1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/page_alloc.c
4 *
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
7 *
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
16 */
17
18#include <linux/stddef.h>
19#include <linux/mm.h>
20#include <linux/highmem.h>
21#include <linux/interrupt.h>
22#include <linux/jiffies.h>
23#include <linux/compiler.h>
24#include <linux/kernel.h>
25#include <linux/kasan.h>
26#include <linux/kmsan.h>
27#include <linux/module.h>
28#include <linux/suspend.h>
29#include <linux/ratelimit.h>
30#include <linux/oom.h>
31#include <linux/topology.h>
32#include <linux/sysctl.h>
33#include <linux/cpu.h>
34#include <linux/cpuset.h>
35#include <linux/pagevec.h>
36#include <linux/memory_hotplug.h>
37#include <linux/nodemask.h>
38#include <linux/vmstat.h>
39#include <linux/fault-inject.h>
40#include <linux/compaction.h>
41#include <trace/events/kmem.h>
42#include <trace/events/oom.h>
43#include <linux/prefetch.h>
44#include <linux/mm_inline.h>
45#include <linux/mmu_notifier.h>
46#include <linux/migrate.h>
47#include <linux/sched/mm.h>
48#include <linux/page_owner.h>
49#include <linux/page_table_check.h>
50#include <linux/memcontrol.h>
51#include <linux/ftrace.h>
52#include <linux/lockdep.h>
53#include <linux/psi.h>
54#include <linux/khugepaged.h>
55#include <linux/delayacct.h>
56#include <linux/cacheinfo.h>
57#include <linux/pgalloc_tag.h>
58#include <asm/div64.h>
59#include "internal.h"
60#include "shuffle.h"
61#include "page_reporting.h"
62
63/* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
64typedef int __bitwise fpi_t;
65
66/* No special request */
67#define FPI_NONE ((__force fpi_t)0)
68
69/*
70 * Skip free page reporting notification for the (possibly merged) page.
71 * This does not hinder free page reporting from grabbing the page,
72 * reporting it and marking it "reported" - it only skips notifying
73 * the free page reporting infrastructure about a newly freed page. For
74 * example, used when temporarily pulling a page from a freelist and
75 * putting it back unmodified.
76 */
77#define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
78
79/*
80 * Place the (possibly merged) page to the tail of the freelist. Will ignore
81 * page shuffling (relevant code - e.g., memory onlining - is expected to
82 * shuffle the whole zone).
83 *
84 * Note: No code should rely on this flag for correctness - it's purely
85 * to allow for optimizations when handing back either fresh pages
86 * (memory onlining) or untouched pages (page isolation, free page
87 * reporting).
88 */
89#define FPI_TO_TAIL ((__force fpi_t)BIT(1))
90
91/* Free the page without taking locks. Rely on trylock only. */
92#define FPI_TRYLOCK ((__force fpi_t)BIT(2))
93
94/* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
95static DEFINE_MUTEX(pcp_batch_high_lock);
96#define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
97
98#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
99/*
100 * On SMP, spin_trylock is sufficient protection.
101 * On PREEMPT_RT, spin_trylock is equivalent on both SMP and UP.
102 */
103#define pcp_trylock_prepare(flags) do { } while (0)
104#define pcp_trylock_finish(flag) do { } while (0)
105#else
106
107/* UP spin_trylock always succeeds so disable IRQs to prevent re-entrancy. */
108#define pcp_trylock_prepare(flags) local_irq_save(flags)
109#define pcp_trylock_finish(flags) local_irq_restore(flags)
110#endif
111
112/*
113 * Locking a pcp requires a PCP lookup followed by a spinlock. To avoid
114 * a migration causing the wrong PCP to be locked and remote memory being
115 * potentially allocated, pin the task to the CPU for the lookup+lock.
116 * preempt_disable is used on !RT because it is faster than migrate_disable.
117 * migrate_disable is used on RT because otherwise RT spinlock usage is
118 * interfered with and a high priority task cannot preempt the allocator.
119 */
120#ifndef CONFIG_PREEMPT_RT
121#define pcpu_task_pin() preempt_disable()
122#define pcpu_task_unpin() preempt_enable()
123#else
124#define pcpu_task_pin() migrate_disable()
125#define pcpu_task_unpin() migrate_enable()
126#endif
127
128/*
129 * Generic helper to lookup and a per-cpu variable with an embedded spinlock.
130 * Return value should be used with equivalent unlock helper.
131 */
132#define pcpu_spin_lock(type, member, ptr) \
133({ \
134 type *_ret; \
135 pcpu_task_pin(); \
136 _ret = this_cpu_ptr(ptr); \
137 spin_lock(&_ret->member); \
138 _ret; \
139})
140
141#define pcpu_spin_trylock(type, member, ptr) \
142({ \
143 type *_ret; \
144 pcpu_task_pin(); \
145 _ret = this_cpu_ptr(ptr); \
146 if (!spin_trylock(&_ret->member)) { \
147 pcpu_task_unpin(); \
148 _ret = NULL; \
149 } \
150 _ret; \
151})
152
153#define pcpu_spin_unlock(member, ptr) \
154({ \
155 spin_unlock(&ptr->member); \
156 pcpu_task_unpin(); \
157})
158
159/* struct per_cpu_pages specific helpers. */
160#define pcp_spin_lock(ptr) \
161 pcpu_spin_lock(struct per_cpu_pages, lock, ptr)
162
163#define pcp_spin_trylock(ptr) \
164 pcpu_spin_trylock(struct per_cpu_pages, lock, ptr)
165
166#define pcp_spin_unlock(ptr) \
167 pcpu_spin_unlock(lock, ptr)
168
169#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
170DEFINE_PER_CPU(int, numa_node);
171EXPORT_PER_CPU_SYMBOL(numa_node);
172#endif
173
174DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
175
176#ifdef CONFIG_HAVE_MEMORYLESS_NODES
177/*
178 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
179 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
180 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
181 * defined in <linux/topology.h>.
182 */
183DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
184EXPORT_PER_CPU_SYMBOL(_numa_mem_);
185#endif
186
187static DEFINE_MUTEX(pcpu_drain_mutex);
188
189#ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
190volatile unsigned long latent_entropy __latent_entropy;
191EXPORT_SYMBOL(latent_entropy);
192#endif
193
194/*
195 * Array of node states.
196 */
197nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
198 [N_POSSIBLE] = NODE_MASK_ALL,
199 [N_ONLINE] = { { [0] = 1UL } },
200#ifndef CONFIG_NUMA
201 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
202#ifdef CONFIG_HIGHMEM
203 [N_HIGH_MEMORY] = { { [0] = 1UL } },
204#endif
205 [N_MEMORY] = { { [0] = 1UL } },
206 [N_CPU] = { { [0] = 1UL } },
207#endif /* NUMA */
208};
209EXPORT_SYMBOL(node_states);
210
211gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
212
213#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
214unsigned int pageblock_order __read_mostly;
215#endif
216
217static void __free_pages_ok(struct page *page, unsigned int order,
218 fpi_t fpi_flags);
219
220/*
221 * results with 256, 32 in the lowmem_reserve sysctl:
222 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
223 * 1G machine -> (16M dma, 784M normal, 224M high)
224 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
225 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
226 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
227 *
228 * TBD: should special case ZONE_DMA32 machines here - in those we normally
229 * don't need any ZONE_NORMAL reservation
230 */
231static int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
232#ifdef CONFIG_ZONE_DMA
233 [ZONE_DMA] = 256,
234#endif
235#ifdef CONFIG_ZONE_DMA32
236 [ZONE_DMA32] = 256,
237#endif
238 [ZONE_NORMAL] = 32,
239#ifdef CONFIG_HIGHMEM
240 [ZONE_HIGHMEM] = 0,
241#endif
242 [ZONE_MOVABLE] = 0,
243};
244
245char * const zone_names[MAX_NR_ZONES] = {
246#ifdef CONFIG_ZONE_DMA
247 "DMA",
248#endif
249#ifdef CONFIG_ZONE_DMA32
250 "DMA32",
251#endif
252 "Normal",
253#ifdef CONFIG_HIGHMEM
254 "HighMem",
255#endif
256 "Movable",
257#ifdef CONFIG_ZONE_DEVICE
258 "Device",
259#endif
260};
261
262const char * const migratetype_names[MIGRATE_TYPES] = {
263 "Unmovable",
264 "Movable",
265 "Reclaimable",
266 "HighAtomic",
267#ifdef CONFIG_CMA
268 "CMA",
269#endif
270#ifdef CONFIG_MEMORY_ISOLATION
271 "Isolate",
272#endif
273};
274
275int min_free_kbytes = 1024;
276int user_min_free_kbytes = -1;
277static int watermark_boost_factor __read_mostly = 15000;
278static int watermark_scale_factor = 10;
279int defrag_mode;
280
281/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
282int movable_zone;
283EXPORT_SYMBOL(movable_zone);
284
285#if MAX_NUMNODES > 1
286unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
287unsigned int nr_online_nodes __read_mostly = 1;
288EXPORT_SYMBOL(nr_node_ids);
289EXPORT_SYMBOL(nr_online_nodes);
290#endif
291
292static bool page_contains_unaccepted(struct page *page, unsigned int order);
293static bool cond_accept_memory(struct zone *zone, unsigned int order,
294 int alloc_flags);
295static bool __free_unaccepted(struct page *page);
296
297int page_group_by_mobility_disabled __read_mostly;
298
299#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
300/*
301 * During boot we initialize deferred pages on-demand, as needed, but once
302 * page_alloc_init_late() has finished, the deferred pages are all initialized,
303 * and we can permanently disable that path.
304 */
305DEFINE_STATIC_KEY_TRUE(deferred_pages);
306
307static inline bool deferred_pages_enabled(void)
308{
309 return static_branch_unlikely(&deferred_pages);
310}
311
312/*
313 * deferred_grow_zone() is __init, but it is called from
314 * get_page_from_freelist() during early boot until deferred_pages permanently
315 * disables this call. This is why we have refdata wrapper to avoid warning,
316 * and to ensure that the function body gets unloaded.
317 */
318static bool __ref
319_deferred_grow_zone(struct zone *zone, unsigned int order)
320{
321 return deferred_grow_zone(zone, order);
322}
323#else
324static inline bool deferred_pages_enabled(void)
325{
326 return false;
327}
328
329static inline bool _deferred_grow_zone(struct zone *zone, unsigned int order)
330{
331 return false;
332}
333#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
334
335/* Return a pointer to the bitmap storing bits affecting a block of pages */
336static inline unsigned long *get_pageblock_bitmap(const struct page *page,
337 unsigned long pfn)
338{
339#ifdef CONFIG_SPARSEMEM
340 return section_to_usemap(ms: __pfn_to_section(pfn));
341#else
342 return page_zone(page)->pageblock_flags;
343#endif /* CONFIG_SPARSEMEM */
344}
345
346static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
347{
348#ifdef CONFIG_SPARSEMEM
349 pfn &= (PAGES_PER_SECTION-1);
350#else
351 pfn = pfn - pageblock_start_pfn(page_zone(page)->zone_start_pfn);
352#endif /* CONFIG_SPARSEMEM */
353 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
354}
355
356static __always_inline bool is_standalone_pb_bit(enum pageblock_bits pb_bit)
357{
358 return pb_bit >= PB_compact_skip && pb_bit < __NR_PAGEBLOCK_BITS;
359}
360
361static __always_inline void
362get_pfnblock_bitmap_bitidx(const struct page *page, unsigned long pfn,
363 unsigned long **bitmap_word, unsigned long *bitidx)
364{
365 unsigned long *bitmap;
366 unsigned long word_bitidx;
367
368#ifdef CONFIG_MEMORY_ISOLATION
369 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 8);
370#else
371 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
372#endif
373 BUILD_BUG_ON(__MIGRATE_TYPE_END > MIGRATETYPE_MASK);
374 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
375
376 bitmap = get_pageblock_bitmap(page, pfn);
377 *bitidx = pfn_to_bitidx(page, pfn);
378 word_bitidx = *bitidx / BITS_PER_LONG;
379 *bitidx &= (BITS_PER_LONG - 1);
380 *bitmap_word = &bitmap[word_bitidx];
381}
382
383
384/**
385 * __get_pfnblock_flags_mask - Return the requested group of flags for
386 * a pageblock_nr_pages block of pages
387 * @page: The page within the block of interest
388 * @pfn: The target page frame number
389 * @mask: mask of bits that the caller is interested in
390 *
391 * Return: pageblock_bits flags
392 */
393static unsigned long __get_pfnblock_flags_mask(const struct page *page,
394 unsigned long pfn,
395 unsigned long mask)
396{
397 unsigned long *bitmap_word;
398 unsigned long bitidx;
399 unsigned long word;
400
401 get_pfnblock_bitmap_bitidx(page, pfn, bitmap_word: &bitmap_word, bitidx: &bitidx);
402 /*
403 * This races, without locks, with set_pfnblock_migratetype(). Ensure
404 * a consistent read of the memory array, so that results, even though
405 * racy, are not corrupted.
406 */
407 word = READ_ONCE(*bitmap_word);
408 return (word >> bitidx) & mask;
409}
410
411/**
412 * get_pfnblock_bit - Check if a standalone bit of a pageblock is set
413 * @page: The page within the block of interest
414 * @pfn: The target page frame number
415 * @pb_bit: pageblock bit to check
416 *
417 * Return: true if the bit is set, otherwise false
418 */
419bool get_pfnblock_bit(const struct page *page, unsigned long pfn,
420 enum pageblock_bits pb_bit)
421{
422 unsigned long *bitmap_word;
423 unsigned long bitidx;
424
425 if (WARN_ON_ONCE(!is_standalone_pb_bit(pb_bit)))
426 return false;
427
428 get_pfnblock_bitmap_bitidx(page, pfn, bitmap_word: &bitmap_word, bitidx: &bitidx);
429
430 return test_bit(bitidx + pb_bit, bitmap_word);
431}
432
433/**
434 * get_pfnblock_migratetype - Return the migratetype of a pageblock
435 * @page: The page within the block of interest
436 * @pfn: The target page frame number
437 *
438 * Return: The migratetype of the pageblock
439 *
440 * Use get_pfnblock_migratetype() if caller already has both @page and @pfn
441 * to save a call to page_to_pfn().
442 */
443__always_inline enum migratetype
444get_pfnblock_migratetype(const struct page *page, unsigned long pfn)
445{
446 unsigned long mask = MIGRATETYPE_AND_ISO_MASK;
447 unsigned long flags;
448
449 flags = __get_pfnblock_flags_mask(page, pfn, mask);
450
451#ifdef CONFIG_MEMORY_ISOLATION
452 if (flags & BIT(PB_migrate_isolate))
453 return MIGRATE_ISOLATE;
454#endif
455 return flags & MIGRATETYPE_MASK;
456}
457
458/**
459 * __set_pfnblock_flags_mask - Set the requested group of flags for
460 * a pageblock_nr_pages block of pages
461 * @page: The page within the block of interest
462 * @pfn: The target page frame number
463 * @flags: The flags to set
464 * @mask: mask of bits that the caller is interested in
465 */
466static void __set_pfnblock_flags_mask(struct page *page, unsigned long pfn,
467 unsigned long flags, unsigned long mask)
468{
469 unsigned long *bitmap_word;
470 unsigned long bitidx;
471 unsigned long word;
472
473 get_pfnblock_bitmap_bitidx(page, pfn, bitmap_word: &bitmap_word, bitidx: &bitidx);
474
475 mask <<= bitidx;
476 flags <<= bitidx;
477
478 word = READ_ONCE(*bitmap_word);
479 do {
480 } while (!try_cmpxchg(bitmap_word, &word, (word & ~mask) | flags));
481}
482
483/**
484 * set_pfnblock_bit - Set a standalone bit of a pageblock
485 * @page: The page within the block of interest
486 * @pfn: The target page frame number
487 * @pb_bit: pageblock bit to set
488 */
489void set_pfnblock_bit(const struct page *page, unsigned long pfn,
490 enum pageblock_bits pb_bit)
491{
492 unsigned long *bitmap_word;
493 unsigned long bitidx;
494
495 if (WARN_ON_ONCE(!is_standalone_pb_bit(pb_bit)))
496 return;
497
498 get_pfnblock_bitmap_bitidx(page, pfn, bitmap_word: &bitmap_word, bitidx: &bitidx);
499
500 set_bit(nr: bitidx + pb_bit, addr: bitmap_word);
501}
502
503/**
504 * clear_pfnblock_bit - Clear a standalone bit of a pageblock
505 * @page: The page within the block of interest
506 * @pfn: The target page frame number
507 * @pb_bit: pageblock bit to clear
508 */
509void clear_pfnblock_bit(const struct page *page, unsigned long pfn,
510 enum pageblock_bits pb_bit)
511{
512 unsigned long *bitmap_word;
513 unsigned long bitidx;
514
515 if (WARN_ON_ONCE(!is_standalone_pb_bit(pb_bit)))
516 return;
517
518 get_pfnblock_bitmap_bitidx(page, pfn, bitmap_word: &bitmap_word, bitidx: &bitidx);
519
520 clear_bit(nr: bitidx + pb_bit, addr: bitmap_word);
521}
522
523/**
524 * set_pageblock_migratetype - Set the migratetype of a pageblock
525 * @page: The page within the block of interest
526 * @migratetype: migratetype to set
527 */
528static void set_pageblock_migratetype(struct page *page,
529 enum migratetype migratetype)
530{
531 if (unlikely(page_group_by_mobility_disabled &&
532 migratetype < MIGRATE_PCPTYPES))
533 migratetype = MIGRATE_UNMOVABLE;
534
535#ifdef CONFIG_MEMORY_ISOLATION
536 if (migratetype == MIGRATE_ISOLATE) {
537 VM_WARN_ONCE(1,
538 "Use set_pageblock_isolate() for pageblock isolation");
539 return;
540 }
541 VM_WARN_ONCE(get_pageblock_isolate(page),
542 "Use clear_pageblock_isolate() to unisolate pageblock");
543 /* MIGRATETYPE_AND_ISO_MASK clears PB_migrate_isolate if it is set */
544#endif
545 __set_pfnblock_flags_mask(page, page_to_pfn(page),
546 flags: (unsigned long)migratetype,
547 MIGRATETYPE_AND_ISO_MASK);
548}
549
550void __meminit init_pageblock_migratetype(struct page *page,
551 enum migratetype migratetype,
552 bool isolate)
553{
554 unsigned long flags;
555
556 if (unlikely(page_group_by_mobility_disabled &&
557 migratetype < MIGRATE_PCPTYPES))
558 migratetype = MIGRATE_UNMOVABLE;
559
560 flags = migratetype;
561
562#ifdef CONFIG_MEMORY_ISOLATION
563 if (migratetype == MIGRATE_ISOLATE) {
564 VM_WARN_ONCE(
565 1,
566 "Set isolate=true to isolate pageblock with a migratetype");
567 return;
568 }
569 if (isolate)
570 flags |= BIT(PB_migrate_isolate);
571#endif
572 __set_pfnblock_flags_mask(page, page_to_pfn(page), flags,
573 MIGRATETYPE_AND_ISO_MASK);
574}
575
576#ifdef CONFIG_DEBUG_VM
577static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
578{
579 int ret;
580 unsigned seq;
581 unsigned long pfn = page_to_pfn(page);
582 unsigned long sp, start_pfn;
583
584 do {
585 seq = zone_span_seqbegin(zone);
586 start_pfn = zone->zone_start_pfn;
587 sp = zone->spanned_pages;
588 ret = !zone_spans_pfn(zone, pfn);
589 } while (zone_span_seqretry(zone, seq));
590
591 if (ret)
592 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
593 pfn, zone_to_nid(zone), zone->name,
594 start_pfn, start_pfn + sp);
595
596 return ret;
597}
598
599/*
600 * Temporary debugging check for pages not lying within a given zone.
601 */
602static bool __maybe_unused bad_range(struct zone *zone, struct page *page)
603{
604 if (page_outside_zone_boundaries(zone, page))
605 return true;
606 if (zone != page_zone(page))
607 return true;
608
609 return false;
610}
611#else
612static inline bool __maybe_unused bad_range(struct zone *zone, struct page *page)
613{
614 return false;
615}
616#endif
617
618static void bad_page(struct page *page, const char *reason)
619{
620 static unsigned long resume;
621 static unsigned long nr_shown;
622 static unsigned long nr_unshown;
623
624 /*
625 * Allow a burst of 60 reports, then keep quiet for that minute;
626 * or allow a steady drip of one report per second.
627 */
628 if (nr_shown == 60) {
629 if (time_before(jiffies, resume)) {
630 nr_unshown++;
631 goto out;
632 }
633 if (nr_unshown) {
634 pr_alert(
635 "BUG: Bad page state: %lu messages suppressed\n",
636 nr_unshown);
637 nr_unshown = 0;
638 }
639 nr_shown = 0;
640 }
641 if (nr_shown++ == 0)
642 resume = jiffies + 60 * HZ;
643
644 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
645 current->comm, page_to_pfn(page));
646 dump_page(page, reason);
647
648 print_modules();
649 dump_stack();
650out:
651 /* Leave bad fields for debug, except PageBuddy could make trouble */
652 if (PageBuddy(page))
653 __ClearPageBuddy(page);
654 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
655}
656
657static inline unsigned int order_to_pindex(int migratetype, int order)
658{
659
660#ifdef CONFIG_TRANSPARENT_HUGEPAGE
661 bool movable;
662 if (order > PAGE_ALLOC_COSTLY_ORDER) {
663 VM_BUG_ON(order != HPAGE_PMD_ORDER);
664
665 movable = migratetype == MIGRATE_MOVABLE;
666
667 return NR_LOWORDER_PCP_LISTS + movable;
668 }
669#else
670 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
671#endif
672
673 return (MIGRATE_PCPTYPES * order) + migratetype;
674}
675
676static inline int pindex_to_order(unsigned int pindex)
677{
678 int order = pindex / MIGRATE_PCPTYPES;
679
680#ifdef CONFIG_TRANSPARENT_HUGEPAGE
681 if (pindex >= NR_LOWORDER_PCP_LISTS)
682 order = HPAGE_PMD_ORDER;
683#else
684 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
685#endif
686
687 return order;
688}
689
690static inline bool pcp_allowed_order(unsigned int order)
691{
692 if (order <= PAGE_ALLOC_COSTLY_ORDER)
693 return true;
694#ifdef CONFIG_TRANSPARENT_HUGEPAGE
695 if (order == HPAGE_PMD_ORDER)
696 return true;
697#endif
698 return false;
699}
700
701/*
702 * Higher-order pages are called "compound pages". They are structured thusly:
703 *
704 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
705 *
706 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
707 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
708 *
709 * The first tail page's ->compound_order holds the order of allocation.
710 * This usage means that zero-order pages may not be compound.
711 */
712
713void prep_compound_page(struct page *page, unsigned int order)
714{
715 int i;
716 int nr_pages = 1 << order;
717
718 __SetPageHead(page);
719 for (i = 1; i < nr_pages; i++)
720 prep_compound_tail(head: page, tail_idx: i);
721
722 prep_compound_head(page, order);
723}
724
725static inline void set_buddy_order(struct page *page, unsigned int order)
726{
727 set_page_private(page, private: order);
728 __SetPageBuddy(page);
729}
730
731#ifdef CONFIG_COMPACTION
732static inline struct capture_control *task_capc(struct zone *zone)
733{
734 struct capture_control *capc = current->capture_control;
735
736 return unlikely(capc) &&
737 !(current->flags & PF_KTHREAD) &&
738 !capc->page &&
739 capc->cc->zone == zone ? capc : NULL;
740}
741
742static inline bool
743compaction_capture(struct capture_control *capc, struct page *page,
744 int order, int migratetype)
745{
746 if (!capc || order != capc->cc->order)
747 return false;
748
749 /* Do not accidentally pollute CMA or isolated regions*/
750 if (is_migrate_cma(migratetype) ||
751 is_migrate_isolate(migratetype))
752 return false;
753
754 /*
755 * Do not let lower order allocations pollute a movable pageblock
756 * unless compaction is also requesting movable pages.
757 * This might let an unmovable request use a reclaimable pageblock
758 * and vice-versa but no more than normal fallback logic which can
759 * have trouble finding a high-order free page.
760 */
761 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE &&
762 capc->cc->migratetype != MIGRATE_MOVABLE)
763 return false;
764
765 if (migratetype != capc->cc->migratetype)
766 trace_mm_page_alloc_extfrag(page, alloc_order: capc->cc->order, fallback_order: order,
767 alloc_migratetype: capc->cc->migratetype, fallback_migratetype: migratetype);
768
769 capc->page = page;
770 return true;
771}
772
773#else
774static inline struct capture_control *task_capc(struct zone *zone)
775{
776 return NULL;
777}
778
779static inline bool
780compaction_capture(struct capture_control *capc, struct page *page,
781 int order, int migratetype)
782{
783 return false;
784}
785#endif /* CONFIG_COMPACTION */
786
787static inline void account_freepages(struct zone *zone, int nr_pages,
788 int migratetype)
789{
790 lockdep_assert_held(&zone->lock);
791
792 if (is_migrate_isolate(migratetype))
793 return;
794
795 __mod_zone_page_state(zone, item: NR_FREE_PAGES, nr_pages);
796
797 if (is_migrate_cma(migratetype))
798 __mod_zone_page_state(zone, item: NR_FREE_CMA_PAGES, nr_pages);
799 else if (migratetype == MIGRATE_HIGHATOMIC)
800 WRITE_ONCE(zone->nr_free_highatomic,
801 zone->nr_free_highatomic + nr_pages);
802}
803
804/* Used for pages not on another list */
805static inline void __add_to_free_list(struct page *page, struct zone *zone,
806 unsigned int order, int migratetype,
807 bool tail)
808{
809 struct free_area *area = &zone->free_area[order];
810 int nr_pages = 1 << order;
811
812 VM_WARN_ONCE(get_pageblock_migratetype(page) != migratetype,
813 "page type is %d, passed migratetype is %d (nr=%d)\n",
814 get_pageblock_migratetype(page), migratetype, nr_pages);
815
816 if (tail)
817 list_add_tail(new: &page->buddy_list, head: &area->free_list[migratetype]);
818 else
819 list_add(new: &page->buddy_list, head: &area->free_list[migratetype]);
820 area->nr_free++;
821
822 if (order >= pageblock_order && !is_migrate_isolate(migratetype))
823 __mod_zone_page_state(zone, item: NR_FREE_PAGES_BLOCKS, nr_pages);
824}
825
826/*
827 * Used for pages which are on another list. Move the pages to the tail
828 * of the list - so the moved pages won't immediately be considered for
829 * allocation again (e.g., optimization for memory onlining).
830 */
831static inline void move_to_free_list(struct page *page, struct zone *zone,
832 unsigned int order, int old_mt, int new_mt)
833{
834 struct free_area *area = &zone->free_area[order];
835 int nr_pages = 1 << order;
836
837 /* Free page moving can fail, so it happens before the type update */
838 VM_WARN_ONCE(get_pageblock_migratetype(page) != old_mt,
839 "page type is %d, passed migratetype is %d (nr=%d)\n",
840 get_pageblock_migratetype(page), old_mt, nr_pages);
841
842 list_move_tail(list: &page->buddy_list, head: &area->free_list[new_mt]);
843
844 account_freepages(zone, nr_pages: -nr_pages, migratetype: old_mt);
845 account_freepages(zone, nr_pages, migratetype: new_mt);
846
847 if (order >= pageblock_order &&
848 is_migrate_isolate(migratetype: old_mt) != is_migrate_isolate(migratetype: new_mt)) {
849 if (!is_migrate_isolate(migratetype: old_mt))
850 nr_pages = -nr_pages;
851 __mod_zone_page_state(zone, item: NR_FREE_PAGES_BLOCKS, nr_pages);
852 }
853}
854
855static inline void __del_page_from_free_list(struct page *page, struct zone *zone,
856 unsigned int order, int migratetype)
857{
858 int nr_pages = 1 << order;
859
860 VM_WARN_ONCE(get_pageblock_migratetype(page) != migratetype,
861 "page type is %d, passed migratetype is %d (nr=%d)\n",
862 get_pageblock_migratetype(page), migratetype, nr_pages);
863
864 /* clear reported state and update reported page count */
865 if (page_reported(page))
866 __ClearPageReported(page);
867
868 list_del(entry: &page->buddy_list);
869 __ClearPageBuddy(page);
870 set_page_private(page, private: 0);
871 zone->free_area[order].nr_free--;
872
873 if (order >= pageblock_order && !is_migrate_isolate(migratetype))
874 __mod_zone_page_state(zone, item: NR_FREE_PAGES_BLOCKS, -nr_pages);
875}
876
877static inline void del_page_from_free_list(struct page *page, struct zone *zone,
878 unsigned int order, int migratetype)
879{
880 __del_page_from_free_list(page, zone, order, migratetype);
881 account_freepages(zone, nr_pages: -(1 << order), migratetype);
882}
883
884static inline struct page *get_page_from_free_area(struct free_area *area,
885 int migratetype)
886{
887 return list_first_entry_or_null(&area->free_list[migratetype],
888 struct page, buddy_list);
889}
890
891/*
892 * If this is less than the 2nd largest possible page, check if the buddy
893 * of the next-higher order is free. If it is, it's possible
894 * that pages are being freed that will coalesce soon. In case,
895 * that is happening, add the free page to the tail of the list
896 * so it's less likely to be used soon and more likely to be merged
897 * as a 2-level higher order page
898 */
899static inline bool
900buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
901 struct page *page, unsigned int order)
902{
903 unsigned long higher_page_pfn;
904 struct page *higher_page;
905
906 if (order >= MAX_PAGE_ORDER - 1)
907 return false;
908
909 higher_page_pfn = buddy_pfn & pfn;
910 higher_page = page + (higher_page_pfn - pfn);
911
912 return find_buddy_page_pfn(page: higher_page, pfn: higher_page_pfn, order: order + 1,
913 NULL) != NULL;
914}
915
916/*
917 * Freeing function for a buddy system allocator.
918 *
919 * The concept of a buddy system is to maintain direct-mapped table
920 * (containing bit values) for memory blocks of various "orders".
921 * The bottom level table contains the map for the smallest allocatable
922 * units of memory (here, pages), and each level above it describes
923 * pairs of units from the levels below, hence, "buddies".
924 * At a high level, all that happens here is marking the table entry
925 * at the bottom level available, and propagating the changes upward
926 * as necessary, plus some accounting needed to play nicely with other
927 * parts of the VM system.
928 * At each level, we keep a list of pages, which are heads of continuous
929 * free pages of length of (1 << order) and marked with PageBuddy.
930 * Page's order is recorded in page_private(page) field.
931 * So when we are allocating or freeing one, we can derive the state of the
932 * other. That is, if we allocate a small block, and both were
933 * free, the remainder of the region must be split into blocks.
934 * If a block is freed, and its buddy is also free, then this
935 * triggers coalescing into a block of larger size.
936 *
937 * -- nyc
938 */
939
940static inline void __free_one_page(struct page *page,
941 unsigned long pfn,
942 struct zone *zone, unsigned int order,
943 int migratetype, fpi_t fpi_flags)
944{
945 struct capture_control *capc = task_capc(zone);
946 unsigned long buddy_pfn = 0;
947 unsigned long combined_pfn;
948 struct page *buddy;
949 bool to_tail;
950
951 VM_BUG_ON(!zone_is_initialized(zone));
952 VM_BUG_ON_PAGE(page->flags.f & PAGE_FLAGS_CHECK_AT_PREP, page);
953
954 VM_BUG_ON(migratetype == -1);
955 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
956 VM_BUG_ON_PAGE(bad_range(zone, page), page);
957
958 account_freepages(zone, nr_pages: 1 << order, migratetype);
959
960 while (order < MAX_PAGE_ORDER) {
961 int buddy_mt = migratetype;
962
963 if (compaction_capture(capc, page, order, migratetype)) {
964 account_freepages(zone, nr_pages: -(1 << order), migratetype);
965 return;
966 }
967
968 buddy = find_buddy_page_pfn(page, pfn, order, buddy_pfn: &buddy_pfn);
969 if (!buddy)
970 goto done_merging;
971
972 if (unlikely(order >= pageblock_order)) {
973 /*
974 * We want to prevent merge between freepages on pageblock
975 * without fallbacks and normal pageblock. Without this,
976 * pageblock isolation could cause incorrect freepage or CMA
977 * accounting or HIGHATOMIC accounting.
978 */
979 buddy_mt = get_pfnblock_migratetype(page: buddy, pfn: buddy_pfn);
980
981 if (migratetype != buddy_mt &&
982 (!migratetype_is_mergeable(mt: migratetype) ||
983 !migratetype_is_mergeable(mt: buddy_mt)))
984 goto done_merging;
985 }
986
987 /*
988 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
989 * merge with it and move up one order.
990 */
991 if (page_is_guard(page: buddy))
992 clear_page_guard(zone, page: buddy, order);
993 else
994 __del_page_from_free_list(page: buddy, zone, order, migratetype: buddy_mt);
995
996 if (unlikely(buddy_mt != migratetype)) {
997 /*
998 * Match buddy type. This ensures that an
999 * expand() down the line puts the sub-blocks
1000 * on the right freelists.
1001 */
1002 set_pageblock_migratetype(page: buddy, migratetype);
1003 }
1004
1005 combined_pfn = buddy_pfn & pfn;
1006 page = page + (combined_pfn - pfn);
1007 pfn = combined_pfn;
1008 order++;
1009 }
1010
1011done_merging:
1012 set_buddy_order(page, order);
1013
1014 if (fpi_flags & FPI_TO_TAIL)
1015 to_tail = true;
1016 else if (is_shuffle_order(order))
1017 to_tail = shuffle_pick_tail();
1018 else
1019 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1020
1021 __add_to_free_list(page, zone, order, migratetype, tail: to_tail);
1022
1023 /* Notify page reporting subsystem of freed page */
1024 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1025 page_reporting_notify_free(order);
1026}
1027
1028/*
1029 * A bad page could be due to a number of fields. Instead of multiple branches,
1030 * try and check multiple fields with one check. The caller must do a detailed
1031 * check if necessary.
1032 */
1033static inline bool page_expected_state(struct page *page,
1034 unsigned long check_flags)
1035{
1036 if (unlikely(atomic_read(&page->_mapcount) != -1))
1037 return false;
1038
1039 if (unlikely((unsigned long)page->mapping |
1040 page_ref_count(page) |
1041#ifdef CONFIG_MEMCG
1042 page->memcg_data |
1043#endif
1044 page_pool_page_is_pp(page) |
1045 (page->flags.f & check_flags)))
1046 return false;
1047
1048 return true;
1049}
1050
1051static const char *page_bad_reason(struct page *page, unsigned long flags)
1052{
1053 const char *bad_reason = NULL;
1054
1055 if (unlikely(atomic_read(&page->_mapcount) != -1))
1056 bad_reason = "nonzero mapcount";
1057 if (unlikely(page->mapping != NULL))
1058 bad_reason = "non-NULL mapping";
1059 if (unlikely(page_ref_count(page) != 0))
1060 bad_reason = "nonzero _refcount";
1061 if (unlikely(page->flags.f & flags)) {
1062 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1063 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1064 else
1065 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1066 }
1067#ifdef CONFIG_MEMCG
1068 if (unlikely(page->memcg_data))
1069 bad_reason = "page still charged to cgroup";
1070#endif
1071 if (unlikely(page_pool_page_is_pp(page)))
1072 bad_reason = "page_pool leak";
1073 return bad_reason;
1074}
1075
1076static inline bool free_page_is_bad(struct page *page)
1077{
1078 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1079 return false;
1080
1081 /* Something has gone sideways, find it */
1082 bad_page(page, reason: page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1083 return true;
1084}
1085
1086static inline bool is_check_pages_enabled(void)
1087{
1088 return static_branch_unlikely(&check_pages_enabled);
1089}
1090
1091static int free_tail_page_prepare(struct page *head_page, struct page *page)
1092{
1093 struct folio *folio = (struct folio *)head_page;
1094 int ret = 1;
1095
1096 /*
1097 * We rely page->lru.next never has bit 0 set, unless the page
1098 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1099 */
1100 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1101
1102 if (!is_check_pages_enabled()) {
1103 ret = 0;
1104 goto out;
1105 }
1106 switch (page - head_page) {
1107 case 1:
1108 /* the first tail page: these may be in place of ->mapping */
1109 if (unlikely(folio_large_mapcount(folio))) {
1110 bad_page(page, reason: "nonzero large_mapcount");
1111 goto out;
1112 }
1113 if (IS_ENABLED(CONFIG_PAGE_MAPCOUNT) &&
1114 unlikely(atomic_read(&folio->_nr_pages_mapped))) {
1115 bad_page(page, reason: "nonzero nr_pages_mapped");
1116 goto out;
1117 }
1118 if (IS_ENABLED(CONFIG_MM_ID)) {
1119 if (unlikely(folio->_mm_id_mapcount[0] != -1)) {
1120 bad_page(page, reason: "nonzero mm mapcount 0");
1121 goto out;
1122 }
1123 if (unlikely(folio->_mm_id_mapcount[1] != -1)) {
1124 bad_page(page, reason: "nonzero mm mapcount 1");
1125 goto out;
1126 }
1127 }
1128 if (IS_ENABLED(CONFIG_64BIT)) {
1129 if (unlikely(atomic_read(&folio->_entire_mapcount) + 1)) {
1130 bad_page(page, reason: "nonzero entire_mapcount");
1131 goto out;
1132 }
1133 if (unlikely(atomic_read(&folio->_pincount))) {
1134 bad_page(page, reason: "nonzero pincount");
1135 goto out;
1136 }
1137 }
1138 break;
1139 case 2:
1140 /* the second tail page: deferred_list overlaps ->mapping */
1141 if (unlikely(!list_empty(&folio->_deferred_list))) {
1142 bad_page(page, reason: "on deferred list");
1143 goto out;
1144 }
1145 if (!IS_ENABLED(CONFIG_64BIT)) {
1146 if (unlikely(atomic_read(&folio->_entire_mapcount) + 1)) {
1147 bad_page(page, reason: "nonzero entire_mapcount");
1148 goto out;
1149 }
1150 if (unlikely(atomic_read(&folio->_pincount))) {
1151 bad_page(page, reason: "nonzero pincount");
1152 goto out;
1153 }
1154 }
1155 break;
1156 case 3:
1157 /* the third tail page: hugetlb specifics overlap ->mappings */
1158 if (IS_ENABLED(CONFIG_HUGETLB_PAGE))
1159 break;
1160 fallthrough;
1161 default:
1162 if (page->mapping != TAIL_MAPPING) {
1163 bad_page(page, reason: "corrupted mapping in tail page");
1164 goto out;
1165 }
1166 break;
1167 }
1168 if (unlikely(!PageTail(page))) {
1169 bad_page(page, reason: "PageTail not set");
1170 goto out;
1171 }
1172 if (unlikely(compound_head(page) != head_page)) {
1173 bad_page(page, reason: "compound_head not consistent");
1174 goto out;
1175 }
1176 ret = 0;
1177out:
1178 page->mapping = NULL;
1179 clear_compound_head(page);
1180 return ret;
1181}
1182
1183/*
1184 * Skip KASAN memory poisoning when either:
1185 *
1186 * 1. For generic KASAN: deferred memory initialization has not yet completed.
1187 * Tag-based KASAN modes skip pages freed via deferred memory initialization
1188 * using page tags instead (see below).
1189 * 2. For tag-based KASAN modes: the page has a match-all KASAN tag, indicating
1190 * that error detection is disabled for accesses via the page address.
1191 *
1192 * Pages will have match-all tags in the following circumstances:
1193 *
1194 * 1. Pages are being initialized for the first time, including during deferred
1195 * memory init; see the call to page_kasan_tag_reset in __init_single_page.
1196 * 2. The allocation was not unpoisoned due to __GFP_SKIP_KASAN, with the
1197 * exception of pages unpoisoned by kasan_unpoison_vmalloc.
1198 * 3. The allocation was excluded from being checked due to sampling,
1199 * see the call to kasan_unpoison_pages.
1200 *
1201 * Poisoning pages during deferred memory init will greatly lengthen the
1202 * process and cause problem in large memory systems as the deferred pages
1203 * initialization is done with interrupt disabled.
1204 *
1205 * Assuming that there will be no reference to those newly initialized
1206 * pages before they are ever allocated, this should have no effect on
1207 * KASAN memory tracking as the poison will be properly inserted at page
1208 * allocation time. The only corner case is when pages are allocated by
1209 * on-demand allocation and then freed again before the deferred pages
1210 * initialization is done, but this is not likely to happen.
1211 */
1212static inline bool should_skip_kasan_poison(struct page *page)
1213{
1214 if (IS_ENABLED(CONFIG_KASAN_GENERIC))
1215 return deferred_pages_enabled();
1216
1217 return page_kasan_tag(page) == KASAN_TAG_KERNEL;
1218}
1219
1220static void kernel_init_pages(struct page *page, int numpages)
1221{
1222 int i;
1223
1224 /* s390's use of memset() could override KASAN redzones. */
1225 kasan_disable_current();
1226 for (i = 0; i < numpages; i++)
1227 clear_highpage_kasan_tagged(page: page + i);
1228 kasan_enable_current();
1229}
1230
1231#ifdef CONFIG_MEM_ALLOC_PROFILING
1232
1233/* Should be called only if mem_alloc_profiling_enabled() */
1234void __clear_page_tag_ref(struct page *page)
1235{
1236 union pgtag_ref_handle handle;
1237 union codetag_ref ref;
1238
1239 if (get_page_tag_ref(page, &ref, &handle)) {
1240 set_codetag_empty(&ref);
1241 update_page_tag_ref(handle, &ref);
1242 put_page_tag_ref(handle);
1243 }
1244}
1245
1246/* Should be called only if mem_alloc_profiling_enabled() */
1247static noinline
1248void __pgalloc_tag_add(struct page *page, struct task_struct *task,
1249 unsigned int nr)
1250{
1251 union pgtag_ref_handle handle;
1252 union codetag_ref ref;
1253
1254 if (get_page_tag_ref(page, &ref, &handle)) {
1255 alloc_tag_add(&ref, task->alloc_tag, PAGE_SIZE * nr);
1256 update_page_tag_ref(handle, &ref);
1257 put_page_tag_ref(handle);
1258 }
1259}
1260
1261static inline void pgalloc_tag_add(struct page *page, struct task_struct *task,
1262 unsigned int nr)
1263{
1264 if (mem_alloc_profiling_enabled())
1265 __pgalloc_tag_add(page, task, nr);
1266}
1267
1268/* Should be called only if mem_alloc_profiling_enabled() */
1269static noinline
1270void __pgalloc_tag_sub(struct page *page, unsigned int nr)
1271{
1272 union pgtag_ref_handle handle;
1273 union codetag_ref ref;
1274
1275 if (get_page_tag_ref(page, &ref, &handle)) {
1276 alloc_tag_sub(&ref, PAGE_SIZE * nr);
1277 update_page_tag_ref(handle, &ref);
1278 put_page_tag_ref(handle);
1279 }
1280}
1281
1282static inline void pgalloc_tag_sub(struct page *page, unsigned int nr)
1283{
1284 if (mem_alloc_profiling_enabled())
1285 __pgalloc_tag_sub(page, nr);
1286}
1287
1288/* When tag is not NULL, assuming mem_alloc_profiling_enabled */
1289static inline void pgalloc_tag_sub_pages(struct alloc_tag *tag, unsigned int nr)
1290{
1291 if (tag)
1292 this_cpu_sub(tag->counters->bytes, PAGE_SIZE * nr);
1293}
1294
1295#else /* CONFIG_MEM_ALLOC_PROFILING */
1296
1297static inline void pgalloc_tag_add(struct page *page, struct task_struct *task,
1298 unsigned int nr) {}
1299static inline void pgalloc_tag_sub(struct page *page, unsigned int nr) {}
1300static inline void pgalloc_tag_sub_pages(struct alloc_tag *tag, unsigned int nr) {}
1301
1302#endif /* CONFIG_MEM_ALLOC_PROFILING */
1303
1304__always_inline bool free_pages_prepare(struct page *page,
1305 unsigned int order)
1306{
1307 int bad = 0;
1308 bool skip_kasan_poison = should_skip_kasan_poison(page);
1309 bool init = want_init_on_free();
1310 bool compound = PageCompound(page);
1311 struct folio *folio = page_folio(page);
1312
1313 VM_BUG_ON_PAGE(PageTail(page), page);
1314
1315 trace_mm_page_free(page, order);
1316 kmsan_free_page(page, order);
1317
1318 if (memcg_kmem_online() && PageMemcgKmem(page))
1319 __memcg_kmem_uncharge_page(page, order);
1320
1321 /*
1322 * In rare cases, when truncation or holepunching raced with
1323 * munlock after VM_LOCKED was cleared, Mlocked may still be
1324 * found set here. This does not indicate a problem, unless
1325 * "unevictable_pgs_cleared" appears worryingly large.
1326 */
1327 if (unlikely(folio_test_mlocked(folio))) {
1328 long nr_pages = folio_nr_pages(folio);
1329
1330 __folio_clear_mlocked(folio);
1331 zone_stat_mod_folio(folio, item: NR_MLOCK, nr: -nr_pages);
1332 count_vm_events(item: UNEVICTABLE_PGCLEARED, delta: nr_pages);
1333 }
1334
1335 if (unlikely(PageHWPoison(page)) && !order) {
1336 /* Do not let hwpoison pages hit pcplists/buddy */
1337 reset_page_owner(page, order);
1338 page_table_check_free(page, order);
1339 pgalloc_tag_sub(page, nr: 1 << order);
1340
1341 /*
1342 * The page is isolated and accounted for.
1343 * Mark the codetag as empty to avoid accounting error
1344 * when the page is freed by unpoison_memory().
1345 */
1346 clear_page_tag_ref(page);
1347 return false;
1348 }
1349
1350 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1351
1352 /*
1353 * Check tail pages before head page information is cleared to
1354 * avoid checking PageCompound for order-0 pages.
1355 */
1356 if (unlikely(order)) {
1357 int i;
1358
1359 if (compound) {
1360 page[1].flags.f &= ~PAGE_FLAGS_SECOND;
1361#ifdef NR_PAGES_IN_LARGE_FOLIO
1362 folio->_nr_pages = 0;
1363#endif
1364 }
1365 for (i = 1; i < (1 << order); i++) {
1366 if (compound)
1367 bad += free_tail_page_prepare(head_page: page, page: page + i);
1368 if (is_check_pages_enabled()) {
1369 if (free_page_is_bad(page: page + i)) {
1370 bad++;
1371 continue;
1372 }
1373 }
1374 (page + i)->flags.f &= ~PAGE_FLAGS_CHECK_AT_PREP;
1375 }
1376 }
1377 if (folio_test_anon(folio)) {
1378 mod_mthp_stat(order, item: MTHP_STAT_NR_ANON, delta: -1);
1379 folio->mapping = NULL;
1380 }
1381 if (unlikely(page_has_type(page)))
1382 /* Reset the page_type (which overlays _mapcount) */
1383 page->page_type = UINT_MAX;
1384
1385 if (is_check_pages_enabled()) {
1386 if (free_page_is_bad(page))
1387 bad++;
1388 if (bad)
1389 return false;
1390 }
1391
1392 page_cpupid_reset_last(page);
1393 page->flags.f &= ~PAGE_FLAGS_CHECK_AT_PREP;
1394 reset_page_owner(page, order);
1395 page_table_check_free(page, order);
1396 pgalloc_tag_sub(page, nr: 1 << order);
1397
1398 if (!PageHighMem(page)) {
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, numpages: 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, numpages: 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, numpages: 1 << order);
1433
1434 return true;
1435}
1436
1437/*
1438 * Frees a number of pages from the PCP lists
1439 * Assumes all pages on list are in same zone.
1440 * count is the number of pages to free.
1441 */
1442static void free_pcppages_bulk(struct zone *zone, int count,
1443 struct per_cpu_pages *pcp,
1444 int pindex)
1445{
1446 unsigned long flags;
1447 unsigned int order;
1448 struct page *page;
1449
1450 /*
1451 * Ensure proper count is passed which otherwise would stuck in the
1452 * below while (list_empty(list)) loop.
1453 */
1454 count = min(pcp->count, count);
1455
1456 /* Ensure requested pindex is drained first. */
1457 pindex = pindex - 1;
1458
1459 spin_lock_irqsave(&zone->lock, flags);
1460
1461 while (count > 0) {
1462 struct list_head *list;
1463 int nr_pages;
1464
1465 /* Remove pages from lists in a round-robin fashion. */
1466 do {
1467 if (++pindex > NR_PCP_LISTS - 1)
1468 pindex = 0;
1469 list = &pcp->lists[pindex];
1470 } while (list_empty(head: list));
1471
1472 order = pindex_to_order(pindex);
1473 nr_pages = 1 << order;
1474 do {
1475 unsigned long pfn;
1476 int mt;
1477
1478 page = list_last_entry(list, struct page, pcp_list);
1479 pfn = page_to_pfn(page);
1480 mt = get_pfnblock_migratetype(page, pfn);
1481
1482 /* must delete to avoid corrupting pcp list */
1483 list_del(entry: &page->pcp_list);
1484 count -= nr_pages;
1485 pcp->count -= nr_pages;
1486
1487 __free_one_page(page, pfn, zone, order, migratetype: mt, FPI_NONE);
1488 trace_mm_page_pcpu_drain(page, order, migratetype: mt);
1489 } while (count > 0 && !list_empty(head: list));
1490 }
1491
1492 spin_unlock_irqrestore(lock: &zone->lock, flags);
1493}
1494
1495/* Split a multi-block free page into its individual pageblocks. */
1496static void split_large_buddy(struct zone *zone, struct page *page,
1497 unsigned long pfn, int order, fpi_t fpi)
1498{
1499 unsigned long end = pfn + (1 << order);
1500
1501 VM_WARN_ON_ONCE(!IS_ALIGNED(pfn, 1 << order));
1502 /* Caller removed page from freelist, buddy info cleared! */
1503 VM_WARN_ON_ONCE(PageBuddy(page));
1504
1505 if (order > pageblock_order)
1506 order = pageblock_order;
1507
1508 do {
1509 int mt = get_pfnblock_migratetype(page, pfn);
1510
1511 __free_one_page(page, pfn, zone, order, migratetype: mt, fpi_flags: fpi);
1512 pfn += 1 << order;
1513 if (pfn == end)
1514 break;
1515 page = pfn_to_page(pfn);
1516 } while (1);
1517}
1518
1519static void add_page_to_zone_llist(struct zone *zone, struct page *page,
1520 unsigned int order)
1521{
1522 /* Remember the order */
1523 page->private = order;
1524 /* Add the page to the free list */
1525 llist_add(new: &page->pcp_llist, head: &zone->trylock_free_pages);
1526}
1527
1528static void free_one_page(struct zone *zone, struct page *page,
1529 unsigned long pfn, unsigned int order,
1530 fpi_t fpi_flags)
1531{
1532 struct llist_head *llhead;
1533 unsigned long flags;
1534
1535 if (unlikely(fpi_flags & FPI_TRYLOCK)) {
1536 if (!spin_trylock_irqsave(&zone->lock, flags)) {
1537 add_page_to_zone_llist(zone, page, order);
1538 return;
1539 }
1540 } else {
1541 spin_lock_irqsave(&zone->lock, flags);
1542 }
1543
1544 /* The lock succeeded. Process deferred pages. */
1545 llhead = &zone->trylock_free_pages;
1546 if (unlikely(!llist_empty(llhead) && !(fpi_flags & FPI_TRYLOCK))) {
1547 struct llist_node *llnode;
1548 struct page *p, *tmp;
1549
1550 llnode = llist_del_all(head: llhead);
1551 llist_for_each_entry_safe(p, tmp, llnode, pcp_llist) {
1552 unsigned int p_order = p->private;
1553
1554 split_large_buddy(zone, page: p, page_to_pfn(p), order: p_order, fpi: fpi_flags);
1555 __count_vm_events(item: PGFREE, delta: 1 << p_order);
1556 }
1557 }
1558 split_large_buddy(zone, page, pfn, order, fpi: fpi_flags);
1559 spin_unlock_irqrestore(lock: &zone->lock, flags);
1560
1561 __count_vm_events(item: PGFREE, delta: 1 << order);
1562}
1563
1564static void __free_pages_ok(struct page *page, unsigned int order,
1565 fpi_t fpi_flags)
1566{
1567 unsigned long pfn = page_to_pfn(page);
1568 struct zone *zone = page_zone(page);
1569
1570 if (free_pages_prepare(page, order))
1571 free_one_page(zone, page, pfn, order, fpi_flags);
1572}
1573
1574void __meminit __free_pages_core(struct page *page, unsigned int order,
1575 enum meminit_context context)
1576{
1577 unsigned int nr_pages = 1 << order;
1578 struct page *p = page;
1579 unsigned int loop;
1580
1581 /*
1582 * When initializing the memmap, __init_single_page() sets the refcount
1583 * of all pages to 1 ("allocated"/"not free"). We have to set the
1584 * refcount of all involved pages to 0.
1585 *
1586 * Note that hotplugged memory pages are initialized to PageOffline().
1587 * Pages freed from memblock might be marked as reserved.
1588 */
1589 if (IS_ENABLED(CONFIG_MEMORY_HOTPLUG) &&
1590 unlikely(context == MEMINIT_HOTPLUG)) {
1591 for (loop = 0; loop < nr_pages; loop++, p++) {
1592 VM_WARN_ON_ONCE(PageReserved(p));
1593 __ClearPageOffline(page: p);
1594 set_page_count(page: p, v: 0);
1595 }
1596
1597 adjust_managed_page_count(page, count: nr_pages);
1598 } else {
1599 for (loop = 0; loop < nr_pages; loop++, p++) {
1600 __ClearPageReserved(page: p);
1601 set_page_count(page: p, v: 0);
1602 }
1603
1604 /* memblock adjusts totalram_pages() manually. */
1605 atomic_long_add(i: nr_pages, v: &page_zone(page)->managed_pages);
1606 }
1607
1608 if (page_contains_unaccepted(page, order)) {
1609 if (order == MAX_PAGE_ORDER && __free_unaccepted(page))
1610 return;
1611
1612 accept_memory(page_to_phys(page), PAGE_SIZE << order);
1613 }
1614
1615 /*
1616 * Bypass PCP and place fresh pages right to the tail, primarily
1617 * relevant for memory onlining.
1618 */
1619 __free_pages_ok(page, order, FPI_TO_TAIL);
1620}
1621
1622/*
1623 * Check that the whole (or subset of) a pageblock given by the interval of
1624 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1625 * with the migration of free compaction scanner.
1626 *
1627 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1628 *
1629 * It's possible on some configurations to have a setup like node0 node1 node0
1630 * i.e. it's possible that all pages within a zones range of pages do not
1631 * belong to a single zone. We assume that a border between node0 and node1
1632 * can occur within a single pageblock, but not a node0 node1 node0
1633 * interleaving within a single pageblock. It is therefore sufficient to check
1634 * the first and last page of a pageblock and avoid checking each individual
1635 * page in a pageblock.
1636 *
1637 * Note: the function may return non-NULL struct page even for a page block
1638 * which contains a memory hole (i.e. there is no physical memory for a subset
1639 * of the pfn range). For example, if the pageblock order is MAX_PAGE_ORDER, which
1640 * will fall into 2 sub-sections, and the end pfn of the pageblock may be hole
1641 * even though the start pfn is online and valid. This should be safe most of
1642 * the time because struct pages are still initialized via init_unavailable_range()
1643 * and pfn walkers shouldn't touch any physical memory range for which they do
1644 * not recognize any specific metadata in struct pages.
1645 */
1646struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1647 unsigned long end_pfn, struct zone *zone)
1648{
1649 struct page *start_page;
1650 struct page *end_page;
1651
1652 /* end_pfn is one past the range we are checking */
1653 end_pfn--;
1654
1655 if (!pfn_valid(pfn: end_pfn))
1656 return NULL;
1657
1658 start_page = pfn_to_online_page(start_pfn);
1659 if (!start_page)
1660 return NULL;
1661
1662 if (page_zone(page: start_page) != zone)
1663 return NULL;
1664
1665 end_page = pfn_to_page(end_pfn);
1666
1667 /* This gives a shorter code than deriving page_zone(end_page) */
1668 if (page_zone_id(page: start_page) != page_zone_id(page: end_page))
1669 return NULL;
1670
1671 return start_page;
1672}
1673
1674/*
1675 * The order of subdivision here is critical for the IO subsystem.
1676 * Please do not alter this order without good reasons and regression
1677 * testing. Specifically, as large blocks of memory are subdivided,
1678 * the order in which smaller blocks are delivered depends on the order
1679 * they're subdivided in this function. This is the primary factor
1680 * influencing the order in which pages are delivered to the IO
1681 * subsystem according to empirical testing, and this is also justified
1682 * by considering the behavior of a buddy system containing a single
1683 * large block of memory acted on by a series of small allocations.
1684 * This behavior is a critical factor in sglist merging's success.
1685 *
1686 * -- nyc
1687 */
1688static inline unsigned int expand(struct zone *zone, struct page *page, int low,
1689 int high, int migratetype)
1690{
1691 unsigned int size = 1 << high;
1692 unsigned int nr_added = 0;
1693
1694 while (high > low) {
1695 high--;
1696 size >>= 1;
1697 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1698
1699 /*
1700 * Mark as guard pages (or page), that will allow to
1701 * merge back to allocator when buddy will be freed.
1702 * Corresponding page table entries will not be touched,
1703 * pages will stay not present in virtual address space
1704 */
1705 if (set_page_guard(zone, page: &page[size], order: high))
1706 continue;
1707
1708 __add_to_free_list(page: &page[size], zone, order: high, migratetype, tail: false);
1709 set_buddy_order(page: &page[size], order: high);
1710 nr_added += size;
1711 }
1712
1713 return nr_added;
1714}
1715
1716static __always_inline void page_del_and_expand(struct zone *zone,
1717 struct page *page, int low,
1718 int high, int migratetype)
1719{
1720 int nr_pages = 1 << high;
1721
1722 __del_page_from_free_list(page, zone, order: high, migratetype);
1723 nr_pages -= expand(zone, page, low, high, migratetype);
1724 account_freepages(zone, nr_pages: -nr_pages, migratetype);
1725}
1726
1727static void check_new_page_bad(struct page *page)
1728{
1729 if (unlikely(PageHWPoison(page))) {
1730 /* Don't complain about hwpoisoned pages */
1731 if (PageBuddy(page))
1732 __ClearPageBuddy(page);
1733 return;
1734 }
1735
1736 bad_page(page,
1737 reason: page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
1738}
1739
1740/*
1741 * This page is about to be returned from the page allocator
1742 */
1743static bool check_new_page(struct page *page)
1744{
1745 if (likely(page_expected_state(page,
1746 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1747 return false;
1748
1749 check_new_page_bad(page);
1750 return true;
1751}
1752
1753static inline bool check_new_pages(struct page *page, unsigned int order)
1754{
1755 if (is_check_pages_enabled()) {
1756 for (int i = 0; i < (1 << order); i++) {
1757 struct page *p = page + i;
1758
1759 if (check_new_page(page: p))
1760 return true;
1761 }
1762 }
1763
1764 return false;
1765}
1766
1767static inline bool should_skip_kasan_unpoison(gfp_t flags)
1768{
1769 /* Don't skip if a software KASAN mode is enabled. */
1770 if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
1771 IS_ENABLED(CONFIG_KASAN_SW_TAGS))
1772 return false;
1773
1774 /* Skip, if hardware tag-based KASAN is not enabled. */
1775 if (!kasan_hw_tags_enabled())
1776 return true;
1777
1778 /*
1779 * With hardware tag-based KASAN enabled, skip if this has been
1780 * requested via __GFP_SKIP_KASAN.
1781 */
1782 return flags & __GFP_SKIP_KASAN;
1783}
1784
1785static inline bool should_skip_init(gfp_t flags)
1786{
1787 /* Don't skip, if hardware tag-based KASAN is not enabled. */
1788 if (!kasan_hw_tags_enabled())
1789 return false;
1790
1791 /* For hardware tag-based KASAN, skip if requested. */
1792 return (flags & __GFP_SKIP_ZERO);
1793}
1794
1795inline void post_alloc_hook(struct page *page, unsigned int order,
1796 gfp_t gfp_flags)
1797{
1798 bool init = !want_init_on_free() && want_init_on_alloc(flags: gfp_flags) &&
1799 !should_skip_init(flags: gfp_flags);
1800 bool zero_tags = init && (gfp_flags & __GFP_ZEROTAGS);
1801 int i;
1802
1803 set_page_private(page, private: 0);
1804
1805 arch_alloc_page(page, order);
1806 debug_pagealloc_map_pages(page, numpages: 1 << order);
1807
1808 /*
1809 * Page unpoisoning must happen before memory initialization.
1810 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
1811 * allocations and the page unpoisoning code will complain.
1812 */
1813 kernel_unpoison_pages(page, numpages: 1 << order);
1814
1815 /*
1816 * As memory initialization might be integrated into KASAN,
1817 * KASAN unpoisoning and memory initializion code must be
1818 * kept together to avoid discrepancies in behavior.
1819 */
1820
1821 /*
1822 * If memory tags should be zeroed
1823 * (which happens only when memory should be initialized as well).
1824 */
1825 if (zero_tags) {
1826 /* Initialize both memory and memory tags. */
1827 for (i = 0; i != 1 << order; ++i)
1828 tag_clear_highpage(page: page + i);
1829
1830 /* Take note that memory was initialized by the loop above. */
1831 init = false;
1832 }
1833 if (!should_skip_kasan_unpoison(flags: 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: page + i);
1845 }
1846 /* If memory is still not initialized, initialize it now. */
1847 if (init)
1848 kernel_init_pages(page, numpages: 1 << order);
1849
1850 set_page_owner(page, order, gfp_mask: gfp_flags);
1851 page_table_check_alloc(page, order);
1852 pgalloc_tag_add(page, current, nr: 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, low: order, high: current_order,
1895 migratetype);
1896 trace_mm_page_alloc_zone_locked(page, order, migratetype,
1897 percpu_refill: 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, pfn: start))
1984 return false;
1985 if (!zone_spans_pfn(zone, pfn: 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: &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), migratetype: 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 void change_pageblock_range(struct page *pageblock_page,
2155 int start_order, int migratetype)
2156{
2157 int nr_pageblocks = 1 << (start_order - pageblock_order);
2158
2159 while (nr_pageblocks--) {
2160 set_pageblock_migratetype(page: pageblock_page, migratetype);
2161 pageblock_page += pageblock_nr_pages;
2162 }
2163}
2164
2165static inline bool boost_watermark(struct zone *zone)
2166{
2167 unsigned long max_boost;
2168
2169 if (!watermark_boost_factor)
2170 return false;
2171 /*
2172 * Don't bother in zones that are unlikely to produce results.
2173 * On small machines, including kdump capture kernels running
2174 * in a small area, boosting the watermark can cause an out of
2175 * memory situation immediately.
2176 */
2177 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2178 return false;
2179
2180 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2181 watermark_boost_factor, 10000);
2182
2183 /*
2184 * high watermark may be uninitialised if fragmentation occurs
2185 * very early in boot so do not boost. We do not fall
2186 * through and boost by pageblock_nr_pages as failing
2187 * allocations that early means that reclaim is not going
2188 * to help and it may even be impossible to reclaim the
2189 * boosted watermark resulting in a hang.
2190 */
2191 if (!max_boost)
2192 return false;
2193
2194 max_boost = max(pageblock_nr_pages, max_boost);
2195
2196 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2197 max_boost);
2198
2199 return true;
2200}
2201
2202/*
2203 * When we are falling back to another migratetype during allocation, should we
2204 * try to claim an entire block to satisfy further allocations, instead of
2205 * polluting multiple pageblocks?
2206 */
2207static bool should_try_claim_block(unsigned int order, int start_mt)
2208{
2209 /*
2210 * Leaving this order check is intended, although there is
2211 * relaxed order check in next check. The reason is that
2212 * we can actually claim the whole pageblock if this condition met,
2213 * but, below check doesn't guarantee it and that is just heuristic
2214 * so could be changed anytime.
2215 */
2216 if (order >= pageblock_order)
2217 return true;
2218
2219 /*
2220 * Above a certain threshold, always try to claim, as it's likely there
2221 * will be more free pages in the pageblock.
2222 */
2223 if (order >= pageblock_order / 2)
2224 return true;
2225
2226 /*
2227 * Unmovable/reclaimable allocations would cause permanent
2228 * fragmentations if they fell back to allocating from a movable block
2229 * (polluting it), so we try to claim the whole block regardless of the
2230 * allocation size. Later movable allocations can always steal from this
2231 * block, which is less problematic.
2232 */
2233 if (start_mt == MIGRATE_RECLAIMABLE || start_mt == MIGRATE_UNMOVABLE)
2234 return true;
2235
2236 if (page_group_by_mobility_disabled)
2237 return true;
2238
2239 /*
2240 * Movable pages won't cause permanent fragmentation, so when you alloc
2241 * small pages, we just need to temporarily steal unmovable or
2242 * reclaimable pages that are closest to the request size. After a
2243 * while, memory compaction may occur to form large contiguous pages,
2244 * and the next movable allocation may not need to steal.
2245 */
2246 return false;
2247}
2248
2249/*
2250 * Check whether there is a suitable fallback freepage with requested order.
2251 * If claimable is true, this function returns fallback_mt only if
2252 * we would do this whole-block claiming. This would help to reduce
2253 * fragmentation due to mixed migratetype pages in one pageblock.
2254 */
2255int find_suitable_fallback(struct free_area *area, unsigned int order,
2256 int migratetype, bool claimable)
2257{
2258 int i;
2259
2260 if (claimable && !should_try_claim_block(order, start_mt: migratetype))
2261 return -2;
2262
2263 if (area->nr_free == 0)
2264 return -1;
2265
2266 for (i = 0; i < MIGRATE_PCPTYPES - 1 ; i++) {
2267 int fallback_mt = fallbacks[migratetype][i];
2268
2269 if (!free_area_empty(area, migratetype: fallback_mt))
2270 return fallback_mt;
2271 }
2272
2273 return -1;
2274}
2275
2276/*
2277 * This function implements actual block claiming behaviour. If order is large
2278 * enough, we can claim the whole pageblock for the requested migratetype. If
2279 * not, we check the pageblock for constituent pages; if at least half of the
2280 * pages are free or compatible, we can still claim the whole block, so pages
2281 * freed in the future will be put on the correct free list.
2282 */
2283static struct page *
2284try_to_claim_block(struct zone *zone, struct page *page,
2285 int current_order, int order, int start_type,
2286 int block_type, unsigned int alloc_flags)
2287{
2288 int free_pages, movable_pages, alike_pages;
2289 unsigned long start_pfn;
2290
2291 /* Take ownership for orders >= pageblock_order */
2292 if (current_order >= pageblock_order) {
2293 unsigned int nr_added;
2294
2295 del_page_from_free_list(page, zone, order: current_order, migratetype: block_type);
2296 change_pageblock_range(pageblock_page: page, start_order: current_order, migratetype: start_type);
2297 nr_added = expand(zone, page, low: order, high: current_order, migratetype: start_type);
2298 account_freepages(zone, nr_pages: nr_added, migratetype: start_type);
2299 return page;
2300 }
2301
2302 /*
2303 * Boost watermarks to increase reclaim pressure to reduce the
2304 * likelihood of future fallbacks. Wake kswapd now as the node
2305 * may be balanced overall and kswapd will not wake naturally.
2306 */
2307 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2308 set_bit(nr: ZONE_BOOSTED_WATERMARK, addr: &zone->flags);
2309
2310 /* moving whole block can fail due to zone boundary conditions */
2311 if (!prep_move_freepages_block(zone, page, start_pfn: &start_pfn, num_free: &free_pages,
2312 num_movable: &movable_pages))
2313 return NULL;
2314
2315 /*
2316 * Determine how many pages are compatible with our allocation.
2317 * For movable allocation, it's the number of movable pages which
2318 * we just obtained. For other types it's a bit more tricky.
2319 */
2320 if (start_type == MIGRATE_MOVABLE) {
2321 alike_pages = movable_pages;
2322 } else {
2323 /*
2324 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2325 * to MOVABLE pageblock, consider all non-movable pages as
2326 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2327 * vice versa, be conservative since we can't distinguish the
2328 * exact migratetype of non-movable pages.
2329 */
2330 if (block_type == MIGRATE_MOVABLE)
2331 alike_pages = pageblock_nr_pages
2332 - (free_pages + movable_pages);
2333 else
2334 alike_pages = 0;
2335 }
2336 /*
2337 * If a sufficient number of pages in the block are either free or of
2338 * compatible migratability as our allocation, claim the whole block.
2339 */
2340 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2341 page_group_by_mobility_disabled) {
2342 __move_freepages_block(zone, start_pfn, old_mt: block_type, new_mt: start_type);
2343 set_pageblock_migratetype(pfn_to_page(start_pfn), migratetype: start_type);
2344 return __rmqueue_smallest(zone, order, migratetype: start_type);
2345 }
2346
2347 return NULL;
2348}
2349
2350/*
2351 * Try to allocate from some fallback migratetype by claiming the entire block,
2352 * i.e. converting it to the allocation's start migratetype.
2353 *
2354 * The use of signed ints for order and current_order is a deliberate
2355 * deviation from the rest of this file, to make the for loop
2356 * condition simpler.
2357 */
2358static __always_inline struct page *
2359__rmqueue_claim(struct zone *zone, int order, int start_migratetype,
2360 unsigned int alloc_flags)
2361{
2362 struct free_area *area;
2363 int current_order;
2364 int min_order = order;
2365 struct page *page;
2366 int fallback_mt;
2367
2368 /*
2369 * Do not steal pages from freelists belonging to other pageblocks
2370 * i.e. orders < pageblock_order. If there are no local zones free,
2371 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2372 */
2373 if (order < pageblock_order && alloc_flags & ALLOC_NOFRAGMENT)
2374 min_order = pageblock_order;
2375
2376 /*
2377 * Find the largest available free page in the other list. This roughly
2378 * approximates finding the pageblock with the most free pages, which
2379 * would be too costly to do exactly.
2380 */
2381 for (current_order = MAX_PAGE_ORDER; current_order >= min_order;
2382 --current_order) {
2383 area = &(zone->free_area[current_order]);
2384 fallback_mt = find_suitable_fallback(area, order: current_order,
2385 migratetype: start_migratetype, claimable: true);
2386
2387 /* No block in that order */
2388 if (fallback_mt == -1)
2389 continue;
2390
2391 /* Advanced into orders too low to claim, abort */
2392 if (fallback_mt == -2)
2393 break;
2394
2395 page = get_page_from_free_area(area, migratetype: fallback_mt);
2396 page = try_to_claim_block(zone, page, current_order, order,
2397 start_type: start_migratetype, block_type: fallback_mt,
2398 alloc_flags);
2399 if (page) {
2400 trace_mm_page_alloc_extfrag(page, alloc_order: order, fallback_order: current_order,
2401 alloc_migratetype: start_migratetype, fallback_migratetype: fallback_mt);
2402 return page;
2403 }
2404 }
2405
2406 return NULL;
2407}
2408
2409/*
2410 * Try to steal a single page from some fallback migratetype. Leave the rest of
2411 * the block as its current migratetype, potentially causing fragmentation.
2412 */
2413static __always_inline struct page *
2414__rmqueue_steal(struct zone *zone, int order, int start_migratetype)
2415{
2416 struct free_area *area;
2417 int current_order;
2418 struct page *page;
2419 int fallback_mt;
2420
2421 for (current_order = order; current_order < NR_PAGE_ORDERS; current_order++) {
2422 area = &(zone->free_area[current_order]);
2423 fallback_mt = find_suitable_fallback(area, order: current_order,
2424 migratetype: start_migratetype, claimable: false);
2425 if (fallback_mt == -1)
2426 continue;
2427
2428 page = get_page_from_free_area(area, migratetype: fallback_mt);
2429 page_del_and_expand(zone, page, low: order, high: current_order, migratetype: fallback_mt);
2430 trace_mm_page_alloc_extfrag(page, alloc_order: order, fallback_order: current_order,
2431 alloc_migratetype: start_migratetype, fallback_migratetype: fallback_mt);
2432 return page;
2433 }
2434
2435 return NULL;
2436}
2437
2438enum rmqueue_mode {
2439 RMQUEUE_NORMAL,
2440 RMQUEUE_CMA,
2441 RMQUEUE_CLAIM,
2442 RMQUEUE_STEAL,
2443};
2444
2445/*
2446 * Do the hard work of removing an element from the buddy allocator.
2447 * Call me with the zone->lock already held.
2448 */
2449static __always_inline struct page *
2450__rmqueue(struct zone *zone, unsigned int order, int migratetype,
2451 unsigned int alloc_flags, enum rmqueue_mode *mode)
2452{
2453 struct page *page;
2454
2455 if (IS_ENABLED(CONFIG_CMA)) {
2456 /*
2457 * Balance movable allocations between regular and CMA areas by
2458 * allocating from CMA when over half of the zone's free memory
2459 * is in the CMA area.
2460 */
2461 if (alloc_flags & ALLOC_CMA &&
2462 zone_page_state(zone, item: NR_FREE_CMA_PAGES) >
2463 zone_page_state(zone, item: NR_FREE_PAGES) / 2) {
2464 page = __rmqueue_cma_fallback(zone, order);
2465 if (page)
2466 return page;
2467 }
2468 }
2469
2470 /*
2471 * First try the freelists of the requested migratetype, then try
2472 * fallbacks modes with increasing levels of fragmentation risk.
2473 *
2474 * The fallback logic is expensive and rmqueue_bulk() calls in
2475 * a loop with the zone->lock held, meaning the freelists are
2476 * not subject to any outside changes. Remember in *mode where
2477 * we found pay dirt, to save us the search on the next call.
2478 */
2479 switch (*mode) {
2480 case RMQUEUE_NORMAL:
2481 page = __rmqueue_smallest(zone, order, migratetype);
2482 if (page)
2483 return page;
2484 fallthrough;
2485 case RMQUEUE_CMA:
2486 if (alloc_flags & ALLOC_CMA) {
2487 page = __rmqueue_cma_fallback(zone, order);
2488 if (page) {
2489 *mode = RMQUEUE_CMA;
2490 return page;
2491 }
2492 }
2493 fallthrough;
2494 case RMQUEUE_CLAIM:
2495 page = __rmqueue_claim(zone, order, start_migratetype: migratetype, alloc_flags);
2496 if (page) {
2497 /* Replenished preferred freelist, back to normal mode. */
2498 *mode = RMQUEUE_NORMAL;
2499 return page;
2500 }
2501 fallthrough;
2502 case RMQUEUE_STEAL:
2503 if (!(alloc_flags & ALLOC_NOFRAGMENT)) {
2504 page = __rmqueue_steal(zone, order, start_migratetype: migratetype);
2505 if (page) {
2506 *mode = RMQUEUE_STEAL;
2507 return page;
2508 }
2509 }
2510 }
2511 return NULL;
2512}
2513
2514/*
2515 * Obtain a specified number of elements from the buddy allocator, all under
2516 * a single hold of the lock, for efficiency. Add them to the supplied list.
2517 * Returns the number of new pages which were placed at *list.
2518 */
2519static int rmqueue_bulk(struct zone *zone, unsigned int order,
2520 unsigned long count, struct list_head *list,
2521 int migratetype, unsigned int alloc_flags)
2522{
2523 enum rmqueue_mode rmqm = RMQUEUE_NORMAL;
2524 unsigned long flags;
2525 int i;
2526
2527 if (unlikely(alloc_flags & ALLOC_TRYLOCK)) {
2528 if (!spin_trylock_irqsave(&zone->lock, flags))
2529 return 0;
2530 } else {
2531 spin_lock_irqsave(&zone->lock, flags);
2532 }
2533 for (i = 0; i < count; ++i) {
2534 struct page *page = __rmqueue(zone, order, migratetype,
2535 alloc_flags, mode: &rmqm);
2536 if (unlikely(page == NULL))
2537 break;
2538
2539 /*
2540 * Split buddy pages returned by expand() are received here in
2541 * physical page order. The page is added to the tail of
2542 * caller's list. From the callers perspective, the linked list
2543 * is ordered by page number under some conditions. This is
2544 * useful for IO devices that can forward direction from the
2545 * head, thus also in the physical page order. This is useful
2546 * for IO devices that can merge IO requests if the physical
2547 * pages are ordered properly.
2548 */
2549 list_add_tail(new: &page->pcp_list, head: list);
2550 }
2551 spin_unlock_irqrestore(lock: &zone->lock, flags);
2552
2553 return i;
2554}
2555
2556/*
2557 * Called from the vmstat counter updater to decay the PCP high.
2558 * Return whether there are addition works to do.
2559 */
2560int decay_pcp_high(struct zone *zone, struct per_cpu_pages *pcp)
2561{
2562 int high_min, to_drain, batch;
2563 int todo = 0;
2564
2565 high_min = READ_ONCE(pcp->high_min);
2566 batch = READ_ONCE(pcp->batch);
2567 /*
2568 * Decrease pcp->high periodically to try to free possible
2569 * idle PCP pages. And, avoid to free too many pages to
2570 * control latency. This caps pcp->high decrement too.
2571 */
2572 if (pcp->high > high_min) {
2573 pcp->high = max3(pcp->count - (batch << CONFIG_PCP_BATCH_SCALE_MAX),
2574 pcp->high - (pcp->high >> 3), high_min);
2575 if (pcp->high > high_min)
2576 todo++;
2577 }
2578
2579 to_drain = pcp->count - pcp->high;
2580 if (to_drain > 0) {
2581 spin_lock(lock: &pcp->lock);
2582 free_pcppages_bulk(zone, count: to_drain, pcp, pindex: 0);
2583 spin_unlock(lock: &pcp->lock);
2584 todo++;
2585 }
2586
2587 return todo;
2588}
2589
2590#ifdef CONFIG_NUMA
2591/*
2592 * Called from the vmstat counter updater to drain pagesets of this
2593 * currently executing processor on remote nodes after they have
2594 * expired.
2595 */
2596void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2597{
2598 int to_drain, batch;
2599
2600 batch = READ_ONCE(pcp->batch);
2601 to_drain = min(pcp->count, batch);
2602 if (to_drain > 0) {
2603 spin_lock(lock: &pcp->lock);
2604 free_pcppages_bulk(zone, count: to_drain, pcp, pindex: 0);
2605 spin_unlock(lock: &pcp->lock);
2606 }
2607}
2608#endif
2609
2610/*
2611 * Drain pcplists of the indicated processor and zone.
2612 */
2613static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2614{
2615 struct per_cpu_pages *pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2616 int count;
2617
2618 do {
2619 spin_lock(lock: &pcp->lock);
2620 count = pcp->count;
2621 if (count) {
2622 int to_drain = min(count,
2623 pcp->batch << CONFIG_PCP_BATCH_SCALE_MAX);
2624
2625 free_pcppages_bulk(zone, count: to_drain, pcp, pindex: 0);
2626 count -= to_drain;
2627 }
2628 spin_unlock(lock: &pcp->lock);
2629 } while (count);
2630}
2631
2632/*
2633 * Drain pcplists of all zones on the indicated processor.
2634 */
2635static void drain_pages(unsigned int cpu)
2636{
2637 struct zone *zone;
2638
2639 for_each_populated_zone(zone) {
2640 drain_pages_zone(cpu, zone);
2641 }
2642}
2643
2644/*
2645 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2646 */
2647void drain_local_pages(struct zone *zone)
2648{
2649 int cpu = smp_processor_id();
2650
2651 if (zone)
2652 drain_pages_zone(cpu, zone);
2653 else
2654 drain_pages(cpu);
2655}
2656
2657/*
2658 * The implementation of drain_all_pages(), exposing an extra parameter to
2659 * drain on all cpus.
2660 *
2661 * drain_all_pages() is optimized to only execute on cpus where pcplists are
2662 * not empty. The check for non-emptiness can however race with a free to
2663 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
2664 * that need the guarantee that every CPU has drained can disable the
2665 * optimizing racy check.
2666 */
2667static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
2668{
2669 int cpu;
2670
2671 /*
2672 * Allocate in the BSS so we won't require allocation in
2673 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2674 */
2675 static cpumask_t cpus_with_pcps;
2676
2677 /*
2678 * Do not drain if one is already in progress unless it's specific to
2679 * a zone. Such callers are primarily CMA and memory hotplug and need
2680 * the drain to be complete when the call returns.
2681 */
2682 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2683 if (!zone)
2684 return;
2685 mutex_lock(lock: &pcpu_drain_mutex);
2686 }
2687
2688 /*
2689 * We don't care about racing with CPU hotplug event
2690 * as offline notification will cause the notified
2691 * cpu to drain that CPU pcps and on_each_cpu_mask
2692 * disables preemption as part of its processing
2693 */
2694 for_each_online_cpu(cpu) {
2695 struct per_cpu_pages *pcp;
2696 struct zone *z;
2697 bool has_pcps = false;
2698
2699 if (force_all_cpus) {
2700 /*
2701 * The pcp.count check is racy, some callers need a
2702 * guarantee that no cpu is missed.
2703 */
2704 has_pcps = true;
2705 } else if (zone) {
2706 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2707 if (pcp->count)
2708 has_pcps = true;
2709 } else {
2710 for_each_populated_zone(z) {
2711 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
2712 if (pcp->count) {
2713 has_pcps = true;
2714 break;
2715 }
2716 }
2717 }
2718
2719 if (has_pcps)
2720 cpumask_set_cpu(cpu, dstp: &cpus_with_pcps);
2721 else
2722 cpumask_clear_cpu(cpu, dstp: &cpus_with_pcps);
2723 }
2724
2725 for_each_cpu(cpu, &cpus_with_pcps) {
2726 if (zone)
2727 drain_pages_zone(cpu, zone);
2728 else
2729 drain_pages(cpu);
2730 }
2731
2732 mutex_unlock(lock: &pcpu_drain_mutex);
2733}
2734
2735/*
2736 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2737 *
2738 * When zone parameter is non-NULL, spill just the single zone's pages.
2739 */
2740void drain_all_pages(struct zone *zone)
2741{
2742 __drain_all_pages(zone, force_all_cpus: false);
2743}
2744
2745static int nr_pcp_free(struct per_cpu_pages *pcp, int batch, int high, bool free_high)
2746{
2747 int min_nr_free, max_nr_free;
2748
2749 /* Free as much as possible if batch freeing high-order pages. */
2750 if (unlikely(free_high))
2751 return min(pcp->count, batch << CONFIG_PCP_BATCH_SCALE_MAX);
2752
2753 /* Check for PCP disabled or boot pageset */
2754 if (unlikely(high < batch))
2755 return 1;
2756
2757 /* Leave at least pcp->batch pages on the list */
2758 min_nr_free = batch;
2759 max_nr_free = high - batch;
2760
2761 /*
2762 * Increase the batch number to the number of the consecutive
2763 * freed pages to reduce zone lock contention.
2764 */
2765 batch = clamp_t(int, pcp->free_count, min_nr_free, max_nr_free);
2766
2767 return batch;
2768}
2769
2770static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
2771 int batch, bool free_high)
2772{
2773 int high, high_min, high_max;
2774
2775 high_min = READ_ONCE(pcp->high_min);
2776 high_max = READ_ONCE(pcp->high_max);
2777 high = pcp->high = clamp(pcp->high, high_min, high_max);
2778
2779 if (unlikely(!high))
2780 return 0;
2781
2782 if (unlikely(free_high)) {
2783 pcp->high = max(high - (batch << CONFIG_PCP_BATCH_SCALE_MAX),
2784 high_min);
2785 return 0;
2786 }
2787
2788 /*
2789 * If reclaim is active, limit the number of pages that can be
2790 * stored on pcp lists
2791 */
2792 if (test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags)) {
2793 int free_count = max_t(int, pcp->free_count, batch);
2794
2795 pcp->high = max(high - free_count, high_min);
2796 return min(batch << 2, pcp->high);
2797 }
2798
2799 if (high_min == high_max)
2800 return high;
2801
2802 if (test_bit(ZONE_BELOW_HIGH, &zone->flags)) {
2803 int free_count = max_t(int, pcp->free_count, batch);
2804
2805 pcp->high = max(high - free_count, high_min);
2806 high = max(pcp->count, high_min);
2807 } else if (pcp->count >= high) {
2808 int need_high = pcp->free_count + batch;
2809
2810 /* pcp->high should be large enough to hold batch freed pages */
2811 if (pcp->high < need_high)
2812 pcp->high = clamp(need_high, high_min, high_max);
2813 }
2814
2815 return high;
2816}
2817
2818static void free_frozen_page_commit(struct zone *zone,
2819 struct per_cpu_pages *pcp, struct page *page, int migratetype,
2820 unsigned int order, fpi_t fpi_flags)
2821{
2822 int high, batch;
2823 int pindex;
2824 bool free_high = false;
2825
2826 /*
2827 * On freeing, reduce the number of pages that are batch allocated.
2828 * See nr_pcp_alloc() where alloc_factor is increased for subsequent
2829 * allocations.
2830 */
2831 pcp->alloc_factor >>= 1;
2832 __count_vm_events(item: PGFREE, delta: 1 << order);
2833 pindex = order_to_pindex(migratetype, order);
2834 list_add(new: &page->pcp_list, head: &pcp->lists[pindex]);
2835 pcp->count += 1 << order;
2836
2837 batch = READ_ONCE(pcp->batch);
2838 /*
2839 * As high-order pages other than THP's stored on PCP can contribute
2840 * to fragmentation, limit the number stored when PCP is heavily
2841 * freeing without allocation. The remainder after bulk freeing
2842 * stops will be drained from vmstat refresh context.
2843 */
2844 if (order && order <= PAGE_ALLOC_COSTLY_ORDER) {
2845 free_high = (pcp->free_count >= (batch + pcp->high_min / 2) &&
2846 (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) &&
2847 (!(pcp->flags & PCPF_FREE_HIGH_BATCH) ||
2848 pcp->count >= batch));
2849 pcp->flags |= PCPF_PREV_FREE_HIGH_ORDER;
2850 } else if (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) {
2851 pcp->flags &= ~PCPF_PREV_FREE_HIGH_ORDER;
2852 }
2853 if (pcp->free_count < (batch << CONFIG_PCP_BATCH_SCALE_MAX))
2854 pcp->free_count += (1 << order);
2855
2856 if (unlikely(fpi_flags & FPI_TRYLOCK)) {
2857 /*
2858 * Do not attempt to take a zone lock. Let pcp->count get
2859 * over high mark temporarily.
2860 */
2861 return;
2862 }
2863
2864 high = nr_pcp_high(pcp, zone, batch, free_high);
2865 if (pcp->count < high)
2866 return;
2867
2868 free_pcppages_bulk(zone, count: nr_pcp_free(pcp, batch, high, free_high),
2869 pcp, pindex);
2870 if (test_bit(ZONE_BELOW_HIGH, &zone->flags) &&
2871 zone_watermark_ok(z: zone, order: 0, mark: high_wmark_pages(z: zone),
2872 highest_zoneidx: ZONE_MOVABLE, alloc_flags: 0)) {
2873 struct pglist_data *pgdat = zone->zone_pgdat;
2874 clear_bit(nr: ZONE_BELOW_HIGH, addr: &zone->flags);
2875
2876 /*
2877 * Assume that memory pressure on this node is gone and may be
2878 * in a reclaimable state. If a memory fallback node exists,
2879 * direct reclaim may not have been triggered, causing a
2880 * 'hopeless node' to stay in that state for a while. Let
2881 * kswapd work again by resetting kswapd_failures.
2882 */
2883 if (atomic_read(v: &pgdat->kswapd_failures) >= MAX_RECLAIM_RETRIES &&
2884 next_memory_node(nid: pgdat->node_id) < MAX_NUMNODES)
2885 atomic_set(v: &pgdat->kswapd_failures, i: 0);
2886 }
2887}
2888
2889/*
2890 * Free a pcp page
2891 */
2892static void __free_frozen_pages(struct page *page, unsigned int order,
2893 fpi_t fpi_flags)
2894{
2895 unsigned long __maybe_unused UP_flags;
2896 struct per_cpu_pages *pcp;
2897 struct zone *zone;
2898 unsigned long pfn = page_to_pfn(page);
2899 int migratetype;
2900
2901 if (!pcp_allowed_order(order)) {
2902 __free_pages_ok(page, order, fpi_flags);
2903 return;
2904 }
2905
2906 if (!free_pages_prepare(page, order))
2907 return;
2908
2909 /*
2910 * We only track unmovable, reclaimable and movable on pcp lists.
2911 * Place ISOLATE pages on the isolated list because they are being
2912 * offlined but treat HIGHATOMIC and CMA as movable pages so we can
2913 * get those areas back if necessary. Otherwise, we may have to free
2914 * excessively into the page allocator
2915 */
2916 zone = page_zone(page);
2917 migratetype = get_pfnblock_migratetype(page, pfn);
2918 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
2919 if (unlikely(is_migrate_isolate(migratetype))) {
2920 free_one_page(zone, page, pfn, order, fpi_flags);
2921 return;
2922 }
2923 migratetype = MIGRATE_MOVABLE;
2924 }
2925
2926 if (unlikely((fpi_flags & FPI_TRYLOCK) && IS_ENABLED(CONFIG_PREEMPT_RT)
2927 && (in_nmi() || in_hardirq()))) {
2928 add_page_to_zone_llist(zone, page, order);
2929 return;
2930 }
2931 pcp_trylock_prepare(UP_flags);
2932 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2933 if (pcp) {
2934 free_frozen_page_commit(zone, pcp, page, migratetype, order, fpi_flags);
2935 pcp_spin_unlock(pcp);
2936 } else {
2937 free_one_page(zone, page, pfn, order, fpi_flags);
2938 }
2939 pcp_trylock_finish(UP_flags);
2940}
2941
2942void free_frozen_pages(struct page *page, unsigned int order)
2943{
2944 __free_frozen_pages(page, order, FPI_NONE);
2945}
2946
2947/*
2948 * Free a batch of folios
2949 */
2950void free_unref_folios(struct folio_batch *folios)
2951{
2952 unsigned long __maybe_unused UP_flags;
2953 struct per_cpu_pages *pcp = NULL;
2954 struct zone *locked_zone = NULL;
2955 int i, j;
2956
2957 /* Prepare folios for freeing */
2958 for (i = 0, j = 0; i < folios->nr; i++) {
2959 struct folio *folio = folios->folios[i];
2960 unsigned long pfn = folio_pfn(folio);
2961 unsigned int order = folio_order(folio);
2962
2963 if (!free_pages_prepare(page: &folio->page, order))
2964 continue;
2965 /*
2966 * Free orders not handled on the PCP directly to the
2967 * allocator.
2968 */
2969 if (!pcp_allowed_order(order)) {
2970 free_one_page(zone: folio_zone(folio), page: &folio->page,
2971 pfn, order, FPI_NONE);
2972 continue;
2973 }
2974 folio->private = (void *)(unsigned long)order;
2975 if (j != i)
2976 folios->folios[j] = folio;
2977 j++;
2978 }
2979 folios->nr = j;
2980
2981 for (i = 0; i < folios->nr; i++) {
2982 struct folio *folio = folios->folios[i];
2983 struct zone *zone = folio_zone(folio);
2984 unsigned long pfn = folio_pfn(folio);
2985 unsigned int order = (unsigned long)folio->private;
2986 int migratetype;
2987
2988 folio->private = NULL;
2989 migratetype = get_pfnblock_migratetype(page: &folio->page, pfn);
2990
2991 /* Different zone requires a different pcp lock */
2992 if (zone != locked_zone ||
2993 is_migrate_isolate(migratetype)) {
2994 if (pcp) {
2995 pcp_spin_unlock(pcp);
2996 pcp_trylock_finish(UP_flags);
2997 locked_zone = NULL;
2998 pcp = NULL;
2999 }
3000
3001 /*
3002 * Free isolated pages directly to the
3003 * allocator, see comment in free_frozen_pages.
3004 */
3005 if (is_migrate_isolate(migratetype)) {
3006 free_one_page(zone, page: &folio->page, pfn,
3007 order, FPI_NONE);
3008 continue;
3009 }
3010
3011 /*
3012 * trylock is necessary as folios may be getting freed
3013 * from IRQ or SoftIRQ context after an IO completion.
3014 */
3015 pcp_trylock_prepare(UP_flags);
3016 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
3017 if (unlikely(!pcp)) {
3018 pcp_trylock_finish(UP_flags);
3019 free_one_page(zone, page: &folio->page, pfn,
3020 order, FPI_NONE);
3021 continue;
3022 }
3023 locked_zone = zone;
3024 }
3025
3026 /*
3027 * Non-isolated types over MIGRATE_PCPTYPES get added
3028 * to the MIGRATE_MOVABLE pcp list.
3029 */
3030 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
3031 migratetype = MIGRATE_MOVABLE;
3032
3033 trace_mm_page_free_batched(page: &folio->page);
3034 free_frozen_page_commit(zone, pcp, page: &folio->page, migratetype,
3035 order, FPI_NONE);
3036 }
3037
3038 if (pcp) {
3039 pcp_spin_unlock(pcp);
3040 pcp_trylock_finish(UP_flags);
3041 }
3042 folio_batch_reinit(fbatch: folios);
3043}
3044
3045/*
3046 * split_page takes a non-compound higher-order page, and splits it into
3047 * n (1<<order) sub-pages: page[0..n]
3048 * Each sub-page must be freed individually.
3049 *
3050 * Note: this is probably too low level an operation for use in drivers.
3051 * Please consult with lkml before using this in your driver.
3052 */
3053void split_page(struct page *page, unsigned int order)
3054{
3055 int i;
3056
3057 VM_BUG_ON_PAGE(PageCompound(page), page);
3058 VM_BUG_ON_PAGE(!page_count(page), page);
3059
3060 for (i = 1; i < (1 << order); i++)
3061 set_page_refcounted(page + i);
3062 split_page_owner(page, old_order: order, new_order: 0);
3063 pgalloc_tag_split(page_folio(page), old_order: order, new_order: 0);
3064 split_page_memcg(first: page, order);
3065}
3066EXPORT_SYMBOL_GPL(split_page);
3067
3068int __isolate_free_page(struct page *page, unsigned int order)
3069{
3070 struct zone *zone = page_zone(page);
3071 int mt = get_pageblock_migratetype(page);
3072
3073 if (!is_migrate_isolate(migratetype: mt)) {
3074 unsigned long watermark;
3075 /*
3076 * Obey watermarks as if the page was being allocated. We can
3077 * emulate a high-order watermark check with a raised order-0
3078 * watermark, because we already know our high-order page
3079 * exists.
3080 */
3081 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3082 if (!zone_watermark_ok(z: zone, order: 0, mark: watermark, highest_zoneidx: 0, ALLOC_CMA))
3083 return 0;
3084 }
3085
3086 del_page_from_free_list(page, zone, order, migratetype: mt);
3087
3088 /*
3089 * Set the pageblock if the isolated page is at least half of a
3090 * pageblock
3091 */
3092 if (order >= pageblock_order - 1) {
3093 struct page *endpage = page + (1 << order) - 1;
3094 for (; page < endpage; page += pageblock_nr_pages) {
3095 int mt = get_pageblock_migratetype(page);
3096 /*
3097 * Only change normal pageblocks (i.e., they can merge
3098 * with others)
3099 */
3100 if (migratetype_is_mergeable(mt))
3101 move_freepages_block(zone, page, old_mt: mt,
3102 new_mt: MIGRATE_MOVABLE);
3103 }
3104 }
3105
3106 return 1UL << order;
3107}
3108
3109/**
3110 * __putback_isolated_page - Return a now-isolated page back where we got it
3111 * @page: Page that was isolated
3112 * @order: Order of the isolated page
3113 * @mt: The page's pageblock's migratetype
3114 *
3115 * This function is meant to return a page pulled from the free lists via
3116 * __isolate_free_page back to the free lists they were pulled from.
3117 */
3118void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3119{
3120 struct zone *zone = page_zone(page);
3121
3122 /* zone lock should be held when this function is called */
3123 lockdep_assert_held(&zone->lock);
3124
3125 /* Return isolated page to tail of freelist. */
3126 __free_one_page(page, page_to_pfn(page), zone, order, migratetype: mt,
3127 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3128}
3129
3130/*
3131 * Update NUMA hit/miss statistics
3132 */
3133static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
3134 long nr_account)
3135{
3136#ifdef CONFIG_NUMA
3137 enum numa_stat_item local_stat = NUMA_LOCAL;
3138
3139 /* skip numa counters update if numa stats is disabled */
3140 if (!static_branch_likely(&vm_numa_stat_key))
3141 return;
3142
3143 if (zone_to_nid(zone: z) != numa_node_id())
3144 local_stat = NUMA_OTHER;
3145
3146 if (zone_to_nid(zone: z) == zone_to_nid(zone: preferred_zone))
3147 __count_numa_events(zone: z, item: NUMA_HIT, delta: nr_account);
3148 else {
3149 __count_numa_events(zone: z, item: NUMA_MISS, delta: nr_account);
3150 __count_numa_events(zone: preferred_zone, item: NUMA_FOREIGN, delta: nr_account);
3151 }
3152 __count_numa_events(zone: z, item: local_stat, delta: nr_account);
3153#endif
3154}
3155
3156static __always_inline
3157struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone,
3158 unsigned int order, unsigned int alloc_flags,
3159 int migratetype)
3160{
3161 struct page *page;
3162 unsigned long flags;
3163
3164 do {
3165 page = NULL;
3166 if (unlikely(alloc_flags & ALLOC_TRYLOCK)) {
3167 if (!spin_trylock_irqsave(&zone->lock, flags))
3168 return NULL;
3169 } else {
3170 spin_lock_irqsave(&zone->lock, flags);
3171 }
3172 if (alloc_flags & ALLOC_HIGHATOMIC)
3173 page = __rmqueue_smallest(zone, order, migratetype: MIGRATE_HIGHATOMIC);
3174 if (!page) {
3175 enum rmqueue_mode rmqm = RMQUEUE_NORMAL;
3176
3177 page = __rmqueue(zone, order, migratetype, alloc_flags, mode: &rmqm);
3178
3179 /*
3180 * If the allocation fails, allow OOM handling and
3181 * order-0 (atomic) allocs access to HIGHATOMIC
3182 * reserves as failing now is worse than failing a
3183 * high-order atomic allocation in the future.
3184 */
3185 if (!page && (alloc_flags & (ALLOC_OOM|ALLOC_NON_BLOCK)))
3186 page = __rmqueue_smallest(zone, order, migratetype: MIGRATE_HIGHATOMIC);
3187
3188 if (!page) {
3189 spin_unlock_irqrestore(lock: &zone->lock, flags);
3190 return NULL;
3191 }
3192 }
3193 spin_unlock_irqrestore(lock: &zone->lock, flags);
3194 } while (check_new_pages(page, order));
3195
3196 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3197 zone_statistics(preferred_zone, z: zone, nr_account: 1);
3198
3199 return page;
3200}
3201
3202static int nr_pcp_alloc(struct per_cpu_pages *pcp, struct zone *zone, int order)
3203{
3204 int high, base_batch, batch, max_nr_alloc;
3205 int high_max, high_min;
3206
3207 base_batch = READ_ONCE(pcp->batch);
3208 high_min = READ_ONCE(pcp->high_min);
3209 high_max = READ_ONCE(pcp->high_max);
3210 high = pcp->high = clamp(pcp->high, high_min, high_max);
3211
3212 /* Check for PCP disabled or boot pageset */
3213 if (unlikely(high < base_batch))
3214 return 1;
3215
3216 if (order)
3217 batch = base_batch;
3218 else
3219 batch = (base_batch << pcp->alloc_factor);
3220
3221 /*
3222 * If we had larger pcp->high, we could avoid to allocate from
3223 * zone.
3224 */
3225 if (high_min != high_max && !test_bit(ZONE_BELOW_HIGH, &zone->flags))
3226 high = pcp->high = min(high + batch, high_max);
3227
3228 if (!order) {
3229 max_nr_alloc = max(high - pcp->count - base_batch, base_batch);
3230 /*
3231 * Double the number of pages allocated each time there is
3232 * subsequent allocation of order-0 pages without any freeing.
3233 */
3234 if (batch <= max_nr_alloc &&
3235 pcp->alloc_factor < CONFIG_PCP_BATCH_SCALE_MAX)
3236 pcp->alloc_factor++;
3237 batch = min(batch, max_nr_alloc);
3238 }
3239
3240 /*
3241 * Scale batch relative to order if batch implies free pages
3242 * can be stored on the PCP. Batch can be 1 for small zones or
3243 * for boot pagesets which should never store free pages as
3244 * the pages may belong to arbitrary zones.
3245 */
3246 if (batch > 1)
3247 batch = max(batch >> order, 2);
3248
3249 return batch;
3250}
3251
3252/* Remove page from the per-cpu list, caller must protect the list */
3253static inline
3254struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
3255 int migratetype,
3256 unsigned int alloc_flags,
3257 struct per_cpu_pages *pcp,
3258 struct list_head *list)
3259{
3260 struct page *page;
3261
3262 do {
3263 if (list_empty(head: list)) {
3264 int batch = nr_pcp_alloc(pcp, zone, order);
3265 int alloced;
3266
3267 alloced = rmqueue_bulk(zone, order,
3268 count: batch, list,
3269 migratetype, alloc_flags);
3270
3271 pcp->count += alloced << order;
3272 if (unlikely(list_empty(list)))
3273 return NULL;
3274 }
3275
3276 page = list_first_entry(list, struct page, pcp_list);
3277 list_del(entry: &page->pcp_list);
3278 pcp->count -= 1 << order;
3279 } while (check_new_pages(page, order));
3280
3281 return page;
3282}
3283
3284/* Lock and remove page from the per-cpu list */
3285static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3286 struct zone *zone, unsigned int order,
3287 int migratetype, unsigned int alloc_flags)
3288{
3289 struct per_cpu_pages *pcp;
3290 struct list_head *list;
3291 struct page *page;
3292 unsigned long __maybe_unused UP_flags;
3293
3294 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
3295 pcp_trylock_prepare(UP_flags);
3296 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
3297 if (!pcp) {
3298 pcp_trylock_finish(UP_flags);
3299 return NULL;
3300 }
3301
3302 /*
3303 * On allocation, reduce the number of pages that are batch freed.
3304 * See nr_pcp_free() where free_factor is increased for subsequent
3305 * frees.
3306 */
3307 pcp->free_count >>= 1;
3308 list = &pcp->lists[order_to_pindex(migratetype, order)];
3309 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
3310 pcp_spin_unlock(pcp);
3311 pcp_trylock_finish(UP_flags);
3312 if (page) {
3313 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3314 zone_statistics(preferred_zone, z: zone, nr_account: 1);
3315 }
3316 return page;
3317}
3318
3319/*
3320 * Allocate a page from the given zone.
3321 * Use pcplists for THP or "cheap" high-order allocations.
3322 */
3323
3324/*
3325 * Do not instrument rmqueue() with KMSAN. This function may call
3326 * __msan_poison_alloca() through a call to set_pfnblock_migratetype().
3327 * If __msan_poison_alloca() attempts to allocate pages for the stack depot, it
3328 * may call rmqueue() again, which will result in a deadlock.
3329 */
3330__no_sanitize_memory
3331static inline
3332struct page *rmqueue(struct zone *preferred_zone,
3333 struct zone *zone, unsigned int order,
3334 gfp_t gfp_flags, unsigned int alloc_flags,
3335 int migratetype)
3336{
3337 struct page *page;
3338
3339 if (likely(pcp_allowed_order(order))) {
3340 page = rmqueue_pcplist(preferred_zone, zone, order,
3341 migratetype, alloc_flags);
3342 if (likely(page))
3343 goto out;
3344 }
3345
3346 page = rmqueue_buddy(preferred_zone, zone, order, alloc_flags,
3347 migratetype);
3348
3349out:
3350 /* Separate test+clear to avoid unnecessary atomics */
3351 if ((alloc_flags & ALLOC_KSWAPD) &&
3352 unlikely(test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags))) {
3353 clear_bit(nr: ZONE_BOOSTED_WATERMARK, addr: &zone->flags);
3354 wakeup_kswapd(zone, gfp_mask: 0, order: 0, zone_idx(zone));
3355 }
3356
3357 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3358 return page;
3359}
3360
3361/*
3362 * Reserve the pageblock(s) surrounding an allocation request for
3363 * exclusive use of high-order atomic allocations if there are no
3364 * empty page blocks that contain a page with a suitable order
3365 */
3366static void reserve_highatomic_pageblock(struct page *page, int order,
3367 struct zone *zone)
3368{
3369 int mt;
3370 unsigned long max_managed, flags;
3371
3372 /*
3373 * The number reserved as: minimum is 1 pageblock, maximum is
3374 * roughly 1% of a zone. But if 1% of a zone falls below a
3375 * pageblock size, then don't reserve any pageblocks.
3376 * Check is race-prone but harmless.
3377 */
3378 if ((zone_managed_pages(zone) / 100) < pageblock_nr_pages)
3379 return;
3380 max_managed = ALIGN((zone_managed_pages(zone) / 100), pageblock_nr_pages);
3381 if (zone->nr_reserved_highatomic >= max_managed)
3382 return;
3383
3384 spin_lock_irqsave(&zone->lock, flags);
3385
3386 /* Recheck the nr_reserved_highatomic limit under the lock */
3387 if (zone->nr_reserved_highatomic >= max_managed)
3388 goto out_unlock;
3389
3390 /* Yoink! */
3391 mt = get_pageblock_migratetype(page);
3392 /* Only reserve normal pageblocks (i.e., they can merge with others) */
3393 if (!migratetype_is_mergeable(mt))
3394 goto out_unlock;
3395
3396 if (order < pageblock_order) {
3397 if (move_freepages_block(zone, page, old_mt: mt, new_mt: MIGRATE_HIGHATOMIC) == -1)
3398 goto out_unlock;
3399 zone->nr_reserved_highatomic += pageblock_nr_pages;
3400 } else {
3401 change_pageblock_range(pageblock_page: page, start_order: order, migratetype: MIGRATE_HIGHATOMIC);
3402 zone->nr_reserved_highatomic += 1 << order;
3403 }
3404
3405out_unlock:
3406 spin_unlock_irqrestore(lock: &zone->lock, flags);
3407}
3408
3409/*
3410 * Used when an allocation is about to fail under memory pressure. This
3411 * potentially hurts the reliability of high-order allocations when under
3412 * intense memory pressure but failed atomic allocations should be easier
3413 * to recover from than an OOM.
3414 *
3415 * If @force is true, try to unreserve pageblocks even though highatomic
3416 * pageblock is exhausted.
3417 */
3418static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
3419 bool force)
3420{
3421 struct zonelist *zonelist = ac->zonelist;
3422 unsigned long flags;
3423 struct zoneref *z;
3424 struct zone *zone;
3425 struct page *page;
3426 int order;
3427 int ret;
3428
3429 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
3430 ac->nodemask) {
3431 /*
3432 * Preserve at least one pageblock unless memory pressure
3433 * is really high.
3434 */
3435 if (!force && zone->nr_reserved_highatomic <=
3436 pageblock_nr_pages)
3437 continue;
3438
3439 spin_lock_irqsave(&zone->lock, flags);
3440 for (order = 0; order < NR_PAGE_ORDERS; order++) {
3441 struct free_area *area = &(zone->free_area[order]);
3442 unsigned long size;
3443
3444 page = get_page_from_free_area(area, migratetype: MIGRATE_HIGHATOMIC);
3445 if (!page)
3446 continue;
3447
3448 size = max(pageblock_nr_pages, 1UL << order);
3449 /*
3450 * It should never happen but changes to
3451 * locking could inadvertently allow a per-cpu
3452 * drain to add pages to MIGRATE_HIGHATOMIC
3453 * while unreserving so be safe and watch for
3454 * underflows.
3455 */
3456 if (WARN_ON_ONCE(size > zone->nr_reserved_highatomic))
3457 size = zone->nr_reserved_highatomic;
3458 zone->nr_reserved_highatomic -= size;
3459
3460 /*
3461 * Convert to ac->migratetype and avoid the normal
3462 * pageblock stealing heuristics. Minimally, the caller
3463 * is doing the work and needs the pages. More
3464 * importantly, if the block was always converted to
3465 * MIGRATE_UNMOVABLE or another type then the number
3466 * of pageblocks that cannot be completely freed
3467 * may increase.
3468 */
3469 if (order < pageblock_order)
3470 ret = move_freepages_block(zone, page,
3471 old_mt: MIGRATE_HIGHATOMIC,
3472 new_mt: ac->migratetype);
3473 else {
3474 move_to_free_list(page, zone, order,
3475 old_mt: MIGRATE_HIGHATOMIC,
3476 new_mt: ac->migratetype);
3477 change_pageblock_range(pageblock_page: page, start_order: order,
3478 migratetype: ac->migratetype);
3479 ret = 1;
3480 }
3481 /*
3482 * Reserving the block(s) already succeeded,
3483 * so this should not fail on zone boundaries.
3484 */
3485 WARN_ON_ONCE(ret == -1);
3486 if (ret > 0) {
3487 spin_unlock_irqrestore(lock: &zone->lock, flags);
3488 return ret;
3489 }
3490 }
3491 spin_unlock_irqrestore(lock: &zone->lock, flags);
3492 }
3493
3494 return false;
3495}
3496
3497static inline long __zone_watermark_unusable_free(struct zone *z,
3498 unsigned int order, unsigned int alloc_flags)
3499{
3500 long unusable_free = (1 << order) - 1;
3501
3502 /*
3503 * If the caller does not have rights to reserves below the min
3504 * watermark then subtract the free pages reserved for highatomic.
3505 */
3506 if (likely(!(alloc_flags & ALLOC_RESERVES)))
3507 unusable_free += READ_ONCE(z->nr_free_highatomic);
3508
3509#ifdef CONFIG_CMA
3510 /* If allocation can't use CMA areas don't use free CMA pages */
3511 if (!(alloc_flags & ALLOC_CMA))
3512 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3513#endif
3514
3515 return unusable_free;
3516}
3517
3518/*
3519 * Return true if free base pages are above 'mark'. For high-order checks it
3520 * will return true of the order-0 watermark is reached and there is at least
3521 * one free page of a suitable size. Checking now avoids taking the zone lock
3522 * to check in the allocation paths if no pages are free.
3523 */
3524bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3525 int highest_zoneidx, unsigned int alloc_flags,
3526 long free_pages)
3527{
3528 long min = mark;
3529 int o;
3530
3531 /* free_pages may go negative - that's OK */
3532 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3533
3534 if (unlikely(alloc_flags & ALLOC_RESERVES)) {
3535 /*
3536 * __GFP_HIGH allows access to 50% of the min reserve as well
3537 * as OOM.
3538 */
3539 if (alloc_flags & ALLOC_MIN_RESERVE) {
3540 min -= min / 2;
3541
3542 /*
3543 * Non-blocking allocations (e.g. GFP_ATOMIC) can
3544 * access more reserves than just __GFP_HIGH. Other
3545 * non-blocking allocations requests such as GFP_NOWAIT
3546 * or (GFP_KERNEL & ~__GFP_DIRECT_RECLAIM) do not get
3547 * access to the min reserve.
3548 */
3549 if (alloc_flags & ALLOC_NON_BLOCK)
3550 min -= min / 4;
3551 }
3552
3553 /*
3554 * OOM victims can try even harder than the normal reserve
3555 * users on the grounds that it's definitely going to be in
3556 * the exit path shortly and free memory. Any allocation it
3557 * makes during the free path will be small and short-lived.
3558 */
3559 if (alloc_flags & ALLOC_OOM)
3560 min -= min / 2;
3561 }
3562
3563 /*
3564 * Check watermarks for an order-0 allocation request. If these
3565 * are not met, then a high-order request also cannot go ahead
3566 * even if a suitable page happened to be free.
3567 */
3568 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3569 return false;
3570
3571 /* If this is an order-0 request then the watermark is fine */
3572 if (!order)
3573 return true;
3574
3575 /* For a high-order request, check at least one suitable page is free */
3576 for (o = order; o < NR_PAGE_ORDERS; o++) {
3577 struct free_area *area = &z->free_area[o];
3578 int mt;
3579
3580 if (!area->nr_free)
3581 continue;
3582
3583 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3584 if (!free_area_empty(area, migratetype: mt))
3585 return true;
3586 }
3587
3588#ifdef CONFIG_CMA
3589 if ((alloc_flags & ALLOC_CMA) &&
3590 !free_area_empty(area, MIGRATE_CMA)) {
3591 return true;
3592 }
3593#endif
3594 if ((alloc_flags & (ALLOC_HIGHATOMIC|ALLOC_OOM)) &&
3595 !free_area_empty(area, migratetype: MIGRATE_HIGHATOMIC)) {
3596 return true;
3597 }
3598 }
3599 return false;
3600}
3601
3602bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3603 int highest_zoneidx, unsigned int alloc_flags)
3604{
3605 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3606 free_pages: zone_page_state(zone: z, item: NR_FREE_PAGES));
3607}
3608
3609static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3610 unsigned long mark, int highest_zoneidx,
3611 unsigned int alloc_flags, gfp_t gfp_mask)
3612{
3613 long free_pages;
3614
3615 free_pages = zone_page_state(zone: z, item: NR_FREE_PAGES);
3616
3617 /*
3618 * Fast check for order-0 only. If this fails then the reserves
3619 * need to be calculated.
3620 */
3621 if (!order) {
3622 long usable_free;
3623 long reserved;
3624
3625 usable_free = free_pages;
3626 reserved = __zone_watermark_unusable_free(z, order: 0, alloc_flags);
3627
3628 /* reserved may over estimate high-atomic reserves. */
3629 usable_free -= min(usable_free, reserved);
3630 if (usable_free > mark + z->lowmem_reserve[highest_zoneidx])
3631 return true;
3632 }
3633
3634 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3635 free_pages))
3636 return true;
3637
3638 /*
3639 * Ignore watermark boosting for __GFP_HIGH order-0 allocations
3640 * when checking the min watermark. The min watermark is the
3641 * point where boosting is ignored so that kswapd is woken up
3642 * when below the low watermark.
3643 */
3644 if (unlikely(!order && (alloc_flags & ALLOC_MIN_RESERVE) && z->watermark_boost
3645 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3646 mark = z->_watermark[WMARK_MIN];
3647 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3648 alloc_flags, free_pages);
3649 }
3650
3651 return false;
3652}
3653
3654#ifdef CONFIG_NUMA
3655int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
3656
3657static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3658{
3659 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3660 node_reclaim_distance;
3661}
3662#else /* CONFIG_NUMA */
3663static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3664{
3665 return true;
3666}
3667#endif /* CONFIG_NUMA */
3668
3669/*
3670 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3671 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3672 * premature use of a lower zone may cause lowmem pressure problems that
3673 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3674 * probably too small. It only makes sense to spread allocations to avoid
3675 * fragmentation between the Normal and DMA32 zones.
3676 */
3677static inline unsigned int
3678alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3679{
3680 unsigned int alloc_flags;
3681
3682 /*
3683 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3684 * to save a branch.
3685 */
3686 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3687
3688 if (defrag_mode) {
3689 alloc_flags |= ALLOC_NOFRAGMENT;
3690 return alloc_flags;
3691 }
3692
3693#ifdef CONFIG_ZONE_DMA32
3694 if (!zone)
3695 return alloc_flags;
3696
3697 if (zone_idx(zone) != ZONE_NORMAL)
3698 return alloc_flags;
3699
3700 /*
3701 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3702 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3703 * on UMA that if Normal is populated then so is DMA32.
3704 */
3705 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3706 if (nr_online_nodes > 1 && !populated_zone(zone: --zone))
3707 return alloc_flags;
3708
3709 alloc_flags |= ALLOC_NOFRAGMENT;
3710#endif /* CONFIG_ZONE_DMA32 */
3711 return alloc_flags;
3712}
3713
3714/* Must be called after current_gfp_context() which can change gfp_mask */
3715static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
3716 unsigned int alloc_flags)
3717{
3718#ifdef CONFIG_CMA
3719 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3720 alloc_flags |= ALLOC_CMA;
3721#endif
3722 return alloc_flags;
3723}
3724
3725/*
3726 * get_page_from_freelist goes through the zonelist trying to allocate
3727 * a page.
3728 */
3729static struct page *
3730get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3731 const struct alloc_context *ac)
3732{
3733 struct zoneref *z;
3734 struct zone *zone;
3735 struct pglist_data *last_pgdat = NULL;
3736 bool last_pgdat_dirty_ok = false;
3737 bool no_fallback;
3738 bool skip_kswapd_nodes = nr_online_nodes > 1;
3739 bool skipped_kswapd_nodes = false;
3740
3741retry:
3742 /*
3743 * Scan zonelist, looking for a zone with enough free.
3744 * See also cpuset_current_node_allowed() comment in kernel/cgroup/cpuset.c.
3745 */
3746 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3747 z = ac->preferred_zoneref;
3748 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3749 ac->nodemask) {
3750 struct page *page;
3751 unsigned long mark;
3752
3753 if (cpusets_enabled() &&
3754 (alloc_flags & ALLOC_CPUSET) &&
3755 !__cpuset_zone_allowed(z: zone, gfp_mask))
3756 continue;
3757 /*
3758 * When allocating a page cache page for writing, we
3759 * want to get it from a node that is within its dirty
3760 * limit, such that no single node holds more than its
3761 * proportional share of globally allowed dirty pages.
3762 * The dirty limits take into account the node's
3763 * lowmem reserves and high watermark so that kswapd
3764 * should be able to balance it without having to
3765 * write pages from its LRU list.
3766 *
3767 * XXX: For now, allow allocations to potentially
3768 * exceed the per-node dirty limit in the slowpath
3769 * (spread_dirty_pages unset) before going into reclaim,
3770 * which is important when on a NUMA setup the allowed
3771 * nodes are together not big enough to reach the
3772 * global limit. The proper fix for these situations
3773 * will require awareness of nodes in the
3774 * dirty-throttling and the flusher threads.
3775 */
3776 if (ac->spread_dirty_pages) {
3777 if (last_pgdat != zone->zone_pgdat) {
3778 last_pgdat = zone->zone_pgdat;
3779 last_pgdat_dirty_ok = node_dirty_ok(pgdat: zone->zone_pgdat);
3780 }
3781
3782 if (!last_pgdat_dirty_ok)
3783 continue;
3784 }
3785
3786 if (no_fallback && !defrag_mode && nr_online_nodes > 1 &&
3787 zone != zonelist_zone(zoneref: ac->preferred_zoneref)) {
3788 int local_nid;
3789
3790 /*
3791 * If moving to a remote node, retry but allow
3792 * fragmenting fallbacks. Locality is more important
3793 * than fragmentation avoidance.
3794 */
3795 local_nid = zonelist_node_idx(zoneref: ac->preferred_zoneref);
3796 if (zone_to_nid(zone) != local_nid) {
3797 alloc_flags &= ~ALLOC_NOFRAGMENT;
3798 goto retry;
3799 }
3800 }
3801
3802 /*
3803 * If kswapd is already active on a node, keep looking
3804 * for other nodes that might be idle. This can happen
3805 * if another process has NUMA bindings and is causing
3806 * kswapd wakeups on only some nodes. Avoid accidental
3807 * "node_reclaim_mode"-like behavior in this case.
3808 */
3809 if (skip_kswapd_nodes &&
3810 !waitqueue_active(wq_head: &zone->zone_pgdat->kswapd_wait)) {
3811 skipped_kswapd_nodes = true;
3812 continue;
3813 }
3814
3815 cond_accept_memory(zone, order, alloc_flags);
3816
3817 /*
3818 * Detect whether the number of free pages is below high
3819 * watermark. If so, we will decrease pcp->high and free
3820 * PCP pages in free path to reduce the possibility of
3821 * premature page reclaiming. Detection is done here to
3822 * avoid to do that in hotter free path.
3823 */
3824 if (test_bit(ZONE_BELOW_HIGH, &zone->flags))
3825 goto check_alloc_wmark;
3826
3827 mark = high_wmark_pages(z: zone);
3828 if (zone_watermark_fast(z: zone, order, mark,
3829 highest_zoneidx: ac->highest_zoneidx, alloc_flags,
3830 gfp_mask))
3831 goto try_this_zone;
3832 else
3833 set_bit(nr: ZONE_BELOW_HIGH, addr: &zone->flags);
3834
3835check_alloc_wmark:
3836 mark = wmark_pages(z: zone, w: alloc_flags & ALLOC_WMARK_MASK);
3837 if (!zone_watermark_fast(z: zone, order, mark,
3838 highest_zoneidx: ac->highest_zoneidx, alloc_flags,
3839 gfp_mask)) {
3840 int ret;
3841
3842 if (cond_accept_memory(zone, order, alloc_flags))
3843 goto try_this_zone;
3844
3845 /*
3846 * Watermark failed for this zone, but see if we can
3847 * grow this zone if it contains deferred pages.
3848 */
3849 if (deferred_pages_enabled()) {
3850 if (_deferred_grow_zone(zone, order))
3851 goto try_this_zone;
3852 }
3853 /* Checked here to keep the fast path fast */
3854 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3855 if (alloc_flags & ALLOC_NO_WATERMARKS)
3856 goto try_this_zone;
3857
3858 if (!node_reclaim_enabled() ||
3859 !zone_allows_reclaim(local_zone: zonelist_zone(zoneref: ac->preferred_zoneref), zone))
3860 continue;
3861
3862 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3863 switch (ret) {
3864 case NODE_RECLAIM_NOSCAN:
3865 /* did not scan */
3866 continue;
3867 case NODE_RECLAIM_FULL:
3868 /* scanned but unreclaimable */
3869 continue;
3870 default:
3871 /* did we reclaim enough */
3872 if (zone_watermark_ok(z: zone, order, mark,
3873 highest_zoneidx: ac->highest_zoneidx, alloc_flags))
3874 goto try_this_zone;
3875
3876 continue;
3877 }
3878 }
3879
3880try_this_zone:
3881 page = rmqueue(preferred_zone: zonelist_zone(zoneref: ac->preferred_zoneref), zone, order,
3882 gfp_flags: gfp_mask, alloc_flags, migratetype: ac->migratetype);
3883 if (page) {
3884 prep_new_page(page, order, gfp_flags: gfp_mask, alloc_flags);
3885
3886 /*
3887 * If this is a high-order atomic allocation then check
3888 * if the pageblock should be reserved for the future
3889 */
3890 if (unlikely(alloc_flags & ALLOC_HIGHATOMIC))
3891 reserve_highatomic_pageblock(page, order, zone);
3892
3893 return page;
3894 } else {
3895 if (cond_accept_memory(zone, order, alloc_flags))
3896 goto try_this_zone;
3897
3898 /* Try again if zone has deferred pages */
3899 if (deferred_pages_enabled()) {
3900 if (_deferred_grow_zone(zone, order))
3901 goto try_this_zone;
3902 }
3903 }
3904 }
3905
3906 /*
3907 * If we skipped over nodes with active kswapds and found no
3908 * idle nodes, retry and place anywhere the watermarks permit.
3909 */
3910 if (skip_kswapd_nodes && skipped_kswapd_nodes) {
3911 skip_kswapd_nodes = false;
3912 goto retry;
3913 }
3914
3915 /*
3916 * It's possible on a UMA machine to get through all zones that are
3917 * fragmented. If avoiding fragmentation, reset and try again.
3918 */
3919 if (no_fallback && !defrag_mode) {
3920 alloc_flags &= ~ALLOC_NOFRAGMENT;
3921 goto retry;
3922 }
3923
3924 return NULL;
3925}
3926
3927static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3928{
3929 unsigned int filter = SHOW_MEM_FILTER_NODES;
3930
3931 /*
3932 * This documents exceptions given to allocations in certain
3933 * contexts that are allowed to allocate outside current's set
3934 * of allowed nodes.
3935 */
3936 if (!(gfp_mask & __GFP_NOMEMALLOC))
3937 if (tsk_is_oom_victim(current) ||
3938 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3939 filter &= ~SHOW_MEM_FILTER_NODES;
3940 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3941 filter &= ~SHOW_MEM_FILTER_NODES;
3942
3943 __show_mem(flags: filter, nodemask, max_zone_idx: gfp_zone(flags: gfp_mask));
3944}
3945
3946void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3947{
3948 struct va_format vaf;
3949 va_list args;
3950 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3951
3952 if ((gfp_mask & __GFP_NOWARN) ||
3953 !__ratelimit(&nopage_rs) ||
3954 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
3955 return;
3956
3957 va_start(args, fmt);
3958 vaf.fmt = fmt;
3959 vaf.va = &args;
3960 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3961 current->comm, &vaf, gfp_mask, &gfp_mask,
3962 nodemask_pr_args(nodemask));
3963 va_end(args);
3964
3965 cpuset_print_current_mems_allowed();
3966 pr_cont("\n");
3967 dump_stack();
3968 warn_alloc_show_mem(gfp_mask, nodemask);
3969}
3970
3971static inline struct page *
3972__alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3973 unsigned int alloc_flags,
3974 const struct alloc_context *ac)
3975{
3976 struct page *page;
3977
3978 page = get_page_from_freelist(gfp_mask, order,
3979 alloc_flags: alloc_flags|ALLOC_CPUSET, ac);
3980 /*
3981 * fallback to ignore cpuset restriction if our nodes
3982 * are depleted
3983 */
3984 if (!page)
3985 page = get_page_from_freelist(gfp_mask, order,
3986 alloc_flags, ac);
3987 return page;
3988}
3989
3990static inline struct page *
3991__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3992 const struct alloc_context *ac, unsigned long *did_some_progress)
3993{
3994 struct oom_control oc = {
3995 .zonelist = ac->zonelist,
3996 .nodemask = ac->nodemask,
3997 .memcg = NULL,
3998 .gfp_mask = gfp_mask,
3999 .order = order,
4000 };
4001 struct page *page;
4002
4003 *did_some_progress = 0;
4004
4005 /*
4006 * Acquire the oom lock. If that fails, somebody else is
4007 * making progress for us.
4008 */
4009 if (!mutex_trylock(lock: &oom_lock)) {
4010 *did_some_progress = 1;
4011 schedule_timeout_uninterruptible(timeout: 1);
4012 return NULL;
4013 }
4014
4015 /*
4016 * Go through the zonelist yet one more time, keep very high watermark
4017 * here, this is only to catch a parallel oom killing, we must fail if
4018 * we're still under heavy pressure. But make sure that this reclaim
4019 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4020 * allocation which will never fail due to oom_lock already held.
4021 */
4022 page = get_page_from_freelist(gfp_mask: (gfp_mask | __GFP_HARDWALL) &
4023 ~__GFP_DIRECT_RECLAIM, order,
4024 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4025 if (page)
4026 goto out;
4027
4028 /* Coredumps can quickly deplete all memory reserves */
4029 if (current->flags & PF_DUMPCORE)
4030 goto out;
4031 /* The OOM killer will not help higher order allocs */
4032 if (order > PAGE_ALLOC_COSTLY_ORDER)
4033 goto out;
4034 /*
4035 * We have already exhausted all our reclaim opportunities without any
4036 * success so it is time to admit defeat. We will skip the OOM killer
4037 * because it is very likely that the caller has a more reasonable
4038 * fallback than shooting a random task.
4039 *
4040 * The OOM killer may not free memory on a specific node.
4041 */
4042 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4043 goto out;
4044 /* The OOM killer does not needlessly kill tasks for lowmem */
4045 if (ac->highest_zoneidx < ZONE_NORMAL)
4046 goto out;
4047 if (pm_suspended_storage())
4048 goto out;
4049 /*
4050 * XXX: GFP_NOFS allocations should rather fail than rely on
4051 * other request to make a forward progress.
4052 * We are in an unfortunate situation where out_of_memory cannot
4053 * do much for this context but let's try it to at least get
4054 * access to memory reserved if the current task is killed (see
4055 * out_of_memory). Once filesystems are ready to handle allocation
4056 * failures more gracefully we should just bail out here.
4057 */
4058
4059 /* Exhausted what can be done so it's blame time */
4060 if (out_of_memory(oc: &oc) ||
4061 WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) {
4062 *did_some_progress = 1;
4063
4064 /*
4065 * Help non-failing allocations by giving them access to memory
4066 * reserves
4067 */
4068 if (gfp_mask & __GFP_NOFAIL)
4069 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4070 ALLOC_NO_WATERMARKS, ac);
4071 }
4072out:
4073 mutex_unlock(lock: &oom_lock);
4074 return page;
4075}
4076
4077/*
4078 * Maximum number of compaction retries with a progress before OOM
4079 * killer is consider as the only way to move forward.
4080 */
4081#define MAX_COMPACT_RETRIES 16
4082
4083#ifdef CONFIG_COMPACTION
4084/* Try memory compaction for high-order allocations before reclaim */
4085static struct page *
4086__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4087 unsigned int alloc_flags, const struct alloc_context *ac,
4088 enum compact_priority prio, enum compact_result *compact_result)
4089{
4090 struct page *page = NULL;
4091 unsigned long pflags;
4092 unsigned int noreclaim_flag;
4093
4094 if (!order)
4095 return NULL;
4096
4097 psi_memstall_enter(flags: &pflags);
4098 delayacct_compact_start();
4099 noreclaim_flag = memalloc_noreclaim_save();
4100
4101 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4102 prio, page: &page);
4103
4104 memalloc_noreclaim_restore(flags: noreclaim_flag);
4105 psi_memstall_leave(flags: &pflags);
4106 delayacct_compact_end();
4107
4108 if (*compact_result == COMPACT_SKIPPED)
4109 return NULL;
4110 /*
4111 * At least in one zone compaction wasn't deferred or skipped, so let's
4112 * count a compaction stall
4113 */
4114 count_vm_event(item: COMPACTSTALL);
4115
4116 /* Prep a captured page if available */
4117 if (page)
4118 prep_new_page(page, order, gfp_flags: gfp_mask, alloc_flags);
4119
4120 /* Try get a page from the freelist if available */
4121 if (!page)
4122 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4123
4124 if (page) {
4125 struct zone *zone = page_zone(page);
4126
4127 zone->compact_blockskip_flush = false;
4128 compaction_defer_reset(zone, order, alloc_success: true);
4129 count_vm_event(item: COMPACTSUCCESS);
4130 return page;
4131 }
4132
4133 /*
4134 * It's bad if compaction run occurs and fails. The most likely reason
4135 * is that pages exist, but not enough to satisfy watermarks.
4136 */
4137 count_vm_event(item: COMPACTFAIL);
4138
4139 cond_resched();
4140
4141 return NULL;
4142}
4143
4144static inline bool
4145should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4146 enum compact_result compact_result,
4147 enum compact_priority *compact_priority,
4148 int *compaction_retries)
4149{
4150 int max_retries = MAX_COMPACT_RETRIES;
4151 int min_priority;
4152 bool ret = false;
4153 int retries = *compaction_retries;
4154 enum compact_priority priority = *compact_priority;
4155
4156 if (!order)
4157 return false;
4158
4159 if (fatal_signal_pending(current))
4160 return false;
4161
4162 /*
4163 * Compaction was skipped due to a lack of free order-0
4164 * migration targets. Continue if reclaim can help.
4165 */
4166 if (compact_result == COMPACT_SKIPPED) {
4167 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4168 goto out;
4169 }
4170
4171 /*
4172 * Compaction managed to coalesce some page blocks, but the
4173 * allocation failed presumably due to a race. Retry some.
4174 */
4175 if (compact_result == COMPACT_SUCCESS) {
4176 /*
4177 * !costly requests are much more important than
4178 * __GFP_RETRY_MAYFAIL costly ones because they are de
4179 * facto nofail and invoke OOM killer to move on while
4180 * costly can fail and users are ready to cope with
4181 * that. 1/4 retries is rather arbitrary but we would
4182 * need much more detailed feedback from compaction to
4183 * make a better decision.
4184 */
4185 if (order > PAGE_ALLOC_COSTLY_ORDER)
4186 max_retries /= 4;
4187
4188 if (++(*compaction_retries) <= max_retries) {
4189 ret = true;
4190 goto out;
4191 }
4192 }
4193
4194 /*
4195 * Compaction failed. Retry with increasing priority.
4196 */
4197 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4198 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4199
4200 if (*compact_priority > min_priority) {
4201 (*compact_priority)--;
4202 *compaction_retries = 0;
4203 ret = true;
4204 }
4205out:
4206 trace_compact_retry(order, priority, result: compact_result, retries, max_retries, ret);
4207 return ret;
4208}
4209#else
4210static inline struct page *
4211__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4212 unsigned int alloc_flags, const struct alloc_context *ac,
4213 enum compact_priority prio, enum compact_result *compact_result)
4214{
4215 *compact_result = COMPACT_SKIPPED;
4216 return NULL;
4217}
4218
4219static inline bool
4220should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4221 enum compact_result compact_result,
4222 enum compact_priority *compact_priority,
4223 int *compaction_retries)
4224{
4225 struct zone *zone;
4226 struct zoneref *z;
4227
4228 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4229 return false;
4230
4231 /*
4232 * There are setups with compaction disabled which would prefer to loop
4233 * inside the allocator rather than hit the oom killer prematurely.
4234 * Let's give them a good hope and keep retrying while the order-0
4235 * watermarks are OK.
4236 */
4237 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4238 ac->highest_zoneidx, ac->nodemask) {
4239 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4240 ac->highest_zoneidx, alloc_flags))
4241 return true;
4242 }
4243 return false;
4244}
4245#endif /* CONFIG_COMPACTION */
4246
4247#ifdef CONFIG_LOCKDEP
4248static struct lockdep_map __fs_reclaim_map =
4249 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4250
4251static bool __need_reclaim(gfp_t gfp_mask)
4252{
4253 /* no reclaim without waiting on it */
4254 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4255 return false;
4256
4257 /* this guy won't enter reclaim */
4258 if (current->flags & PF_MEMALLOC)
4259 return false;
4260
4261 if (gfp_mask & __GFP_NOLOCKDEP)
4262 return false;
4263
4264 return true;
4265}
4266
4267void __fs_reclaim_acquire(unsigned long ip)
4268{
4269 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
4270}
4271
4272void __fs_reclaim_release(unsigned long ip)
4273{
4274 lock_release(&__fs_reclaim_map, ip);
4275}
4276
4277void fs_reclaim_acquire(gfp_t gfp_mask)
4278{
4279 gfp_mask = current_gfp_context(gfp_mask);
4280
4281 if (__need_reclaim(gfp_mask)) {
4282 if (gfp_mask & __GFP_FS)
4283 __fs_reclaim_acquire(_RET_IP_);
4284
4285#ifdef CONFIG_MMU_NOTIFIER
4286 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4287 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4288#endif
4289
4290 }
4291}
4292EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4293
4294void fs_reclaim_release(gfp_t gfp_mask)
4295{
4296 gfp_mask = current_gfp_context(gfp_mask);
4297
4298 if (__need_reclaim(gfp_mask)) {
4299 if (gfp_mask & __GFP_FS)
4300 __fs_reclaim_release(_RET_IP_);
4301 }
4302}
4303EXPORT_SYMBOL_GPL(fs_reclaim_release);
4304#endif
4305
4306/*
4307 * Zonelists may change due to hotplug during allocation. Detect when zonelists
4308 * have been rebuilt so allocation retries. Reader side does not lock and
4309 * retries the allocation if zonelist changes. Writer side is protected by the
4310 * embedded spin_lock.
4311 */
4312static DEFINE_SEQLOCK(zonelist_update_seq);
4313
4314static unsigned int zonelist_iter_begin(void)
4315{
4316 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
4317 return read_seqbegin(sl: &zonelist_update_seq);
4318
4319 return 0;
4320}
4321
4322static unsigned int check_retry_zonelist(unsigned int seq)
4323{
4324 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
4325 return read_seqretry(sl: &zonelist_update_seq, start: seq);
4326
4327 return seq;
4328}
4329
4330/* Perform direct synchronous page reclaim */
4331static unsigned long
4332__perform_reclaim(gfp_t gfp_mask, unsigned int order,
4333 const struct alloc_context *ac)
4334{
4335 unsigned int noreclaim_flag;
4336 unsigned long progress;
4337
4338 cond_resched();
4339
4340 /* We now go into synchronous reclaim */
4341 cpuset_memory_pressure_bump();
4342 fs_reclaim_acquire(gfp_mask);
4343 noreclaim_flag = memalloc_noreclaim_save();
4344
4345 progress = try_to_free_pages(zonelist: ac->zonelist, order, gfp_mask,
4346 mask: ac->nodemask);
4347
4348 memalloc_noreclaim_restore(flags: noreclaim_flag);
4349 fs_reclaim_release(gfp_mask);
4350
4351 cond_resched();
4352
4353 return progress;
4354}
4355
4356/* The really slow allocator path where we enter direct reclaim */
4357static inline struct page *
4358__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4359 unsigned int alloc_flags, const struct alloc_context *ac,
4360 unsigned long *did_some_progress)
4361{
4362 struct page *page = NULL;
4363 unsigned long pflags;
4364 bool drained = false;
4365
4366 psi_memstall_enter(flags: &pflags);
4367 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4368 if (unlikely(!(*did_some_progress)))
4369 goto out;
4370
4371retry:
4372 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4373
4374 /*
4375 * If an allocation failed after direct reclaim, it could be because
4376 * pages are pinned on the per-cpu lists or in high alloc reserves.
4377 * Shrink them and try again
4378 */
4379 if (!page && !drained) {
4380 unreserve_highatomic_pageblock(ac, force: false);
4381 drain_all_pages(NULL);
4382 drained = true;
4383 goto retry;
4384 }
4385out:
4386 psi_memstall_leave(flags: &pflags);
4387
4388 return page;
4389}
4390
4391static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4392 const struct alloc_context *ac)
4393{
4394 struct zoneref *z;
4395 struct zone *zone;
4396 pg_data_t *last_pgdat = NULL;
4397 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4398 unsigned int reclaim_order;
4399
4400 if (defrag_mode)
4401 reclaim_order = max(order, pageblock_order);
4402 else
4403 reclaim_order = order;
4404
4405 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4406 ac->nodemask) {
4407 if (!managed_zone(zone))
4408 continue;
4409 if (last_pgdat == zone->zone_pgdat)
4410 continue;
4411 wakeup_kswapd(zone, gfp_mask, order: reclaim_order, highest_zoneidx);
4412 last_pgdat = zone->zone_pgdat;
4413 }
4414}
4415
4416static inline unsigned int
4417gfp_to_alloc_flags(gfp_t gfp_mask, unsigned int order)
4418{
4419 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4420
4421 /*
4422 * __GFP_HIGH is assumed to be the same as ALLOC_MIN_RESERVE
4423 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4424 * to save two branches.
4425 */
4426 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_MIN_RESERVE);
4427 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4428
4429 /*
4430 * The caller may dip into page reserves a bit more if the caller
4431 * cannot run direct reclaim, or if the caller has realtime scheduling
4432 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4433 * set both ALLOC_NON_BLOCK and ALLOC_MIN_RESERVE(__GFP_HIGH).
4434 */
4435 alloc_flags |= (__force int)
4436 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4437
4438 if (!(gfp_mask & __GFP_DIRECT_RECLAIM)) {
4439 /*
4440 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4441 * if it can't schedule.
4442 */
4443 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
4444 alloc_flags |= ALLOC_NON_BLOCK;
4445
4446 if (order > 0 && (alloc_flags & ALLOC_MIN_RESERVE))
4447 alloc_flags |= ALLOC_HIGHATOMIC;
4448 }
4449
4450 /*
4451 * Ignore cpuset mems for non-blocking __GFP_HIGH (probably
4452 * GFP_ATOMIC) rather than fail, see the comment for
4453 * cpuset_current_node_allowed().
4454 */
4455 if (alloc_flags & ALLOC_MIN_RESERVE)
4456 alloc_flags &= ~ALLOC_CPUSET;
4457 } else if (unlikely(rt_or_dl_task(current)) && in_task())
4458 alloc_flags |= ALLOC_MIN_RESERVE;
4459
4460 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4461
4462 if (defrag_mode)
4463 alloc_flags |= ALLOC_NOFRAGMENT;
4464
4465 return alloc_flags;
4466}
4467
4468static bool oom_reserves_allowed(struct task_struct *tsk)
4469{
4470 if (!tsk_is_oom_victim(tsk))
4471 return false;
4472
4473 /*
4474 * !MMU doesn't have oom reaper so give access to memory reserves
4475 * only to the thread with TIF_MEMDIE set
4476 */
4477 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4478 return false;
4479
4480 return true;
4481}
4482
4483/*
4484 * Distinguish requests which really need access to full memory
4485 * reserves from oom victims which can live with a portion of it
4486 */
4487static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4488{
4489 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4490 return 0;
4491 if (gfp_mask & __GFP_MEMALLOC)
4492 return ALLOC_NO_WATERMARKS;
4493 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4494 return ALLOC_NO_WATERMARKS;
4495 if (!in_interrupt()) {
4496 if (current->flags & PF_MEMALLOC)
4497 return ALLOC_NO_WATERMARKS;
4498 else if (oom_reserves_allowed(current))
4499 return ALLOC_OOM;
4500 }
4501
4502 return 0;
4503}
4504
4505bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4506{
4507 return !!__gfp_pfmemalloc_flags(gfp_mask);
4508}
4509
4510/*
4511 * Checks whether it makes sense to retry the reclaim to make a forward progress
4512 * for the given allocation request.
4513 *
4514 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4515 * without success, or when we couldn't even meet the watermark if we
4516 * reclaimed all remaining pages on the LRU lists.
4517 *
4518 * Returns true if a retry is viable or false to enter the oom path.
4519 */
4520static inline bool
4521should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4522 struct alloc_context *ac, int alloc_flags,
4523 bool did_some_progress, int *no_progress_loops)
4524{
4525 struct zone *zone;
4526 struct zoneref *z;
4527 bool ret = false;
4528
4529 /*
4530 * Costly allocations might have made a progress but this doesn't mean
4531 * their order will become available due to high fragmentation so
4532 * always increment the no progress counter for them
4533 */
4534 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4535 *no_progress_loops = 0;
4536 else
4537 (*no_progress_loops)++;
4538
4539 if (*no_progress_loops > MAX_RECLAIM_RETRIES)
4540 goto out;
4541
4542
4543 /*
4544 * Keep reclaiming pages while there is a chance this will lead
4545 * somewhere. If none of the target zones can satisfy our allocation
4546 * request even if all reclaimable pages are considered then we are
4547 * screwed and have to go OOM.
4548 */
4549 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4550 ac->highest_zoneidx, ac->nodemask) {
4551 unsigned long available;
4552 unsigned long reclaimable;
4553 unsigned long min_wmark = min_wmark_pages(z: zone);
4554 bool wmark;
4555
4556 if (cpusets_enabled() &&
4557 (alloc_flags & ALLOC_CPUSET) &&
4558 !__cpuset_zone_allowed(z: zone, gfp_mask))
4559 continue;
4560
4561 available = reclaimable = zone_reclaimable_pages(zone);
4562 available += zone_page_state_snapshot(zone, item: NR_FREE_PAGES);
4563
4564 /*
4565 * Would the allocation succeed if we reclaimed all
4566 * reclaimable pages?
4567 */
4568 wmark = __zone_watermark_ok(z: zone, order, mark: min_wmark,
4569 highest_zoneidx: ac->highest_zoneidx, alloc_flags, free_pages: available);
4570 trace_reclaim_retry_zone(zoneref: z, order, reclaimable,
4571 available, min_wmark, no_progress_loops: *no_progress_loops, wmark_check: wmark);
4572 if (wmark) {
4573 ret = true;
4574 break;
4575 }
4576 }
4577
4578 /*
4579 * Memory allocation/reclaim might be called from a WQ context and the
4580 * current implementation of the WQ concurrency control doesn't
4581 * recognize that a particular WQ is congested if the worker thread is
4582 * looping without ever sleeping. Therefore we have to do a short sleep
4583 * here rather than calling cond_resched().
4584 */
4585 if (current->flags & PF_WQ_WORKER)
4586 schedule_timeout_uninterruptible(timeout: 1);
4587 else
4588 cond_resched();
4589out:
4590 /* Before OOM, exhaust highatomic_reserve */
4591 if (!ret)
4592 return unreserve_highatomic_pageblock(ac, force: true);
4593
4594 return ret;
4595}
4596
4597static inline bool
4598check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4599{
4600 /*
4601 * It's possible that cpuset's mems_allowed and the nodemask from
4602 * mempolicy don't intersect. This should be normally dealt with by
4603 * policy_nodemask(), but it's possible to race with cpuset update in
4604 * such a way the check therein was true, and then it became false
4605 * before we got our cpuset_mems_cookie here.
4606 * This assumes that for all allocations, ac->nodemask can come only
4607 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4608 * when it does not intersect with the cpuset restrictions) or the
4609 * caller can deal with a violated nodemask.
4610 */
4611 if (cpusets_enabled() && ac->nodemask &&
4612 !cpuset_nodemask_valid_mems_allowed(nodemask: ac->nodemask)) {
4613 ac->nodemask = NULL;
4614 return true;
4615 }
4616
4617 /*
4618 * When updating a task's mems_allowed or mempolicy nodemask, it is
4619 * possible to race with parallel threads in such a way that our
4620 * allocation can fail while the mask is being updated. If we are about
4621 * to fail, check if the cpuset changed during allocation and if so,
4622 * retry.
4623 */
4624 if (read_mems_allowed_retry(seq: cpuset_mems_cookie))
4625 return true;
4626
4627 return false;
4628}
4629
4630static inline struct page *
4631__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4632 struct alloc_context *ac)
4633{
4634 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4635 bool can_compact = gfp_compaction_allowed(gfp_mask);
4636 bool nofail = gfp_mask & __GFP_NOFAIL;
4637 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4638 struct page *page = NULL;
4639 unsigned int alloc_flags;
4640 unsigned long did_some_progress;
4641 enum compact_priority compact_priority;
4642 enum compact_result compact_result;
4643 int compaction_retries;
4644 int no_progress_loops;
4645 unsigned int cpuset_mems_cookie;
4646 unsigned int zonelist_iter_cookie;
4647 int reserve_flags;
4648
4649 if (unlikely(nofail)) {
4650 /*
4651 * We most definitely don't want callers attempting to
4652 * allocate greater than order-1 page units with __GFP_NOFAIL.
4653 */
4654 WARN_ON_ONCE(order > 1);
4655 /*
4656 * Also we don't support __GFP_NOFAIL without __GFP_DIRECT_RECLAIM,
4657 * otherwise, we may result in lockup.
4658 */
4659 WARN_ON_ONCE(!can_direct_reclaim);
4660 /*
4661 * PF_MEMALLOC request from this context is rather bizarre
4662 * because we cannot reclaim anything and only can loop waiting
4663 * for somebody to do a work for us.
4664 */
4665 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4666 }
4667
4668restart:
4669 compaction_retries = 0;
4670 no_progress_loops = 0;
4671 compact_result = COMPACT_SKIPPED;
4672 compact_priority = DEF_COMPACT_PRIORITY;
4673 cpuset_mems_cookie = read_mems_allowed_begin();
4674 zonelist_iter_cookie = zonelist_iter_begin();
4675
4676 /*
4677 * The fast path uses conservative alloc_flags to succeed only until
4678 * kswapd needs to be woken up, and to avoid the cost of setting up
4679 * alloc_flags precisely. So we do that now.
4680 */
4681 alloc_flags = gfp_to_alloc_flags(gfp_mask, order);
4682
4683 /*
4684 * We need to recalculate the starting point for the zonelist iterator
4685 * because we might have used different nodemask in the fast path, or
4686 * there was a cpuset modification and we are retrying - otherwise we
4687 * could end up iterating over non-eligible zones endlessly.
4688 */
4689 ac->preferred_zoneref = first_zones_zonelist(zonelist: ac->zonelist,
4690 highest_zoneidx: ac->highest_zoneidx, nodes: ac->nodemask);
4691 if (!zonelist_zone(zoneref: ac->preferred_zoneref))
4692 goto nopage;
4693
4694 /*
4695 * Check for insane configurations where the cpuset doesn't contain
4696 * any suitable zone to satisfy the request - e.g. non-movable
4697 * GFP_HIGHUSER allocations from MOVABLE nodes only.
4698 */
4699 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
4700 struct zoneref *z = first_zones_zonelist(zonelist: ac->zonelist,
4701 highest_zoneidx: ac->highest_zoneidx,
4702 nodes: &cpuset_current_mems_allowed);
4703 if (!zonelist_zone(zoneref: z))
4704 goto nopage;
4705 }
4706
4707 if (alloc_flags & ALLOC_KSWAPD)
4708 wake_all_kswapds(order, gfp_mask, ac);
4709
4710 /*
4711 * The adjusted alloc_flags might result in immediate success, so try
4712 * that first
4713 */
4714 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4715 if (page)
4716 goto got_pg;
4717
4718 /*
4719 * For costly allocations, try direct compaction first, as it's likely
4720 * that we have enough base pages and don't need to reclaim. For non-
4721 * movable high-order allocations, do that as well, as compaction will
4722 * try prevent permanent fragmentation by migrating from blocks of the
4723 * same migratetype.
4724 * Don't try this for allocations that are allowed to ignore
4725 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4726 */
4727 if (can_direct_reclaim && can_compact &&
4728 (costly_order ||
4729 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4730 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4731 page = __alloc_pages_direct_compact(gfp_mask, order,
4732 alloc_flags, ac,
4733 prio: INIT_COMPACT_PRIORITY,
4734 compact_result: &compact_result);
4735 if (page)
4736 goto got_pg;
4737
4738 /*
4739 * Checks for costly allocations with __GFP_NORETRY, which
4740 * includes some THP page fault allocations
4741 */
4742 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4743 /*
4744 * If allocating entire pageblock(s) and compaction
4745 * failed because all zones are below low watermarks
4746 * or is prohibited because it recently failed at this
4747 * order, fail immediately unless the allocator has
4748 * requested compaction and reclaim retry.
4749 *
4750 * Reclaim is
4751 * - potentially very expensive because zones are far
4752 * below their low watermarks or this is part of very
4753 * bursty high order allocations,
4754 * - not guaranteed to help because isolate_freepages()
4755 * may not iterate over freed pages as part of its
4756 * linear scan, and
4757 * - unlikely to make entire pageblocks free on its
4758 * own.
4759 */
4760 if (compact_result == COMPACT_SKIPPED ||
4761 compact_result == COMPACT_DEFERRED)
4762 goto nopage;
4763
4764 /*
4765 * Looks like reclaim/compaction is worth trying, but
4766 * sync compaction could be very expensive, so keep
4767 * using async compaction.
4768 */
4769 compact_priority = INIT_COMPACT_PRIORITY;
4770 }
4771 }
4772
4773retry:
4774 /*
4775 * Deal with possible cpuset update races or zonelist updates to avoid
4776 * infinite retries.
4777 */
4778 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4779 check_retry_zonelist(seq: zonelist_iter_cookie))
4780 goto restart;
4781
4782 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4783 if (alloc_flags & ALLOC_KSWAPD)
4784 wake_all_kswapds(order, gfp_mask, ac);
4785
4786 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4787 if (reserve_flags)
4788 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags: reserve_flags) |
4789 (alloc_flags & ALLOC_KSWAPD);
4790
4791 /*
4792 * Reset the nodemask and zonelist iterators if memory policies can be
4793 * ignored. These allocations are high priority and system rather than
4794 * user oriented.
4795 */
4796 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4797 ac->nodemask = NULL;
4798 ac->preferred_zoneref = first_zones_zonelist(zonelist: ac->zonelist,
4799 highest_zoneidx: ac->highest_zoneidx, nodes: ac->nodemask);
4800 }
4801
4802 /* Attempt with potentially adjusted zonelist and alloc_flags */
4803 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4804 if (page)
4805 goto got_pg;
4806
4807 /* Caller is not willing to reclaim, we can't balance anything */
4808 if (!can_direct_reclaim)
4809 goto nopage;
4810
4811 /* Avoid recursion of direct reclaim */
4812 if (current->flags & PF_MEMALLOC)
4813 goto nopage;
4814
4815 /* Try direct reclaim and then allocating */
4816 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4817 did_some_progress: &did_some_progress);
4818 if (page)
4819 goto got_pg;
4820
4821 /* Try direct compaction and then allocating */
4822 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4823 prio: compact_priority, compact_result: &compact_result);
4824 if (page)
4825 goto got_pg;
4826
4827 /* Do not loop if specifically requested */
4828 if (gfp_mask & __GFP_NORETRY)
4829 goto nopage;
4830
4831 /*
4832 * Do not retry costly high order allocations unless they are
4833 * __GFP_RETRY_MAYFAIL and we can compact
4834 */
4835 if (costly_order && (!can_compact ||
4836 !(gfp_mask & __GFP_RETRY_MAYFAIL)))
4837 goto nopage;
4838
4839 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4840 did_some_progress: did_some_progress > 0, no_progress_loops: &no_progress_loops))
4841 goto retry;
4842
4843 /*
4844 * It doesn't make any sense to retry for the compaction if the order-0
4845 * reclaim is not able to make any progress because the current
4846 * implementation of the compaction depends on the sufficient amount
4847 * of free memory (see __compaction_suitable)
4848 */
4849 if (did_some_progress > 0 && can_compact &&
4850 should_compact_retry(ac, order, alloc_flags,
4851 compact_result, compact_priority: &compact_priority,
4852 compaction_retries: &compaction_retries))
4853 goto retry;
4854
4855 /* Reclaim/compaction failed to prevent the fallback */
4856 if (defrag_mode && (alloc_flags & ALLOC_NOFRAGMENT)) {
4857 alloc_flags &= ~ALLOC_NOFRAGMENT;
4858 goto retry;
4859 }
4860
4861 /*
4862 * Deal with possible cpuset update races or zonelist updates to avoid
4863 * a unnecessary OOM kill.
4864 */
4865 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4866 check_retry_zonelist(seq: zonelist_iter_cookie))
4867 goto restart;
4868
4869 /* Reclaim has failed us, start killing things */
4870 page = __alloc_pages_may_oom(gfp_mask, order, ac, did_some_progress: &did_some_progress);
4871 if (page)
4872 goto got_pg;
4873
4874 /* Avoid allocations with no watermarks from looping endlessly */
4875 if (tsk_is_oom_victim(current) &&
4876 (alloc_flags & ALLOC_OOM ||
4877 (gfp_mask & __GFP_NOMEMALLOC)))
4878 goto nopage;
4879
4880 /* Retry as long as the OOM killer is making progress */
4881 if (did_some_progress) {
4882 no_progress_loops = 0;
4883 goto retry;
4884 }
4885
4886nopage:
4887 /*
4888 * Deal with possible cpuset update races or zonelist updates to avoid
4889 * a unnecessary OOM kill.
4890 */
4891 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4892 check_retry_zonelist(seq: zonelist_iter_cookie))
4893 goto restart;
4894
4895 /*
4896 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4897 * we always retry
4898 */
4899 if (unlikely(nofail)) {
4900 /*
4901 * Lacking direct_reclaim we can't do anything to reclaim memory,
4902 * we disregard these unreasonable nofail requests and still
4903 * return NULL
4904 */
4905 if (!can_direct_reclaim)
4906 goto fail;
4907
4908 /*
4909 * Help non-failing allocations by giving some access to memory
4910 * reserves normally used for high priority non-blocking
4911 * allocations but do not use ALLOC_NO_WATERMARKS because this
4912 * could deplete whole memory reserves which would just make
4913 * the situation worse.
4914 */
4915 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_MIN_RESERVE, ac);
4916 if (page)
4917 goto got_pg;
4918
4919 cond_resched();
4920 goto retry;
4921 }
4922fail:
4923 warn_alloc(gfp_mask, nodemask: ac->nodemask,
4924 fmt: "page allocation failure: order:%u", order);
4925got_pg:
4926 return page;
4927}
4928
4929static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4930 int preferred_nid, nodemask_t *nodemask,
4931 struct alloc_context *ac, gfp_t *alloc_gfp,
4932 unsigned int *alloc_flags)
4933{
4934 ac->highest_zoneidx = gfp_zone(flags: gfp_mask);
4935 ac->zonelist = node_zonelist(nid: preferred_nid, flags: gfp_mask);
4936 ac->nodemask = nodemask;
4937 ac->migratetype = gfp_migratetype(gfp_flags: gfp_mask);
4938
4939 if (cpusets_enabled()) {
4940 *alloc_gfp |= __GFP_HARDWALL;
4941 /*
4942 * When we are in the interrupt context, it is irrelevant
4943 * to the current task context. It means that any node ok.
4944 */
4945 if (in_task() && !ac->nodemask)
4946 ac->nodemask = &cpuset_current_mems_allowed;
4947 else
4948 *alloc_flags |= ALLOC_CPUSET;
4949 }
4950
4951 might_alloc(gfp_mask);
4952
4953 /*
4954 * Don't invoke should_fail logic, since it may call
4955 * get_random_u32() and printk() which need to spin_lock.
4956 */
4957 if (!(*alloc_flags & ALLOC_TRYLOCK) &&
4958 should_fail_alloc_page(gfp_mask, order))
4959 return false;
4960
4961 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags: *alloc_flags);
4962
4963 /* Dirty zone balancing only done in the fast path */
4964 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4965
4966 /*
4967 * The preferred zone is used for statistics but crucially it is
4968 * also used as the starting point for the zonelist iterator. It
4969 * may get reset for allocations that ignore memory policies.
4970 */
4971 ac->preferred_zoneref = first_zones_zonelist(zonelist: ac->zonelist,
4972 highest_zoneidx: ac->highest_zoneidx, nodes: ac->nodemask);
4973
4974 return true;
4975}
4976
4977/*
4978 * __alloc_pages_bulk - Allocate a number of order-0 pages to an array
4979 * @gfp: GFP flags for the allocation
4980 * @preferred_nid: The preferred NUMA node ID to allocate from
4981 * @nodemask: Set of nodes to allocate from, may be NULL
4982 * @nr_pages: The number of pages desired in the array
4983 * @page_array: Array to store the pages
4984 *
4985 * This is a batched version of the page allocator that attempts to
4986 * allocate nr_pages quickly. Pages are added to the page_array.
4987 *
4988 * Note that only NULL elements are populated with pages and nr_pages
4989 * is the maximum number of pages that will be stored in the array.
4990 *
4991 * Returns the number of pages in the array.
4992 */
4993unsigned long alloc_pages_bulk_noprof(gfp_t gfp, int preferred_nid,
4994 nodemask_t *nodemask, int nr_pages,
4995 struct page **page_array)
4996{
4997 struct page *page;
4998 unsigned long __maybe_unused UP_flags;
4999 struct zone *zone;
5000 struct zoneref *z;
5001 struct per_cpu_pages *pcp;
5002 struct list_head *pcp_list;
5003 struct alloc_context ac;
5004 gfp_t alloc_gfp;
5005 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5006 int nr_populated = 0, nr_account = 0;
5007
5008 /*
5009 * Skip populated array elements to determine if any pages need
5010 * to be allocated before disabling IRQs.
5011 */
5012 while (nr_populated < nr_pages && page_array[nr_populated])
5013 nr_populated++;
5014
5015 /* No pages requested? */
5016 if (unlikely(nr_pages <= 0))
5017 goto out;
5018
5019 /* Already populated array? */
5020 if (unlikely(nr_pages - nr_populated == 0))
5021 goto out;
5022
5023 /* Bulk allocator does not support memcg accounting. */
5024 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT))
5025 goto failed;
5026
5027 /* Use the single page allocator for one page. */
5028 if (nr_pages - nr_populated == 1)
5029 goto failed;
5030
5031#ifdef CONFIG_PAGE_OWNER
5032 /*
5033 * PAGE_OWNER may recurse into the allocator to allocate space to
5034 * save the stack with pagesets.lock held. Releasing/reacquiring
5035 * removes much of the performance benefit of bulk allocation so
5036 * force the caller to allocate one page at a time as it'll have
5037 * similar performance to added complexity to the bulk allocator.
5038 */
5039 if (static_branch_unlikely(&page_owner_inited))
5040 goto failed;
5041#endif
5042
5043 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5044 gfp &= gfp_allowed_mask;
5045 alloc_gfp = gfp;
5046 if (!prepare_alloc_pages(gfp_mask: gfp, order: 0, preferred_nid, nodemask, ac: &ac, alloc_gfp: &alloc_gfp, alloc_flags: &alloc_flags))
5047 goto out;
5048 gfp = alloc_gfp;
5049
5050 /* Find an allowed local zone that meets the low watermark. */
5051 z = ac.preferred_zoneref;
5052 for_next_zone_zonelist_nodemask(zone, z, ac.highest_zoneidx, ac.nodemask) {
5053 unsigned long mark;
5054
5055 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5056 !__cpuset_zone_allowed(z: zone, gfp_mask: gfp)) {
5057 continue;
5058 }
5059
5060 if (nr_online_nodes > 1 && zone != zonelist_zone(zoneref: ac.preferred_zoneref) &&
5061 zone_to_nid(zone) != zonelist_node_idx(zoneref: ac.preferred_zoneref)) {
5062 goto failed;
5063 }
5064
5065 cond_accept_memory(zone, order: 0, alloc_flags);
5066retry_this_zone:
5067 mark = wmark_pages(z: zone, w: alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
5068 if (zone_watermark_fast(z: zone, order: 0, mark,
5069 highest_zoneidx: zonelist_zone_idx(zoneref: ac.preferred_zoneref),
5070 alloc_flags, gfp_mask: gfp)) {
5071 break;
5072 }
5073
5074 if (cond_accept_memory(zone, order: 0, alloc_flags))
5075 goto retry_this_zone;
5076
5077 /* Try again if zone has deferred pages */
5078 if (deferred_pages_enabled()) {
5079 if (_deferred_grow_zone(zone, order: 0))
5080 goto retry_this_zone;
5081 }
5082 }
5083
5084 /*
5085 * If there are no allowed local zones that meets the watermarks then
5086 * try to allocate a single page and reclaim if necessary.
5087 */
5088 if (unlikely(!zone))
5089 goto failed;
5090
5091 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
5092 pcp_trylock_prepare(UP_flags);
5093 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
5094 if (!pcp)
5095 goto failed_irq;
5096
5097 /* Attempt the batch allocation */
5098 pcp_list = &pcp->lists[order_to_pindex(migratetype: ac.migratetype, order: 0)];
5099 while (nr_populated < nr_pages) {
5100
5101 /* Skip existing pages */
5102 if (page_array[nr_populated]) {
5103 nr_populated++;
5104 continue;
5105 }
5106
5107 page = __rmqueue_pcplist(zone, order: 0, migratetype: ac.migratetype, alloc_flags,
5108 pcp, list: pcp_list);
5109 if (unlikely(!page)) {
5110 /* Try and allocate at least one page */
5111 if (!nr_account) {
5112 pcp_spin_unlock(pcp);
5113 goto failed_irq;
5114 }
5115 break;
5116 }
5117 nr_account++;
5118
5119 prep_new_page(page, order: 0, gfp_flags: gfp, alloc_flags: 0);
5120 set_page_refcounted(page);
5121 page_array[nr_populated++] = page;
5122 }
5123
5124 pcp_spin_unlock(pcp);
5125 pcp_trylock_finish(UP_flags);
5126
5127 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
5128 zone_statistics(preferred_zone: zonelist_zone(zoneref: ac.preferred_zoneref), z: zone, nr_account);
5129
5130out:
5131 return nr_populated;
5132
5133failed_irq:
5134 pcp_trylock_finish(UP_flags);
5135
5136failed:
5137 page = __alloc_pages_noprof(gfp, order: 0, preferred_nid, nodemask);
5138 if (page)
5139 page_array[nr_populated++] = page;
5140 goto out;
5141}
5142EXPORT_SYMBOL_GPL(alloc_pages_bulk_noprof);
5143
5144/*
5145 * This is the 'heart' of the zoned buddy allocator.
5146 */
5147struct page *__alloc_frozen_pages_noprof(gfp_t gfp, unsigned int order,
5148 int preferred_nid, nodemask_t *nodemask)
5149{
5150 struct page *page;
5151 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5152 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5153 struct alloc_context ac = { };
5154
5155 /*
5156 * There are several places where we assume that the order value is sane
5157 * so bail out early if the request is out of bound.
5158 */
5159 if (WARN_ON_ONCE_GFP(order > MAX_PAGE_ORDER, gfp))
5160 return NULL;
5161
5162 gfp &= gfp_allowed_mask;
5163 /*
5164 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5165 * resp. GFP_NOIO which has to be inherited for all allocation requests
5166 * from a particular context which has been marked by
5167 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5168 * movable zones are not used during allocation.
5169 */
5170 gfp = current_gfp_context(flags: gfp);
5171 alloc_gfp = gfp;
5172 if (!prepare_alloc_pages(gfp_mask: gfp, order, preferred_nid, nodemask, ac: &ac,
5173 alloc_gfp: &alloc_gfp, alloc_flags: &alloc_flags))
5174 return NULL;
5175
5176 /*
5177 * Forbid the first pass from falling back to types that fragment
5178 * memory until all local zones are considered.
5179 */
5180 alloc_flags |= alloc_flags_nofragment(zone: zonelist_zone(zoneref: ac.preferred_zoneref), gfp_mask: gfp);
5181
5182 /* First allocation attempt */
5183 page = get_page_from_freelist(gfp_mask: alloc_gfp, order, alloc_flags, ac: &ac);
5184 if (likely(page))
5185 goto out;
5186
5187 alloc_gfp = gfp;
5188 ac.spread_dirty_pages = false;
5189
5190 /*
5191 * Restore the original nodemask if it was potentially replaced with
5192 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5193 */
5194 ac.nodemask = nodemask;
5195
5196 page = __alloc_pages_slowpath(gfp_mask: alloc_gfp, order, ac: &ac);
5197
5198out:
5199 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT) && page &&
5200 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5201 free_frozen_pages(page, order);
5202 page = NULL;
5203 }
5204
5205 trace_mm_page_alloc(page, order, gfp_flags: alloc_gfp, migratetype: ac.migratetype);
5206 kmsan_alloc_page(page, order, flags: alloc_gfp);
5207
5208 return page;
5209}
5210EXPORT_SYMBOL(__alloc_frozen_pages_noprof);
5211
5212struct page *__alloc_pages_noprof(gfp_t gfp, unsigned int order,
5213 int preferred_nid, nodemask_t *nodemask)
5214{
5215 struct page *page;
5216
5217 page = __alloc_frozen_pages_noprof(gfp, order, preferred_nid, nodemask);
5218 if (page)
5219 set_page_refcounted(page);
5220 return page;
5221}
5222EXPORT_SYMBOL(__alloc_pages_noprof);
5223
5224struct folio *__folio_alloc_noprof(gfp_t gfp, unsigned int order, int preferred_nid,
5225 nodemask_t *nodemask)
5226{
5227 struct page *page = __alloc_pages_noprof(gfp | __GFP_COMP, order,
5228 preferred_nid, nodemask);
5229 return page_rmappable_folio(page);
5230}
5231EXPORT_SYMBOL(__folio_alloc_noprof);
5232
5233/*
5234 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5235 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5236 * you need to access high mem.
5237 */
5238unsigned long get_free_pages_noprof(gfp_t gfp_mask, unsigned int order)
5239{
5240 struct page *page;
5241
5242 page = alloc_pages_noprof(gfp: gfp_mask & ~__GFP_HIGHMEM, order);
5243 if (!page)
5244 return 0;
5245 return (unsigned long) page_address(page);
5246}
5247EXPORT_SYMBOL(get_free_pages_noprof);
5248
5249unsigned long get_zeroed_page_noprof(gfp_t gfp_mask)
5250{
5251 return get_free_pages_noprof(gfp_mask | __GFP_ZERO, 0);
5252}
5253EXPORT_SYMBOL(get_zeroed_page_noprof);
5254
5255static void ___free_pages(struct page *page, unsigned int order,
5256 fpi_t fpi_flags)
5257{
5258 /* get PageHead before we drop reference */
5259 int head = PageHead(page);
5260 /* get alloc tag in case the page is released by others */
5261 struct alloc_tag *tag = pgalloc_tag_get(page);
5262
5263 if (put_page_testzero(page))
5264 __free_frozen_pages(page, order, fpi_flags);
5265 else if (!head) {
5266 pgalloc_tag_sub_pages(tag, nr: (1 << order) - 1);
5267 while (order-- > 0) {
5268 /*
5269 * The "tail" pages of this non-compound high-order
5270 * page will have no code tags, so to avoid warnings
5271 * mark them as empty.
5272 */
5273 clear_page_tag_ref(page: page + (1 << order));
5274 __free_frozen_pages(page: page + (1 << order), order,
5275 fpi_flags);
5276 }
5277 }
5278}
5279
5280/**
5281 * __free_pages - Free pages allocated with alloc_pages().
5282 * @page: The page pointer returned from alloc_pages().
5283 * @order: The order of the allocation.
5284 *
5285 * This function can free multi-page allocations that are not compound
5286 * pages. It does not check that the @order passed in matches that of
5287 * the allocation, so it is easy to leak memory. Freeing more memory
5288 * than was allocated will probably emit a warning.
5289 *
5290 * If the last reference to this page is speculative, it will be released
5291 * by put_page() which only frees the first page of a non-compound
5292 * allocation. To prevent the remaining pages from being leaked, we free
5293 * the subsequent pages here. If you want to use the page's reference
5294 * count to decide when to free the allocation, you should allocate a
5295 * compound page, and use put_page() instead of __free_pages().
5296 *
5297 * Context: May be called in interrupt context or while holding a normal
5298 * spinlock, but not in NMI context or while holding a raw spinlock.
5299 */
5300void __free_pages(struct page *page, unsigned int order)
5301{
5302 ___free_pages(page, order, FPI_NONE);
5303}
5304EXPORT_SYMBOL(__free_pages);
5305
5306/*
5307 * Can be called while holding raw_spin_lock or from IRQ and NMI for any
5308 * page type (not only those that came from alloc_pages_nolock)
5309 */
5310void free_pages_nolock(struct page *page, unsigned int order)
5311{
5312 ___free_pages(page, order, FPI_TRYLOCK);
5313}
5314
5315/**
5316 * free_pages - Free pages allocated with __get_free_pages().
5317 * @addr: The virtual address tied to a page returned from __get_free_pages().
5318 * @order: The order of the allocation.
5319 *
5320 * This function behaves the same as __free_pages(). Use this function
5321 * to free pages when you only have a valid virtual address. If you have
5322 * the page, call __free_pages() instead.
5323 */
5324void free_pages(unsigned long addr, unsigned int order)
5325{
5326 if (addr != 0) {
5327 VM_BUG_ON(!virt_addr_valid((void *)addr));
5328 __free_pages(virt_to_page((void *)addr), order);
5329 }
5330}
5331
5332EXPORT_SYMBOL(free_pages);
5333
5334static void *make_alloc_exact(unsigned long addr, unsigned int order,
5335 size_t size)
5336{
5337 if (addr) {
5338 unsigned long nr = DIV_ROUND_UP(size, PAGE_SIZE);
5339 struct page *page = virt_to_page((void *)addr);
5340 struct page *last = page + nr;
5341
5342 split_page_owner(page, old_order: order, new_order: 0);
5343 pgalloc_tag_split(page_folio(page), old_order: order, new_order: 0);
5344 split_page_memcg(first: page, order);
5345 while (page < --last)
5346 set_page_refcounted(last);
5347
5348 last = page + (1UL << order);
5349 for (page += nr; page < last; page++)
5350 __free_pages_ok(page, order: 0, FPI_TO_TAIL);
5351 }
5352 return (void *)addr;
5353}
5354
5355/**
5356 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5357 * @size: the number of bytes to allocate
5358 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5359 *
5360 * This function is similar to alloc_pages(), except that it allocates the
5361 * minimum number of pages to satisfy the request. alloc_pages() can only
5362 * allocate memory in power-of-two pages.
5363 *
5364 * This function is also limited by MAX_PAGE_ORDER.
5365 *
5366 * Memory allocated by this function must be released by free_pages_exact().
5367 *
5368 * Return: pointer to the allocated area or %NULL in case of error.
5369 */
5370void *alloc_pages_exact_noprof(size_t size, gfp_t gfp_mask)
5371{
5372 unsigned int order = get_order(size);
5373 unsigned long addr;
5374
5375 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5376 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5377
5378 addr = get_free_pages_noprof(gfp_mask, order);
5379 return make_alloc_exact(addr, order, size);
5380}
5381EXPORT_SYMBOL(alloc_pages_exact_noprof);
5382
5383/**
5384 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5385 * pages on a node.
5386 * @nid: the preferred node ID where memory should be allocated
5387 * @size: the number of bytes to allocate
5388 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5389 *
5390 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5391 * back.
5392 *
5393 * Return: pointer to the allocated area or %NULL in case of error.
5394 */
5395void * __meminit alloc_pages_exact_nid_noprof(int nid, size_t size, gfp_t gfp_mask)
5396{
5397 unsigned int order = get_order(size);
5398 struct page *p;
5399
5400 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5401 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5402
5403 p = alloc_pages_node_noprof(nid, gfp_mask, order);
5404 if (!p)
5405 return NULL;
5406 return make_alloc_exact(addr: (unsigned long)page_address(p), order, size);
5407}
5408
5409/**
5410 * free_pages_exact - release memory allocated via alloc_pages_exact()
5411 * @virt: the value returned by alloc_pages_exact.
5412 * @size: size of allocation, same value as passed to alloc_pages_exact().
5413 *
5414 * Release the memory allocated by a previous call to alloc_pages_exact.
5415 */
5416void free_pages_exact(void *virt, size_t size)
5417{
5418 unsigned long addr = (unsigned long)virt;
5419 unsigned long end = addr + PAGE_ALIGN(size);
5420
5421 while (addr < end) {
5422 free_page(addr);
5423 addr += PAGE_SIZE;
5424 }
5425}
5426EXPORT_SYMBOL(free_pages_exact);
5427
5428/**
5429 * nr_free_zone_pages - count number of pages beyond high watermark
5430 * @offset: The zone index of the highest zone
5431 *
5432 * nr_free_zone_pages() counts the number of pages which are beyond the
5433 * high watermark within all zones at or below a given zone index. For each
5434 * zone, the number of pages is calculated as:
5435 *
5436 * nr_free_zone_pages = managed_pages - high_pages
5437 *
5438 * Return: number of pages beyond high watermark.
5439 */
5440static unsigned long nr_free_zone_pages(int offset)
5441{
5442 struct zoneref *z;
5443 struct zone *zone;
5444
5445 /* Just pick one node, since fallback list is circular */
5446 unsigned long sum = 0;
5447
5448 struct zonelist *zonelist = node_zonelist(nid: numa_node_id(), GFP_KERNEL);
5449
5450 for_each_zone_zonelist(zone, z, zonelist, offset) {
5451 unsigned long size = zone_managed_pages(zone);
5452 unsigned long high = high_wmark_pages(z: zone);
5453 if (size > high)
5454 sum += size - high;
5455 }
5456
5457 return sum;
5458}
5459
5460/**
5461 * nr_free_buffer_pages - count number of pages beyond high watermark
5462 *
5463 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5464 * watermark within ZONE_DMA and ZONE_NORMAL.
5465 *
5466 * Return: number of pages beyond high watermark within ZONE_DMA and
5467 * ZONE_NORMAL.
5468 */
5469unsigned long nr_free_buffer_pages(void)
5470{
5471 return nr_free_zone_pages(offset: gfp_zone(GFP_USER));
5472}
5473EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5474
5475static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5476{
5477 zoneref->zone = zone;
5478 zoneref->zone_idx = zone_idx(zone);
5479}
5480
5481/*
5482 * Builds allocation fallback zone lists.
5483 *
5484 * Add all populated zones of a node to the zonelist.
5485 */
5486static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5487{
5488 struct zone *zone;
5489 enum zone_type zone_type = MAX_NR_ZONES;
5490 int nr_zones = 0;
5491
5492 do {
5493 zone_type--;
5494 zone = pgdat->node_zones + zone_type;
5495 if (populated_zone(zone)) {
5496 zoneref_set_zone(zone, zoneref: &zonerefs[nr_zones++]);
5497 check_highest_zone(k: zone_type);
5498 }
5499 } while (zone_type);
5500
5501 return nr_zones;
5502}
5503
5504#ifdef CONFIG_NUMA
5505
5506static int __parse_numa_zonelist_order(char *s)
5507{
5508 /*
5509 * We used to support different zonelists modes but they turned
5510 * out to be just not useful. Let's keep the warning in place
5511 * if somebody still use the cmd line parameter so that we do
5512 * not fail it silently
5513 */
5514 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5515 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5516 return -EINVAL;
5517 }
5518 return 0;
5519}
5520
5521static char numa_zonelist_order[] = "Node";
5522#define NUMA_ZONELIST_ORDER_LEN 16
5523/*
5524 * sysctl handler for numa_zonelist_order
5525 */
5526static int numa_zonelist_order_handler(const struct ctl_table *table, int write,
5527 void *buffer, size_t *length, loff_t *ppos)
5528{
5529 if (write)
5530 return __parse_numa_zonelist_order(s: buffer);
5531 return proc_dostring(table, write, buffer, length, ppos);
5532}
5533
5534static int node_load[MAX_NUMNODES];
5535
5536/**
5537 * find_next_best_node - find the next node that should appear in a given node's fallback list
5538 * @node: node whose fallback list we're appending
5539 * @used_node_mask: nodemask_t of already used nodes
5540 *
5541 * We use a number of factors to determine which is the next node that should
5542 * appear on a given node's fallback list. The node should not have appeared
5543 * already in @node's fallback list, and it should be the next closest node
5544 * according to the distance array (which contains arbitrary distance values
5545 * from each node to each node in the system), and should also prefer nodes
5546 * with no CPUs, since presumably they'll have very little allocation pressure
5547 * on them otherwise.
5548 *
5549 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5550 */
5551int find_next_best_node(int node, nodemask_t *used_node_mask)
5552{
5553 int n, val;
5554 int min_val = INT_MAX;
5555 int best_node = NUMA_NO_NODE;
5556
5557 /*
5558 * Use the local node if we haven't already, but for memoryless local
5559 * node, we should skip it and fall back to other nodes.
5560 */
5561 if (!node_isset(node, *used_node_mask) && node_state(node, state: N_MEMORY)) {
5562 node_set(node, *used_node_mask);
5563 return node;
5564 }
5565
5566 for_each_node_state(n, N_MEMORY) {
5567
5568 /* Don't want a node to appear more than once */
5569 if (node_isset(n, *used_node_mask))
5570 continue;
5571
5572 /* Use the distance array to find the distance */
5573 val = node_distance(node, n);
5574
5575 /* Penalize nodes under us ("prefer the next node") */
5576 val += (n < node);
5577
5578 /* Give preference to headless and unused nodes */
5579 if (!cpumask_empty(srcp: cpumask_of_node(node: n)))
5580 val += PENALTY_FOR_NODE_WITH_CPUS;
5581
5582 /* Slight preference for less loaded node */
5583 val *= MAX_NUMNODES;
5584 val += node_load[n];
5585
5586 if (val < min_val) {
5587 min_val = val;
5588 best_node = n;
5589 }
5590 }
5591
5592 if (best_node >= 0)
5593 node_set(best_node, *used_node_mask);
5594
5595 return best_node;
5596}
5597
5598
5599/*
5600 * Build zonelists ordered by node and zones within node.
5601 * This results in maximum locality--normal zone overflows into local
5602 * DMA zone, if any--but risks exhausting DMA zone.
5603 */
5604static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5605 unsigned nr_nodes)
5606{
5607 struct zoneref *zonerefs;
5608 int i;
5609
5610 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5611
5612 for (i = 0; i < nr_nodes; i++) {
5613 int nr_zones;
5614
5615 pg_data_t *node = NODE_DATA(node_order[i]);
5616
5617 nr_zones = build_zonerefs_node(pgdat: node, zonerefs);
5618 zonerefs += nr_zones;
5619 }
5620 zonerefs->zone = NULL;
5621 zonerefs->zone_idx = 0;
5622}
5623
5624/*
5625 * Build __GFP_THISNODE zonelists
5626 */
5627static void build_thisnode_zonelists(pg_data_t *pgdat)
5628{
5629 struct zoneref *zonerefs;
5630 int nr_zones;
5631
5632 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5633 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5634 zonerefs += nr_zones;
5635 zonerefs->zone = NULL;
5636 zonerefs->zone_idx = 0;
5637}
5638
5639static void build_zonelists(pg_data_t *pgdat)
5640{
5641 static int node_order[MAX_NUMNODES];
5642 int node, nr_nodes = 0;
5643 nodemask_t used_mask = NODE_MASK_NONE;
5644 int local_node, prev_node;
5645
5646 /* NUMA-aware ordering of nodes */
5647 local_node = pgdat->node_id;
5648 prev_node = local_node;
5649
5650 memset(s: node_order, c: 0, n: sizeof(node_order));
5651 while ((node = find_next_best_node(node: local_node, used_node_mask: &used_mask)) >= 0) {
5652 /*
5653 * We don't want to pressure a particular node.
5654 * So adding penalty to the first node in same
5655 * distance group to make it round-robin.
5656 */
5657 if (node_distance(local_node, node) !=
5658 node_distance(local_node, prev_node))
5659 node_load[node] += 1;
5660
5661 node_order[nr_nodes++] = node;
5662 prev_node = node;
5663 }
5664
5665 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5666 build_thisnode_zonelists(pgdat);
5667 pr_info("Fallback order for Node %d: ", local_node);
5668 for (node = 0; node < nr_nodes; node++)
5669 pr_cont("%d ", node_order[node]);
5670 pr_cont("\n");
5671}
5672
5673#ifdef CONFIG_HAVE_MEMORYLESS_NODES
5674/*
5675 * Return node id of node used for "local" allocations.
5676 * I.e., first node id of first zone in arg node's generic zonelist.
5677 * Used for initializing percpu 'numa_mem', which is used primarily
5678 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5679 */
5680int local_memory_node(int node)
5681{
5682 struct zoneref *z;
5683
5684 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5685 gfp_zone(GFP_KERNEL),
5686 NULL);
5687 return zonelist_node_idx(z);
5688}
5689#endif
5690
5691static void setup_min_unmapped_ratio(void);
5692static void setup_min_slab_ratio(void);
5693#else /* CONFIG_NUMA */
5694
5695static void build_zonelists(pg_data_t *pgdat)
5696{
5697 struct zoneref *zonerefs;
5698 int nr_zones;
5699
5700 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5701 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5702 zonerefs += nr_zones;
5703
5704 zonerefs->zone = NULL;
5705 zonerefs->zone_idx = 0;
5706}
5707
5708#endif /* CONFIG_NUMA */
5709
5710/*
5711 * Boot pageset table. One per cpu which is going to be used for all
5712 * zones and all nodes. The parameters will be set in such a way
5713 * that an item put on a list will immediately be handed over to
5714 * the buddy list. This is safe since pageset manipulation is done
5715 * with interrupts disabled.
5716 *
5717 * The boot_pagesets must be kept even after bootup is complete for
5718 * unused processors and/or zones. They do play a role for bootstrapping
5719 * hotplugged processors.
5720 *
5721 * zoneinfo_show() and maybe other functions do
5722 * not check if the processor is online before following the pageset pointer.
5723 * Other parts of the kernel may not check if the zone is available.
5724 */
5725static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
5726/* These effectively disable the pcplists in the boot pageset completely */
5727#define BOOT_PAGESET_HIGH 0
5728#define BOOT_PAGESET_BATCH 1
5729static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
5730static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
5731
5732static void __build_all_zonelists(void *data)
5733{
5734 int nid;
5735 int __maybe_unused cpu;
5736 pg_data_t *self = data;
5737 unsigned long flags;
5738
5739 /*
5740 * The zonelist_update_seq must be acquired with irqsave because the
5741 * reader can be invoked from IRQ with GFP_ATOMIC.
5742 */
5743 write_seqlock_irqsave(&zonelist_update_seq, flags);
5744 /*
5745 * Also disable synchronous printk() to prevent any printk() from
5746 * trying to hold port->lock, for
5747 * tty_insert_flip_string_and_push_buffer() on other CPU might be
5748 * calling kmalloc(GFP_ATOMIC | __GFP_NOWARN) with port->lock held.
5749 */
5750 printk_deferred_enter();
5751
5752#ifdef CONFIG_NUMA
5753 memset(s: node_load, c: 0, n: sizeof(node_load));
5754#endif
5755
5756 /*
5757 * This node is hotadded and no memory is yet present. So just
5758 * building zonelists is fine - no need to touch other nodes.
5759 */
5760 if (self && !node_online(self->node_id)) {
5761 build_zonelists(pgdat: self);
5762 } else {
5763 /*
5764 * All possible nodes have pgdat preallocated
5765 * in free_area_init
5766 */
5767 for_each_node(nid) {
5768 pg_data_t *pgdat = NODE_DATA(nid);
5769
5770 build_zonelists(pgdat);
5771 }
5772
5773#ifdef CONFIG_HAVE_MEMORYLESS_NODES
5774 /*
5775 * We now know the "local memory node" for each node--
5776 * i.e., the node of the first zone in the generic zonelist.
5777 * Set up numa_mem percpu variable for on-line cpus. During
5778 * boot, only the boot cpu should be on-line; we'll init the
5779 * secondary cpus' numa_mem as they come on-line. During
5780 * node/memory hotplug, we'll fixup all on-line cpus.
5781 */
5782 for_each_online_cpu(cpu)
5783 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5784#endif
5785 }
5786
5787 printk_deferred_exit();
5788 write_sequnlock_irqrestore(sl: &zonelist_update_seq, flags);
5789}
5790
5791static noinline void __init
5792build_all_zonelists_init(void)
5793{
5794 int cpu;
5795
5796 __build_all_zonelists(NULL);
5797
5798 /*
5799 * Initialize the boot_pagesets that are going to be used
5800 * for bootstrapping processors. The real pagesets for
5801 * each zone will be allocated later when the per cpu
5802 * allocator is available.
5803 *
5804 * boot_pagesets are used also for bootstrapping offline
5805 * cpus if the system is already booted because the pagesets
5806 * are needed to initialize allocators on a specific cpu too.
5807 * F.e. the percpu allocator needs the page allocator which
5808 * needs the percpu allocator in order to allocate its pagesets
5809 * (a chicken-egg dilemma).
5810 */
5811 for_each_possible_cpu(cpu)
5812 per_cpu_pages_init(pcp: &per_cpu(boot_pageset, cpu), pzstats: &per_cpu(boot_zonestats, cpu));
5813
5814 mminit_verify_zonelist();
5815 cpuset_init_current_mems_allowed();
5816}
5817
5818/*
5819 * unless system_state == SYSTEM_BOOTING.
5820 *
5821 * __ref due to call of __init annotated helper build_all_zonelists_init
5822 * [protected by SYSTEM_BOOTING].
5823 */
5824void __ref build_all_zonelists(pg_data_t *pgdat)
5825{
5826 unsigned long vm_total_pages;
5827
5828 if (system_state == SYSTEM_BOOTING) {
5829 build_all_zonelists_init();
5830 } else {
5831 __build_all_zonelists(data: pgdat);
5832 /* cpuset refresh routine should be here */
5833 }
5834 /* Get the number of free pages beyond high watermark in all zones. */
5835 vm_total_pages = nr_free_zone_pages(offset: gfp_zone(GFP_HIGHUSER_MOVABLE));
5836 /*
5837 * Disable grouping by mobility if the number of pages in the
5838 * system is too low to allow the mechanism to work. It would be
5839 * more accurate, but expensive to check per-zone. This check is
5840 * made on memory-hotadd so a system can start with mobility
5841 * disabled and enable it later
5842 */
5843 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5844 page_group_by_mobility_disabled = 1;
5845 else
5846 page_group_by_mobility_disabled = 0;
5847
5848 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5849 nr_online_nodes,
5850 str_off_on(page_group_by_mobility_disabled),
5851 vm_total_pages);
5852#ifdef CONFIG_NUMA
5853 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5854#endif
5855}
5856
5857static int zone_batchsize(struct zone *zone)
5858{
5859#ifdef CONFIG_MMU
5860 int batch;
5861
5862 /*
5863 * The number of pages to batch allocate is either ~0.1%
5864 * of the zone or 1MB, whichever is smaller. The batch
5865 * size is striking a balance between allocation latency
5866 * and zone lock contention.
5867 */
5868 batch = min(zone_managed_pages(zone) >> 10, SZ_1M / PAGE_SIZE);
5869 batch /= 4; /* We effectively *= 4 below */
5870 if (batch < 1)
5871 batch = 1;
5872
5873 /*
5874 * Clamp the batch to a 2^n - 1 value. Having a power
5875 * of 2 value was found to be more likely to have
5876 * suboptimal cache aliasing properties in some cases.
5877 *
5878 * For example if 2 tasks are alternately allocating
5879 * batches of pages, one task can end up with a lot
5880 * of pages of one half of the possible page colors
5881 * and the other with pages of the other colors.
5882 */
5883 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5884
5885 return batch;
5886
5887#else
5888 /* The deferral and batching of frees should be suppressed under NOMMU
5889 * conditions.
5890 *
5891 * The problem is that NOMMU needs to be able to allocate large chunks
5892 * of contiguous memory as there's no hardware page translation to
5893 * assemble apparent contiguous memory from discontiguous pages.
5894 *
5895 * Queueing large contiguous runs of pages for batching, however,
5896 * causes the pages to actually be freed in smaller chunks. As there
5897 * can be a significant delay between the individual batches being
5898 * recycled, this leads to the once large chunks of space being
5899 * fragmented and becoming unavailable for high-order allocations.
5900 */
5901 return 0;
5902#endif
5903}
5904
5905static int percpu_pagelist_high_fraction;
5906static int zone_highsize(struct zone *zone, int batch, int cpu_online,
5907 int high_fraction)
5908{
5909#ifdef CONFIG_MMU
5910 int high;
5911 int nr_split_cpus;
5912 unsigned long total_pages;
5913
5914 if (!high_fraction) {
5915 /*
5916 * By default, the high value of the pcp is based on the zone
5917 * low watermark so that if they are full then background
5918 * reclaim will not be started prematurely.
5919 */
5920 total_pages = low_wmark_pages(z: zone);
5921 } else {
5922 /*
5923 * If percpu_pagelist_high_fraction is configured, the high
5924 * value is based on a fraction of the managed pages in the
5925 * zone.
5926 */
5927 total_pages = zone_managed_pages(zone) / high_fraction;
5928 }
5929
5930 /*
5931 * Split the high value across all online CPUs local to the zone. Note
5932 * that early in boot that CPUs may not be online yet and that during
5933 * CPU hotplug that the cpumask is not yet updated when a CPU is being
5934 * onlined. For memory nodes that have no CPUs, split the high value
5935 * across all online CPUs to mitigate the risk that reclaim is triggered
5936 * prematurely due to pages stored on pcp lists.
5937 */
5938 nr_split_cpus = cpumask_weight(srcp: cpumask_of_node(node: zone_to_nid(zone))) + cpu_online;
5939 if (!nr_split_cpus)
5940 nr_split_cpus = num_online_cpus();
5941 high = total_pages / nr_split_cpus;
5942
5943 /*
5944 * Ensure high is at least batch*4. The multiple is based on the
5945 * historical relationship between high and batch.
5946 */
5947 high = max(high, batch << 2);
5948
5949 return high;
5950#else
5951 return 0;
5952#endif
5953}
5954
5955/*
5956 * pcp->high and pcp->batch values are related and generally batch is lower
5957 * than high. They are also related to pcp->count such that count is lower
5958 * than high, and as soon as it reaches high, the pcplist is flushed.
5959 *
5960 * However, guaranteeing these relations at all times would require e.g. write
5961 * barriers here but also careful usage of read barriers at the read side, and
5962 * thus be prone to error and bad for performance. Thus the update only prevents
5963 * store tearing. Any new users of pcp->batch, pcp->high_min and pcp->high_max
5964 * should ensure they can cope with those fields changing asynchronously, and
5965 * fully trust only the pcp->count field on the local CPU with interrupts
5966 * disabled.
5967 *
5968 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5969 * outside of boot time (or some other assurance that no concurrent updaters
5970 * exist).
5971 */
5972static void pageset_update(struct per_cpu_pages *pcp, unsigned long high_min,
5973 unsigned long high_max, unsigned long batch)
5974{
5975 WRITE_ONCE(pcp->batch, batch);
5976 WRITE_ONCE(pcp->high_min, high_min);
5977 WRITE_ONCE(pcp->high_max, high_max);
5978}
5979
5980static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
5981{
5982 int pindex;
5983
5984 memset(s: pcp, c: 0, n: sizeof(*pcp));
5985 memset(s: pzstats, c: 0, n: sizeof(*pzstats));
5986
5987 spin_lock_init(&pcp->lock);
5988 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
5989 INIT_LIST_HEAD(list: &pcp->lists[pindex]);
5990
5991 /*
5992 * Set batch and high values safe for a boot pageset. A true percpu
5993 * pageset's initialization will update them subsequently. Here we don't
5994 * need to be as careful as pageset_update() as nobody can access the
5995 * pageset yet.
5996 */
5997 pcp->high_min = BOOT_PAGESET_HIGH;
5998 pcp->high_max = BOOT_PAGESET_HIGH;
5999 pcp->batch = BOOT_PAGESET_BATCH;
6000}
6001
6002static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high_min,
6003 unsigned long high_max, unsigned long batch)
6004{
6005 struct per_cpu_pages *pcp;
6006 int cpu;
6007
6008 for_each_possible_cpu(cpu) {
6009 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6010 pageset_update(pcp, high_min, high_max, batch);
6011 }
6012}
6013
6014/*
6015 * Calculate and set new high and batch values for all per-cpu pagesets of a
6016 * zone based on the zone's size.
6017 */
6018static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
6019{
6020 int new_high_min, new_high_max, new_batch;
6021
6022 new_batch = max(1, zone_batchsize(zone));
6023 if (percpu_pagelist_high_fraction) {
6024 new_high_min = zone_highsize(zone, batch: new_batch, cpu_online,
6025 high_fraction: percpu_pagelist_high_fraction);
6026 /*
6027 * PCP high is tuned manually, disable auto-tuning via
6028 * setting high_min and high_max to the manual value.
6029 */
6030 new_high_max = new_high_min;
6031 } else {
6032 new_high_min = zone_highsize(zone, batch: new_batch, cpu_online, high_fraction: 0);
6033 new_high_max = zone_highsize(zone, batch: new_batch, cpu_online,
6034 MIN_PERCPU_PAGELIST_HIGH_FRACTION);
6035 }
6036
6037 if (zone->pageset_high_min == new_high_min &&
6038 zone->pageset_high_max == new_high_max &&
6039 zone->pageset_batch == new_batch)
6040 return;
6041
6042 zone->pageset_high_min = new_high_min;
6043 zone->pageset_high_max = new_high_max;
6044 zone->pageset_batch = new_batch;
6045
6046 __zone_set_pageset_high_and_batch(zone, high_min: new_high_min, high_max: new_high_max,
6047 batch: new_batch);
6048}
6049
6050void __meminit setup_zone_pageset(struct zone *zone)
6051{
6052 int cpu;
6053
6054 /* Size may be 0 on !SMP && !NUMA */
6055 if (sizeof(struct per_cpu_zonestat) > 0)
6056 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
6057
6058 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
6059 for_each_possible_cpu(cpu) {
6060 struct per_cpu_pages *pcp;
6061 struct per_cpu_zonestat *pzstats;
6062
6063 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6064 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
6065 per_cpu_pages_init(pcp, pzstats);
6066 }
6067
6068 zone_set_pageset_high_and_batch(zone, cpu_online: 0);
6069}
6070
6071/*
6072 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6073 * page high values need to be recalculated.
6074 */
6075static void zone_pcp_update(struct zone *zone, int cpu_online)
6076{
6077 mutex_lock(lock: &pcp_batch_high_lock);
6078 zone_set_pageset_high_and_batch(zone, cpu_online);
6079 mutex_unlock(lock: &pcp_batch_high_lock);
6080}
6081
6082static void zone_pcp_update_cacheinfo(struct zone *zone, unsigned int cpu)
6083{
6084 struct per_cpu_pages *pcp;
6085 struct cpu_cacheinfo *cci;
6086
6087 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6088 cci = get_cpu_cacheinfo(cpu);
6089 /*
6090 * If data cache slice of CPU is large enough, "pcp->batch"
6091 * pages can be preserved in PCP before draining PCP for
6092 * consecutive high-order pages freeing without allocation.
6093 * This can reduce zone lock contention without hurting
6094 * cache-hot pages sharing.
6095 */
6096 spin_lock(lock: &pcp->lock);
6097 if ((cci->per_cpu_data_slice_size >> PAGE_SHIFT) > 3 * pcp->batch)
6098 pcp->flags |= PCPF_FREE_HIGH_BATCH;
6099 else
6100 pcp->flags &= ~PCPF_FREE_HIGH_BATCH;
6101 spin_unlock(lock: &pcp->lock);
6102}
6103
6104void setup_pcp_cacheinfo(unsigned int cpu)
6105{
6106 struct zone *zone;
6107
6108 for_each_populated_zone(zone)
6109 zone_pcp_update_cacheinfo(zone, cpu);
6110}
6111
6112/*
6113 * Allocate per cpu pagesets and initialize them.
6114 * Before this call only boot pagesets were available.
6115 */
6116void __init setup_per_cpu_pageset(void)
6117{
6118 struct pglist_data *pgdat;
6119 struct zone *zone;
6120 int __maybe_unused cpu;
6121
6122 for_each_populated_zone(zone)
6123 setup_zone_pageset(zone);
6124
6125#ifdef CONFIG_NUMA
6126 /*
6127 * Unpopulated zones continue using the boot pagesets.
6128 * The numa stats for these pagesets need to be reset.
6129 * Otherwise, they will end up skewing the stats of
6130 * the nodes these zones are associated with.
6131 */
6132 for_each_possible_cpu(cpu) {
6133 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
6134 memset(s: pzstats->vm_numa_event, c: 0,
6135 n: sizeof(pzstats->vm_numa_event));
6136 }
6137#endif
6138
6139 for_each_online_pgdat(pgdat)
6140 pgdat->per_cpu_nodestats =
6141 alloc_percpu(struct per_cpu_nodestat);
6142}
6143
6144__meminit void zone_pcp_init(struct zone *zone)
6145{
6146 /*
6147 * per cpu subsystem is not up at this point. The following code
6148 * relies on the ability of the linker to provide the
6149 * offset of a (static) per cpu variable into the per cpu area.
6150 */
6151 zone->per_cpu_pageset = &boot_pageset;
6152 zone->per_cpu_zonestats = &boot_zonestats;
6153 zone->pageset_high_min = BOOT_PAGESET_HIGH;
6154 zone->pageset_high_max = BOOT_PAGESET_HIGH;
6155 zone->pageset_batch = BOOT_PAGESET_BATCH;
6156
6157 if (populated_zone(zone))
6158 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
6159 zone->present_pages, zone_batchsize(zone));
6160}
6161
6162static void setup_per_zone_lowmem_reserve(void);
6163
6164void adjust_managed_page_count(struct page *page, long count)
6165{
6166 atomic_long_add(i: count, v: &page_zone(page)->managed_pages);
6167 totalram_pages_add(count);
6168 setup_per_zone_lowmem_reserve();
6169}
6170EXPORT_SYMBOL(adjust_managed_page_count);
6171
6172unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
6173{
6174 void *pos;
6175 unsigned long pages = 0;
6176
6177 start = (void *)PAGE_ALIGN((unsigned long)start);
6178 end = (void *)((unsigned long)end & PAGE_MASK);
6179 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6180 struct page *page = virt_to_page(pos);
6181 void *direct_map_addr;
6182
6183 /*
6184 * 'direct_map_addr' might be different from 'pos'
6185 * because some architectures' virt_to_page()
6186 * work with aliases. Getting the direct map
6187 * address ensures that we get a _writeable_
6188 * alias for the memset().
6189 */
6190 direct_map_addr = page_address(page);
6191 /*
6192 * Perform a kasan-unchecked memset() since this memory
6193 * has not been initialized.
6194 */
6195 direct_map_addr = kasan_reset_tag(addr: direct_map_addr);
6196 if ((unsigned int)poison <= 0xFF)
6197 memset(s: direct_map_addr, c: poison, PAGE_SIZE);
6198
6199 free_reserved_page(page);
6200 }
6201
6202 if (pages && s)
6203 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
6204
6205 return pages;
6206}
6207
6208void free_reserved_page(struct page *page)
6209{
6210 clear_page_tag_ref(page);
6211 ClearPageReserved(page);
6212 init_page_count(page);
6213 __free_page(page);
6214 adjust_managed_page_count(page, 1);
6215}
6216EXPORT_SYMBOL(free_reserved_page);
6217
6218static int page_alloc_cpu_dead(unsigned int cpu)
6219{
6220 struct zone *zone;
6221
6222 lru_add_drain_cpu(cpu);
6223 mlock_drain_remote(cpu);
6224 drain_pages(cpu);
6225
6226 /*
6227 * Spill the event counters of the dead processor
6228 * into the current processors event counters.
6229 * This artificially elevates the count of the current
6230 * processor.
6231 */
6232 vm_events_fold_cpu(cpu);
6233
6234 /*
6235 * Zero the differential counters of the dead processor
6236 * so that the vm statistics are consistent.
6237 *
6238 * This is only okay since the processor is dead and cannot
6239 * race with what we are doing.
6240 */
6241 cpu_vm_stats_fold(cpu);
6242
6243 for_each_populated_zone(zone)
6244 zone_pcp_update(zone, cpu_online: 0);
6245
6246 return 0;
6247}
6248
6249static int page_alloc_cpu_online(unsigned int cpu)
6250{
6251 struct zone *zone;
6252
6253 for_each_populated_zone(zone)
6254 zone_pcp_update(zone, cpu_online: 1);
6255 return 0;
6256}
6257
6258void __init page_alloc_init_cpuhp(void)
6259{
6260 int ret;
6261
6262 ret = cpuhp_setup_state_nocalls(state: CPUHP_PAGE_ALLOC,
6263 name: "mm/page_alloc:pcp",
6264 startup: page_alloc_cpu_online,
6265 teardown: page_alloc_cpu_dead);
6266 WARN_ON(ret < 0);
6267}
6268
6269/*
6270 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6271 * or min_free_kbytes changes.
6272 */
6273static void calculate_totalreserve_pages(void)
6274{
6275 struct pglist_data *pgdat;
6276 unsigned long reserve_pages = 0;
6277 enum zone_type i, j;
6278
6279 for_each_online_pgdat(pgdat) {
6280
6281 pgdat->totalreserve_pages = 0;
6282
6283 for (i = 0; i < MAX_NR_ZONES; i++) {
6284 struct zone *zone = pgdat->node_zones + i;
6285 long max = 0;
6286 unsigned long managed_pages = zone_managed_pages(zone);
6287
6288 /* Find valid and maximum lowmem_reserve in the zone */
6289 for (j = i; j < MAX_NR_ZONES; j++)
6290 max = max(max, zone->lowmem_reserve[j]);
6291
6292 /* we treat the high watermark as reserved pages. */
6293 max += high_wmark_pages(z: zone);
6294
6295 max = min_t(unsigned long, max, managed_pages);
6296
6297 pgdat->totalreserve_pages += max;
6298
6299 reserve_pages += max;
6300 }
6301 }
6302 totalreserve_pages = reserve_pages;
6303 trace_mm_calculate_totalreserve_pages(totalreserve_pages);
6304}
6305
6306/*
6307 * setup_per_zone_lowmem_reserve - called whenever
6308 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6309 * has a correct pages reserved value, so an adequate number of
6310 * pages are left in the zone after a successful __alloc_pages().
6311 */
6312static void setup_per_zone_lowmem_reserve(void)
6313{
6314 struct pglist_data *pgdat;
6315 enum zone_type i, j;
6316
6317 for_each_online_pgdat(pgdat) {
6318 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
6319 struct zone *zone = &pgdat->node_zones[i];
6320 int ratio = sysctl_lowmem_reserve_ratio[i];
6321 bool clear = !ratio || !zone_managed_pages(zone);
6322 unsigned long managed_pages = 0;
6323
6324 for (j = i + 1; j < MAX_NR_ZONES; j++) {
6325 struct zone *upper_zone = &pgdat->node_zones[j];
6326
6327 managed_pages += zone_managed_pages(zone: upper_zone);
6328
6329 if (clear)
6330 zone->lowmem_reserve[j] = 0;
6331 else
6332 zone->lowmem_reserve[j] = managed_pages / ratio;
6333 trace_mm_setup_per_zone_lowmem_reserve(zone, upper_zone,
6334 lowmem_reserve: zone->lowmem_reserve[j]);
6335 }
6336 }
6337 }
6338
6339 /* update totalreserve_pages */
6340 calculate_totalreserve_pages();
6341}
6342
6343static void __setup_per_zone_wmarks(void)
6344{
6345 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6346 unsigned long lowmem_pages = 0;
6347 struct zone *zone;
6348 unsigned long flags;
6349
6350 /* Calculate total number of !ZONE_HIGHMEM and !ZONE_MOVABLE pages */
6351 for_each_zone(zone) {
6352 if (!is_highmem(zone) && zone_idx(zone) != ZONE_MOVABLE)
6353 lowmem_pages += zone_managed_pages(zone);
6354 }
6355
6356 for_each_zone(zone) {
6357 u64 tmp;
6358
6359 spin_lock_irqsave(&zone->lock, flags);
6360 tmp = (u64)pages_min * zone_managed_pages(zone);
6361 tmp = div64_ul(tmp, lowmem_pages);
6362 if (is_highmem(zone) || zone_idx(zone) == ZONE_MOVABLE) {
6363 /*
6364 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6365 * need highmem and movable zones pages, so cap pages_min
6366 * to a small value here.
6367 *
6368 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6369 * deltas control async page reclaim, and so should
6370 * not be capped for highmem and movable zones.
6371 */
6372 unsigned long min_pages;
6373
6374 min_pages = zone_managed_pages(zone) / 1024;
6375 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6376 zone->_watermark[WMARK_MIN] = min_pages;
6377 } else {
6378 /*
6379 * If it's a lowmem zone, reserve a number of pages
6380 * proportionate to the zone's size.
6381 */
6382 zone->_watermark[WMARK_MIN] = tmp;
6383 }
6384
6385 /*
6386 * Set the kswapd watermarks distance according to the
6387 * scale factor in proportion to available memory, but
6388 * ensure a minimum size on small systems.
6389 */
6390 tmp = max_t(u64, tmp >> 2,
6391 mult_frac(zone_managed_pages(zone),
6392 watermark_scale_factor, 10000));
6393
6394 zone->watermark_boost = 0;
6395 zone->_watermark[WMARK_LOW] = min_wmark_pages(z: zone) + tmp;
6396 zone->_watermark[WMARK_HIGH] = low_wmark_pages(z: zone) + tmp;
6397 zone->_watermark[WMARK_PROMO] = high_wmark_pages(z: zone) + tmp;
6398 trace_mm_setup_per_zone_wmarks(zone);
6399
6400 spin_unlock_irqrestore(lock: &zone->lock, flags);
6401 }
6402
6403 /* update totalreserve_pages */
6404 calculate_totalreserve_pages();
6405}
6406
6407/**
6408 * setup_per_zone_wmarks - called when min_free_kbytes changes
6409 * or when memory is hot-{added|removed}
6410 *
6411 * Ensures that the watermark[min,low,high] values for each zone are set
6412 * correctly with respect to min_free_kbytes.
6413 */
6414void setup_per_zone_wmarks(void)
6415{
6416 struct zone *zone;
6417 static DEFINE_SPINLOCK(lock);
6418
6419 spin_lock(lock: &lock);
6420 __setup_per_zone_wmarks();
6421 spin_unlock(lock: &lock);
6422
6423 /*
6424 * The watermark size have changed so update the pcpu batch
6425 * and high limits or the limits may be inappropriate.
6426 */
6427 for_each_zone(zone)
6428 zone_pcp_update(zone, cpu_online: 0);
6429}
6430
6431/*
6432 * Initialise min_free_kbytes.
6433 *
6434 * For small machines we want it small (128k min). For large machines
6435 * we want it large (256MB max). But it is not linear, because network
6436 * bandwidth does not increase linearly with machine size. We use
6437 *
6438 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6439 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6440 *
6441 * which yields
6442 *
6443 * 16MB: 512k
6444 * 32MB: 724k
6445 * 64MB: 1024k
6446 * 128MB: 1448k
6447 * 256MB: 2048k
6448 * 512MB: 2896k
6449 * 1024MB: 4096k
6450 * 2048MB: 5792k
6451 * 4096MB: 8192k
6452 * 8192MB: 11584k
6453 * 16384MB: 16384k
6454 */
6455void calculate_min_free_kbytes(void)
6456{
6457 unsigned long lowmem_kbytes;
6458 int new_min_free_kbytes;
6459
6460 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6461 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6462
6463 if (new_min_free_kbytes > user_min_free_kbytes)
6464 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
6465 else
6466 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6467 new_min_free_kbytes, user_min_free_kbytes);
6468
6469}
6470
6471int __meminit init_per_zone_wmark_min(void)
6472{
6473 calculate_min_free_kbytes();
6474 setup_per_zone_wmarks();
6475 refresh_zone_stat_thresholds();
6476 setup_per_zone_lowmem_reserve();
6477
6478#ifdef CONFIG_NUMA
6479 setup_min_unmapped_ratio();
6480 setup_min_slab_ratio();
6481#endif
6482
6483 khugepaged_min_free_kbytes_update();
6484
6485 return 0;
6486}
6487postcore_initcall(init_per_zone_wmark_min)
6488
6489/*
6490 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6491 * that we can call two helper functions whenever min_free_kbytes
6492 * changes.
6493 */
6494static int min_free_kbytes_sysctl_handler(const struct ctl_table *table, int write,
6495 void *buffer, size_t *length, loff_t *ppos)
6496{
6497 int rc;
6498
6499 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6500 if (rc)
6501 return rc;
6502
6503 if (write) {
6504 user_min_free_kbytes = min_free_kbytes;
6505 setup_per_zone_wmarks();
6506 }
6507 return 0;
6508}
6509
6510static int watermark_scale_factor_sysctl_handler(const struct ctl_table *table, int write,
6511 void *buffer, size_t *length, loff_t *ppos)
6512{
6513 int rc;
6514
6515 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6516 if (rc)
6517 return rc;
6518
6519 if (write)
6520 setup_per_zone_wmarks();
6521
6522 return 0;
6523}
6524
6525#ifdef CONFIG_NUMA
6526static void setup_min_unmapped_ratio(void)
6527{
6528 pg_data_t *pgdat;
6529 struct zone *zone;
6530
6531 for_each_online_pgdat(pgdat)
6532 pgdat->min_unmapped_pages = 0;
6533
6534 for_each_zone(zone)
6535 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
6536 sysctl_min_unmapped_ratio) / 100;
6537}
6538
6539
6540static int sysctl_min_unmapped_ratio_sysctl_handler(const struct ctl_table *table, int write,
6541 void *buffer, size_t *length, loff_t *ppos)
6542{
6543 int rc;
6544
6545 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6546 if (rc)
6547 return rc;
6548
6549 setup_min_unmapped_ratio();
6550
6551 return 0;
6552}
6553
6554static void setup_min_slab_ratio(void)
6555{
6556 pg_data_t *pgdat;
6557 struct zone *zone;
6558
6559 for_each_online_pgdat(pgdat)
6560 pgdat->min_slab_pages = 0;
6561
6562 for_each_zone(zone)
6563 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
6564 sysctl_min_slab_ratio) / 100;
6565}
6566
6567static int sysctl_min_slab_ratio_sysctl_handler(const struct ctl_table *table, int write,
6568 void *buffer, size_t *length, loff_t *ppos)
6569{
6570 int rc;
6571
6572 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6573 if (rc)
6574 return rc;
6575
6576 setup_min_slab_ratio();
6577
6578 return 0;
6579}
6580#endif
6581
6582/*
6583 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6584 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6585 * whenever sysctl_lowmem_reserve_ratio changes.
6586 *
6587 * The reserve ratio obviously has absolutely no relation with the
6588 * minimum watermarks. The lowmem reserve ratio can only make sense
6589 * if in function of the boot time zone sizes.
6590 */
6591static int lowmem_reserve_ratio_sysctl_handler(const struct ctl_table *table,
6592 int write, void *buffer, size_t *length, loff_t *ppos)
6593{
6594 int i;
6595
6596 proc_dointvec_minmax(table, write, buffer, length, ppos);
6597
6598 for (i = 0; i < MAX_NR_ZONES; i++) {
6599 if (sysctl_lowmem_reserve_ratio[i] < 1)
6600 sysctl_lowmem_reserve_ratio[i] = 0;
6601 }
6602
6603 setup_per_zone_lowmem_reserve();
6604 return 0;
6605}
6606
6607/*
6608 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
6609 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6610 * pagelist can have before it gets flushed back to buddy allocator.
6611 */
6612static int percpu_pagelist_high_fraction_sysctl_handler(const struct ctl_table *table,
6613 int write, void *buffer, size_t *length, loff_t *ppos)
6614{
6615 struct zone *zone;
6616 int old_percpu_pagelist_high_fraction;
6617 int ret;
6618
6619 mutex_lock(lock: &pcp_batch_high_lock);
6620 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
6621
6622 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6623 if (!write || ret < 0)
6624 goto out;
6625
6626 /* Sanity checking to avoid pcp imbalance */
6627 if (percpu_pagelist_high_fraction &&
6628 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
6629 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
6630 ret = -EINVAL;
6631 goto out;
6632 }
6633
6634 /* No change? */
6635 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
6636 goto out;
6637
6638 for_each_populated_zone(zone)
6639 zone_set_pageset_high_and_batch(zone, cpu_online: 0);
6640out:
6641 mutex_unlock(lock: &pcp_batch_high_lock);
6642 return ret;
6643}
6644
6645static const struct ctl_table page_alloc_sysctl_table[] = {
6646 {
6647 .procname = "min_free_kbytes",
6648 .data = &min_free_kbytes,
6649 .maxlen = sizeof(min_free_kbytes),
6650 .mode = 0644,
6651 .proc_handler = min_free_kbytes_sysctl_handler,
6652 .extra1 = SYSCTL_ZERO,
6653 },
6654 {
6655 .procname = "watermark_boost_factor",
6656 .data = &watermark_boost_factor,
6657 .maxlen = sizeof(watermark_boost_factor),
6658 .mode = 0644,
6659 .proc_handler = proc_dointvec_minmax,
6660 .extra1 = SYSCTL_ZERO,
6661 },
6662 {
6663 .procname = "watermark_scale_factor",
6664 .data = &watermark_scale_factor,
6665 .maxlen = sizeof(watermark_scale_factor),
6666 .mode = 0644,
6667 .proc_handler = watermark_scale_factor_sysctl_handler,
6668 .extra1 = SYSCTL_ONE,
6669 .extra2 = SYSCTL_THREE_THOUSAND,
6670 },
6671 {
6672 .procname = "defrag_mode",
6673 .data = &defrag_mode,
6674 .maxlen = sizeof(defrag_mode),
6675 .mode = 0644,
6676 .proc_handler = proc_dointvec_minmax,
6677 .extra1 = SYSCTL_ZERO,
6678 .extra2 = SYSCTL_ONE,
6679 },
6680 {
6681 .procname = "percpu_pagelist_high_fraction",
6682 .data = &percpu_pagelist_high_fraction,
6683 .maxlen = sizeof(percpu_pagelist_high_fraction),
6684 .mode = 0644,
6685 .proc_handler = percpu_pagelist_high_fraction_sysctl_handler,
6686 .extra1 = SYSCTL_ZERO,
6687 },
6688 {
6689 .procname = "lowmem_reserve_ratio",
6690 .data = &sysctl_lowmem_reserve_ratio,
6691 .maxlen = sizeof(sysctl_lowmem_reserve_ratio),
6692 .mode = 0644,
6693 .proc_handler = lowmem_reserve_ratio_sysctl_handler,
6694 },
6695#ifdef CONFIG_NUMA
6696 {
6697 .procname = "numa_zonelist_order",
6698 .data = &numa_zonelist_order,
6699 .maxlen = NUMA_ZONELIST_ORDER_LEN,
6700 .mode = 0644,
6701 .proc_handler = numa_zonelist_order_handler,
6702 },
6703 {
6704 .procname = "min_unmapped_ratio",
6705 .data = &sysctl_min_unmapped_ratio,
6706 .maxlen = sizeof(sysctl_min_unmapped_ratio),
6707 .mode = 0644,
6708 .proc_handler = sysctl_min_unmapped_ratio_sysctl_handler,
6709 .extra1 = SYSCTL_ZERO,
6710 .extra2 = SYSCTL_ONE_HUNDRED,
6711 },
6712 {
6713 .procname = "min_slab_ratio",
6714 .data = &sysctl_min_slab_ratio,
6715 .maxlen = sizeof(sysctl_min_slab_ratio),
6716 .mode = 0644,
6717 .proc_handler = sysctl_min_slab_ratio_sysctl_handler,
6718 .extra1 = SYSCTL_ZERO,
6719 .extra2 = SYSCTL_ONE_HUNDRED,
6720 },
6721#endif
6722};
6723
6724void __init page_alloc_sysctl_init(void)
6725{
6726 register_sysctl_init("vm", page_alloc_sysctl_table);
6727}
6728
6729#ifdef CONFIG_CONTIG_ALLOC
6730/* Usage: See admin-guide/dynamic-debug-howto.rst */
6731static void alloc_contig_dump_pages(struct list_head *page_list)
6732{
6733 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
6734
6735 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
6736 struct page *page;
6737
6738 dump_stack();
6739 list_for_each_entry(page, page_list, lru)
6740 dump_page(page, "migration failure");
6741 }
6742}
6743
6744/* [start, end) must belong to a single zone. */
6745static int __alloc_contig_migrate_range(struct compact_control *cc,
6746 unsigned long start, unsigned long end)
6747{
6748 /* This function is based on compact_zone() from compaction.c. */
6749 unsigned int nr_reclaimed;
6750 unsigned long pfn = start;
6751 unsigned int tries = 0;
6752 int ret = 0;
6753 struct migration_target_control mtc = {
6754 .nid = zone_to_nid(cc->zone),
6755 .gfp_mask = cc->gfp_mask,
6756 .reason = MR_CONTIG_RANGE,
6757 };
6758
6759 lru_cache_disable();
6760
6761 while (pfn < end || !list_empty(&cc->migratepages)) {
6762 if (fatal_signal_pending(current)) {
6763 ret = -EINTR;
6764 break;
6765 }
6766
6767 if (list_empty(&cc->migratepages)) {
6768 cc->nr_migratepages = 0;
6769 ret = isolate_migratepages_range(cc, pfn, end);
6770 if (ret && ret != -EAGAIN)
6771 break;
6772 pfn = cc->migrate_pfn;
6773 tries = 0;
6774 } else if (++tries == 5) {
6775 ret = -EBUSY;
6776 break;
6777 }
6778
6779 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6780 &cc->migratepages);
6781 cc->nr_migratepages -= nr_reclaimed;
6782
6783 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
6784 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
6785
6786 /*
6787 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
6788 * to retry again over this error, so do the same here.
6789 */
6790 if (ret == -ENOMEM)
6791 break;
6792 }
6793
6794 lru_cache_enable();
6795 if (ret < 0) {
6796 if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY)
6797 alloc_contig_dump_pages(&cc->migratepages);
6798 putback_movable_pages(&cc->migratepages);
6799 }
6800
6801 return (ret < 0) ? ret : 0;
6802}
6803
6804static void split_free_pages(struct list_head *list, gfp_t gfp_mask)
6805{
6806 int order;
6807
6808 for (order = 0; order < NR_PAGE_ORDERS; order++) {
6809 struct page *page, *next;
6810 int nr_pages = 1 << order;
6811
6812 list_for_each_entry_safe(page, next, &list[order], lru) {
6813 int i;
6814
6815 post_alloc_hook(page, order, gfp_mask);
6816 set_page_refcounted(page);
6817 if (!order)
6818 continue;
6819
6820 split_page(page, order);
6821
6822 /* Add all subpages to the order-0 head, in sequence. */
6823 list_del(&page->lru);
6824 for (i = 0; i < nr_pages; i++)
6825 list_add_tail(&page[i].lru, &list[0]);
6826 }
6827 }
6828}
6829
6830static int __alloc_contig_verify_gfp_mask(gfp_t gfp_mask, gfp_t *gfp_cc_mask)
6831{
6832 const gfp_t reclaim_mask = __GFP_IO | __GFP_FS | __GFP_RECLAIM;
6833 const gfp_t action_mask = __GFP_COMP | __GFP_RETRY_MAYFAIL | __GFP_NOWARN |
6834 __GFP_ZERO | __GFP_ZEROTAGS | __GFP_SKIP_ZERO;
6835 const gfp_t cc_action_mask = __GFP_RETRY_MAYFAIL | __GFP_NOWARN;
6836
6837 /*
6838 * We are given the range to allocate; node, mobility and placement
6839 * hints are irrelevant at this point. We'll simply ignore them.
6840 */
6841 gfp_mask &= ~(GFP_ZONEMASK | __GFP_RECLAIMABLE | __GFP_WRITE |
6842 __GFP_HARDWALL | __GFP_THISNODE | __GFP_MOVABLE);
6843
6844 /*
6845 * We only support most reclaim flags (but not NOFAIL/NORETRY), and
6846 * selected action flags.
6847 */
6848 if (gfp_mask & ~(reclaim_mask | action_mask))
6849 return -EINVAL;
6850
6851 /*
6852 * Flags to control page compaction/migration/reclaim, to free up our
6853 * page range. Migratable pages are movable, __GFP_MOVABLE is implied
6854 * for them.
6855 *
6856 * Traditionally we always had __GFP_RETRY_MAYFAIL set, keep doing that
6857 * to not degrade callers.
6858 */
6859 *gfp_cc_mask = (gfp_mask & (reclaim_mask | cc_action_mask)) |
6860 __GFP_MOVABLE | __GFP_RETRY_MAYFAIL;
6861 return 0;
6862}
6863
6864/**
6865 * alloc_contig_range() -- tries to allocate given range of pages
6866 * @start: start PFN to allocate
6867 * @end: one-past-the-last PFN to allocate
6868 * @alloc_flags: allocation information
6869 * @gfp_mask: GFP mask. Node/zone/placement hints are ignored; only some
6870 * action and reclaim modifiers are supported. Reclaim modifiers
6871 * control allocation behavior during compaction/migration/reclaim.
6872 *
6873 * The PFN range does not have to be pageblock aligned. The PFN range must
6874 * belong to a single zone.
6875 *
6876 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
6877 * pageblocks in the range. Once isolated, the pageblocks should not
6878 * be modified by others.
6879 *
6880 * Return: zero on success or negative error code. On success all
6881 * pages which PFN is in [start, end) are allocated for the caller and
6882 * need to be freed with free_contig_range().
6883 */
6884int alloc_contig_range_noprof(unsigned long start, unsigned long end,
6885 acr_flags_t alloc_flags, gfp_t gfp_mask)
6886{
6887 const unsigned int order = ilog2(end - start);
6888 unsigned long outer_start, outer_end;
6889 int ret = 0;
6890
6891 struct compact_control cc = {
6892 .nr_migratepages = 0,
6893 .order = -1,
6894 .zone = page_zone(pfn_to_page(start)),
6895 .mode = MIGRATE_SYNC,
6896 .ignore_skip_hint = true,
6897 .no_set_skip_hint = true,
6898 .alloc_contig = true,
6899 };
6900 INIT_LIST_HEAD(&cc.migratepages);
6901 enum pb_isolate_mode mode = (alloc_flags & ACR_FLAGS_CMA) ?
6902 PB_ISOLATE_MODE_CMA_ALLOC :
6903 PB_ISOLATE_MODE_OTHER;
6904
6905 /*
6906 * In contrast to the buddy, we allow for orders here that exceed
6907 * MAX_PAGE_ORDER, so we must manually make sure that we are not
6908 * exceeding the maximum folio order.
6909 */
6910 if (WARN_ON_ONCE((gfp_mask & __GFP_COMP) && order > MAX_FOLIO_ORDER))
6911 return -EINVAL;
6912
6913 gfp_mask = current_gfp_context(gfp_mask);
6914 if (__alloc_contig_verify_gfp_mask(gfp_mask, (gfp_t *)&cc.gfp_mask))
6915 return -EINVAL;
6916
6917 /*
6918 * What we do here is we mark all pageblocks in range as
6919 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6920 * have different sizes, and due to the way page allocator
6921 * work, start_isolate_page_range() has special handlings for this.
6922 *
6923 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6924 * migrate the pages from an unaligned range (ie. pages that
6925 * we are interested in). This will put all the pages in
6926 * range back to page allocator as MIGRATE_ISOLATE.
6927 *
6928 * When this is done, we take the pages in range from page
6929 * allocator removing them from the buddy system. This way
6930 * page allocator will never consider using them.
6931 *
6932 * This lets us mark the pageblocks back as
6933 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6934 * aligned range but not in the unaligned, original range are
6935 * put back to page allocator so that buddy can use them.
6936 */
6937
6938 ret = start_isolate_page_range(start, end, mode);
6939 if (ret)
6940 goto done;
6941
6942 drain_all_pages(cc.zone);
6943
6944 /*
6945 * In case of -EBUSY, we'd like to know which page causes problem.
6946 * So, just fall through. test_pages_isolated() has a tracepoint
6947 * which will report the busy page.
6948 *
6949 * It is possible that busy pages could become available before
6950 * the call to test_pages_isolated, and the range will actually be
6951 * allocated. So, if we fall through be sure to clear ret so that
6952 * -EBUSY is not accidentally used or returned to caller.
6953 */
6954 ret = __alloc_contig_migrate_range(&cc, start, end);
6955 if (ret && ret != -EBUSY)
6956 goto done;
6957
6958 /*
6959 * When in-use hugetlb pages are migrated, they may simply be released
6960 * back into the free hugepage pool instead of being returned to the
6961 * buddy system. After the migration of in-use huge pages is completed,
6962 * we will invoke replace_free_hugepage_folios() to ensure that these
6963 * hugepages are properly released to the buddy system.
6964 */
6965 ret = replace_free_hugepage_folios(start, end);
6966 if (ret)
6967 goto done;
6968
6969 /*
6970 * Pages from [start, end) are within a pageblock_nr_pages
6971 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6972 * more, all pages in [start, end) are free in page allocator.
6973 * What we are going to do is to allocate all pages from
6974 * [start, end) (that is remove them from page allocator).
6975 *
6976 * The only problem is that pages at the beginning and at the
6977 * end of interesting range may be not aligned with pages that
6978 * page allocator holds, ie. they can be part of higher order
6979 * pages. Because of this, we reserve the bigger range and
6980 * once this is done free the pages we are not interested in.
6981 *
6982 * We don't have to hold zone->lock here because the pages are
6983 * isolated thus they won't get removed from buddy.
6984 */
6985 outer_start = find_large_buddy(start);
6986
6987 /* Make sure the range is really isolated. */
6988 if (test_pages_isolated(outer_start, end, mode)) {
6989 ret = -EBUSY;
6990 goto done;
6991 }
6992
6993 /* Grab isolated pages from freelists. */
6994 outer_end = isolate_freepages_range(&cc, outer_start, end);
6995 if (!outer_end) {
6996 ret = -EBUSY;
6997 goto done;
6998 }
6999
7000 if (!(gfp_mask & __GFP_COMP)) {
7001 split_free_pages(cc.freepages, gfp_mask);
7002
7003 /* Free head and tail (if any) */
7004 if (start != outer_start)
7005 free_contig_range(outer_start, start - outer_start);
7006 if (end != outer_end)
7007 free_contig_range(end, outer_end - end);
7008 } else if (start == outer_start && end == outer_end && is_power_of_2(end - start)) {
7009 struct page *head = pfn_to_page(start);
7010
7011 check_new_pages(head, order);
7012 prep_new_page(head, order, gfp_mask, 0);
7013 set_page_refcounted(head);
7014 } else {
7015 ret = -EINVAL;
7016 WARN(true, "PFN range: requested [%lu, %lu), allocated [%lu, %lu)\n",
7017 start, end, outer_start, outer_end);
7018 }
7019done:
7020 undo_isolate_page_range(start, end);
7021 return ret;
7022}
7023EXPORT_SYMBOL(alloc_contig_range_noprof);
7024
7025static int __alloc_contig_pages(unsigned long start_pfn,
7026 unsigned long nr_pages, gfp_t gfp_mask)
7027{
7028 unsigned long end_pfn = start_pfn + nr_pages;
7029
7030 return alloc_contig_range_noprof(start_pfn, end_pfn, ACR_FLAGS_NONE,
7031 gfp_mask);
7032}
7033
7034static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
7035 unsigned long nr_pages)
7036{
7037 unsigned long i, end_pfn = start_pfn + nr_pages;
7038 struct page *page;
7039
7040 for (i = start_pfn; i < end_pfn; i++) {
7041 page = pfn_to_online_page(i);
7042 if (!page)
7043 return false;
7044
7045 if (page_zone(page) != z)
7046 return false;
7047
7048 if (PageReserved(page))
7049 return false;
7050
7051 if (PageHuge(page))
7052 return false;
7053 }
7054 return true;
7055}
7056
7057static bool zone_spans_last_pfn(const struct zone *zone,
7058 unsigned long start_pfn, unsigned long nr_pages)
7059{
7060 unsigned long last_pfn = start_pfn + nr_pages - 1;
7061
7062 return zone_spans_pfn(zone, last_pfn);
7063}
7064
7065/**
7066 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
7067 * @nr_pages: Number of contiguous pages to allocate
7068 * @gfp_mask: GFP mask. Node/zone/placement hints limit the search; only some
7069 * action and reclaim modifiers are supported. Reclaim modifiers
7070 * control allocation behavior during compaction/migration/reclaim.
7071 * @nid: Target node
7072 * @nodemask: Mask for other possible nodes
7073 *
7074 * This routine is a wrapper around alloc_contig_range(). It scans over zones
7075 * on an applicable zonelist to find a contiguous pfn range which can then be
7076 * tried for allocation with alloc_contig_range(). This routine is intended
7077 * for allocation requests which can not be fulfilled with the buddy allocator.
7078 *
7079 * The allocated memory is always aligned to a page boundary. If nr_pages is a
7080 * power of two, then allocated range is also guaranteed to be aligned to same
7081 * nr_pages (e.g. 1GB request would be aligned to 1GB).
7082 *
7083 * Allocated pages can be freed with free_contig_range() or by manually calling
7084 * __free_page() on each allocated page.
7085 *
7086 * Return: pointer to contiguous pages on success, or NULL if not successful.
7087 */
7088struct page *alloc_contig_pages_noprof(unsigned long nr_pages, gfp_t gfp_mask,
7089 int nid, nodemask_t *nodemask)
7090{
7091 unsigned long ret, pfn, flags;
7092 struct zonelist *zonelist;
7093 struct zone *zone;
7094 struct zoneref *z;
7095
7096 zonelist = node_zonelist(nid, gfp_mask);
7097 for_each_zone_zonelist_nodemask(zone, z, zonelist,
7098 gfp_zone(gfp_mask), nodemask) {
7099 spin_lock_irqsave(&zone->lock, flags);
7100
7101 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
7102 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
7103 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
7104 /*
7105 * We release the zone lock here because
7106 * alloc_contig_range() will also lock the zone
7107 * at some point. If there's an allocation
7108 * spinning on this lock, it may win the race
7109 * and cause alloc_contig_range() to fail...
7110 */
7111 spin_unlock_irqrestore(&zone->lock, flags);
7112 ret = __alloc_contig_pages(pfn, nr_pages,
7113 gfp_mask);
7114 if (!ret)
7115 return pfn_to_page(pfn);
7116 spin_lock_irqsave(&zone->lock, flags);
7117 }
7118 pfn += nr_pages;
7119 }
7120 spin_unlock_irqrestore(&zone->lock, flags);
7121 }
7122 return NULL;
7123}
7124#endif /* CONFIG_CONTIG_ALLOC */
7125
7126void free_contig_range(unsigned long pfn, unsigned long nr_pages)
7127{
7128 unsigned long count = 0;
7129 struct folio *folio = pfn_folio(pfn);
7130
7131 if (folio_test_large(folio)) {
7132 int expected = folio_nr_pages(folio);
7133
7134 if (nr_pages == expected)
7135 folio_put(folio);
7136 else
7137 WARN(true, "PFN %lu: nr_pages %lu != expected %d\n",
7138 pfn, nr_pages, expected);
7139 return;
7140 }
7141
7142 for (; nr_pages--; pfn++) {
7143 struct page *page = pfn_to_page(pfn);
7144
7145 count += page_count(page) != 1;
7146 __free_page(page);
7147 }
7148 WARN(count != 0, "%lu pages are still in use!\n", count);
7149}
7150EXPORT_SYMBOL(free_contig_range);
7151
7152/*
7153 * Effectively disable pcplists for the zone by setting the high limit to 0
7154 * and draining all cpus. A concurrent page freeing on another CPU that's about
7155 * to put the page on pcplist will either finish before the drain and the page
7156 * will be drained, or observe the new high limit and skip the pcplist.
7157 *
7158 * Must be paired with a call to zone_pcp_enable().
7159 */
7160void zone_pcp_disable(struct zone *zone)
7161{
7162 mutex_lock(lock: &pcp_batch_high_lock);
7163 __zone_set_pageset_high_and_batch(zone, high_min: 0, high_max: 0, batch: 1);
7164 __drain_all_pages(zone, force_all_cpus: true);
7165}
7166
7167void zone_pcp_enable(struct zone *zone)
7168{
7169 __zone_set_pageset_high_and_batch(zone, high_min: zone->pageset_high_min,
7170 high_max: zone->pageset_high_max, batch: zone->pageset_batch);
7171 mutex_unlock(lock: &pcp_batch_high_lock);
7172}
7173
7174void zone_pcp_reset(struct zone *zone)
7175{
7176 int cpu;
7177 struct per_cpu_zonestat *pzstats;
7178
7179 if (zone->per_cpu_pageset != &boot_pageset) {
7180 for_each_online_cpu(cpu) {
7181 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
7182 drain_zonestat(zone, pzstats);
7183 }
7184 free_percpu(pdata: zone->per_cpu_pageset);
7185 zone->per_cpu_pageset = &boot_pageset;
7186 if (zone->per_cpu_zonestats != &boot_zonestats) {
7187 free_percpu(pdata: zone->per_cpu_zonestats);
7188 zone->per_cpu_zonestats = &boot_zonestats;
7189 }
7190 }
7191}
7192
7193#ifdef CONFIG_MEMORY_HOTREMOVE
7194/*
7195 * All pages in the range must be in a single zone, must not contain holes,
7196 * must span full sections, and must be isolated before calling this function.
7197 *
7198 * Returns the number of managed (non-PageOffline()) pages in the range: the
7199 * number of pages for which memory offlining code must adjust managed page
7200 * counters using adjust_managed_page_count().
7201 */
7202unsigned long __offline_isolated_pages(unsigned long start_pfn,
7203 unsigned long end_pfn)
7204{
7205 unsigned long already_offline = 0, flags;
7206 unsigned long pfn = start_pfn;
7207 struct page *page;
7208 struct zone *zone;
7209 unsigned int order;
7210
7211 offline_mem_sections(pfn, end_pfn);
7212 zone = page_zone(pfn_to_page(pfn));
7213 spin_lock_irqsave(&zone->lock, flags);
7214 while (pfn < end_pfn) {
7215 page = pfn_to_page(pfn);
7216 /*
7217 * The HWPoisoned page may be not in buddy system, and
7218 * page_count() is not 0.
7219 */
7220 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7221 pfn++;
7222 continue;
7223 }
7224 /*
7225 * At this point all remaining PageOffline() pages have a
7226 * reference count of 0 and can simply be skipped.
7227 */
7228 if (PageOffline(page)) {
7229 BUG_ON(page_count(page));
7230 BUG_ON(PageBuddy(page));
7231 already_offline++;
7232 pfn++;
7233 continue;
7234 }
7235
7236 BUG_ON(page_count(page));
7237 BUG_ON(!PageBuddy(page));
7238 VM_WARN_ON(get_pageblock_migratetype(page) != MIGRATE_ISOLATE);
7239 order = buddy_order(page);
7240 del_page_from_free_list(page, zone, order, MIGRATE_ISOLATE);
7241 pfn += (1 << order);
7242 }
7243 spin_unlock_irqrestore(&zone->lock, flags);
7244
7245 return end_pfn - start_pfn - already_offline;
7246}
7247#endif
7248
7249/*
7250 * This function returns a stable result only if called under zone lock.
7251 */
7252bool is_free_buddy_page(const struct page *page)
7253{
7254 unsigned long pfn = page_to_pfn(page);
7255 unsigned int order;
7256
7257 for (order = 0; order < NR_PAGE_ORDERS; order++) {
7258 const struct page *head = page - (pfn & ((1 << order) - 1));
7259
7260 if (PageBuddy(page: head) &&
7261 buddy_order_unsafe(head) >= order)
7262 break;
7263 }
7264
7265 return order <= MAX_PAGE_ORDER;
7266}
7267EXPORT_SYMBOL(is_free_buddy_page);
7268
7269#ifdef CONFIG_MEMORY_FAILURE
7270static inline void add_to_free_list(struct page *page, struct zone *zone,
7271 unsigned int order, int migratetype,
7272 bool tail)
7273{
7274 __add_to_free_list(page, zone, order, migratetype, tail);
7275 account_freepages(zone, 1 << order, migratetype);
7276}
7277
7278/*
7279 * Break down a higher-order page in sub-pages, and keep our target out of
7280 * buddy allocator.
7281 */
7282static void break_down_buddy_pages(struct zone *zone, struct page *page,
7283 struct page *target, int low, int high,
7284 int migratetype)
7285{
7286 unsigned long size = 1 << high;
7287 struct page *current_buddy;
7288
7289 while (high > low) {
7290 high--;
7291 size >>= 1;
7292
7293 if (target >= &page[size]) {
7294 current_buddy = page;
7295 page = page + size;
7296 } else {
7297 current_buddy = page + size;
7298 }
7299
7300 if (set_page_guard(zone, current_buddy, high))
7301 continue;
7302
7303 add_to_free_list(current_buddy, zone, high, migratetype, false);
7304 set_buddy_order(current_buddy, high);
7305 }
7306}
7307
7308/*
7309 * Take a page that will be marked as poisoned off the buddy allocator.
7310 */
7311bool take_page_off_buddy(struct page *page)
7312{
7313 struct zone *zone = page_zone(page);
7314 unsigned long pfn = page_to_pfn(page);
7315 unsigned long flags;
7316 unsigned int order;
7317 bool ret = false;
7318
7319 spin_lock_irqsave(&zone->lock, flags);
7320 for (order = 0; order < NR_PAGE_ORDERS; order++) {
7321 struct page *page_head = page - (pfn & ((1 << order) - 1));
7322 int page_order = buddy_order(page_head);
7323
7324 if (PageBuddy(page_head) && page_order >= order) {
7325 unsigned long pfn_head = page_to_pfn(page_head);
7326 int migratetype = get_pfnblock_migratetype(page_head,
7327 pfn_head);
7328
7329 del_page_from_free_list(page_head, zone, page_order,
7330 migratetype);
7331 break_down_buddy_pages(zone, page_head, page, 0,
7332 page_order, migratetype);
7333 SetPageHWPoisonTakenOff(page);
7334 ret = true;
7335 break;
7336 }
7337 if (page_count(page_head) > 0)
7338 break;
7339 }
7340 spin_unlock_irqrestore(&zone->lock, flags);
7341 return ret;
7342}
7343
7344/*
7345 * Cancel takeoff done by take_page_off_buddy().
7346 */
7347bool put_page_back_buddy(struct page *page)
7348{
7349 struct zone *zone = page_zone(page);
7350 unsigned long flags;
7351 bool ret = false;
7352
7353 spin_lock_irqsave(&zone->lock, flags);
7354 if (put_page_testzero(page)) {
7355 unsigned long pfn = page_to_pfn(page);
7356 int migratetype = get_pfnblock_migratetype(page, pfn);
7357
7358 ClearPageHWPoisonTakenOff(page);
7359 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
7360 if (TestClearPageHWPoison(page)) {
7361 ret = true;
7362 }
7363 }
7364 spin_unlock_irqrestore(&zone->lock, flags);
7365
7366 return ret;
7367}
7368#endif
7369
7370#ifdef CONFIG_ZONE_DMA
7371bool has_managed_dma(void)
7372{
7373 struct pglist_data *pgdat;
7374
7375 for_each_online_pgdat(pgdat) {
7376 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
7377
7378 if (managed_zone(zone))
7379 return true;
7380 }
7381 return false;
7382}
7383#endif /* CONFIG_ZONE_DMA */
7384
7385#ifdef CONFIG_UNACCEPTED_MEMORY
7386
7387static bool lazy_accept = true;
7388
7389static int __init accept_memory_parse(char *p)
7390{
7391 if (!strcmp(p, "lazy")) {
7392 lazy_accept = true;
7393 return 0;
7394 } else if (!strcmp(p, "eager")) {
7395 lazy_accept = false;
7396 return 0;
7397 } else {
7398 return -EINVAL;
7399 }
7400}
7401early_param("accept_memory", accept_memory_parse);
7402
7403static bool page_contains_unaccepted(struct page *page, unsigned int order)
7404{
7405 phys_addr_t start = page_to_phys(page);
7406
7407 return range_contains_unaccepted_memory(start, PAGE_SIZE << order);
7408}
7409
7410static void __accept_page(struct zone *zone, unsigned long *flags,
7411 struct page *page)
7412{
7413 list_del(&page->lru);
7414 account_freepages(zone, -MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
7415 __mod_zone_page_state(zone, NR_UNACCEPTED, -MAX_ORDER_NR_PAGES);
7416 __ClearPageUnaccepted(page);
7417 spin_unlock_irqrestore(&zone->lock, *flags);
7418
7419 accept_memory(page_to_phys(page), PAGE_SIZE << MAX_PAGE_ORDER);
7420
7421 __free_pages_ok(page, MAX_PAGE_ORDER, FPI_TO_TAIL);
7422}
7423
7424void accept_page(struct page *page)
7425{
7426 struct zone *zone = page_zone(page);
7427 unsigned long flags;
7428
7429 spin_lock_irqsave(&zone->lock, flags);
7430 if (!PageUnaccepted(page)) {
7431 spin_unlock_irqrestore(&zone->lock, flags);
7432 return;
7433 }
7434
7435 /* Unlocks zone->lock */
7436 __accept_page(zone, &flags, page);
7437}
7438
7439static bool try_to_accept_memory_one(struct zone *zone)
7440{
7441 unsigned long flags;
7442 struct page *page;
7443
7444 spin_lock_irqsave(&zone->lock, flags);
7445 page = list_first_entry_or_null(&zone->unaccepted_pages,
7446 struct page, lru);
7447 if (!page) {
7448 spin_unlock_irqrestore(&zone->lock, flags);
7449 return false;
7450 }
7451
7452 /* Unlocks zone->lock */
7453 __accept_page(zone, &flags, page);
7454
7455 return true;
7456}
7457
7458static bool cond_accept_memory(struct zone *zone, unsigned int order,
7459 int alloc_flags)
7460{
7461 long to_accept, wmark;
7462 bool ret = false;
7463
7464 if (list_empty(&zone->unaccepted_pages))
7465 return false;
7466
7467 /* Bailout, since try_to_accept_memory_one() needs to take a lock */
7468 if (alloc_flags & ALLOC_TRYLOCK)
7469 return false;
7470
7471 wmark = promo_wmark_pages(zone);
7472
7473 /*
7474 * Watermarks have not been initialized yet.
7475 *
7476 * Accepting one MAX_ORDER page to ensure progress.
7477 */
7478 if (!wmark)
7479 return try_to_accept_memory_one(zone);
7480
7481 /* How much to accept to get to promo watermark? */
7482 to_accept = wmark -
7483 (zone_page_state(zone, NR_FREE_PAGES) -
7484 __zone_watermark_unusable_free(zone, order, 0) -
7485 zone_page_state(zone, NR_UNACCEPTED));
7486
7487 while (to_accept > 0) {
7488 if (!try_to_accept_memory_one(zone))
7489 break;
7490 ret = true;
7491 to_accept -= MAX_ORDER_NR_PAGES;
7492 }
7493
7494 return ret;
7495}
7496
7497static bool __free_unaccepted(struct page *page)
7498{
7499 struct zone *zone = page_zone(page);
7500 unsigned long flags;
7501
7502 if (!lazy_accept)
7503 return false;
7504
7505 spin_lock_irqsave(&zone->lock, flags);
7506 list_add_tail(&page->lru, &zone->unaccepted_pages);
7507 account_freepages(zone, MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
7508 __mod_zone_page_state(zone, NR_UNACCEPTED, MAX_ORDER_NR_PAGES);
7509 __SetPageUnaccepted(page);
7510 spin_unlock_irqrestore(&zone->lock, flags);
7511
7512 return true;
7513}
7514
7515#else
7516
7517static bool page_contains_unaccepted(struct page *page, unsigned int order)
7518{
7519 return false;
7520}
7521
7522static bool cond_accept_memory(struct zone *zone, unsigned int order,
7523 int alloc_flags)
7524{
7525 return false;
7526}
7527
7528static bool __free_unaccepted(struct page *page)
7529{
7530 BUILD_BUG();
7531 return false;
7532}
7533
7534#endif /* CONFIG_UNACCEPTED_MEMORY */
7535
7536struct page *alloc_frozen_pages_nolock_noprof(gfp_t gfp_flags, int nid, unsigned int order)
7537{
7538 /*
7539 * Do not specify __GFP_DIRECT_RECLAIM, since direct claim is not allowed.
7540 * Do not specify __GFP_KSWAPD_RECLAIM either, since wake up of kswapd
7541 * is not safe in arbitrary context.
7542 *
7543 * These two are the conditions for gfpflags_allow_spinning() being true.
7544 *
7545 * Specify __GFP_NOWARN since failing alloc_pages_nolock() is not a reason
7546 * to warn. Also warn would trigger printk() which is unsafe from
7547 * various contexts. We cannot use printk_deferred_enter() to mitigate,
7548 * since the running context is unknown.
7549 *
7550 * Specify __GFP_ZERO to make sure that call to kmsan_alloc_page() below
7551 * is safe in any context. Also zeroing the page is mandatory for
7552 * BPF use cases.
7553 *
7554 * Though __GFP_NOMEMALLOC is not checked in the code path below,
7555 * specify it here to highlight that alloc_pages_nolock()
7556 * doesn't want to deplete reserves.
7557 */
7558 gfp_t alloc_gfp = __GFP_NOWARN | __GFP_ZERO | __GFP_NOMEMALLOC | __GFP_COMP
7559 | gfp_flags;
7560 unsigned int alloc_flags = ALLOC_TRYLOCK;
7561 struct alloc_context ac = { };
7562 struct page *page;
7563
7564 VM_WARN_ON_ONCE(gfp_flags & ~__GFP_ACCOUNT);
7565 /*
7566 * In PREEMPT_RT spin_trylock() will call raw_spin_lock() which is
7567 * unsafe in NMI. If spin_trylock() is called from hard IRQ the current
7568 * task may be waiting for one rt_spin_lock, but rt_spin_trylock() will
7569 * mark the task as the owner of another rt_spin_lock which will
7570 * confuse PI logic, so return immediately if called form hard IRQ or
7571 * NMI.
7572 *
7573 * Note, irqs_disabled() case is ok. This function can be called
7574 * from raw_spin_lock_irqsave region.
7575 */
7576 if (IS_ENABLED(CONFIG_PREEMPT_RT) && (in_nmi() || in_hardirq()))
7577 return NULL;
7578 if (!pcp_allowed_order(order))
7579 return NULL;
7580
7581 /* Bailout, since _deferred_grow_zone() needs to take a lock */
7582 if (deferred_pages_enabled())
7583 return NULL;
7584
7585 if (nid == NUMA_NO_NODE)
7586 nid = numa_node_id();
7587
7588 prepare_alloc_pages(gfp_mask: alloc_gfp, order, preferred_nid: nid, NULL, ac: &ac,
7589 alloc_gfp: &alloc_gfp, alloc_flags: &alloc_flags);
7590
7591 /*
7592 * Best effort allocation from percpu free list.
7593 * If it's empty attempt to spin_trylock zone->lock.
7594 */
7595 page = get_page_from_freelist(gfp_mask: alloc_gfp, order, alloc_flags, ac: &ac);
7596
7597 /* Unlike regular alloc_pages() there is no __alloc_pages_slowpath(). */
7598
7599 if (memcg_kmem_online() && page && (gfp_flags & __GFP_ACCOUNT) &&
7600 unlikely(__memcg_kmem_charge_page(page, alloc_gfp, order) != 0)) {
7601 __free_frozen_pages(page, order, FPI_TRYLOCK);
7602 page = NULL;
7603 }
7604 trace_mm_page_alloc(page, order, gfp_flags: alloc_gfp, migratetype: ac.migratetype);
7605 kmsan_alloc_page(page, order, flags: alloc_gfp);
7606 return page;
7607}
7608/**
7609 * alloc_pages_nolock - opportunistic reentrant allocation from any context
7610 * @gfp_flags: GFP flags. Only __GFP_ACCOUNT allowed.
7611 * @nid: node to allocate from
7612 * @order: allocation order size
7613 *
7614 * Allocates pages of a given order from the given node. This is safe to
7615 * call from any context (from atomic, NMI, and also reentrant
7616 * allocator -> tracepoint -> alloc_pages_nolock_noprof).
7617 * Allocation is best effort and to be expected to fail easily so nobody should
7618 * rely on the success. Failures are not reported via warn_alloc().
7619 * See always fail conditions below.
7620 *
7621 * Return: allocated page or NULL on failure. NULL does not mean EBUSY or EAGAIN.
7622 * It means ENOMEM. There is no reason to call it again and expect !NULL.
7623 */
7624struct page *alloc_pages_nolock_noprof(gfp_t gfp_flags, int nid, unsigned int order)
7625{
7626 struct page *page;
7627
7628 page = alloc_frozen_pages_nolock_noprof(gfp_flags, nid, order);
7629 if (page)
7630 set_page_refcounted(page);
7631 return page;
7632}
7633EXPORT_SYMBOL_GPL(alloc_pages_nolock_noprof);
7634