3716
3717void fs_reclaim_release(gfp_t gfp_mask)
3718{
3719 gfp_mask = current_gfp_context(gfp_mask);
3720
3721 if (__need_reclaim(gfp_mask)) {
3722 if (gfp_mask & __GFP_FS)
3723 __fs_reclaim_release(_RET_IP_);
3724 }
3725}
3726EXPORT_SYMBOL_GPL(fs_reclaim_release);
3727#endif
3728
3729/*
3730 * Zonelists may change due to hotplug during allocation. Detect when zonelists
3731 * have been rebuilt so allocation retries. Reader side does not lock and
3732 * retries the allocation if zonelist changes. Writer side is protected by the
3733 * embedded spin_lock.
3734 */
3735static DEFINE_SEQLOCK(zonelist_update_seq);
3736
3737static unsigned int zonelist_iter_begin(void)
3738{
3739 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
3740 return read_seqbegin(&zonelist_update_seq);
3741
3742 return 0;
3743}
3744
3745static unsigned int check_retry_zonelist(unsigned int seq)
3746{
3747 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
3748 return read_seqretry(&zonelist_update_seq, seq);
3749
3750 return seq;
3751}
3752
3753/* Perform direct synchronous page reclaim */
3754static unsigned long
3755__perform_reclaim(gfp_t gfp_mask, unsigned int order,
3756 const struct alloc_context *ac)
3757{
3758 unsigned int noreclaim_flag;
3759 unsigned long progress;
3760
3761 cond_resched();
3762
3763 /* We now go into synchronous reclaim */
3764 cpuset_memory_pressure_bump();
3765 fs_reclaim_acquire(gfp_mask);
3766 noreclaim_flag = memalloc_noreclaim_save();
3767
3768 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3769 ac->nodemask);
3770
3771 memalloc_noreclaim_restore(noreclaim_flag);
3772 fs_reclaim_release(gfp_mask);
3773
3774 cond_resched();
3775
3776 return progress;
3777}
3778
3779/* The really slow allocator path where we enter direct reclaim */
3780static inline struct page *
3781__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3782 unsigned int alloc_flags, const struct alloc_context *ac,
3783 unsigned long *did_some_progress)
3784{
3785 struct page *page = NULL;
3786 unsigned long pflags;
3787 bool drained = false;
3788
3789 psi_memstall_enter(&pflags);
3790 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3791 if (unlikely(!(*did_some_progress)))
3792 goto out;
3793
3794retry:
3795 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3796
3797 /*
3798 * If an allocation failed after direct reclaim, it could be because
3799 * pages are pinned on the per-cpu lists or in high alloc reserves.
3800 * Shrink them and try again
3801 */
3802 if (!page && !drained) {
3803 unreserve_highatomic_pageblock(ac, false);
3804 drain_all_pages(NULL);
3805 drained = true;
3806 goto retry;
3807 }
3808out:
3809 psi_memstall_leave(&pflags);
3810
3811 return page;
3812}
3813
3814static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
3815 const struct alloc_context *ac)
3816{
3817 struct zoneref *z;
3818 struct zone *zone;
3819 pg_data_t *last_pgdat = NULL;
3820 enum zone_type highest_zoneidx = ac->highest_zoneidx;
3821
3822 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
3823 ac->nodemask) {
3824 if (!managed_zone(zone))
3825 continue;
3826 if (last_pgdat != zone->zone_pgdat) {
3827 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
3828 last_pgdat = zone->zone_pgdat;
3829 }
3830 }
3831}
3832
3833static inline unsigned int
3834gfp_to_alloc_flags(gfp_t gfp_mask, unsigned int order)
3835{
3836 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3837
3838 /*
3839 * __GFP_HIGH is assumed to be the same as ALLOC_MIN_RESERVE
3840 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3841 * to save two branches.
3842 */
3843 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_MIN_RESERVE);
3844 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
3845
3846 /*
3847 * The caller may dip into page reserves a bit more if the caller
3848 * cannot run direct reclaim, or if the caller has realtime scheduling
3849 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3850 * set both ALLOC_NON_BLOCK and ALLOC_MIN_RESERVE(__GFP_HIGH).
3851 */
3852 alloc_flags |= (__force int)
3853 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
3854
3855 if (!(gfp_mask & __GFP_DIRECT_RECLAIM)) {
3856 /*
3857 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3858 * if it can't schedule.
3859 */
3860 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
3861 alloc_flags |= ALLOC_NON_BLOCK;
3862
3863 if (order > 0)
3864 alloc_flags |= ALLOC_HIGHATOMIC;
3865 }
3866
3867 /*
3868 * Ignore cpuset mems for non-blocking __GFP_HIGH (probably
3869 * GFP_ATOMIC) rather than fail, see the comment for
3870 * cpuset_node_allowed().
3871 */
3872 if (alloc_flags & ALLOC_MIN_RESERVE)
3873 alloc_flags &= ~ALLOC_CPUSET;
3874 } else if (unlikely(rt_task(current)) && in_task())
3875 alloc_flags |= ALLOC_MIN_RESERVE;
3876
3877 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
3878
3879 return alloc_flags;
3880}
3881
3882static bool oom_reserves_allowed(struct task_struct *tsk)
3883{
3884 if (!tsk_is_oom_victim(tsk))
3885 return false;
3886
3887 /*
3888 * !MMU doesn't have oom reaper so give access to memory reserves
3889 * only to the thread with TIF_MEMDIE set
3890 */
3891 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3892 return false;
3893
3894 return true;
3895}
3896
3897/*
3898 * Distinguish requests which really need access to full memory
3899 * reserves from oom victims which can live with a portion of it
3900 */
3901static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
3902{
3903 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3904 return 0;
3905 if (gfp_mask & __GFP_MEMALLOC)
3906 return ALLOC_NO_WATERMARKS;
3907 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3908 return ALLOC_NO_WATERMARKS;
3909 if (!in_interrupt()) {
3910 if (current->flags & PF_MEMALLOC)
3911 return ALLOC_NO_WATERMARKS;
3912 else if (oom_reserves_allowed(current))
3913 return ALLOC_OOM;
3914 }
3915
3916 return 0;
3917}
3918
3919bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3920{
3921 return !!__gfp_pfmemalloc_flags(gfp_mask);
3922}
3923
3924/*
3925 * Checks whether it makes sense to retry the reclaim to make a forward progress
3926 * for the given allocation request.
3927 *
3928 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3929 * without success, or when we couldn't even meet the watermark if we
3930 * reclaimed all remaining pages on the LRU lists.
3931 *
3932 * Returns true if a retry is viable or false to enter the oom path.
3933 */
3934static inline bool
3935should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3936 struct alloc_context *ac, int alloc_flags,
3937 bool did_some_progress, int *no_progress_loops)
3938{
3939 struct zone *zone;
3940 struct zoneref *z;
3941 bool ret = false;
3942
3943 /*
3944 * Costly allocations might have made a progress but this doesn't mean
3945 * their order will become available due to high fragmentation so
3946 * always increment the no progress counter for them
3947 */
3948 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3949 *no_progress_loops = 0;
3950 else
3951 (*no_progress_loops)++;
3952
3953 if (*no_progress_loops > MAX_RECLAIM_RETRIES)
3954 goto out;
3955
3956
3957 /*
3958 * Keep reclaiming pages while there is a chance this will lead
3959 * somewhere. If none of the target zones can satisfy our allocation
3960 * request even if all reclaimable pages are considered then we are
3961 * screwed and have to go OOM.
3962 */
3963 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3964 ac->highest_zoneidx, ac->nodemask) {
3965 unsigned long available;
3966 unsigned long reclaimable;
3967 unsigned long min_wmark = min_wmark_pages(zone);
3968 bool wmark;
3969
3970 available = reclaimable = zone_reclaimable_pages(zone);
3971 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3972
3973 /*
3974 * Would the allocation succeed if we reclaimed all
3975 * reclaimable pages?
3976 */
3977 wmark = __zone_watermark_ok(zone, order, min_wmark,
3978 ac->highest_zoneidx, alloc_flags, available);
3979 trace_reclaim_retry_zone(z, order, reclaimable,
3980 available, min_wmark, *no_progress_loops, wmark);
3981 if (wmark) {
3982 ret = true;
3983 break;
3984 }
3985 }
3986
3987 /*
3988 * Memory allocation/reclaim might be called from a WQ context and the
3989 * current implementation of the WQ concurrency control doesn't
3990 * recognize that a particular WQ is congested if the worker thread is
3991 * looping without ever sleeping. Therefore we have to do a short sleep
3992 * here rather than calling cond_resched().
3993 */
3994 if (current->flags & PF_WQ_WORKER)
3995 schedule_timeout_uninterruptible(1);
3996 else
3997 cond_resched();
3998out:
3999 /* Before OOM, exhaust highatomic_reserve */
4000 if (!ret)
4001 return unreserve_highatomic_pageblock(ac, true);
4002
4003 return ret;
4004}
4005
4006static inline bool
4007check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4008{
4009 /*
4010 * It's possible that cpuset's mems_allowed and the nodemask from
4011 * mempolicy don't intersect. This should be normally dealt with by
4012 * policy_nodemask(), but it's possible to race with cpuset update in
4013 * such a way the check therein was true, and then it became false
4014 * before we got our cpuset_mems_cookie here.
4015 * This assumes that for all allocations, ac->nodemask can come only
4016 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4017 * when it does not intersect with the cpuset restrictions) or the
4018 * caller can deal with a violated nodemask.
4019 */
4020 if (cpusets_enabled() && ac->nodemask &&
4021 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4022 ac->nodemask = NULL;
4023 return true;
4024 }
4025
4026 /*
4027 * When updating a task's mems_allowed or mempolicy nodemask, it is
4028 * possible to race with parallel threads in such a way that our
4029 * allocation can fail while the mask is being updated. If we are about
4030 * to fail, check if the cpuset changed during allocation and if so,
4031 * retry.
4032 */
4033 if (read_mems_allowed_retry(cpuset_mems_cookie))
4034 return true;
4035
4036 return false;
4037}
4038
4039static inline struct page *
4040__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4041 struct alloc_context *ac)
4042{
4043 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4044 bool can_compact = gfp_compaction_allowed(gfp_mask);
4045 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4046 struct page *page = NULL;
4047 unsigned int alloc_flags;
4048 unsigned long did_some_progress;
4049 enum compact_priority compact_priority;
4050 enum compact_result compact_result;
4051 int compaction_retries;
4052 int no_progress_loops;
4053 unsigned int cpuset_mems_cookie;
4054 unsigned int zonelist_iter_cookie;
4055 int reserve_flags;
4056
4057restart:
4058 compaction_retries = 0;
4059 no_progress_loops = 0;
4060 compact_priority = DEF_COMPACT_PRIORITY;
4061 cpuset_mems_cookie = read_mems_allowed_begin();
4062 zonelist_iter_cookie = zonelist_iter_begin();
4063
4064 /*
4065 * The fast path uses conservative alloc_flags to succeed only until
4066 * kswapd needs to be woken up, and to avoid the cost of setting up
4067 * alloc_flags precisely. So we do that now.
4068 */
4069 alloc_flags = gfp_to_alloc_flags(gfp_mask, order);
4070
4071 /*
4072 * We need to recalculate the starting point for the zonelist iterator
4073 * because we might have used different nodemask in the fast path, or
4074 * there was a cpuset modification and we are retrying - otherwise we
4075 * could end up iterating over non-eligible zones endlessly.
4076 */
4077 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4078 ac->highest_zoneidx, ac->nodemask);
4079 if (!ac->preferred_zoneref->zone)
4080 goto nopage;
4081
4082 /*
4083 * Check for insane configurations where the cpuset doesn't contain
4084 * any suitable zone to satisfy the request - e.g. non-movable
4085 * GFP_HIGHUSER allocations from MOVABLE nodes only.
4086 */
4087 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
4088 struct zoneref *z = first_zones_zonelist(ac->zonelist,
4089 ac->highest_zoneidx,
4090 &cpuset_current_mems_allowed);
4091 if (!z->zone)
4092 goto nopage;
4093 }
4094
4095 if (alloc_flags & ALLOC_KSWAPD)
4096 wake_all_kswapds(order, gfp_mask, ac);
4097
4098 /*
4099 * The adjusted alloc_flags might result in immediate success, so try
4100 * that first
4101 */
4102 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4103 if (page)
4104 goto got_pg;
4105
4106 /*
4107 * For costly allocations, try direct compaction first, as it's likely
4108 * that we have enough base pages and don't need to reclaim. For non-
4109 * movable high-order allocations, do that as well, as compaction will
4110 * try prevent permanent fragmentation by migrating from blocks of the
4111 * same migratetype.
4112 * Don't try this for allocations that are allowed to ignore
4113 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4114 */
4115 if (can_direct_reclaim && can_compact &&
4116 (costly_order ||
4117 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4118 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4119 page = __alloc_pages_direct_compact(gfp_mask, order,
4120 alloc_flags, ac,
4121 INIT_COMPACT_PRIORITY,
4122 &compact_result);
4123 if (page)
4124 goto got_pg;
4125
4126 /*
4127 * Checks for costly allocations with __GFP_NORETRY, which
4128 * includes some THP page fault allocations
4129 */
4130 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4131 /*
4132 * If allocating entire pageblock(s) and compaction
4133 * failed because all zones are below low watermarks
4134 * or is prohibited because it recently failed at this
4135 * order, fail immediately unless the allocator has
4136 * requested compaction and reclaim retry.
4137 *
4138 * Reclaim is
4139 * - potentially very expensive because zones are far
4140 * below their low watermarks or this is part of very
4141 * bursty high order allocations,
4142 * - not guaranteed to help because isolate_freepages()
4143 * may not iterate over freed pages as part of its
4144 * linear scan, and
4145 * - unlikely to make entire pageblocks free on its
4146 * own.
4147 */
4148 if (compact_result == COMPACT_SKIPPED ||
4149 compact_result == COMPACT_DEFERRED)
4150 goto nopage;
4151
4152 /*
4153 * Looks like reclaim/compaction is worth trying, but
4154 * sync compaction could be very expensive, so keep
4155 * using async compaction.
4156 */
4157 compact_priority = INIT_COMPACT_PRIORITY;
4158 }
4159 }
4160
4161retry:
4162 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4163 if (alloc_flags & ALLOC_KSWAPD)
4164 wake_all_kswapds(order, gfp_mask, ac);
4165
4166 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4167 if (reserve_flags)
4168 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) |
4169 (alloc_flags & ALLOC_KSWAPD);
4170
4171 /*
4172 * Reset the nodemask and zonelist iterators if memory policies can be
4173 * ignored. These allocations are high priority and system rather than
4174 * user oriented.
4175 */
4176 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4177 ac->nodemask = NULL;
4178 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4179 ac->highest_zoneidx, ac->nodemask);
4180 }
4181
4182 /* Attempt with potentially adjusted zonelist and alloc_flags */
4183 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4184 if (page)
4185 goto got_pg;
4186
4187 /* Caller is not willing to reclaim, we can't balance anything */
4188 if (!can_direct_reclaim)
4189 goto nopage;
4190
4191 /* Avoid recursion of direct reclaim */
4192 if (current->flags & PF_MEMALLOC)
4193 goto nopage;
4194
4195 /* Try direct reclaim and then allocating */
4196 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4197 &did_some_progress);
4198 if (page)
4199 goto got_pg;
4200
4201 /* Try direct compaction and then allocating */
4202 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4203 compact_priority, &compact_result);
4204 if (page)
4205 goto got_pg;
4206
4207 /* Do not loop if specifically requested */
4208 if (gfp_mask & __GFP_NORETRY)
4209 goto nopage;
4210
4211 /*
4212 * Do not retry costly high order allocations unless they are
4213 * __GFP_RETRY_MAYFAIL and we can compact
4214 */
4215 if (costly_order && (!can_compact ||
4216 !(gfp_mask & __GFP_RETRY_MAYFAIL)))
4217 goto nopage;
4218
4219 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4220 did_some_progress > 0, &no_progress_loops))
4221 goto retry;
4222
4223 /*
4224 * It doesn't make any sense to retry for the compaction if the order-0
4225 * reclaim is not able to make any progress because the current
4226 * implementation of the compaction depends on the sufficient amount
4227 * of free memory (see __compaction_suitable)
4228 */
4229 if (did_some_progress > 0 && can_compact &&
4230 should_compact_retry(ac, order, alloc_flags,
4231 compact_result, &compact_priority,
4232 &compaction_retries))
4233 goto retry;
4234
4235
4236 /*
4237 * Deal with possible cpuset update races or zonelist updates to avoid
4238 * a unnecessary OOM kill.
4239 */
4240 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4241 check_retry_zonelist(zonelist_iter_cookie))
4242 goto restart;
4243
4244 /* Reclaim has failed us, start killing things */
4245 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4246 if (page)
4247 goto got_pg;
4248
4249 /* Avoid allocations with no watermarks from looping endlessly */
4250 if (tsk_is_oom_victim(current) &&
4251 (alloc_flags & ALLOC_OOM ||
4252 (gfp_mask & __GFP_NOMEMALLOC)))
4253 goto nopage;
4254
4255 /* Retry as long as the OOM killer is making progress */
4256 if (did_some_progress) {
4257 no_progress_loops = 0;
4258 goto retry;
4259 }
4260
4261nopage:
4262 /*
4263 * Deal with possible cpuset update races or zonelist updates to avoid
4264 * a unnecessary OOM kill.
4265 */
4266 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4267 check_retry_zonelist(zonelist_iter_cookie))
4268 goto restart;
4269
4270 /*
4271 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4272 * we always retry
4273 */
4274 if (gfp_mask & __GFP_NOFAIL) {
4275 /*
4276 * All existing users of the __GFP_NOFAIL are blockable, so warn
4277 * of any new users that actually require GFP_NOWAIT
4278 */
4279 if (WARN_ON_ONCE_GFP(!can_direct_reclaim, gfp_mask))
4280 goto fail;
4281
4282 /*
4283 * PF_MEMALLOC request from this context is rather bizarre
4284 * because we cannot reclaim anything and only can loop waiting
4285 * for somebody to do a work for us
4286 */
4287 WARN_ON_ONCE_GFP(current->flags & PF_MEMALLOC, gfp_mask);
4288
4289 /*
4290 * non failing costly orders are a hard requirement which we
4291 * are not prepared for much so let's warn about these users
4292 * so that we can identify them and convert them to something
4293 * else.
4294 */
4295 WARN_ON_ONCE_GFP(costly_order, gfp_mask);
4296
4297 /*
4298 * Help non-failing allocations by giving some access to memory
4299 * reserves normally used for high priority non-blocking
4300 * allocations but do not use ALLOC_NO_WATERMARKS because this
4301 * could deplete whole memory reserves which would just make
4302 * the situation worse.
4303 */
4304 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_MIN_RESERVE, ac);
4305 if (page)
4306 goto got_pg;
4307
4308 cond_resched();
4309 goto retry;
4310 }
4311fail:
4312 warn_alloc(gfp_mask, ac->nodemask,
4313 "page allocation failure: order:%u", order);
4314got_pg:
4315 return page;
4316}
4317
4318static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4319 int preferred_nid, nodemask_t *nodemask,
4320 struct alloc_context *ac, gfp_t *alloc_gfp,
4321 unsigned int *alloc_flags)
4322{
4323 ac->highest_zoneidx = gfp_zone(gfp_mask);
4324 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4325 ac->nodemask = nodemask;
4326 ac->migratetype = gfp_migratetype(gfp_mask);
4327
4328 if (cpusets_enabled()) {
4329 *alloc_gfp |= __GFP_HARDWALL;
4330 /*
4331 * When we are in the interrupt context, it is irrelevant
4332 * to the current task context. It means that any node ok.
4333 */
4334 if (in_task() && !ac->nodemask)
4335 ac->nodemask = &cpuset_current_mems_allowed;
4336 else
4337 *alloc_flags |= ALLOC_CPUSET;
4338 }
4339
4340 might_alloc(gfp_mask);
4341
4342 if (should_fail_alloc_page(gfp_mask, order))
4343 return false;
4344
4345 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
4346
4347 /* Dirty zone balancing only done in the fast path */
4348 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4349
4350 /*
4351 * The preferred zone is used for statistics but crucially it is
4352 * also used as the starting point for the zonelist iterator. It
4353 * may get reset for allocations that ignore memory policies.
4354 */
4355 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4356 ac->highest_zoneidx, ac->nodemask);
4357
4358 return true;
4359}
4360
4361/*
4362 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
4363 * @gfp: GFP flags for the allocation
4364 * @preferred_nid: The preferred NUMA node ID to allocate from
4365 * @nodemask: Set of nodes to allocate from, may be NULL
4366 * @nr_pages: The number of pages desired on the list or array
4367 * @page_list: Optional list to store the allocated pages
4368 * @page_array: Optional array to store the pages
4369 *
4370 * This is a batched version of the page allocator that attempts to
4371 * allocate nr_pages quickly. Pages are added to page_list if page_list
4372 * is not NULL, otherwise it is assumed that the page_array is valid.
4373 *
4374 * For lists, nr_pages is the number of pages that should be allocated.
4375 *
4376 * For arrays, only NULL elements are populated with pages and nr_pages
4377 * is the maximum number of pages that will be stored in the array.
4378 *
4379 * Returns the number of pages on the list or array.
4380 */
4381unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
4382 nodemask_t *nodemask, int nr_pages,
4383 struct list_head *page_list,
4384 struct page **page_array)
4385{
4386 struct page *page;
4387 unsigned long __maybe_unused UP_flags;
4388 struct zone *zone;
4389 struct zoneref *z;
4390 struct per_cpu_pages *pcp;
4391 struct list_head *pcp_list;
4392 struct alloc_context ac;
4393 gfp_t alloc_gfp;
4394 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4395 int nr_populated = 0, nr_account = 0;
4396
4397 /*
4398 * Skip populated array elements to determine if any pages need
4399 * to be allocated before disabling IRQs.
4400 */
4401 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
4402 nr_populated++;
4403
4404 /* No pages requested? */
4405 if (unlikely(nr_pages <= 0))
4406 goto out;
4407
4408 /* Already populated array? */
4409 if (unlikely(page_array && nr_pages - nr_populated == 0))
4410 goto out;
4411
4412 /* Bulk allocator does not support memcg accounting. */
4413 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT))
4414 goto failed;
4415
4416 /* Use the single page allocator for one page. */
4417 if (nr_pages - nr_populated == 1)
4418 goto failed;
4419
4420#ifdef CONFIG_PAGE_OWNER
4421 /*
4422 * PAGE_OWNER may recurse into the allocator to allocate space to
4423 * save the stack with pagesets.lock held. Releasing/reacquiring
4424 * removes much of the performance benefit of bulk allocation so
4425 * force the caller to allocate one page at a time as it'll have
4426 * similar performance to added complexity to the bulk allocator.
4427 */
4428 if (static_branch_unlikely(&page_owner_inited))
4429 goto failed;
4430#endif
4431
4432 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
4433 gfp &= gfp_allowed_mask;
4434 alloc_gfp = gfp;
4435 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
4436 goto out;
4437 gfp = alloc_gfp;
4438
4439 /* Find an allowed local zone that meets the low watermark. */
4440 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
4441 unsigned long mark;
4442
4443 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
4444 !__cpuset_zone_allowed(zone, gfp)) {
4445 continue;
4446 }
4447
4448 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
4449 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
4450 goto failed;
4451 }
4452
4453 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
4454 if (zone_watermark_fast(zone, 0, mark,
4455 zonelist_zone_idx(ac.preferred_zoneref),
4456 alloc_flags, gfp)) {
4457 break;
4458 }
4459 }
4460
4461 /*
4462 * If there are no allowed local zones that meets the watermarks then
4463 * try to allocate a single page and reclaim if necessary.
4464 */
4465 if (unlikely(!zone))
4466 goto failed;
4467
4468 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
4469 pcp_trylock_prepare(UP_flags);
4470 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
4471 if (!pcp)
4472 goto failed_irq;
4473
4474 /* Attempt the batch allocation */
4475 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
4476 while (nr_populated < nr_pages) {
4477
4478 /* Skip existing pages */
4479 if (page_array && page_array[nr_populated]) {
4480 nr_populated++;
4481 continue;
4482 }
4483
4484 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
4485 pcp, pcp_list);
4486 if (unlikely(!page)) {
4487 /* Try and allocate at least one page */
4488 if (!nr_account) {
4489 pcp_spin_unlock(pcp);
4490 goto failed_irq;
4491 }
4492 break;
4493 }
4494 nr_account++;
4495
4496 prep_new_page(page, 0, gfp, 0);
4497 if (page_list)
4498 list_add(&page->lru, page_list);
4499 else
4500 page_array[nr_populated] = page;
4501 nr_populated++;
4502 }
4503
4504 pcp_spin_unlock(pcp);
4505 pcp_trylock_finish(UP_flags);
4506
4507 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
4508 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
4509
4510out:
4511 return nr_populated;
4512
4513failed_irq:
4514 pcp_trylock_finish(UP_flags);
4515
4516failed:
4517 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
4518 if (page) {
4519 if (page_list)
4520 list_add(&page->lru, page_list);
4521 else
4522 page_array[nr_populated] = page;
4523 nr_populated++;
4524 }
4525
4526 goto out;
4527}
4528EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
4529
4530/*
4531 * This is the 'heart' of the zoned buddy allocator.
4532 */
4533struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
4534 nodemask_t *nodemask)
4535{
4536 struct page *page;
4537 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4538 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
4539 struct alloc_context ac = { };
4540
4541 /*
4542 * There are several places where we assume that the order value is sane
4543 * so bail out early if the request is out of bound.
4544 */
4545 if (WARN_ON_ONCE_GFP(order > MAX_PAGE_ORDER, gfp))
4546 return NULL;
4547
4548 gfp &= gfp_allowed_mask;
4549 /*
4550 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4551 * resp. GFP_NOIO which has to be inherited for all allocation requests
4552 * from a particular context which has been marked by
4553 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
4554 * movable zones are not used during allocation.
4555 */
4556 gfp = current_gfp_context(gfp);
4557 alloc_gfp = gfp;
4558 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
4559 &alloc_gfp, &alloc_flags))
4560 return NULL;
4561
4562 /*
4563 * Forbid the first pass from falling back to types that fragment
4564 * memory until all local zones are considered.
4565 */
4566 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
4567
4568 /* First allocation attempt */
4569 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
4570 if (likely(page))
4571 goto out;
4572
4573 alloc_gfp = gfp;
4574 ac.spread_dirty_pages = false;
4575
4576 /*
4577 * Restore the original nodemask if it was potentially replaced with
4578 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4579 */
4580 ac.nodemask = nodemask;
4581
4582 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
4583
4584out:
4585 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT) && page &&
4586 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
4587 __free_pages(page, order);
4588 page = NULL;
4589 }
4590
4591 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
4592 kmsan_alloc_page(page, order, alloc_gfp);
4593
4594 return page;
4595}
4596EXPORT_SYMBOL(__alloc_pages);
4597
4598struct folio *__folio_alloc(gfp_t gfp, unsigned int order, int preferred_nid,
4599 nodemask_t *nodemask)
4600{
4601 struct page *page = __alloc_pages(gfp | __GFP_COMP, order,
4602 preferred_nid, nodemask);
4603 return page_rmappable_folio(page);
4604}
4605EXPORT_SYMBOL(__folio_alloc);
4606
4607/*
4608 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4609 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4610 * you need to access high mem.
4611 */
4612unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4613{
4614 struct page *page;
4615
4616 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4617 if (!page)
4618 return 0;
4619 return (unsigned long) page_address(page);
4620}
4621EXPORT_SYMBOL(__get_free_pages);
4622
4623unsigned long get_zeroed_page(gfp_t gfp_mask)
4624{
4625 return __get_free_page(gfp_mask | __GFP_ZERO);
4626}
4627EXPORT_SYMBOL(get_zeroed_page);
4628
4629/**
4630 * __free_pages - Free pages allocated with alloc_pages().
4631 * @page: The page pointer returned from alloc_pages().
4632 * @order: The order of the allocation.
4633 *
4634 * This function can free multi-page allocations that are not compound
4635 * pages. It does not check that the @order passed in matches that of
4636 * the allocation, so it is easy to leak memory. Freeing more memory
4637 * than was allocated will probably emit a warning.
4638 *
4639 * If the last reference to this page is speculative, it will be released
4640 * by put_page() which only frees the first page of a non-compound
4641 * allocation. To prevent the remaining pages from being leaked, we free
4642 * the subsequent pages here. If you want to use the page's reference
4643 * count to decide when to free the allocation, you should allocate a
4644 * compound page, and use put_page() instead of __free_pages().
4645 *
4646 * Context: May be called in interrupt context or while holding a normal
4647 * spinlock, but not in NMI context or while holding a raw spinlock.
4648 */
4649void __free_pages(struct page *page, unsigned int order)
4650{
4651 /* get PageHead before we drop reference */
4652 int head = PageHead(page);
4653
4654 if (put_page_testzero(page))
4655 free_the_page(page, order);
4656 else if (!head)
4657 while (order-- > 0)
4658 free_the_page(page + (1 << order), order);
4659}
4660EXPORT_SYMBOL(__free_pages);
4661
4662void free_pages(unsigned long addr, unsigned int order)
4663{
4664 if (addr != 0) {
4665 VM_BUG_ON(!virt_addr_valid((void *)addr));
4666 __free_pages(virt_to_page((void *)addr), order);
4667 }
4668}
4669
4670EXPORT_SYMBOL(free_pages);
4671
4672/*
4673 * Page Fragment:
4674 * An arbitrary-length arbitrary-offset area of memory which resides
4675 * within a 0 or higher order page. Multiple fragments within that page
4676 * are individually refcounted, in the page's reference counter.
4677 *
4678 * The page_frag functions below provide a simple allocation framework for
4679 * page fragments. This is used by the network stack and network device
4680 * drivers to provide a backing region of memory for use as either an
4681 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4682 */
4683static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4684 gfp_t gfp_mask)
4685{
4686 struct page *page = NULL;
4687 gfp_t gfp = gfp_mask;
4688
4689#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4690 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4691 __GFP_NOMEMALLOC;
4692 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4693 PAGE_FRAG_CACHE_MAX_ORDER);
4694 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4695#endif
4696 if (unlikely(!page))
4697 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4698
4699 nc->va = page ? page_address(page) : NULL;
4700
4701 return page;
4702}
4703
4704void __page_frag_cache_drain(struct page *page, unsigned int count)
4705{
4706 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4707
4708 if (page_ref_sub_and_test(page, count))
4709 free_the_page(page, compound_order(page));
4710}
4711EXPORT_SYMBOL(__page_frag_cache_drain);
4712
4713void *page_frag_alloc_align(struct page_frag_cache *nc,
4714 unsigned int fragsz, gfp_t gfp_mask,
4715 unsigned int align_mask)
4716{
4717 unsigned int size = PAGE_SIZE;
4718 struct page *page;
4719 int offset;
4720
4721 if (unlikely(!nc->va)) {
4722refill:
4723 page = __page_frag_cache_refill(nc, gfp_mask);
4724 if (!page)
4725 return NULL;
4726
4727#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4728 /* if size can vary use size else just use PAGE_SIZE */
4729 size = nc->size;
4730#endif
4731 /* Even if we own the page, we do not use atomic_set().
4732 * This would break get_page_unless_zero() users.
4733 */
4734 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4735
4736 /* reset page count bias and offset to start of new frag */
4737 nc->pfmemalloc = page_is_pfmemalloc(page);
4738 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4739 nc->offset = size;
4740 }
4741
4742 offset = nc->offset - fragsz;
4743 if (unlikely(offset < 0)) {
4744 page = virt_to_page(nc->va);
4745
4746 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4747 goto refill;
4748
4749 if (unlikely(nc->pfmemalloc)) {
4750 free_the_page(page, compound_order(page));
4751 goto refill;
4752 }
4753
4754#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4755 /* if size can vary use size else just use PAGE_SIZE */
4756 size = nc->size;
4757#endif
4758 /* OK, page count is 0, we can safely set it */
4759 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4760
4761 /* reset page count bias and offset to start of new frag */
4762 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4763 offset = size - fragsz;
4764 if (unlikely(offset < 0)) {
4765 /*
4766 * The caller is trying to allocate a fragment
4767 * with fragsz > PAGE_SIZE but the cache isn't big
4768 * enough to satisfy the request, this may
4769 * happen in low memory conditions.
4770 * We don't release the cache page because
4771 * it could make memory pressure worse
4772 * so we simply return NULL here.
4773 */
4774 return NULL;
4775 }
4776 }
4777
4778 nc->pagecnt_bias--;
4779 offset &= align_mask;
4780 nc->offset = offset;
4781
4782 return nc->va + offset;
4783}
4784EXPORT_SYMBOL(page_frag_alloc_align);
4785
4786/*
4787 * Frees a page fragment allocated out of either a compound or order 0 page.
4788 */
4789void page_frag_free(void *addr)
4790{
4791 struct page *page = virt_to_head_page(addr);
4792
4793 if (unlikely(put_page_testzero(page)))
4794 free_the_page(page, compound_order(page));
4795}
4796EXPORT_SYMBOL(page_frag_free);
4797
4798static void *make_alloc_exact(unsigned long addr, unsigned int order,
4799 size_t size)
4800{
4801 if (addr) {
4802 unsigned long nr = DIV_ROUND_UP(size, PAGE_SIZE);
4803 struct page *page = virt_to_page((void *)addr);
4804 struct page *last = page + nr;
4805
4806 split_page_owner(page, 1 << order);
4807 split_page_memcg(page, 1 << order);
4808 while (page < --last)
4809 set_page_refcounted(last);
4810
4811 last = page + (1UL << order);
4812 for (page += nr; page < last; page++)
4813 __free_pages_ok(page, 0, FPI_TO_TAIL);
4814 }
4815 return (void *)addr;
4816}
4817
4818/**
4819 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4820 * @size: the number of bytes to allocate
4821 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4822 *
4823 * This function is similar to alloc_pages(), except that it allocates the
4824 * minimum number of pages to satisfy the request. alloc_pages() can only
4825 * allocate memory in power-of-two pages.
4826 *
4827 * This function is also limited by MAX_PAGE_ORDER.
4828 *
4829 * Memory allocated by this function must be released by free_pages_exact().
4830 *
4831 * Return: pointer to the allocated area or %NULL in case of error.
4832 */
4833void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4834{
4835 unsigned int order = get_order(size);
4836 unsigned long addr;
4837
4838 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
4839 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
4840
4841 addr = __get_free_pages(gfp_mask, order);
4842 return make_alloc_exact(addr, order, size);
4843}
4844EXPORT_SYMBOL(alloc_pages_exact);
4845
4846/**
4847 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4848 * pages on a node.
4849 * @nid: the preferred node ID where memory should be allocated
4850 * @size: the number of bytes to allocate
4851 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4852 *
4853 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4854 * back.
4855 *
4856 * Return: pointer to the allocated area or %NULL in case of error.
4857 */
4858void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4859{
4860 unsigned int order = get_order(size);
4861 struct page *p;
4862
4863 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
4864 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
4865
4866 p = alloc_pages_node(nid, gfp_mask, order);
4867 if (!p)
4868 return NULL;
4869 return make_alloc_exact((unsigned long)page_address(p), order, size);
4870}
4871
4872/**
4873 * free_pages_exact - release memory allocated via alloc_pages_exact()
4874 * @virt: the value returned by alloc_pages_exact.
4875 * @size: size of allocation, same value as passed to alloc_pages_exact().
4876 *
4877 * Release the memory allocated by a previous call to alloc_pages_exact.
4878 */
4879void free_pages_exact(void *virt, size_t size)
4880{
4881 unsigned long addr = (unsigned long)virt;
4882 unsigned long end = addr + PAGE_ALIGN(size);
4883
4884 while (addr < end) {
4885 free_page(addr);
4886 addr += PAGE_SIZE;
4887 }
4888}
4889EXPORT_SYMBOL(free_pages_exact);
4890
4891/**
4892 * nr_free_zone_pages - count number of pages beyond high watermark
4893 * @offset: The zone index of the highest zone
4894 *
4895 * nr_free_zone_pages() counts the number of pages which are beyond the
4896 * high watermark within all zones at or below a given zone index. For each
4897 * zone, the number of pages is calculated as:
4898 *
4899 * nr_free_zone_pages = managed_pages - high_pages
4900 *
4901 * Return: number of pages beyond high watermark.
4902 */
4903static unsigned long nr_free_zone_pages(int offset)
4904{
4905 struct zoneref *z;
4906 struct zone *zone;
4907
4908 /* Just pick one node, since fallback list is circular */
4909 unsigned long sum = 0;
4910
4911 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4912
4913 for_each_zone_zonelist(zone, z, zonelist, offset) {
4914 unsigned long size = zone_managed_pages(zone);
4915 unsigned long high = high_wmark_pages(zone);
4916 if (size > high)
4917 sum += size - high;
4918 }
4919
4920 return sum;
4921}
4922
4923/**
4924 * nr_free_buffer_pages - count number of pages beyond high watermark
4925 *
4926 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4927 * watermark within ZONE_DMA and ZONE_NORMAL.
4928 *
4929 * Return: number of pages beyond high watermark within ZONE_DMA and
4930 * ZONE_NORMAL.
4931 */
4932unsigned long nr_free_buffer_pages(void)
4933{
4934 return nr_free_zone_pages(gfp_zone(GFP_USER));
4935}
4936EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4937
4938static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4939{
4940 zoneref->zone = zone;
4941 zoneref->zone_idx = zone_idx(zone);
4942}
4943
4944/*
4945 * Builds allocation fallback zone lists.
4946 *
4947 * Add all populated zones of a node to the zonelist.
4948 */
4949static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
4950{
4951 struct zone *zone;
4952 enum zone_type zone_type = MAX_NR_ZONES;
4953 int nr_zones = 0;
4954
4955 do {
4956 zone_type--;
4957 zone = pgdat->node_zones + zone_type;
4958 if (populated_zone(zone)) {
4959 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
4960 check_highest_zone(zone_type);
4961 }
4962 } while (zone_type);
4963
4964 return nr_zones;
4965}
4966
4967#ifdef CONFIG_NUMA
4968
4969static int __parse_numa_zonelist_order(char *s)
4970{
4971 /*
4972 * We used to support different zonelists modes but they turned
4973 * out to be just not useful. Let's keep the warning in place
4974 * if somebody still use the cmd line parameter so that we do
4975 * not fail it silently
4976 */
4977 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
4978 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
4979 return -EINVAL;
4980 }
4981 return 0;
4982}
4983
4984static char numa_zonelist_order[] = "Node";
4985#define NUMA_ZONELIST_ORDER_LEN 16
4986/*
4987 * sysctl handler for numa_zonelist_order
4988 */
4989static int numa_zonelist_order_handler(struct ctl_table *table, int write,
4990 void *buffer, size_t *length, loff_t *ppos)
4991{
4992 if (write)
4993 return __parse_numa_zonelist_order(buffer);
4994 return proc_dostring(table, write, buffer, length, ppos);
4995}
4996
4997static int node_load[MAX_NUMNODES];
4998
4999/**
5000 * find_next_best_node - find the next node that should appear in a given node's fallback list
5001 * @node: node whose fallback list we're appending
5002 * @used_node_mask: nodemask_t of already used nodes
5003 *
5004 * We use a number of factors to determine which is the next node that should
5005 * appear on a given node's fallback list. The node should not have appeared
5006 * already in @node's fallback list, and it should be the next closest node
5007 * according to the distance array (which contains arbitrary distance values
5008 * from each node to each node in the system), and should also prefer nodes
5009 * with no CPUs, since presumably they'll have very little allocation pressure
5010 * on them otherwise.
5011 *
5012 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5013 */
5014int find_next_best_node(int node, nodemask_t *used_node_mask)
5015{
5016 int n, val;
5017 int min_val = INT_MAX;
5018 int best_node = NUMA_NO_NODE;
5019
5020 /*
5021 * Use the local node if we haven't already, but for memoryless local
5022 * node, we should skip it and fall back to other nodes.
5023 */
5024 if (!node_isset(node, *used_node_mask) && node_state(node, N_MEMORY)) {
5025 node_set(node, *used_node_mask);
5026 return node;
5027 }
5028
5029 for_each_node_state(n, N_MEMORY) {
5030
5031 /* Don't want a node to appear more than once */
5032 if (node_isset(n, *used_node_mask))
5033 continue;
5034
5035 /* Use the distance array to find the distance */
5036 val = node_distance(node, n);
5037
5038 /* Penalize nodes under us ("prefer the next node") */
5039 val += (n < node);
5040
5041 /* Give preference to headless and unused nodes */
5042 if (!cpumask_empty(cpumask_of_node(n)))
5043 val += PENALTY_FOR_NODE_WITH_CPUS;
5044
5045 /* Slight preference for less loaded node */
5046 val *= MAX_NUMNODES;
5047 val += node_load[n];
5048
5049 if (val < min_val) {
5050 min_val = val;
5051 best_node = n;
5052 }
5053 }
5054
5055 if (best_node >= 0)
5056 node_set(best_node, *used_node_mask);
5057
5058 return best_node;
5059}
5060
5061
5062/*
5063 * Build zonelists ordered by node and zones within node.
5064 * This results in maximum locality--normal zone overflows into local
5065 * DMA zone, if any--but risks exhausting DMA zone.
5066 */
5067static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5068 unsigned nr_nodes)
5069{
5070 struct zoneref *zonerefs;
5071 int i;
5072
5073 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5074
5075 for (i = 0; i < nr_nodes; i++) {
5076 int nr_zones;
5077
5078 pg_data_t *node = NODE_DATA(node_order[i]);
5079
5080 nr_zones = build_zonerefs_node(node, zonerefs);
5081 zonerefs += nr_zones;
5082 }
5083 zonerefs->zone = NULL;
5084 zonerefs->zone_idx = 0;
5085}
5086
5087/*
5088 * Build gfp_thisnode zonelists
5089 */
5090static void build_thisnode_zonelists(pg_data_t *pgdat)
5091{
5092 struct zoneref *zonerefs;
5093 int nr_zones;
5094
5095 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5096 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5097 zonerefs += nr_zones;
5098 zonerefs->zone = NULL;
5099 zonerefs->zone_idx = 0;
5100}
5101
5102/*
5103 * Build zonelists ordered by zone and nodes within zones.
5104 * This results in conserving DMA zone[s] until all Normal memory is
5105 * exhausted, but results in overflowing to remote node while memory
5106 * may still exist in local DMA zone.
5107 */
5108
5109static void build_zonelists(pg_data_t *pgdat)
5110{
5111 static int node_order[MAX_NUMNODES];
5112 int node, nr_nodes = 0;
5113 nodemask_t used_mask = NODE_MASK_NONE;
5114 int local_node, prev_node;
5115
5116 /* NUMA-aware ordering of nodes */
5117 local_node = pgdat->node_id;
5118 prev_node = local_node;
5119
5120 memset(node_order, 0, sizeof(node_order));
5121 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5122 /*
5123 * We don't want to pressure a particular node.
5124 * So adding penalty to the first node in same
5125 * distance group to make it round-robin.
5126 */
5127 if (node_distance(local_node, node) !=
5128 node_distance(local_node, prev_node))
5129 node_load[node] += 1;
5130
5131 node_order[nr_nodes++] = node;
5132 prev_node = node;
5133 }
5134
5135 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5136 build_thisnode_zonelists(pgdat);
5137 pr_info("Fallback order for Node %d: ", local_node);
5138 for (node = 0; node < nr_nodes; node++)
5139 pr_cont("%d ", node_order[node]);
5140 pr_cont("\n");
5141}
5142
5143#ifdef CONFIG_HAVE_MEMORYLESS_NODES
5144/*
5145 * Return node id of node used for "local" allocations.
5146 * I.e., first node id of first zone in arg node's generic zonelist.
5147 * Used for initializing percpu 'numa_mem', which is used primarily
5148 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5149 */
5150int local_memory_node(int node)
5151{
5152 struct zoneref *z;
5153
5154 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5155 gfp_zone(GFP_KERNEL),
5156 NULL);
5157 return zone_to_nid(z->zone);
5158}
5159#endif
5160
5161static void setup_min_unmapped_ratio(void);
5162static void setup_min_slab_ratio(void);
5163#else /* CONFIG_NUMA */
5164
5165static void build_zonelists(pg_data_t *pgdat)
5166{
5167 int node, local_node;
5168 struct zoneref *zonerefs;
5169 int nr_zones;
5170
5171 local_node = pgdat->node_id;
5172
5173 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5174 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5175 zonerefs += nr_zones;
5176
5177 /*
5178 * Now we build the zonelist so that it contains the zones
5179 * of all the other nodes.
5180 * We don't want to pressure a particular node, so when
5181 * building the zones for node N, we make sure that the
5182 * zones coming right after the local ones are those from
5183 * node N+1 (modulo N)
5184 */
5185 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5186 if (!node_online(node))
5187 continue;
5188 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5189 zonerefs += nr_zones;
5190 }
5191 for (node = 0; node < local_node; node++) {
5192 if (!node_online(node))
5193 continue;
5194 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5195 zonerefs += nr_zones;
5196 }
5197
5198 zonerefs->zone = NULL;
5199 zonerefs->zone_idx = 0;
5200}
5201
5202#endif /* CONFIG_NUMA */
5203
5204/*
5205 * Boot pageset table. One per cpu which is going to be used for all
5206 * zones and all nodes. The parameters will be set in such a way
5207 * that an item put on a list will immediately be handed over to
5208 * the buddy list. This is safe since pageset manipulation is done
5209 * with interrupts disabled.
5210 *
5211 * The boot_pagesets must be kept even after bootup is complete for
5212 * unused processors and/or zones. They do play a role for bootstrapping
5213 * hotplugged processors.
5214 *
5215 * zoneinfo_show() and maybe other functions do
5216 * not check if the processor is online before following the pageset pointer.
5217 * Other parts of the kernel may not check if the zone is available.
5218 */
5219static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
5220/* These effectively disable the pcplists in the boot pageset completely */
5221#define BOOT_PAGESET_HIGH 0
5222#define BOOT_PAGESET_BATCH 1
5223static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
5224static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
5225
5226static void __build_all_zonelists(void *data)
5227{
5228 int nid;
5229 int __maybe_unused cpu;
5230 pg_data_t *self = data;
5231 unsigned long flags;
5232
5233 /*
5234 * The zonelist_update_seq must be acquired with irqsave because the
5235 * reader can be invoked from IRQ with GFP_ATOMIC.
5236 */
5237 write_seqlock_irqsave(&zonelist_update_seq, flags);
5238 /*
5239 * Also disable synchronous printk() to prevent any printk() from
5240 * trying to hold port->lock, for
5241 * tty_insert_flip_string_and_push_buffer() on other CPU might be
5242 * calling kmalloc(GFP_ATOMIC | __GFP_NOWARN) with port->lock held.
5243 */
5244 printk_deferred_enter();
5245
5246#ifdef CONFIG_NUMA
5247 memset(node_load, 0, sizeof(node_load));
5248#endif
5249
5250 /*
5251 * This node is hotadded and no memory is yet present. So just
5252 * building zonelists is fine - no need to touch other nodes.
5253 */
5254 if (self && !node_online(self->node_id)) {
5255 build_zonelists(self);
5256 } else {
5257 /*
5258 * All possible nodes have pgdat preallocated
5259 * in free_area_init
5260 */
5261 for_each_node(nid) {
5262 pg_data_t *pgdat = NODE_DATA(nid);
5263
5264 build_zonelists(pgdat);
5265 }
5266
5267#ifdef CONFIG_HAVE_MEMORYLESS_NODES
5268 /*
5269 * We now know the "local memory node" for each node--
5270 * i.e., the node of the first zone in the generic zonelist.
5271 * Set up numa_mem percpu variable for on-line cpus. During
5272 * boot, only the boot cpu should be on-line; we'll init the
5273 * secondary cpus' numa_mem as they come on-line. During
5274 * node/memory hotplug, we'll fixup all on-line cpus.
5275 */
5276 for_each_online_cpu(cpu)
5277 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5278#endif
5279 }
5280
5281 printk_deferred_exit();
5282 write_sequnlock_irqrestore(&zonelist_update_seq, flags);
5283}
5284
5285static noinline void __init
5286build_all_zonelists_init(void)
5287{
5288 int cpu;
5289
5290 __build_all_zonelists(NULL);
5291
5292 /*
5293 * Initialize the boot_pagesets that are going to be used
5294 * for bootstrapping processors. The real pagesets for
5295 * each zone will be allocated later when the per cpu
5296 * allocator is available.
5297 *
5298 * boot_pagesets are used also for bootstrapping offline
5299 * cpus if the system is already booted because the pagesets
5300 * are needed to initialize allocators on a specific cpu too.
5301 * F.e. the percpu allocator needs the page allocator which
5302 * needs the percpu allocator in order to allocate its pagesets
5303 * (a chicken-egg dilemma).
5304 */
5305 for_each_possible_cpu(cpu)
5306 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
5307
5308 mminit_verify_zonelist();
5309 cpuset_init_current_mems_allowed();
5310}
5311
5312/*
5313 * unless system_state == SYSTEM_BOOTING.
5314 *
5315 * __ref due to call of __init annotated helper build_all_zonelists_init
5316 * [protected by SYSTEM_BOOTING].
5317 */
5318void __ref build_all_zonelists(pg_data_t *pgdat)
5319{
5320 unsigned long vm_total_pages;
5321
5322 if (system_state == SYSTEM_BOOTING) {
5323 build_all_zonelists_init();
5324 } else {
5325 __build_all_zonelists(pgdat);
5326 /* cpuset refresh routine should be here */
5327 }
5328 /* Get the number of free pages beyond high watermark in all zones. */
5329 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5330 /*
5331 * Disable grouping by mobility if the number of pages in the
5332 * system is too low to allow the mechanism to work. It would be
5333 * more accurate, but expensive to check per-zone. This check is
5334 * made on memory-hotadd so a system can start with mobility
5335 * disabled and enable it later
5336 */
5337 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5338 page_group_by_mobility_disabled = 1;
5339 else
5340 page_group_by_mobility_disabled = 0;
5341
5342 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5343 nr_online_nodes,
5344 page_group_by_mobility_disabled ? "off" : "on",
5345 vm_total_pages);
5346#ifdef CONFIG_NUMA
5347 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5348#endif
5349}
5350
5351static int zone_batchsize(struct zone *zone)
5352{
5353#ifdef CONFIG_MMU
5354 int batch;
5355
5356 /*
5357 * The number of pages to batch allocate is either ~0.1%
5358 * of the zone or 1MB, whichever is smaller. The batch
5359 * size is striking a balance between allocation latency
5360 * and zone lock contention.
5361 */
5362 batch = min(zone_managed_pages(zone) >> 10, SZ_1M / PAGE_SIZE);
5363 batch /= 4; /* We effectively *= 4 below */
5364 if (batch < 1)
5365 batch = 1;
5366
5367 /*
5368 * Clamp the batch to a 2^n - 1 value. Having a power
5369 * of 2 value was found to be more likely to have
5370 * suboptimal cache aliasing properties in some cases.
5371 *
5372 * For example if 2 tasks are alternately allocating
5373 * batches of pages, one task can end up with a lot
5374 * of pages of one half of the possible page colors
5375 * and the other with pages of the other colors.
5376 */
5377 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5378
5379 return batch;
5380
5381#else
5382 /* The deferral and batching of frees should be suppressed under NOMMU
5383 * conditions.
5384 *
5385 * The problem is that NOMMU needs to be able to allocate large chunks
5386 * of contiguous memory as there's no hardware page translation to
5387 * assemble apparent contiguous memory from discontiguous pages.
5388 *
5389 * Queueing large contiguous runs of pages for batching, however,
5390 * causes the pages to actually be freed in smaller chunks. As there
5391 * can be a significant delay between the individual batches being
5392 * recycled, this leads to the once large chunks of space being
5393 * fragmented and becoming unavailable for high-order allocations.
5394 */
5395 return 0;
5396#endif
5397}
5398
5399static int percpu_pagelist_high_fraction;
5400static int zone_highsize(struct zone *zone, int batch, int cpu_online,
5401 int high_fraction)
5402{
5403#ifdef CONFIG_MMU
5404 int high;
5405 int nr_split_cpus;
5406 unsigned long total_pages;
5407
5408 if (!high_fraction) {
5409 /*
5410 * By default, the high value of the pcp is based on the zone
5411 * low watermark so that if they are full then background
5412 * reclaim will not be started prematurely.
5413 */
5414 total_pages = low_wmark_pages(zone);
5415 } else {
5416 /*
5417 * If percpu_pagelist_high_fraction is configured, the high
5418 * value is based on a fraction of the managed pages in the
5419 * zone.
5420 */
5421 total_pages = zone_managed_pages(zone) / high_fraction;
5422 }
5423
5424 /*
5425 * Split the high value across all online CPUs local to the zone. Note
5426 * that early in boot that CPUs may not be online yet and that during
5427 * CPU hotplug that the cpumask is not yet updated when a CPU is being
5428 * onlined. For memory nodes that have no CPUs, split the high value
5429 * across all online CPUs to mitigate the risk that reclaim is triggered
5430 * prematurely due to pages stored on pcp lists.
5431 */
5432 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
5433 if (!nr_split_cpus)
5434 nr_split_cpus = num_online_cpus();
5435 high = total_pages / nr_split_cpus;
5436
5437 /*
5438 * Ensure high is at least batch*4. The multiple is based on the
5439 * historical relationship between high and batch.
5440 */
5441 high = max(high, batch << 2);
5442
5443 return high;
5444#else
5445 return 0;
5446#endif
5447}
5448
5449/*
5450 * pcp->high and pcp->batch values are related and generally batch is lower
5451 * than high. They are also related to pcp->count such that count is lower
5452 * than high, and as soon as it reaches high, the pcplist is flushed.
5453 *
5454 * However, guaranteeing these relations at all times would require e.g. write
5455 * barriers here but also careful usage of read barriers at the read side, and
5456 * thus be prone to error and bad for performance. Thus the update only prevents
5457 * store tearing. Any new users of pcp->batch, pcp->high_min and pcp->high_max
5458 * should ensure they can cope with those fields changing asynchronously, and
5459 * fully trust only the pcp->count field on the local CPU with interrupts
5460 * disabled.
5461 *
5462 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5463 * outside of boot time (or some other assurance that no concurrent updaters
5464 * exist).
5465 */
5466static void pageset_update(struct per_cpu_pages *pcp, unsigned long high_min,
5467 unsigned long high_max, unsigned long batch)
5468{
5469 WRITE_ONCE(pcp->batch, batch);
5470 WRITE_ONCE(pcp->high_min, high_min);
5471 WRITE_ONCE(pcp->high_max, high_max);
5472}
5473
5474static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
5475{
5476 int pindex;
5477
5478 memset(pcp, 0, sizeof(*pcp));
5479 memset(pzstats, 0, sizeof(*pzstats));
5480
5481 spin_lock_init(&pcp->lock);
5482 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
5483 INIT_LIST_HEAD(&pcp->lists[pindex]);
5484
5485 /*
5486 * Set batch and high values safe for a boot pageset. A true percpu
5487 * pageset's initialization will update them subsequently. Here we don't
5488 * need to be as careful as pageset_update() as nobody can access the
5489 * pageset yet.
5490 */
5491 pcp->high_min = BOOT_PAGESET_HIGH;
5492 pcp->high_max = BOOT_PAGESET_HIGH;
5493 pcp->batch = BOOT_PAGESET_BATCH;
5494 pcp->free_count = 0;
5495}
5496
5497static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high_min,
5498 unsigned long high_max, unsigned long batch)
5499{
5500 struct per_cpu_pages *pcp;
5501 int cpu;
5502
5503 for_each_possible_cpu(cpu) {
5504 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5505 pageset_update(pcp, high_min, high_max, batch);
5506 }
5507}
5508
5509/*
5510 * Calculate and set new high and batch values for all per-cpu pagesets of a
5511 * zone based on the zone's size.
5512 */
5513static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
5514{
5515 int new_high_min, new_high_max, new_batch;
5516
5517 new_batch = max(1, zone_batchsize(zone));
5518 if (percpu_pagelist_high_fraction) {
5519 new_high_min = zone_highsize(zone, new_batch, cpu_online,
5520 percpu_pagelist_high_fraction);
5521 /*
5522 * PCP high is tuned manually, disable auto-tuning via
5523 * setting high_min and high_max to the manual value.
5524 */
5525 new_high_max = new_high_min;
5526 } else {
5527 new_high_min = zone_highsize(zone, new_batch, cpu_online, 0);
5528 new_high_max = zone_highsize(zone, new_batch, cpu_online,
5529 MIN_PERCPU_PAGELIST_HIGH_FRACTION);
5530 }
5531
5532 if (zone->pageset_high_min == new_high_min &&
5533 zone->pageset_high_max == new_high_max &&
5534 zone->pageset_batch == new_batch)
5535 return;
5536
5537 zone->pageset_high_min = new_high_min;
5538 zone->pageset_high_max = new_high_max;
5539 zone->pageset_batch = new_batch;
5540
5541 __zone_set_pageset_high_and_batch(zone, new_high_min, new_high_max,
5542 new_batch);
5543}
5544
5545void __meminit setup_zone_pageset(struct zone *zone)
5546{
5547 int cpu;
5548
5549 /* Size may be 0 on !SMP && !NUMA */
5550 if (sizeof(struct per_cpu_zonestat) > 0)
5551 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
5552
5553 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
5554 for_each_possible_cpu(cpu) {
5555 struct per_cpu_pages *pcp;
5556 struct per_cpu_zonestat *pzstats;
5557
5558 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5559 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
5560 per_cpu_pages_init(pcp, pzstats);
5561 }
5562
5563 zone_set_pageset_high_and_batch(zone, 0);
5564}
5565
5566/*
5567 * The zone indicated has a new number of managed_pages; batch sizes and percpu
5568 * page high values need to be recalculated.
5569 */
5570static void zone_pcp_update(struct zone *zone, int cpu_online)
5571{
5572 mutex_lock(&pcp_batch_high_lock);
5573 zone_set_pageset_high_and_batch(zone, cpu_online);
5574 mutex_unlock(&pcp_batch_high_lock);
5575}
5576
5577static void zone_pcp_update_cacheinfo(struct zone *zone)
5578{
5579 int cpu;
5580 struct per_cpu_pages *pcp;
5581 struct cpu_cacheinfo *cci;
5582
5583 for_each_online_cpu(cpu) {
5584 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5585 cci = get_cpu_cacheinfo(cpu);
5586 /*
5587 * If data cache slice of CPU is large enough, "pcp->batch"
5588 * pages can be preserved in PCP before draining PCP for
5589 * consecutive high-order pages freeing without allocation.
5590 * This can reduce zone lock contention without hurting
5591 * cache-hot pages sharing.
5592 */
5593 spin_lock(&pcp->lock);
5594 if ((cci->per_cpu_data_slice_size >> PAGE_SHIFT) > 3 * pcp->batch)
5595 pcp->flags |= PCPF_FREE_HIGH_BATCH;
5596 else
5597 pcp->flags &= ~PCPF_FREE_HIGH_BATCH;
5598 spin_unlock(&pcp->lock);
5599 }
5600}
5601
5602void setup_pcp_cacheinfo(void)
5603{
5604 struct zone *zone;
5605
5606 for_each_populated_zone(zone)
5607 zone_pcp_update_cacheinfo(zone);
5608}
5609
5610/*
5611 * Allocate per cpu pagesets and initialize them.
5612 * Before this call only boot pagesets were available.
5613 */
5614void __init setup_per_cpu_pageset(void)
5615{
5616 struct pglist_data *pgdat;
5617 struct zone *zone;
5618 int __maybe_unused cpu;
5619
5620 for_each_populated_zone(zone)
5621 setup_zone_pageset(zone);
5622
5623#ifdef CONFIG_NUMA
5624 /*
5625 * Unpopulated zones continue using the boot pagesets.
5626 * The numa stats for these pagesets need to be reset.
5627 * Otherwise, they will end up skewing the stats of
5628 * the nodes these zones are associated with.
5629 */
5630 for_each_possible_cpu(cpu) {
5631 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
5632 memset(pzstats->vm_numa_event, 0,
5633 sizeof(pzstats->vm_numa_event));
5634 }
5635#endif
5636
5637 for_each_online_pgdat(pgdat)
5638 pgdat->per_cpu_nodestats =
5639 alloc_percpu(struct per_cpu_nodestat);
5640}
5641
5642__meminit void zone_pcp_init(struct zone *zone)
5643{
5644 /*
5645 * per cpu subsystem is not up at this point. The following code
5646 * relies on the ability of the linker to provide the
5647 * offset of a (static) per cpu variable into the per cpu area.
5648 */
5649 zone->per_cpu_pageset = &boot_pageset;
5650 zone->per_cpu_zonestats = &boot_zonestats;
5651 zone->pageset_high_min = BOOT_PAGESET_HIGH;
5652 zone->pageset_high_max = BOOT_PAGESET_HIGH;
5653 zone->pageset_batch = BOOT_PAGESET_BATCH;
5654
5655 if (populated_zone(zone))
5656 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
5657 zone->present_pages, zone_batchsize(zone));
5658}
5659
5660void adjust_managed_page_count(struct page *page, long count)
5661{
5662 atomic_long_add(count, &page_zone(page)->managed_pages);
5663 totalram_pages_add(count);
5664#ifdef CONFIG_HIGHMEM
5665 if (PageHighMem(page))
5666 totalhigh_pages_add(count);
5667#endif
5668}
5669EXPORT_SYMBOL(adjust_managed_page_count);
5670
5671unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
5672{
5673 void *pos;
5674 unsigned long pages = 0;
5675
5676 start = (void *)PAGE_ALIGN((unsigned long)start);
5677 end = (void *)((unsigned long)end & PAGE_MASK);
5678 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5679 struct page *page = virt_to_page(pos);
5680 void *direct_map_addr;
5681
5682 /*
5683 * 'direct_map_addr' might be different from 'pos'
5684 * because some architectures' virt_to_page()
5685 * work with aliases. Getting the direct map
5686 * address ensures that we get a _writeable_
5687 * alias for the memset().
5688 */
5689 direct_map_addr = page_address(page);
5690 /*
5691 * Perform a kasan-unchecked memset() since this memory
5692 * has not been initialized.
5693 */
5694 direct_map_addr = kasan_reset_tag(direct_map_addr);
5695 if ((unsigned int)poison <= 0xFF)
5696 memset(direct_map_addr, poison, PAGE_SIZE);
5697
5698 free_reserved_page(page);
5699 }
5700
5701 if (pages && s)
5702 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
5703
5704 return pages;
5705}
5706
5707static int page_alloc_cpu_dead(unsigned int cpu)
5708{
5709 struct zone *zone;
5710
5711 lru_add_drain_cpu(cpu);
5712 mlock_drain_remote(cpu);
5713 drain_pages(cpu);
5714
5715 /*
5716 * Spill the event counters of the dead processor
5717 * into the current processors event counters.
5718 * This artificially elevates the count of the current
5719 * processor.
5720 */
5721 vm_events_fold_cpu(cpu);
5722
5723 /*
5724 * Zero the differential counters of the dead processor
5725 * so that the vm statistics are consistent.
5726 *
5727 * This is only okay since the processor is dead and cannot
5728 * race with what we are doing.
5729 */
5730 cpu_vm_stats_fold(cpu);
5731
5732 for_each_populated_zone(zone)
5733 zone_pcp_update(zone, 0);
5734
5735 return 0;
5736}
5737
5738static int page_alloc_cpu_online(unsigned int cpu)
5739{
5740 struct zone *zone;
5741
5742 for_each_populated_zone(zone)
5743 zone_pcp_update(zone, 1);
5744 return 0;
5745}
5746
5747void __init page_alloc_init_cpuhp(void)
5748{
5749 int ret;
5750
5751 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
5752 "mm/page_alloc:pcp",
5753 page_alloc_cpu_online,
5754 page_alloc_cpu_dead);
5755 WARN_ON(ret < 0);
5756}
5757
5758/*
5759 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5760 * or min_free_kbytes changes.
5761 */
5762static void calculate_totalreserve_pages(void)
5763{
5764 struct pglist_data *pgdat;
5765 unsigned long reserve_pages = 0;
5766 enum zone_type i, j;
5767
5768 for_each_online_pgdat(pgdat) {
5769
5770 pgdat->totalreserve_pages = 0;
5771
5772 for (i = 0; i < MAX_NR_ZONES; i++) {
5773 struct zone *zone = pgdat->node_zones + i;
5774 long max = 0;
5775 unsigned long managed_pages = zone_managed_pages(zone);
5776
5777 /* Find valid and maximum lowmem_reserve in the zone */
5778 for (j = i; j < MAX_NR_ZONES; j++) {
5779 if (zone->lowmem_reserve[j] > max)
5780 max = zone->lowmem_reserve[j];
5781 }
5782
5783 /* we treat the high watermark as reserved pages. */
5784 max += high_wmark_pages(zone);
5785
5786 if (max > managed_pages)
5787 max = managed_pages;
5788
5789 pgdat->totalreserve_pages += max;
5790
5791 reserve_pages += max;
5792 }
5793 }
5794 totalreserve_pages = reserve_pages;
5795}
5796
5797/*
5798 * setup_per_zone_lowmem_reserve - called whenever
5799 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
5800 * has a correct pages reserved value, so an adequate number of
5801 * pages are left in the zone after a successful __alloc_pages().
5802 */
5803static void setup_per_zone_lowmem_reserve(void)
5804{
5805 struct pglist_data *pgdat;
5806 enum zone_type i, j;
5807
5808 for_each_online_pgdat(pgdat) {
5809 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
5810 struct zone *zone = &pgdat->node_zones[i];
5811 int ratio = sysctl_lowmem_reserve_ratio[i];
5812 bool clear = !ratio || !zone_managed_pages(zone);
5813 unsigned long managed_pages = 0;
5814
5815 for (j = i + 1; j < MAX_NR_ZONES; j++) {
5816 struct zone *upper_zone = &pgdat->node_zones[j];
5817
5818 managed_pages += zone_managed_pages(upper_zone);
5819
5820 if (clear)
5821 zone->lowmem_reserve[j] = 0;
5822 else
5823 zone->lowmem_reserve[j] = managed_pages / ratio;
5824 }
5825 }
5826 }
5827
5828 /* update totalreserve_pages */
5829 calculate_totalreserve_pages();
5830}
5831
5832static void __setup_per_zone_wmarks(void)
5833{
5834 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5835 unsigned long lowmem_pages = 0;
5836 struct zone *zone;
5837 unsigned long flags;
5838
5839 /* Calculate total number of !ZONE_HIGHMEM and !ZONE_MOVABLE pages */
5840 for_each_zone(zone) {
5841 if (!is_highmem(zone) && zone_idx(zone) != ZONE_MOVABLE)
5842 lowmem_pages += zone_managed_pages(zone);
5843 }
5844
5845 for_each_zone(zone) {
5846 u64 tmp;
5847
5848 spin_lock_irqsave(&zone->lock, flags);
5849 tmp = (u64)pages_min * zone_managed_pages(zone);
5850 do_div(tmp, lowmem_pages);
5851 if (is_highmem(zone) || zone_idx(zone) == ZONE_MOVABLE) {
5852 /*
5853 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5854 * need highmem and movable zones pages, so cap pages_min
5855 * to a small value here.
5856 *
5857 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5858 * deltas control async page reclaim, and so should
5859 * not be capped for highmem and movable zones.
5860 */
5861 unsigned long min_pages;
5862
5863 min_pages = zone_managed_pages(zone) / 1024;
5864 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5865 zone->_watermark[WMARK_MIN] = min_pages;
5866 } else {
5867 /*
5868 * If it's a lowmem zone, reserve a number of pages
5869 * proportionate to the zone's size.
5870 */
5871 zone->_watermark[WMARK_MIN] = tmp;
5872 }
5873
5874 /*
5875 * Set the kswapd watermarks distance according to the
5876 * scale factor in proportion to available memory, but
5877 * ensure a minimum size on small systems.
5878 */
5879 tmp = max_t(u64, tmp >> 2,
5880 mult_frac(zone_managed_pages(zone),
5881 watermark_scale_factor, 10000));
5882
5883 zone->watermark_boost = 0;
5884 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
5885 zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
5886 zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
5887
5888 spin_unlock_irqrestore(&zone->lock, flags);
5889 }
5890
5891 /* update totalreserve_pages */
5892 calculate_totalreserve_pages();
5893}
5894
5895/**
5896 * setup_per_zone_wmarks - called when min_free_kbytes changes
5897 * or when memory is hot-{added|removed}
5898 *
5899 * Ensures that the watermark[min,low,high] values for each zone are set
5900 * correctly with respect to min_free_kbytes.
5901 */
5902void setup_per_zone_wmarks(void)
5903{
5904 struct zone *zone;
5905 static DEFINE_SPINLOCK(lock);
5906
5907 spin_lock(&lock);
5908 __setup_per_zone_wmarks();
5909 spin_unlock(&lock);
5910
5911 /*
5912 * The watermark size have changed so update the pcpu batch
5913 * and high limits or the limits may be inappropriate.
5914 */
5915 for_each_zone(zone)
5916 zone_pcp_update(zone, 0);
5917}
5918
5919/*
5920 * Initialise min_free_kbytes.
5921 *
5922 * For small machines we want it small (128k min). For large machines
5923 * we want it large (256MB max). But it is not linear, because network
5924 * bandwidth does not increase linearly with machine size. We use
5925 *
5926 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5927 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5928 *
5929 * which yields
5930 *
5931 * 16MB: 512k
5932 * 32MB: 724k
5933 * 64MB: 1024k
5934 * 128MB: 1448k
5935 * 256MB: 2048k
5936 * 512MB: 2896k
5937 * 1024MB: 4096k
5938 * 2048MB: 5792k
5939 * 4096MB: 8192k
5940 * 8192MB: 11584k
5941 * 16384MB: 16384k
5942 */
5943void calculate_min_free_kbytes(void)
5944{
5945 unsigned long lowmem_kbytes;
5946 int new_min_free_kbytes;
5947
5948 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5949 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5950
5951 if (new_min_free_kbytes > user_min_free_kbytes)
5952 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
5953 else
5954 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5955 new_min_free_kbytes, user_min_free_kbytes);
5956
5957}
5958
5959int __meminit init_per_zone_wmark_min(void)
5960{
5961 calculate_min_free_kbytes();
5962 setup_per_zone_wmarks();
5963 refresh_zone_stat_thresholds();
5964 setup_per_zone_lowmem_reserve();
5965
5966#ifdef CONFIG_NUMA
5967 setup_min_unmapped_ratio();
5968 setup_min_slab_ratio();
5969#endif
5970
5971 khugepaged_min_free_kbytes_update();
5972
5973 return 0;
5974}
5975postcore_initcall(init_per_zone_wmark_min)
5976
5977/*
5978 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5979 * that we can call two helper functions whenever min_free_kbytes
5980 * changes.
5981 */
5982static int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
5983 void *buffer, size_t *length, loff_t *ppos)
5984{
5985 int rc;
5986
5987 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5988 if (rc)
5989 return rc;
5990
5991 if (write) {
5992 user_min_free_kbytes = min_free_kbytes;
5993 setup_per_zone_wmarks();
5994 }
5995 return 0;
5996}
5997
5998static int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
5999 void *buffer, size_t *length, loff_t *ppos)
6000{
6001 int rc;
6002
6003 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6004 if (rc)
6005 return rc;
6006
6007 if (write)
6008 setup_per_zone_wmarks();
6009
6010 return 0;
6011}
6012
6013#ifdef CONFIG_NUMA
6014static void setup_min_unmapped_ratio(void)
6015{
6016 pg_data_t *pgdat;
6017 struct zone *zone;
6018
6019 for_each_online_pgdat(pgdat)
6020 pgdat->min_unmapped_pages = 0;
6021
6022 for_each_zone(zone)
6023 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
6024 sysctl_min_unmapped_ratio) / 100;
6025}
6026
6027
6028static int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6029 void *buffer, size_t *length, loff_t *ppos)
6030{
6031 int rc;
6032
6033 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6034 if (rc)
6035 return rc;
6036
6037 setup_min_unmapped_ratio();
6038
6039 return 0;
6040}
6041
6042static void setup_min_slab_ratio(void)
6043{
6044 pg_data_t *pgdat;
6045 struct zone *zone;
6046
6047 for_each_online_pgdat(pgdat)
6048 pgdat->min_slab_pages = 0;
6049
6050 for_each_zone(zone)
6051 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
6052 sysctl_min_slab_ratio) / 100;
6053}
6054
6055static int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6056 void *buffer, size_t *length, loff_t *ppos)
6057{
6058 int rc;
6059
6060 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6061 if (rc)
6062 return rc;
6063
6064 setup_min_slab_ratio();
6065
6066 return 0;
6067}
6068#endif
6069
6070/*
6071 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6072 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6073 * whenever sysctl_lowmem_reserve_ratio changes.
6074 *
6075 * The reserve ratio obviously has absolutely no relation with the
6076 * minimum watermarks. The lowmem reserve ratio can only make sense
6077 * if in function of the boot time zone sizes.
6078 */
6079static int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table,
6080 int write, void *buffer, size_t *length, loff_t *ppos)
6081{
6082 int i;
6083
6084 proc_dointvec_minmax(table, write, buffer, length, ppos);
6085
6086 for (i = 0; i < MAX_NR_ZONES; i++) {
6087 if (sysctl_lowmem_reserve_ratio[i] < 1)
6088 sysctl_lowmem_reserve_ratio[i] = 0;
6089 }
6090
6091 setup_per_zone_lowmem_reserve();
6092 return 0;
6093}
6094
6095/*
6096 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
6097 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6098 * pagelist can have before it gets flushed back to buddy allocator.
6099 */
6100static int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
6101 int write, void *buffer, size_t *length, loff_t *ppos)
6102{
6103 struct zone *zone;
6104 int old_percpu_pagelist_high_fraction;
6105 int ret;
6106
6107 mutex_lock(&pcp_batch_high_lock);
6108 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
6109
6110 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6111 if (!write || ret < 0)
6112 goto out;
6113
6114 /* Sanity checking to avoid pcp imbalance */
6115 if (percpu_pagelist_high_fraction &&
6116 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
6117 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
6118 ret = -EINVAL;
6119 goto out;
6120 }
6121
6122 /* No change? */
6123 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
6124 goto out;
6125
6126 for_each_populated_zone(zone)
6127 zone_set_pageset_high_and_batch(zone, 0);
6128out:
6129 mutex_unlock(&pcp_batch_high_lock);
6130 return ret;
6131}
6132
6133static struct ctl_table page_alloc_sysctl_table[] = {
6134 {
6135 .procname = "min_free_kbytes",
6136 .data = &min_free_kbytes,
6137 .maxlen = sizeof(min_free_kbytes),
6138 .mode = 0644,
6139 .proc_handler = min_free_kbytes_sysctl_handler,
6140 .extra1 = SYSCTL_ZERO,
6141 },
6142 {
6143 .procname = "watermark_boost_factor",
6144 .data = &watermark_boost_factor,
6145 .maxlen = sizeof(watermark_boost_factor),
6146 .mode = 0644,
6147 .proc_handler = proc_dointvec_minmax,
6148 .extra1 = SYSCTL_ZERO,
6149 },
6150 {
6151 .procname = "watermark_scale_factor",
6152 .data = &watermark_scale_factor,
6153 .maxlen = sizeof(watermark_scale_factor),
6154 .mode = 0644,
6155 .proc_handler = watermark_scale_factor_sysctl_handler,
6156 .extra1 = SYSCTL_ONE,
6157 .extra2 = SYSCTL_THREE_THOUSAND,
6158 },
6159 {
6160 .procname = "percpu_pagelist_high_fraction",
6161 .data = &percpu_pagelist_high_fraction,
6162 .maxlen = sizeof(percpu_pagelist_high_fraction),
6163 .mode = 0644,
6164 .proc_handler = percpu_pagelist_high_fraction_sysctl_handler,
6165 .extra1 = SYSCTL_ZERO,
6166 },
6167 {
6168 .procname = "lowmem_reserve_ratio",
6169 .data = &sysctl_lowmem_reserve_ratio,
6170 .maxlen = sizeof(sysctl_lowmem_reserve_ratio),
6171 .mode = 0644,
6172 .proc_handler = lowmem_reserve_ratio_sysctl_handler,
6173 },
6174#ifdef CONFIG_NUMA
6175 {
6176 .procname = "numa_zonelist_order",
6177 .data = &numa_zonelist_order,
6178 .maxlen = NUMA_ZONELIST_ORDER_LEN,
6179 .mode = 0644,
6180 .proc_handler = numa_zonelist_order_handler,
6181 },
6182 {
6183 .procname = "min_unmapped_ratio",
6184 .data = &sysctl_min_unmapped_ratio,
6185 .maxlen = sizeof(sysctl_min_unmapped_ratio),
6186 .mode = 0644,
6187 .proc_handler = sysctl_min_unmapped_ratio_sysctl_handler,
6188 .extra1 = SYSCTL_ZERO,
6189 .extra2 = SYSCTL_ONE_HUNDRED,
6190 },
6191 {
6192 .procname = "min_slab_ratio",
6193 .data = &sysctl_min_slab_ratio,
6194 .maxlen = sizeof(sysctl_min_slab_ratio),
6195 .mode = 0644,
6196 .proc_handler = sysctl_min_slab_ratio_sysctl_handler,
6197 .extra1 = SYSCTL_ZERO,
6198 .extra2 = SYSCTL_ONE_HUNDRED,
6199 },
6200#endif
6201 {}
6202};
6203
6204void __init page_alloc_sysctl_init(void)
6205{
6206 register_sysctl_init("vm", page_alloc_sysctl_table);
6207}
6208
6209#ifdef CONFIG_CONTIG_ALLOC
6210/* Usage: See admin-guide/dynamic-debug-howto.rst */
6211static void alloc_contig_dump_pages(struct list_head *page_list)
6212{
6213 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
6214
6215 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
6216 struct page *page;
6217
6218 dump_stack();
6219 list_for_each_entry(page, page_list, lru)
6220 dump_page(page, "migration failure");
6221 }
6222}
6223
6224/* [start, end) must belong to a single zone. */
6225int __alloc_contig_migrate_range(struct compact_control *cc,
6226 unsigned long start, unsigned long end)
6227{
6228 /* This function is based on compact_zone() from compaction.c. */
6229 unsigned int nr_reclaimed;
6230 unsigned long pfn = start;
6231 unsigned int tries = 0;
6232 int ret = 0;
6233 struct migration_target_control mtc = {
6234 .nid = zone_to_nid(cc->zone),
6235 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
6236 };
6237
6238 lru_cache_disable();
6239
6240 while (pfn < end || !list_empty(&cc->migratepages)) {
6241 if (fatal_signal_pending(current)) {
6242 ret = -EINTR;
6243 break;
6244 }
6245
6246 if (list_empty(&cc->migratepages)) {
6247 cc->nr_migratepages = 0;
6248 ret = isolate_migratepages_range(cc, pfn, end);
6249 if (ret && ret != -EAGAIN)
6250 break;
6251 pfn = cc->migrate_pfn;
6252 tries = 0;
6253 } else if (++tries == 5) {
6254 ret = -EBUSY;
6255 break;
6256 }
6257
6258 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6259 &cc->migratepages);
6260 cc->nr_migratepages -= nr_reclaimed;
6261
6262 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
6263 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
6264
6265 /*
6266 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
6267 * to retry again over this error, so do the same here.
6268 */
6269 if (ret == -ENOMEM)
6270 break;
6271 }
6272
6273 lru_cache_enable();
6274 if (ret < 0) {
6275 if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY)
6276 alloc_contig_dump_pages(&cc->migratepages);
6277 putback_movable_pages(&cc->migratepages);
6278 return ret;
6279 }
6280 return 0;
6281}
6282
6283/**
6284 * alloc_contig_range() -- tries to allocate given range of pages
6285 * @start: start PFN to allocate
6286 * @end: one-past-the-last PFN to allocate
6287 * @migratetype: migratetype of the underlying pageblocks (either
6288 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6289 * in range must have the same migratetype and it must
6290 * be either of the two.
6291 * @gfp_mask: GFP mask to use during compaction
6292 *
6293 * The PFN range does not have to be pageblock aligned. The PFN range must
6294 * belong to a single zone.
6295 *
6296 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
6297 * pageblocks in the range. Once isolated, the pageblocks should not
6298 * be modified by others.
6299 *
6300 * Return: zero on success or negative error code. On success all
6301 * pages which PFN is in [start, end) are allocated for the caller and
6302 * need to be freed with free_contig_range().
6303 */
6304int alloc_contig_range(unsigned long start, unsigned long end,
6305 unsigned migratetype, gfp_t gfp_mask)
6306{
6307 unsigned long outer_start, outer_end;
6308 int order;
6309 int ret = 0;
6310
6311 struct compact_control cc = {
6312 .nr_migratepages = 0,
6313 .order = -1,
6314 .zone = page_zone(pfn_to_page(start)),
6315 .mode = MIGRATE_SYNC,
6316 .ignore_skip_hint = true,
6317 .no_set_skip_hint = true,
6318 .gfp_mask = current_gfp_context(gfp_mask),
6319 .alloc_contig = true,
6320 };
6321 INIT_LIST_HEAD(&cc.migratepages);
6322
6323 /*
6324 * What we do here is we mark all pageblocks in range as
6325 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6326 * have different sizes, and due to the way page allocator
6327 * work, start_isolate_page_range() has special handlings for this.
6328 *
6329 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6330 * migrate the pages from an unaligned range (ie. pages that
6331 * we are interested in). This will put all the pages in
6332 * range back to page allocator as MIGRATE_ISOLATE.
6333 *
6334 * When this is done, we take the pages in range from page
6335 * allocator removing them from the buddy system. This way
6336 * page allocator will never consider using them.
6337 *
6338 * This lets us mark the pageblocks back as
6339 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6340 * aligned range but not in the unaligned, original range are
6341 * put back to page allocator so that buddy can use them.
6342 */
6343
6344 ret = start_isolate_page_range(start, end, migratetype, 0, gfp_mask);
6345 if (ret)
6346 goto done;
6347
6348 drain_all_pages(cc.zone);
6349
6350 /*
6351 * In case of -EBUSY, we'd like to know which page causes problem.
6352 * So, just fall through. test_pages_isolated() has a tracepoint
6353 * which will report the busy page.
6354 *
6355 * It is possible that busy pages could become available before
6356 * the call to test_pages_isolated, and the range will actually be
6357 * allocated. So, if we fall through be sure to clear ret so that
6358 * -EBUSY is not accidentally used or returned to caller.
6359 */
6360 ret = __alloc_contig_migrate_range(&cc, start, end);
6361 if (ret && ret != -EBUSY)
6362 goto done;
6363 ret = 0;
6364
6365 /*
6366 * Pages from [start, end) are within a pageblock_nr_pages
6367 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6368 * more, all pages in [start, end) are free in page allocator.
6369 * What we are going to do is to allocate all pages from
6370 * [start, end) (that is remove them from page allocator).
6371 *
6372 * The only problem is that pages at the beginning and at the
6373 * end of interesting range may be not aligned with pages that
6374 * page allocator holds, ie. they can be part of higher order
6375 * pages. Because of this, we reserve the bigger range and
6376 * once this is done free the pages we are not interested in.
6377 *
6378 * We don't have to hold zone->lock here because the pages are
6379 * isolated thus they won't get removed from buddy.
6380 */
6381
6382 order = 0;
6383 outer_start = start;
6384 while (!PageBuddy(pfn_to_page(outer_start))) {
6385 if (++order > MAX_PAGE_ORDER) {
6386 outer_start = start;
6387 break;
6388 }
6389 outer_start &= ~0UL << order;
6390 }
6391
6392 if (outer_start != start) {
6393 order = buddy_order(pfn_to_page(outer_start));
6394
6395 /*
6396 * outer_start page could be small order buddy page and
6397 * it doesn't include start page. Adjust outer_start
6398 * in this case to report failed page properly
6399 * on tracepoint in test_pages_isolated()
6400 */
6401 if (outer_start + (1UL << order) <= start)
6402 outer_start = start;
6403 }
6404
6405 /* Make sure the range is really isolated. */
6406 if (test_pages_isolated(outer_start, end, 0)) {
6407 ret = -EBUSY;
6408 goto done;
6409 }
6410
6411 /* Grab isolated pages from freelists. */
6412 outer_end = isolate_freepages_range(&cc, outer_start, end);
6413 if (!outer_end) {
6414 ret = -EBUSY;
6415 goto done;
6416 }
6417
6418 /* Free head and tail (if any) */
6419 if (start != outer_start)
6420 free_contig_range(outer_start, start - outer_start);
6421 if (end != outer_end)
6422 free_contig_range(end, outer_end - end);
6423
6424done:
6425 undo_isolate_page_range(start, end, migratetype);
6426 return ret;
6427}
6428EXPORT_SYMBOL(alloc_contig_range);
6429
6430static int __alloc_contig_pages(unsigned long start_pfn,
6431 unsigned long nr_pages, gfp_t gfp_mask)
6432{
6433 unsigned long end_pfn = start_pfn + nr_pages;
6434
6435 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
6436 gfp_mask);
6437}
6438
6439static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
6440 unsigned long nr_pages)
6441{
6442 unsigned long i, end_pfn = start_pfn + nr_pages;
6443 struct page *page;
6444
6445 for (i = start_pfn; i < end_pfn; i++) {
6446 page = pfn_to_online_page(i);
6447 if (!page)
6448 return false;
6449
6450 if (page_zone(page) != z)
6451 return false;
6452
6453 if (PageReserved(page))
6454 return false;
6455
6456 if (PageHuge(page))
6457 return false;
6458 }
6459 return true;
6460}
6461
6462static bool zone_spans_last_pfn(const struct zone *zone,
6463 unsigned long start_pfn, unsigned long nr_pages)
6464{
6465 unsigned long last_pfn = start_pfn + nr_pages - 1;
6466
6467 return zone_spans_pfn(zone, last_pfn);
6468}
6469
6470/**
6471 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
6472 * @nr_pages: Number of contiguous pages to allocate
6473 * @gfp_mask: GFP mask to limit search and used during compaction
6474 * @nid: Target node
6475 * @nodemask: Mask for other possible nodes
6476 *
6477 * This routine is a wrapper around alloc_contig_range(). It scans over zones
6478 * on an applicable zonelist to find a contiguous pfn range which can then be
6479 * tried for allocation with alloc_contig_range(). This routine is intended
6480 * for allocation requests which can not be fulfilled with the buddy allocator.
6481 *
6482 * The allocated memory is always aligned to a page boundary. If nr_pages is a
6483 * power of two, then allocated range is also guaranteed to be aligned to same
6484 * nr_pages (e.g. 1GB request would be aligned to 1GB).
6485 *
6486 * Allocated pages can be freed with free_contig_range() or by manually calling
6487 * __free_page() on each allocated page.
6488 *
6489 * Return: pointer to contiguous pages on success, or NULL if not successful.
6490 */
6491struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
6492 int nid, nodemask_t *nodemask)
6493{
6494 unsigned long ret, pfn, flags;
6495 struct zonelist *zonelist;
6496 struct zone *zone;
6497 struct zoneref *z;
6498
6499 zonelist = node_zonelist(nid, gfp_mask);
6500 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6501 gfp_zone(gfp_mask), nodemask) {
6502 spin_lock_irqsave(&zone->lock, flags);
6503
6504 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
6505 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
6506 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
6507 /*
6508 * We release the zone lock here because
6509 * alloc_contig_range() will also lock the zone
6510 * at some point. If there's an allocation
6511 * spinning on this lock, it may win the race
6512 * and cause alloc_contig_range() to fail...
6513 */
6514 spin_unlock_irqrestore(&zone->lock, flags);
6515 ret = __alloc_contig_pages(pfn, nr_pages,
6516 gfp_mask);
6517 if (!ret)
6518 return pfn_to_page(pfn);
6519 spin_lock_irqsave(&zone->lock, flags);
6520 }
6521 pfn += nr_pages;
6522 }
6523 spin_unlock_irqrestore(&zone->lock, flags);
6524 }
6525 return NULL;
6526}
6527#endif /* CONFIG_CONTIG_ALLOC */
6528
6529void free_contig_range(unsigned long pfn, unsigned long nr_pages)
6530{
6531 unsigned long count = 0;
6532
6533 for (; nr_pages--; pfn++) {
6534 struct page *page = pfn_to_page(pfn);
6535
6536 count += page_count(page) != 1;
6537 __free_page(page);
6538 }
6539 WARN(count != 0, "%lu pages are still in use!\n", count);
6540}
6541EXPORT_SYMBOL(free_contig_range);
6542
6543/*
6544 * Effectively disable pcplists for the zone by setting the high limit to 0
6545 * and draining all cpus. A concurrent page freeing on another CPU that's about
6546 * to put the page on pcplist will either finish before the drain and the page
6547 * will be drained, or observe the new high limit and skip the pcplist.
6548 *
6549 * Must be paired with a call to zone_pcp_enable().
6550 */
6551void zone_pcp_disable(struct zone *zone)
6552{
6553 mutex_lock(&pcp_batch_high_lock);
6554 __zone_set_pageset_high_and_batch(zone, 0, 0, 1);
6555 __drain_all_pages(zone, true);
6556}
6557
6558void zone_pcp_enable(struct zone *zone)
6559{
6560 __zone_set_pageset_high_and_batch(zone, zone->pageset_high_min,
6561 zone->pageset_high_max, zone->pageset_batch);
6562 mutex_unlock(&pcp_batch_high_lock);
6563}
6564
6565void zone_pcp_reset(struct zone *zone)
6566{
6567 int cpu;
6568 struct per_cpu_zonestat *pzstats;
6569
6570 if (zone->per_cpu_pageset != &boot_pageset) {
6571 for_each_online_cpu(cpu) {
6572 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
6573 drain_zonestat(zone, pzstats);
6574 }
6575 free_percpu(zone->per_cpu_pageset);
6576 zone->per_cpu_pageset = &boot_pageset;
6577 if (zone->per_cpu_zonestats != &boot_zonestats) {
6578 free_percpu(zone->per_cpu_zonestats);
6579 zone->per_cpu_zonestats = &boot_zonestats;
6580 }
6581 }
6582}
6583
6584#ifdef CONFIG_MEMORY_HOTREMOVE
6585/*
6586 * All pages in the range must be in a single zone, must not contain holes,
6587 * must span full sections, and must be isolated before calling this function.
6588 */
6589void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6590{
6591 unsigned long pfn = start_pfn;
6592 struct page *page;
6593 struct zone *zone;
6594 unsigned int order;
6595 unsigned long flags;
6596
6597 offline_mem_sections(pfn, end_pfn);
6598 zone = page_zone(pfn_to_page(pfn));
6599 spin_lock_irqsave(&zone->lock, flags);
6600 while (pfn < end_pfn) {
6601 page = pfn_to_page(pfn);
6602 /*
6603 * The HWPoisoned page may be not in buddy system, and
6604 * page_count() is not 0.
6605 */
6606 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6607 pfn++;
6608 continue;
6609 }
6610 /*
6611 * At this point all remaining PageOffline() pages have a
6612 * reference count of 0 and can simply be skipped.
6613 */
6614 if (PageOffline(page)) {
6615 BUG_ON(page_count(page));
6616 BUG_ON(PageBuddy(page));
6617 pfn++;
6618 continue;
6619 }
6620
6621 BUG_ON(page_count(page));
6622 BUG_ON(!PageBuddy(page));
6623 order = buddy_order(page);
6624 del_page_from_free_list(page, zone, order);
6625 pfn += (1 << order);
6626 }
6627 spin_unlock_irqrestore(&zone->lock, flags);
6628}
6629#endif
6630
6631/*
6632 * This function returns a stable result only if called under zone lock.
6633 */
6634bool is_free_buddy_page(struct page *page)
6635{
6636 unsigned long pfn = page_to_pfn(page);
6637 unsigned int order;
6638
6639 for (order = 0; order < NR_PAGE_ORDERS; order++) {
6640 struct page *page_head = page - (pfn & ((1 << order) - 1));
6641
6642 if (PageBuddy(page_head) &&
6643 buddy_order_unsafe(page_head) >= order)
6644 break;
6645 }
6646
6647 return order <= MAX_PAGE_ORDER;
6648}
6649EXPORT_SYMBOL(is_free_buddy_page);
6650
6651#ifdef CONFIG_MEMORY_FAILURE
6652/*
6653 * Break down a higher-order page in sub-pages, and keep our target out of
6654 * buddy allocator.
6655 */
6656static void break_down_buddy_pages(struct zone *zone, struct page *page,
6657 struct page *target, int low, int high,
6658 int migratetype)
6659{
6660 unsigned long size = 1 << high;
6661 struct page *current_buddy;
6662
6663 while (high > low) {
6664 high--;
6665 size >>= 1;
6666
6667 if (target >= &page[size]) {
6668 current_buddy = page;
6669 page = page + size;
6670 } else {
6671 current_buddy = page + size;
6672 }
6673
6674 if (set_page_guard(zone, current_buddy, high, migratetype))
6675 continue;
6676
6677 add_to_free_list(current_buddy, zone, high, migratetype);
6678 set_buddy_order(current_buddy, high);
6679 }
6680}
6681
6682/*
6683 * Take a page that will be marked as poisoned off the buddy allocator.
6684 */
6685bool take_page_off_buddy(struct page *page)
6686{
6687 struct zone *zone = page_zone(page);
6688 unsigned long pfn = page_to_pfn(page);
6689 unsigned long flags;
6690 unsigned int order;
6691 bool ret = false;
6692
6693 spin_lock_irqsave(&zone->lock, flags);
6694 for (order = 0; order < NR_PAGE_ORDERS; order++) {
6695 struct page *page_head = page - (pfn & ((1 << order) - 1));
6696 int page_order = buddy_order(page_head);
6697
6698 if (PageBuddy(page_head) && page_order >= order) {
6699 unsigned long pfn_head = page_to_pfn(page_head);
6700 int migratetype = get_pfnblock_migratetype(page_head,
6701 pfn_head);
6702
6703 del_page_from_free_list(page_head, zone, page_order);
6704 break_down_buddy_pages(zone, page_head, page, 0,
6705 page_order, migratetype);
6706 SetPageHWPoisonTakenOff(page);
6707 if (!is_migrate_isolate(migratetype))
6708 __mod_zone_freepage_state(zone, -1, migratetype);
6709 ret = true;
6710 break;
6711 }
6712 if (page_count(page_head) > 0)
6713 break;
6714 }
6715 spin_unlock_irqrestore(&zone->lock, flags);
6716 return ret;
6717}
6718
6719/*
6720 * Cancel takeoff done by take_page_off_buddy().
6721 */
6722bool put_page_back_buddy(struct page *page)
6723{
6724 struct zone *zone = page_zone(page);
6725 unsigned long pfn = page_to_pfn(page);
6726 unsigned long flags;
6727 int migratetype = get_pfnblock_migratetype(page, pfn);
6728 bool ret = false;
6729
6730 spin_lock_irqsave(&zone->lock, flags);
6731 if (put_page_testzero(page)) {
6732 ClearPageHWPoisonTakenOff(page);
6733 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
6734 if (TestClearPageHWPoison(page)) {
6735 ret = true;
6736 }
6737 }
6738 spin_unlock_irqrestore(&zone->lock, flags);
6739
6740 return ret;
6741}
6742#endif
6743
6744#ifdef CONFIG_ZONE_DMA
6745bool has_managed_dma(void)
6746{
6747 struct pglist_data *pgdat;
6748
6749 for_each_online_pgdat(pgdat) {
6750 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
6751
6752 if (managed_zone(zone))
6753 return true;
6754 }
6755 return false;
6756}
6757#endif /* CONFIG_ZONE_DMA */
6758
6759#ifdef CONFIG_UNACCEPTED_MEMORY
6760
6761/* Counts number of zones with unaccepted pages. */
6762static DEFINE_STATIC_KEY_FALSE(zones_with_unaccepted_pages);
6763
6764static bool lazy_accept = true;
6765
6766static int __init accept_memory_parse(char *p)
6767{
6768 if (!strcmp(p, "lazy")) {
6769 lazy_accept = true;
6770 return 0;
6771 } else if (!strcmp(p, "eager")) {
6772 lazy_accept = false;
6773 return 0;
6774 } else {
6775 return -EINVAL;
6776 }
6777}
6778early_param("accept_memory", accept_memory_parse);
6779
6780static bool page_contains_unaccepted(struct page *page, unsigned int order)
6781{
6782 phys_addr_t start = page_to_phys(page);
6783 phys_addr_t end = start + (PAGE_SIZE << order);
6784
6785 return range_contains_unaccepted_memory(start, end);
6786}
6787
6788static void accept_page(struct page *page, unsigned int order)
6789{
6790 phys_addr_t start = page_to_phys(page);
6791
6792 accept_memory(start, start + (PAGE_SIZE << order));
6793}
6794
6795static bool try_to_accept_memory_one(struct zone *zone)
6796{
6797 unsigned long flags;
6798 struct page *page;
6799 bool last;
6800
6801 if (list_empty(&zone->unaccepted_pages))
6802 return false;
6803
6804 spin_lock_irqsave(&zone->lock, flags);
6805 page = list_first_entry_or_null(&zone->unaccepted_pages,
6806 struct page, lru);
6807 if (!page) {
6808 spin_unlock_irqrestore(&zone->lock, flags);
6809 return false;
6810 }
6811
6812 list_del(&page->lru);
6813 last = list_empty(&zone->unaccepted_pages);
6814
6815 __mod_zone_freepage_state(zone, -MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
6816 __mod_zone_page_state(zone, NR_UNACCEPTED, -MAX_ORDER_NR_PAGES);
6817 spin_unlock_irqrestore(&zone->lock, flags);
6818
6819 accept_page(page, MAX_PAGE_ORDER);
6820
6821 __free_pages_ok(page, MAX_PAGE_ORDER, FPI_TO_TAIL);
6822
6823 if (last)
6824 static_branch_dec(&zones_with_unaccepted_pages);
6825
6826 return true;
6827}
6828
6829static bool try_to_accept_memory(struct zone *zone, unsigned int order)
6830{
6831 long to_accept;
6832 int ret = false;
6833
6834 /* How much to accept to get to high watermark? */
6835 to_accept = high_wmark_pages(zone) -
6836 (zone_page_state(zone, NR_FREE_PAGES) -
6837 __zone_watermark_unusable_free(zone, order, 0));
6838
6839 /* Accept at least one page */
6840 do {
6841 if (!try_to_accept_memory_one(zone))
6842 break;
6843 ret = true;
6844 to_accept -= MAX_ORDER_NR_PAGES;
6845 } while (to_accept > 0);
6846
6847 return ret;
6848}
6849
6850static inline bool has_unaccepted_memory(void)
6851{
6852 return static_branch_unlikely(&zones_with_unaccepted_pages);
6853}
6854
6855static bool __free_unaccepted(struct page *page)
6856{
6857 struct zone *zone = page_zone(page);
6858 unsigned long flags;
6859 bool first = false;
6860
6861 if (!lazy_accept)
6862 return false;
6863
6864 spin_lock_irqsave(&zone->lock, flags);
6865 first = list_empty(&zone->unaccepted_pages);
6866 list_add_tail(&page->lru, &zone->unaccepted_pages);
6867 __mod_zone_freepage_state(zone, MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
6868 __mod_zone_page_state(zone, NR_UNACCEPTED, MAX_ORDER_NR_PAGES);
6869 spin_unlock_irqrestore(&zone->lock, flags);
6870
6871 if (first)
6872 static_branch_inc(&zones_with_unaccepted_pages);
6873
6874 return true;
6875}
6876
6877#else
6878
6879static bool page_contains_unaccepted(struct page *page, unsigned int order)
6880{
6881 return false;
6882}
6883
6884static void accept_page(struct page *page, unsigned int order)
6885{
6886}
6887
6888static bool try_to_accept_memory(struct zone *zone, unsigned int order)
6889{
6890 return false;
6891}
6892
6893static inline bool has_unaccepted_memory(void)
6894{
6895 return false;
6896}
6897
6898static bool __free_unaccepted(struct page *page)
6899{
6900 BUILD_BUG();
6901 return false;
6902}
6903
6904#endif /* CONFIG_UNACCEPTED_MEMORY */