Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
1// SPDX-License-Identifier: GPL-2.0-only
2/*
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/highmem.h>
20#include <linux/interrupt.h>
21#include <linux/jiffies.h>
22#include <linux/compiler.h>
23#include <linux/kernel.h>
24#include <linux/kasan.h>
25#include <linux/kmsan.h>
26#include <linux/module.h>
27#include <linux/suspend.h>
28#include <linux/ratelimit.h>
29#include <linux/oom.h>
30#include <linux/topology.h>
31#include <linux/sysctl.h>
32#include <linux/cpu.h>
33#include <linux/cpuset.h>
34#include <linux/folio_batch.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 <linux/pgalloc_tag.h>
57#include <asm/div64.h>
58#include "internal.h"
59#include "shuffle.h"
60#include "page_reporting.h"
61
62/* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
63typedef int __bitwise fpi_t;
64
65/* No special request */
66#define FPI_NONE ((__force fpi_t)0)
67
68/*
69 * Skip free page reporting notification for the (possibly merged) page.
70 * This does not hinder free page reporting from grabbing the page,
71 * reporting it and marking it "reported" - it only skips notifying
72 * the free page reporting infrastructure about a newly freed page. For
73 * example, used when temporarily pulling a page from a freelist and
74 * putting it back unmodified.
75 */
76#define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
77
78/*
79 * Place the (possibly merged) page to the tail of the freelist. Will ignore
80 * page shuffling (relevant code - e.g., memory onlining - is expected to
81 * shuffle the whole zone).
82 *
83 * Note: No code should rely on this flag for correctness - it's purely
84 * to allow for optimizations when handing back either fresh pages
85 * (memory onlining) or untouched pages (page isolation, free page
86 * reporting).
87 */
88#define FPI_TO_TAIL ((__force fpi_t)BIT(1))
89
90/* Free the page without taking locks. Rely on trylock only. */
91#define FPI_TRYLOCK ((__force fpi_t)BIT(2))
92
93/* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
94static DEFINE_MUTEX(pcp_batch_high_lock);
95#define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
96
97/*
98 * Locking a pcp requires a PCP lookup followed by a spinlock. To avoid
99 * a migration causing the wrong PCP to be locked and remote memory being
100 * potentially allocated, pin the task to the CPU for the lookup+lock.
101 * preempt_disable is used on !RT because it is faster than migrate_disable.
102 * migrate_disable is used on RT because otherwise RT spinlock usage is
103 * interfered with and a high priority task cannot preempt the allocator.
104 */
105#ifndef CONFIG_PREEMPT_RT
106#define pcpu_task_pin() preempt_disable()
107#define pcpu_task_unpin() preempt_enable()
108#else
109#define pcpu_task_pin() migrate_disable()
110#define pcpu_task_unpin() migrate_enable()
111#endif
112
113/*
114 * A helper to lookup and trylock pcp with embedded spinlock.
115 * The return value should be used with the unlock helper.
116 * NULL return value means the trylock failed.
117 */
118#ifdef CONFIG_SMP
119#define pcp_spin_trylock(ptr) \
120({ \
121 struct per_cpu_pages *_ret; \
122 pcpu_task_pin(); \
123 _ret = this_cpu_ptr(ptr); \
124 if (!spin_trylock(&_ret->lock)) { \
125 pcpu_task_unpin(); \
126 _ret = NULL; \
127 } \
128 _ret; \
129})
130
131#define pcp_spin_unlock(ptr) \
132({ \
133 spin_unlock(&ptr->lock); \
134 pcpu_task_unpin(); \
135})
136
137/*
138 * On CONFIG_SMP=n the UP implementation of spin_trylock() never fails and thus
139 * is not compatible with our locking scheme. However we do not need pcp for
140 * scalability in the first place, so just make all the trylocks fail and take
141 * the slow path unconditionally.
142 */
143#else
144#define pcp_spin_trylock(ptr) \
145 NULL
146
147#define pcp_spin_unlock(ptr) \
148 BUG_ON(1)
149#endif
150
151/*
152 * In some cases we do not need to pin the task to the CPU because we are
153 * already given a specific cpu's pcp pointer.
154 */
155#define pcp_spin_lock_nopin(ptr) \
156 spin_lock(&(ptr)->lock)
157#define pcp_spin_unlock_nopin(ptr) \
158 spin_unlock(&(ptr)->lock)
159
160#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
161DEFINE_PER_CPU(int, numa_node);
162EXPORT_PER_CPU_SYMBOL(numa_node);
163#endif
164
165DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
166
167#ifdef CONFIG_HAVE_MEMORYLESS_NODES
168/*
169 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
170 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
171 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
172 * defined in <linux/topology.h>.
173 */
174DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
175EXPORT_PER_CPU_SYMBOL(_numa_mem_);
176#endif
177
178static DEFINE_MUTEX(pcpu_drain_mutex);
179
180#ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
181volatile unsigned long latent_entropy __latent_entropy;
182EXPORT_SYMBOL(latent_entropy);
183#endif
184
185/*
186 * Array of node states.
187 */
188nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
189 [N_POSSIBLE] = NODE_MASK_ALL,
190 [N_ONLINE] = { { [0] = 1UL } },
191#ifndef CONFIG_NUMA
192 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
193#ifdef CONFIG_HIGHMEM
194 [N_HIGH_MEMORY] = { { [0] = 1UL } },
195#endif
196 [N_MEMORY] = { { [0] = 1UL } },
197 [N_CPU] = { { [0] = 1UL } },
198#endif /* NUMA */
199};
200EXPORT_SYMBOL(node_states);
201
202gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
203
204#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
205unsigned int pageblock_order __read_mostly;
206#endif
207
208static void __free_pages_ok(struct page *page, unsigned int order,
209 fpi_t fpi_flags);
210static void reserve_highatomic_pageblock(struct page *page, int order,
211 struct zone *zone);
212
213/*
214 * results with 256, 32 in the lowmem_reserve sysctl:
215 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
216 * 1G machine -> (16M dma, 784M normal, 224M high)
217 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
218 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
219 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
220 *
221 * TBD: should special case ZONE_DMA32 machines here - in those we normally
222 * don't need any ZONE_NORMAL reservation
223 */
224static int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
225#ifdef CONFIG_ZONE_DMA
226 [ZONE_DMA] = 256,
227#endif
228#ifdef CONFIG_ZONE_DMA32
229 [ZONE_DMA32] = 256,
230#endif
231 [ZONE_NORMAL] = 32,
232#ifdef CONFIG_HIGHMEM
233 [ZONE_HIGHMEM] = 0,
234#endif
235 [ZONE_MOVABLE] = 0,
236};
237
238char * const zone_names[MAX_NR_ZONES] = {
239#ifdef CONFIG_ZONE_DMA
240 "DMA",
241#endif
242#ifdef CONFIG_ZONE_DMA32
243 "DMA32",
244#endif
245 "Normal",
246#ifdef CONFIG_HIGHMEM
247 "HighMem",
248#endif
249 "Movable",
250#ifdef CONFIG_ZONE_DEVICE
251 "Device",
252#endif
253};
254
255const char * const migratetype_names[MIGRATE_TYPES] = {
256 "Unmovable",
257 "Movable",
258 "Reclaimable",
259 "HighAtomic",
260#ifdef CONFIG_CMA
261 "CMA",
262#endif
263#ifdef CONFIG_MEMORY_ISOLATION
264 "Isolate",
265#endif
266};
267
268int min_free_kbytes = 1024;
269int user_min_free_kbytes = -1;
270static int watermark_boost_factor __read_mostly = 15000;
271static int watermark_scale_factor = 10;
272int defrag_mode;
273
274/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
275int movable_zone;
276EXPORT_SYMBOL(movable_zone);
277
278#if MAX_NUMNODES > 1
279unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
280unsigned int nr_online_nodes __read_mostly = 1;
281EXPORT_SYMBOL(nr_node_ids);
282EXPORT_SYMBOL(nr_online_nodes);
283#endif
284
285static bool page_contains_unaccepted(struct page *page, unsigned int order);
286static bool cond_accept_memory(struct zone *zone, unsigned int order,
287 int alloc_flags);
288static bool __free_unaccepted(struct page *page);
289
290int page_group_by_mobility_disabled __read_mostly;
291
292#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
293/*
294 * During boot we initialize deferred pages on-demand, as needed, but once
295 * page_alloc_init_late() has finished, the deferred pages are all initialized,
296 * and we can permanently disable that path.
297 */
298DEFINE_STATIC_KEY_TRUE(deferred_pages);
299
300/*
301 * deferred_grow_zone() is __init, but it is called from
302 * get_page_from_freelist() during early boot until deferred_pages permanently
303 * disables this call. This is why we have refdata wrapper to avoid warning,
304 * and to ensure that the function body gets unloaded.
305 */
306static bool __ref
307_deferred_grow_zone(struct zone *zone, unsigned int order)
308{
309 return deferred_grow_zone(zone, order);
310}
311#else
312static inline bool _deferred_grow_zone(struct zone *zone, unsigned int order)
313{
314 return false;
315}
316#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
317
318/* Return a pointer to the bitmap storing bits affecting a block of pages */
319static inline unsigned long *get_pageblock_bitmap(const struct page *page,
320 unsigned long pfn)
321{
322#ifdef CONFIG_SPARSEMEM
323 return section_to_usemap(__pfn_to_section(pfn));
324#else
325 return page_zone(page)->pageblock_flags;
326#endif /* CONFIG_SPARSEMEM */
327}
328
329static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
330{
331#ifdef CONFIG_SPARSEMEM
332 pfn &= (PAGES_PER_SECTION-1);
333#else
334 pfn = pfn - pageblock_start_pfn(page_zone(page)->zone_start_pfn);
335#endif /* CONFIG_SPARSEMEM */
336 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
337}
338
339static __always_inline bool is_standalone_pb_bit(enum pageblock_bits pb_bit)
340{
341 return pb_bit >= PB_compact_skip && pb_bit < __NR_PAGEBLOCK_BITS;
342}
343
344static __always_inline void
345get_pfnblock_bitmap_bitidx(const struct page *page, unsigned long pfn,
346 unsigned long **bitmap_word, unsigned long *bitidx)
347{
348 unsigned long *bitmap;
349 unsigned long word_bitidx;
350
351#ifdef CONFIG_MEMORY_ISOLATION
352 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 8);
353#else
354 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
355#endif
356 BUILD_BUG_ON(__MIGRATE_TYPE_END > MIGRATETYPE_MASK);
357 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
358
359 bitmap = get_pageblock_bitmap(page, pfn);
360 *bitidx = pfn_to_bitidx(page, pfn);
361 word_bitidx = *bitidx / BITS_PER_LONG;
362 *bitidx &= (BITS_PER_LONG - 1);
363 *bitmap_word = &bitmap[word_bitidx];
364}
365
366
367/**
368 * __get_pfnblock_flags_mask - Return the requested group of flags for
369 * a pageblock_nr_pages block of pages
370 * @page: The page within the block of interest
371 * @pfn: The target page frame number
372 * @mask: mask of bits that the caller is interested in
373 *
374 * Return: pageblock_bits flags
375 */
376static unsigned long __get_pfnblock_flags_mask(const struct page *page,
377 unsigned long pfn,
378 unsigned long mask)
379{
380 unsigned long *bitmap_word;
381 unsigned long bitidx;
382 unsigned long word;
383
384 get_pfnblock_bitmap_bitidx(page, pfn, &bitmap_word, &bitidx);
385 /*
386 * This races, without locks, with set_pfnblock_migratetype(). Ensure
387 * a consistent read of the memory array, so that results, even though
388 * racy, are not corrupted.
389 */
390 word = READ_ONCE(*bitmap_word);
391 return (word >> bitidx) & mask;
392}
393
394/**
395 * get_pfnblock_bit - Check if a standalone bit of a pageblock is set
396 * @page: The page within the block of interest
397 * @pfn: The target page frame number
398 * @pb_bit: pageblock bit to check
399 *
400 * Return: true if the bit is set, otherwise false
401 */
402bool get_pfnblock_bit(const struct page *page, unsigned long pfn,
403 enum pageblock_bits pb_bit)
404{
405 unsigned long *bitmap_word;
406 unsigned long bitidx;
407
408 if (WARN_ON_ONCE(!is_standalone_pb_bit(pb_bit)))
409 return false;
410
411 get_pfnblock_bitmap_bitidx(page, pfn, &bitmap_word, &bitidx);
412
413 return test_bit(bitidx + pb_bit, bitmap_word);
414}
415
416/**
417 * get_pfnblock_migratetype - Return the migratetype of a pageblock
418 * @page: The page within the block of interest
419 * @pfn: The target page frame number
420 *
421 * Return: The migratetype of the pageblock
422 *
423 * Use get_pfnblock_migratetype() if caller already has both @page and @pfn
424 * to save a call to page_to_pfn().
425 */
426__always_inline enum migratetype
427get_pfnblock_migratetype(const struct page *page, unsigned long pfn)
428{
429 unsigned long mask = MIGRATETYPE_AND_ISO_MASK;
430 unsigned long flags;
431
432 flags = __get_pfnblock_flags_mask(page, pfn, mask);
433
434#ifdef CONFIG_MEMORY_ISOLATION
435 if (flags & BIT(PB_migrate_isolate))
436 return MIGRATE_ISOLATE;
437#endif
438 return flags & MIGRATETYPE_MASK;
439}
440
441/**
442 * __set_pfnblock_flags_mask - Set the requested group of flags for
443 * a pageblock_nr_pages block of pages
444 * @page: The page within the block of interest
445 * @pfn: The target page frame number
446 * @flags: The flags to set
447 * @mask: mask of bits that the caller is interested in
448 */
449static void __set_pfnblock_flags_mask(struct page *page, unsigned long pfn,
450 unsigned long flags, unsigned long mask)
451{
452 unsigned long *bitmap_word;
453 unsigned long bitidx;
454 unsigned long word;
455
456 get_pfnblock_bitmap_bitidx(page, pfn, &bitmap_word, &bitidx);
457
458 mask <<= bitidx;
459 flags <<= bitidx;
460
461 word = READ_ONCE(*bitmap_word);
462 do {
463 } while (!try_cmpxchg(bitmap_word, &word, (word & ~mask) | flags));
464}
465
466/**
467 * set_pfnblock_bit - Set a standalone bit of a pageblock
468 * @page: The page within the block of interest
469 * @pfn: The target page frame number
470 * @pb_bit: pageblock bit to set
471 */
472void set_pfnblock_bit(const struct page *page, unsigned long pfn,
473 enum pageblock_bits pb_bit)
474{
475 unsigned long *bitmap_word;
476 unsigned long bitidx;
477
478 if (WARN_ON_ONCE(!is_standalone_pb_bit(pb_bit)))
479 return;
480
481 get_pfnblock_bitmap_bitidx(page, pfn, &bitmap_word, &bitidx);
482
483 set_bit(bitidx + pb_bit, bitmap_word);
484}
485
486/**
487 * clear_pfnblock_bit - Clear a standalone bit of a pageblock
488 * @page: The page within the block of interest
489 * @pfn: The target page frame number
490 * @pb_bit: pageblock bit to clear
491 */
492void clear_pfnblock_bit(const struct page *page, unsigned long pfn,
493 enum pageblock_bits pb_bit)
494{
495 unsigned long *bitmap_word;
496 unsigned long bitidx;
497
498 if (WARN_ON_ONCE(!is_standalone_pb_bit(pb_bit)))
499 return;
500
501 get_pfnblock_bitmap_bitidx(page, pfn, &bitmap_word, &bitidx);
502
503 clear_bit(bitidx + pb_bit, bitmap_word);
504}
505
506/**
507 * set_pageblock_migratetype - Set the migratetype of a pageblock
508 * @page: The page within the block of interest
509 * @migratetype: migratetype to set
510 */
511static void set_pageblock_migratetype(struct page *page,
512 enum migratetype migratetype)
513{
514 if (unlikely(page_group_by_mobility_disabled &&
515 migratetype < MIGRATE_PCPTYPES))
516 migratetype = MIGRATE_UNMOVABLE;
517
518#ifdef CONFIG_MEMORY_ISOLATION
519 if (migratetype == MIGRATE_ISOLATE) {
520 VM_WARN_ONCE(1,
521 "Use set_pageblock_isolate() for pageblock isolation");
522 return;
523 }
524 VM_WARN_ONCE(get_pageblock_isolate(page),
525 "Use clear_pageblock_isolate() to unisolate pageblock");
526 /* MIGRATETYPE_AND_ISO_MASK clears PB_migrate_isolate if it is set */
527#endif
528 __set_pfnblock_flags_mask(page, page_to_pfn(page),
529 (unsigned long)migratetype,
530 MIGRATETYPE_AND_ISO_MASK);
531}
532
533void __meminit init_pageblock_migratetype(struct page *page,
534 enum migratetype migratetype,
535 bool isolate)
536{
537 unsigned long flags;
538
539 if (unlikely(page_group_by_mobility_disabled &&
540 migratetype < MIGRATE_PCPTYPES))
541 migratetype = MIGRATE_UNMOVABLE;
542
543 flags = migratetype;
544
545#ifdef CONFIG_MEMORY_ISOLATION
546 if (migratetype == MIGRATE_ISOLATE) {
547 VM_WARN_ONCE(
548 1,
549 "Set isolate=true to isolate pageblock with a migratetype");
550 return;
551 }
552 if (isolate)
553 flags |= BIT(PB_migrate_isolate);
554#endif
555 __set_pfnblock_flags_mask(page, page_to_pfn(page), flags,
556 MIGRATETYPE_AND_ISO_MASK);
557}
558
559#ifdef CONFIG_DEBUG_VM
560static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
561{
562 int ret;
563 unsigned seq;
564 unsigned long pfn = page_to_pfn(page);
565 unsigned long sp, start_pfn;
566
567 do {
568 seq = zone_span_seqbegin(zone);
569 start_pfn = zone->zone_start_pfn;
570 sp = zone->spanned_pages;
571 ret = !zone_spans_pfn(zone, pfn);
572 } while (zone_span_seqretry(zone, seq));
573
574 if (ret)
575 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
576 pfn, zone_to_nid(zone), zone->name,
577 start_pfn, start_pfn + sp);
578
579 return ret;
580}
581
582/*
583 * Temporary debugging check for pages not lying within a given zone.
584 */
585static bool __maybe_unused bad_range(struct zone *zone, struct page *page)
586{
587 if (page_outside_zone_boundaries(zone, page))
588 return true;
589 if (zone != page_zone(page))
590 return true;
591
592 return false;
593}
594#else
595static inline bool __maybe_unused bad_range(struct zone *zone, struct page *page)
596{
597 return false;
598}
599#endif
600
601static void bad_page(struct page *page, const char *reason)
602{
603 static unsigned long resume;
604 static unsigned long nr_shown;
605 static unsigned long nr_unshown;
606
607 /*
608 * Allow a burst of 60 reports, then keep quiet for that minute;
609 * or allow a steady drip of one report per second.
610 */
611 if (nr_shown == 60) {
612 if (time_before(jiffies, resume)) {
613 nr_unshown++;
614 goto out;
615 }
616 if (nr_unshown) {
617 pr_alert(
618 "BUG: Bad page state: %lu messages suppressed\n",
619 nr_unshown);
620 nr_unshown = 0;
621 }
622 nr_shown = 0;
623 }
624 if (nr_shown++ == 0)
625 resume = jiffies + 60 * HZ;
626
627 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
628 current->comm, page_to_pfn(page));
629 dump_page(page, reason);
630
631 print_modules();
632 dump_stack();
633out:
634 /* Leave bad fields for debug, except PageBuddy could make trouble */
635 if (PageBuddy(page))
636 __ClearPageBuddy(page);
637 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
638}
639
640static inline unsigned int order_to_pindex(int migratetype, int order)
641{
642
643#ifdef CONFIG_TRANSPARENT_HUGEPAGE
644 bool movable;
645 if (order > PAGE_ALLOC_COSTLY_ORDER) {
646 VM_BUG_ON(!is_pmd_order(order));
647
648 movable = migratetype == MIGRATE_MOVABLE;
649
650 return NR_LOWORDER_PCP_LISTS + movable;
651 }
652#else
653 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
654#endif
655
656 return (MIGRATE_PCPTYPES * order) + migratetype;
657}
658
659static inline int pindex_to_order(unsigned int pindex)
660{
661 int order = pindex / MIGRATE_PCPTYPES;
662
663#ifdef CONFIG_TRANSPARENT_HUGEPAGE
664 if (pindex >= NR_LOWORDER_PCP_LISTS)
665 order = HPAGE_PMD_ORDER;
666#else
667 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
668#endif
669
670 return order;
671}
672
673static inline bool pcp_allowed_order(unsigned int order)
674{
675 if (order <= PAGE_ALLOC_COSTLY_ORDER)
676 return true;
677#ifdef CONFIG_TRANSPARENT_HUGEPAGE
678 if (is_pmd_order(order))
679 return true;
680#endif
681 return false;
682}
683
684/*
685 * Higher-order pages are called "compound pages". They are structured thusly:
686 *
687 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
688 *
689 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
690 * in bit 0 of page->compound_info. The rest of bits is pointer to head page.
691 *
692 * The first tail page's ->compound_order holds the order of allocation.
693 * This usage means that zero-order pages may not be compound.
694 */
695
696void prep_compound_page(struct page *page, unsigned int order)
697{
698 int i;
699 int nr_pages = 1 << order;
700
701 __SetPageHead(page);
702 for (i = 1; i < nr_pages; i++)
703 prep_compound_tail(page + i, page, order);
704
705 prep_compound_head(page, order);
706}
707
708static inline void set_buddy_order(struct page *page, unsigned int order)
709{
710 set_page_private(page, order);
711 __SetPageBuddy(page);
712}
713
714#ifdef CONFIG_COMPACTION
715static inline struct capture_control *task_capc(struct zone *zone)
716{
717 struct capture_control *capc = current->capture_control;
718
719 return unlikely(capc) &&
720 !(current->flags & PF_KTHREAD) &&
721 !capc->page &&
722 capc->cc->zone == zone ? capc : NULL;
723}
724
725static inline bool
726compaction_capture(struct capture_control *capc, struct page *page,
727 int order, int migratetype)
728{
729 if (!capc || order != capc->cc->order)
730 return false;
731
732 /* Do not accidentally pollute CMA or isolated regions*/
733 if (is_migrate_cma(migratetype) ||
734 is_migrate_isolate(migratetype))
735 return false;
736
737 /*
738 * Do not let lower order allocations pollute a movable pageblock
739 * unless compaction is also requesting movable pages.
740 * This might let an unmovable request use a reclaimable pageblock
741 * and vice-versa but no more than normal fallback logic which can
742 * have trouble finding a high-order free page.
743 */
744 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE &&
745 capc->cc->migratetype != MIGRATE_MOVABLE)
746 return false;
747
748 if (migratetype != capc->cc->migratetype)
749 trace_mm_page_alloc_extfrag(page, capc->cc->order, order,
750 capc->cc->migratetype, migratetype);
751
752 capc->page = page;
753 return true;
754}
755
756#else
757static inline struct capture_control *task_capc(struct zone *zone)
758{
759 return NULL;
760}
761
762static inline bool
763compaction_capture(struct capture_control *capc, struct page *page,
764 int order, int migratetype)
765{
766 return false;
767}
768#endif /* CONFIG_COMPACTION */
769
770static inline void account_freepages(struct zone *zone, int nr_pages,
771 int migratetype)
772{
773 lockdep_assert_held(&zone->lock);
774
775 if (is_migrate_isolate(migratetype))
776 return;
777
778 __mod_zone_page_state(zone, NR_FREE_PAGES, nr_pages);
779
780 if (is_migrate_cma(migratetype))
781 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, nr_pages);
782 else if (migratetype == MIGRATE_HIGHATOMIC)
783 WRITE_ONCE(zone->nr_free_highatomic,
784 zone->nr_free_highatomic + nr_pages);
785}
786
787/* Used for pages not on another list */
788static inline void __add_to_free_list(struct page *page, struct zone *zone,
789 unsigned int order, int migratetype,
790 bool tail)
791{
792 struct free_area *area = &zone->free_area[order];
793 int nr_pages = 1 << order;
794
795 VM_WARN_ONCE(get_pageblock_migratetype(page) != migratetype,
796 "page type is %d, passed migratetype is %d (nr=%d)\n",
797 get_pageblock_migratetype(page), migratetype, nr_pages);
798
799 if (tail)
800 list_add_tail(&page->buddy_list, &area->free_list[migratetype]);
801 else
802 list_add(&page->buddy_list, &area->free_list[migratetype]);
803 area->nr_free++;
804
805 if (order >= pageblock_order && !is_migrate_isolate(migratetype))
806 __mod_zone_page_state(zone, NR_FREE_PAGES_BLOCKS, nr_pages);
807}
808
809/*
810 * Used for pages which are on another list. Move the pages to the tail
811 * of the list - so the moved pages won't immediately be considered for
812 * allocation again (e.g., optimization for memory onlining).
813 */
814static inline void move_to_free_list(struct page *page, struct zone *zone,
815 unsigned int order, int old_mt, int new_mt)
816{
817 struct free_area *area = &zone->free_area[order];
818 int nr_pages = 1 << order;
819
820 /* Free page moving can fail, so it happens before the type update */
821 VM_WARN_ONCE(get_pageblock_migratetype(page) != old_mt,
822 "page type is %d, passed migratetype is %d (nr=%d)\n",
823 get_pageblock_migratetype(page), old_mt, nr_pages);
824
825 list_move_tail(&page->buddy_list, &area->free_list[new_mt]);
826
827 account_freepages(zone, -nr_pages, old_mt);
828 account_freepages(zone, nr_pages, new_mt);
829
830 if (order >= pageblock_order &&
831 is_migrate_isolate(old_mt) != is_migrate_isolate(new_mt)) {
832 if (!is_migrate_isolate(old_mt))
833 nr_pages = -nr_pages;
834 __mod_zone_page_state(zone, NR_FREE_PAGES_BLOCKS, nr_pages);
835 }
836}
837
838static inline void __del_page_from_free_list(struct page *page, struct zone *zone,
839 unsigned int order, int migratetype)
840{
841 int nr_pages = 1 << order;
842
843 VM_WARN_ONCE(get_pageblock_migratetype(page) != migratetype,
844 "page type is %d, passed migratetype is %d (nr=%d)\n",
845 get_pageblock_migratetype(page), migratetype, nr_pages);
846
847 /* clear reported state and update reported page count */
848 if (page_reported(page))
849 __ClearPageReported(page);
850
851 list_del(&page->buddy_list);
852 __ClearPageBuddy(page);
853 set_page_private(page, 0);
854 zone->free_area[order].nr_free--;
855
856 if (order >= pageblock_order && !is_migrate_isolate(migratetype))
857 __mod_zone_page_state(zone, NR_FREE_PAGES_BLOCKS, -nr_pages);
858}
859
860static inline void del_page_from_free_list(struct page *page, struct zone *zone,
861 unsigned int order, int migratetype)
862{
863 __del_page_from_free_list(page, zone, order, migratetype);
864 account_freepages(zone, -(1 << order), migratetype);
865}
866
867static inline struct page *get_page_from_free_area(struct free_area *area,
868 int migratetype)
869{
870 return list_first_entry_or_null(&area->free_list[migratetype],
871 struct page, buddy_list);
872}
873
874/*
875 * If this is less than the 2nd largest possible page, check if the buddy
876 * of the next-higher order is free. If it is, it's possible
877 * that pages are being freed that will coalesce soon. In case,
878 * that is happening, add the free page to the tail of the list
879 * so it's less likely to be used soon and more likely to be merged
880 * as a 2-level higher order page
881 */
882static inline bool
883buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
884 struct page *page, unsigned int order)
885{
886 unsigned long higher_page_pfn;
887 struct page *higher_page;
888
889 if (order >= MAX_PAGE_ORDER - 1)
890 return false;
891
892 higher_page_pfn = buddy_pfn & pfn;
893 higher_page = page + (higher_page_pfn - pfn);
894
895 return find_buddy_page_pfn(higher_page, higher_page_pfn, order + 1,
896 NULL) != NULL;
897}
898
899static void change_pageblock_range(struct page *pageblock_page,
900 int start_order, int migratetype)
901{
902 int nr_pageblocks = 1 << (start_order - pageblock_order);
903
904 while (nr_pageblocks--) {
905 set_pageblock_migratetype(pageblock_page, migratetype);
906 pageblock_page += pageblock_nr_pages;
907 }
908}
909
910/*
911 * Freeing function for a buddy system allocator.
912 *
913 * The concept of a buddy system is to maintain direct-mapped table
914 * (containing bit values) for memory blocks of various "orders".
915 * The bottom level table contains the map for the smallest allocatable
916 * units of memory (here, pages), and each level above it describes
917 * pairs of units from the levels below, hence, "buddies".
918 * At a high level, all that happens here is marking the table entry
919 * at the bottom level available, and propagating the changes upward
920 * as necessary, plus some accounting needed to play nicely with other
921 * parts of the VM system.
922 * At each level, we keep a list of pages, which are heads of continuous
923 * free pages of length of (1 << order) and marked with PageBuddy.
924 * Page's order is recorded in page_private(page) field.
925 * So when we are allocating or freeing one, we can derive the state of the
926 * other. That is, if we allocate a small block, and both were
927 * free, the remainder of the region must be split into blocks.
928 * If a block is freed, and its buddy is also free, then this
929 * triggers coalescing into a block of larger size.
930 *
931 * -- nyc
932 */
933
934static inline void __free_one_page(struct page *page,
935 unsigned long pfn,
936 struct zone *zone, unsigned int order,
937 int migratetype, fpi_t fpi_flags)
938{
939 struct capture_control *capc = task_capc(zone);
940 unsigned long buddy_pfn = 0;
941 unsigned long combined_pfn;
942 struct page *buddy;
943 bool to_tail;
944
945 VM_BUG_ON(!zone_is_initialized(zone));
946 VM_BUG_ON_PAGE(page->flags.f & PAGE_FLAGS_CHECK_AT_PREP, page);
947
948 VM_BUG_ON(migratetype == -1);
949 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
950 VM_BUG_ON_PAGE(bad_range(zone, page), page);
951
952 account_freepages(zone, 1 << order, migratetype);
953
954 while (order < MAX_PAGE_ORDER) {
955 int buddy_mt = migratetype;
956
957 if (compaction_capture(capc, page, order, migratetype)) {
958 account_freepages(zone, -(1 << order), migratetype);
959 return;
960 }
961
962 buddy = find_buddy_page_pfn(page, pfn, order, &buddy_pfn);
963 if (!buddy)
964 goto done_merging;
965
966 if (unlikely(order >= pageblock_order)) {
967 /*
968 * We want to prevent merge between freepages on pageblock
969 * without fallbacks and normal pageblock. Without this,
970 * pageblock isolation could cause incorrect freepage or CMA
971 * accounting or HIGHATOMIC accounting.
972 */
973 buddy_mt = get_pfnblock_migratetype(buddy, buddy_pfn);
974
975 if (migratetype != buddy_mt &&
976 (!migratetype_is_mergeable(migratetype) ||
977 !migratetype_is_mergeable(buddy_mt)))
978 goto done_merging;
979 }
980
981 /*
982 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
983 * merge with it and move up one order.
984 */
985 if (page_is_guard(buddy))
986 clear_page_guard(zone, buddy, order);
987 else
988 __del_page_from_free_list(buddy, zone, order, buddy_mt);
989
990 if (unlikely(buddy_mt != migratetype)) {
991 /*
992 * Match buddy type. This ensures that an
993 * expand() down the line puts the sub-blocks
994 * on the right freelists.
995 */
996 change_pageblock_range(buddy, order, migratetype);
997 }
998
999 combined_pfn = buddy_pfn & pfn;
1000 page = page + (combined_pfn - pfn);
1001 pfn = combined_pfn;
1002 order++;
1003 }
1004
1005done_merging:
1006 set_buddy_order(page, order);
1007
1008 if (fpi_flags & FPI_TO_TAIL)
1009 to_tail = true;
1010 else if (is_shuffle_order(order))
1011 to_tail = shuffle_pick_tail();
1012 else
1013 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1014
1015 __add_to_free_list(page, zone, order, migratetype, to_tail);
1016
1017 /* Notify page reporting subsystem of freed page */
1018 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1019 page_reporting_notify_free(order);
1020}
1021
1022/*
1023 * A bad page could be due to a number of fields. Instead of multiple branches,
1024 * try and check multiple fields with one check. The caller must do a detailed
1025 * check if necessary.
1026 */
1027static inline bool page_expected_state(struct page *page,
1028 unsigned long check_flags)
1029{
1030 if (unlikely(atomic_read(&page->_mapcount) != -1))
1031 return false;
1032
1033 if (unlikely((unsigned long)page->mapping |
1034 page_ref_count(page) |
1035#ifdef CONFIG_MEMCG
1036 page->memcg_data |
1037#endif
1038 (page->flags.f & check_flags)))
1039 return false;
1040
1041 return true;
1042}
1043
1044static const char *page_bad_reason(struct page *page, unsigned long flags)
1045{
1046 const char *bad_reason = NULL;
1047
1048 if (unlikely(atomic_read(&page->_mapcount) != -1))
1049 bad_reason = "nonzero mapcount";
1050 if (unlikely(page->mapping != NULL))
1051 bad_reason = "non-NULL mapping";
1052 if (unlikely(page_ref_count(page) != 0))
1053 bad_reason = "nonzero _refcount";
1054 if (unlikely(page->flags.f & flags)) {
1055 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1056 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1057 else
1058 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1059 }
1060#ifdef CONFIG_MEMCG
1061 if (unlikely(page->memcg_data))
1062 bad_reason = "page still charged to cgroup";
1063#endif
1064 return bad_reason;
1065}
1066
1067static inline bool free_page_is_bad(struct page *page)
1068{
1069 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1070 return false;
1071
1072 /* Something has gone sideways, find it */
1073 bad_page(page, page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1074 return true;
1075}
1076
1077static inline bool is_check_pages_enabled(void)
1078{
1079 return static_branch_unlikely(&check_pages_enabled);
1080}
1081
1082static int free_tail_page_prepare(struct page *head_page, struct page *page)
1083{
1084 struct folio *folio = (struct folio *)head_page;
1085 int ret = 1;
1086
1087 /*
1088 * We rely page->lru.next never has bit 0 set, unless the page
1089 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1090 */
1091 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1092
1093 if (!is_check_pages_enabled()) {
1094 ret = 0;
1095 goto out;
1096 }
1097 switch (page - head_page) {
1098 case 1:
1099 /* the first tail page: these may be in place of ->mapping */
1100 if (unlikely(folio_large_mapcount(folio))) {
1101 bad_page(page, "nonzero large_mapcount");
1102 goto out;
1103 }
1104 if (IS_ENABLED(CONFIG_PAGE_MAPCOUNT) &&
1105 unlikely(atomic_read(&folio->_nr_pages_mapped))) {
1106 bad_page(page, "nonzero nr_pages_mapped");
1107 goto out;
1108 }
1109 if (IS_ENABLED(CONFIG_MM_ID)) {
1110 if (unlikely(folio->_mm_id_mapcount[0] != -1)) {
1111 bad_page(page, "nonzero mm mapcount 0");
1112 goto out;
1113 }
1114 if (unlikely(folio->_mm_id_mapcount[1] != -1)) {
1115 bad_page(page, "nonzero mm mapcount 1");
1116 goto out;
1117 }
1118 }
1119 if (IS_ENABLED(CONFIG_64BIT)) {
1120 if (unlikely(atomic_read(&folio->_entire_mapcount) + 1)) {
1121 bad_page(page, "nonzero entire_mapcount");
1122 goto out;
1123 }
1124 if (unlikely(atomic_read(&folio->_pincount))) {
1125 bad_page(page, "nonzero pincount");
1126 goto out;
1127 }
1128 }
1129 break;
1130 case 2:
1131 /* the second tail page: deferred_list overlaps ->mapping */
1132 if (unlikely(!list_empty(&folio->_deferred_list))) {
1133 bad_page(page, "on deferred list");
1134 goto out;
1135 }
1136 if (!IS_ENABLED(CONFIG_64BIT)) {
1137 if (unlikely(atomic_read(&folio->_entire_mapcount) + 1)) {
1138 bad_page(page, "nonzero entire_mapcount");
1139 goto out;
1140 }
1141 if (unlikely(atomic_read(&folio->_pincount))) {
1142 bad_page(page, "nonzero pincount");
1143 goto out;
1144 }
1145 }
1146 break;
1147 case 3:
1148 /* the third tail page: hugetlb specifics overlap ->mappings */
1149 if (IS_ENABLED(CONFIG_HUGETLB_PAGE))
1150 break;
1151 fallthrough;
1152 default:
1153 if (page->mapping != TAIL_MAPPING) {
1154 bad_page(page, "corrupted mapping in tail page");
1155 goto out;
1156 }
1157 break;
1158 }
1159 if (unlikely(!PageTail(page))) {
1160 bad_page(page, "PageTail not set");
1161 goto out;
1162 }
1163 if (unlikely(compound_head(page) != head_page)) {
1164 bad_page(page, "compound_head not consistent");
1165 goto out;
1166 }
1167 ret = 0;
1168out:
1169 page->mapping = NULL;
1170 clear_compound_head(page);
1171 return ret;
1172}
1173
1174/*
1175 * Skip KASAN memory poisoning when either:
1176 *
1177 * 1. For generic KASAN: deferred memory initialization has not yet completed.
1178 * Tag-based KASAN modes skip pages freed via deferred memory initialization
1179 * using page tags instead (see below).
1180 * 2. For tag-based KASAN modes: the page has a match-all KASAN tag, indicating
1181 * that error detection is disabled for accesses via the page address.
1182 *
1183 * Pages will have match-all tags in the following circumstances:
1184 *
1185 * 1. Pages are being initialized for the first time, including during deferred
1186 * memory init; see the call to page_kasan_tag_reset in __init_single_page.
1187 * 2. The allocation was not unpoisoned due to __GFP_SKIP_KASAN, with the
1188 * exception of pages unpoisoned by kasan_unpoison_vmalloc.
1189 * 3. The allocation was excluded from being checked due to sampling,
1190 * see the call to kasan_unpoison_pages.
1191 *
1192 * Poisoning pages during deferred memory init will greatly lengthen the
1193 * process and cause problem in large memory systems as the deferred pages
1194 * initialization is done with interrupt disabled.
1195 *
1196 * Assuming that there will be no reference to those newly initialized
1197 * pages before they are ever allocated, this should have no effect on
1198 * KASAN memory tracking as the poison will be properly inserted at page
1199 * allocation time. The only corner case is when pages are allocated by
1200 * on-demand allocation and then freed again before the deferred pages
1201 * initialization is done, but this is not likely to happen.
1202 */
1203static inline bool should_skip_kasan_poison(struct page *page)
1204{
1205 if (IS_ENABLED(CONFIG_KASAN_GENERIC))
1206 return deferred_pages_enabled();
1207
1208 return page_kasan_tag(page) == KASAN_TAG_KERNEL;
1209}
1210
1211static void kernel_init_pages(struct page *page, int numpages)
1212{
1213 int i;
1214
1215 /* s390's use of memset() could override KASAN redzones. */
1216 kasan_disable_current();
1217 for (i = 0; i < numpages; i++)
1218 clear_highpage_kasan_tagged(page + i);
1219 kasan_enable_current();
1220}
1221
1222#ifdef CONFIG_MEM_ALLOC_PROFILING
1223
1224/* Should be called only if mem_alloc_profiling_enabled() */
1225void __clear_page_tag_ref(struct page *page)
1226{
1227 union pgtag_ref_handle handle;
1228 union codetag_ref ref;
1229
1230 if (get_page_tag_ref(page, &ref, &handle)) {
1231 set_codetag_empty(&ref);
1232 update_page_tag_ref(handle, &ref);
1233 put_page_tag_ref(handle);
1234 }
1235}
1236
1237/* Should be called only if mem_alloc_profiling_enabled() */
1238static noinline
1239void __pgalloc_tag_add(struct page *page, struct task_struct *task,
1240 unsigned int nr)
1241{
1242 union pgtag_ref_handle handle;
1243 union codetag_ref ref;
1244
1245 if (get_page_tag_ref(page, &ref, &handle)) {
1246 alloc_tag_add(&ref, task->alloc_tag, PAGE_SIZE * nr);
1247 update_page_tag_ref(handle, &ref);
1248 put_page_tag_ref(handle);
1249 }
1250}
1251
1252static inline void pgalloc_tag_add(struct page *page, struct task_struct *task,
1253 unsigned int nr)
1254{
1255 if (mem_alloc_profiling_enabled())
1256 __pgalloc_tag_add(page, task, nr);
1257}
1258
1259/* Should be called only if mem_alloc_profiling_enabled() */
1260static noinline
1261void __pgalloc_tag_sub(struct page *page, unsigned int nr)
1262{
1263 union pgtag_ref_handle handle;
1264 union codetag_ref ref;
1265
1266 if (get_page_tag_ref(page, &ref, &handle)) {
1267 alloc_tag_sub(&ref, PAGE_SIZE * nr);
1268 update_page_tag_ref(handle, &ref);
1269 put_page_tag_ref(handle);
1270 }
1271}
1272
1273static inline void pgalloc_tag_sub(struct page *page, unsigned int nr)
1274{
1275 if (mem_alloc_profiling_enabled())
1276 __pgalloc_tag_sub(page, nr);
1277}
1278
1279/* When tag is not NULL, assuming mem_alloc_profiling_enabled */
1280static inline void pgalloc_tag_sub_pages(struct alloc_tag *tag, unsigned int nr)
1281{
1282 if (tag)
1283 this_cpu_sub(tag->counters->bytes, PAGE_SIZE * nr);
1284}
1285
1286#else /* CONFIG_MEM_ALLOC_PROFILING */
1287
1288static inline void pgalloc_tag_add(struct page *page, struct task_struct *task,
1289 unsigned int nr) {}
1290static inline void pgalloc_tag_sub(struct page *page, unsigned int nr) {}
1291static inline void pgalloc_tag_sub_pages(struct alloc_tag *tag, unsigned int nr) {}
1292
1293#endif /* CONFIG_MEM_ALLOC_PROFILING */
1294
1295__always_inline bool __free_pages_prepare(struct page *page,
1296 unsigned int order, fpi_t fpi_flags)
1297{
1298 int bad = 0;
1299 bool skip_kasan_poison = should_skip_kasan_poison(page);
1300 bool init = want_init_on_free();
1301 bool compound = PageCompound(page);
1302 struct folio *folio = page_folio(page);
1303
1304 VM_BUG_ON_PAGE(PageTail(page), page);
1305
1306 trace_mm_page_free(page, order);
1307 kmsan_free_page(page, order);
1308
1309 if (memcg_kmem_online() && PageMemcgKmem(page))
1310 __memcg_kmem_uncharge_page(page, order);
1311
1312 /*
1313 * In rare cases, when truncation or holepunching raced with
1314 * munlock after VM_LOCKED was cleared, Mlocked may still be
1315 * found set here. This does not indicate a problem, unless
1316 * "unevictable_pgs_cleared" appears worryingly large.
1317 */
1318 if (unlikely(folio_test_mlocked(folio))) {
1319 long nr_pages = folio_nr_pages(folio);
1320
1321 __folio_clear_mlocked(folio);
1322 zone_stat_mod_folio(folio, NR_MLOCK, -nr_pages);
1323 count_vm_events(UNEVICTABLE_PGCLEARED, nr_pages);
1324 }
1325
1326 if (unlikely(PageHWPoison(page)) && !order) {
1327 /* Do not let hwpoison pages hit pcplists/buddy */
1328 reset_page_owner(page, order);
1329 page_table_check_free(page, order);
1330 pgalloc_tag_sub(page, 1 << order);
1331
1332 /*
1333 * The page is isolated and accounted for.
1334 * Mark the codetag as empty to avoid accounting error
1335 * when the page is freed by unpoison_memory().
1336 */
1337 clear_page_tag_ref(page);
1338 return false;
1339 }
1340
1341 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1342
1343 /*
1344 * Check tail pages before head page information is cleared to
1345 * avoid checking PageCompound for order-0 pages.
1346 */
1347 if (unlikely(order)) {
1348 int i;
1349
1350 if (compound) {
1351 page[1].flags.f &= ~PAGE_FLAGS_SECOND;
1352#ifdef NR_PAGES_IN_LARGE_FOLIO
1353 folio->_nr_pages = 0;
1354#endif
1355 }
1356 for (i = 1; i < (1 << order); i++) {
1357 if (compound)
1358 bad += free_tail_page_prepare(page, page + i);
1359 if (is_check_pages_enabled()) {
1360 if (free_page_is_bad(page + i)) {
1361 bad++;
1362 continue;
1363 }
1364 }
1365 (page + i)->flags.f &= ~PAGE_FLAGS_CHECK_AT_PREP;
1366 }
1367 }
1368 if (folio_test_anon(folio)) {
1369 mod_mthp_stat(order, MTHP_STAT_NR_ANON, -1);
1370 folio->mapping = NULL;
1371 }
1372 if (unlikely(page_has_type(page))) {
1373 /* networking expects to clear its page type before releasing */
1374 if (is_check_pages_enabled()) {
1375 if (unlikely(PageNetpp(page))) {
1376 bad_page(page, "page_pool leak");
1377 return false;
1378 }
1379 }
1380 /* Reset the page_type (which overlays _mapcount) */
1381 page->page_type = UINT_MAX;
1382 }
1383
1384 if (is_check_pages_enabled()) {
1385 if (free_page_is_bad(page))
1386 bad++;
1387 if (bad)
1388 return false;
1389 }
1390
1391 page_cpupid_reset_last(page);
1392 page->flags.f &= ~PAGE_FLAGS_CHECK_AT_PREP;
1393 page->private = 0;
1394 reset_page_owner(page, order);
1395 page_table_check_free(page, order);
1396 pgalloc_tag_sub(page, 1 << order);
1397
1398 if (!PageHighMem(page) && !(fpi_flags & FPI_TRYLOCK)) {
1399 debug_check_no_locks_freed(page_address(page),
1400 PAGE_SIZE << order);
1401 debug_check_no_obj_freed(page_address(page),
1402 PAGE_SIZE << order);
1403 }
1404
1405 kernel_poison_pages(page, 1 << order);
1406
1407 /*
1408 * As memory initialization might be integrated into KASAN,
1409 * KASAN poisoning and memory initialization code must be
1410 * kept together to avoid discrepancies in behavior.
1411 *
1412 * With hardware tag-based KASAN, memory tags must be set before the
1413 * page becomes unavailable via debug_pagealloc or arch_free_page.
1414 */
1415 if (!skip_kasan_poison) {
1416 kasan_poison_pages(page, order, init);
1417
1418 /* Memory is already initialized if KASAN did it internally. */
1419 if (kasan_has_integrated_init())
1420 init = false;
1421 }
1422 if (init)
1423 kernel_init_pages(page, 1 << order);
1424
1425 /*
1426 * arch_free_page() can make the page's contents inaccessible. s390
1427 * does this. So nothing which can access the page's contents should
1428 * happen after this.
1429 */
1430 arch_free_page(page, order);
1431
1432 debug_pagealloc_unmap_pages(page, 1 << order);
1433
1434 return true;
1435}
1436
1437bool free_pages_prepare(struct page *page, unsigned int order)
1438{
1439 return __free_pages_prepare(page, order, FPI_NONE);
1440}
1441
1442/*
1443 * Frees a number of pages from the PCP lists
1444 * Assumes all pages on list are in same zone.
1445 * count is the number of pages to free.
1446 */
1447static void free_pcppages_bulk(struct zone *zone, int count,
1448 struct per_cpu_pages *pcp,
1449 int pindex)
1450{
1451 unsigned long flags;
1452 unsigned int order;
1453 struct page *page;
1454
1455 /*
1456 * Ensure proper count is passed which otherwise would stuck in the
1457 * below while (list_empty(list)) loop.
1458 */
1459 count = min(pcp->count, count);
1460
1461 /* Ensure requested pindex is drained first. */
1462 pindex = pindex - 1;
1463
1464 spin_lock_irqsave(&zone->lock, flags);
1465
1466 while (count > 0) {
1467 struct list_head *list;
1468 int nr_pages;
1469
1470 /* Remove pages from lists in a round-robin fashion. */
1471 do {
1472 if (++pindex > NR_PCP_LISTS - 1)
1473 pindex = 0;
1474 list = &pcp->lists[pindex];
1475 } while (list_empty(list));
1476
1477 order = pindex_to_order(pindex);
1478 nr_pages = 1 << order;
1479 do {
1480 unsigned long pfn;
1481 int mt;
1482
1483 page = list_last_entry(list, struct page, pcp_list);
1484 pfn = page_to_pfn(page);
1485 mt = get_pfnblock_migratetype(page, pfn);
1486
1487 /* must delete to avoid corrupting pcp list */
1488 list_del(&page->pcp_list);
1489 count -= nr_pages;
1490 pcp->count -= nr_pages;
1491
1492 __free_one_page(page, pfn, zone, order, mt, FPI_NONE);
1493 trace_mm_page_pcpu_drain(page, order, mt);
1494 } while (count > 0 && !list_empty(list));
1495 }
1496
1497 spin_unlock_irqrestore(&zone->lock, flags);
1498}
1499
1500/* Split a multi-block free page into its individual pageblocks. */
1501static void split_large_buddy(struct zone *zone, struct page *page,
1502 unsigned long pfn, int order, fpi_t fpi)
1503{
1504 unsigned long end = pfn + (1 << order);
1505
1506 VM_WARN_ON_ONCE(!IS_ALIGNED(pfn, 1 << order));
1507 /* Caller removed page from freelist, buddy info cleared! */
1508 VM_WARN_ON_ONCE(PageBuddy(page));
1509
1510 if (order > pageblock_order)
1511 order = pageblock_order;
1512
1513 do {
1514 int mt = get_pfnblock_migratetype(page, pfn);
1515
1516 __free_one_page(page, pfn, zone, order, mt, fpi);
1517 pfn += 1 << order;
1518 if (pfn == end)
1519 break;
1520 page = pfn_to_page(pfn);
1521 } while (1);
1522}
1523
1524static void add_page_to_zone_llist(struct zone *zone, struct page *page,
1525 unsigned int order)
1526{
1527 /* Remember the order */
1528 page->private = order;
1529 /* Add the page to the free list */
1530 llist_add(&page->pcp_llist, &zone->trylock_free_pages);
1531}
1532
1533static void free_one_page(struct zone *zone, struct page *page,
1534 unsigned long pfn, unsigned int order,
1535 fpi_t fpi_flags)
1536{
1537 struct llist_head *llhead;
1538 unsigned long flags;
1539
1540 if (unlikely(fpi_flags & FPI_TRYLOCK)) {
1541 if (!spin_trylock_irqsave(&zone->lock, flags)) {
1542 add_page_to_zone_llist(zone, page, order);
1543 return;
1544 }
1545 } else {
1546 spin_lock_irqsave(&zone->lock, flags);
1547 }
1548
1549 /* The lock succeeded. Process deferred pages. */
1550 llhead = &zone->trylock_free_pages;
1551 if (unlikely(!llist_empty(llhead) && !(fpi_flags & FPI_TRYLOCK))) {
1552 struct llist_node *llnode;
1553 struct page *p, *tmp;
1554
1555 llnode = llist_del_all(llhead);
1556 llist_for_each_entry_safe(p, tmp, llnode, pcp_llist) {
1557 unsigned int p_order = p->private;
1558
1559 split_large_buddy(zone, p, page_to_pfn(p), p_order, fpi_flags);
1560 __count_vm_events(PGFREE, 1 << p_order);
1561 }
1562 }
1563 split_large_buddy(zone, page, pfn, order, fpi_flags);
1564 spin_unlock_irqrestore(&zone->lock, flags);
1565
1566 __count_vm_events(PGFREE, 1 << order);
1567}
1568
1569static void __free_pages_ok(struct page *page, unsigned int order,
1570 fpi_t fpi_flags)
1571{
1572 unsigned long pfn = page_to_pfn(page);
1573 struct zone *zone = page_zone(page);
1574
1575 if (__free_pages_prepare(page, order, fpi_flags))
1576 free_one_page(zone, page, pfn, order, fpi_flags);
1577}
1578
1579void __meminit __free_pages_core(struct page *page, unsigned int order,
1580 enum meminit_context context)
1581{
1582 unsigned int nr_pages = 1 << order;
1583 struct page *p = page;
1584 unsigned int loop;
1585
1586 /*
1587 * When initializing the memmap, __init_single_page() sets the refcount
1588 * of all pages to 1 ("allocated"/"not free"). We have to set the
1589 * refcount of all involved pages to 0.
1590 *
1591 * Note that hotplugged memory pages are initialized to PageOffline().
1592 * Pages freed from memblock might be marked as reserved.
1593 */
1594 if (IS_ENABLED(CONFIG_MEMORY_HOTPLUG) &&
1595 unlikely(context == MEMINIT_HOTPLUG)) {
1596 for (loop = 0; loop < nr_pages; loop++, p++) {
1597 VM_WARN_ON_ONCE(PageReserved(p));
1598 __ClearPageOffline(p);
1599 set_page_count(p, 0);
1600 }
1601
1602 adjust_managed_page_count(page, nr_pages);
1603 } else {
1604 for (loop = 0; loop < nr_pages; loop++, p++) {
1605 __ClearPageReserved(p);
1606 set_page_count(p, 0);
1607 }
1608
1609 /* memblock adjusts totalram_pages() manually. */
1610 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1611 }
1612
1613 if (page_contains_unaccepted(page, order)) {
1614 if (order == MAX_PAGE_ORDER && __free_unaccepted(page))
1615 return;
1616
1617 accept_memory(page_to_phys(page), PAGE_SIZE << order);
1618 }
1619
1620 /*
1621 * Bypass PCP and place fresh pages right to the tail, primarily
1622 * relevant for memory onlining.
1623 */
1624 __free_pages_ok(page, order, FPI_TO_TAIL);
1625}
1626
1627/*
1628 * Check that the whole (or subset of) a pageblock given by the interval of
1629 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1630 * with the migration of free compaction scanner.
1631 *
1632 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1633 *
1634 * It's possible on some configurations to have a setup like node0 node1 node0
1635 * i.e. it's possible that all pages within a zones range of pages do not
1636 * belong to a single zone. We assume that a border between node0 and node1
1637 * can occur within a single pageblock, but not a node0 node1 node0
1638 * interleaving within a single pageblock. It is therefore sufficient to check
1639 * the first and last page of a pageblock and avoid checking each individual
1640 * page in a pageblock.
1641 *
1642 * Note: the function may return non-NULL struct page even for a page block
1643 * which contains a memory hole (i.e. there is no physical memory for a subset
1644 * of the pfn range). For example, if the pageblock order is MAX_PAGE_ORDER, which
1645 * will fall into 2 sub-sections, and the end pfn of the pageblock may be hole
1646 * even though the start pfn is online and valid. This should be safe most of
1647 * the time because struct pages are still initialized via init_unavailable_range()
1648 * and pfn walkers shouldn't touch any physical memory range for which they do
1649 * not recognize any specific metadata in struct pages.
1650 */
1651struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1652 unsigned long end_pfn, struct zone *zone)
1653{
1654 struct page *start_page;
1655 struct page *end_page;
1656
1657 /* end_pfn is one past the range we are checking */
1658 end_pfn--;
1659
1660 if (!pfn_valid(end_pfn))
1661 return NULL;
1662
1663 start_page = pfn_to_online_page(start_pfn);
1664 if (!start_page)
1665 return NULL;
1666
1667 if (page_zone(start_page) != zone)
1668 return NULL;
1669
1670 end_page = pfn_to_page(end_pfn);
1671
1672 /* This gives a shorter code than deriving page_zone(end_page) */
1673 if (page_zone_id(start_page) != page_zone_id(end_page))
1674 return NULL;
1675
1676 return start_page;
1677}
1678
1679/*
1680 * The order of subdivision here is critical for the IO subsystem.
1681 * Please do not alter this order without good reasons and regression
1682 * testing. Specifically, as large blocks of memory are subdivided,
1683 * the order in which smaller blocks are delivered depends on the order
1684 * they're subdivided in this function. This is the primary factor
1685 * influencing the order in which pages are delivered to the IO
1686 * subsystem according to empirical testing, and this is also justified
1687 * by considering the behavior of a buddy system containing a single
1688 * large block of memory acted on by a series of small allocations.
1689 * This behavior is a critical factor in sglist merging's success.
1690 *
1691 * -- nyc
1692 */
1693static inline unsigned int expand(struct zone *zone, struct page *page, int low,
1694 int high, int migratetype)
1695{
1696 unsigned int size = 1 << high;
1697 unsigned int nr_added = 0;
1698
1699 while (high > low) {
1700 high--;
1701 size >>= 1;
1702 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1703
1704 /*
1705 * Mark as guard pages (or page), that will allow to
1706 * merge back to allocator when buddy will be freed.
1707 * Corresponding page table entries will not be touched,
1708 * pages will stay not present in virtual address space
1709 */
1710 if (set_page_guard(zone, &page[size], high))
1711 continue;
1712
1713 __add_to_free_list(&page[size], zone, high, migratetype, false);
1714 set_buddy_order(&page[size], high);
1715 nr_added += size;
1716 }
1717
1718 return nr_added;
1719}
1720
1721static __always_inline void page_del_and_expand(struct zone *zone,
1722 struct page *page, int low,
1723 int high, int migratetype)
1724{
1725 int nr_pages = 1 << high;
1726
1727 __del_page_from_free_list(page, zone, high, migratetype);
1728 nr_pages -= expand(zone, page, low, high, migratetype);
1729 account_freepages(zone, -nr_pages, migratetype);
1730}
1731
1732static void check_new_page_bad(struct page *page)
1733{
1734 if (unlikely(PageHWPoison(page))) {
1735 /* Don't complain about hwpoisoned pages */
1736 if (PageBuddy(page))
1737 __ClearPageBuddy(page);
1738 return;
1739 }
1740
1741 bad_page(page,
1742 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
1743}
1744
1745/*
1746 * This page is about to be returned from the page allocator
1747 */
1748static bool check_new_page(struct page *page)
1749{
1750 if (likely(page_expected_state(page,
1751 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1752 return false;
1753
1754 check_new_page_bad(page);
1755 return true;
1756}
1757
1758static inline bool check_new_pages(struct page *page, unsigned int order)
1759{
1760 if (is_check_pages_enabled()) {
1761 for (int i = 0; i < (1 << order); i++) {
1762 struct page *p = page + i;
1763
1764 if (check_new_page(p))
1765 return true;
1766 }
1767 }
1768
1769 return false;
1770}
1771
1772static inline bool should_skip_kasan_unpoison(gfp_t flags)
1773{
1774 /* Don't skip if a software KASAN mode is enabled. */
1775 if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
1776 IS_ENABLED(CONFIG_KASAN_SW_TAGS))
1777 return false;
1778
1779 /* Skip, if hardware tag-based KASAN is not enabled. */
1780 if (!kasan_hw_tags_enabled())
1781 return true;
1782
1783 /*
1784 * With hardware tag-based KASAN enabled, skip if this has been
1785 * requested via __GFP_SKIP_KASAN.
1786 */
1787 return flags & __GFP_SKIP_KASAN;
1788}
1789
1790static inline bool should_skip_init(gfp_t flags)
1791{
1792 /* Don't skip, if hardware tag-based KASAN is not enabled. */
1793 if (!kasan_hw_tags_enabled())
1794 return false;
1795
1796 /* For hardware tag-based KASAN, skip if requested. */
1797 return (flags & __GFP_SKIP_ZERO);
1798}
1799
1800inline void post_alloc_hook(struct page *page, unsigned int order,
1801 gfp_t gfp_flags)
1802{
1803 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) &&
1804 !should_skip_init(gfp_flags);
1805 bool zero_tags = init && (gfp_flags & __GFP_ZEROTAGS);
1806 int i;
1807
1808 set_page_private(page, 0);
1809
1810 arch_alloc_page(page, order);
1811 debug_pagealloc_map_pages(page, 1 << order);
1812
1813 /*
1814 * Page unpoisoning must happen before memory initialization.
1815 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
1816 * allocations and the page unpoisoning code will complain.
1817 */
1818 kernel_unpoison_pages(page, 1 << order);
1819
1820 /*
1821 * As memory initialization might be integrated into KASAN,
1822 * KASAN unpoisoning and memory initialization code must be
1823 * kept together to avoid discrepancies in behavior.
1824 */
1825
1826 /*
1827 * If memory tags should be zeroed
1828 * (which happens only when memory should be initialized as well).
1829 */
1830 if (zero_tags)
1831 init = !tag_clear_highpages(page, 1 << order);
1832
1833 if (!should_skip_kasan_unpoison(gfp_flags) &&
1834 kasan_unpoison_pages(page, order, init)) {
1835 /* Take note that memory was initialized by KASAN. */
1836 if (kasan_has_integrated_init())
1837 init = false;
1838 } else {
1839 /*
1840 * If memory tags have not been set by KASAN, reset the page
1841 * tags to ensure page_address() dereferencing does not fault.
1842 */
1843 for (i = 0; i != 1 << order; ++i)
1844 page_kasan_tag_reset(page + i);
1845 }
1846 /* If memory is still not initialized, initialize it now. */
1847 if (init)
1848 kernel_init_pages(page, 1 << order);
1849
1850 set_page_owner(page, order, gfp_flags);
1851 page_table_check_alloc(page, order);
1852 pgalloc_tag_add(page, current, 1 << order);
1853}
1854
1855static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1856 unsigned int alloc_flags)
1857{
1858 post_alloc_hook(page, order, gfp_flags);
1859
1860 if (order && (gfp_flags & __GFP_COMP))
1861 prep_compound_page(page, order);
1862
1863 /*
1864 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1865 * allocate the page. The expectation is that the caller is taking
1866 * steps that will free more memory. The caller should avoid the page
1867 * being used for !PFMEMALLOC purposes.
1868 */
1869 if (alloc_flags & ALLOC_NO_WATERMARKS)
1870 set_page_pfmemalloc(page);
1871 else
1872 clear_page_pfmemalloc(page);
1873}
1874
1875/*
1876 * Go through the free lists for the given migratetype and remove
1877 * the smallest available page from the freelists
1878 */
1879static __always_inline
1880struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1881 int migratetype)
1882{
1883 unsigned int current_order;
1884 struct free_area *area;
1885 struct page *page;
1886
1887 /* Find a page of the appropriate size in the preferred list */
1888 for (current_order = order; current_order < NR_PAGE_ORDERS; ++current_order) {
1889 area = &(zone->free_area[current_order]);
1890 page = get_page_from_free_area(area, migratetype);
1891 if (!page)
1892 continue;
1893
1894 page_del_and_expand(zone, page, order, current_order,
1895 migratetype);
1896 trace_mm_page_alloc_zone_locked(page, order, migratetype,
1897 pcp_allowed_order(order) &&
1898 migratetype < MIGRATE_PCPTYPES);
1899 return page;
1900 }
1901
1902 return NULL;
1903}
1904
1905
1906/*
1907 * This array describes the order lists are fallen back to when
1908 * the free lists for the desirable migrate type are depleted
1909 *
1910 * The other migratetypes do not have fallbacks.
1911 */
1912static int fallbacks[MIGRATE_PCPTYPES][MIGRATE_PCPTYPES - 1] = {
1913 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE },
1914 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE },
1915 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE },
1916};
1917
1918#ifdef CONFIG_CMA
1919static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1920 unsigned int order)
1921{
1922 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1923}
1924#else
1925static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1926 unsigned int order) { return NULL; }
1927#endif
1928
1929/*
1930 * Move all free pages of a block to new type's freelist. Caller needs to
1931 * change the block type.
1932 */
1933static int __move_freepages_block(struct zone *zone, unsigned long start_pfn,
1934 int old_mt, int new_mt)
1935{
1936 struct page *page;
1937 unsigned long pfn, end_pfn;
1938 unsigned int order;
1939 int pages_moved = 0;
1940
1941 VM_WARN_ON(start_pfn & (pageblock_nr_pages - 1));
1942 end_pfn = pageblock_end_pfn(start_pfn);
1943
1944 for (pfn = start_pfn; pfn < end_pfn;) {
1945 page = pfn_to_page(pfn);
1946 if (!PageBuddy(page)) {
1947 pfn++;
1948 continue;
1949 }
1950
1951 /* Make sure we are not inadvertently changing nodes */
1952 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1953 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
1954
1955 order = buddy_order(page);
1956
1957 move_to_free_list(page, zone, order, old_mt, new_mt);
1958
1959 pfn += 1 << order;
1960 pages_moved += 1 << order;
1961 }
1962
1963 return pages_moved;
1964}
1965
1966static bool prep_move_freepages_block(struct zone *zone, struct page *page,
1967 unsigned long *start_pfn,
1968 int *num_free, int *num_movable)
1969{
1970 unsigned long pfn, start, end;
1971
1972 pfn = page_to_pfn(page);
1973 start = pageblock_start_pfn(pfn);
1974 end = pageblock_end_pfn(pfn);
1975
1976 /*
1977 * The caller only has the lock for @zone, don't touch ranges
1978 * that straddle into other zones. While we could move part of
1979 * the range that's inside the zone, this call is usually
1980 * accompanied by other operations such as migratetype updates
1981 * which also should be locked.
1982 */
1983 if (!zone_spans_pfn(zone, start))
1984 return false;
1985 if (!zone_spans_pfn(zone, end - 1))
1986 return false;
1987
1988 *start_pfn = start;
1989
1990 if (num_free) {
1991 *num_free = 0;
1992 *num_movable = 0;
1993 for (pfn = start; pfn < end;) {
1994 page = pfn_to_page(pfn);
1995 if (PageBuddy(page)) {
1996 int nr = 1 << buddy_order(page);
1997
1998 *num_free += nr;
1999 pfn += nr;
2000 continue;
2001 }
2002 /*
2003 * We assume that pages that could be isolated for
2004 * migration are movable. But we don't actually try
2005 * isolating, as that would be expensive.
2006 */
2007 if (PageLRU(page) || page_has_movable_ops(page))
2008 (*num_movable)++;
2009 pfn++;
2010 }
2011 }
2012
2013 return true;
2014}
2015
2016static int move_freepages_block(struct zone *zone, struct page *page,
2017 int old_mt, int new_mt)
2018{
2019 unsigned long start_pfn;
2020 int res;
2021
2022 if (!prep_move_freepages_block(zone, page, &start_pfn, NULL, NULL))
2023 return -1;
2024
2025 res = __move_freepages_block(zone, start_pfn, old_mt, new_mt);
2026 set_pageblock_migratetype(pfn_to_page(start_pfn), new_mt);
2027
2028 return res;
2029
2030}
2031
2032#ifdef CONFIG_MEMORY_ISOLATION
2033/* Look for a buddy that straddles start_pfn */
2034static unsigned long find_large_buddy(unsigned long start_pfn)
2035{
2036 /*
2037 * If start_pfn is not an order-0 PageBuddy, next PageBuddy containing
2038 * start_pfn has minimal order of __ffs(start_pfn) + 1. Start checking
2039 * the order with __ffs(start_pfn). If start_pfn is order-0 PageBuddy,
2040 * the starting order does not matter.
2041 */
2042 int order = start_pfn ? __ffs(start_pfn) : MAX_PAGE_ORDER;
2043 struct page *page;
2044 unsigned long pfn = start_pfn;
2045
2046 while (!PageBuddy(page = pfn_to_page(pfn))) {
2047 /* Nothing found */
2048 if (++order > MAX_PAGE_ORDER)
2049 return start_pfn;
2050 pfn &= ~0UL << order;
2051 }
2052
2053 /*
2054 * Found a preceding buddy, but does it straddle?
2055 */
2056 if (pfn + (1 << buddy_order(page)) > start_pfn)
2057 return pfn;
2058
2059 /* Nothing found */
2060 return start_pfn;
2061}
2062
2063static inline void toggle_pageblock_isolate(struct page *page, bool isolate)
2064{
2065 if (isolate)
2066 set_pageblock_isolate(page);
2067 else
2068 clear_pageblock_isolate(page);
2069}
2070
2071/**
2072 * __move_freepages_block_isolate - move free pages in block for page isolation
2073 * @zone: the zone
2074 * @page: the pageblock page
2075 * @isolate: to isolate the given pageblock or unisolate it
2076 *
2077 * This is similar to move_freepages_block(), but handles the special
2078 * case encountered in page isolation, where the block of interest
2079 * might be part of a larger buddy spanning multiple pageblocks.
2080 *
2081 * Unlike the regular page allocator path, which moves pages while
2082 * stealing buddies off the freelist, page isolation is interested in
2083 * arbitrary pfn ranges that may have overlapping buddies on both ends.
2084 *
2085 * This function handles that. Straddling buddies are split into
2086 * individual pageblocks. Only the block of interest is moved.
2087 *
2088 * Returns %true if pages could be moved, %false otherwise.
2089 */
2090static bool __move_freepages_block_isolate(struct zone *zone,
2091 struct page *page, bool isolate)
2092{
2093 unsigned long start_pfn, buddy_pfn;
2094 int from_mt;
2095 int to_mt;
2096 struct page *buddy;
2097
2098 if (isolate == get_pageblock_isolate(page)) {
2099 VM_WARN_ONCE(1, "%s a pageblock that is already in that state",
2100 isolate ? "Isolate" : "Unisolate");
2101 return false;
2102 }
2103
2104 if (!prep_move_freepages_block(zone, page, &start_pfn, NULL, NULL))
2105 return false;
2106
2107 /* No splits needed if buddies can't span multiple blocks */
2108 if (pageblock_order == MAX_PAGE_ORDER)
2109 goto move;
2110
2111 buddy_pfn = find_large_buddy(start_pfn);
2112 buddy = pfn_to_page(buddy_pfn);
2113 /* We're a part of a larger buddy */
2114 if (PageBuddy(buddy) && buddy_order(buddy) > pageblock_order) {
2115 int order = buddy_order(buddy);
2116
2117 del_page_from_free_list(buddy, zone, order,
2118 get_pfnblock_migratetype(buddy, buddy_pfn));
2119 toggle_pageblock_isolate(page, isolate);
2120 split_large_buddy(zone, buddy, buddy_pfn, order, FPI_NONE);
2121 return true;
2122 }
2123
2124move:
2125 /* Use MIGRATETYPE_MASK to get non-isolate migratetype */
2126 if (isolate) {
2127 from_mt = __get_pfnblock_flags_mask(page, page_to_pfn(page),
2128 MIGRATETYPE_MASK);
2129 to_mt = MIGRATE_ISOLATE;
2130 } else {
2131 from_mt = MIGRATE_ISOLATE;
2132 to_mt = __get_pfnblock_flags_mask(page, page_to_pfn(page),
2133 MIGRATETYPE_MASK);
2134 }
2135
2136 __move_freepages_block(zone, start_pfn, from_mt, to_mt);
2137 toggle_pageblock_isolate(pfn_to_page(start_pfn), isolate);
2138
2139 return true;
2140}
2141
2142bool pageblock_isolate_and_move_free_pages(struct zone *zone, struct page *page)
2143{
2144 return __move_freepages_block_isolate(zone, page, true);
2145}
2146
2147bool pageblock_unisolate_and_move_free_pages(struct zone *zone, struct page *page)
2148{
2149 return __move_freepages_block_isolate(zone, page, false);
2150}
2151
2152#endif /* CONFIG_MEMORY_ISOLATION */
2153
2154static inline bool boost_watermark(struct zone *zone)
2155{
2156 unsigned long max_boost;
2157
2158 if (!watermark_boost_factor)
2159 return false;
2160 /*
2161 * Don't bother in zones that are unlikely to produce results.
2162 * On small machines, including kdump capture kernels running
2163 * in a small area, boosting the watermark can cause an out of
2164 * memory situation immediately.
2165 */
2166 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2167 return false;
2168
2169 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2170 watermark_boost_factor, 10000);
2171
2172 /*
2173 * high watermark may be uninitialised if fragmentation occurs
2174 * very early in boot so do not boost. We do not fall
2175 * through and boost by pageblock_nr_pages as failing
2176 * allocations that early means that reclaim is not going
2177 * to help and it may even be impossible to reclaim the
2178 * boosted watermark resulting in a hang.
2179 */
2180 if (!max_boost)
2181 return false;
2182
2183 max_boost = max(pageblock_nr_pages, max_boost);
2184
2185 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2186 max_boost);
2187
2188 return true;
2189}
2190
2191/*
2192 * When we are falling back to another migratetype during allocation, should we
2193 * try to claim an entire block to satisfy further allocations, instead of
2194 * polluting multiple pageblocks?
2195 */
2196static bool should_try_claim_block(unsigned int order, int start_mt)
2197{
2198 /*
2199 * Leaving this order check is intended, although there is
2200 * relaxed order check in next check. The reason is that
2201 * we can actually claim the whole pageblock if this condition met,
2202 * but, below check doesn't guarantee it and that is just heuristic
2203 * so could be changed anytime.
2204 */
2205 if (order >= pageblock_order)
2206 return true;
2207
2208 /*
2209 * Above a certain threshold, always try to claim, as it's likely there
2210 * will be more free pages in the pageblock.
2211 */
2212 if (order >= pageblock_order / 2)
2213 return true;
2214
2215 /*
2216 * Unmovable/reclaimable allocations would cause permanent
2217 * fragmentations if they fell back to allocating from a movable block
2218 * (polluting it), so we try to claim the whole block regardless of the
2219 * allocation size. Later movable allocations can always steal from this
2220 * block, which is less problematic.
2221 */
2222 if (start_mt == MIGRATE_RECLAIMABLE || start_mt == MIGRATE_UNMOVABLE)
2223 return true;
2224
2225 if (page_group_by_mobility_disabled)
2226 return true;
2227
2228 /*
2229 * Movable pages won't cause permanent fragmentation, so when you alloc
2230 * small pages, we just need to temporarily steal unmovable or
2231 * reclaimable pages that are closest to the request size. After a
2232 * while, memory compaction may occur to form large contiguous pages,
2233 * and the next movable allocation may not need to steal.
2234 */
2235 return false;
2236}
2237
2238/*
2239 * Check whether there is a suitable fallback freepage with requested order.
2240 * If claimable is true, this function returns fallback_mt only if
2241 * we would do this whole-block claiming. This would help to reduce
2242 * fragmentation due to mixed migratetype pages in one pageblock.
2243 */
2244int find_suitable_fallback(struct free_area *area, unsigned int order,
2245 int migratetype, bool claimable)
2246{
2247 int i;
2248
2249 if (claimable && !should_try_claim_block(order, migratetype))
2250 return -2;
2251
2252 if (area->nr_free == 0)
2253 return -1;
2254
2255 for (i = 0; i < MIGRATE_PCPTYPES - 1 ; i++) {
2256 int fallback_mt = fallbacks[migratetype][i];
2257
2258 if (!free_area_empty(area, fallback_mt))
2259 return fallback_mt;
2260 }
2261
2262 return -1;
2263}
2264
2265/*
2266 * This function implements actual block claiming behaviour. If order is large
2267 * enough, we can claim the whole pageblock for the requested migratetype. If
2268 * not, we check the pageblock for constituent pages; if at least half of the
2269 * pages are free or compatible, we can still claim the whole block, so pages
2270 * freed in the future will be put on the correct free list.
2271 */
2272static struct page *
2273try_to_claim_block(struct zone *zone, struct page *page,
2274 int current_order, int order, int start_type,
2275 int block_type, unsigned int alloc_flags)
2276{
2277 int free_pages, movable_pages, alike_pages;
2278 unsigned long start_pfn;
2279
2280 /* Take ownership for orders >= pageblock_order */
2281 if (current_order >= pageblock_order) {
2282 unsigned int nr_added;
2283
2284 del_page_from_free_list(page, zone, current_order, block_type);
2285 change_pageblock_range(page, current_order, start_type);
2286 nr_added = expand(zone, page, order, current_order, start_type);
2287 account_freepages(zone, nr_added, start_type);
2288 return page;
2289 }
2290
2291 /*
2292 * Boost watermarks to increase reclaim pressure to reduce the
2293 * likelihood of future fallbacks. Wake kswapd now as the node
2294 * may be balanced overall and kswapd will not wake naturally.
2295 */
2296 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2297 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2298
2299 /* moving whole block can fail due to zone boundary conditions */
2300 if (!prep_move_freepages_block(zone, page, &start_pfn, &free_pages,
2301 &movable_pages))
2302 return NULL;
2303
2304 /*
2305 * Determine how many pages are compatible with our allocation.
2306 * For movable allocation, it's the number of movable pages which
2307 * we just obtained. For other types it's a bit more tricky.
2308 */
2309 if (start_type == MIGRATE_MOVABLE) {
2310 alike_pages = movable_pages;
2311 } else {
2312 /*
2313 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2314 * to MOVABLE pageblock, consider all non-movable pages as
2315 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2316 * vice versa, be conservative since we can't distinguish the
2317 * exact migratetype of non-movable pages.
2318 */
2319 if (block_type == MIGRATE_MOVABLE)
2320 alike_pages = pageblock_nr_pages
2321 - (free_pages + movable_pages);
2322 else
2323 alike_pages = 0;
2324 }
2325 /*
2326 * If a sufficient number of pages in the block are either free or of
2327 * compatible migratability as our allocation, claim the whole block.
2328 */
2329 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2330 page_group_by_mobility_disabled) {
2331 __move_freepages_block(zone, start_pfn, block_type, start_type);
2332 set_pageblock_migratetype(pfn_to_page(start_pfn), start_type);
2333 return __rmqueue_smallest(zone, order, start_type);
2334 }
2335
2336 return NULL;
2337}
2338
2339/*
2340 * Try to allocate from some fallback migratetype by claiming the entire block,
2341 * i.e. converting it to the allocation's start migratetype.
2342 *
2343 * The use of signed ints for order and current_order is a deliberate
2344 * deviation from the rest of this file, to make the for loop
2345 * condition simpler.
2346 */
2347static __always_inline struct page *
2348__rmqueue_claim(struct zone *zone, int order, int start_migratetype,
2349 unsigned int alloc_flags)
2350{
2351 struct free_area *area;
2352 int current_order;
2353 int min_order = order;
2354 struct page *page;
2355 int fallback_mt;
2356
2357 /*
2358 * Do not steal pages from freelists belonging to other pageblocks
2359 * i.e. orders < pageblock_order. If there are no local zones free,
2360 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2361 */
2362 if (order < pageblock_order && alloc_flags & ALLOC_NOFRAGMENT)
2363 min_order = pageblock_order;
2364
2365 /*
2366 * Find the largest available free page in the other list. This roughly
2367 * approximates finding the pageblock with the most free pages, which
2368 * would be too costly to do exactly.
2369 */
2370 for (current_order = MAX_PAGE_ORDER; current_order >= min_order;
2371 --current_order) {
2372 area = &(zone->free_area[current_order]);
2373 fallback_mt = find_suitable_fallback(area, current_order,
2374 start_migratetype, true);
2375
2376 /* No block in that order */
2377 if (fallback_mt == -1)
2378 continue;
2379
2380 /* Advanced into orders too low to claim, abort */
2381 if (fallback_mt == -2)
2382 break;
2383
2384 page = get_page_from_free_area(area, fallback_mt);
2385 page = try_to_claim_block(zone, page, current_order, order,
2386 start_migratetype, fallback_mt,
2387 alloc_flags);
2388 if (page) {
2389 trace_mm_page_alloc_extfrag(page, order, current_order,
2390 start_migratetype, fallback_mt);
2391 return page;
2392 }
2393 }
2394
2395 return NULL;
2396}
2397
2398/*
2399 * Try to steal a single page from some fallback migratetype. Leave the rest of
2400 * the block as its current migratetype, potentially causing fragmentation.
2401 */
2402static __always_inline struct page *
2403__rmqueue_steal(struct zone *zone, int order, int start_migratetype)
2404{
2405 struct free_area *area;
2406 int current_order;
2407 struct page *page;
2408 int fallback_mt;
2409
2410 for (current_order = order; current_order < NR_PAGE_ORDERS; current_order++) {
2411 area = &(zone->free_area[current_order]);
2412 fallback_mt = find_suitable_fallback(area, current_order,
2413 start_migratetype, false);
2414 if (fallback_mt == -1)
2415 continue;
2416
2417 page = get_page_from_free_area(area, fallback_mt);
2418 page_del_and_expand(zone, page, order, current_order, fallback_mt);
2419 trace_mm_page_alloc_extfrag(page, order, current_order,
2420 start_migratetype, fallback_mt);
2421 return page;
2422 }
2423
2424 return NULL;
2425}
2426
2427enum rmqueue_mode {
2428 RMQUEUE_NORMAL,
2429 RMQUEUE_CMA,
2430 RMQUEUE_CLAIM,
2431 RMQUEUE_STEAL,
2432};
2433
2434/*
2435 * Do the hard work of removing an element from the buddy allocator.
2436 * Call me with the zone->lock already held.
2437 */
2438static __always_inline struct page *
2439__rmqueue(struct zone *zone, unsigned int order, int migratetype,
2440 unsigned int alloc_flags, enum rmqueue_mode *mode)
2441{
2442 struct page *page;
2443
2444 if (IS_ENABLED(CONFIG_CMA)) {
2445 /*
2446 * Balance movable allocations between regular and CMA areas by
2447 * allocating from CMA when over half of the zone's free memory
2448 * is in the CMA area.
2449 */
2450 if (alloc_flags & ALLOC_CMA &&
2451 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2452 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2453 page = __rmqueue_cma_fallback(zone, order);
2454 if (page)
2455 return page;
2456 }
2457 }
2458
2459 /*
2460 * First try the freelists of the requested migratetype, then try
2461 * fallbacks modes with increasing levels of fragmentation risk.
2462 *
2463 * The fallback logic is expensive and rmqueue_bulk() calls in
2464 * a loop with the zone->lock held, meaning the freelists are
2465 * not subject to any outside changes. Remember in *mode where
2466 * we found pay dirt, to save us the search on the next call.
2467 */
2468 switch (*mode) {
2469 case RMQUEUE_NORMAL:
2470 page = __rmqueue_smallest(zone, order, migratetype);
2471 if (page)
2472 return page;
2473 fallthrough;
2474 case RMQUEUE_CMA:
2475 if (alloc_flags & ALLOC_CMA) {
2476 page = __rmqueue_cma_fallback(zone, order);
2477 if (page) {
2478 *mode = RMQUEUE_CMA;
2479 return page;
2480 }
2481 }
2482 fallthrough;
2483 case RMQUEUE_CLAIM:
2484 page = __rmqueue_claim(zone, order, migratetype, alloc_flags);
2485 if (page) {
2486 /* Replenished preferred freelist, back to normal mode. */
2487 *mode = RMQUEUE_NORMAL;
2488 return page;
2489 }
2490 fallthrough;
2491 case RMQUEUE_STEAL:
2492 if (!(alloc_flags & ALLOC_NOFRAGMENT)) {
2493 page = __rmqueue_steal(zone, order, migratetype);
2494 if (page) {
2495 *mode = RMQUEUE_STEAL;
2496 return page;
2497 }
2498 }
2499 }
2500 return NULL;
2501}
2502
2503/*
2504 * Obtain a specified number of elements from the buddy allocator, all under
2505 * a single hold of the lock, for efficiency. Add them to the supplied list.
2506 * Returns the number of new pages which were placed at *list.
2507 */
2508static int rmqueue_bulk(struct zone *zone, unsigned int order,
2509 unsigned long count, struct list_head *list,
2510 int migratetype, unsigned int alloc_flags)
2511{
2512 enum rmqueue_mode rmqm = RMQUEUE_NORMAL;
2513 unsigned long flags;
2514 int i;
2515
2516 if (unlikely(alloc_flags & ALLOC_TRYLOCK)) {
2517 if (!spin_trylock_irqsave(&zone->lock, flags))
2518 return 0;
2519 } else {
2520 spin_lock_irqsave(&zone->lock, flags);
2521 }
2522 for (i = 0; i < count; ++i) {
2523 struct page *page = __rmqueue(zone, order, migratetype,
2524 alloc_flags, &rmqm);
2525 if (unlikely(page == NULL))
2526 break;
2527
2528 /*
2529 * Split buddy pages returned by expand() are received here in
2530 * physical page order. The page is added to the tail of
2531 * caller's list. From the callers perspective, the linked list
2532 * is ordered by page number under some conditions. This is
2533 * useful for IO devices that can forward direction from the
2534 * head, thus also in the physical page order. This is useful
2535 * for IO devices that can merge IO requests if the physical
2536 * pages are ordered properly.
2537 */
2538 list_add_tail(&page->pcp_list, list);
2539 }
2540 spin_unlock_irqrestore(&zone->lock, flags);
2541
2542 return i;
2543}
2544
2545/*
2546 * Called from the vmstat counter updater to decay the PCP high.
2547 * Return whether there are addition works to do.
2548 */
2549bool decay_pcp_high(struct zone *zone, struct per_cpu_pages *pcp)
2550{
2551 int high_min, to_drain, to_drain_batched, batch;
2552 bool todo = false;
2553
2554 high_min = READ_ONCE(pcp->high_min);
2555 batch = READ_ONCE(pcp->batch);
2556 /*
2557 * Decrease pcp->high periodically to try to free possible
2558 * idle PCP pages. And, avoid to free too many pages to
2559 * control latency. This caps pcp->high decrement too.
2560 */
2561 if (pcp->high > high_min) {
2562 pcp->high = max3(pcp->count - (batch << CONFIG_PCP_BATCH_SCALE_MAX),
2563 pcp->high - (pcp->high >> 3), high_min);
2564 if (pcp->high > high_min)
2565 todo = true;
2566 }
2567
2568 to_drain = pcp->count - pcp->high;
2569 while (to_drain > 0) {
2570 to_drain_batched = min(to_drain, batch);
2571 pcp_spin_lock_nopin(pcp);
2572 free_pcppages_bulk(zone, to_drain_batched, pcp, 0);
2573 pcp_spin_unlock_nopin(pcp);
2574 todo = true;
2575
2576 to_drain -= to_drain_batched;
2577 }
2578
2579 return todo;
2580}
2581
2582#ifdef CONFIG_NUMA
2583/*
2584 * Called from the vmstat counter updater to drain pagesets of this
2585 * currently executing processor on remote nodes after they have
2586 * expired.
2587 */
2588void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2589{
2590 int to_drain, batch;
2591
2592 batch = READ_ONCE(pcp->batch);
2593 to_drain = min(pcp->count, batch);
2594 if (to_drain > 0) {
2595 pcp_spin_lock_nopin(pcp);
2596 free_pcppages_bulk(zone, to_drain, pcp, 0);
2597 pcp_spin_unlock_nopin(pcp);
2598 }
2599}
2600#endif
2601
2602/*
2603 * Drain pcplists of the indicated processor and zone.
2604 */
2605static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2606{
2607 struct per_cpu_pages *pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2608 int count;
2609
2610 do {
2611 pcp_spin_lock_nopin(pcp);
2612 count = pcp->count;
2613 if (count) {
2614 int to_drain = min(count,
2615 pcp->batch << CONFIG_PCP_BATCH_SCALE_MAX);
2616
2617 free_pcppages_bulk(zone, to_drain, pcp, 0);
2618 count -= to_drain;
2619 }
2620 pcp_spin_unlock_nopin(pcp);
2621 } while (count);
2622}
2623
2624/*
2625 * Drain pcplists of all zones on the indicated processor.
2626 */
2627static void drain_pages(unsigned int cpu)
2628{
2629 struct zone *zone;
2630
2631 for_each_populated_zone(zone) {
2632 drain_pages_zone(cpu, zone);
2633 }
2634}
2635
2636/*
2637 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2638 */
2639void drain_local_pages(struct zone *zone)
2640{
2641 int cpu = smp_processor_id();
2642
2643 if (zone)
2644 drain_pages_zone(cpu, zone);
2645 else
2646 drain_pages(cpu);
2647}
2648
2649/*
2650 * The implementation of drain_all_pages(), exposing an extra parameter to
2651 * drain on all cpus.
2652 *
2653 * drain_all_pages() is optimized to only execute on cpus where pcplists are
2654 * not empty. The check for non-emptiness can however race with a free to
2655 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
2656 * that need the guarantee that every CPU has drained can disable the
2657 * optimizing racy check.
2658 */
2659static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
2660{
2661 int cpu;
2662
2663 /*
2664 * Allocate in the BSS so we won't require allocation in
2665 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2666 */
2667 static cpumask_t cpus_with_pcps;
2668
2669 /*
2670 * Do not drain if one is already in progress unless it's specific to
2671 * a zone. Such callers are primarily CMA and memory hotplug and need
2672 * the drain to be complete when the call returns.
2673 */
2674 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2675 if (!zone)
2676 return;
2677 mutex_lock(&pcpu_drain_mutex);
2678 }
2679
2680 /*
2681 * We don't care about racing with CPU hotplug event
2682 * as offline notification will cause the notified
2683 * cpu to drain that CPU pcps and on_each_cpu_mask
2684 * disables preemption as part of its processing
2685 */
2686 for_each_online_cpu(cpu) {
2687 struct per_cpu_pages *pcp;
2688 struct zone *z;
2689 bool has_pcps = false;
2690
2691 if (force_all_cpus) {
2692 /*
2693 * The pcp.count check is racy, some callers need a
2694 * guarantee that no cpu is missed.
2695 */
2696 has_pcps = true;
2697 } else if (zone) {
2698 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2699 if (pcp->count)
2700 has_pcps = true;
2701 } else {
2702 for_each_populated_zone(z) {
2703 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
2704 if (pcp->count) {
2705 has_pcps = true;
2706 break;
2707 }
2708 }
2709 }
2710
2711 if (has_pcps)
2712 cpumask_set_cpu(cpu, &cpus_with_pcps);
2713 else
2714 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2715 }
2716
2717 for_each_cpu(cpu, &cpus_with_pcps) {
2718 if (zone)
2719 drain_pages_zone(cpu, zone);
2720 else
2721 drain_pages(cpu);
2722 }
2723
2724 mutex_unlock(&pcpu_drain_mutex);
2725}
2726
2727/*
2728 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2729 *
2730 * When zone parameter is non-NULL, spill just the single zone's pages.
2731 */
2732void drain_all_pages(struct zone *zone)
2733{
2734 __drain_all_pages(zone, false);
2735}
2736
2737static int nr_pcp_free(struct per_cpu_pages *pcp, int batch, int high, bool free_high)
2738{
2739 int min_nr_free, max_nr_free;
2740
2741 /* Free as much as possible if batch freeing high-order pages. */
2742 if (unlikely(free_high))
2743 return min(pcp->count, batch << CONFIG_PCP_BATCH_SCALE_MAX);
2744
2745 /* Check for PCP disabled or boot pageset */
2746 if (unlikely(high < batch))
2747 return 1;
2748
2749 /* Leave at least pcp->batch pages on the list */
2750 min_nr_free = batch;
2751 max_nr_free = high - batch;
2752
2753 /*
2754 * Increase the batch number to the number of the consecutive
2755 * freed pages to reduce zone lock contention.
2756 */
2757 batch = clamp_t(int, pcp->free_count, min_nr_free, max_nr_free);
2758
2759 return batch;
2760}
2761
2762static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
2763 int batch, bool free_high)
2764{
2765 int high, high_min, high_max;
2766
2767 high_min = READ_ONCE(pcp->high_min);
2768 high_max = READ_ONCE(pcp->high_max);
2769 high = pcp->high = clamp(pcp->high, high_min, high_max);
2770
2771 if (unlikely(!high))
2772 return 0;
2773
2774 if (unlikely(free_high)) {
2775 pcp->high = max(high - (batch << CONFIG_PCP_BATCH_SCALE_MAX),
2776 high_min);
2777 return 0;
2778 }
2779
2780 /*
2781 * If reclaim is active, limit the number of pages that can be
2782 * stored on pcp lists
2783 */
2784 if (test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags)) {
2785 int free_count = max_t(int, pcp->free_count, batch);
2786
2787 pcp->high = max(high - free_count, high_min);
2788 return min(batch << 2, pcp->high);
2789 }
2790
2791 if (high_min == high_max)
2792 return high;
2793
2794 if (test_bit(ZONE_BELOW_HIGH, &zone->flags)) {
2795 int free_count = max_t(int, pcp->free_count, batch);
2796
2797 pcp->high = max(high - free_count, high_min);
2798 high = max(pcp->count, high_min);
2799 } else if (pcp->count >= high) {
2800 int need_high = pcp->free_count + batch;
2801
2802 /* pcp->high should be large enough to hold batch freed pages */
2803 if (pcp->high < need_high)
2804 pcp->high = clamp(need_high, high_min, high_max);
2805 }
2806
2807 return high;
2808}
2809
2810/*
2811 * Tune pcp alloc factor and adjust count & free_count. Free pages to bring the
2812 * pcp's watermarks below high.
2813 *
2814 * May return a freed pcp, if during page freeing the pcp spinlock cannot be
2815 * reacquired. Return true if pcp is locked, false otherwise.
2816 */
2817static bool free_frozen_page_commit(struct zone *zone,
2818 struct per_cpu_pages *pcp, struct page *page, int migratetype,
2819 unsigned int order, fpi_t fpi_flags)
2820{
2821 int high, batch;
2822 int to_free, to_free_batched;
2823 int pindex;
2824 int cpu = smp_processor_id();
2825 int ret = true;
2826 bool free_high = false;
2827
2828 /*
2829 * On freeing, reduce the number of pages that are batch allocated.
2830 * See nr_pcp_alloc() where alloc_factor is increased for subsequent
2831 * allocations.
2832 */
2833 pcp->alloc_factor >>= 1;
2834 __count_vm_events(PGFREE, 1 << order);
2835 pindex = order_to_pindex(migratetype, order);
2836 list_add(&page->pcp_list, &pcp->lists[pindex]);
2837 pcp->count += 1 << order;
2838
2839 batch = READ_ONCE(pcp->batch);
2840 /*
2841 * As high-order pages other than THP's stored on PCP can contribute
2842 * to fragmentation, limit the number stored when PCP is heavily
2843 * freeing without allocation. The remainder after bulk freeing
2844 * stops will be drained from vmstat refresh context.
2845 */
2846 if (order && order <= PAGE_ALLOC_COSTLY_ORDER) {
2847 free_high = (pcp->free_count >= (batch + pcp->high_min / 2) &&
2848 (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) &&
2849 (!(pcp->flags & PCPF_FREE_HIGH_BATCH) ||
2850 pcp->count >= batch));
2851 pcp->flags |= PCPF_PREV_FREE_HIGH_ORDER;
2852 } else if (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) {
2853 pcp->flags &= ~PCPF_PREV_FREE_HIGH_ORDER;
2854 }
2855 if (pcp->free_count < (batch << CONFIG_PCP_BATCH_SCALE_MAX))
2856 pcp->free_count += (1 << order);
2857
2858 if (unlikely(fpi_flags & FPI_TRYLOCK)) {
2859 /*
2860 * Do not attempt to take a zone lock. Let pcp->count get
2861 * over high mark temporarily.
2862 */
2863 return true;
2864 }
2865
2866 high = nr_pcp_high(pcp, zone, batch, free_high);
2867 if (pcp->count < high)
2868 return true;
2869
2870 to_free = nr_pcp_free(pcp, batch, high, free_high);
2871 while (to_free > 0 && pcp->count > 0) {
2872 to_free_batched = min(to_free, batch);
2873 free_pcppages_bulk(zone, to_free_batched, pcp, pindex);
2874 to_free -= to_free_batched;
2875
2876 if (to_free == 0 || pcp->count == 0)
2877 break;
2878
2879 pcp_spin_unlock(pcp);
2880
2881 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2882 if (!pcp) {
2883 ret = false;
2884 break;
2885 }
2886
2887 /*
2888 * Check if this thread has been migrated to a different CPU.
2889 * If that is the case, give up and indicate that the pcp is
2890 * returned in an unlocked state.
2891 */
2892 if (smp_processor_id() != cpu) {
2893 pcp_spin_unlock(pcp);
2894 ret = false;
2895 break;
2896 }
2897 }
2898
2899 if (test_bit(ZONE_BELOW_HIGH, &zone->flags) &&
2900 zone_watermark_ok(zone, 0, high_wmark_pages(zone),
2901 ZONE_MOVABLE, 0)) {
2902 struct pglist_data *pgdat = zone->zone_pgdat;
2903 clear_bit(ZONE_BELOW_HIGH, &zone->flags);
2904
2905 /*
2906 * Assume that memory pressure on this node is gone and may be
2907 * in a reclaimable state. If a memory fallback node exists,
2908 * direct reclaim may not have been triggered, causing a
2909 * 'hopeless node' to stay in that state for a while. Let
2910 * kswapd work again by resetting kswapd_failures.
2911 */
2912 if (kswapd_test_hopeless(pgdat) &&
2913 next_memory_node(pgdat->node_id) < MAX_NUMNODES)
2914 kswapd_clear_hopeless(pgdat, KSWAPD_CLEAR_HOPELESS_PCP);
2915 }
2916 return ret;
2917}
2918
2919/*
2920 * Free a pcp page
2921 */
2922static void __free_frozen_pages(struct page *page, unsigned int order,
2923 fpi_t fpi_flags)
2924{
2925 struct per_cpu_pages *pcp;
2926 struct zone *zone;
2927 unsigned long pfn = page_to_pfn(page);
2928 int migratetype;
2929
2930 if (!pcp_allowed_order(order)) {
2931 __free_pages_ok(page, order, fpi_flags);
2932 return;
2933 }
2934
2935 if (!__free_pages_prepare(page, order, fpi_flags))
2936 return;
2937
2938 /*
2939 * We only track unmovable, reclaimable and movable on pcp lists.
2940 * Place ISOLATE pages on the isolated list because they are being
2941 * offlined but treat HIGHATOMIC and CMA as movable pages so we can
2942 * get those areas back if necessary. Otherwise, we may have to free
2943 * excessively into the page allocator
2944 */
2945 zone = page_zone(page);
2946 migratetype = get_pfnblock_migratetype(page, pfn);
2947 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
2948 if (unlikely(is_migrate_isolate(migratetype))) {
2949 free_one_page(zone, page, pfn, order, fpi_flags);
2950 return;
2951 }
2952 migratetype = MIGRATE_MOVABLE;
2953 }
2954
2955 if (unlikely((fpi_flags & FPI_TRYLOCK) && IS_ENABLED(CONFIG_PREEMPT_RT)
2956 && (in_nmi() || in_hardirq()))) {
2957 add_page_to_zone_llist(zone, page, order);
2958 return;
2959 }
2960 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2961 if (pcp) {
2962 if (!free_frozen_page_commit(zone, pcp, page, migratetype,
2963 order, fpi_flags))
2964 return;
2965 pcp_spin_unlock(pcp);
2966 } else {
2967 free_one_page(zone, page, pfn, order, fpi_flags);
2968 }
2969}
2970
2971void free_frozen_pages(struct page *page, unsigned int order)
2972{
2973 __free_frozen_pages(page, order, FPI_NONE);
2974}
2975
2976void free_frozen_pages_nolock(struct page *page, unsigned int order)
2977{
2978 __free_frozen_pages(page, order, FPI_TRYLOCK);
2979}
2980
2981/*
2982 * Free a batch of folios
2983 */
2984void free_unref_folios(struct folio_batch *folios)
2985{
2986 struct per_cpu_pages *pcp = NULL;
2987 struct zone *locked_zone = NULL;
2988 int i, j;
2989
2990 /* Prepare folios for freeing */
2991 for (i = 0, j = 0; i < folios->nr; i++) {
2992 struct folio *folio = folios->folios[i];
2993 unsigned long pfn = folio_pfn(folio);
2994 unsigned int order = folio_order(folio);
2995
2996 if (!__free_pages_prepare(&folio->page, order, FPI_NONE))
2997 continue;
2998 /*
2999 * Free orders not handled on the PCP directly to the
3000 * allocator.
3001 */
3002 if (!pcp_allowed_order(order)) {
3003 free_one_page(folio_zone(folio), &folio->page,
3004 pfn, order, FPI_NONE);
3005 continue;
3006 }
3007 folio->private = (void *)(unsigned long)order;
3008 if (j != i)
3009 folios->folios[j] = folio;
3010 j++;
3011 }
3012 folios->nr = j;
3013
3014 for (i = 0; i < folios->nr; i++) {
3015 struct folio *folio = folios->folios[i];
3016 struct zone *zone = folio_zone(folio);
3017 unsigned long pfn = folio_pfn(folio);
3018 unsigned int order = (unsigned long)folio->private;
3019 int migratetype;
3020
3021 folio->private = NULL;
3022 migratetype = get_pfnblock_migratetype(&folio->page, pfn);
3023
3024 /* Different zone requires a different pcp lock */
3025 if (zone != locked_zone ||
3026 is_migrate_isolate(migratetype)) {
3027 if (pcp) {
3028 pcp_spin_unlock(pcp);
3029 locked_zone = NULL;
3030 pcp = NULL;
3031 }
3032
3033 /*
3034 * Free isolated pages directly to the
3035 * allocator, see comment in free_frozen_pages.
3036 */
3037 if (is_migrate_isolate(migratetype)) {
3038 free_one_page(zone, &folio->page, pfn,
3039 order, FPI_NONE);
3040 continue;
3041 }
3042
3043 /*
3044 * trylock is necessary as folios may be getting freed
3045 * from IRQ or SoftIRQ context after an IO completion.
3046 */
3047 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
3048 if (unlikely(!pcp)) {
3049 free_one_page(zone, &folio->page, pfn,
3050 order, FPI_NONE);
3051 continue;
3052 }
3053 locked_zone = zone;
3054 }
3055
3056 /*
3057 * Non-isolated types over MIGRATE_PCPTYPES get added
3058 * to the MIGRATE_MOVABLE pcp list.
3059 */
3060 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
3061 migratetype = MIGRATE_MOVABLE;
3062
3063 trace_mm_page_free_batched(&folio->page);
3064 if (!free_frozen_page_commit(zone, pcp, &folio->page,
3065 migratetype, order, FPI_NONE)) {
3066 pcp = NULL;
3067 locked_zone = NULL;
3068 }
3069 }
3070
3071 if (pcp)
3072 pcp_spin_unlock(pcp);
3073 folio_batch_reinit(folios);
3074}
3075
3076static void __split_page(struct page *page, unsigned int order)
3077{
3078 VM_WARN_ON_PAGE(PageCompound(page), page);
3079
3080 split_page_owner(page, order, 0);
3081 pgalloc_tag_split(page_folio(page), order, 0);
3082 split_page_memcg(page, order);
3083}
3084
3085/*
3086 * split_page takes a non-compound higher-order page, and splits it into
3087 * n (1<<order) sub-pages: page[0..n]
3088 * Each sub-page must be freed individually.
3089 *
3090 * Note: this is probably too low level an operation for use in drivers.
3091 * Please consult with lkml before using this in your driver.
3092 */
3093void split_page(struct page *page, unsigned int order)
3094{
3095 int i;
3096
3097 VM_WARN_ON_PAGE(!page_count(page), page);
3098
3099 for (i = 1; i < (1 << order); i++)
3100 set_page_refcounted(page + i);
3101
3102 __split_page(page, order);
3103}
3104EXPORT_SYMBOL_GPL(split_page);
3105
3106int __isolate_free_page(struct page *page, unsigned int order)
3107{
3108 struct zone *zone = page_zone(page);
3109 int mt = get_pageblock_migratetype(page);
3110
3111 if (!is_migrate_isolate(mt)) {
3112 unsigned long watermark;
3113 /*
3114 * Obey watermarks as if the page was being allocated. We can
3115 * emulate a high-order watermark check with a raised order-0
3116 * watermark, because we already know our high-order page
3117 * exists.
3118 */
3119 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3120 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3121 return 0;
3122 }
3123
3124 del_page_from_free_list(page, zone, order, mt);
3125
3126 /*
3127 * Set the pageblock if the isolated page is at least half of a
3128 * pageblock
3129 */
3130 if (order >= pageblock_order - 1) {
3131 struct page *endpage = page + (1 << order) - 1;
3132 for (; page < endpage; page += pageblock_nr_pages) {
3133 int mt = get_pageblock_migratetype(page);
3134 /*
3135 * Only change normal pageblocks (i.e., they can merge
3136 * with others)
3137 */
3138 if (migratetype_is_mergeable(mt))
3139 move_freepages_block(zone, page, mt,
3140 MIGRATE_MOVABLE);
3141 }
3142 }
3143
3144 return 1UL << order;
3145}
3146
3147/**
3148 * __putback_isolated_page - Return a now-isolated page back where we got it
3149 * @page: Page that was isolated
3150 * @order: Order of the isolated page
3151 * @mt: The page's pageblock's migratetype
3152 *
3153 * This function is meant to return a page pulled from the free lists via
3154 * __isolate_free_page back to the free lists they were pulled from.
3155 */
3156void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3157{
3158 struct zone *zone = page_zone(page);
3159
3160 /* zone lock should be held when this function is called */
3161 lockdep_assert_held(&zone->lock);
3162
3163 /* Return isolated page to tail of freelist. */
3164 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3165 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3166}
3167
3168/*
3169 * Update NUMA hit/miss statistics
3170 */
3171static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
3172 long nr_account)
3173{
3174#ifdef CONFIG_NUMA
3175 enum numa_stat_item local_stat = NUMA_LOCAL;
3176
3177 /* skip numa counters update if numa stats is disabled */
3178 if (!static_branch_likely(&vm_numa_stat_key))
3179 return;
3180
3181 if (zone_to_nid(z) != numa_node_id())
3182 local_stat = NUMA_OTHER;
3183
3184 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3185 __count_numa_events(z, NUMA_HIT, nr_account);
3186 else {
3187 __count_numa_events(z, NUMA_MISS, nr_account);
3188 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
3189 }
3190 __count_numa_events(z, local_stat, nr_account);
3191#endif
3192}
3193
3194static __always_inline
3195struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone,
3196 unsigned int order, unsigned int alloc_flags,
3197 int migratetype)
3198{
3199 struct page *page;
3200 unsigned long flags;
3201
3202 do {
3203 page = NULL;
3204 if (unlikely(alloc_flags & ALLOC_TRYLOCK)) {
3205 if (!spin_trylock_irqsave(&zone->lock, flags))
3206 return NULL;
3207 } else {
3208 spin_lock_irqsave(&zone->lock, flags);
3209 }
3210 if (alloc_flags & ALLOC_HIGHATOMIC)
3211 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3212 if (!page) {
3213 enum rmqueue_mode rmqm = RMQUEUE_NORMAL;
3214
3215 page = __rmqueue(zone, order, migratetype, alloc_flags, &rmqm);
3216
3217 /*
3218 * If the allocation fails, allow OOM handling and
3219 * order-0 (atomic) allocs access to HIGHATOMIC
3220 * reserves as failing now is worse than failing a
3221 * high-order atomic allocation in the future.
3222 */
3223 if (!page && (alloc_flags & (ALLOC_OOM|ALLOC_NON_BLOCK)))
3224 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3225
3226 if (!page) {
3227 spin_unlock_irqrestore(&zone->lock, flags);
3228 return NULL;
3229 }
3230 }
3231 spin_unlock_irqrestore(&zone->lock, flags);
3232 } while (check_new_pages(page, order));
3233
3234 /*
3235 * If this is a high-order atomic allocation then check
3236 * if the pageblock should be reserved for the future
3237 */
3238 if (unlikely(alloc_flags & ALLOC_HIGHATOMIC))
3239 reserve_highatomic_pageblock(page, order, zone);
3240
3241 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3242 zone_statistics(preferred_zone, zone, 1);
3243
3244 return page;
3245}
3246
3247static int nr_pcp_alloc(struct per_cpu_pages *pcp, struct zone *zone, int order)
3248{
3249 int high, base_batch, batch, max_nr_alloc;
3250 int high_max, high_min;
3251
3252 base_batch = READ_ONCE(pcp->batch);
3253 high_min = READ_ONCE(pcp->high_min);
3254 high_max = READ_ONCE(pcp->high_max);
3255 high = pcp->high = clamp(pcp->high, high_min, high_max);
3256
3257 /* Check for PCP disabled or boot pageset */
3258 if (unlikely(high < base_batch))
3259 return 1;
3260
3261 if (order)
3262 batch = base_batch;
3263 else
3264 batch = (base_batch << pcp->alloc_factor);
3265
3266 /*
3267 * If we had larger pcp->high, we could avoid to allocate from
3268 * zone.
3269 */
3270 if (high_min != high_max && !test_bit(ZONE_BELOW_HIGH, &zone->flags))
3271 high = pcp->high = min(high + batch, high_max);
3272
3273 if (!order) {
3274 max_nr_alloc = max(high - pcp->count - base_batch, base_batch);
3275 /*
3276 * Double the number of pages allocated each time there is
3277 * subsequent allocation of order-0 pages without any freeing.
3278 */
3279 if (batch <= max_nr_alloc &&
3280 pcp->alloc_factor < CONFIG_PCP_BATCH_SCALE_MAX)
3281 pcp->alloc_factor++;
3282 batch = min(batch, max_nr_alloc);
3283 }
3284
3285 /*
3286 * Scale batch relative to order if batch implies free pages
3287 * can be stored on the PCP. Batch can be 1 for small zones or
3288 * for boot pagesets which should never store free pages as
3289 * the pages may belong to arbitrary zones.
3290 */
3291 if (batch > 1)
3292 batch = max(batch >> order, 2);
3293
3294 return batch;
3295}
3296
3297/* Remove page from the per-cpu list, caller must protect the list */
3298static inline
3299struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
3300 int migratetype,
3301 unsigned int alloc_flags,
3302 struct per_cpu_pages *pcp,
3303 struct list_head *list)
3304{
3305 struct page *page;
3306
3307 do {
3308 if (list_empty(list)) {
3309 int batch = nr_pcp_alloc(pcp, zone, order);
3310 int alloced;
3311
3312 /*
3313 * Don't refill the list for a higher order atomic
3314 * allocation under memory pressure, as this would
3315 * not build up any HIGHATOMIC reserves, which
3316 * might be needed soon.
3317 *
3318 * Instead, direct it towards the reserves by
3319 * returning NULL, which will make the caller fall
3320 * back to rmqueue_buddy. This will try to use the
3321 * reserves first and grow them if needed.
3322 */
3323 if (alloc_flags & ALLOC_HIGHATOMIC)
3324 return NULL;
3325
3326 alloced = rmqueue_bulk(zone, order,
3327 batch, list,
3328 migratetype, alloc_flags);
3329
3330 pcp->count += alloced << order;
3331 if (unlikely(list_empty(list)))
3332 return NULL;
3333 }
3334
3335 page = list_first_entry(list, struct page, pcp_list);
3336 list_del(&page->pcp_list);
3337 pcp->count -= 1 << order;
3338 } while (check_new_pages(page, order));
3339
3340 return page;
3341}
3342
3343/* Lock and remove page from the per-cpu list */
3344static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3345 struct zone *zone, unsigned int order,
3346 int migratetype, unsigned int alloc_flags)
3347{
3348 struct per_cpu_pages *pcp;
3349 struct list_head *list;
3350 struct page *page;
3351
3352 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
3353 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
3354 if (!pcp)
3355 return NULL;
3356
3357 /*
3358 * On allocation, reduce the number of pages that are batch freed.
3359 * See nr_pcp_free() where free_factor is increased for subsequent
3360 * frees.
3361 */
3362 pcp->free_count >>= 1;
3363 list = &pcp->lists[order_to_pindex(migratetype, order)];
3364 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
3365 pcp_spin_unlock(pcp);
3366 if (page) {
3367 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3368 zone_statistics(preferred_zone, zone, 1);
3369 }
3370 return page;
3371}
3372
3373/*
3374 * Allocate a page from the given zone.
3375 * Use pcplists for THP or "cheap" high-order allocations.
3376 */
3377
3378/*
3379 * Do not instrument rmqueue() with KMSAN. This function may call
3380 * __msan_poison_alloca() through a call to set_pfnblock_migratetype().
3381 * If __msan_poison_alloca() attempts to allocate pages for the stack depot, it
3382 * may call rmqueue() again, which will result in a deadlock.
3383 */
3384__no_sanitize_memory
3385static inline
3386struct page *rmqueue(struct zone *preferred_zone,
3387 struct zone *zone, unsigned int order,
3388 gfp_t gfp_flags, unsigned int alloc_flags,
3389 int migratetype)
3390{
3391 struct page *page;
3392
3393 if (likely(pcp_allowed_order(order))) {
3394 page = rmqueue_pcplist(preferred_zone, zone, order,
3395 migratetype, alloc_flags);
3396 if (likely(page))
3397 goto out;
3398 }
3399
3400 page = rmqueue_buddy(preferred_zone, zone, order, alloc_flags,
3401 migratetype);
3402
3403out:
3404 /* Separate test+clear to avoid unnecessary atomics */
3405 if ((alloc_flags & ALLOC_KSWAPD) &&
3406 unlikely(test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags))) {
3407 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3408 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3409 }
3410
3411 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3412 return page;
3413}
3414
3415/*
3416 * Reserve the pageblock(s) surrounding an allocation request for
3417 * exclusive use of high-order atomic allocations if there are no
3418 * empty page blocks that contain a page with a suitable order
3419 */
3420static void reserve_highatomic_pageblock(struct page *page, int order,
3421 struct zone *zone)
3422{
3423 int mt;
3424 unsigned long max_managed, flags;
3425
3426 /*
3427 * The number reserved as: minimum is 1 pageblock, maximum is
3428 * roughly 1% of a zone. But if 1% of a zone falls below a
3429 * pageblock size, then don't reserve any pageblocks.
3430 * Check is race-prone but harmless.
3431 */
3432 if ((zone_managed_pages(zone) / 100) < pageblock_nr_pages)
3433 return;
3434 max_managed = ALIGN((zone_managed_pages(zone) / 100), pageblock_nr_pages);
3435 if (zone->nr_reserved_highatomic >= max_managed)
3436 return;
3437
3438 spin_lock_irqsave(&zone->lock, flags);
3439
3440 /* Recheck the nr_reserved_highatomic limit under the lock */
3441 if (zone->nr_reserved_highatomic >= max_managed)
3442 goto out_unlock;
3443
3444 /* Yoink! */
3445 mt = get_pageblock_migratetype(page);
3446 /* Only reserve normal pageblocks (i.e., they can merge with others) */
3447 if (!migratetype_is_mergeable(mt))
3448 goto out_unlock;
3449
3450 if (order < pageblock_order) {
3451 if (move_freepages_block(zone, page, mt, MIGRATE_HIGHATOMIC) == -1)
3452 goto out_unlock;
3453 zone->nr_reserved_highatomic += pageblock_nr_pages;
3454 } else {
3455 change_pageblock_range(page, order, MIGRATE_HIGHATOMIC);
3456 zone->nr_reserved_highatomic += 1 << order;
3457 }
3458
3459out_unlock:
3460 spin_unlock_irqrestore(&zone->lock, flags);
3461}
3462
3463/*
3464 * Used when an allocation is about to fail under memory pressure. This
3465 * potentially hurts the reliability of high-order allocations when under
3466 * intense memory pressure but failed atomic allocations should be easier
3467 * to recover from than an OOM.
3468 *
3469 * If @force is true, try to unreserve pageblocks even though highatomic
3470 * pageblock is exhausted.
3471 */
3472static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
3473 bool force)
3474{
3475 struct zonelist *zonelist = ac->zonelist;
3476 unsigned long flags;
3477 struct zoneref *z;
3478 struct zone *zone;
3479 struct page *page;
3480 int order;
3481 int ret;
3482
3483 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
3484 ac->nodemask) {
3485 /*
3486 * Preserve at least one pageblock unless memory pressure
3487 * is really high.
3488 */
3489 if (!force && zone->nr_reserved_highatomic <=
3490 pageblock_nr_pages)
3491 continue;
3492
3493 spin_lock_irqsave(&zone->lock, flags);
3494 for (order = 0; order < NR_PAGE_ORDERS; order++) {
3495 struct free_area *area = &(zone->free_area[order]);
3496 unsigned long size;
3497
3498 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
3499 if (!page)
3500 continue;
3501
3502 size = max(pageblock_nr_pages, 1UL << order);
3503 /*
3504 * It should never happen but changes to
3505 * locking could inadvertently allow a per-cpu
3506 * drain to add pages to MIGRATE_HIGHATOMIC
3507 * while unreserving so be safe and watch for
3508 * underflows.
3509 */
3510 if (WARN_ON_ONCE(size > zone->nr_reserved_highatomic))
3511 size = zone->nr_reserved_highatomic;
3512 zone->nr_reserved_highatomic -= size;
3513
3514 /*
3515 * Convert to ac->migratetype and avoid the normal
3516 * pageblock stealing heuristics. Minimally, the caller
3517 * is doing the work and needs the pages. More
3518 * importantly, if the block was always converted to
3519 * MIGRATE_UNMOVABLE or another type then the number
3520 * of pageblocks that cannot be completely freed
3521 * may increase.
3522 */
3523 if (order < pageblock_order)
3524 ret = move_freepages_block(zone, page,
3525 MIGRATE_HIGHATOMIC,
3526 ac->migratetype);
3527 else {
3528 move_to_free_list(page, zone, order,
3529 MIGRATE_HIGHATOMIC,
3530 ac->migratetype);
3531 change_pageblock_range(page, order,
3532 ac->migratetype);
3533 ret = 1;
3534 }
3535 /*
3536 * Reserving the block(s) already succeeded,
3537 * so this should not fail on zone boundaries.
3538 */
3539 WARN_ON_ONCE(ret == -1);
3540 if (ret > 0) {
3541 spin_unlock_irqrestore(&zone->lock, flags);
3542 return ret;
3543 }
3544 }
3545 spin_unlock_irqrestore(&zone->lock, flags);
3546 }
3547
3548 return false;
3549}
3550
3551static inline long __zone_watermark_unusable_free(struct zone *z,
3552 unsigned int order, unsigned int alloc_flags)
3553{
3554 long unusable_free = (1 << order) - 1;
3555
3556 /*
3557 * If the caller does not have rights to reserves below the min
3558 * watermark then subtract the free pages reserved for highatomic.
3559 */
3560 if (likely(!(alloc_flags & ALLOC_RESERVES)))
3561 unusable_free += READ_ONCE(z->nr_free_highatomic);
3562
3563#ifdef CONFIG_CMA
3564 /* If allocation can't use CMA areas don't use free CMA pages */
3565 if (!(alloc_flags & ALLOC_CMA))
3566 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3567#endif
3568
3569 return unusable_free;
3570}
3571
3572/*
3573 * Return true if free base pages are above 'mark'. For high-order checks it
3574 * will return true of the order-0 watermark is reached and there is at least
3575 * one free page of a suitable size. Checking now avoids taking the zone lock
3576 * to check in the allocation paths if no pages are free.
3577 */
3578bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3579 int highest_zoneidx, unsigned int alloc_flags,
3580 long free_pages)
3581{
3582 long min = mark;
3583 int o;
3584
3585 /* free_pages may go negative - that's OK */
3586 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3587
3588 if (unlikely(alloc_flags & ALLOC_RESERVES)) {
3589 /*
3590 * __GFP_HIGH allows access to 50% of the min reserve as well
3591 * as OOM.
3592 */
3593 if (alloc_flags & ALLOC_MIN_RESERVE) {
3594 min -= min / 2;
3595
3596 /*
3597 * Non-blocking allocations (e.g. GFP_ATOMIC) can
3598 * access more reserves than just __GFP_HIGH. Other
3599 * non-blocking allocations requests such as GFP_NOWAIT
3600 * or (GFP_KERNEL & ~__GFP_DIRECT_RECLAIM) do not get
3601 * access to the min reserve.
3602 */
3603 if (alloc_flags & ALLOC_NON_BLOCK)
3604 min -= min / 4;
3605 }
3606
3607 /*
3608 * OOM victims can try even harder than the normal reserve
3609 * users on the grounds that it's definitely going to be in
3610 * the exit path shortly and free memory. Any allocation it
3611 * makes during the free path will be small and short-lived.
3612 */
3613 if (alloc_flags & ALLOC_OOM)
3614 min -= min / 2;
3615 }
3616
3617 /*
3618 * Check watermarks for an order-0 allocation request. If these
3619 * are not met, then a high-order request also cannot go ahead
3620 * even if a suitable page happened to be free.
3621 */
3622 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3623 return false;
3624
3625 /* If this is an order-0 request then the watermark is fine */
3626 if (!order)
3627 return true;
3628
3629 /* For a high-order request, check at least one suitable page is free */
3630 for (o = order; o < NR_PAGE_ORDERS; o++) {
3631 struct free_area *area = &z->free_area[o];
3632 int mt;
3633
3634 if (!area->nr_free)
3635 continue;
3636
3637 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3638 if (!free_area_empty(area, mt))
3639 return true;
3640 }
3641
3642#ifdef CONFIG_CMA
3643 if ((alloc_flags & ALLOC_CMA) &&
3644 !free_area_empty(area, MIGRATE_CMA)) {
3645 return true;
3646 }
3647#endif
3648 if ((alloc_flags & (ALLOC_HIGHATOMIC|ALLOC_OOM)) &&
3649 !free_area_empty(area, MIGRATE_HIGHATOMIC)) {
3650 return true;
3651 }
3652 }
3653 return false;
3654}
3655
3656bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3657 int highest_zoneidx, unsigned int alloc_flags)
3658{
3659 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3660 zone_page_state(z, NR_FREE_PAGES));
3661}
3662
3663static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3664 unsigned long mark, int highest_zoneidx,
3665 unsigned int alloc_flags, gfp_t gfp_mask)
3666{
3667 long free_pages;
3668
3669 free_pages = zone_page_state(z, NR_FREE_PAGES);
3670
3671 /*
3672 * Fast check for order-0 only. If this fails then the reserves
3673 * need to be calculated.
3674 */
3675 if (!order) {
3676 long usable_free;
3677 long reserved;
3678
3679 usable_free = free_pages;
3680 reserved = __zone_watermark_unusable_free(z, 0, alloc_flags);
3681
3682 /* reserved may over estimate high-atomic reserves. */
3683 usable_free -= min(usable_free, reserved);
3684 if (usable_free > mark + z->lowmem_reserve[highest_zoneidx])
3685 return true;
3686 }
3687
3688 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3689 free_pages))
3690 return true;
3691
3692 /*
3693 * Ignore watermark boosting for __GFP_HIGH order-0 allocations
3694 * when checking the min watermark. The min watermark is the
3695 * point where boosting is ignored so that kswapd is woken up
3696 * when below the low watermark.
3697 */
3698 if (unlikely(!order && (alloc_flags & ALLOC_MIN_RESERVE) && z->watermark_boost
3699 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3700 mark = z->_watermark[WMARK_MIN];
3701 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3702 alloc_flags, free_pages);
3703 }
3704
3705 return false;
3706}
3707
3708#ifdef CONFIG_NUMA
3709int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
3710
3711static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3712{
3713 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3714 node_reclaim_distance;
3715}
3716#else /* CONFIG_NUMA */
3717static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3718{
3719 return true;
3720}
3721#endif /* CONFIG_NUMA */
3722
3723/*
3724 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3725 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3726 * premature use of a lower zone may cause lowmem pressure problems that
3727 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3728 * probably too small. It only makes sense to spread allocations to avoid
3729 * fragmentation between the Normal and DMA32 zones.
3730 */
3731static inline unsigned int
3732alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3733{
3734 unsigned int alloc_flags;
3735
3736 /*
3737 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3738 * to save a branch.
3739 */
3740 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3741
3742 if (defrag_mode) {
3743 alloc_flags |= ALLOC_NOFRAGMENT;
3744 return alloc_flags;
3745 }
3746
3747#ifdef CONFIG_ZONE_DMA32
3748 if (!zone)
3749 return alloc_flags;
3750
3751 if (zone_idx(zone) != ZONE_NORMAL)
3752 return alloc_flags;
3753
3754 /*
3755 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3756 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3757 * on UMA that if Normal is populated then so is DMA32.
3758 */
3759 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3760 if (nr_online_nodes > 1 && !populated_zone(--zone))
3761 return alloc_flags;
3762
3763 alloc_flags |= ALLOC_NOFRAGMENT;
3764#endif /* CONFIG_ZONE_DMA32 */
3765 return alloc_flags;
3766}
3767
3768/* Must be called after current_gfp_context() which can change gfp_mask */
3769static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
3770 unsigned int alloc_flags)
3771{
3772#ifdef CONFIG_CMA
3773 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3774 alloc_flags |= ALLOC_CMA;
3775#endif
3776 return alloc_flags;
3777}
3778
3779/*
3780 * get_page_from_freelist goes through the zonelist trying to allocate
3781 * a page.
3782 */
3783static struct page *
3784get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3785 const struct alloc_context *ac)
3786{
3787 struct zoneref *z;
3788 struct zone *zone;
3789 struct pglist_data *last_pgdat = NULL;
3790 bool last_pgdat_dirty_ok = false;
3791 bool no_fallback;
3792 bool skip_kswapd_nodes = nr_online_nodes > 1;
3793 bool skipped_kswapd_nodes = false;
3794
3795retry:
3796 /*
3797 * Scan zonelist, looking for a zone with enough free.
3798 * See also cpuset_current_node_allowed() comment in kernel/cgroup/cpuset.c.
3799 */
3800 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3801 z = ac->preferred_zoneref;
3802 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3803 ac->nodemask) {
3804 struct page *page;
3805 unsigned long mark;
3806
3807 if (cpusets_enabled() &&
3808 (alloc_flags & ALLOC_CPUSET) &&
3809 !__cpuset_zone_allowed(zone, gfp_mask))
3810 continue;
3811 /*
3812 * When allocating a page cache page for writing, we
3813 * want to get it from a node that is within its dirty
3814 * limit, such that no single node holds more than its
3815 * proportional share of globally allowed dirty pages.
3816 * The dirty limits take into account the node's
3817 * lowmem reserves and high watermark so that kswapd
3818 * should be able to balance it without having to
3819 * write pages from its LRU list.
3820 *
3821 * XXX: For now, allow allocations to potentially
3822 * exceed the per-node dirty limit in the slowpath
3823 * (spread_dirty_pages unset) before going into reclaim,
3824 * which is important when on a NUMA setup the allowed
3825 * nodes are together not big enough to reach the
3826 * global limit. The proper fix for these situations
3827 * will require awareness of nodes in the
3828 * dirty-throttling and the flusher threads.
3829 */
3830 if (ac->spread_dirty_pages) {
3831 if (last_pgdat != zone->zone_pgdat) {
3832 last_pgdat = zone->zone_pgdat;
3833 last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat);
3834 }
3835
3836 if (!last_pgdat_dirty_ok)
3837 continue;
3838 }
3839
3840 if (no_fallback && !defrag_mode && nr_online_nodes > 1 &&
3841 zone != zonelist_zone(ac->preferred_zoneref)) {
3842 int local_nid;
3843
3844 /*
3845 * If moving to a remote node, retry but allow
3846 * fragmenting fallbacks. Locality is more important
3847 * than fragmentation avoidance.
3848 */
3849 local_nid = zonelist_node_idx(ac->preferred_zoneref);
3850 if (zone_to_nid(zone) != local_nid) {
3851 alloc_flags &= ~ALLOC_NOFRAGMENT;
3852 goto retry;
3853 }
3854 }
3855
3856 /*
3857 * If kswapd is already active on a node, keep looking
3858 * for other nodes that might be idle. This can happen
3859 * if another process has NUMA bindings and is causing
3860 * kswapd wakeups on only some nodes. Avoid accidental
3861 * "node_reclaim_mode"-like behavior in this case.
3862 */
3863 if (skip_kswapd_nodes &&
3864 !waitqueue_active(&zone->zone_pgdat->kswapd_wait)) {
3865 skipped_kswapd_nodes = true;
3866 continue;
3867 }
3868
3869 cond_accept_memory(zone, order, alloc_flags);
3870
3871 /*
3872 * Detect whether the number of free pages is below high
3873 * watermark. If so, we will decrease pcp->high and free
3874 * PCP pages in free path to reduce the possibility of
3875 * premature page reclaiming. Detection is done here to
3876 * avoid to do that in hotter free path.
3877 */
3878 if (test_bit(ZONE_BELOW_HIGH, &zone->flags))
3879 goto check_alloc_wmark;
3880
3881 mark = high_wmark_pages(zone);
3882 if (zone_watermark_fast(zone, order, mark,
3883 ac->highest_zoneidx, alloc_flags,
3884 gfp_mask))
3885 goto try_this_zone;
3886 else
3887 set_bit(ZONE_BELOW_HIGH, &zone->flags);
3888
3889check_alloc_wmark:
3890 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3891 if (!zone_watermark_fast(zone, order, mark,
3892 ac->highest_zoneidx, alloc_flags,
3893 gfp_mask)) {
3894 int ret;
3895
3896 if (cond_accept_memory(zone, order, alloc_flags))
3897 goto try_this_zone;
3898
3899 /*
3900 * Watermark failed for this zone, but see if we can
3901 * grow this zone if it contains deferred pages.
3902 */
3903 if (deferred_pages_enabled()) {
3904 if (_deferred_grow_zone(zone, order))
3905 goto try_this_zone;
3906 }
3907 /* Checked here to keep the fast path fast */
3908 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3909 if (alloc_flags & ALLOC_NO_WATERMARKS)
3910 goto try_this_zone;
3911
3912 if (!node_reclaim_enabled() ||
3913 !zone_allows_reclaim(zonelist_zone(ac->preferred_zoneref), zone))
3914 continue;
3915
3916 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3917 switch (ret) {
3918 case NODE_RECLAIM_NOSCAN:
3919 /* did not scan */
3920 continue;
3921 case NODE_RECLAIM_FULL:
3922 /* scanned but unreclaimable */
3923 continue;
3924 default:
3925 /* did we reclaim enough */
3926 if (zone_watermark_ok(zone, order, mark,
3927 ac->highest_zoneidx, alloc_flags))
3928 goto try_this_zone;
3929
3930 continue;
3931 }
3932 }
3933
3934try_this_zone:
3935 page = rmqueue(zonelist_zone(ac->preferred_zoneref), zone, order,
3936 gfp_mask, alloc_flags, ac->migratetype);
3937 if (page) {
3938 prep_new_page(page, order, gfp_mask, alloc_flags);
3939
3940 return page;
3941 } else {
3942 if (cond_accept_memory(zone, order, alloc_flags))
3943 goto try_this_zone;
3944
3945 /* Try again if zone has deferred pages */
3946 if (deferred_pages_enabled()) {
3947 if (_deferred_grow_zone(zone, order))
3948 goto try_this_zone;
3949 }
3950 }
3951 }
3952
3953 /*
3954 * If we skipped over nodes with active kswapds and found no
3955 * idle nodes, retry and place anywhere the watermarks permit.
3956 */
3957 if (skip_kswapd_nodes && skipped_kswapd_nodes) {
3958 skip_kswapd_nodes = false;
3959 goto retry;
3960 }
3961
3962 /*
3963 * It's possible on a UMA machine to get through all zones that are
3964 * fragmented. If avoiding fragmentation, reset and try again.
3965 */
3966 if (no_fallback && !defrag_mode) {
3967 alloc_flags &= ~ALLOC_NOFRAGMENT;
3968 goto retry;
3969 }
3970
3971 return NULL;
3972}
3973
3974static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3975{
3976 unsigned int filter = SHOW_MEM_FILTER_NODES;
3977
3978 /*
3979 * This documents exceptions given to allocations in certain
3980 * contexts that are allowed to allocate outside current's set
3981 * of allowed nodes.
3982 */
3983 if (!(gfp_mask & __GFP_NOMEMALLOC))
3984 if (tsk_is_oom_victim(current) ||
3985 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3986 filter &= ~SHOW_MEM_FILTER_NODES;
3987 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3988 filter &= ~SHOW_MEM_FILTER_NODES;
3989
3990 __show_mem(filter, nodemask, gfp_zone(gfp_mask));
3991 mem_cgroup_show_protected_memory(NULL);
3992}
3993
3994void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3995{
3996 struct va_format vaf;
3997 va_list args;
3998 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3999
4000 if ((gfp_mask & __GFP_NOWARN) ||
4001 !__ratelimit(&nopage_rs) ||
4002 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
4003 return;
4004
4005 va_start(args, fmt);
4006 vaf.fmt = fmt;
4007 vaf.va = &args;
4008 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4009 current->comm, &vaf, gfp_mask, &gfp_mask,
4010 nodemask_pr_args(nodemask));
4011 va_end(args);
4012
4013 cpuset_print_current_mems_allowed();
4014 pr_cont("\n");
4015 dump_stack();
4016 warn_alloc_show_mem(gfp_mask, nodemask);
4017}
4018
4019static inline struct page *
4020__alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4021 unsigned int alloc_flags,
4022 const struct alloc_context *ac)
4023{
4024 struct page *page;
4025
4026 page = get_page_from_freelist(gfp_mask, order,
4027 alloc_flags|ALLOC_CPUSET, ac);
4028 /*
4029 * fallback to ignore cpuset restriction if our nodes
4030 * are depleted
4031 */
4032 if (!page)
4033 page = get_page_from_freelist(gfp_mask, order,
4034 alloc_flags, ac);
4035 return page;
4036}
4037
4038static inline struct page *
4039__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4040 const struct alloc_context *ac, unsigned long *did_some_progress)
4041{
4042 struct oom_control oc = {
4043 .zonelist = ac->zonelist,
4044 .nodemask = ac->nodemask,
4045 .memcg = NULL,
4046 .gfp_mask = gfp_mask,
4047 .order = order,
4048 };
4049 struct page *page;
4050
4051 *did_some_progress = 0;
4052
4053 /*
4054 * Acquire the oom lock. If that fails, somebody else is
4055 * making progress for us.
4056 */
4057 if (!mutex_trylock(&oom_lock)) {
4058 *did_some_progress = 1;
4059 schedule_timeout_uninterruptible(1);
4060 return NULL;
4061 }
4062
4063 /*
4064 * Go through the zonelist yet one more time, keep very high watermark
4065 * here, this is only to catch a parallel oom killing, we must fail if
4066 * we're still under heavy pressure. But make sure that this reclaim
4067 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4068 * allocation which will never fail due to oom_lock already held.
4069 */
4070 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4071 ~__GFP_DIRECT_RECLAIM, order,
4072 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4073 if (page)
4074 goto out;
4075
4076 /* Coredumps can quickly deplete all memory reserves */
4077 if (current->flags & PF_DUMPCORE)
4078 goto out;
4079 /* The OOM killer will not help higher order allocs */
4080 if (order > PAGE_ALLOC_COSTLY_ORDER)
4081 goto out;
4082 /*
4083 * We have already exhausted all our reclaim opportunities without any
4084 * success so it is time to admit defeat. We will skip the OOM killer
4085 * because it is very likely that the caller has a more reasonable
4086 * fallback than shooting a random task.
4087 *
4088 * The OOM killer may not free memory on a specific node.
4089 */
4090 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4091 goto out;
4092 /* The OOM killer does not needlessly kill tasks for lowmem */
4093 if (ac->highest_zoneidx < ZONE_NORMAL)
4094 goto out;
4095 if (pm_suspended_storage())
4096 goto out;
4097 /*
4098 * XXX: GFP_NOFS allocations should rather fail than rely on
4099 * other request to make a forward progress.
4100 * We are in an unfortunate situation where out_of_memory cannot
4101 * do much for this context but let's try it to at least get
4102 * access to memory reserved if the current task is killed (see
4103 * out_of_memory). Once filesystems are ready to handle allocation
4104 * failures more gracefully we should just bail out here.
4105 */
4106
4107 /* Exhausted what can be done so it's blame time */
4108 if (out_of_memory(&oc) ||
4109 WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) {
4110 *did_some_progress = 1;
4111
4112 /*
4113 * Help non-failing allocations by giving them access to memory
4114 * reserves
4115 */
4116 if (gfp_mask & __GFP_NOFAIL)
4117 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4118 ALLOC_NO_WATERMARKS, ac);
4119 }
4120out:
4121 mutex_unlock(&oom_lock);
4122 return page;
4123}
4124
4125/*
4126 * Maximum number of compaction retries with a progress before OOM
4127 * killer is consider as the only way to move forward.
4128 */
4129#define MAX_COMPACT_RETRIES 16
4130
4131#ifdef CONFIG_COMPACTION
4132/* Try memory compaction for high-order allocations before reclaim */
4133static struct page *
4134__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4135 unsigned int alloc_flags, const struct alloc_context *ac,
4136 enum compact_priority prio, enum compact_result *compact_result)
4137{
4138 struct page *page = NULL;
4139 unsigned long pflags;
4140 unsigned int noreclaim_flag;
4141
4142 if (!order)
4143 return NULL;
4144
4145 psi_memstall_enter(&pflags);
4146 delayacct_compact_start();
4147 noreclaim_flag = memalloc_noreclaim_save();
4148
4149 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4150 prio, &page);
4151
4152 memalloc_noreclaim_restore(noreclaim_flag);
4153 psi_memstall_leave(&pflags);
4154 delayacct_compact_end();
4155
4156 if (*compact_result == COMPACT_SKIPPED)
4157 return NULL;
4158 /*
4159 * At least in one zone compaction wasn't deferred or skipped, so let's
4160 * count a compaction stall
4161 */
4162 count_vm_event(COMPACTSTALL);
4163
4164 /* Prep a captured page if available */
4165 if (page)
4166 prep_new_page(page, order, gfp_mask, alloc_flags);
4167
4168 /* Try get a page from the freelist if available */
4169 if (!page)
4170 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4171
4172 if (page) {
4173 struct zone *zone = page_zone(page);
4174
4175 zone->compact_blockskip_flush = false;
4176 compaction_defer_reset(zone, order, true);
4177 count_vm_event(COMPACTSUCCESS);
4178 return page;
4179 }
4180
4181 /*
4182 * It's bad if compaction run occurs and fails. The most likely reason
4183 * is that pages exist, but not enough to satisfy watermarks.
4184 */
4185 count_vm_event(COMPACTFAIL);
4186
4187 cond_resched();
4188
4189 return NULL;
4190}
4191
4192static inline bool
4193should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4194 enum compact_result compact_result,
4195 enum compact_priority *compact_priority,
4196 int *compaction_retries)
4197{
4198 int max_retries = MAX_COMPACT_RETRIES;
4199 int min_priority;
4200 bool ret = false;
4201 int retries = *compaction_retries;
4202 enum compact_priority priority = *compact_priority;
4203
4204 if (!order)
4205 return false;
4206
4207 if (fatal_signal_pending(current))
4208 return false;
4209
4210 /*
4211 * Compaction was skipped due to a lack of free order-0
4212 * migration targets. Continue if reclaim can help.
4213 */
4214 if (compact_result == COMPACT_SKIPPED) {
4215 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4216 goto out;
4217 }
4218
4219 /*
4220 * Compaction managed to coalesce some page blocks, but the
4221 * allocation failed presumably due to a race. Retry some.
4222 */
4223 if (compact_result == COMPACT_SUCCESS) {
4224 /*
4225 * !costly requests are much more important than
4226 * __GFP_RETRY_MAYFAIL costly ones because they are de
4227 * facto nofail and invoke OOM killer to move on while
4228 * costly can fail and users are ready to cope with
4229 * that. 1/4 retries is rather arbitrary but we would
4230 * need much more detailed feedback from compaction to
4231 * make a better decision.
4232 */
4233 if (order > PAGE_ALLOC_COSTLY_ORDER)
4234 max_retries /= 4;
4235
4236 if (++(*compaction_retries) <= max_retries) {
4237 ret = true;
4238 goto out;
4239 }
4240 }
4241
4242 /*
4243 * Compaction failed. Retry with increasing priority.
4244 */
4245 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4246 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4247
4248 if (*compact_priority > min_priority) {
4249 (*compact_priority)--;
4250 *compaction_retries = 0;
4251 ret = true;
4252 }
4253out:
4254 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4255 return ret;
4256}
4257#else
4258static inline struct page *
4259__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4260 unsigned int alloc_flags, const struct alloc_context *ac,
4261 enum compact_priority prio, enum compact_result *compact_result)
4262{
4263 *compact_result = COMPACT_SKIPPED;
4264 return NULL;
4265}
4266
4267static inline bool
4268should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4269 enum compact_result compact_result,
4270 enum compact_priority *compact_priority,
4271 int *compaction_retries)
4272{
4273 struct zone *zone;
4274 struct zoneref *z;
4275
4276 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4277 return false;
4278
4279 /*
4280 * There are setups with compaction disabled which would prefer to loop
4281 * inside the allocator rather than hit the oom killer prematurely.
4282 * Let's give them a good hope and keep retrying while the order-0
4283 * watermarks are OK.
4284 */
4285 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4286 ac->highest_zoneidx, ac->nodemask) {
4287 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4288 ac->highest_zoneidx, alloc_flags))
4289 return true;
4290 }
4291 return false;
4292}
4293#endif /* CONFIG_COMPACTION */
4294
4295#ifdef CONFIG_LOCKDEP
4296static struct lockdep_map __fs_reclaim_map =
4297 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4298
4299static bool __need_reclaim(gfp_t gfp_mask)
4300{
4301 /* no reclaim without waiting on it */
4302 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4303 return false;
4304
4305 /* this guy won't enter reclaim */
4306 if (current->flags & PF_MEMALLOC)
4307 return false;
4308
4309 if (gfp_mask & __GFP_NOLOCKDEP)
4310 return false;
4311
4312 return true;
4313}
4314
4315void __fs_reclaim_acquire(unsigned long ip)
4316{
4317 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
4318}
4319
4320void __fs_reclaim_release(unsigned long ip)
4321{
4322 lock_release(&__fs_reclaim_map, ip);
4323}
4324
4325void fs_reclaim_acquire(gfp_t gfp_mask)
4326{
4327 gfp_mask = current_gfp_context(gfp_mask);
4328
4329 if (__need_reclaim(gfp_mask)) {
4330 if (gfp_mask & __GFP_FS)
4331 __fs_reclaim_acquire(_RET_IP_);
4332
4333#ifdef CONFIG_MMU_NOTIFIER
4334 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4335 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4336#endif
4337
4338 }
4339}
4340EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4341
4342void fs_reclaim_release(gfp_t gfp_mask)
4343{
4344 gfp_mask = current_gfp_context(gfp_mask);
4345
4346 if (__need_reclaim(gfp_mask)) {
4347 if (gfp_mask & __GFP_FS)
4348 __fs_reclaim_release(_RET_IP_);
4349 }
4350}
4351EXPORT_SYMBOL_GPL(fs_reclaim_release);
4352#endif
4353
4354/*
4355 * Zonelists may change due to hotplug during allocation. Detect when zonelists
4356 * have been rebuilt so allocation retries. Reader side does not lock and
4357 * retries the allocation if zonelist changes. Writer side is protected by the
4358 * embedded spin_lock.
4359 */
4360static DEFINE_SEQLOCK(zonelist_update_seq);
4361
4362static unsigned int zonelist_iter_begin(void)
4363{
4364 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
4365 return read_seqbegin(&zonelist_update_seq);
4366
4367 return 0;
4368}
4369
4370static unsigned int check_retry_zonelist(unsigned int seq)
4371{
4372 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
4373 return read_seqretry(&zonelist_update_seq, seq);
4374
4375 return seq;
4376}
4377
4378/* Perform direct synchronous page reclaim */
4379static unsigned long
4380__perform_reclaim(gfp_t gfp_mask, unsigned int order,
4381 const struct alloc_context *ac)
4382{
4383 unsigned int noreclaim_flag;
4384 unsigned long progress;
4385
4386 cond_resched();
4387
4388 /* We now go into synchronous reclaim */
4389 cpuset_memory_pressure_bump();
4390 fs_reclaim_acquire(gfp_mask);
4391 noreclaim_flag = memalloc_noreclaim_save();
4392
4393 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4394 ac->nodemask);
4395
4396 memalloc_noreclaim_restore(noreclaim_flag);
4397 fs_reclaim_release(gfp_mask);
4398
4399 cond_resched();
4400
4401 return progress;
4402}
4403
4404/* The really slow allocator path where we enter direct reclaim */
4405static inline struct page *
4406__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4407 unsigned int alloc_flags, const struct alloc_context *ac,
4408 unsigned long *did_some_progress)
4409{
4410 struct page *page = NULL;
4411 unsigned long pflags;
4412 bool drained = false;
4413
4414 psi_memstall_enter(&pflags);
4415 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4416 if (unlikely(!(*did_some_progress)))
4417 goto out;
4418
4419retry:
4420 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4421
4422 /*
4423 * If an allocation failed after direct reclaim, it could be because
4424 * pages are pinned on the per-cpu lists or in high alloc reserves.
4425 * Shrink them and try again
4426 */
4427 if (!page && !drained) {
4428 unreserve_highatomic_pageblock(ac, false);
4429 drain_all_pages(NULL);
4430 drained = true;
4431 goto retry;
4432 }
4433out:
4434 psi_memstall_leave(&pflags);
4435
4436 return page;
4437}
4438
4439static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4440 const struct alloc_context *ac)
4441{
4442 struct zoneref *z;
4443 struct zone *zone;
4444 pg_data_t *last_pgdat = NULL;
4445 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4446 unsigned int reclaim_order;
4447
4448 if (defrag_mode)
4449 reclaim_order = max(order, pageblock_order);
4450 else
4451 reclaim_order = order;
4452
4453 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4454 ac->nodemask) {
4455 if (!managed_zone(zone))
4456 continue;
4457 if (last_pgdat == zone->zone_pgdat)
4458 continue;
4459 wakeup_kswapd(zone, gfp_mask, reclaim_order, highest_zoneidx);
4460 last_pgdat = zone->zone_pgdat;
4461 }
4462}
4463
4464static inline unsigned int
4465gfp_to_alloc_flags(gfp_t gfp_mask, unsigned int order)
4466{
4467 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4468
4469 /*
4470 * __GFP_HIGH is assumed to be the same as ALLOC_MIN_RESERVE
4471 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4472 * to save two branches.
4473 */
4474 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_MIN_RESERVE);
4475 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4476
4477 /*
4478 * The caller may dip into page reserves a bit more if the caller
4479 * cannot run direct reclaim, or if the caller has realtime scheduling
4480 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4481 * set both ALLOC_NON_BLOCK and ALLOC_MIN_RESERVE(__GFP_HIGH).
4482 */
4483 alloc_flags |= (__force int)
4484 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4485
4486 if (!(gfp_mask & __GFP_DIRECT_RECLAIM)) {
4487 /*
4488 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4489 * if it can't schedule.
4490 */
4491 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
4492 alloc_flags |= ALLOC_NON_BLOCK;
4493
4494 if (order > 0 && (alloc_flags & ALLOC_MIN_RESERVE))
4495 alloc_flags |= ALLOC_HIGHATOMIC;
4496 }
4497
4498 /*
4499 * Ignore cpuset mems for non-blocking __GFP_HIGH (probably
4500 * GFP_ATOMIC) rather than fail, see the comment for
4501 * cpuset_current_node_allowed().
4502 */
4503 if (alloc_flags & ALLOC_MIN_RESERVE)
4504 alloc_flags &= ~ALLOC_CPUSET;
4505 } else if (unlikely(rt_or_dl_task(current)) && in_task())
4506 alloc_flags |= ALLOC_MIN_RESERVE;
4507
4508 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4509
4510 if (defrag_mode)
4511 alloc_flags |= ALLOC_NOFRAGMENT;
4512
4513 return alloc_flags;
4514}
4515
4516static bool oom_reserves_allowed(struct task_struct *tsk)
4517{
4518 if (!tsk_is_oom_victim(tsk))
4519 return false;
4520
4521 /*
4522 * !MMU doesn't have oom reaper so give access to memory reserves
4523 * only to the thread with TIF_MEMDIE set
4524 */
4525 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4526 return false;
4527
4528 return true;
4529}
4530
4531/*
4532 * Distinguish requests which really need access to full memory
4533 * reserves from oom victims which can live with a portion of it
4534 */
4535static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4536{
4537 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4538 return 0;
4539 if (gfp_mask & __GFP_MEMALLOC)
4540 return ALLOC_NO_WATERMARKS;
4541 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4542 return ALLOC_NO_WATERMARKS;
4543 if (!in_interrupt()) {
4544 if (current->flags & PF_MEMALLOC)
4545 return ALLOC_NO_WATERMARKS;
4546 else if (oom_reserves_allowed(current))
4547 return ALLOC_OOM;
4548 }
4549
4550 return 0;
4551}
4552
4553bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4554{
4555 return !!__gfp_pfmemalloc_flags(gfp_mask);
4556}
4557
4558/*
4559 * Checks whether it makes sense to retry the reclaim to make a forward progress
4560 * for the given allocation request.
4561 *
4562 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4563 * without success, or when we couldn't even meet the watermark if we
4564 * reclaimed all remaining pages on the LRU lists.
4565 *
4566 * Returns true if a retry is viable or false to enter the oom path.
4567 */
4568static inline bool
4569should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4570 struct alloc_context *ac, int alloc_flags,
4571 bool did_some_progress, int *no_progress_loops)
4572{
4573 struct zone *zone;
4574 struct zoneref *z;
4575 bool ret = false;
4576
4577 /*
4578 * Costly allocations might have made a progress but this doesn't mean
4579 * their order will become available due to high fragmentation so
4580 * always increment the no progress counter for them
4581 */
4582 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4583 *no_progress_loops = 0;
4584 else
4585 (*no_progress_loops)++;
4586
4587 if (*no_progress_loops > MAX_RECLAIM_RETRIES)
4588 goto out;
4589
4590
4591 /*
4592 * Keep reclaiming pages while there is a chance this will lead
4593 * somewhere. If none of the target zones can satisfy our allocation
4594 * request even if all reclaimable pages are considered then we are
4595 * screwed and have to go OOM.
4596 */
4597 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4598 ac->highest_zoneidx, ac->nodemask) {
4599 unsigned long available;
4600 unsigned long reclaimable;
4601 unsigned long min_wmark = min_wmark_pages(zone);
4602 bool wmark;
4603
4604 if (cpusets_enabled() &&
4605 (alloc_flags & ALLOC_CPUSET) &&
4606 !__cpuset_zone_allowed(zone, gfp_mask))
4607 continue;
4608
4609 available = reclaimable = zone_reclaimable_pages(zone);
4610 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4611
4612 /*
4613 * Would the allocation succeed if we reclaimed all
4614 * reclaimable pages?
4615 */
4616 wmark = __zone_watermark_ok(zone, order, min_wmark,
4617 ac->highest_zoneidx, alloc_flags, available);
4618 trace_reclaim_retry_zone(z, order, reclaimable,
4619 available, min_wmark, *no_progress_loops, wmark);
4620 if (wmark) {
4621 ret = true;
4622 break;
4623 }
4624 }
4625
4626 /*
4627 * Memory allocation/reclaim might be called from a WQ context and the
4628 * current implementation of the WQ concurrency control doesn't
4629 * recognize that a particular WQ is congested if the worker thread is
4630 * looping without ever sleeping. Therefore we have to do a short sleep
4631 * here rather than calling cond_resched().
4632 */
4633 if (current->flags & PF_WQ_WORKER)
4634 schedule_timeout_uninterruptible(1);
4635 else
4636 cond_resched();
4637out:
4638 /* Before OOM, exhaust highatomic_reserve */
4639 if (!ret)
4640 return unreserve_highatomic_pageblock(ac, true);
4641
4642 return ret;
4643}
4644
4645static inline bool
4646check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4647{
4648 /*
4649 * It's possible that cpuset's mems_allowed and the nodemask from
4650 * mempolicy don't intersect. This should be normally dealt with by
4651 * policy_nodemask(), but it's possible to race with cpuset update in
4652 * such a way the check therein was true, and then it became false
4653 * before we got our cpuset_mems_cookie here.
4654 * This assumes that for all allocations, ac->nodemask can come only
4655 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4656 * when it does not intersect with the cpuset restrictions) or the
4657 * caller can deal with a violated nodemask.
4658 */
4659 if (cpusets_enabled() && ac->nodemask &&
4660 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4661 ac->nodemask = NULL;
4662 return true;
4663 }
4664
4665 /*
4666 * When updating a task's mems_allowed or mempolicy nodemask, it is
4667 * possible to race with parallel threads in such a way that our
4668 * allocation can fail while the mask is being updated. If we are about
4669 * to fail, check if the cpuset changed during allocation and if so,
4670 * retry.
4671 */
4672 if (read_mems_allowed_retry(cpuset_mems_cookie))
4673 return true;
4674
4675 return false;
4676}
4677
4678static inline struct page *
4679__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4680 struct alloc_context *ac)
4681{
4682 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4683 bool can_compact = can_direct_reclaim && gfp_compaction_allowed(gfp_mask);
4684 bool nofail = gfp_mask & __GFP_NOFAIL;
4685 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4686 struct page *page = NULL;
4687 unsigned int alloc_flags;
4688 unsigned long did_some_progress;
4689 enum compact_priority compact_priority;
4690 enum compact_result compact_result;
4691 int compaction_retries;
4692 int no_progress_loops;
4693 unsigned int cpuset_mems_cookie;
4694 unsigned int zonelist_iter_cookie;
4695 int reserve_flags;
4696 bool compact_first = false;
4697 bool can_retry_reserves = true;
4698
4699 if (unlikely(nofail)) {
4700 /*
4701 * Also we don't support __GFP_NOFAIL without __GFP_DIRECT_RECLAIM,
4702 * otherwise, we may result in lockup.
4703 */
4704 WARN_ON_ONCE(!can_direct_reclaim);
4705 /*
4706 * PF_MEMALLOC request from this context is rather bizarre
4707 * because we cannot reclaim anything and only can loop waiting
4708 * for somebody to do a work for us.
4709 */
4710 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4711 }
4712
4713restart:
4714 compaction_retries = 0;
4715 no_progress_loops = 0;
4716 compact_result = COMPACT_SKIPPED;
4717 compact_priority = DEF_COMPACT_PRIORITY;
4718 cpuset_mems_cookie = read_mems_allowed_begin();
4719 zonelist_iter_cookie = zonelist_iter_begin();
4720
4721 /*
4722 * For costly allocations, try direct compaction first, as it's likely
4723 * that we have enough base pages and don't need to reclaim. For non-
4724 * movable high-order allocations, do that as well, as compaction will
4725 * try prevent permanent fragmentation by migrating from blocks of the
4726 * same migratetype.
4727 */
4728 if (can_compact && (costly_order || (order > 0 &&
4729 ac->migratetype != MIGRATE_MOVABLE))) {
4730 compact_first = true;
4731 compact_priority = INIT_COMPACT_PRIORITY;
4732 }
4733
4734 /*
4735 * The fast path uses conservative alloc_flags to succeed only until
4736 * kswapd needs to be woken up, and to avoid the cost of setting up
4737 * alloc_flags precisely. So we do that now.
4738 */
4739 alloc_flags = gfp_to_alloc_flags(gfp_mask, order);
4740
4741 /*
4742 * We need to recalculate the starting point for the zonelist iterator
4743 * because we might have used different nodemask in the fast path, or
4744 * there was a cpuset modification and we are retrying - otherwise we
4745 * could end up iterating over non-eligible zones endlessly.
4746 */
4747 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4748 ac->highest_zoneidx, ac->nodemask);
4749 if (!zonelist_zone(ac->preferred_zoneref))
4750 goto nopage;
4751
4752 /*
4753 * Check for insane configurations where the cpuset doesn't contain
4754 * any suitable zone to satisfy the request - e.g. non-movable
4755 * GFP_HIGHUSER allocations from MOVABLE nodes only.
4756 */
4757 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
4758 struct zoneref *z = first_zones_zonelist(ac->zonelist,
4759 ac->highest_zoneidx,
4760 &cpuset_current_mems_allowed);
4761 if (!zonelist_zone(z))
4762 goto nopage;
4763 }
4764
4765retry:
4766 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4767 if (alloc_flags & ALLOC_KSWAPD)
4768 wake_all_kswapds(order, gfp_mask, ac);
4769
4770 /*
4771 * The adjusted alloc_flags might result in immediate success, so try
4772 * that first
4773 */
4774 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4775 if (page)
4776 goto got_pg;
4777
4778 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4779 if (reserve_flags)
4780 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) |
4781 (alloc_flags & ALLOC_KSWAPD);
4782
4783 /*
4784 * Reset the nodemask and zonelist iterators if memory policies can be
4785 * ignored. These allocations are high priority and system rather than
4786 * user oriented.
4787 */
4788 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4789 ac->nodemask = NULL;
4790 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4791 ac->highest_zoneidx, ac->nodemask);
4792
4793 /*
4794 * The first time we adjust anything due to being allowed to
4795 * ignore memory policies or watermarks, retry immediately. This
4796 * allows us to keep the first allocation attempt optimistic so
4797 * it can succeed in a zone that is still above watermarks.
4798 */
4799 if (can_retry_reserves) {
4800 can_retry_reserves = false;
4801 goto retry;
4802 }
4803 }
4804
4805 /* Caller is not willing to reclaim, we can't balance anything */
4806 if (!can_direct_reclaim)
4807 goto nopage;
4808
4809 /* Avoid recursion of direct reclaim */
4810 if (current->flags & PF_MEMALLOC)
4811 goto nopage;
4812
4813 /* Try direct reclaim and then allocating */
4814 if (!compact_first) {
4815 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags,
4816 ac, &did_some_progress);
4817 if (page)
4818 goto got_pg;
4819 }
4820
4821 /* Try direct compaction and then allocating */
4822 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4823 compact_priority, &compact_result);
4824 if (page)
4825 goto got_pg;
4826
4827 if (compact_first) {
4828 /*
4829 * THP page faults may attempt local node only first, but are
4830 * then allowed to only compact, not reclaim, see
4831 * alloc_pages_mpol().
4832 *
4833 * Compaction has failed above and we don't want such THP
4834 * allocations to put reclaim pressure on a single node in a
4835 * situation where other nodes might have plenty of available
4836 * memory.
4837 */
4838 if (gfp_has_flags(gfp_mask, __GFP_NORETRY | __GFP_THISNODE))
4839 goto nopage;
4840
4841 /*
4842 * For the initial compaction attempt we have lowered its
4843 * priority. Restore it for further retries, if those are
4844 * allowed. With __GFP_NORETRY there will be a single round of
4845 * reclaim and compaction with the lowered priority.
4846 */
4847 if (!(gfp_mask & __GFP_NORETRY))
4848 compact_priority = DEF_COMPACT_PRIORITY;
4849
4850 compact_first = false;
4851 goto retry;
4852 }
4853
4854 /* Do not loop if specifically requested */
4855 if (gfp_mask & __GFP_NORETRY)
4856 goto nopage;
4857
4858 /*
4859 * Do not retry costly high order allocations unless they are
4860 * __GFP_RETRY_MAYFAIL and we can compact
4861 */
4862 if (costly_order && (!can_compact ||
4863 !(gfp_mask & __GFP_RETRY_MAYFAIL)))
4864 goto nopage;
4865
4866 /*
4867 * Deal with possible cpuset update races or zonelist updates to avoid
4868 * infinite retries. No "goto retry;" can be placed above this check
4869 * unless it can execute just once.
4870 */
4871 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4872 check_retry_zonelist(zonelist_iter_cookie))
4873 goto restart;
4874
4875 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4876 did_some_progress > 0, &no_progress_loops))
4877 goto retry;
4878
4879 /*
4880 * It doesn't make any sense to retry for the compaction if the order-0
4881 * reclaim is not able to make any progress because the current
4882 * implementation of the compaction depends on the sufficient amount
4883 * of free memory (see __compaction_suitable)
4884 */
4885 if (did_some_progress > 0 && can_compact &&
4886 should_compact_retry(ac, order, alloc_flags,
4887 compact_result, &compact_priority,
4888 &compaction_retries))
4889 goto retry;
4890
4891 /* Reclaim/compaction failed to prevent the fallback */
4892 if (defrag_mode && (alloc_flags & ALLOC_NOFRAGMENT)) {
4893 alloc_flags &= ~ALLOC_NOFRAGMENT;
4894 goto retry;
4895 }
4896
4897 /*
4898 * Deal with possible cpuset update races or zonelist updates to avoid
4899 * a unnecessary OOM kill.
4900 */
4901 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4902 check_retry_zonelist(zonelist_iter_cookie))
4903 goto restart;
4904
4905 /* Reclaim has failed us, start killing things */
4906 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4907 if (page)
4908 goto got_pg;
4909
4910 /* Avoid allocations with no watermarks from looping endlessly */
4911 if (tsk_is_oom_victim(current) &&
4912 (alloc_flags & ALLOC_OOM ||
4913 (gfp_mask & __GFP_NOMEMALLOC)))
4914 goto nopage;
4915
4916 /* Retry as long as the OOM killer is making progress */
4917 if (did_some_progress) {
4918 no_progress_loops = 0;
4919 goto retry;
4920 }
4921
4922nopage:
4923 /*
4924 * Deal with possible cpuset update races or zonelist updates to avoid
4925 * a unnecessary OOM kill.
4926 */
4927 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4928 check_retry_zonelist(zonelist_iter_cookie))
4929 goto restart;
4930
4931 /*
4932 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4933 * we always retry
4934 */
4935 if (unlikely(nofail)) {
4936 /*
4937 * Lacking direct_reclaim we can't do anything to reclaim memory,
4938 * we disregard these unreasonable nofail requests and still
4939 * return NULL
4940 */
4941 if (!can_direct_reclaim)
4942 goto fail;
4943
4944 /*
4945 * Help non-failing allocations by giving some access to memory
4946 * reserves normally used for high priority non-blocking
4947 * allocations but do not use ALLOC_NO_WATERMARKS because this
4948 * could deplete whole memory reserves which would just make
4949 * the situation worse.
4950 */
4951 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_MIN_RESERVE, ac);
4952 if (page)
4953 goto got_pg;
4954
4955 cond_resched();
4956 goto retry;
4957 }
4958fail:
4959 warn_alloc(gfp_mask, ac->nodemask,
4960 "page allocation failure: order:%u", order);
4961got_pg:
4962 return page;
4963}
4964
4965static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4966 int preferred_nid, nodemask_t *nodemask,
4967 struct alloc_context *ac, gfp_t *alloc_gfp,
4968 unsigned int *alloc_flags)
4969{
4970 ac->highest_zoneidx = gfp_zone(gfp_mask);
4971 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4972 ac->nodemask = nodemask;
4973 ac->migratetype = gfp_migratetype(gfp_mask);
4974
4975 if (cpusets_enabled()) {
4976 *alloc_gfp |= __GFP_HARDWALL;
4977 /*
4978 * When we are in the interrupt context, it is irrelevant
4979 * to the current task context. It means that any node ok.
4980 */
4981 if (in_task() && !ac->nodemask)
4982 ac->nodemask = &cpuset_current_mems_allowed;
4983 else
4984 *alloc_flags |= ALLOC_CPUSET;
4985 }
4986
4987 might_alloc(gfp_mask);
4988
4989 /*
4990 * Don't invoke should_fail logic, since it may call
4991 * get_random_u32() and printk() which need to spin_lock.
4992 */
4993 if (!(*alloc_flags & ALLOC_TRYLOCK) &&
4994 should_fail_alloc_page(gfp_mask, order))
4995 return false;
4996
4997 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
4998
4999 /* Dirty zone balancing only done in the fast path */
5000 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5001
5002 /*
5003 * The preferred zone is used for statistics but crucially it is
5004 * also used as the starting point for the zonelist iterator. It
5005 * may get reset for allocations that ignore memory policies.
5006 */
5007 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5008 ac->highest_zoneidx, ac->nodemask);
5009
5010 return true;
5011}
5012
5013/*
5014 * __alloc_pages_bulk - Allocate a number of order-0 pages to an array
5015 * @gfp: GFP flags for the allocation
5016 * @preferred_nid: The preferred NUMA node ID to allocate from
5017 * @nodemask: Set of nodes to allocate from, may be NULL
5018 * @nr_pages: The number of pages desired in the array
5019 * @page_array: Array to store the pages
5020 *
5021 * This is a batched version of the page allocator that attempts to allocate
5022 * @nr_pages quickly. Pages are added to @page_array.
5023 *
5024 * Note that only the elements in @page_array that were cleared to %NULL on
5025 * entry are populated with newly allocated pages. @nr_pages is the maximum
5026 * number of pages that will be stored in the array.
5027 *
5028 * Returns the number of pages in @page_array, including ones already
5029 * allocated on entry. This can be less than the number requested in @nr_pages,
5030 * but all empty slots are filled from the beginning. I.e., if all slots in
5031 * @page_array were set to %NULL on entry, the slots from 0 to the return value
5032 * - 1 will be filled.
5033 */
5034unsigned long alloc_pages_bulk_noprof(gfp_t gfp, int preferred_nid,
5035 nodemask_t *nodemask, int nr_pages,
5036 struct page **page_array)
5037{
5038 struct page *page;
5039 struct zone *zone;
5040 struct zoneref *z;
5041 struct per_cpu_pages *pcp;
5042 struct list_head *pcp_list;
5043 struct alloc_context ac;
5044 gfp_t alloc_gfp;
5045 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5046 int nr_populated = 0, nr_account = 0;
5047
5048 /*
5049 * Skip populated array elements to determine if any pages need
5050 * to be allocated before disabling IRQs.
5051 */
5052 while (nr_populated < nr_pages && page_array[nr_populated])
5053 nr_populated++;
5054
5055 /* No pages requested? */
5056 if (unlikely(nr_pages <= 0))
5057 goto out;
5058
5059 /* Already populated array? */
5060 if (unlikely(nr_pages - nr_populated == 0))
5061 goto out;
5062
5063 /* Bulk allocator does not support memcg accounting. */
5064 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT))
5065 goto failed;
5066
5067 /* Use the single page allocator for one page. */
5068 if (nr_pages - nr_populated == 1)
5069 goto failed;
5070
5071#ifdef CONFIG_PAGE_OWNER
5072 /*
5073 * PAGE_OWNER may recurse into the allocator to allocate space to
5074 * save the stack with pagesets.lock held. Releasing/reacquiring
5075 * removes much of the performance benefit of bulk allocation so
5076 * force the caller to allocate one page at a time as it'll have
5077 * similar performance to added complexity to the bulk allocator.
5078 */
5079 if (static_branch_unlikely(&page_owner_inited))
5080 goto failed;
5081#endif
5082
5083 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5084 gfp &= gfp_allowed_mask;
5085 alloc_gfp = gfp;
5086 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
5087 goto out;
5088 gfp = alloc_gfp;
5089
5090 /* Find an allowed local zone that meets the low watermark. */
5091 z = ac.preferred_zoneref;
5092 for_next_zone_zonelist_nodemask(zone, z, ac.highest_zoneidx, ac.nodemask) {
5093 unsigned long mark;
5094
5095 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5096 !__cpuset_zone_allowed(zone, gfp)) {
5097 continue;
5098 }
5099
5100 if (nr_online_nodes > 1 && zone != zonelist_zone(ac.preferred_zoneref) &&
5101 zone_to_nid(zone) != zonelist_node_idx(ac.preferred_zoneref)) {
5102 goto failed;
5103 }
5104
5105 cond_accept_memory(zone, 0, alloc_flags);
5106retry_this_zone:
5107 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages - nr_populated;
5108 if (zone_watermark_fast(zone, 0, mark,
5109 zonelist_zone_idx(ac.preferred_zoneref),
5110 alloc_flags, gfp)) {
5111 break;
5112 }
5113
5114 if (cond_accept_memory(zone, 0, alloc_flags))
5115 goto retry_this_zone;
5116
5117 /* Try again if zone has deferred pages */
5118 if (deferred_pages_enabled()) {
5119 if (_deferred_grow_zone(zone, 0))
5120 goto retry_this_zone;
5121 }
5122 }
5123
5124 /*
5125 * If there are no allowed local zones that meets the watermarks then
5126 * try to allocate a single page and reclaim if necessary.
5127 */
5128 if (unlikely(!zone))
5129 goto failed;
5130
5131 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
5132 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
5133 if (!pcp)
5134 goto failed;
5135
5136 /* Attempt the batch allocation */
5137 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
5138 while (nr_populated < nr_pages) {
5139
5140 /* Skip existing pages */
5141 if (page_array[nr_populated]) {
5142 nr_populated++;
5143 continue;
5144 }
5145
5146 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
5147 pcp, pcp_list);
5148 if (unlikely(!page)) {
5149 /* Try and allocate at least one page */
5150 if (!nr_account) {
5151 pcp_spin_unlock(pcp);
5152 goto failed;
5153 }
5154 break;
5155 }
5156 nr_account++;
5157
5158 prep_new_page(page, 0, gfp, 0);
5159 set_page_refcounted(page);
5160 page_array[nr_populated++] = page;
5161 }
5162
5163 pcp_spin_unlock(pcp);
5164
5165 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
5166 zone_statistics(zonelist_zone(ac.preferred_zoneref), zone, nr_account);
5167
5168out:
5169 return nr_populated;
5170
5171failed:
5172 page = __alloc_pages_noprof(gfp, 0, preferred_nid, nodemask);
5173 if (page)
5174 page_array[nr_populated++] = page;
5175 goto out;
5176}
5177EXPORT_SYMBOL_GPL(alloc_pages_bulk_noprof);
5178
5179/*
5180 * This is the 'heart' of the zoned buddy allocator.
5181 */
5182struct page *__alloc_frozen_pages_noprof(gfp_t gfp, unsigned int order,
5183 int preferred_nid, nodemask_t *nodemask)
5184{
5185 struct page *page;
5186 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5187 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5188 struct alloc_context ac = { };
5189
5190 /*
5191 * There are several places where we assume that the order value is sane
5192 * so bail out early if the request is out of bound.
5193 */
5194 if (WARN_ON_ONCE_GFP(order > MAX_PAGE_ORDER, gfp))
5195 return NULL;
5196
5197 gfp &= gfp_allowed_mask;
5198 /*
5199 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5200 * resp. GFP_NOIO which has to be inherited for all allocation requests
5201 * from a particular context which has been marked by
5202 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5203 * movable zones are not used during allocation.
5204 */
5205 gfp = current_gfp_context(gfp);
5206 alloc_gfp = gfp;
5207 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5208 &alloc_gfp, &alloc_flags))
5209 return NULL;
5210
5211 /*
5212 * Forbid the first pass from falling back to types that fragment
5213 * memory until all local zones are considered.
5214 */
5215 alloc_flags |= alloc_flags_nofragment(zonelist_zone(ac.preferred_zoneref), gfp);
5216
5217 /* First allocation attempt */
5218 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5219 if (likely(page))
5220 goto out;
5221
5222 alloc_gfp = gfp;
5223 ac.spread_dirty_pages = false;
5224
5225 /*
5226 * Restore the original nodemask if it was potentially replaced with
5227 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5228 */
5229 ac.nodemask = nodemask;
5230
5231 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5232
5233out:
5234 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT) && page &&
5235 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5236 free_frozen_pages(page, order);
5237 page = NULL;
5238 }
5239
5240 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5241 kmsan_alloc_page(page, order, alloc_gfp);
5242
5243 return page;
5244}
5245EXPORT_SYMBOL(__alloc_frozen_pages_noprof);
5246
5247struct page *__alloc_pages_noprof(gfp_t gfp, unsigned int order,
5248 int preferred_nid, nodemask_t *nodemask)
5249{
5250 struct page *page;
5251
5252 page = __alloc_frozen_pages_noprof(gfp, order, preferred_nid, nodemask);
5253 if (page)
5254 set_page_refcounted(page);
5255 return page;
5256}
5257EXPORT_SYMBOL(__alloc_pages_noprof);
5258
5259struct folio *__folio_alloc_noprof(gfp_t gfp, unsigned int order, int preferred_nid,
5260 nodemask_t *nodemask)
5261{
5262 struct page *page = __alloc_pages_noprof(gfp | __GFP_COMP, order,
5263 preferred_nid, nodemask);
5264 return page_rmappable_folio(page);
5265}
5266EXPORT_SYMBOL(__folio_alloc_noprof);
5267
5268/*
5269 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5270 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5271 * you need to access high mem.
5272 */
5273unsigned long get_free_pages_noprof(gfp_t gfp_mask, unsigned int order)
5274{
5275 struct page *page;
5276
5277 page = alloc_pages_noprof(gfp_mask & ~__GFP_HIGHMEM, order);
5278 if (!page)
5279 return 0;
5280 return (unsigned long) page_address(page);
5281}
5282EXPORT_SYMBOL(get_free_pages_noprof);
5283
5284unsigned long get_zeroed_page_noprof(gfp_t gfp_mask)
5285{
5286 return get_free_pages_noprof(gfp_mask | __GFP_ZERO, 0);
5287}
5288EXPORT_SYMBOL(get_zeroed_page_noprof);
5289
5290static void ___free_pages(struct page *page, unsigned int order,
5291 fpi_t fpi_flags)
5292{
5293 /* get PageHead before we drop reference */
5294 int head = PageHead(page);
5295 /* get alloc tag in case the page is released by others */
5296 struct alloc_tag *tag = pgalloc_tag_get(page);
5297
5298 if (put_page_testzero(page))
5299 __free_frozen_pages(page, order, fpi_flags);
5300 else if (!head) {
5301 pgalloc_tag_sub_pages(tag, (1 << order) - 1);
5302 while (order-- > 0) {
5303 /*
5304 * The "tail" pages of this non-compound high-order
5305 * page will have no code tags, so to avoid warnings
5306 * mark them as empty.
5307 */
5308 clear_page_tag_ref(page + (1 << order));
5309 __free_frozen_pages(page + (1 << order), order,
5310 fpi_flags);
5311 }
5312 }
5313}
5314
5315/**
5316 * __free_pages - Free pages allocated with alloc_pages().
5317 * @page: The page pointer returned from alloc_pages().
5318 * @order: The order of the allocation.
5319 *
5320 * This function can free multi-page allocations that are not compound
5321 * pages. It does not check that the @order passed in matches that of
5322 * the allocation, so it is easy to leak memory. Freeing more memory
5323 * than was allocated will probably emit a warning.
5324 *
5325 * If the last reference to this page is speculative, it will be released
5326 * by put_page() which only frees the first page of a non-compound
5327 * allocation. To prevent the remaining pages from being leaked, we free
5328 * the subsequent pages here. If you want to use the page's reference
5329 * count to decide when to free the allocation, you should allocate a
5330 * compound page, and use put_page() instead of __free_pages().
5331 *
5332 * Context: May be called in interrupt context or while holding a normal
5333 * spinlock, but not in NMI context or while holding a raw spinlock.
5334 */
5335void __free_pages(struct page *page, unsigned int order)
5336{
5337 ___free_pages(page, order, FPI_NONE);
5338}
5339EXPORT_SYMBOL(__free_pages);
5340
5341/*
5342 * Can be called while holding raw_spin_lock or from IRQ and NMI for any
5343 * page type (not only those that came from alloc_pages_nolock)
5344 */
5345void free_pages_nolock(struct page *page, unsigned int order)
5346{
5347 ___free_pages(page, order, FPI_TRYLOCK);
5348}
5349
5350/**
5351 * free_pages - Free pages allocated with __get_free_pages().
5352 * @addr: The virtual address tied to a page returned from __get_free_pages().
5353 * @order: The order of the allocation.
5354 *
5355 * This function behaves the same as __free_pages(). Use this function
5356 * to free pages when you only have a valid virtual address. If you have
5357 * the page, call __free_pages() instead.
5358 */
5359void free_pages(unsigned long addr, unsigned int order)
5360{
5361 if (addr != 0) {
5362 VM_BUG_ON(!virt_addr_valid((void *)addr));
5363 __free_pages(virt_to_page((void *)addr), order);
5364 }
5365}
5366
5367EXPORT_SYMBOL(free_pages);
5368
5369static void *make_alloc_exact(unsigned long addr, unsigned int order,
5370 size_t size)
5371{
5372 if (addr) {
5373 unsigned long nr = DIV_ROUND_UP(size, PAGE_SIZE);
5374 struct page *page = virt_to_page((void *)addr);
5375 struct page *last = page + nr;
5376
5377 __split_page(page, order);
5378 while (page < --last)
5379 set_page_refcounted(last);
5380
5381 last = page + (1UL << order);
5382 for (page += nr; page < last; page++)
5383 __free_pages_ok(page, 0, FPI_TO_TAIL);
5384 }
5385 return (void *)addr;
5386}
5387
5388/**
5389 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5390 * @size: the number of bytes to allocate
5391 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5392 *
5393 * This function is similar to alloc_pages(), except that it allocates the
5394 * minimum number of pages to satisfy the request. alloc_pages() can only
5395 * allocate memory in power-of-two pages.
5396 *
5397 * This function is also limited by MAX_PAGE_ORDER.
5398 *
5399 * Memory allocated by this function must be released by free_pages_exact().
5400 *
5401 * Return: pointer to the allocated area or %NULL in case of error.
5402 */
5403void *alloc_pages_exact_noprof(size_t size, gfp_t gfp_mask)
5404{
5405 unsigned int order = get_order(size);
5406 unsigned long addr;
5407
5408 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5409 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5410
5411 addr = get_free_pages_noprof(gfp_mask, order);
5412 return make_alloc_exact(addr, order, size);
5413}
5414EXPORT_SYMBOL(alloc_pages_exact_noprof);
5415
5416/**
5417 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5418 * pages on a node.
5419 * @nid: the preferred node ID where memory should be allocated
5420 * @size: the number of bytes to allocate
5421 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5422 *
5423 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5424 * back.
5425 *
5426 * Return: pointer to the allocated area or %NULL in case of error.
5427 */
5428void * __meminit alloc_pages_exact_nid_noprof(int nid, size_t size, gfp_t gfp_mask)
5429{
5430 unsigned int order = get_order(size);
5431 struct page *p;
5432
5433 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5434 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5435
5436 p = alloc_pages_node_noprof(nid, gfp_mask, order);
5437 if (!p)
5438 return NULL;
5439 return make_alloc_exact((unsigned long)page_address(p), order, size);
5440}
5441
5442/**
5443 * free_pages_exact - release memory allocated via alloc_pages_exact()
5444 * @virt: the value returned by alloc_pages_exact.
5445 * @size: size of allocation, same value as passed to alloc_pages_exact().
5446 *
5447 * Release the memory allocated by a previous call to alloc_pages_exact.
5448 */
5449void free_pages_exact(void *virt, size_t size)
5450{
5451 unsigned long addr = (unsigned long)virt;
5452 unsigned long end = addr + PAGE_ALIGN(size);
5453
5454 while (addr < end) {
5455 free_page(addr);
5456 addr += PAGE_SIZE;
5457 }
5458}
5459EXPORT_SYMBOL(free_pages_exact);
5460
5461/**
5462 * nr_free_zone_pages - count number of pages beyond high watermark
5463 * @offset: The zone index of the highest zone
5464 *
5465 * nr_free_zone_pages() counts the number of pages which are beyond the
5466 * high watermark within all zones at or below a given zone index. For each
5467 * zone, the number of pages is calculated as:
5468 *
5469 * nr_free_zone_pages = managed_pages - high_pages
5470 *
5471 * Return: number of pages beyond high watermark.
5472 */
5473static unsigned long nr_free_zone_pages(int offset)
5474{
5475 struct zoneref *z;
5476 struct zone *zone;
5477
5478 /* Just pick one node, since fallback list is circular */
5479 unsigned long sum = 0;
5480
5481 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5482
5483 for_each_zone_zonelist(zone, z, zonelist, offset) {
5484 unsigned long size = zone_managed_pages(zone);
5485 unsigned long high = high_wmark_pages(zone);
5486 if (size > high)
5487 sum += size - high;
5488 }
5489
5490 return sum;
5491}
5492
5493/**
5494 * nr_free_buffer_pages - count number of pages beyond high watermark
5495 *
5496 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5497 * watermark within ZONE_DMA and ZONE_NORMAL.
5498 *
5499 * Return: number of pages beyond high watermark within ZONE_DMA and
5500 * ZONE_NORMAL.
5501 */
5502unsigned long nr_free_buffer_pages(void)
5503{
5504 return nr_free_zone_pages(gfp_zone(GFP_USER));
5505}
5506EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5507
5508static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5509{
5510 zoneref->zone = zone;
5511 zoneref->zone_idx = zone_idx(zone);
5512}
5513
5514/*
5515 * Builds allocation fallback zone lists.
5516 *
5517 * Add all populated zones of a node to the zonelist.
5518 */
5519static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5520{
5521 struct zone *zone;
5522 enum zone_type zone_type = MAX_NR_ZONES;
5523 int nr_zones = 0;
5524
5525 do {
5526 zone_type--;
5527 zone = pgdat->node_zones + zone_type;
5528 if (populated_zone(zone)) {
5529 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5530 check_highest_zone(zone_type);
5531 }
5532 } while (zone_type);
5533
5534 return nr_zones;
5535}
5536
5537#ifdef CONFIG_NUMA
5538
5539static int __parse_numa_zonelist_order(char *s)
5540{
5541 /*
5542 * We used to support different zonelists modes but they turned
5543 * out to be just not useful. Let's keep the warning in place
5544 * if somebody still use the cmd line parameter so that we do
5545 * not fail it silently
5546 */
5547 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5548 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5549 return -EINVAL;
5550 }
5551 return 0;
5552}
5553
5554static char numa_zonelist_order[] = "Node";
5555#define NUMA_ZONELIST_ORDER_LEN 16
5556/*
5557 * sysctl handler for numa_zonelist_order
5558 */
5559static int numa_zonelist_order_handler(const struct ctl_table *table, int write,
5560 void *buffer, size_t *length, loff_t *ppos)
5561{
5562 if (write)
5563 return __parse_numa_zonelist_order(buffer);
5564 return proc_dostring(table, write, buffer, length, ppos);
5565}
5566
5567static int node_load[MAX_NUMNODES];
5568
5569/**
5570 * find_next_best_node - find the next node that should appear in a given node's fallback list
5571 * @node: node whose fallback list we're appending
5572 * @used_node_mask: nodemask_t of already used nodes
5573 *
5574 * We use a number of factors to determine which is the next node that should
5575 * appear on a given node's fallback list. The node should not have appeared
5576 * already in @node's fallback list, and it should be the next closest node
5577 * according to the distance array (which contains arbitrary distance values
5578 * from each node to each node in the system), and should also prefer nodes
5579 * with no CPUs, since presumably they'll have very little allocation pressure
5580 * on them otherwise.
5581 *
5582 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5583 */
5584int find_next_best_node(int node, nodemask_t *used_node_mask)
5585{
5586 int n, val;
5587 int min_val = INT_MAX;
5588 int best_node = NUMA_NO_NODE;
5589
5590 /*
5591 * Use the local node if we haven't already, but for memoryless local
5592 * node, we should skip it and fall back to other nodes.
5593 */
5594 if (!node_isset(node, *used_node_mask) && node_state(node, N_MEMORY)) {
5595 node_set(node, *used_node_mask);
5596 return node;
5597 }
5598
5599 for_each_node_state(n, N_MEMORY) {
5600
5601 /* Don't want a node to appear more than once */
5602 if (node_isset(n, *used_node_mask))
5603 continue;
5604
5605 /* Use the distance array to find the distance */
5606 val = node_distance(node, n);
5607
5608 /* Penalize nodes under us ("prefer the next node") */
5609 val += (n < node);
5610
5611 /* Give preference to headless and unused nodes */
5612 if (!cpumask_empty(cpumask_of_node(n)))
5613 val += PENALTY_FOR_NODE_WITH_CPUS;
5614
5615 /* Slight preference for less loaded node */
5616 val *= MAX_NUMNODES;
5617 val += node_load[n];
5618
5619 if (val < min_val) {
5620 min_val = val;
5621 best_node = n;
5622 }
5623 }
5624
5625 if (best_node >= 0)
5626 node_set(best_node, *used_node_mask);
5627
5628 return best_node;
5629}
5630
5631
5632/*
5633 * Build zonelists ordered by node and zones within node.
5634 * This results in maximum locality--normal zone overflows into local
5635 * DMA zone, if any--but risks exhausting DMA zone.
5636 */
5637static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5638 unsigned nr_nodes)
5639{
5640 struct zoneref *zonerefs;
5641 int i;
5642
5643 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5644
5645 for (i = 0; i < nr_nodes; i++) {
5646 int nr_zones;
5647
5648 pg_data_t *node = NODE_DATA(node_order[i]);
5649
5650 nr_zones = build_zonerefs_node(node, zonerefs);
5651 zonerefs += nr_zones;
5652 }
5653 zonerefs->zone = NULL;
5654 zonerefs->zone_idx = 0;
5655}
5656
5657/*
5658 * Build __GFP_THISNODE zonelists
5659 */
5660static void build_thisnode_zonelists(pg_data_t *pgdat)
5661{
5662 struct zoneref *zonerefs;
5663 int nr_zones;
5664
5665 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5666 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5667 zonerefs += nr_zones;
5668 zonerefs->zone = NULL;
5669 zonerefs->zone_idx = 0;
5670}
5671
5672static void build_zonelists(pg_data_t *pgdat)
5673{
5674 static int node_order[MAX_NUMNODES];
5675 int node, nr_nodes = 0;
5676 nodemask_t used_mask = NODE_MASK_NONE;
5677 int local_node, prev_node;
5678
5679 /* NUMA-aware ordering of nodes */
5680 local_node = pgdat->node_id;
5681 prev_node = local_node;
5682
5683 memset(node_order, 0, sizeof(node_order));
5684 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5685 /*
5686 * We don't want to pressure a particular node.
5687 * So adding penalty to the first node in same
5688 * distance group to make it round-robin.
5689 */
5690 if (node_distance(local_node, node) !=
5691 node_distance(local_node, prev_node))
5692 node_load[node] += 1;
5693
5694 node_order[nr_nodes++] = node;
5695 prev_node = node;
5696 }
5697
5698 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5699 build_thisnode_zonelists(pgdat);
5700 pr_info("Fallback order for Node %d: ", local_node);
5701 for (node = 0; node < nr_nodes; node++)
5702 pr_cont("%d ", node_order[node]);
5703 pr_cont("\n");
5704}
5705
5706#ifdef CONFIG_HAVE_MEMORYLESS_NODES
5707/*
5708 * Return node id of node used for "local" allocations.
5709 * I.e., first node id of first zone in arg node's generic zonelist.
5710 * Used for initializing percpu 'numa_mem', which is used primarily
5711 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5712 */
5713int local_memory_node(int node)
5714{
5715 struct zoneref *z;
5716
5717 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5718 gfp_zone(GFP_KERNEL),
5719 NULL);
5720 return zonelist_node_idx(z);
5721}
5722#endif
5723
5724static void setup_min_unmapped_ratio(void);
5725static void setup_min_slab_ratio(void);
5726#else /* CONFIG_NUMA */
5727
5728static void build_zonelists(pg_data_t *pgdat)
5729{
5730 struct zoneref *zonerefs;
5731 int nr_zones;
5732
5733 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5734 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5735 zonerefs += nr_zones;
5736
5737 zonerefs->zone = NULL;
5738 zonerefs->zone_idx = 0;
5739}
5740
5741#endif /* CONFIG_NUMA */
5742
5743/*
5744 * Boot pageset table. One per cpu which is going to be used for all
5745 * zones and all nodes. The parameters will be set in such a way
5746 * that an item put on a list will immediately be handed over to
5747 * the buddy list. This is safe since pageset manipulation is done
5748 * with interrupts disabled.
5749 *
5750 * The boot_pagesets must be kept even after bootup is complete for
5751 * unused processors and/or zones. They do play a role for bootstrapping
5752 * hotplugged processors.
5753 *
5754 * zoneinfo_show() and maybe other functions do
5755 * not check if the processor is online before following the pageset pointer.
5756 * Other parts of the kernel may not check if the zone is available.
5757 */
5758static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
5759/* These effectively disable the pcplists in the boot pageset completely */
5760#define BOOT_PAGESET_HIGH 0
5761#define BOOT_PAGESET_BATCH 1
5762static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
5763static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
5764
5765static void __build_all_zonelists(void *data)
5766{
5767 int nid;
5768 int __maybe_unused cpu;
5769 pg_data_t *self = data;
5770 unsigned long flags;
5771
5772 /*
5773 * The zonelist_update_seq must be acquired with irqsave because the
5774 * reader can be invoked from IRQ with GFP_ATOMIC.
5775 */
5776 write_seqlock_irqsave(&zonelist_update_seq, flags);
5777 /*
5778 * Also disable synchronous printk() to prevent any printk() from
5779 * trying to hold port->lock, for
5780 * tty_insert_flip_string_and_push_buffer() on other CPU might be
5781 * calling kmalloc(GFP_ATOMIC | __GFP_NOWARN) with port->lock held.
5782 */
5783 printk_deferred_enter();
5784
5785#ifdef CONFIG_NUMA
5786 memset(node_load, 0, sizeof(node_load));
5787#endif
5788
5789 /*
5790 * This node is hotadded and no memory is yet present. So just
5791 * building zonelists is fine - no need to touch other nodes.
5792 */
5793 if (self && !node_online(self->node_id)) {
5794 build_zonelists(self);
5795 } else {
5796 /*
5797 * All possible nodes have pgdat preallocated
5798 * in free_area_init
5799 */
5800 for_each_node(nid) {
5801 pg_data_t *pgdat = NODE_DATA(nid);
5802
5803 build_zonelists(pgdat);
5804 }
5805
5806#ifdef CONFIG_HAVE_MEMORYLESS_NODES
5807 /*
5808 * We now know the "local memory node" for each node--
5809 * i.e., the node of the first zone in the generic zonelist.
5810 * Set up numa_mem percpu variable for on-line cpus. During
5811 * boot, only the boot cpu should be on-line; we'll init the
5812 * secondary cpus' numa_mem as they come on-line. During
5813 * node/memory hotplug, we'll fixup all on-line cpus.
5814 */
5815 for_each_online_cpu(cpu)
5816 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5817#endif
5818 }
5819
5820 printk_deferred_exit();
5821 write_sequnlock_irqrestore(&zonelist_update_seq, flags);
5822}
5823
5824static noinline void __init
5825build_all_zonelists_init(void)
5826{
5827 int cpu;
5828
5829 __build_all_zonelists(NULL);
5830
5831 /*
5832 * Initialize the boot_pagesets that are going to be used
5833 * for bootstrapping processors. The real pagesets for
5834 * each zone will be allocated later when the per cpu
5835 * allocator is available.
5836 *
5837 * boot_pagesets are used also for bootstrapping offline
5838 * cpus if the system is already booted because the pagesets
5839 * are needed to initialize allocators on a specific cpu too.
5840 * F.e. the percpu allocator needs the page allocator which
5841 * needs the percpu allocator in order to allocate its pagesets
5842 * (a chicken-egg dilemma).
5843 */
5844 for_each_possible_cpu(cpu)
5845 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
5846
5847 mminit_verify_zonelist();
5848 cpuset_init_current_mems_allowed();
5849}
5850
5851/*
5852 * unless system_state == SYSTEM_BOOTING.
5853 *
5854 * __ref due to call of __init annotated helper build_all_zonelists_init
5855 * [protected by SYSTEM_BOOTING].
5856 */
5857void __ref build_all_zonelists(pg_data_t *pgdat)
5858{
5859 unsigned long vm_total_pages;
5860
5861 if (system_state == SYSTEM_BOOTING) {
5862 build_all_zonelists_init();
5863 } else {
5864 __build_all_zonelists(pgdat);
5865 /* cpuset refresh routine should be here */
5866 }
5867 /* Get the number of free pages beyond high watermark in all zones. */
5868 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5869 /*
5870 * Disable grouping by mobility if the number of pages in the
5871 * system is too low to allow the mechanism to work. It would be
5872 * more accurate, but expensive to check per-zone. This check is
5873 * made on memory-hotadd so a system can start with mobility
5874 * disabled and enable it later
5875 */
5876 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5877 page_group_by_mobility_disabled = 1;
5878 else
5879 page_group_by_mobility_disabled = 0;
5880
5881 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5882 nr_online_nodes,
5883 str_off_on(page_group_by_mobility_disabled),
5884 vm_total_pages);
5885#ifdef CONFIG_NUMA
5886 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5887#endif
5888}
5889
5890static int zone_batchsize(struct zone *zone)
5891{
5892#ifdef CONFIG_MMU
5893 int batch;
5894
5895 /*
5896 * The number of pages to batch allocate is either ~0.025%
5897 * of the zone or 256KB, whichever is smaller. The batch
5898 * size is striking a balance between allocation latency
5899 * and zone lock contention.
5900 */
5901 batch = min(zone_managed_pages(zone) >> 12, SZ_256K / PAGE_SIZE);
5902 if (batch <= 1)
5903 return 1;
5904
5905 /*
5906 * Clamp the batch to a 2^n - 1 value. Having a power
5907 * of 2 value was found to be more likely to have
5908 * suboptimal cache aliasing properties in some cases.
5909 *
5910 * For example if 2 tasks are alternately allocating
5911 * batches of pages, one task can end up with a lot
5912 * of pages of one half of the possible page colors
5913 * and the other with pages of the other colors.
5914 */
5915 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5916
5917 return batch;
5918
5919#else
5920 /* The deferral and batching of frees should be suppressed under NOMMU
5921 * conditions.
5922 *
5923 * The problem is that NOMMU needs to be able to allocate large chunks
5924 * of contiguous memory as there's no hardware page translation to
5925 * assemble apparent contiguous memory from discontiguous pages.
5926 *
5927 * Queueing large contiguous runs of pages for batching, however,
5928 * causes the pages to actually be freed in smaller chunks. As there
5929 * can be a significant delay between the individual batches being
5930 * recycled, this leads to the once large chunks of space being
5931 * fragmented and becoming unavailable for high-order allocations.
5932 */
5933 return 1;
5934#endif
5935}
5936
5937static int percpu_pagelist_high_fraction;
5938static int zone_highsize(struct zone *zone, int batch, int cpu_online,
5939 int high_fraction)
5940{
5941#ifdef CONFIG_MMU
5942 int high;
5943 int nr_split_cpus;
5944 unsigned long total_pages;
5945
5946 if (!high_fraction) {
5947 /*
5948 * By default, the high value of the pcp is based on the zone
5949 * low watermark so that if they are full then background
5950 * reclaim will not be started prematurely.
5951 */
5952 total_pages = low_wmark_pages(zone);
5953 } else {
5954 /*
5955 * If percpu_pagelist_high_fraction is configured, the high
5956 * value is based on a fraction of the managed pages in the
5957 * zone.
5958 */
5959 total_pages = zone_managed_pages(zone) / high_fraction;
5960 }
5961
5962 /*
5963 * Split the high value across all online CPUs local to the zone. Note
5964 * that early in boot that CPUs may not be online yet and that during
5965 * CPU hotplug that the cpumask is not yet updated when a CPU is being
5966 * onlined. For memory nodes that have no CPUs, split the high value
5967 * across all online CPUs to mitigate the risk that reclaim is triggered
5968 * prematurely due to pages stored on pcp lists.
5969 */
5970 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
5971 if (!nr_split_cpus)
5972 nr_split_cpus = num_online_cpus();
5973 high = total_pages / nr_split_cpus;
5974
5975 /*
5976 * Ensure high is at least batch*4. The multiple is based on the
5977 * historical relationship between high and batch.
5978 */
5979 high = max(high, batch << 2);
5980
5981 return high;
5982#else
5983 return 0;
5984#endif
5985}
5986
5987/*
5988 * pcp->high and pcp->batch values are related and generally batch is lower
5989 * than high. They are also related to pcp->count such that count is lower
5990 * than high, and as soon as it reaches high, the pcplist is flushed.
5991 *
5992 * However, guaranteeing these relations at all times would require e.g. write
5993 * barriers here but also careful usage of read barriers at the read side, and
5994 * thus be prone to error and bad for performance. Thus the update only prevents
5995 * store tearing. Any new users of pcp->batch, pcp->high_min and pcp->high_max
5996 * should ensure they can cope with those fields changing asynchronously, and
5997 * fully trust only the pcp->count field on the local CPU with interrupts
5998 * disabled.
5999 *
6000 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6001 * outside of boot time (or some other assurance that no concurrent updaters
6002 * exist).
6003 */
6004static void pageset_update(struct per_cpu_pages *pcp, unsigned long high_min,
6005 unsigned long high_max, unsigned long batch)
6006{
6007 WRITE_ONCE(pcp->batch, batch);
6008 WRITE_ONCE(pcp->high_min, high_min);
6009 WRITE_ONCE(pcp->high_max, high_max);
6010}
6011
6012static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
6013{
6014 int pindex;
6015
6016 memset(pcp, 0, sizeof(*pcp));
6017 memset(pzstats, 0, sizeof(*pzstats));
6018
6019 spin_lock_init(&pcp->lock);
6020 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
6021 INIT_LIST_HEAD(&pcp->lists[pindex]);
6022
6023 /*
6024 * Set batch and high values safe for a boot pageset. A true percpu
6025 * pageset's initialization will update them subsequently. Here we don't
6026 * need to be as careful as pageset_update() as nobody can access the
6027 * pageset yet.
6028 */
6029 pcp->high_min = BOOT_PAGESET_HIGH;
6030 pcp->high_max = BOOT_PAGESET_HIGH;
6031 pcp->batch = BOOT_PAGESET_BATCH;
6032}
6033
6034static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high_min,
6035 unsigned long high_max, unsigned long batch)
6036{
6037 struct per_cpu_pages *pcp;
6038 int cpu;
6039
6040 for_each_possible_cpu(cpu) {
6041 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6042 pageset_update(pcp, high_min, high_max, batch);
6043 }
6044}
6045
6046/*
6047 * Calculate and set new high and batch values for all per-cpu pagesets of a
6048 * zone based on the zone's size.
6049 */
6050static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
6051{
6052 int new_high_min, new_high_max, new_batch;
6053
6054 new_batch = zone_batchsize(zone);
6055 if (percpu_pagelist_high_fraction) {
6056 new_high_min = zone_highsize(zone, new_batch, cpu_online,
6057 percpu_pagelist_high_fraction);
6058 /*
6059 * PCP high is tuned manually, disable auto-tuning via
6060 * setting high_min and high_max to the manual value.
6061 */
6062 new_high_max = new_high_min;
6063 } else {
6064 new_high_min = zone_highsize(zone, new_batch, cpu_online, 0);
6065 new_high_max = zone_highsize(zone, new_batch, cpu_online,
6066 MIN_PERCPU_PAGELIST_HIGH_FRACTION);
6067 }
6068
6069 if (zone->pageset_high_min == new_high_min &&
6070 zone->pageset_high_max == new_high_max &&
6071 zone->pageset_batch == new_batch)
6072 return;
6073
6074 zone->pageset_high_min = new_high_min;
6075 zone->pageset_high_max = new_high_max;
6076 zone->pageset_batch = new_batch;
6077
6078 __zone_set_pageset_high_and_batch(zone, new_high_min, new_high_max,
6079 new_batch);
6080}
6081
6082void __meminit setup_zone_pageset(struct zone *zone)
6083{
6084 int cpu;
6085
6086 /* Size may be 0 on !SMP && !NUMA */
6087 if (sizeof(struct per_cpu_zonestat) > 0)
6088 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
6089
6090 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
6091 for_each_possible_cpu(cpu) {
6092 struct per_cpu_pages *pcp;
6093 struct per_cpu_zonestat *pzstats;
6094
6095 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6096 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
6097 per_cpu_pages_init(pcp, pzstats);
6098 }
6099
6100 zone_set_pageset_high_and_batch(zone, 0);
6101}
6102
6103/*
6104 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6105 * page high values need to be recalculated.
6106 */
6107static void zone_pcp_update(struct zone *zone, int cpu_online)
6108{
6109 mutex_lock(&pcp_batch_high_lock);
6110 zone_set_pageset_high_and_batch(zone, cpu_online);
6111 mutex_unlock(&pcp_batch_high_lock);
6112}
6113
6114static void zone_pcp_update_cacheinfo(struct zone *zone, unsigned int cpu)
6115{
6116 struct per_cpu_pages *pcp;
6117 struct cpu_cacheinfo *cci;
6118
6119 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6120 cci = get_cpu_cacheinfo(cpu);
6121 /*
6122 * If data cache slice of CPU is large enough, "pcp->batch"
6123 * pages can be preserved in PCP before draining PCP for
6124 * consecutive high-order pages freeing without allocation.
6125 * This can reduce zone lock contention without hurting
6126 * cache-hot pages sharing.
6127 */
6128 pcp_spin_lock_nopin(pcp);
6129 if ((cci->per_cpu_data_slice_size >> PAGE_SHIFT) > 3 * pcp->batch)
6130 pcp->flags |= PCPF_FREE_HIGH_BATCH;
6131 else
6132 pcp->flags &= ~PCPF_FREE_HIGH_BATCH;
6133 pcp_spin_unlock_nopin(pcp);
6134}
6135
6136void setup_pcp_cacheinfo(unsigned int cpu)
6137{
6138 struct zone *zone;
6139
6140 for_each_populated_zone(zone)
6141 zone_pcp_update_cacheinfo(zone, cpu);
6142}
6143
6144/*
6145 * Allocate per cpu pagesets and initialize them.
6146 * Before this call only boot pagesets were available.
6147 */
6148void __init setup_per_cpu_pageset(void)
6149{
6150 struct pglist_data *pgdat;
6151 struct zone *zone;
6152 int __maybe_unused cpu;
6153
6154 for_each_populated_zone(zone)
6155 setup_zone_pageset(zone);
6156
6157#ifdef CONFIG_NUMA
6158 /*
6159 * Unpopulated zones continue using the boot pagesets.
6160 * The numa stats for these pagesets need to be reset.
6161 * Otherwise, they will end up skewing the stats of
6162 * the nodes these zones are associated with.
6163 */
6164 for_each_possible_cpu(cpu) {
6165 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
6166 memset(pzstats->vm_numa_event, 0,
6167 sizeof(pzstats->vm_numa_event));
6168 }
6169#endif
6170
6171 for_each_online_pgdat(pgdat)
6172 pgdat->per_cpu_nodestats =
6173 alloc_percpu(struct per_cpu_nodestat);
6174}
6175
6176__meminit void zone_pcp_init(struct zone *zone)
6177{
6178 /*
6179 * per cpu subsystem is not up at this point. The following code
6180 * relies on the ability of the linker to provide the
6181 * offset of a (static) per cpu variable into the per cpu area.
6182 */
6183 zone->per_cpu_pageset = &boot_pageset;
6184 zone->per_cpu_zonestats = &boot_zonestats;
6185 zone->pageset_high_min = BOOT_PAGESET_HIGH;
6186 zone->pageset_high_max = BOOT_PAGESET_HIGH;
6187 zone->pageset_batch = BOOT_PAGESET_BATCH;
6188
6189 if (populated_zone(zone))
6190 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
6191 zone->present_pages, zone_batchsize(zone));
6192}
6193
6194static void setup_per_zone_lowmem_reserve(void);
6195
6196void adjust_managed_page_count(struct page *page, long count)
6197{
6198 atomic_long_add(count, &page_zone(page)->managed_pages);
6199 totalram_pages_add(count);
6200 setup_per_zone_lowmem_reserve();
6201}
6202EXPORT_SYMBOL(adjust_managed_page_count);
6203
6204void free_reserved_page(struct page *page)
6205{
6206 clear_page_tag_ref(page);
6207 ClearPageReserved(page);
6208 init_page_count(page);
6209 __free_page(page);
6210 adjust_managed_page_count(page, 1);
6211}
6212EXPORT_SYMBOL(free_reserved_page);
6213
6214static int page_alloc_cpu_dead(unsigned int cpu)
6215{
6216 struct zone *zone;
6217
6218 lru_add_drain_cpu(cpu);
6219 mlock_drain_remote(cpu);
6220 drain_pages(cpu);
6221
6222 /*
6223 * Spill the event counters of the dead processor
6224 * into the current processors event counters.
6225 * This artificially elevates the count of the current
6226 * processor.
6227 */
6228 vm_events_fold_cpu(cpu);
6229
6230 /*
6231 * Zero the differential counters of the dead processor
6232 * so that the vm statistics are consistent.
6233 *
6234 * This is only okay since the processor is dead and cannot
6235 * race with what we are doing.
6236 */
6237 cpu_vm_stats_fold(cpu);
6238
6239 for_each_populated_zone(zone)
6240 zone_pcp_update(zone, 0);
6241
6242 return 0;
6243}
6244
6245static int page_alloc_cpu_online(unsigned int cpu)
6246{
6247 struct zone *zone;
6248
6249 for_each_populated_zone(zone)
6250 zone_pcp_update(zone, 1);
6251 return 0;
6252}
6253
6254void __init page_alloc_init_cpuhp(void)
6255{
6256 int ret;
6257
6258 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
6259 "mm/page_alloc:pcp",
6260 page_alloc_cpu_online,
6261 page_alloc_cpu_dead);
6262 WARN_ON(ret < 0);
6263}
6264
6265/*
6266 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6267 * or min_free_kbytes changes.
6268 */
6269static void calculate_totalreserve_pages(void)
6270{
6271 struct pglist_data *pgdat;
6272 unsigned long reserve_pages = 0;
6273 enum zone_type i, j;
6274
6275 for_each_online_pgdat(pgdat) {
6276
6277 pgdat->totalreserve_pages = 0;
6278
6279 for (i = 0; i < MAX_NR_ZONES; i++) {
6280 struct zone *zone = pgdat->node_zones + i;
6281 long max = 0;
6282 unsigned long managed_pages = zone_managed_pages(zone);
6283
6284 /*
6285 * lowmem_reserve[j] is monotonically non-decreasing
6286 * in j for a given zone (see
6287 * setup_per_zone_lowmem_reserve()). The maximum
6288 * valid reserve lives at the highest index with a
6289 * non-zero value, so scan backwards and stop at the
6290 * first hit.
6291 */
6292 for (j = MAX_NR_ZONES - 1; j > i; j--) {
6293 if (!zone->lowmem_reserve[j])
6294 continue;
6295
6296 max = zone->lowmem_reserve[j];
6297 break;
6298 }
6299 /* we treat the high watermark as reserved pages. */
6300 max += high_wmark_pages(zone);
6301
6302 max = min_t(unsigned long, max, managed_pages);
6303
6304 pgdat->totalreserve_pages += max;
6305
6306 reserve_pages += max;
6307 }
6308 }
6309 totalreserve_pages = reserve_pages;
6310 trace_mm_calculate_totalreserve_pages(totalreserve_pages);
6311}
6312
6313/*
6314 * setup_per_zone_lowmem_reserve - called whenever
6315 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6316 * has a correct pages reserved value, so an adequate number of
6317 * pages are left in the zone after a successful __alloc_pages().
6318 */
6319static void setup_per_zone_lowmem_reserve(void)
6320{
6321 struct pglist_data *pgdat;
6322 enum zone_type i, j;
6323 /*
6324 * For a given zone node_zones[i], lowmem_reserve[j] (j > i)
6325 * represents how many pages in zone i must effectively be kept
6326 * in reserve when deciding whether an allocation class that is
6327 * allowed to allocate from zones up to j may fall back into
6328 * zone i.
6329 *
6330 * As j increases, the allocation class can use a strictly larger
6331 * set of fallback zones and therefore must not be allowed to
6332 * deplete low zones more aggressively than a less flexible one.
6333 * As a result, lowmem_reserve[j] is required to be monotonically
6334 * non-decreasing in j for each zone i. Callers such as
6335 * calculate_totalreserve_pages() rely on this monotonicity when
6336 * selecting the maximum reserve entry.
6337 */
6338 for_each_online_pgdat(pgdat) {
6339 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
6340 struct zone *zone = &pgdat->node_zones[i];
6341 int ratio = sysctl_lowmem_reserve_ratio[i];
6342 bool clear = !ratio || !zone_managed_pages(zone);
6343 unsigned long managed_pages = 0;
6344
6345 for (j = i + 1; j < MAX_NR_ZONES; j++) {
6346 struct zone *upper_zone = &pgdat->node_zones[j];
6347
6348 managed_pages += zone_managed_pages(upper_zone);
6349
6350 if (clear)
6351 zone->lowmem_reserve[j] = 0;
6352 else
6353 zone->lowmem_reserve[j] = managed_pages / ratio;
6354 trace_mm_setup_per_zone_lowmem_reserve(zone, upper_zone,
6355 zone->lowmem_reserve[j]);
6356 }
6357 }
6358 }
6359
6360 /* update totalreserve_pages */
6361 calculate_totalreserve_pages();
6362}
6363
6364static void __setup_per_zone_wmarks(void)
6365{
6366 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6367 unsigned long lowmem_pages = 0;
6368 struct zone *zone;
6369 unsigned long flags;
6370
6371 /* Calculate total number of !ZONE_HIGHMEM and !ZONE_MOVABLE pages */
6372 for_each_zone(zone) {
6373 if (!is_highmem(zone) && zone_idx(zone) != ZONE_MOVABLE)
6374 lowmem_pages += zone_managed_pages(zone);
6375 }
6376
6377 for_each_zone(zone) {
6378 u64 tmp;
6379
6380 spin_lock_irqsave(&zone->lock, flags);
6381 tmp = (u64)pages_min * zone_managed_pages(zone);
6382 tmp = div64_ul(tmp, lowmem_pages);
6383 if (is_highmem(zone) || zone_idx(zone) == ZONE_MOVABLE) {
6384 /*
6385 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6386 * need highmem and movable zones pages, so cap pages_min
6387 * to a small value here.
6388 *
6389 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6390 * deltas control async page reclaim, and so should
6391 * not be capped for highmem and movable zones.
6392 */
6393 unsigned long min_pages;
6394
6395 min_pages = zone_managed_pages(zone) / 1024;
6396 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6397 zone->_watermark[WMARK_MIN] = min_pages;
6398 } else {
6399 /*
6400 * If it's a lowmem zone, reserve a number of pages
6401 * proportionate to the zone's size.
6402 */
6403 zone->_watermark[WMARK_MIN] = tmp;
6404 }
6405
6406 /*
6407 * Set the kswapd watermarks distance according to the
6408 * scale factor in proportion to available memory, but
6409 * ensure a minimum size on small systems.
6410 */
6411 tmp = max_t(u64, tmp >> 2,
6412 mult_frac(zone_managed_pages(zone),
6413 watermark_scale_factor, 10000));
6414
6415 zone->watermark_boost = 0;
6416 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6417 zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
6418 zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
6419 trace_mm_setup_per_zone_wmarks(zone);
6420
6421 spin_unlock_irqrestore(&zone->lock, flags);
6422 }
6423
6424 /* update totalreserve_pages */
6425 calculate_totalreserve_pages();
6426}
6427
6428/**
6429 * setup_per_zone_wmarks - called when min_free_kbytes changes
6430 * or when memory is hot-{added|removed}
6431 *
6432 * Ensures that the watermark[min,low,high] values for each zone are set
6433 * correctly with respect to min_free_kbytes.
6434 */
6435void setup_per_zone_wmarks(void)
6436{
6437 struct zone *zone;
6438 static DEFINE_SPINLOCK(lock);
6439
6440 spin_lock(&lock);
6441 __setup_per_zone_wmarks();
6442 spin_unlock(&lock);
6443
6444 /*
6445 * The watermark size have changed so update the pcpu batch
6446 * and high limits or the limits may be inappropriate.
6447 */
6448 for_each_zone(zone)
6449 zone_pcp_update(zone, 0);
6450}
6451
6452/*
6453 * Initialise min_free_kbytes.
6454 *
6455 * For small machines we want it small (128k min). For large machines
6456 * we want it large (256MB max). But it is not linear, because network
6457 * bandwidth does not increase linearly with machine size. We use
6458 *
6459 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6460 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6461 *
6462 * which yields
6463 *
6464 * 16MB: 512k
6465 * 32MB: 724k
6466 * 64MB: 1024k
6467 * 128MB: 1448k
6468 * 256MB: 2048k
6469 * 512MB: 2896k
6470 * 1024MB: 4096k
6471 * 2048MB: 5792k
6472 * 4096MB: 8192k
6473 * 8192MB: 11584k
6474 * 16384MB: 16384k
6475 */
6476void calculate_min_free_kbytes(void)
6477{
6478 unsigned long lowmem_kbytes;
6479 int new_min_free_kbytes;
6480
6481 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6482 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6483
6484 if (new_min_free_kbytes > user_min_free_kbytes)
6485 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
6486 else
6487 pr_warn_ratelimited("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6488 new_min_free_kbytes, user_min_free_kbytes);
6489
6490}
6491
6492int __meminit init_per_zone_wmark_min(void)
6493{
6494 calculate_min_free_kbytes();
6495 setup_per_zone_wmarks();
6496 refresh_zone_stat_thresholds();
6497 setup_per_zone_lowmem_reserve();
6498
6499#ifdef CONFIG_NUMA
6500 setup_min_unmapped_ratio();
6501 setup_min_slab_ratio();
6502#endif
6503
6504 khugepaged_min_free_kbytes_update();
6505
6506 return 0;
6507}
6508postcore_initcall(init_per_zone_wmark_min)
6509
6510/*
6511 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6512 * that we can call two helper functions whenever min_free_kbytes
6513 * changes.
6514 */
6515static int min_free_kbytes_sysctl_handler(const struct ctl_table *table, int write,
6516 void *buffer, size_t *length, loff_t *ppos)
6517{
6518 int rc;
6519
6520 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6521 if (rc)
6522 return rc;
6523
6524 if (write) {
6525 user_min_free_kbytes = min_free_kbytes;
6526 setup_per_zone_wmarks();
6527 }
6528 return 0;
6529}
6530
6531static int watermark_scale_factor_sysctl_handler(const struct ctl_table *table, int write,
6532 void *buffer, size_t *length, loff_t *ppos)
6533{
6534 int rc;
6535
6536 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6537 if (rc)
6538 return rc;
6539
6540 if (write)
6541 setup_per_zone_wmarks();
6542
6543 return 0;
6544}
6545
6546#ifdef CONFIG_NUMA
6547static void setup_min_unmapped_ratio(void)
6548{
6549 pg_data_t *pgdat;
6550 struct zone *zone;
6551
6552 for_each_online_pgdat(pgdat)
6553 pgdat->min_unmapped_pages = 0;
6554
6555 for_each_zone(zone)
6556 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
6557 sysctl_min_unmapped_ratio) / 100;
6558}
6559
6560
6561static int sysctl_min_unmapped_ratio_sysctl_handler(const struct ctl_table *table, int write,
6562 void *buffer, size_t *length, loff_t *ppos)
6563{
6564 int rc;
6565
6566 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6567 if (rc)
6568 return rc;
6569
6570 setup_min_unmapped_ratio();
6571
6572 return 0;
6573}
6574
6575static void setup_min_slab_ratio(void)
6576{
6577 pg_data_t *pgdat;
6578 struct zone *zone;
6579
6580 for_each_online_pgdat(pgdat)
6581 pgdat->min_slab_pages = 0;
6582
6583 for_each_zone(zone)
6584 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
6585 sysctl_min_slab_ratio) / 100;
6586}
6587
6588static int sysctl_min_slab_ratio_sysctl_handler(const struct ctl_table *table, int write,
6589 void *buffer, size_t *length, loff_t *ppos)
6590{
6591 int rc;
6592
6593 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6594 if (rc)
6595 return rc;
6596
6597 setup_min_slab_ratio();
6598
6599 return 0;
6600}
6601#endif
6602
6603/*
6604 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6605 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6606 * whenever sysctl_lowmem_reserve_ratio changes.
6607 *
6608 * The reserve ratio obviously has absolutely no relation with the
6609 * minimum watermarks. The lowmem reserve ratio can only make sense
6610 * if in function of the boot time zone sizes.
6611 */
6612static int lowmem_reserve_ratio_sysctl_handler(const struct ctl_table *table,
6613 int write, void *buffer, size_t *length, loff_t *ppos)
6614{
6615 int i;
6616
6617 proc_dointvec_minmax(table, write, buffer, length, ppos);
6618
6619 for (i = 0; i < MAX_NR_ZONES; i++) {
6620 if (sysctl_lowmem_reserve_ratio[i] < 1)
6621 sysctl_lowmem_reserve_ratio[i] = 0;
6622 }
6623
6624 setup_per_zone_lowmem_reserve();
6625 return 0;
6626}
6627
6628/*
6629 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
6630 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6631 * pagelist can have before it gets flushed back to buddy allocator.
6632 */
6633static int percpu_pagelist_high_fraction_sysctl_handler(const struct ctl_table *table,
6634 int write, void *buffer, size_t *length, loff_t *ppos)
6635{
6636 struct zone *zone;
6637 int old_percpu_pagelist_high_fraction;
6638 int ret;
6639
6640 /*
6641 * Avoid using pcp_batch_high_lock for reads as the value is read
6642 * atomically and a race with offlining is harmless.
6643 */
6644
6645 if (!write)
6646 return proc_dointvec_minmax(table, write, buffer, length, ppos);
6647
6648 mutex_lock(&pcp_batch_high_lock);
6649 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
6650
6651 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6652 if (ret < 0)
6653 goto out;
6654
6655 /* Sanity checking to avoid pcp imbalance */
6656 if (percpu_pagelist_high_fraction &&
6657 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
6658 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
6659 ret = -EINVAL;
6660 goto out;
6661 }
6662
6663 /* No change? */
6664 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
6665 goto out;
6666
6667 for_each_populated_zone(zone)
6668 zone_set_pageset_high_and_batch(zone, 0);
6669out:
6670 mutex_unlock(&pcp_batch_high_lock);
6671 return ret;
6672}
6673
6674static const struct ctl_table page_alloc_sysctl_table[] = {
6675 {
6676 .procname = "min_free_kbytes",
6677 .data = &min_free_kbytes,
6678 .maxlen = sizeof(min_free_kbytes),
6679 .mode = 0644,
6680 .proc_handler = min_free_kbytes_sysctl_handler,
6681 .extra1 = SYSCTL_ZERO,
6682 },
6683 {
6684 .procname = "watermark_boost_factor",
6685 .data = &watermark_boost_factor,
6686 .maxlen = sizeof(watermark_boost_factor),
6687 .mode = 0644,
6688 .proc_handler = proc_dointvec_minmax,
6689 .extra1 = SYSCTL_ZERO,
6690 },
6691 {
6692 .procname = "watermark_scale_factor",
6693 .data = &watermark_scale_factor,
6694 .maxlen = sizeof(watermark_scale_factor),
6695 .mode = 0644,
6696 .proc_handler = watermark_scale_factor_sysctl_handler,
6697 .extra1 = SYSCTL_ONE,
6698 .extra2 = SYSCTL_THREE_THOUSAND,
6699 },
6700 {
6701 .procname = "defrag_mode",
6702 .data = &defrag_mode,
6703 .maxlen = sizeof(defrag_mode),
6704 .mode = 0644,
6705 .proc_handler = proc_dointvec_minmax,
6706 .extra1 = SYSCTL_ZERO,
6707 .extra2 = SYSCTL_ONE,
6708 },
6709 {
6710 .procname = "percpu_pagelist_high_fraction",
6711 .data = &percpu_pagelist_high_fraction,
6712 .maxlen = sizeof(percpu_pagelist_high_fraction),
6713 .mode = 0644,
6714 .proc_handler = percpu_pagelist_high_fraction_sysctl_handler,
6715 .extra1 = SYSCTL_ZERO,
6716 },
6717 {
6718 .procname = "lowmem_reserve_ratio",
6719 .data = &sysctl_lowmem_reserve_ratio,
6720 .maxlen = sizeof(sysctl_lowmem_reserve_ratio),
6721 .mode = 0644,
6722 .proc_handler = lowmem_reserve_ratio_sysctl_handler,
6723 },
6724#ifdef CONFIG_NUMA
6725 {
6726 .procname = "numa_zonelist_order",
6727 .data = &numa_zonelist_order,
6728 .maxlen = NUMA_ZONELIST_ORDER_LEN,
6729 .mode = 0644,
6730 .proc_handler = numa_zonelist_order_handler,
6731 },
6732 {
6733 .procname = "min_unmapped_ratio",
6734 .data = &sysctl_min_unmapped_ratio,
6735 .maxlen = sizeof(sysctl_min_unmapped_ratio),
6736 .mode = 0644,
6737 .proc_handler = sysctl_min_unmapped_ratio_sysctl_handler,
6738 .extra1 = SYSCTL_ZERO,
6739 .extra2 = SYSCTL_ONE_HUNDRED,
6740 },
6741 {
6742 .procname = "min_slab_ratio",
6743 .data = &sysctl_min_slab_ratio,
6744 .maxlen = sizeof(sysctl_min_slab_ratio),
6745 .mode = 0644,
6746 .proc_handler = sysctl_min_slab_ratio_sysctl_handler,
6747 .extra1 = SYSCTL_ZERO,
6748 .extra2 = SYSCTL_ONE_HUNDRED,
6749 },
6750#endif
6751};
6752
6753void __init page_alloc_sysctl_init(void)
6754{
6755 register_sysctl_init("vm", page_alloc_sysctl_table);
6756}
6757
6758#ifdef CONFIG_CONTIG_ALLOC
6759/* Usage: See admin-guide/dynamic-debug-howto.rst */
6760static void alloc_contig_dump_pages(struct list_head *page_list)
6761{
6762 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
6763
6764 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
6765 struct page *page;
6766
6767 dump_stack();
6768 list_for_each_entry(page, page_list, lru)
6769 dump_page(page, "migration failure");
6770 }
6771}
6772
6773/* [start, end) must belong to a single zone. */
6774static int __alloc_contig_migrate_range(struct compact_control *cc,
6775 unsigned long start, unsigned long end)
6776{
6777 /* This function is based on compact_zone() from compaction.c. */
6778 unsigned int nr_reclaimed;
6779 unsigned long pfn = start;
6780 unsigned int tries = 0;
6781 int ret = 0;
6782 struct migration_target_control mtc = {
6783 .nid = zone_to_nid(cc->zone),
6784 .gfp_mask = cc->gfp_mask,
6785 .reason = MR_CONTIG_RANGE,
6786 };
6787
6788 lru_cache_disable();
6789
6790 while (pfn < end || !list_empty(&cc->migratepages)) {
6791 if (fatal_signal_pending(current)) {
6792 ret = -EINTR;
6793 break;
6794 }
6795
6796 if (list_empty(&cc->migratepages)) {
6797 cc->nr_migratepages = 0;
6798 ret = isolate_migratepages_range(cc, pfn, end);
6799 if (ret && ret != -EAGAIN)
6800 break;
6801 pfn = cc->migrate_pfn;
6802 tries = 0;
6803 } else if (++tries == 5) {
6804 ret = -EBUSY;
6805 break;
6806 }
6807
6808 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6809 &cc->migratepages);
6810 cc->nr_migratepages -= nr_reclaimed;
6811
6812 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
6813 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
6814
6815 /*
6816 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
6817 * to retry again over this error, so do the same here.
6818 */
6819 if (ret == -ENOMEM)
6820 break;
6821 }
6822
6823 lru_cache_enable();
6824 if (ret < 0) {
6825 if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY)
6826 alloc_contig_dump_pages(&cc->migratepages);
6827 putback_movable_pages(&cc->migratepages);
6828 }
6829
6830 return (ret < 0) ? ret : 0;
6831}
6832
6833static void split_free_frozen_pages(struct list_head *list, gfp_t gfp_mask)
6834{
6835 int order;
6836
6837 for (order = 0; order < NR_PAGE_ORDERS; order++) {
6838 struct page *page, *next;
6839 int nr_pages = 1 << order;
6840
6841 list_for_each_entry_safe(page, next, &list[order], lru) {
6842 int i;
6843
6844 post_alloc_hook(page, order, gfp_mask);
6845 if (!order)
6846 continue;
6847
6848 __split_page(page, order);
6849
6850 /* Add all subpages to the order-0 head, in sequence. */
6851 list_del(&page->lru);
6852 for (i = 0; i < nr_pages; i++)
6853 list_add_tail(&page[i].lru, &list[0]);
6854 }
6855 }
6856}
6857
6858static int __alloc_contig_verify_gfp_mask(gfp_t gfp_mask, gfp_t *gfp_cc_mask)
6859{
6860 const gfp_t reclaim_mask = __GFP_IO | __GFP_FS | __GFP_RECLAIM;
6861 const gfp_t action_mask = __GFP_COMP | __GFP_RETRY_MAYFAIL | __GFP_NOWARN |
6862 __GFP_ZERO | __GFP_ZEROTAGS | __GFP_SKIP_ZERO |
6863 __GFP_SKIP_KASAN;
6864 const gfp_t cc_action_mask = __GFP_RETRY_MAYFAIL | __GFP_NOWARN;
6865
6866 /*
6867 * We are given the range to allocate; node, mobility and placement
6868 * hints are irrelevant at this point. We'll simply ignore them.
6869 */
6870 gfp_mask &= ~(GFP_ZONEMASK | __GFP_RECLAIMABLE | __GFP_WRITE |
6871 __GFP_HARDWALL | __GFP_THISNODE | __GFP_MOVABLE);
6872
6873 /*
6874 * We only support most reclaim flags (but not NOFAIL/NORETRY), and
6875 * selected action flags.
6876 */
6877 if (gfp_mask & ~(reclaim_mask | action_mask))
6878 return -EINVAL;
6879
6880 /*
6881 * Flags to control page compaction/migration/reclaim, to free up our
6882 * page range. Migratable pages are movable, __GFP_MOVABLE is implied
6883 * for them.
6884 *
6885 * Traditionally we always had __GFP_RETRY_MAYFAIL set, keep doing that
6886 * to not degrade callers.
6887 */
6888 *gfp_cc_mask = (gfp_mask & (reclaim_mask | cc_action_mask)) |
6889 __GFP_MOVABLE | __GFP_RETRY_MAYFAIL;
6890 return 0;
6891}
6892
6893static void __free_contig_frozen_range(unsigned long pfn, unsigned long nr_pages)
6894{
6895 for (; nr_pages--; pfn++)
6896 free_frozen_pages(pfn_to_page(pfn), 0);
6897}
6898
6899/**
6900 * alloc_contig_frozen_range() -- tries to allocate given range of frozen pages
6901 * @start: start PFN to allocate
6902 * @end: one-past-the-last PFN to allocate
6903 * @alloc_flags: allocation information
6904 * @gfp_mask: GFP mask. Node/zone/placement hints are ignored; only some
6905 * action and reclaim modifiers are supported. Reclaim modifiers
6906 * control allocation behavior during compaction/migration/reclaim.
6907 *
6908 * The PFN range does not have to be pageblock aligned. The PFN range must
6909 * belong to a single zone.
6910 *
6911 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
6912 * pageblocks in the range. Once isolated, the pageblocks should not
6913 * be modified by others.
6914 *
6915 * All frozen pages which PFN is in [start, end) are allocated for the
6916 * caller, and they could be freed with free_contig_frozen_range(),
6917 * free_frozen_pages() also could be used to free compound frozen pages
6918 * directly.
6919 *
6920 * Return: zero on success or negative error code.
6921 */
6922int alloc_contig_frozen_range_noprof(unsigned long start, unsigned long end,
6923 acr_flags_t alloc_flags, gfp_t gfp_mask)
6924{
6925 const unsigned int order = ilog2(end - start);
6926 unsigned long outer_start, outer_end;
6927 int ret = 0;
6928
6929 struct compact_control cc = {
6930 .nr_migratepages = 0,
6931 .order = -1,
6932 .zone = page_zone(pfn_to_page(start)),
6933 .mode = MIGRATE_SYNC,
6934 .ignore_skip_hint = true,
6935 .no_set_skip_hint = true,
6936 .alloc_contig = true,
6937 };
6938 INIT_LIST_HEAD(&cc.migratepages);
6939 enum pb_isolate_mode mode = (alloc_flags & ACR_FLAGS_CMA) ?
6940 PB_ISOLATE_MODE_CMA_ALLOC :
6941 PB_ISOLATE_MODE_OTHER;
6942
6943 /*
6944 * In contrast to the buddy, we allow for orders here that exceed
6945 * MAX_PAGE_ORDER, so we must manually make sure that we are not
6946 * exceeding the maximum folio order.
6947 */
6948 if (WARN_ON_ONCE((gfp_mask & __GFP_COMP) && order > MAX_FOLIO_ORDER))
6949 return -EINVAL;
6950
6951 gfp_mask = current_gfp_context(gfp_mask);
6952 if (__alloc_contig_verify_gfp_mask(gfp_mask, (gfp_t *)&cc.gfp_mask))
6953 return -EINVAL;
6954
6955 /*
6956 * What we do here is we mark all pageblocks in range as
6957 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6958 * have different sizes, and due to the way page allocator
6959 * work, start_isolate_page_range() has special handlings for this.
6960 *
6961 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6962 * migrate the pages from an unaligned range (ie. pages that
6963 * we are interested in). This will put all the pages in
6964 * range back to page allocator as MIGRATE_ISOLATE.
6965 *
6966 * When this is done, we take the pages in range from page
6967 * allocator removing them from the buddy system. This way
6968 * page allocator will never consider using them.
6969 *
6970 * This lets us mark the pageblocks back as
6971 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6972 * aligned range but not in the unaligned, original range are
6973 * put back to page allocator so that buddy can use them.
6974 */
6975
6976 ret = start_isolate_page_range(start, end, mode);
6977 if (ret)
6978 goto done;
6979
6980 drain_all_pages(cc.zone);
6981
6982 /*
6983 * In case of -EBUSY, we'd like to know which page causes problem.
6984 * So, just fall through. test_pages_isolated() has a tracepoint
6985 * which will report the busy page.
6986 *
6987 * It is possible that busy pages could become available before
6988 * the call to test_pages_isolated, and the range will actually be
6989 * allocated. So, if we fall through be sure to clear ret so that
6990 * -EBUSY is not accidentally used or returned to caller.
6991 */
6992 ret = __alloc_contig_migrate_range(&cc, start, end);
6993 if (ret && ret != -EBUSY)
6994 goto done;
6995
6996 /*
6997 * When in-use hugetlb pages are migrated, they may simply be released
6998 * back into the free hugepage pool instead of being returned to the
6999 * buddy system. After the migration of in-use huge pages is completed,
7000 * we will invoke replace_free_hugepage_folios() to ensure that these
7001 * hugepages are properly released to the buddy system.
7002 */
7003 ret = replace_free_hugepage_folios(start, end);
7004 if (ret)
7005 goto done;
7006
7007 /*
7008 * Pages from [start, end) are within a pageblock_nr_pages
7009 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7010 * more, all pages in [start, end) are free in page allocator.
7011 * What we are going to do is to allocate all pages from
7012 * [start, end) (that is remove them from page allocator).
7013 *
7014 * The only problem is that pages at the beginning and at the
7015 * end of interesting range may be not aligned with pages that
7016 * page allocator holds, ie. they can be part of higher order
7017 * pages. Because of this, we reserve the bigger range and
7018 * once this is done free the pages we are not interested in.
7019 *
7020 * We don't have to hold zone->lock here because the pages are
7021 * isolated thus they won't get removed from buddy.
7022 */
7023 outer_start = find_large_buddy(start);
7024
7025 /* Make sure the range is really isolated. */
7026 if (test_pages_isolated(outer_start, end, mode)) {
7027 ret = -EBUSY;
7028 goto done;
7029 }
7030
7031 /* Grab isolated pages from freelists. */
7032 outer_end = isolate_freepages_range(&cc, outer_start, end);
7033 if (!outer_end) {
7034 ret = -EBUSY;
7035 goto done;
7036 }
7037
7038 if (!(gfp_mask & __GFP_COMP)) {
7039 split_free_frozen_pages(cc.freepages, gfp_mask);
7040
7041 /* Free head and tail (if any) */
7042 if (start != outer_start)
7043 __free_contig_frozen_range(outer_start, start - outer_start);
7044 if (end != outer_end)
7045 __free_contig_frozen_range(end, outer_end - end);
7046 } else if (start == outer_start && end == outer_end && is_power_of_2(end - start)) {
7047 struct page *head = pfn_to_page(start);
7048
7049 check_new_pages(head, order);
7050 prep_new_page(head, order, gfp_mask, 0);
7051 } else {
7052 ret = -EINVAL;
7053 WARN(true, "PFN range: requested [%lu, %lu), allocated [%lu, %lu)\n",
7054 start, end, outer_start, outer_end);
7055 }
7056done:
7057 undo_isolate_page_range(start, end);
7058 return ret;
7059}
7060EXPORT_SYMBOL(alloc_contig_frozen_range_noprof);
7061
7062/**
7063 * alloc_contig_range() -- tries to allocate given range of pages
7064 * @start: start PFN to allocate
7065 * @end: one-past-the-last PFN to allocate
7066 * @alloc_flags: allocation information
7067 * @gfp_mask: GFP mask.
7068 *
7069 * This routine is a wrapper around alloc_contig_frozen_range(), it can't
7070 * be used to allocate compound pages, the refcount of each allocated page
7071 * will be set to one.
7072 *
7073 * All pages which PFN is in [start, end) are allocated for the caller,
7074 * and should be freed with free_contig_range() or by manually calling
7075 * __free_page() on each allocated page.
7076 *
7077 * Return: zero on success or negative error code.
7078 */
7079int alloc_contig_range_noprof(unsigned long start, unsigned long end,
7080 acr_flags_t alloc_flags, gfp_t gfp_mask)
7081{
7082 int ret;
7083
7084 if (WARN_ON(gfp_mask & __GFP_COMP))
7085 return -EINVAL;
7086
7087 ret = alloc_contig_frozen_range_noprof(start, end, alloc_flags, gfp_mask);
7088 if (!ret)
7089 set_pages_refcounted(pfn_to_page(start), end - start);
7090
7091 return ret;
7092}
7093EXPORT_SYMBOL(alloc_contig_range_noprof);
7094
7095static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
7096 unsigned long nr_pages, bool skip_hugetlb,
7097 bool *skipped_hugetlb)
7098{
7099 unsigned long end_pfn = start_pfn + nr_pages;
7100 struct page *page;
7101
7102 while (start_pfn < end_pfn) {
7103 unsigned long step = 1;
7104
7105 page = pfn_to_online_page(start_pfn);
7106 if (!page)
7107 return false;
7108
7109 if (page_zone(page) != z)
7110 return false;
7111
7112 if (page_is_unmovable(z, page, PB_ISOLATE_MODE_OTHER, &step))
7113 return false;
7114
7115 /*
7116 * Only consider ranges containing hugepages if those pages are
7117 * smaller than the requested contiguous region. e.g.:
7118 * Move 2MB pages to free up a 1GB range.
7119 * Don't move 1GB pages to free up a 2MB range.
7120 *
7121 * This makes contiguous allocation more reliable if multiple
7122 * hugepage sizes are used without causing needless movement.
7123 */
7124 if (PageHuge(page)) {
7125 unsigned int order;
7126
7127 if (skip_hugetlb) {
7128 *skipped_hugetlb = true;
7129 return false;
7130 }
7131
7132 page = compound_head(page);
7133 order = compound_order(page);
7134 if ((order >= MAX_FOLIO_ORDER) ||
7135 (nr_pages <= (1 << order)))
7136 return false;
7137 }
7138
7139 start_pfn += step;
7140 }
7141 return true;
7142}
7143
7144static bool zone_spans_last_pfn(const struct zone *zone,
7145 unsigned long start_pfn, unsigned long nr_pages)
7146{
7147 unsigned long last_pfn = start_pfn + nr_pages - 1;
7148
7149 return zone_spans_pfn(zone, last_pfn);
7150}
7151
7152/**
7153 * alloc_contig_frozen_pages() -- tries to find and allocate contiguous range of frozen pages
7154 * @nr_pages: Number of contiguous pages to allocate
7155 * @gfp_mask: GFP mask. Node/zone/placement hints limit the search; only some
7156 * action and reclaim modifiers are supported. Reclaim modifiers
7157 * control allocation behavior during compaction/migration/reclaim.
7158 * @nid: Target node
7159 * @nodemask: Mask for other possible nodes
7160 *
7161 * This routine is a wrapper around alloc_contig_frozen_range(). It scans over
7162 * zones on an applicable zonelist to find a contiguous pfn range which can then
7163 * be tried for allocation with alloc_contig_frozen_range(). This routine is
7164 * intended for allocation requests which can not be fulfilled with the buddy
7165 * allocator.
7166 *
7167 * The allocated memory is always aligned to a page boundary. If nr_pages is a
7168 * power of two, then allocated range is also guaranteed to be aligned to same
7169 * nr_pages (e.g. 1GB request would be aligned to 1GB).
7170 *
7171 * Allocated frozen pages need be freed with free_contig_frozen_range(),
7172 * or by manually calling free_frozen_pages() on each allocated frozen
7173 * non-compound page, for compound frozen pages could be freed with
7174 * free_frozen_pages() directly.
7175 *
7176 * Return: pointer to contiguous frozen pages on success, or NULL if not successful.
7177 */
7178struct page *alloc_contig_frozen_pages_noprof(unsigned long nr_pages,
7179 gfp_t gfp_mask, int nid, nodemask_t *nodemask)
7180{
7181 unsigned long ret, pfn, flags;
7182 struct zonelist *zonelist;
7183 struct zone *zone;
7184 struct zoneref *z;
7185 bool skip_hugetlb = true;
7186 bool skipped_hugetlb = false;
7187
7188retry:
7189 zonelist = node_zonelist(nid, gfp_mask);
7190 for_each_zone_zonelist_nodemask(zone, z, zonelist,
7191 gfp_zone(gfp_mask), nodemask) {
7192 spin_lock_irqsave(&zone->lock, flags);
7193
7194 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
7195 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
7196 if (pfn_range_valid_contig(zone, pfn, nr_pages,
7197 skip_hugetlb,
7198 &skipped_hugetlb)) {
7199 /*
7200 * We release the zone lock here because
7201 * alloc_contig_frozen_range() will also lock
7202 * the zone at some point. If there's an
7203 * allocation spinning on this lock, it may
7204 * win the race and cause allocation to fail.
7205 */
7206 spin_unlock_irqrestore(&zone->lock, flags);
7207 ret = alloc_contig_frozen_range_noprof(pfn,
7208 pfn + nr_pages,
7209 ACR_FLAGS_NONE,
7210 gfp_mask);
7211 if (!ret)
7212 return pfn_to_page(pfn);
7213 spin_lock_irqsave(&zone->lock, flags);
7214 }
7215 pfn += nr_pages;
7216 }
7217 spin_unlock_irqrestore(&zone->lock, flags);
7218 }
7219 /*
7220 * If we failed, retry the search, but treat regions with HugeTLB pages
7221 * as valid targets. This retains fast-allocations on first pass
7222 * without trying to migrate HugeTLB pages (which may fail). On the
7223 * second pass, we will try moving HugeTLB pages when those pages are
7224 * smaller than the requested contiguous region size.
7225 */
7226 if (skip_hugetlb && skipped_hugetlb) {
7227 skip_hugetlb = false;
7228 goto retry;
7229 }
7230 return NULL;
7231}
7232EXPORT_SYMBOL(alloc_contig_frozen_pages_noprof);
7233
7234/**
7235 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
7236 * @nr_pages: Number of contiguous pages to allocate
7237 * @gfp_mask: GFP mask.
7238 * @nid: Target node
7239 * @nodemask: Mask for other possible nodes
7240 *
7241 * This routine is a wrapper around alloc_contig_frozen_pages(), it can't
7242 * be used to allocate compound pages, the refcount of each allocated page
7243 * will be set to one.
7244 *
7245 * Allocated pages can be freed with free_contig_range() or by manually
7246 * calling __free_page() on each allocated page.
7247 *
7248 * Return: pointer to contiguous pages on success, or NULL if not successful.
7249 */
7250struct page *alloc_contig_pages_noprof(unsigned long nr_pages, gfp_t gfp_mask,
7251 int nid, nodemask_t *nodemask)
7252{
7253 struct page *page;
7254
7255 if (WARN_ON(gfp_mask & __GFP_COMP))
7256 return NULL;
7257
7258 page = alloc_contig_frozen_pages_noprof(nr_pages, gfp_mask, nid,
7259 nodemask);
7260 if (page)
7261 set_pages_refcounted(page, nr_pages);
7262
7263 return page;
7264}
7265EXPORT_SYMBOL(alloc_contig_pages_noprof);
7266
7267/**
7268 * free_contig_frozen_range() -- free the contiguous range of frozen pages
7269 * @pfn: start PFN to free
7270 * @nr_pages: Number of contiguous frozen pages to free
7271 *
7272 * This can be used to free the allocated compound/non-compound frozen pages.
7273 */
7274void free_contig_frozen_range(unsigned long pfn, unsigned long nr_pages)
7275{
7276 struct page *first_page = pfn_to_page(pfn);
7277 const unsigned int order = ilog2(nr_pages);
7278
7279 if (WARN_ON_ONCE(first_page != compound_head(first_page)))
7280 return;
7281
7282 if (PageHead(first_page)) {
7283 WARN_ON_ONCE(order != compound_order(first_page));
7284 free_frozen_pages(first_page, order);
7285 return;
7286 }
7287
7288 __free_contig_frozen_range(pfn, nr_pages);
7289}
7290EXPORT_SYMBOL(free_contig_frozen_range);
7291
7292/**
7293 * free_contig_range() -- free the contiguous range of pages
7294 * @pfn: start PFN to free
7295 * @nr_pages: Number of contiguous pages to free
7296 *
7297 * This can be only used to free the allocated non-compound pages.
7298 */
7299void free_contig_range(unsigned long pfn, unsigned long nr_pages)
7300{
7301 if (WARN_ON_ONCE(PageHead(pfn_to_page(pfn))))
7302 return;
7303
7304 for (; nr_pages--; pfn++)
7305 __free_page(pfn_to_page(pfn));
7306}
7307EXPORT_SYMBOL(free_contig_range);
7308#endif /* CONFIG_CONTIG_ALLOC */
7309
7310/*
7311 * Effectively disable pcplists for the zone by setting the high limit to 0
7312 * and draining all cpus. A concurrent page freeing on another CPU that's about
7313 * to put the page on pcplist will either finish before the drain and the page
7314 * will be drained, or observe the new high limit and skip the pcplist.
7315 *
7316 * Must be paired with a call to zone_pcp_enable().
7317 */
7318void zone_pcp_disable(struct zone *zone)
7319{
7320 mutex_lock(&pcp_batch_high_lock);
7321 __zone_set_pageset_high_and_batch(zone, 0, 0, 1);
7322 __drain_all_pages(zone, true);
7323}
7324
7325void zone_pcp_enable(struct zone *zone)
7326{
7327 __zone_set_pageset_high_and_batch(zone, zone->pageset_high_min,
7328 zone->pageset_high_max, zone->pageset_batch);
7329 mutex_unlock(&pcp_batch_high_lock);
7330}
7331
7332void zone_pcp_reset(struct zone *zone)
7333{
7334 int cpu;
7335 struct per_cpu_zonestat *pzstats;
7336
7337 if (zone->per_cpu_pageset != &boot_pageset) {
7338 for_each_online_cpu(cpu) {
7339 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
7340 drain_zonestat(zone, pzstats);
7341 }
7342 free_percpu(zone->per_cpu_pageset);
7343 zone->per_cpu_pageset = &boot_pageset;
7344 if (zone->per_cpu_zonestats != &boot_zonestats) {
7345 free_percpu(zone->per_cpu_zonestats);
7346 zone->per_cpu_zonestats = &boot_zonestats;
7347 }
7348 }
7349}
7350
7351#ifdef CONFIG_MEMORY_HOTREMOVE
7352/*
7353 * All pages in the range must be in a single zone, must not contain holes,
7354 * must span full sections, and must be isolated before calling this function.
7355 *
7356 * Returns the number of managed (non-PageOffline()) pages in the range: the
7357 * number of pages for which memory offlining code must adjust managed page
7358 * counters using adjust_managed_page_count().
7359 */
7360unsigned long __offline_isolated_pages(unsigned long start_pfn,
7361 unsigned long end_pfn)
7362{
7363 unsigned long already_offline = 0, flags;
7364 unsigned long pfn = start_pfn;
7365 struct page *page;
7366 struct zone *zone;
7367 unsigned int order;
7368
7369 offline_mem_sections(pfn, end_pfn);
7370 zone = page_zone(pfn_to_page(pfn));
7371 spin_lock_irqsave(&zone->lock, flags);
7372 while (pfn < end_pfn) {
7373 page = pfn_to_page(pfn);
7374 /*
7375 * The HWPoisoned page may be not in buddy system, and
7376 * page_count() is not 0.
7377 */
7378 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7379 pfn++;
7380 continue;
7381 }
7382 /*
7383 * At this point all remaining PageOffline() pages have a
7384 * reference count of 0 and can simply be skipped.
7385 */
7386 if (PageOffline(page)) {
7387 BUG_ON(page_count(page));
7388 BUG_ON(PageBuddy(page));
7389 already_offline++;
7390 pfn++;
7391 continue;
7392 }
7393
7394 BUG_ON(page_count(page));
7395 BUG_ON(!PageBuddy(page));
7396 VM_WARN_ON(get_pageblock_migratetype(page) != MIGRATE_ISOLATE);
7397 order = buddy_order(page);
7398 del_page_from_free_list(page, zone, order, MIGRATE_ISOLATE);
7399 pfn += (1 << order);
7400 }
7401 spin_unlock_irqrestore(&zone->lock, flags);
7402
7403 return end_pfn - start_pfn - already_offline;
7404}
7405#endif
7406
7407/*
7408 * This function returns a stable result only if called under zone lock.
7409 */
7410bool is_free_buddy_page(const struct page *page)
7411{
7412 unsigned long pfn = page_to_pfn(page);
7413 unsigned int order;
7414
7415 for (order = 0; order < NR_PAGE_ORDERS; order++) {
7416 const struct page *head = page - (pfn & ((1 << order) - 1));
7417
7418 if (PageBuddy(head) &&
7419 buddy_order_unsafe(head) >= order)
7420 break;
7421 }
7422
7423 return order <= MAX_PAGE_ORDER;
7424}
7425EXPORT_SYMBOL(is_free_buddy_page);
7426
7427#ifdef CONFIG_MEMORY_FAILURE
7428static inline void add_to_free_list(struct page *page, struct zone *zone,
7429 unsigned int order, int migratetype,
7430 bool tail)
7431{
7432 __add_to_free_list(page, zone, order, migratetype, tail);
7433 account_freepages(zone, 1 << order, migratetype);
7434}
7435
7436/*
7437 * Break down a higher-order page in sub-pages, and keep our target out of
7438 * buddy allocator.
7439 */
7440static void break_down_buddy_pages(struct zone *zone, struct page *page,
7441 struct page *target, int low, int high,
7442 int migratetype)
7443{
7444 unsigned long size = 1 << high;
7445 struct page *current_buddy;
7446
7447 while (high > low) {
7448 high--;
7449 size >>= 1;
7450
7451 if (target >= &page[size]) {
7452 current_buddy = page;
7453 page = page + size;
7454 } else {
7455 current_buddy = page + size;
7456 }
7457
7458 if (set_page_guard(zone, current_buddy, high))
7459 continue;
7460
7461 add_to_free_list(current_buddy, zone, high, migratetype, false);
7462 set_buddy_order(current_buddy, high);
7463 }
7464}
7465
7466/*
7467 * Take a page that will be marked as poisoned off the buddy allocator.
7468 */
7469bool take_page_off_buddy(struct page *page)
7470{
7471 struct zone *zone = page_zone(page);
7472 unsigned long pfn = page_to_pfn(page);
7473 unsigned long flags;
7474 unsigned int order;
7475 bool ret = false;
7476
7477 spin_lock_irqsave(&zone->lock, flags);
7478 for (order = 0; order < NR_PAGE_ORDERS; order++) {
7479 struct page *page_head = page - (pfn & ((1 << order) - 1));
7480 int page_order = buddy_order(page_head);
7481
7482 if (PageBuddy(page_head) && page_order >= order) {
7483 unsigned long pfn_head = page_to_pfn(page_head);
7484 int migratetype = get_pfnblock_migratetype(page_head,
7485 pfn_head);
7486
7487 del_page_from_free_list(page_head, zone, page_order,
7488 migratetype);
7489 break_down_buddy_pages(zone, page_head, page, 0,
7490 page_order, migratetype);
7491 SetPageHWPoisonTakenOff(page);
7492 ret = true;
7493 break;
7494 }
7495 if (page_count(page_head) > 0)
7496 break;
7497 }
7498 spin_unlock_irqrestore(&zone->lock, flags);
7499 return ret;
7500}
7501
7502/*
7503 * Cancel takeoff done by take_page_off_buddy().
7504 */
7505bool put_page_back_buddy(struct page *page)
7506{
7507 struct zone *zone = page_zone(page);
7508 unsigned long flags;
7509 bool ret = false;
7510
7511 spin_lock_irqsave(&zone->lock, flags);
7512 if (put_page_testzero(page)) {
7513 unsigned long pfn = page_to_pfn(page);
7514 int migratetype = get_pfnblock_migratetype(page, pfn);
7515
7516 ClearPageHWPoisonTakenOff(page);
7517 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
7518 if (TestClearPageHWPoison(page)) {
7519 ret = true;
7520 }
7521 }
7522 spin_unlock_irqrestore(&zone->lock, flags);
7523
7524 return ret;
7525}
7526#endif
7527
7528bool has_managed_zone(enum zone_type zone)
7529{
7530 struct pglist_data *pgdat;
7531
7532 for_each_online_pgdat(pgdat) {
7533 if (managed_zone(&pgdat->node_zones[zone]))
7534 return true;
7535 }
7536 return false;
7537}
7538
7539#ifdef CONFIG_UNACCEPTED_MEMORY
7540
7541static bool lazy_accept = true;
7542
7543static int __init accept_memory_parse(char *p)
7544{
7545 if (!strcmp(p, "lazy")) {
7546 lazy_accept = true;
7547 return 0;
7548 } else if (!strcmp(p, "eager")) {
7549 lazy_accept = false;
7550 return 0;
7551 } else {
7552 return -EINVAL;
7553 }
7554}
7555early_param("accept_memory", accept_memory_parse);
7556
7557static bool page_contains_unaccepted(struct page *page, unsigned int order)
7558{
7559 phys_addr_t start = page_to_phys(page);
7560
7561 return range_contains_unaccepted_memory(start, PAGE_SIZE << order);
7562}
7563
7564static void __accept_page(struct zone *zone, unsigned long *flags,
7565 struct page *page)
7566{
7567 list_del(&page->lru);
7568 account_freepages(zone, -MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
7569 __mod_zone_page_state(zone, NR_UNACCEPTED, -MAX_ORDER_NR_PAGES);
7570 __ClearPageUnaccepted(page);
7571 spin_unlock_irqrestore(&zone->lock, *flags);
7572
7573 accept_memory(page_to_phys(page), PAGE_SIZE << MAX_PAGE_ORDER);
7574
7575 __free_pages_ok(page, MAX_PAGE_ORDER, FPI_TO_TAIL);
7576}
7577
7578void accept_page(struct page *page)
7579{
7580 struct zone *zone = page_zone(page);
7581 unsigned long flags;
7582
7583 spin_lock_irqsave(&zone->lock, flags);
7584 if (!PageUnaccepted(page)) {
7585 spin_unlock_irqrestore(&zone->lock, flags);
7586 return;
7587 }
7588
7589 /* Unlocks zone->lock */
7590 __accept_page(zone, &flags, page);
7591}
7592
7593static bool try_to_accept_memory_one(struct zone *zone)
7594{
7595 unsigned long flags;
7596 struct page *page;
7597
7598 spin_lock_irqsave(&zone->lock, flags);
7599 page = list_first_entry_or_null(&zone->unaccepted_pages,
7600 struct page, lru);
7601 if (!page) {
7602 spin_unlock_irqrestore(&zone->lock, flags);
7603 return false;
7604 }
7605
7606 /* Unlocks zone->lock */
7607 __accept_page(zone, &flags, page);
7608
7609 return true;
7610}
7611
7612static bool cond_accept_memory(struct zone *zone, unsigned int order,
7613 int alloc_flags)
7614{
7615 long to_accept, wmark;
7616 bool ret = false;
7617
7618 if (list_empty(&zone->unaccepted_pages))
7619 return false;
7620
7621 /* Bailout, since try_to_accept_memory_one() needs to take a lock */
7622 if (alloc_flags & ALLOC_TRYLOCK)
7623 return false;
7624
7625 wmark = promo_wmark_pages(zone);
7626
7627 /*
7628 * Watermarks have not been initialized yet.
7629 *
7630 * Accepting one MAX_ORDER page to ensure progress.
7631 */
7632 if (!wmark)
7633 return try_to_accept_memory_one(zone);
7634
7635 /* How much to accept to get to promo watermark? */
7636 to_accept = wmark -
7637 (zone_page_state(zone, NR_FREE_PAGES) -
7638 __zone_watermark_unusable_free(zone, order, 0) -
7639 zone_page_state(zone, NR_UNACCEPTED));
7640
7641 while (to_accept > 0) {
7642 if (!try_to_accept_memory_one(zone))
7643 break;
7644 ret = true;
7645 to_accept -= MAX_ORDER_NR_PAGES;
7646 }
7647
7648 return ret;
7649}
7650
7651static bool __free_unaccepted(struct page *page)
7652{
7653 struct zone *zone = page_zone(page);
7654 unsigned long flags;
7655
7656 if (!lazy_accept)
7657 return false;
7658
7659 spin_lock_irqsave(&zone->lock, flags);
7660 list_add_tail(&page->lru, &zone->unaccepted_pages);
7661 account_freepages(zone, MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
7662 __mod_zone_page_state(zone, NR_UNACCEPTED, MAX_ORDER_NR_PAGES);
7663 __SetPageUnaccepted(page);
7664 spin_unlock_irqrestore(&zone->lock, flags);
7665
7666 return true;
7667}
7668
7669#else
7670
7671static bool page_contains_unaccepted(struct page *page, unsigned int order)
7672{
7673 return false;
7674}
7675
7676static bool cond_accept_memory(struct zone *zone, unsigned int order,
7677 int alloc_flags)
7678{
7679 return false;
7680}
7681
7682static bool __free_unaccepted(struct page *page)
7683{
7684 BUILD_BUG();
7685 return false;
7686}
7687
7688#endif /* CONFIG_UNACCEPTED_MEMORY */
7689
7690struct page *alloc_frozen_pages_nolock_noprof(gfp_t gfp_flags, int nid, unsigned int order)
7691{
7692 /*
7693 * Do not specify __GFP_DIRECT_RECLAIM, since direct claim is not allowed.
7694 * Do not specify __GFP_KSWAPD_RECLAIM either, since wake up of kswapd
7695 * is not safe in arbitrary context.
7696 *
7697 * These two are the conditions for gfpflags_allow_spinning() being true.
7698 *
7699 * Specify __GFP_NOWARN since failing alloc_pages_nolock() is not a reason
7700 * to warn. Also warn would trigger printk() which is unsafe from
7701 * various contexts. We cannot use printk_deferred_enter() to mitigate,
7702 * since the running context is unknown.
7703 *
7704 * Specify __GFP_ZERO to make sure that call to kmsan_alloc_page() below
7705 * is safe in any context. Also zeroing the page is mandatory for
7706 * BPF use cases.
7707 *
7708 * Though __GFP_NOMEMALLOC is not checked in the code path below,
7709 * specify it here to highlight that alloc_pages_nolock()
7710 * doesn't want to deplete reserves.
7711 */
7712 gfp_t alloc_gfp = __GFP_NOWARN | __GFP_ZERO | __GFP_NOMEMALLOC | __GFP_COMP
7713 | gfp_flags;
7714 unsigned int alloc_flags = ALLOC_TRYLOCK;
7715 struct alloc_context ac = { };
7716 struct page *page;
7717
7718 VM_WARN_ON_ONCE(gfp_flags & ~__GFP_ACCOUNT);
7719 /*
7720 * In PREEMPT_RT spin_trylock() will call raw_spin_lock() which is
7721 * unsafe in NMI. If spin_trylock() is called from hard IRQ the current
7722 * task may be waiting for one rt_spin_lock, but rt_spin_trylock() will
7723 * mark the task as the owner of another rt_spin_lock which will
7724 * confuse PI logic, so return immediately if called from hard IRQ or
7725 * NMI.
7726 *
7727 * Note, irqs_disabled() case is ok. This function can be called
7728 * from raw_spin_lock_irqsave region.
7729 */
7730 if (IS_ENABLED(CONFIG_PREEMPT_RT) && (in_nmi() || in_hardirq()))
7731 return NULL;
7732 if (!pcp_allowed_order(order))
7733 return NULL;
7734
7735 /* Bailout, since _deferred_grow_zone() needs to take a lock */
7736 if (deferred_pages_enabled())
7737 return NULL;
7738
7739 if (nid == NUMA_NO_NODE)
7740 nid = numa_node_id();
7741
7742 prepare_alloc_pages(alloc_gfp, order, nid, NULL, &ac,
7743 &alloc_gfp, &alloc_flags);
7744
7745 /*
7746 * Best effort allocation from percpu free list.
7747 * If it's empty attempt to spin_trylock zone->lock.
7748 */
7749 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
7750
7751 /* Unlike regular alloc_pages() there is no __alloc_pages_slowpath(). */
7752
7753 if (memcg_kmem_online() && page && (gfp_flags & __GFP_ACCOUNT) &&
7754 unlikely(__memcg_kmem_charge_page(page, alloc_gfp, order) != 0)) {
7755 __free_frozen_pages(page, order, FPI_TRYLOCK);
7756 page = NULL;
7757 }
7758 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
7759 kmsan_alloc_page(page, order, alloc_gfp);
7760 return page;
7761}
7762/**
7763 * alloc_pages_nolock - opportunistic reentrant allocation from any context
7764 * @gfp_flags: GFP flags. Only __GFP_ACCOUNT allowed.
7765 * @nid: node to allocate from
7766 * @order: allocation order size
7767 *
7768 * Allocates pages of a given order from the given node. This is safe to
7769 * call from any context (from atomic, NMI, and also reentrant
7770 * allocator -> tracepoint -> alloc_pages_nolock_noprof).
7771 * Allocation is best effort and to be expected to fail easily so nobody should
7772 * rely on the success. Failures are not reported via warn_alloc().
7773 * See always fail conditions below.
7774 *
7775 * Return: allocated page or NULL on failure. NULL does not mean EBUSY or EAGAIN.
7776 * It means ENOMEM. There is no reason to call it again and expect !NULL.
7777 */
7778struct page *alloc_pages_nolock_noprof(gfp_t gfp_flags, int nid, unsigned int order)
7779{
7780 struct page *page;
7781
7782 page = alloc_frozen_pages_nolock_noprof(gfp_flags, nid, order);
7783 if (page)
7784 set_page_refcounted(page);
7785 return page;
7786}
7787EXPORT_SYMBOL_GPL(alloc_pages_nolock_noprof);