1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/page_alloc.c
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
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)
18 #include <linux/stddef.h>
20 #include <linux/highmem.h>
21 #include <linux/swap.h>
22 #include <linux/swapops.h>
23 #include <linux/interrupt.h>
24 #include <linux/pagemap.h>
25 #include <linux/jiffies.h>
26 #include <linux/memblock.h>
27 #include <linux/compiler.h>
28 #include <linux/kernel.h>
29 #include <linux/kasan.h>
30 #include <linux/kmsan.h>
31 #include <linux/module.h>
32 #include <linux/suspend.h>
33 #include <linux/pagevec.h>
34 #include <linux/blkdev.h>
35 #include <linux/slab.h>
36 #include <linux/ratelimit.h>
37 #include <linux/oom.h>
38 #include <linux/topology.h>
39 #include <linux/sysctl.h>
40 #include <linux/cpu.h>
41 #include <linux/cpuset.h>
42 #include <linux/memory_hotplug.h>
43 #include <linux/nodemask.h>
44 #include <linux/vmalloc.h>
45 #include <linux/vmstat.h>
46 #include <linux/mempolicy.h>
47 #include <linux/memremap.h>
48 #include <linux/stop_machine.h>
49 #include <linux/random.h>
50 #include <linux/sort.h>
51 #include <linux/pfn.h>
52 #include <linux/backing-dev.h>
53 #include <linux/fault-inject.h>
54 #include <linux/page-isolation.h>
55 #include <linux/debugobjects.h>
56 #include <linux/kmemleak.h>
57 #include <linux/compaction.h>
58 #include <trace/events/kmem.h>
59 #include <trace/events/oom.h>
60 #include <linux/prefetch.h>
61 #include <linux/mm_inline.h>
62 #include <linux/mmu_notifier.h>
63 #include <linux/migrate.h>
64 #include <linux/hugetlb.h>
65 #include <linux/sched/rt.h>
66 #include <linux/sched/mm.h>
67 #include <linux/page_owner.h>
68 #include <linux/page_table_check.h>
69 #include <linux/kthread.h>
70 #include <linux/memcontrol.h>
71 #include <linux/ftrace.h>
72 #include <linux/lockdep.h>
73 #include <linux/nmi.h>
74 #include <linux/psi.h>
75 #include <linux/padata.h>
76 #include <linux/khugepaged.h>
77 #include <linux/buffer_head.h>
78 #include <linux/delayacct.h>
79 #include <asm/sections.h>
80 #include <asm/tlbflush.h>
81 #include <asm/div64.h>
84 #include "page_reporting.h"
87 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
88 typedef int __bitwise fpi_t;
90 /* No special request */
91 #define FPI_NONE ((__force fpi_t)0)
94 * Skip free page reporting notification for the (possibly merged) page.
95 * This does not hinder free page reporting from grabbing the page,
96 * reporting it and marking it "reported" - it only skips notifying
97 * the free page reporting infrastructure about a newly freed page. For
98 * example, used when temporarily pulling a page from a freelist and
99 * putting it back unmodified.
101 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
104 * Place the (possibly merged) page to the tail of the freelist. Will ignore
105 * page shuffling (relevant code - e.g., memory onlining - is expected to
106 * shuffle the whole zone).
108 * Note: No code should rely on this flag for correctness - it's purely
109 * to allow for optimizations when handing back either fresh pages
110 * (memory onlining) or untouched pages (page isolation, free page
113 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
116 * Don't poison memory with KASAN (only for the tag-based modes).
117 * During boot, all non-reserved memblock memory is exposed to page_alloc.
118 * Poisoning all that memory lengthens boot time, especially on systems with
119 * large amount of RAM. This flag is used to skip that poisoning.
120 * This is only done for the tag-based KASAN modes, as those are able to
121 * detect memory corruptions with the memory tags assigned by default.
122 * All memory allocated normally after boot gets poisoned as usual.
124 #define FPI_SKIP_KASAN_POISON ((__force fpi_t)BIT(2))
126 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
127 static DEFINE_MUTEX(pcp_batch_high_lock);
128 #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
130 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
132 * On SMP, spin_trylock is sufficient protection.
133 * On PREEMPT_RT, spin_trylock is equivalent on both SMP and UP.
135 #define pcp_trylock_prepare(flags) do { } while (0)
136 #define pcp_trylock_finish(flag) do { } while (0)
139 /* UP spin_trylock always succeeds so disable IRQs to prevent re-entrancy. */
140 #define pcp_trylock_prepare(flags) local_irq_save(flags)
141 #define pcp_trylock_finish(flags) local_irq_restore(flags)
145 * Locking a pcp requires a PCP lookup followed by a spinlock. To avoid
146 * a migration causing the wrong PCP to be locked and remote memory being
147 * potentially allocated, pin the task to the CPU for the lookup+lock.
148 * preempt_disable is used on !RT because it is faster than migrate_disable.
149 * migrate_disable is used on RT because otherwise RT spinlock usage is
150 * interfered with and a high priority task cannot preempt the allocator.
152 #ifndef CONFIG_PREEMPT_RT
153 #define pcpu_task_pin() preempt_disable()
154 #define pcpu_task_unpin() preempt_enable()
156 #define pcpu_task_pin() migrate_disable()
157 #define pcpu_task_unpin() migrate_enable()
161 * Generic helper to lookup and a per-cpu variable with an embedded spinlock.
162 * Return value should be used with equivalent unlock helper.
164 #define pcpu_spin_lock(type, member, ptr) \
168 _ret = this_cpu_ptr(ptr); \
169 spin_lock(&_ret->member); \
173 #define pcpu_spin_trylock(type, member, ptr) \
177 _ret = this_cpu_ptr(ptr); \
178 if (!spin_trylock(&_ret->member)) { \
185 #define pcpu_spin_unlock(member, ptr) \
187 spin_unlock(&ptr->member); \
191 /* struct per_cpu_pages specific helpers. */
192 #define pcp_spin_lock(ptr) \
193 pcpu_spin_lock(struct per_cpu_pages, lock, ptr)
195 #define pcp_spin_trylock(ptr) \
196 pcpu_spin_trylock(struct per_cpu_pages, lock, ptr)
198 #define pcp_spin_unlock(ptr) \
199 pcpu_spin_unlock(lock, ptr)
201 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
202 DEFINE_PER_CPU(int, numa_node);
203 EXPORT_PER_CPU_SYMBOL(numa_node);
206 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
208 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
210 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
211 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
212 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
213 * defined in <linux/topology.h>.
215 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
216 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
219 static DEFINE_MUTEX(pcpu_drain_mutex);
221 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
222 volatile unsigned long latent_entropy __latent_entropy;
223 EXPORT_SYMBOL(latent_entropy);
227 * Array of node states.
229 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
230 [N_POSSIBLE] = NODE_MASK_ALL,
231 [N_ONLINE] = { { [0] = 1UL } },
233 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
234 #ifdef CONFIG_HIGHMEM
235 [N_HIGH_MEMORY] = { { [0] = 1UL } },
237 [N_MEMORY] = { { [0] = 1UL } },
238 [N_CPU] = { { [0] = 1UL } },
241 EXPORT_SYMBOL(node_states);
243 atomic_long_t _totalram_pages __read_mostly;
244 EXPORT_SYMBOL(_totalram_pages);
245 unsigned long totalreserve_pages __read_mostly;
246 unsigned long totalcma_pages __read_mostly;
248 int percpu_pagelist_high_fraction;
249 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
250 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_ALLOC_DEFAULT_ON, init_on_alloc);
251 EXPORT_SYMBOL(init_on_alloc);
253 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_FREE_DEFAULT_ON, init_on_free);
254 EXPORT_SYMBOL(init_on_free);
256 static bool _init_on_alloc_enabled_early __read_mostly
257 = IS_ENABLED(CONFIG_INIT_ON_ALLOC_DEFAULT_ON);
258 static int __init early_init_on_alloc(char *buf)
261 return kstrtobool(buf, &_init_on_alloc_enabled_early);
263 early_param("init_on_alloc", early_init_on_alloc);
265 static bool _init_on_free_enabled_early __read_mostly
266 = IS_ENABLED(CONFIG_INIT_ON_FREE_DEFAULT_ON);
267 static int __init early_init_on_free(char *buf)
269 return kstrtobool(buf, &_init_on_free_enabled_early);
271 early_param("init_on_free", early_init_on_free);
274 * A cached value of the page's pageblock's migratetype, used when the page is
275 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
276 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
277 * Also the migratetype set in the page does not necessarily match the pcplist
278 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
279 * other index - this ensures that it will be put on the correct CMA freelist.
281 static inline int get_pcppage_migratetype(struct page *page)
286 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
288 page->index = migratetype;
291 #ifdef CONFIG_PM_SLEEP
293 * The following functions are used by the suspend/hibernate code to temporarily
294 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
295 * while devices are suspended. To avoid races with the suspend/hibernate code,
296 * they should always be called with system_transition_mutex held
297 * (gfp_allowed_mask also should only be modified with system_transition_mutex
298 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
299 * with that modification).
302 static gfp_t saved_gfp_mask;
304 void pm_restore_gfp_mask(void)
306 WARN_ON(!mutex_is_locked(&system_transition_mutex));
307 if (saved_gfp_mask) {
308 gfp_allowed_mask = saved_gfp_mask;
313 void pm_restrict_gfp_mask(void)
315 WARN_ON(!mutex_is_locked(&system_transition_mutex));
316 WARN_ON(saved_gfp_mask);
317 saved_gfp_mask = gfp_allowed_mask;
318 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
321 bool pm_suspended_storage(void)
323 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
327 #endif /* CONFIG_PM_SLEEP */
329 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
330 unsigned int pageblock_order __read_mostly;
333 static void __free_pages_ok(struct page *page, unsigned int order,
337 * results with 256, 32 in the lowmem_reserve sysctl:
338 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
339 * 1G machine -> (16M dma, 784M normal, 224M high)
340 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
341 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
342 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
344 * TBD: should special case ZONE_DMA32 machines here - in those we normally
345 * don't need any ZONE_NORMAL reservation
347 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
348 #ifdef CONFIG_ZONE_DMA
351 #ifdef CONFIG_ZONE_DMA32
355 #ifdef CONFIG_HIGHMEM
361 static char * const zone_names[MAX_NR_ZONES] = {
362 #ifdef CONFIG_ZONE_DMA
365 #ifdef CONFIG_ZONE_DMA32
369 #ifdef CONFIG_HIGHMEM
373 #ifdef CONFIG_ZONE_DEVICE
378 const char * const migratetype_names[MIGRATE_TYPES] = {
386 #ifdef CONFIG_MEMORY_ISOLATION
391 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
392 [NULL_COMPOUND_DTOR] = NULL,
393 [COMPOUND_PAGE_DTOR] = free_compound_page,
394 #ifdef CONFIG_HUGETLB_PAGE
395 [HUGETLB_PAGE_DTOR] = free_huge_page,
397 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
398 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
402 int min_free_kbytes = 1024;
403 int user_min_free_kbytes = -1;
404 int watermark_boost_factor __read_mostly = 15000;
405 int watermark_scale_factor = 10;
407 static unsigned long nr_kernel_pages __initdata;
408 static unsigned long nr_all_pages __initdata;
409 static unsigned long dma_reserve __initdata;
411 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
412 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
413 static unsigned long required_kernelcore __initdata;
414 static unsigned long required_kernelcore_percent __initdata;
415 static unsigned long required_movablecore __initdata;
416 static unsigned long required_movablecore_percent __initdata;
417 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
418 bool mirrored_kernelcore __initdata_memblock;
420 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
422 EXPORT_SYMBOL(movable_zone);
425 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
426 unsigned int nr_online_nodes __read_mostly = 1;
427 EXPORT_SYMBOL(nr_node_ids);
428 EXPORT_SYMBOL(nr_online_nodes);
431 int page_group_by_mobility_disabled __read_mostly;
433 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
435 * During boot we initialize deferred pages on-demand, as needed, but once
436 * page_alloc_init_late() has finished, the deferred pages are all initialized,
437 * and we can permanently disable that path.
439 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
441 static inline bool deferred_pages_enabled(void)
443 return static_branch_unlikely(&deferred_pages);
446 /* Returns true if the struct page for the pfn is uninitialised */
447 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
449 int nid = early_pfn_to_nid(pfn);
451 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
458 * Returns true when the remaining initialisation should be deferred until
459 * later in the boot cycle when it can be parallelised.
461 static bool __meminit
462 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
464 static unsigned long prev_end_pfn, nr_initialised;
466 if (early_page_ext_enabled())
469 * prev_end_pfn static that contains the end of previous zone
470 * No need to protect because called very early in boot before smp_init.
472 if (prev_end_pfn != end_pfn) {
473 prev_end_pfn = end_pfn;
477 /* Always populate low zones for address-constrained allocations */
478 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
481 if (NODE_DATA(nid)->first_deferred_pfn != ULONG_MAX)
484 * We start only with one section of pages, more pages are added as
485 * needed until the rest of deferred pages are initialized.
488 if ((nr_initialised > PAGES_PER_SECTION) &&
489 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
490 NODE_DATA(nid)->first_deferred_pfn = pfn;
496 static inline bool deferred_pages_enabled(void)
501 static inline bool early_page_uninitialised(unsigned long pfn)
506 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
512 /* Return a pointer to the bitmap storing bits affecting a block of pages */
513 static inline unsigned long *get_pageblock_bitmap(const struct page *page,
516 #ifdef CONFIG_SPARSEMEM
517 return section_to_usemap(__pfn_to_section(pfn));
519 return page_zone(page)->pageblock_flags;
520 #endif /* CONFIG_SPARSEMEM */
523 static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
525 #ifdef CONFIG_SPARSEMEM
526 pfn &= (PAGES_PER_SECTION-1);
528 pfn = pfn - pageblock_start_pfn(page_zone(page)->zone_start_pfn);
529 #endif /* CONFIG_SPARSEMEM */
530 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
533 static __always_inline
534 unsigned long __get_pfnblock_flags_mask(const struct page *page,
538 unsigned long *bitmap;
539 unsigned long bitidx, word_bitidx;
542 bitmap = get_pageblock_bitmap(page, pfn);
543 bitidx = pfn_to_bitidx(page, pfn);
544 word_bitidx = bitidx / BITS_PER_LONG;
545 bitidx &= (BITS_PER_LONG-1);
547 * This races, without locks, with set_pfnblock_flags_mask(). Ensure
548 * a consistent read of the memory array, so that results, even though
549 * racy, are not corrupted.
551 word = READ_ONCE(bitmap[word_bitidx]);
552 return (word >> bitidx) & mask;
556 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
557 * @page: The page within the block of interest
558 * @pfn: The target page frame number
559 * @mask: mask of bits that the caller is interested in
561 * Return: pageblock_bits flags
563 unsigned long get_pfnblock_flags_mask(const struct page *page,
564 unsigned long pfn, unsigned long mask)
566 return __get_pfnblock_flags_mask(page, pfn, mask);
569 static __always_inline int get_pfnblock_migratetype(const struct page *page,
572 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
576 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
577 * @page: The page within the block of interest
578 * @flags: The flags to set
579 * @pfn: The target page frame number
580 * @mask: mask of bits that the caller is interested in
582 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
586 unsigned long *bitmap;
587 unsigned long bitidx, word_bitidx;
590 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
591 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
593 bitmap = get_pageblock_bitmap(page, pfn);
594 bitidx = pfn_to_bitidx(page, pfn);
595 word_bitidx = bitidx / BITS_PER_LONG;
596 bitidx &= (BITS_PER_LONG-1);
598 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
603 word = READ_ONCE(bitmap[word_bitidx]);
605 } while (!try_cmpxchg(&bitmap[word_bitidx], &word, (word & ~mask) | flags));
608 void set_pageblock_migratetype(struct page *page, int migratetype)
610 if (unlikely(page_group_by_mobility_disabled &&
611 migratetype < MIGRATE_PCPTYPES))
612 migratetype = MIGRATE_UNMOVABLE;
614 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
615 page_to_pfn(page), MIGRATETYPE_MASK);
618 #ifdef CONFIG_DEBUG_VM
619 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
623 unsigned long pfn = page_to_pfn(page);
624 unsigned long sp, start_pfn;
627 seq = zone_span_seqbegin(zone);
628 start_pfn = zone->zone_start_pfn;
629 sp = zone->spanned_pages;
630 if (!zone_spans_pfn(zone, pfn))
632 } while (zone_span_seqretry(zone, seq));
635 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
636 pfn, zone_to_nid(zone), zone->name,
637 start_pfn, start_pfn + sp);
642 static int page_is_consistent(struct zone *zone, struct page *page)
644 if (zone != page_zone(page))
650 * Temporary debugging check for pages not lying within a given zone.
652 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
654 if (page_outside_zone_boundaries(zone, page))
656 if (!page_is_consistent(zone, page))
662 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
668 static void bad_page(struct page *page, const char *reason)
670 static unsigned long resume;
671 static unsigned long nr_shown;
672 static unsigned long nr_unshown;
675 * Allow a burst of 60 reports, then keep quiet for that minute;
676 * or allow a steady drip of one report per second.
678 if (nr_shown == 60) {
679 if (time_before(jiffies, resume)) {
685 "BUG: Bad page state: %lu messages suppressed\n",
692 resume = jiffies + 60 * HZ;
694 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
695 current->comm, page_to_pfn(page));
696 dump_page(page, reason);
701 /* Leave bad fields for debug, except PageBuddy could make trouble */
702 page_mapcount_reset(page); /* remove PageBuddy */
703 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
706 static inline unsigned int order_to_pindex(int migratetype, int order)
710 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
711 if (order > PAGE_ALLOC_COSTLY_ORDER) {
712 VM_BUG_ON(order != pageblock_order);
713 return NR_LOWORDER_PCP_LISTS;
716 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
719 return (MIGRATE_PCPTYPES * base) + migratetype;
722 static inline int pindex_to_order(unsigned int pindex)
724 int order = pindex / MIGRATE_PCPTYPES;
726 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
727 if (pindex == NR_LOWORDER_PCP_LISTS)
728 order = pageblock_order;
730 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
736 static inline bool pcp_allowed_order(unsigned int order)
738 if (order <= PAGE_ALLOC_COSTLY_ORDER)
740 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
741 if (order == pageblock_order)
747 static inline void free_the_page(struct page *page, unsigned int order)
749 if (pcp_allowed_order(order)) /* Via pcp? */
750 free_unref_page(page, order);
752 __free_pages_ok(page, order, FPI_NONE);
756 * Higher-order pages are called "compound pages". They are structured thusly:
758 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
760 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
761 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
763 * The first tail page's ->compound_dtor holds the offset in array of compound
764 * page destructors. See compound_page_dtors.
766 * The first tail page's ->compound_order holds the order of allocation.
767 * This usage means that zero-order pages may not be compound.
770 void free_compound_page(struct page *page)
772 mem_cgroup_uncharge(page_folio(page));
773 free_the_page(page, compound_order(page));
776 static void prep_compound_head(struct page *page, unsigned int order)
778 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
779 set_compound_order(page, order);
780 atomic_set(compound_mapcount_ptr(page), -1);
781 atomic_set(compound_pincount_ptr(page), 0);
784 static void prep_compound_tail(struct page *head, int tail_idx)
786 struct page *p = head + tail_idx;
788 p->mapping = TAIL_MAPPING;
789 set_compound_head(p, head);
790 set_page_private(p, 0);
793 void prep_compound_page(struct page *page, unsigned int order)
796 int nr_pages = 1 << order;
799 for (i = 1; i < nr_pages; i++)
800 prep_compound_tail(page, i);
802 prep_compound_head(page, order);
805 void destroy_large_folio(struct folio *folio)
807 enum compound_dtor_id dtor = folio_page(folio, 1)->compound_dtor;
809 VM_BUG_ON_FOLIO(dtor >= NR_COMPOUND_DTORS, folio);
810 compound_page_dtors[dtor](&folio->page);
813 #ifdef CONFIG_DEBUG_PAGEALLOC
814 unsigned int _debug_guardpage_minorder;
816 bool _debug_pagealloc_enabled_early __read_mostly
817 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
818 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
819 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
820 EXPORT_SYMBOL(_debug_pagealloc_enabled);
822 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
824 static int __init early_debug_pagealloc(char *buf)
826 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
828 early_param("debug_pagealloc", early_debug_pagealloc);
830 static int __init debug_guardpage_minorder_setup(char *buf)
834 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
835 pr_err("Bad debug_guardpage_minorder value\n");
838 _debug_guardpage_minorder = res;
839 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
842 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
844 static inline bool set_page_guard(struct zone *zone, struct page *page,
845 unsigned int order, int migratetype)
847 if (!debug_guardpage_enabled())
850 if (order >= debug_guardpage_minorder())
853 __SetPageGuard(page);
854 INIT_LIST_HEAD(&page->buddy_list);
855 set_page_private(page, order);
856 /* Guard pages are not available for any usage */
857 if (!is_migrate_isolate(migratetype))
858 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
863 static inline void clear_page_guard(struct zone *zone, struct page *page,
864 unsigned int order, int migratetype)
866 if (!debug_guardpage_enabled())
869 __ClearPageGuard(page);
871 set_page_private(page, 0);
872 if (!is_migrate_isolate(migratetype))
873 __mod_zone_freepage_state(zone, (1 << order), migratetype);
876 static inline bool set_page_guard(struct zone *zone, struct page *page,
877 unsigned int order, int migratetype) { return false; }
878 static inline void clear_page_guard(struct zone *zone, struct page *page,
879 unsigned int order, int migratetype) {}
883 * Enable static keys related to various memory debugging and hardening options.
884 * Some override others, and depend on early params that are evaluated in the
885 * order of appearance. So we need to first gather the full picture of what was
886 * enabled, and then make decisions.
888 void __init init_mem_debugging_and_hardening(void)
890 bool page_poisoning_requested = false;
892 #ifdef CONFIG_PAGE_POISONING
894 * Page poisoning is debug page alloc for some arches. If
895 * either of those options are enabled, enable poisoning.
897 if (page_poisoning_enabled() ||
898 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
899 debug_pagealloc_enabled())) {
900 static_branch_enable(&_page_poisoning_enabled);
901 page_poisoning_requested = true;
905 if ((_init_on_alloc_enabled_early || _init_on_free_enabled_early) &&
906 page_poisoning_requested) {
907 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
908 "will take precedence over init_on_alloc and init_on_free\n");
909 _init_on_alloc_enabled_early = false;
910 _init_on_free_enabled_early = false;
913 if (_init_on_alloc_enabled_early)
914 static_branch_enable(&init_on_alloc);
916 static_branch_disable(&init_on_alloc);
918 if (_init_on_free_enabled_early)
919 static_branch_enable(&init_on_free);
921 static_branch_disable(&init_on_free);
923 if (IS_ENABLED(CONFIG_KMSAN) &&
924 (_init_on_alloc_enabled_early || _init_on_free_enabled_early))
925 pr_info("mem auto-init: please make sure init_on_alloc and init_on_free are disabled when running KMSAN\n");
927 #ifdef CONFIG_DEBUG_PAGEALLOC
928 if (!debug_pagealloc_enabled())
931 static_branch_enable(&_debug_pagealloc_enabled);
933 if (!debug_guardpage_minorder())
936 static_branch_enable(&_debug_guardpage_enabled);
940 static inline void set_buddy_order(struct page *page, unsigned int order)
942 set_page_private(page, order);
943 __SetPageBuddy(page);
946 #ifdef CONFIG_COMPACTION
947 static inline struct capture_control *task_capc(struct zone *zone)
949 struct capture_control *capc = current->capture_control;
951 return unlikely(capc) &&
952 !(current->flags & PF_KTHREAD) &&
954 capc->cc->zone == zone ? capc : NULL;
958 compaction_capture(struct capture_control *capc, struct page *page,
959 int order, int migratetype)
961 if (!capc || order != capc->cc->order)
964 /* Do not accidentally pollute CMA or isolated regions*/
965 if (is_migrate_cma(migratetype) ||
966 is_migrate_isolate(migratetype))
970 * Do not let lower order allocations pollute a movable pageblock.
971 * This might let an unmovable request use a reclaimable pageblock
972 * and vice-versa but no more than normal fallback logic which can
973 * have trouble finding a high-order free page.
975 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
983 static inline struct capture_control *task_capc(struct zone *zone)
989 compaction_capture(struct capture_control *capc, struct page *page,
990 int order, int migratetype)
994 #endif /* CONFIG_COMPACTION */
996 /* Used for pages not on another list */
997 static inline void add_to_free_list(struct page *page, struct zone *zone,
998 unsigned int order, int migratetype)
1000 struct free_area *area = &zone->free_area[order];
1002 list_add(&page->buddy_list, &area->free_list[migratetype]);
1006 /* Used for pages not on another list */
1007 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
1008 unsigned int order, int migratetype)
1010 struct free_area *area = &zone->free_area[order];
1012 list_add_tail(&page->buddy_list, &area->free_list[migratetype]);
1017 * Used for pages which are on another list. Move the pages to the tail
1018 * of the list - so the moved pages won't immediately be considered for
1019 * allocation again (e.g., optimization for memory onlining).
1021 static inline void move_to_free_list(struct page *page, struct zone *zone,
1022 unsigned int order, int migratetype)
1024 struct free_area *area = &zone->free_area[order];
1026 list_move_tail(&page->buddy_list, &area->free_list[migratetype]);
1029 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
1032 /* clear reported state and update reported page count */
1033 if (page_reported(page))
1034 __ClearPageReported(page);
1036 list_del(&page->buddy_list);
1037 __ClearPageBuddy(page);
1038 set_page_private(page, 0);
1039 zone->free_area[order].nr_free--;
1043 * If this is not the largest possible page, check if the buddy
1044 * of the next-highest order is free. If it is, it's possible
1045 * that pages are being freed that will coalesce soon. In case,
1046 * that is happening, add the free page to the tail of the list
1047 * so it's less likely to be used soon and more likely to be merged
1048 * as a higher order page
1051 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
1052 struct page *page, unsigned int order)
1054 unsigned long higher_page_pfn;
1055 struct page *higher_page;
1057 if (order >= MAX_ORDER - 2)
1060 higher_page_pfn = buddy_pfn & pfn;
1061 higher_page = page + (higher_page_pfn - pfn);
1063 return find_buddy_page_pfn(higher_page, higher_page_pfn, order + 1,
1068 * Freeing function for a buddy system allocator.
1070 * The concept of a buddy system is to maintain direct-mapped table
1071 * (containing bit values) for memory blocks of various "orders".
1072 * The bottom level table contains the map for the smallest allocatable
1073 * units of memory (here, pages), and each level above it describes
1074 * pairs of units from the levels below, hence, "buddies".
1075 * At a high level, all that happens here is marking the table entry
1076 * at the bottom level available, and propagating the changes upward
1077 * as necessary, plus some accounting needed to play nicely with other
1078 * parts of the VM system.
1079 * At each level, we keep a list of pages, which are heads of continuous
1080 * free pages of length of (1 << order) and marked with PageBuddy.
1081 * Page's order is recorded in page_private(page) field.
1082 * So when we are allocating or freeing one, we can derive the state of the
1083 * other. That is, if we allocate a small block, and both were
1084 * free, the remainder of the region must be split into blocks.
1085 * If a block is freed, and its buddy is also free, then this
1086 * triggers coalescing into a block of larger size.
1091 static inline void __free_one_page(struct page *page,
1093 struct zone *zone, unsigned int order,
1094 int migratetype, fpi_t fpi_flags)
1096 struct capture_control *capc = task_capc(zone);
1097 unsigned long buddy_pfn = 0;
1098 unsigned long combined_pfn;
1102 VM_BUG_ON(!zone_is_initialized(zone));
1103 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1105 VM_BUG_ON(migratetype == -1);
1106 if (likely(!is_migrate_isolate(migratetype)))
1107 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1109 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1110 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1112 while (order < MAX_ORDER - 1) {
1113 if (compaction_capture(capc, page, order, migratetype)) {
1114 __mod_zone_freepage_state(zone, -(1 << order),
1119 buddy = find_buddy_page_pfn(page, pfn, order, &buddy_pfn);
1123 if (unlikely(order >= pageblock_order)) {
1125 * We want to prevent merge between freepages on pageblock
1126 * without fallbacks and normal pageblock. Without this,
1127 * pageblock isolation could cause incorrect freepage or CMA
1128 * accounting or HIGHATOMIC accounting.
1130 int buddy_mt = get_pageblock_migratetype(buddy);
1132 if (migratetype != buddy_mt
1133 && (!migratetype_is_mergeable(migratetype) ||
1134 !migratetype_is_mergeable(buddy_mt)))
1139 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1140 * merge with it and move up one order.
1142 if (page_is_guard(buddy))
1143 clear_page_guard(zone, buddy, order, migratetype);
1145 del_page_from_free_list(buddy, zone, order);
1146 combined_pfn = buddy_pfn & pfn;
1147 page = page + (combined_pfn - pfn);
1153 set_buddy_order(page, order);
1155 if (fpi_flags & FPI_TO_TAIL)
1157 else if (is_shuffle_order(order))
1158 to_tail = shuffle_pick_tail();
1160 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1163 add_to_free_list_tail(page, zone, order, migratetype);
1165 add_to_free_list(page, zone, order, migratetype);
1167 /* Notify page reporting subsystem of freed page */
1168 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1169 page_reporting_notify_free(order);
1173 * split_free_page() -- split a free page at split_pfn_offset
1174 * @free_page: the original free page
1175 * @order: the order of the page
1176 * @split_pfn_offset: split offset within the page
1178 * Return -ENOENT if the free page is changed, otherwise 0
1180 * It is used when the free page crosses two pageblocks with different migratetypes
1181 * at split_pfn_offset within the page. The split free page will be put into
1182 * separate migratetype lists afterwards. Otherwise, the function achieves
1185 int split_free_page(struct page *free_page,
1186 unsigned int order, unsigned long split_pfn_offset)
1188 struct zone *zone = page_zone(free_page);
1189 unsigned long free_page_pfn = page_to_pfn(free_page);
1191 unsigned long flags;
1192 int free_page_order;
1196 if (split_pfn_offset == 0)
1199 spin_lock_irqsave(&zone->lock, flags);
1201 if (!PageBuddy(free_page) || buddy_order(free_page) != order) {
1206 mt = get_pageblock_migratetype(free_page);
1207 if (likely(!is_migrate_isolate(mt)))
1208 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1210 del_page_from_free_list(free_page, zone, order);
1211 for (pfn = free_page_pfn;
1212 pfn < free_page_pfn + (1UL << order);) {
1213 int mt = get_pfnblock_migratetype(pfn_to_page(pfn), pfn);
1215 free_page_order = min_t(unsigned int,
1216 pfn ? __ffs(pfn) : order,
1217 __fls(split_pfn_offset));
1218 __free_one_page(pfn_to_page(pfn), pfn, zone, free_page_order,
1220 pfn += 1UL << free_page_order;
1221 split_pfn_offset -= (1UL << free_page_order);
1222 /* we have done the first part, now switch to second part */
1223 if (split_pfn_offset == 0)
1224 split_pfn_offset = (1UL << order) - (pfn - free_page_pfn);
1227 spin_unlock_irqrestore(&zone->lock, flags);
1231 * A bad page could be due to a number of fields. Instead of multiple branches,
1232 * try and check multiple fields with one check. The caller must do a detailed
1233 * check if necessary.
1235 static inline bool page_expected_state(struct page *page,
1236 unsigned long check_flags)
1238 if (unlikely(atomic_read(&page->_mapcount) != -1))
1241 if (unlikely((unsigned long)page->mapping |
1242 page_ref_count(page) |
1246 (page->flags & check_flags)))
1252 static const char *page_bad_reason(struct page *page, unsigned long flags)
1254 const char *bad_reason = NULL;
1256 if (unlikely(atomic_read(&page->_mapcount) != -1))
1257 bad_reason = "nonzero mapcount";
1258 if (unlikely(page->mapping != NULL))
1259 bad_reason = "non-NULL mapping";
1260 if (unlikely(page_ref_count(page) != 0))
1261 bad_reason = "nonzero _refcount";
1262 if (unlikely(page->flags & flags)) {
1263 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1264 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1266 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1269 if (unlikely(page->memcg_data))
1270 bad_reason = "page still charged to cgroup";
1275 static void free_page_is_bad_report(struct page *page)
1278 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1281 static inline bool free_page_is_bad(struct page *page)
1283 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1286 /* Something has gone sideways, find it */
1287 free_page_is_bad_report(page);
1291 static int free_tail_pages_check(struct page *head_page, struct page *page)
1296 * We rely page->lru.next never has bit 0 set, unless the page
1297 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1299 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1301 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1305 switch (page - head_page) {
1307 /* the first tail page: ->mapping may be compound_mapcount() */
1308 if (unlikely(compound_mapcount(page))) {
1309 bad_page(page, "nonzero compound_mapcount");
1315 * the second tail page: ->mapping is
1316 * deferred_list.next -- ignore value.
1320 if (page->mapping != TAIL_MAPPING) {
1321 bad_page(page, "corrupted mapping in tail page");
1326 if (unlikely(!PageTail(page))) {
1327 bad_page(page, "PageTail not set");
1330 if (unlikely(compound_head(page) != head_page)) {
1331 bad_page(page, "compound_head not consistent");
1336 page->mapping = NULL;
1337 clear_compound_head(page);
1342 * Skip KASAN memory poisoning when either:
1344 * 1. Deferred memory initialization has not yet completed,
1345 * see the explanation below.
1346 * 2. Skipping poisoning is requested via FPI_SKIP_KASAN_POISON,
1347 * see the comment next to it.
1348 * 3. Skipping poisoning is requested via __GFP_SKIP_KASAN_POISON,
1349 * see the comment next to it.
1351 * Poisoning pages during deferred memory init will greatly lengthen the
1352 * process and cause problem in large memory systems as the deferred pages
1353 * initialization is done with interrupt disabled.
1355 * Assuming that there will be no reference to those newly initialized
1356 * pages before they are ever allocated, this should have no effect on
1357 * KASAN memory tracking as the poison will be properly inserted at page
1358 * allocation time. The only corner case is when pages are allocated by
1359 * on-demand allocation and then freed again before the deferred pages
1360 * initialization is done, but this is not likely to happen.
1362 static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
1364 return deferred_pages_enabled() ||
1365 (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
1366 (fpi_flags & FPI_SKIP_KASAN_POISON)) ||
1367 PageSkipKASanPoison(page);
1370 static void kernel_init_pages(struct page *page, int numpages)
1374 /* s390's use of memset() could override KASAN redzones. */
1375 kasan_disable_current();
1376 for (i = 0; i < numpages; i++)
1377 clear_highpage_kasan_tagged(page + i);
1378 kasan_enable_current();
1381 static __always_inline bool free_pages_prepare(struct page *page,
1382 unsigned int order, bool check_free, fpi_t fpi_flags)
1385 bool skip_kasan_poison = should_skip_kasan_poison(page, fpi_flags);
1386 bool init = want_init_on_free();
1388 VM_BUG_ON_PAGE(PageTail(page), page);
1390 trace_mm_page_free(page, order);
1391 kmsan_free_page(page, order);
1393 if (unlikely(PageHWPoison(page)) && !order) {
1395 * Do not let hwpoison pages hit pcplists/buddy
1396 * Untie memcg state and reset page's owner
1398 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1399 __memcg_kmem_uncharge_page(page, order);
1400 reset_page_owner(page, order);
1401 page_table_check_free(page, order);
1406 * Check tail pages before head page information is cleared to
1407 * avoid checking PageCompound for order-0 pages.
1409 if (unlikely(order)) {
1410 bool compound = PageCompound(page);
1413 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1416 ClearPageDoubleMap(page);
1417 ClearPageHasHWPoisoned(page);
1419 for (i = 1; i < (1 << order); i++) {
1421 bad += free_tail_pages_check(page, page + i);
1422 if (unlikely(free_page_is_bad(page + i))) {
1426 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1429 if (PageMappingFlags(page))
1430 page->mapping = NULL;
1431 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1432 __memcg_kmem_uncharge_page(page, order);
1433 if (check_free && free_page_is_bad(page))
1438 page_cpupid_reset_last(page);
1439 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1440 reset_page_owner(page, order);
1441 page_table_check_free(page, order);
1443 if (!PageHighMem(page)) {
1444 debug_check_no_locks_freed(page_address(page),
1445 PAGE_SIZE << order);
1446 debug_check_no_obj_freed(page_address(page),
1447 PAGE_SIZE << order);
1450 kernel_poison_pages(page, 1 << order);
1453 * As memory initialization might be integrated into KASAN,
1454 * KASAN poisoning and memory initialization code must be
1455 * kept together to avoid discrepancies in behavior.
1457 * With hardware tag-based KASAN, memory tags must be set before the
1458 * page becomes unavailable via debug_pagealloc or arch_free_page.
1460 if (!skip_kasan_poison) {
1461 kasan_poison_pages(page, order, init);
1463 /* Memory is already initialized if KASAN did it internally. */
1464 if (kasan_has_integrated_init())
1468 kernel_init_pages(page, 1 << order);
1471 * arch_free_page() can make the page's contents inaccessible. s390
1472 * does this. So nothing which can access the page's contents should
1473 * happen after this.
1475 arch_free_page(page, order);
1477 debug_pagealloc_unmap_pages(page, 1 << order);
1482 #ifdef CONFIG_DEBUG_VM
1484 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1485 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1486 * moved from pcp lists to free lists.
1488 static bool free_pcp_prepare(struct page *page, unsigned int order)
1490 return free_pages_prepare(page, order, true, FPI_NONE);
1493 /* return true if this page has an inappropriate state */
1494 static bool bulkfree_pcp_prepare(struct page *page)
1496 if (debug_pagealloc_enabled_static())
1497 return free_page_is_bad(page);
1503 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1504 * moving from pcp lists to free list in order to reduce overhead. With
1505 * debug_pagealloc enabled, they are checked also immediately when being freed
1508 static bool free_pcp_prepare(struct page *page, unsigned int order)
1510 if (debug_pagealloc_enabled_static())
1511 return free_pages_prepare(page, order, true, FPI_NONE);
1513 return free_pages_prepare(page, order, false, FPI_NONE);
1516 static bool bulkfree_pcp_prepare(struct page *page)
1518 return free_page_is_bad(page);
1520 #endif /* CONFIG_DEBUG_VM */
1523 * Frees a number of pages from the PCP lists
1524 * Assumes all pages on list are in same zone.
1525 * count is the number of pages to free.
1527 static void free_pcppages_bulk(struct zone *zone, int count,
1528 struct per_cpu_pages *pcp,
1531 unsigned long flags;
1533 int max_pindex = NR_PCP_LISTS - 1;
1535 bool isolated_pageblocks;
1539 * Ensure proper count is passed which otherwise would stuck in the
1540 * below while (list_empty(list)) loop.
1542 count = min(pcp->count, count);
1544 /* Ensure requested pindex is drained first. */
1545 pindex = pindex - 1;
1547 spin_lock_irqsave(&zone->lock, flags);
1548 isolated_pageblocks = has_isolate_pageblock(zone);
1551 struct list_head *list;
1554 /* Remove pages from lists in a round-robin fashion. */
1556 if (++pindex > max_pindex)
1557 pindex = min_pindex;
1558 list = &pcp->lists[pindex];
1559 if (!list_empty(list))
1562 if (pindex == max_pindex)
1564 if (pindex == min_pindex)
1568 order = pindex_to_order(pindex);
1569 nr_pages = 1 << order;
1573 page = list_last_entry(list, struct page, pcp_list);
1574 mt = get_pcppage_migratetype(page);
1576 /* must delete to avoid corrupting pcp list */
1577 list_del(&page->pcp_list);
1579 pcp->count -= nr_pages;
1581 if (bulkfree_pcp_prepare(page))
1584 /* MIGRATE_ISOLATE page should not go to pcplists */
1585 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1586 /* Pageblock could have been isolated meanwhile */
1587 if (unlikely(isolated_pageblocks))
1588 mt = get_pageblock_migratetype(page);
1590 __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1591 trace_mm_page_pcpu_drain(page, order, mt);
1592 } while (count > 0 && !list_empty(list));
1595 spin_unlock_irqrestore(&zone->lock, flags);
1598 static void free_one_page(struct zone *zone,
1599 struct page *page, unsigned long pfn,
1601 int migratetype, fpi_t fpi_flags)
1603 unsigned long flags;
1605 spin_lock_irqsave(&zone->lock, flags);
1606 if (unlikely(has_isolate_pageblock(zone) ||
1607 is_migrate_isolate(migratetype))) {
1608 migratetype = get_pfnblock_migratetype(page, pfn);
1610 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1611 spin_unlock_irqrestore(&zone->lock, flags);
1614 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1615 unsigned long zone, int nid)
1617 mm_zero_struct_page(page);
1618 set_page_links(page, zone, nid, pfn);
1619 init_page_count(page);
1620 page_mapcount_reset(page);
1621 page_cpupid_reset_last(page);
1622 page_kasan_tag_reset(page);
1624 INIT_LIST_HEAD(&page->lru);
1625 #ifdef WANT_PAGE_VIRTUAL
1626 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1627 if (!is_highmem_idx(zone))
1628 set_page_address(page, __va(pfn << PAGE_SHIFT));
1632 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1633 static void __meminit init_reserved_page(unsigned long pfn)
1638 if (!early_page_uninitialised(pfn))
1641 nid = early_pfn_to_nid(pfn);
1642 pgdat = NODE_DATA(nid);
1644 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1645 struct zone *zone = &pgdat->node_zones[zid];
1647 if (zone_spans_pfn(zone, pfn))
1650 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1653 static inline void init_reserved_page(unsigned long pfn)
1656 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1659 * Initialised pages do not have PageReserved set. This function is
1660 * called for each range allocated by the bootmem allocator and
1661 * marks the pages PageReserved. The remaining valid pages are later
1662 * sent to the buddy page allocator.
1664 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1666 unsigned long start_pfn = PFN_DOWN(start);
1667 unsigned long end_pfn = PFN_UP(end);
1669 for (; start_pfn < end_pfn; start_pfn++) {
1670 if (pfn_valid(start_pfn)) {
1671 struct page *page = pfn_to_page(start_pfn);
1673 init_reserved_page(start_pfn);
1675 /* Avoid false-positive PageTail() */
1676 INIT_LIST_HEAD(&page->lru);
1679 * no need for atomic set_bit because the struct
1680 * page is not visible yet so nobody should
1683 __SetPageReserved(page);
1688 static void __free_pages_ok(struct page *page, unsigned int order,
1691 unsigned long flags;
1693 unsigned long pfn = page_to_pfn(page);
1694 struct zone *zone = page_zone(page);
1696 if (!free_pages_prepare(page, order, true, fpi_flags))
1699 migratetype = get_pfnblock_migratetype(page, pfn);
1701 spin_lock_irqsave(&zone->lock, flags);
1702 if (unlikely(has_isolate_pageblock(zone) ||
1703 is_migrate_isolate(migratetype))) {
1704 migratetype = get_pfnblock_migratetype(page, pfn);
1706 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1707 spin_unlock_irqrestore(&zone->lock, flags);
1709 __count_vm_events(PGFREE, 1 << order);
1712 void __free_pages_core(struct page *page, unsigned int order)
1714 unsigned int nr_pages = 1 << order;
1715 struct page *p = page;
1719 * When initializing the memmap, __init_single_page() sets the refcount
1720 * of all pages to 1 ("allocated"/"not free"). We have to set the
1721 * refcount of all involved pages to 0.
1724 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1726 __ClearPageReserved(p);
1727 set_page_count(p, 0);
1729 __ClearPageReserved(p);
1730 set_page_count(p, 0);
1732 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1735 * Bypass PCP and place fresh pages right to the tail, primarily
1736 * relevant for memory onlining.
1738 __free_pages_ok(page, order, FPI_TO_TAIL | FPI_SKIP_KASAN_POISON);
1744 * During memory init memblocks map pfns to nids. The search is expensive and
1745 * this caches recent lookups. The implementation of __early_pfn_to_nid
1746 * treats start/end as pfns.
1748 struct mminit_pfnnid_cache {
1749 unsigned long last_start;
1750 unsigned long last_end;
1754 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1757 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1759 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1760 struct mminit_pfnnid_cache *state)
1762 unsigned long start_pfn, end_pfn;
1765 if (state->last_start <= pfn && pfn < state->last_end)
1766 return state->last_nid;
1768 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1769 if (nid != NUMA_NO_NODE) {
1770 state->last_start = start_pfn;
1771 state->last_end = end_pfn;
1772 state->last_nid = nid;
1778 int __meminit early_pfn_to_nid(unsigned long pfn)
1780 static DEFINE_SPINLOCK(early_pfn_lock);
1783 spin_lock(&early_pfn_lock);
1784 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1786 nid = first_online_node;
1787 spin_unlock(&early_pfn_lock);
1791 #endif /* CONFIG_NUMA */
1793 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1796 if (early_page_uninitialised(pfn))
1798 if (!kmsan_memblock_free_pages(page, order)) {
1799 /* KMSAN will take care of these pages. */
1802 __free_pages_core(page, order);
1806 * Check that the whole (or subset of) a pageblock given by the interval of
1807 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1808 * with the migration of free compaction scanner.
1810 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1812 * It's possible on some configurations to have a setup like node0 node1 node0
1813 * i.e. it's possible that all pages within a zones range of pages do not
1814 * belong to a single zone. We assume that a border between node0 and node1
1815 * can occur within a single pageblock, but not a node0 node1 node0
1816 * interleaving within a single pageblock. It is therefore sufficient to check
1817 * the first and last page of a pageblock and avoid checking each individual
1818 * page in a pageblock.
1820 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1821 unsigned long end_pfn, struct zone *zone)
1823 struct page *start_page;
1824 struct page *end_page;
1826 /* end_pfn is one past the range we are checking */
1829 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1832 start_page = pfn_to_online_page(start_pfn);
1836 if (page_zone(start_page) != zone)
1839 end_page = pfn_to_page(end_pfn);
1841 /* This gives a shorter code than deriving page_zone(end_page) */
1842 if (page_zone_id(start_page) != page_zone_id(end_page))
1848 void set_zone_contiguous(struct zone *zone)
1850 unsigned long block_start_pfn = zone->zone_start_pfn;
1851 unsigned long block_end_pfn;
1853 block_end_pfn = pageblock_end_pfn(block_start_pfn);
1854 for (; block_start_pfn < zone_end_pfn(zone);
1855 block_start_pfn = block_end_pfn,
1856 block_end_pfn += pageblock_nr_pages) {
1858 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1860 if (!__pageblock_pfn_to_page(block_start_pfn,
1861 block_end_pfn, zone))
1866 /* We confirm that there is no hole */
1867 zone->contiguous = true;
1870 void clear_zone_contiguous(struct zone *zone)
1872 zone->contiguous = false;
1875 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1876 static void __init deferred_free_range(unsigned long pfn,
1877 unsigned long nr_pages)
1885 page = pfn_to_page(pfn);
1887 /* Free a large naturally-aligned chunk if possible */
1888 if (nr_pages == pageblock_nr_pages && pageblock_aligned(pfn)) {
1889 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1890 __free_pages_core(page, pageblock_order);
1894 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1895 if (pageblock_aligned(pfn))
1896 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1897 __free_pages_core(page, 0);
1901 /* Completion tracking for deferred_init_memmap() threads */
1902 static atomic_t pgdat_init_n_undone __initdata;
1903 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1905 static inline void __init pgdat_init_report_one_done(void)
1907 if (atomic_dec_and_test(&pgdat_init_n_undone))
1908 complete(&pgdat_init_all_done_comp);
1912 * Returns true if page needs to be initialized or freed to buddy allocator.
1914 * We check if a current large page is valid by only checking the validity
1917 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1919 if (pageblock_aligned(pfn) && !pfn_valid(pfn))
1925 * Free pages to buddy allocator. Try to free aligned pages in
1926 * pageblock_nr_pages sizes.
1928 static void __init deferred_free_pages(unsigned long pfn,
1929 unsigned long end_pfn)
1931 unsigned long nr_free = 0;
1933 for (; pfn < end_pfn; pfn++) {
1934 if (!deferred_pfn_valid(pfn)) {
1935 deferred_free_range(pfn - nr_free, nr_free);
1937 } else if (pageblock_aligned(pfn)) {
1938 deferred_free_range(pfn - nr_free, nr_free);
1944 /* Free the last block of pages to allocator */
1945 deferred_free_range(pfn - nr_free, nr_free);
1949 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1950 * by performing it only once every pageblock_nr_pages.
1951 * Return number of pages initialized.
1953 static unsigned long __init deferred_init_pages(struct zone *zone,
1955 unsigned long end_pfn)
1957 int nid = zone_to_nid(zone);
1958 unsigned long nr_pages = 0;
1959 int zid = zone_idx(zone);
1960 struct page *page = NULL;
1962 for (; pfn < end_pfn; pfn++) {
1963 if (!deferred_pfn_valid(pfn)) {
1966 } else if (!page || pageblock_aligned(pfn)) {
1967 page = pfn_to_page(pfn);
1971 __init_single_page(page, pfn, zid, nid);
1978 * This function is meant to pre-load the iterator for the zone init.
1979 * Specifically it walks through the ranges until we are caught up to the
1980 * first_init_pfn value and exits there. If we never encounter the value we
1981 * return false indicating there are no valid ranges left.
1984 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1985 unsigned long *spfn, unsigned long *epfn,
1986 unsigned long first_init_pfn)
1991 * Start out by walking through the ranges in this zone that have
1992 * already been initialized. We don't need to do anything with them
1993 * so we just need to flush them out of the system.
1995 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1996 if (*epfn <= first_init_pfn)
1998 if (*spfn < first_init_pfn)
1999 *spfn = first_init_pfn;
2008 * Initialize and free pages. We do it in two loops: first we initialize
2009 * struct page, then free to buddy allocator, because while we are
2010 * freeing pages we can access pages that are ahead (computing buddy
2011 * page in __free_one_page()).
2013 * In order to try and keep some memory in the cache we have the loop
2014 * broken along max page order boundaries. This way we will not cause
2015 * any issues with the buddy page computation.
2017 static unsigned long __init
2018 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
2019 unsigned long *end_pfn)
2021 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
2022 unsigned long spfn = *start_pfn, epfn = *end_pfn;
2023 unsigned long nr_pages = 0;
2026 /* First we loop through and initialize the page values */
2027 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
2030 if (mo_pfn <= *start_pfn)
2033 t = min(mo_pfn, *end_pfn);
2034 nr_pages += deferred_init_pages(zone, *start_pfn, t);
2036 if (mo_pfn < *end_pfn) {
2037 *start_pfn = mo_pfn;
2042 /* Reset values and now loop through freeing pages as needed */
2045 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
2051 t = min(mo_pfn, epfn);
2052 deferred_free_pages(spfn, t);
2062 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
2065 unsigned long spfn, epfn;
2066 struct zone *zone = arg;
2069 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
2072 * Initialize and free pages in MAX_ORDER sized increments so that we
2073 * can avoid introducing any issues with the buddy allocator.
2075 while (spfn < end_pfn) {
2076 deferred_init_maxorder(&i, zone, &spfn, &epfn);
2081 /* An arch may override for more concurrency. */
2083 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
2088 /* Initialise remaining memory on a node */
2089 static int __init deferred_init_memmap(void *data)
2091 pg_data_t *pgdat = data;
2092 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2093 unsigned long spfn = 0, epfn = 0;
2094 unsigned long first_init_pfn, flags;
2095 unsigned long start = jiffies;
2097 int zid, max_threads;
2100 /* Bind memory initialisation thread to a local node if possible */
2101 if (!cpumask_empty(cpumask))
2102 set_cpus_allowed_ptr(current, cpumask);
2104 pgdat_resize_lock(pgdat, &flags);
2105 first_init_pfn = pgdat->first_deferred_pfn;
2106 if (first_init_pfn == ULONG_MAX) {
2107 pgdat_resize_unlock(pgdat, &flags);
2108 pgdat_init_report_one_done();
2112 /* Sanity check boundaries */
2113 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
2114 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
2115 pgdat->first_deferred_pfn = ULONG_MAX;
2118 * Once we unlock here, the zone cannot be grown anymore, thus if an
2119 * interrupt thread must allocate this early in boot, zone must be
2120 * pre-grown prior to start of deferred page initialization.
2122 pgdat_resize_unlock(pgdat, &flags);
2124 /* Only the highest zone is deferred so find it */
2125 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2126 zone = pgdat->node_zones + zid;
2127 if (first_init_pfn < zone_end_pfn(zone))
2131 /* If the zone is empty somebody else may have cleared out the zone */
2132 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2136 max_threads = deferred_page_init_max_threads(cpumask);
2138 while (spfn < epfn) {
2139 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
2140 struct padata_mt_job job = {
2141 .thread_fn = deferred_init_memmap_chunk,
2144 .size = epfn_align - spfn,
2145 .align = PAGES_PER_SECTION,
2146 .min_chunk = PAGES_PER_SECTION,
2147 .max_threads = max_threads,
2150 padata_do_multithreaded(&job);
2151 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2155 /* Sanity check that the next zone really is unpopulated */
2156 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2158 pr_info("node %d deferred pages initialised in %ums\n",
2159 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2161 pgdat_init_report_one_done();
2166 * If this zone has deferred pages, try to grow it by initializing enough
2167 * deferred pages to satisfy the allocation specified by order, rounded up to
2168 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2169 * of SECTION_SIZE bytes by initializing struct pages in increments of
2170 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2172 * Return true when zone was grown, otherwise return false. We return true even
2173 * when we grow less than requested, to let the caller decide if there are
2174 * enough pages to satisfy the allocation.
2176 * Note: We use noinline because this function is needed only during boot, and
2177 * it is called from a __ref function _deferred_grow_zone. This way we are
2178 * making sure that it is not inlined into permanent text section.
2180 static noinline bool __init
2181 deferred_grow_zone(struct zone *zone, unsigned int order)
2183 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2184 pg_data_t *pgdat = zone->zone_pgdat;
2185 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2186 unsigned long spfn, epfn, flags;
2187 unsigned long nr_pages = 0;
2190 /* Only the last zone may have deferred pages */
2191 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2194 pgdat_resize_lock(pgdat, &flags);
2197 * If someone grew this zone while we were waiting for spinlock, return
2198 * true, as there might be enough pages already.
2200 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2201 pgdat_resize_unlock(pgdat, &flags);
2205 /* If the zone is empty somebody else may have cleared out the zone */
2206 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2207 first_deferred_pfn)) {
2208 pgdat->first_deferred_pfn = ULONG_MAX;
2209 pgdat_resize_unlock(pgdat, &flags);
2210 /* Retry only once. */
2211 return first_deferred_pfn != ULONG_MAX;
2215 * Initialize and free pages in MAX_ORDER sized increments so
2216 * that we can avoid introducing any issues with the buddy
2219 while (spfn < epfn) {
2220 /* update our first deferred PFN for this section */
2221 first_deferred_pfn = spfn;
2223 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2224 touch_nmi_watchdog();
2226 /* We should only stop along section boundaries */
2227 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2230 /* If our quota has been met we can stop here */
2231 if (nr_pages >= nr_pages_needed)
2235 pgdat->first_deferred_pfn = spfn;
2236 pgdat_resize_unlock(pgdat, &flags);
2238 return nr_pages > 0;
2242 * deferred_grow_zone() is __init, but it is called from
2243 * get_page_from_freelist() during early boot until deferred_pages permanently
2244 * disables this call. This is why we have refdata wrapper to avoid warning,
2245 * and to ensure that the function body gets unloaded.
2248 _deferred_grow_zone(struct zone *zone, unsigned int order)
2250 return deferred_grow_zone(zone, order);
2253 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2255 void __init page_alloc_init_late(void)
2260 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2262 /* There will be num_node_state(N_MEMORY) threads */
2263 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2264 for_each_node_state(nid, N_MEMORY) {
2265 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2268 /* Block until all are initialised */
2269 wait_for_completion(&pgdat_init_all_done_comp);
2272 * We initialized the rest of the deferred pages. Permanently disable
2273 * on-demand struct page initialization.
2275 static_branch_disable(&deferred_pages);
2277 /* Reinit limits that are based on free pages after the kernel is up */
2278 files_maxfiles_init();
2283 /* Discard memblock private memory */
2286 for_each_node_state(nid, N_MEMORY)
2287 shuffle_free_memory(NODE_DATA(nid));
2289 for_each_populated_zone(zone)
2290 set_zone_contiguous(zone);
2294 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2295 void __init init_cma_reserved_pageblock(struct page *page)
2297 unsigned i = pageblock_nr_pages;
2298 struct page *p = page;
2301 __ClearPageReserved(p);
2302 set_page_count(p, 0);
2305 set_pageblock_migratetype(page, MIGRATE_CMA);
2306 set_page_refcounted(page);
2307 __free_pages(page, pageblock_order);
2309 adjust_managed_page_count(page, pageblock_nr_pages);
2310 page_zone(page)->cma_pages += pageblock_nr_pages;
2315 * The order of subdivision here is critical for the IO subsystem.
2316 * Please do not alter this order without good reasons and regression
2317 * testing. Specifically, as large blocks of memory are subdivided,
2318 * the order in which smaller blocks are delivered depends on the order
2319 * they're subdivided in this function. This is the primary factor
2320 * influencing the order in which pages are delivered to the IO
2321 * subsystem according to empirical testing, and this is also justified
2322 * by considering the behavior of a buddy system containing a single
2323 * large block of memory acted on by a series of small allocations.
2324 * This behavior is a critical factor in sglist merging's success.
2328 static inline void expand(struct zone *zone, struct page *page,
2329 int low, int high, int migratetype)
2331 unsigned long size = 1 << high;
2333 while (high > low) {
2336 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2339 * Mark as guard pages (or page), that will allow to
2340 * merge back to allocator when buddy will be freed.
2341 * Corresponding page table entries will not be touched,
2342 * pages will stay not present in virtual address space
2344 if (set_page_guard(zone, &page[size], high, migratetype))
2347 add_to_free_list(&page[size], zone, high, migratetype);
2348 set_buddy_order(&page[size], high);
2352 static void check_new_page_bad(struct page *page)
2354 if (unlikely(page->flags & __PG_HWPOISON)) {
2355 /* Don't complain about hwpoisoned pages */
2356 page_mapcount_reset(page); /* remove PageBuddy */
2361 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2365 * This page is about to be returned from the page allocator
2367 static inline int check_new_page(struct page *page)
2369 if (likely(page_expected_state(page,
2370 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2373 check_new_page_bad(page);
2377 static bool check_new_pages(struct page *page, unsigned int order)
2380 for (i = 0; i < (1 << order); i++) {
2381 struct page *p = page + i;
2383 if (unlikely(check_new_page(p)))
2390 #ifdef CONFIG_DEBUG_VM
2392 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2393 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2394 * also checked when pcp lists are refilled from the free lists.
2396 static inline bool check_pcp_refill(struct page *page, unsigned int order)
2398 if (debug_pagealloc_enabled_static())
2399 return check_new_pages(page, order);
2404 static inline bool check_new_pcp(struct page *page, unsigned int order)
2406 return check_new_pages(page, order);
2410 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2411 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2412 * enabled, they are also checked when being allocated from the pcp lists.
2414 static inline bool check_pcp_refill(struct page *page, unsigned int order)
2416 return check_new_pages(page, order);
2418 static inline bool check_new_pcp(struct page *page, unsigned int order)
2420 if (debug_pagealloc_enabled_static())
2421 return check_new_pages(page, order);
2425 #endif /* CONFIG_DEBUG_VM */
2427 static inline bool should_skip_kasan_unpoison(gfp_t flags)
2429 /* Don't skip if a software KASAN mode is enabled. */
2430 if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
2431 IS_ENABLED(CONFIG_KASAN_SW_TAGS))
2434 /* Skip, if hardware tag-based KASAN is not enabled. */
2435 if (!kasan_hw_tags_enabled())
2439 * With hardware tag-based KASAN enabled, skip if this has been
2440 * requested via __GFP_SKIP_KASAN_UNPOISON.
2442 return flags & __GFP_SKIP_KASAN_UNPOISON;
2445 static inline bool should_skip_init(gfp_t flags)
2447 /* Don't skip, if hardware tag-based KASAN is not enabled. */
2448 if (!kasan_hw_tags_enabled())
2451 /* For hardware tag-based KASAN, skip if requested. */
2452 return (flags & __GFP_SKIP_ZERO);
2455 inline void post_alloc_hook(struct page *page, unsigned int order,
2458 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) &&
2459 !should_skip_init(gfp_flags);
2460 bool init_tags = init && (gfp_flags & __GFP_ZEROTAGS);
2463 set_page_private(page, 0);
2464 set_page_refcounted(page);
2466 arch_alloc_page(page, order);
2467 debug_pagealloc_map_pages(page, 1 << order);
2470 * Page unpoisoning must happen before memory initialization.
2471 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2472 * allocations and the page unpoisoning code will complain.
2474 kernel_unpoison_pages(page, 1 << order);
2477 * As memory initialization might be integrated into KASAN,
2478 * KASAN unpoisoning and memory initializion code must be
2479 * kept together to avoid discrepancies in behavior.
2483 * If memory tags should be zeroed (which happens only when memory
2484 * should be initialized as well).
2487 /* Initialize both memory and tags. */
2488 for (i = 0; i != 1 << order; ++i)
2489 tag_clear_highpage(page + i);
2491 /* Note that memory is already initialized by the loop above. */
2494 if (!should_skip_kasan_unpoison(gfp_flags)) {
2495 /* Unpoison shadow memory or set memory tags. */
2496 kasan_unpoison_pages(page, order, init);
2498 /* Note that memory is already initialized by KASAN. */
2499 if (kasan_has_integrated_init())
2502 /* Ensure page_address() dereferencing does not fault. */
2503 for (i = 0; i != 1 << order; ++i)
2504 page_kasan_tag_reset(page + i);
2506 /* If memory is still not initialized, do it now. */
2508 kernel_init_pages(page, 1 << order);
2509 /* Propagate __GFP_SKIP_KASAN_POISON to page flags. */
2510 if (kasan_hw_tags_enabled() && (gfp_flags & __GFP_SKIP_KASAN_POISON))
2511 SetPageSkipKASanPoison(page);
2513 set_page_owner(page, order, gfp_flags);
2514 page_table_check_alloc(page, order);
2517 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2518 unsigned int alloc_flags)
2520 post_alloc_hook(page, order, gfp_flags);
2522 if (order && (gfp_flags & __GFP_COMP))
2523 prep_compound_page(page, order);
2526 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2527 * allocate the page. The expectation is that the caller is taking
2528 * steps that will free more memory. The caller should avoid the page
2529 * being used for !PFMEMALLOC purposes.
2531 if (alloc_flags & ALLOC_NO_WATERMARKS)
2532 set_page_pfmemalloc(page);
2534 clear_page_pfmemalloc(page);
2538 * Go through the free lists for the given migratetype and remove
2539 * the smallest available page from the freelists
2541 static __always_inline
2542 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2545 unsigned int current_order;
2546 struct free_area *area;
2549 /* Find a page of the appropriate size in the preferred list */
2550 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2551 area = &(zone->free_area[current_order]);
2552 page = get_page_from_free_area(area, migratetype);
2555 del_page_from_free_list(page, zone, current_order);
2556 expand(zone, page, order, current_order, migratetype);
2557 set_pcppage_migratetype(page, migratetype);
2558 trace_mm_page_alloc_zone_locked(page, order, migratetype,
2559 pcp_allowed_order(order) &&
2560 migratetype < MIGRATE_PCPTYPES);
2569 * This array describes the order lists are fallen back to when
2570 * the free lists for the desirable migrate type are depleted
2572 * The other migratetypes do not have fallbacks.
2574 static int fallbacks[MIGRATE_TYPES][3] = {
2575 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2576 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2577 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2581 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2584 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2587 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2588 unsigned int order) { return NULL; }
2592 * Move the free pages in a range to the freelist tail of the requested type.
2593 * Note that start_page and end_pages are not aligned on a pageblock
2594 * boundary. If alignment is required, use move_freepages_block()
2596 static int move_freepages(struct zone *zone,
2597 unsigned long start_pfn, unsigned long end_pfn,
2598 int migratetype, int *num_movable)
2603 int pages_moved = 0;
2605 for (pfn = start_pfn; pfn <= end_pfn;) {
2606 page = pfn_to_page(pfn);
2607 if (!PageBuddy(page)) {
2609 * We assume that pages that could be isolated for
2610 * migration are movable. But we don't actually try
2611 * isolating, as that would be expensive.
2614 (PageLRU(page) || __PageMovable(page)))
2620 /* Make sure we are not inadvertently changing nodes */
2621 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2622 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2624 order = buddy_order(page);
2625 move_to_free_list(page, zone, order, migratetype);
2627 pages_moved += 1 << order;
2633 int move_freepages_block(struct zone *zone, struct page *page,
2634 int migratetype, int *num_movable)
2636 unsigned long start_pfn, end_pfn, pfn;
2641 pfn = page_to_pfn(page);
2642 start_pfn = pageblock_start_pfn(pfn);
2643 end_pfn = pageblock_end_pfn(pfn) - 1;
2645 /* Do not cross zone boundaries */
2646 if (!zone_spans_pfn(zone, start_pfn))
2648 if (!zone_spans_pfn(zone, end_pfn))
2651 return move_freepages(zone, start_pfn, end_pfn, migratetype,
2655 static void change_pageblock_range(struct page *pageblock_page,
2656 int start_order, int migratetype)
2658 int nr_pageblocks = 1 << (start_order - pageblock_order);
2660 while (nr_pageblocks--) {
2661 set_pageblock_migratetype(pageblock_page, migratetype);
2662 pageblock_page += pageblock_nr_pages;
2667 * When we are falling back to another migratetype during allocation, try to
2668 * steal extra free pages from the same pageblocks to satisfy further
2669 * allocations, instead of polluting multiple pageblocks.
2671 * If we are stealing a relatively large buddy page, it is likely there will
2672 * be more free pages in the pageblock, so try to steal them all. For
2673 * reclaimable and unmovable allocations, we steal regardless of page size,
2674 * as fragmentation caused by those allocations polluting movable pageblocks
2675 * is worse than movable allocations stealing from unmovable and reclaimable
2678 static bool can_steal_fallback(unsigned int order, int start_mt)
2681 * Leaving this order check is intended, although there is
2682 * relaxed order check in next check. The reason is that
2683 * we can actually steal whole pageblock if this condition met,
2684 * but, below check doesn't guarantee it and that is just heuristic
2685 * so could be changed anytime.
2687 if (order >= pageblock_order)
2690 if (order >= pageblock_order / 2 ||
2691 start_mt == MIGRATE_RECLAIMABLE ||
2692 start_mt == MIGRATE_UNMOVABLE ||
2693 page_group_by_mobility_disabled)
2699 static inline bool boost_watermark(struct zone *zone)
2701 unsigned long max_boost;
2703 if (!watermark_boost_factor)
2706 * Don't bother in zones that are unlikely to produce results.
2707 * On small machines, including kdump capture kernels running
2708 * in a small area, boosting the watermark can cause an out of
2709 * memory situation immediately.
2711 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2714 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2715 watermark_boost_factor, 10000);
2718 * high watermark may be uninitialised if fragmentation occurs
2719 * very early in boot so do not boost. We do not fall
2720 * through and boost by pageblock_nr_pages as failing
2721 * allocations that early means that reclaim is not going
2722 * to help and it may even be impossible to reclaim the
2723 * boosted watermark resulting in a hang.
2728 max_boost = max(pageblock_nr_pages, max_boost);
2730 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2737 * This function implements actual steal behaviour. If order is large enough,
2738 * we can steal whole pageblock. If not, we first move freepages in this
2739 * pageblock to our migratetype and determine how many already-allocated pages
2740 * are there in the pageblock with a compatible migratetype. If at least half
2741 * of pages are free or compatible, we can change migratetype of the pageblock
2742 * itself, so pages freed in the future will be put on the correct free list.
2744 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2745 unsigned int alloc_flags, int start_type, bool whole_block)
2747 unsigned int current_order = buddy_order(page);
2748 int free_pages, movable_pages, alike_pages;
2751 old_block_type = get_pageblock_migratetype(page);
2754 * This can happen due to races and we want to prevent broken
2755 * highatomic accounting.
2757 if (is_migrate_highatomic(old_block_type))
2760 /* Take ownership for orders >= pageblock_order */
2761 if (current_order >= pageblock_order) {
2762 change_pageblock_range(page, current_order, start_type);
2767 * Boost watermarks to increase reclaim pressure to reduce the
2768 * likelihood of future fallbacks. Wake kswapd now as the node
2769 * may be balanced overall and kswapd will not wake naturally.
2771 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2772 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2774 /* We are not allowed to try stealing from the whole block */
2778 free_pages = move_freepages_block(zone, page, start_type,
2781 * Determine how many pages are compatible with our allocation.
2782 * For movable allocation, it's the number of movable pages which
2783 * we just obtained. For other types it's a bit more tricky.
2785 if (start_type == MIGRATE_MOVABLE) {
2786 alike_pages = movable_pages;
2789 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2790 * to MOVABLE pageblock, consider all non-movable pages as
2791 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2792 * vice versa, be conservative since we can't distinguish the
2793 * exact migratetype of non-movable pages.
2795 if (old_block_type == MIGRATE_MOVABLE)
2796 alike_pages = pageblock_nr_pages
2797 - (free_pages + movable_pages);
2802 /* moving whole block can fail due to zone boundary conditions */
2807 * If a sufficient number of pages in the block are either free or of
2808 * comparable migratability as our allocation, claim the whole block.
2810 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2811 page_group_by_mobility_disabled)
2812 set_pageblock_migratetype(page, start_type);
2817 move_to_free_list(page, zone, current_order, start_type);
2821 * Check whether there is a suitable fallback freepage with requested order.
2822 * If only_stealable is true, this function returns fallback_mt only if
2823 * we can steal other freepages all together. This would help to reduce
2824 * fragmentation due to mixed migratetype pages in one pageblock.
2826 int find_suitable_fallback(struct free_area *area, unsigned int order,
2827 int migratetype, bool only_stealable, bool *can_steal)
2832 if (area->nr_free == 0)
2837 fallback_mt = fallbacks[migratetype][i];
2838 if (fallback_mt == MIGRATE_TYPES)
2841 if (free_area_empty(area, fallback_mt))
2844 if (can_steal_fallback(order, migratetype))
2847 if (!only_stealable)
2858 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2859 * there are no empty page blocks that contain a page with a suitable order
2861 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2862 unsigned int alloc_order)
2865 unsigned long max_managed, flags;
2868 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2869 * Check is race-prone but harmless.
2871 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2872 if (zone->nr_reserved_highatomic >= max_managed)
2875 spin_lock_irqsave(&zone->lock, flags);
2877 /* Recheck the nr_reserved_highatomic limit under the lock */
2878 if (zone->nr_reserved_highatomic >= max_managed)
2882 mt = get_pageblock_migratetype(page);
2883 /* Only reserve normal pageblocks (i.e., they can merge with others) */
2884 if (migratetype_is_mergeable(mt)) {
2885 zone->nr_reserved_highatomic += pageblock_nr_pages;
2886 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2887 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2891 spin_unlock_irqrestore(&zone->lock, flags);
2895 * Used when an allocation is about to fail under memory pressure. This
2896 * potentially hurts the reliability of high-order allocations when under
2897 * intense memory pressure but failed atomic allocations should be easier
2898 * to recover from than an OOM.
2900 * If @force is true, try to unreserve a pageblock even though highatomic
2901 * pageblock is exhausted.
2903 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2906 struct zonelist *zonelist = ac->zonelist;
2907 unsigned long flags;
2914 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2917 * Preserve at least one pageblock unless memory pressure
2920 if (!force && zone->nr_reserved_highatomic <=
2924 spin_lock_irqsave(&zone->lock, flags);
2925 for (order = 0; order < MAX_ORDER; order++) {
2926 struct free_area *area = &(zone->free_area[order]);
2928 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2933 * In page freeing path, migratetype change is racy so
2934 * we can counter several free pages in a pageblock
2935 * in this loop although we changed the pageblock type
2936 * from highatomic to ac->migratetype. So we should
2937 * adjust the count once.
2939 if (is_migrate_highatomic_page(page)) {
2941 * It should never happen but changes to
2942 * locking could inadvertently allow a per-cpu
2943 * drain to add pages to MIGRATE_HIGHATOMIC
2944 * while unreserving so be safe and watch for
2947 zone->nr_reserved_highatomic -= min(
2949 zone->nr_reserved_highatomic);
2953 * Convert to ac->migratetype and avoid the normal
2954 * pageblock stealing heuristics. Minimally, the caller
2955 * is doing the work and needs the pages. More
2956 * importantly, if the block was always converted to
2957 * MIGRATE_UNMOVABLE or another type then the number
2958 * of pageblocks that cannot be completely freed
2961 set_pageblock_migratetype(page, ac->migratetype);
2962 ret = move_freepages_block(zone, page, ac->migratetype,
2965 spin_unlock_irqrestore(&zone->lock, flags);
2969 spin_unlock_irqrestore(&zone->lock, flags);
2976 * Try finding a free buddy page on the fallback list and put it on the free
2977 * list of requested migratetype, possibly along with other pages from the same
2978 * block, depending on fragmentation avoidance heuristics. Returns true if
2979 * fallback was found so that __rmqueue_smallest() can grab it.
2981 * The use of signed ints for order and current_order is a deliberate
2982 * deviation from the rest of this file, to make the for loop
2983 * condition simpler.
2985 static __always_inline bool
2986 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2987 unsigned int alloc_flags)
2989 struct free_area *area;
2991 int min_order = order;
2997 * Do not steal pages from freelists belonging to other pageblocks
2998 * i.e. orders < pageblock_order. If there are no local zones free,
2999 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
3001 if (order < pageblock_order && alloc_flags & ALLOC_NOFRAGMENT)
3002 min_order = pageblock_order;
3005 * Find the largest available free page in the other list. This roughly
3006 * approximates finding the pageblock with the most free pages, which
3007 * would be too costly to do exactly.
3009 for (current_order = MAX_ORDER - 1; current_order >= min_order;
3011 area = &(zone->free_area[current_order]);
3012 fallback_mt = find_suitable_fallback(area, current_order,
3013 start_migratetype, false, &can_steal);
3014 if (fallback_mt == -1)
3018 * We cannot steal all free pages from the pageblock and the
3019 * requested migratetype is movable. In that case it's better to
3020 * steal and split the smallest available page instead of the
3021 * largest available page, because even if the next movable
3022 * allocation falls back into a different pageblock than this
3023 * one, it won't cause permanent fragmentation.
3025 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
3026 && current_order > order)
3035 for (current_order = order; current_order < MAX_ORDER;
3037 area = &(zone->free_area[current_order]);
3038 fallback_mt = find_suitable_fallback(area, current_order,
3039 start_migratetype, false, &can_steal);
3040 if (fallback_mt != -1)
3045 * This should not happen - we already found a suitable fallback
3046 * when looking for the largest page.
3048 VM_BUG_ON(current_order == MAX_ORDER);
3051 page = get_page_from_free_area(area, fallback_mt);
3053 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
3056 trace_mm_page_alloc_extfrag(page, order, current_order,
3057 start_migratetype, fallback_mt);
3064 * Do the hard work of removing an element from the buddy allocator.
3065 * Call me with the zone->lock already held.
3067 static __always_inline struct page *
3068 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
3069 unsigned int alloc_flags)
3073 if (IS_ENABLED(CONFIG_CMA)) {
3075 * Balance movable allocations between regular and CMA areas by
3076 * allocating from CMA when over half of the zone's free memory
3077 * is in the CMA area.
3079 if (alloc_flags & ALLOC_CMA &&
3080 zone_page_state(zone, NR_FREE_CMA_PAGES) >
3081 zone_page_state(zone, NR_FREE_PAGES) / 2) {
3082 page = __rmqueue_cma_fallback(zone, order);
3088 page = __rmqueue_smallest(zone, order, migratetype);
3089 if (unlikely(!page)) {
3090 if (alloc_flags & ALLOC_CMA)
3091 page = __rmqueue_cma_fallback(zone, order);
3093 if (!page && __rmqueue_fallback(zone, order, migratetype,
3101 * Obtain a specified number of elements from the buddy allocator, all under
3102 * a single hold of the lock, for efficiency. Add them to the supplied list.
3103 * Returns the number of new pages which were placed at *list.
3105 static int rmqueue_bulk(struct zone *zone, unsigned int order,
3106 unsigned long count, struct list_head *list,
3107 int migratetype, unsigned int alloc_flags)
3109 unsigned long flags;
3110 int i, allocated = 0;
3112 spin_lock_irqsave(&zone->lock, flags);
3113 for (i = 0; i < count; ++i) {
3114 struct page *page = __rmqueue(zone, order, migratetype,
3116 if (unlikely(page == NULL))
3119 if (unlikely(check_pcp_refill(page, order)))
3123 * Split buddy pages returned by expand() are received here in
3124 * physical page order. The page is added to the tail of
3125 * caller's list. From the callers perspective, the linked list
3126 * is ordered by page number under some conditions. This is
3127 * useful for IO devices that can forward direction from the
3128 * head, thus also in the physical page order. This is useful
3129 * for IO devices that can merge IO requests if the physical
3130 * pages are ordered properly.
3132 list_add_tail(&page->pcp_list, list);
3134 if (is_migrate_cma(get_pcppage_migratetype(page)))
3135 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
3140 * i pages were removed from the buddy list even if some leak due
3141 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
3142 * on i. Do not confuse with 'allocated' which is the number of
3143 * pages added to the pcp list.
3145 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
3146 spin_unlock_irqrestore(&zone->lock, flags);
3152 * Called from the vmstat counter updater to drain pagesets of this
3153 * currently executing processor on remote nodes after they have
3156 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
3158 int to_drain, batch;
3160 batch = READ_ONCE(pcp->batch);
3161 to_drain = min(pcp->count, batch);
3163 spin_lock(&pcp->lock);
3164 free_pcppages_bulk(zone, to_drain, pcp, 0);
3165 spin_unlock(&pcp->lock);
3171 * Drain pcplists of the indicated processor and zone.
3173 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
3175 struct per_cpu_pages *pcp;
3177 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3179 spin_lock(&pcp->lock);
3180 free_pcppages_bulk(zone, pcp->count, pcp, 0);
3181 spin_unlock(&pcp->lock);
3186 * Drain pcplists of all zones on the indicated processor.
3188 static void drain_pages(unsigned int cpu)
3192 for_each_populated_zone(zone) {
3193 drain_pages_zone(cpu, zone);
3198 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3200 void drain_local_pages(struct zone *zone)
3202 int cpu = smp_processor_id();
3205 drain_pages_zone(cpu, zone);
3211 * The implementation of drain_all_pages(), exposing an extra parameter to
3212 * drain on all cpus.
3214 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3215 * not empty. The check for non-emptiness can however race with a free to
3216 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3217 * that need the guarantee that every CPU has drained can disable the
3218 * optimizing racy check.
3220 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3225 * Allocate in the BSS so we won't require allocation in
3226 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3228 static cpumask_t cpus_with_pcps;
3231 * Do not drain if one is already in progress unless it's specific to
3232 * a zone. Such callers are primarily CMA and memory hotplug and need
3233 * the drain to be complete when the call returns.
3235 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3238 mutex_lock(&pcpu_drain_mutex);
3242 * We don't care about racing with CPU hotplug event
3243 * as offline notification will cause the notified
3244 * cpu to drain that CPU pcps and on_each_cpu_mask
3245 * disables preemption as part of its processing
3247 for_each_online_cpu(cpu) {
3248 struct per_cpu_pages *pcp;
3250 bool has_pcps = false;
3252 if (force_all_cpus) {
3254 * The pcp.count check is racy, some callers need a
3255 * guarantee that no cpu is missed.
3259 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3263 for_each_populated_zone(z) {
3264 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
3273 cpumask_set_cpu(cpu, &cpus_with_pcps);
3275 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3278 for_each_cpu(cpu, &cpus_with_pcps) {
3280 drain_pages_zone(cpu, zone);
3285 mutex_unlock(&pcpu_drain_mutex);
3289 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3291 * When zone parameter is non-NULL, spill just the single zone's pages.
3293 void drain_all_pages(struct zone *zone)
3295 __drain_all_pages(zone, false);
3298 #ifdef CONFIG_HIBERNATION
3301 * Touch the watchdog for every WD_PAGE_COUNT pages.
3303 #define WD_PAGE_COUNT (128*1024)
3305 void mark_free_pages(struct zone *zone)
3307 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3308 unsigned long flags;
3309 unsigned int order, t;
3312 if (zone_is_empty(zone))
3315 spin_lock_irqsave(&zone->lock, flags);
3317 max_zone_pfn = zone_end_pfn(zone);
3318 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3319 if (pfn_valid(pfn)) {
3320 page = pfn_to_page(pfn);
3322 if (!--page_count) {
3323 touch_nmi_watchdog();
3324 page_count = WD_PAGE_COUNT;
3327 if (page_zone(page) != zone)
3330 if (!swsusp_page_is_forbidden(page))
3331 swsusp_unset_page_free(page);
3334 for_each_migratetype_order(order, t) {
3335 list_for_each_entry(page,
3336 &zone->free_area[order].free_list[t], buddy_list) {
3339 pfn = page_to_pfn(page);
3340 for (i = 0; i < (1UL << order); i++) {
3341 if (!--page_count) {
3342 touch_nmi_watchdog();
3343 page_count = WD_PAGE_COUNT;
3345 swsusp_set_page_free(pfn_to_page(pfn + i));
3349 spin_unlock_irqrestore(&zone->lock, flags);
3351 #endif /* CONFIG_PM */
3353 static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
3358 if (!free_pcp_prepare(page, order))
3361 migratetype = get_pfnblock_migratetype(page, pfn);
3362 set_pcppage_migratetype(page, migratetype);
3366 static int nr_pcp_free(struct per_cpu_pages *pcp, int high, int batch,
3369 int min_nr_free, max_nr_free;
3371 /* Free everything if batch freeing high-order pages. */
3372 if (unlikely(free_high))
3375 /* Check for PCP disabled or boot pageset */
3376 if (unlikely(high < batch))
3379 /* Leave at least pcp->batch pages on the list */
3380 min_nr_free = batch;
3381 max_nr_free = high - batch;
3384 * Double the number of pages freed each time there is subsequent
3385 * freeing of pages without any allocation.
3387 batch <<= pcp->free_factor;
3388 if (batch < max_nr_free)
3390 batch = clamp(batch, min_nr_free, max_nr_free);
3395 static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
3398 int high = READ_ONCE(pcp->high);
3400 if (unlikely(!high || free_high))
3403 if (!test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags))
3407 * If reclaim is active, limit the number of pages that can be
3408 * stored on pcp lists
3410 return min(READ_ONCE(pcp->batch) << 2, high);
3413 static void free_unref_page_commit(struct zone *zone, struct per_cpu_pages *pcp,
3414 struct page *page, int migratetype,
3421 __count_vm_events(PGFREE, 1 << order);
3422 pindex = order_to_pindex(migratetype, order);
3423 list_add(&page->pcp_list, &pcp->lists[pindex]);
3424 pcp->count += 1 << order;
3427 * As high-order pages other than THP's stored on PCP can contribute
3428 * to fragmentation, limit the number stored when PCP is heavily
3429 * freeing without allocation. The remainder after bulk freeing
3430 * stops will be drained from vmstat refresh context.
3432 free_high = (pcp->free_factor && order && order <= PAGE_ALLOC_COSTLY_ORDER);
3434 high = nr_pcp_high(pcp, zone, free_high);
3435 if (pcp->count >= high) {
3436 int batch = READ_ONCE(pcp->batch);
3438 free_pcppages_bulk(zone, nr_pcp_free(pcp, high, batch, free_high), pcp, pindex);
3445 void free_unref_page(struct page *page, unsigned int order)
3447 unsigned long __maybe_unused UP_flags;
3448 struct per_cpu_pages *pcp;
3450 unsigned long pfn = page_to_pfn(page);
3451 int migratetype, pcpmigratetype;
3453 if (!free_unref_page_prepare(page, pfn, order))
3457 * We only track unmovable, reclaimable and movable on pcp lists.
3458 * Place ISOLATE pages on the isolated list because they are being
3459 * offlined but treat HIGHATOMIC and CMA as movable pages so we can
3460 * get those areas back if necessary. Otherwise, we may have to free
3461 * excessively into the page allocator
3463 migratetype = pcpmigratetype = get_pcppage_migratetype(page);
3464 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
3465 if (unlikely(is_migrate_isolate(migratetype))) {
3466 free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
3469 pcpmigratetype = MIGRATE_MOVABLE;
3472 zone = page_zone(page);
3473 pcp_trylock_prepare(UP_flags);
3474 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
3476 free_unref_page_commit(zone, pcp, page, pcpmigratetype, order);
3477 pcp_spin_unlock(pcp);
3479 free_one_page(zone, page, pfn, order, migratetype, FPI_NONE);
3481 pcp_trylock_finish(UP_flags);
3485 * Free a list of 0-order pages
3487 void free_unref_page_list(struct list_head *list)
3489 unsigned long __maybe_unused UP_flags;
3490 struct page *page, *next;
3491 struct per_cpu_pages *pcp = NULL;
3492 struct zone *locked_zone = NULL;
3493 int batch_count = 0;
3496 /* Prepare pages for freeing */
3497 list_for_each_entry_safe(page, next, list, lru) {
3498 unsigned long pfn = page_to_pfn(page);
3499 if (!free_unref_page_prepare(page, pfn, 0)) {
3500 list_del(&page->lru);
3505 * Free isolated pages directly to the allocator, see
3506 * comment in free_unref_page.
3508 migratetype = get_pcppage_migratetype(page);
3509 if (unlikely(is_migrate_isolate(migratetype))) {
3510 list_del(&page->lru);
3511 free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
3516 list_for_each_entry_safe(page, next, list, lru) {
3517 struct zone *zone = page_zone(page);
3519 list_del(&page->lru);
3520 migratetype = get_pcppage_migratetype(page);
3522 /* Different zone, different pcp lock. */
3523 if (zone != locked_zone) {
3525 pcp_spin_unlock(pcp);
3526 pcp_trylock_finish(UP_flags);
3530 * trylock is necessary as pages may be getting freed
3531 * from IRQ or SoftIRQ context after an IO completion.
3533 pcp_trylock_prepare(UP_flags);
3534 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
3535 if (unlikely(!pcp)) {
3536 pcp_trylock_finish(UP_flags);
3537 free_one_page(zone, page, page_to_pfn(page),
3538 0, migratetype, FPI_NONE);
3547 * Non-isolated types over MIGRATE_PCPTYPES get added
3548 * to the MIGRATE_MOVABLE pcp list.
3550 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
3551 migratetype = MIGRATE_MOVABLE;
3553 trace_mm_page_free_batched(page);
3554 free_unref_page_commit(zone, pcp, page, migratetype, 0);
3557 * Guard against excessive lock hold times when freeing
3558 * a large list of pages. Lock will be reacquired if
3559 * necessary on the next iteration.
3561 if (++batch_count == SWAP_CLUSTER_MAX) {
3562 pcp_spin_unlock(pcp);
3563 pcp_trylock_finish(UP_flags);
3571 pcp_spin_unlock(pcp);
3572 pcp_trylock_finish(UP_flags);
3577 * split_page takes a non-compound higher-order page, and splits it into
3578 * n (1<<order) sub-pages: page[0..n]
3579 * Each sub-page must be freed individually.
3581 * Note: this is probably too low level an operation for use in drivers.
3582 * Please consult with lkml before using this in your driver.
3584 void split_page(struct page *page, unsigned int order)
3588 VM_BUG_ON_PAGE(PageCompound(page), page);
3589 VM_BUG_ON_PAGE(!page_count(page), page);
3591 for (i = 1; i < (1 << order); i++)
3592 set_page_refcounted(page + i);
3593 split_page_owner(page, 1 << order);
3594 split_page_memcg(page, 1 << order);
3596 EXPORT_SYMBOL_GPL(split_page);
3598 int __isolate_free_page(struct page *page, unsigned int order)
3600 struct zone *zone = page_zone(page);
3601 int mt = get_pageblock_migratetype(page);
3603 if (!is_migrate_isolate(mt)) {
3604 unsigned long watermark;
3606 * Obey watermarks as if the page was being allocated. We can
3607 * emulate a high-order watermark check with a raised order-0
3608 * watermark, because we already know our high-order page
3611 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3612 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3615 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3618 del_page_from_free_list(page, zone, order);
3621 * Set the pageblock if the isolated page is at least half of a
3624 if (order >= pageblock_order - 1) {
3625 struct page *endpage = page + (1 << order) - 1;
3626 for (; page < endpage; page += pageblock_nr_pages) {
3627 int mt = get_pageblock_migratetype(page);
3629 * Only change normal pageblocks (i.e., they can merge
3632 if (migratetype_is_mergeable(mt))
3633 set_pageblock_migratetype(page,
3638 return 1UL << order;
3642 * __putback_isolated_page - Return a now-isolated page back where we got it
3643 * @page: Page that was isolated
3644 * @order: Order of the isolated page
3645 * @mt: The page's pageblock's migratetype
3647 * This function is meant to return a page pulled from the free lists via
3648 * __isolate_free_page back to the free lists they were pulled from.
3650 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3652 struct zone *zone = page_zone(page);
3654 /* zone lock should be held when this function is called */
3655 lockdep_assert_held(&zone->lock);
3657 /* Return isolated page to tail of freelist. */
3658 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3659 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3663 * Update NUMA hit/miss statistics
3665 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
3669 enum numa_stat_item local_stat = NUMA_LOCAL;
3671 /* skip numa counters update if numa stats is disabled */
3672 if (!static_branch_likely(&vm_numa_stat_key))
3675 if (zone_to_nid(z) != numa_node_id())
3676 local_stat = NUMA_OTHER;
3678 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3679 __count_numa_events(z, NUMA_HIT, nr_account);
3681 __count_numa_events(z, NUMA_MISS, nr_account);
3682 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
3684 __count_numa_events(z, local_stat, nr_account);
3688 static __always_inline
3689 struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone,
3690 unsigned int order, unsigned int alloc_flags,
3694 unsigned long flags;
3698 spin_lock_irqsave(&zone->lock, flags);
3700 * order-0 request can reach here when the pcplist is skipped
3701 * due to non-CMA allocation context. HIGHATOMIC area is
3702 * reserved for high-order atomic allocation, so order-0
3703 * request should skip it.
3705 if (order > 0 && alloc_flags & ALLOC_HARDER)
3706 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3708 page = __rmqueue(zone, order, migratetype, alloc_flags);
3710 spin_unlock_irqrestore(&zone->lock, flags);
3714 __mod_zone_freepage_state(zone, -(1 << order),
3715 get_pcppage_migratetype(page));
3716 spin_unlock_irqrestore(&zone->lock, flags);
3717 } while (check_new_pages(page, order));
3719 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3720 zone_statistics(preferred_zone, zone, 1);
3725 /* Remove page from the per-cpu list, caller must protect the list */
3727 struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
3729 unsigned int alloc_flags,
3730 struct per_cpu_pages *pcp,
3731 struct list_head *list)
3736 if (list_empty(list)) {
3737 int batch = READ_ONCE(pcp->batch);
3741 * Scale batch relative to order if batch implies
3742 * free pages can be stored on the PCP. Batch can
3743 * be 1 for small zones or for boot pagesets which
3744 * should never store free pages as the pages may
3745 * belong to arbitrary zones.
3748 batch = max(batch >> order, 2);
3749 alloced = rmqueue_bulk(zone, order,
3751 migratetype, alloc_flags);
3753 pcp->count += alloced << order;
3754 if (unlikely(list_empty(list)))
3758 page = list_first_entry(list, struct page, pcp_list);
3759 list_del(&page->pcp_list);
3760 pcp->count -= 1 << order;
3761 } while (check_new_pcp(page, order));
3766 /* Lock and remove page from the per-cpu list */
3767 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3768 struct zone *zone, unsigned int order,
3769 int migratetype, unsigned int alloc_flags)
3771 struct per_cpu_pages *pcp;
3772 struct list_head *list;
3774 unsigned long __maybe_unused UP_flags;
3776 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
3777 pcp_trylock_prepare(UP_flags);
3778 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
3780 pcp_trylock_finish(UP_flags);
3785 * On allocation, reduce the number of pages that are batch freed.
3786 * See nr_pcp_free() where free_factor is increased for subsequent
3789 pcp->free_factor >>= 1;
3790 list = &pcp->lists[order_to_pindex(migratetype, order)];
3791 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
3792 pcp_spin_unlock(pcp);
3793 pcp_trylock_finish(UP_flags);
3795 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3796 zone_statistics(preferred_zone, zone, 1);
3802 * Allocate a page from the given zone.
3803 * Use pcplists for THP or "cheap" high-order allocations.
3807 * Do not instrument rmqueue() with KMSAN. This function may call
3808 * __msan_poison_alloca() through a call to set_pfnblock_flags_mask().
3809 * If __msan_poison_alloca() attempts to allocate pages for the stack depot, it
3810 * may call rmqueue() again, which will result in a deadlock.
3812 __no_sanitize_memory
3814 struct page *rmqueue(struct zone *preferred_zone,
3815 struct zone *zone, unsigned int order,
3816 gfp_t gfp_flags, unsigned int alloc_flags,
3822 * We most definitely don't want callers attempting to
3823 * allocate greater than order-1 page units with __GFP_NOFAIL.
3825 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3827 if (likely(pcp_allowed_order(order))) {
3829 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3830 * we need to skip it when CMA area isn't allowed.
3832 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3833 migratetype != MIGRATE_MOVABLE) {
3834 page = rmqueue_pcplist(preferred_zone, zone, order,
3835 migratetype, alloc_flags);
3841 page = rmqueue_buddy(preferred_zone, zone, order, alloc_flags,
3845 /* Separate test+clear to avoid unnecessary atomics */
3846 if (unlikely(test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags))) {
3847 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3848 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3851 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3855 #ifdef CONFIG_FAIL_PAGE_ALLOC
3858 struct fault_attr attr;
3860 bool ignore_gfp_highmem;
3861 bool ignore_gfp_reclaim;
3863 } fail_page_alloc = {
3864 .attr = FAULT_ATTR_INITIALIZER,
3865 .ignore_gfp_reclaim = true,
3866 .ignore_gfp_highmem = true,
3870 static int __init setup_fail_page_alloc(char *str)
3872 return setup_fault_attr(&fail_page_alloc.attr, str);
3874 __setup("fail_page_alloc=", setup_fail_page_alloc);
3876 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3880 if (order < fail_page_alloc.min_order)
3882 if (gfp_mask & __GFP_NOFAIL)
3884 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3886 if (fail_page_alloc.ignore_gfp_reclaim &&
3887 (gfp_mask & __GFP_DIRECT_RECLAIM))
3890 /* See comment in __should_failslab() */
3891 if (gfp_mask & __GFP_NOWARN)
3892 flags |= FAULT_NOWARN;
3894 return should_fail_ex(&fail_page_alloc.attr, 1 << order, flags);
3897 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3899 static int __init fail_page_alloc_debugfs(void)
3901 umode_t mode = S_IFREG | 0600;
3904 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3905 &fail_page_alloc.attr);
3907 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3908 &fail_page_alloc.ignore_gfp_reclaim);
3909 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3910 &fail_page_alloc.ignore_gfp_highmem);
3911 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3916 late_initcall(fail_page_alloc_debugfs);
3918 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3920 #else /* CONFIG_FAIL_PAGE_ALLOC */
3922 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3927 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3929 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3931 return __should_fail_alloc_page(gfp_mask, order);
3933 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3935 static inline long __zone_watermark_unusable_free(struct zone *z,
3936 unsigned int order, unsigned int alloc_flags)
3938 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3939 long unusable_free = (1 << order) - 1;
3942 * If the caller does not have rights to ALLOC_HARDER then subtract
3943 * the high-atomic reserves. This will over-estimate the size of the
3944 * atomic reserve but it avoids a search.
3946 if (likely(!alloc_harder))
3947 unusable_free += z->nr_reserved_highatomic;
3950 /* If allocation can't use CMA areas don't use free CMA pages */
3951 if (!(alloc_flags & ALLOC_CMA))
3952 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3955 return unusable_free;
3959 * Return true if free base pages are above 'mark'. For high-order checks it
3960 * will return true of the order-0 watermark is reached and there is at least
3961 * one free page of a suitable size. Checking now avoids taking the zone lock
3962 * to check in the allocation paths if no pages are free.
3964 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3965 int highest_zoneidx, unsigned int alloc_flags,
3970 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3972 /* free_pages may go negative - that's OK */
3973 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3975 if (alloc_flags & ALLOC_HIGH)
3978 if (unlikely(alloc_harder)) {
3980 * OOM victims can try even harder than normal ALLOC_HARDER
3981 * users on the grounds that it's definitely going to be in
3982 * the exit path shortly and free memory. Any allocation it
3983 * makes during the free path will be small and short-lived.
3985 if (alloc_flags & ALLOC_OOM)
3992 * Check watermarks for an order-0 allocation request. If these
3993 * are not met, then a high-order request also cannot go ahead
3994 * even if a suitable page happened to be free.
3996 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3999 /* If this is an order-0 request then the watermark is fine */
4003 /* For a high-order request, check at least one suitable page is free */
4004 for (o = order; o < MAX_ORDER; o++) {
4005 struct free_area *area = &z->free_area[o];
4011 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
4012 if (!free_area_empty(area, mt))
4017 if ((alloc_flags & ALLOC_CMA) &&
4018 !free_area_empty(area, MIGRATE_CMA)) {
4022 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
4028 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
4029 int highest_zoneidx, unsigned int alloc_flags)
4031 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
4032 zone_page_state(z, NR_FREE_PAGES));
4035 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
4036 unsigned long mark, int highest_zoneidx,
4037 unsigned int alloc_flags, gfp_t gfp_mask)
4041 free_pages = zone_page_state(z, NR_FREE_PAGES);
4044 * Fast check for order-0 only. If this fails then the reserves
4045 * need to be calculated.
4051 usable_free = free_pages;
4052 reserved = __zone_watermark_unusable_free(z, 0, alloc_flags);
4054 /* reserved may over estimate high-atomic reserves. */
4055 usable_free -= min(usable_free, reserved);
4056 if (usable_free > mark + z->lowmem_reserve[highest_zoneidx])
4060 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
4064 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
4065 * when checking the min watermark. The min watermark is the
4066 * point where boosting is ignored so that kswapd is woken up
4067 * when below the low watermark.
4069 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
4070 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
4071 mark = z->_watermark[WMARK_MIN];
4072 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
4073 alloc_flags, free_pages);
4079 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
4080 unsigned long mark, int highest_zoneidx)
4082 long free_pages = zone_page_state(z, NR_FREE_PAGES);
4084 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
4085 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
4087 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
4092 int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
4094 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
4096 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
4097 node_reclaim_distance;
4099 #else /* CONFIG_NUMA */
4100 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
4104 #endif /* CONFIG_NUMA */
4107 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
4108 * fragmentation is subtle. If the preferred zone was HIGHMEM then
4109 * premature use of a lower zone may cause lowmem pressure problems that
4110 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
4111 * probably too small. It only makes sense to spread allocations to avoid
4112 * fragmentation between the Normal and DMA32 zones.
4114 static inline unsigned int
4115 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
4117 unsigned int alloc_flags;
4120 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4123 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
4125 #ifdef CONFIG_ZONE_DMA32
4129 if (zone_idx(zone) != ZONE_NORMAL)
4133 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
4134 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
4135 * on UMA that if Normal is populated then so is DMA32.
4137 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
4138 if (nr_online_nodes > 1 && !populated_zone(--zone))
4141 alloc_flags |= ALLOC_NOFRAGMENT;
4142 #endif /* CONFIG_ZONE_DMA32 */
4146 /* Must be called after current_gfp_context() which can change gfp_mask */
4147 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
4148 unsigned int alloc_flags)
4151 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4152 alloc_flags |= ALLOC_CMA;
4158 * get_page_from_freelist goes through the zonelist trying to allocate
4161 static struct page *
4162 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
4163 const struct alloc_context *ac)
4167 struct pglist_data *last_pgdat = NULL;
4168 bool last_pgdat_dirty_ok = false;
4173 * Scan zonelist, looking for a zone with enough free.
4174 * See also __cpuset_node_allowed() comment in kernel/cgroup/cpuset.c.
4176 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
4177 z = ac->preferred_zoneref;
4178 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
4183 if (cpusets_enabled() &&
4184 (alloc_flags & ALLOC_CPUSET) &&
4185 !__cpuset_zone_allowed(zone, gfp_mask))
4188 * When allocating a page cache page for writing, we
4189 * want to get it from a node that is within its dirty
4190 * limit, such that no single node holds more than its
4191 * proportional share of globally allowed dirty pages.
4192 * The dirty limits take into account the node's
4193 * lowmem reserves and high watermark so that kswapd
4194 * should be able to balance it without having to
4195 * write pages from its LRU list.
4197 * XXX: For now, allow allocations to potentially
4198 * exceed the per-node dirty limit in the slowpath
4199 * (spread_dirty_pages unset) before going into reclaim,
4200 * which is important when on a NUMA setup the allowed
4201 * nodes are together not big enough to reach the
4202 * global limit. The proper fix for these situations
4203 * will require awareness of nodes in the
4204 * dirty-throttling and the flusher threads.
4206 if (ac->spread_dirty_pages) {
4207 if (last_pgdat != zone->zone_pgdat) {
4208 last_pgdat = zone->zone_pgdat;
4209 last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat);
4212 if (!last_pgdat_dirty_ok)
4216 if (no_fallback && nr_online_nodes > 1 &&
4217 zone != ac->preferred_zoneref->zone) {
4221 * If moving to a remote node, retry but allow
4222 * fragmenting fallbacks. Locality is more important
4223 * than fragmentation avoidance.
4225 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
4226 if (zone_to_nid(zone) != local_nid) {
4227 alloc_flags &= ~ALLOC_NOFRAGMENT;
4232 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
4233 if (!zone_watermark_fast(zone, order, mark,
4234 ac->highest_zoneidx, alloc_flags,
4238 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4240 * Watermark failed for this zone, but see if we can
4241 * grow this zone if it contains deferred pages.
4243 if (static_branch_unlikely(&deferred_pages)) {
4244 if (_deferred_grow_zone(zone, order))
4248 /* Checked here to keep the fast path fast */
4249 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
4250 if (alloc_flags & ALLOC_NO_WATERMARKS)
4253 if (!node_reclaim_enabled() ||
4254 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
4257 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
4259 case NODE_RECLAIM_NOSCAN:
4262 case NODE_RECLAIM_FULL:
4263 /* scanned but unreclaimable */
4266 /* did we reclaim enough */
4267 if (zone_watermark_ok(zone, order, mark,
4268 ac->highest_zoneidx, alloc_flags))
4276 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
4277 gfp_mask, alloc_flags, ac->migratetype);
4279 prep_new_page(page, order, gfp_mask, alloc_flags);
4282 * If this is a high-order atomic allocation then check
4283 * if the pageblock should be reserved for the future
4285 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
4286 reserve_highatomic_pageblock(page, zone, order);
4290 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4291 /* Try again if zone has deferred pages */
4292 if (static_branch_unlikely(&deferred_pages)) {
4293 if (_deferred_grow_zone(zone, order))
4301 * It's possible on a UMA machine to get through all zones that are
4302 * fragmented. If avoiding fragmentation, reset and try again.
4305 alloc_flags &= ~ALLOC_NOFRAGMENT;
4312 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4314 unsigned int filter = SHOW_MEM_FILTER_NODES;
4317 * This documents exceptions given to allocations in certain
4318 * contexts that are allowed to allocate outside current's set
4321 if (!(gfp_mask & __GFP_NOMEMALLOC))
4322 if (tsk_is_oom_victim(current) ||
4323 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4324 filter &= ~SHOW_MEM_FILTER_NODES;
4325 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4326 filter &= ~SHOW_MEM_FILTER_NODES;
4328 __show_mem(filter, nodemask, gfp_zone(gfp_mask));
4331 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4333 struct va_format vaf;
4335 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4337 if ((gfp_mask & __GFP_NOWARN) ||
4338 !__ratelimit(&nopage_rs) ||
4339 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
4342 va_start(args, fmt);
4345 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4346 current->comm, &vaf, gfp_mask, &gfp_mask,
4347 nodemask_pr_args(nodemask));
4350 cpuset_print_current_mems_allowed();
4353 warn_alloc_show_mem(gfp_mask, nodemask);
4356 static inline struct page *
4357 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4358 unsigned int alloc_flags,
4359 const struct alloc_context *ac)
4363 page = get_page_from_freelist(gfp_mask, order,
4364 alloc_flags|ALLOC_CPUSET, ac);
4366 * fallback to ignore cpuset restriction if our nodes
4370 page = get_page_from_freelist(gfp_mask, order,
4376 static inline struct page *
4377 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4378 const struct alloc_context *ac, unsigned long *did_some_progress)
4380 struct oom_control oc = {
4381 .zonelist = ac->zonelist,
4382 .nodemask = ac->nodemask,
4384 .gfp_mask = gfp_mask,
4389 *did_some_progress = 0;
4392 * Acquire the oom lock. If that fails, somebody else is
4393 * making progress for us.
4395 if (!mutex_trylock(&oom_lock)) {
4396 *did_some_progress = 1;
4397 schedule_timeout_uninterruptible(1);
4402 * Go through the zonelist yet one more time, keep very high watermark
4403 * here, this is only to catch a parallel oom killing, we must fail if
4404 * we're still under heavy pressure. But make sure that this reclaim
4405 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4406 * allocation which will never fail due to oom_lock already held.
4408 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4409 ~__GFP_DIRECT_RECLAIM, order,
4410 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4414 /* Coredumps can quickly deplete all memory reserves */
4415 if (current->flags & PF_DUMPCORE)
4417 /* The OOM killer will not help higher order allocs */
4418 if (order > PAGE_ALLOC_COSTLY_ORDER)
4421 * We have already exhausted all our reclaim opportunities without any
4422 * success so it is time to admit defeat. We will skip the OOM killer
4423 * because it is very likely that the caller has a more reasonable
4424 * fallback than shooting a random task.
4426 * The OOM killer may not free memory on a specific node.
4428 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4430 /* The OOM killer does not needlessly kill tasks for lowmem */
4431 if (ac->highest_zoneidx < ZONE_NORMAL)
4433 if (pm_suspended_storage())
4436 * XXX: GFP_NOFS allocations should rather fail than rely on
4437 * other request to make a forward progress.
4438 * We are in an unfortunate situation where out_of_memory cannot
4439 * do much for this context but let's try it to at least get
4440 * access to memory reserved if the current task is killed (see
4441 * out_of_memory). Once filesystems are ready to handle allocation
4442 * failures more gracefully we should just bail out here.
4445 /* Exhausted what can be done so it's blame time */
4446 if (out_of_memory(&oc) ||
4447 WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) {
4448 *did_some_progress = 1;
4451 * Help non-failing allocations by giving them access to memory
4454 if (gfp_mask & __GFP_NOFAIL)
4455 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4456 ALLOC_NO_WATERMARKS, ac);
4459 mutex_unlock(&oom_lock);
4464 * Maximum number of compaction retries with a progress before OOM
4465 * killer is consider as the only way to move forward.
4467 #define MAX_COMPACT_RETRIES 16
4469 #ifdef CONFIG_COMPACTION
4470 /* Try memory compaction for high-order allocations before reclaim */
4471 static struct page *
4472 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4473 unsigned int alloc_flags, const struct alloc_context *ac,
4474 enum compact_priority prio, enum compact_result *compact_result)
4476 struct page *page = NULL;
4477 unsigned long pflags;
4478 unsigned int noreclaim_flag;
4483 psi_memstall_enter(&pflags);
4484 delayacct_compact_start();
4485 noreclaim_flag = memalloc_noreclaim_save();
4487 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4490 memalloc_noreclaim_restore(noreclaim_flag);
4491 psi_memstall_leave(&pflags);
4492 delayacct_compact_end();
4494 if (*compact_result == COMPACT_SKIPPED)
4497 * At least in one zone compaction wasn't deferred or skipped, so let's
4498 * count a compaction stall
4500 count_vm_event(COMPACTSTALL);
4502 /* Prep a captured page if available */
4504 prep_new_page(page, order, gfp_mask, alloc_flags);
4506 /* Try get a page from the freelist if available */
4508 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4511 struct zone *zone = page_zone(page);
4513 zone->compact_blockskip_flush = false;
4514 compaction_defer_reset(zone, order, true);
4515 count_vm_event(COMPACTSUCCESS);
4520 * It's bad if compaction run occurs and fails. The most likely reason
4521 * is that pages exist, but not enough to satisfy watermarks.
4523 count_vm_event(COMPACTFAIL);
4531 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4532 enum compact_result compact_result,
4533 enum compact_priority *compact_priority,
4534 int *compaction_retries)
4536 int max_retries = MAX_COMPACT_RETRIES;
4539 int retries = *compaction_retries;
4540 enum compact_priority priority = *compact_priority;
4545 if (fatal_signal_pending(current))
4548 if (compaction_made_progress(compact_result))
4549 (*compaction_retries)++;
4552 * compaction considers all the zone as desperately out of memory
4553 * so it doesn't really make much sense to retry except when the
4554 * failure could be caused by insufficient priority
4556 if (compaction_failed(compact_result))
4557 goto check_priority;
4560 * compaction was skipped because there are not enough order-0 pages
4561 * to work with, so we retry only if it looks like reclaim can help.
4563 if (compaction_needs_reclaim(compact_result)) {
4564 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4569 * make sure the compaction wasn't deferred or didn't bail out early
4570 * due to locks contention before we declare that we should give up.
4571 * But the next retry should use a higher priority if allowed, so
4572 * we don't just keep bailing out endlessly.
4574 if (compaction_withdrawn(compact_result)) {
4575 goto check_priority;
4579 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4580 * costly ones because they are de facto nofail and invoke OOM
4581 * killer to move on while costly can fail and users are ready
4582 * to cope with that. 1/4 retries is rather arbitrary but we
4583 * would need much more detailed feedback from compaction to
4584 * make a better decision.
4586 if (order > PAGE_ALLOC_COSTLY_ORDER)
4588 if (*compaction_retries <= max_retries) {
4594 * Make sure there are attempts at the highest priority if we exhausted
4595 * all retries or failed at the lower priorities.
4598 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4599 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4601 if (*compact_priority > min_priority) {
4602 (*compact_priority)--;
4603 *compaction_retries = 0;
4607 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4611 static inline struct page *
4612 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4613 unsigned int alloc_flags, const struct alloc_context *ac,
4614 enum compact_priority prio, enum compact_result *compact_result)
4616 *compact_result = COMPACT_SKIPPED;
4621 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4622 enum compact_result compact_result,
4623 enum compact_priority *compact_priority,
4624 int *compaction_retries)
4629 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4633 * There are setups with compaction disabled which would prefer to loop
4634 * inside the allocator rather than hit the oom killer prematurely.
4635 * Let's give them a good hope and keep retrying while the order-0
4636 * watermarks are OK.
4638 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4639 ac->highest_zoneidx, ac->nodemask) {
4640 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4641 ac->highest_zoneidx, alloc_flags))
4646 #endif /* CONFIG_COMPACTION */
4648 #ifdef CONFIG_LOCKDEP
4649 static struct lockdep_map __fs_reclaim_map =
4650 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4652 static bool __need_reclaim(gfp_t gfp_mask)
4654 /* no reclaim without waiting on it */
4655 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4658 /* this guy won't enter reclaim */
4659 if (current->flags & PF_MEMALLOC)
4662 if (gfp_mask & __GFP_NOLOCKDEP)
4668 void __fs_reclaim_acquire(unsigned long ip)
4670 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
4673 void __fs_reclaim_release(unsigned long ip)
4675 lock_release(&__fs_reclaim_map, ip);
4678 void fs_reclaim_acquire(gfp_t gfp_mask)
4680 gfp_mask = current_gfp_context(gfp_mask);
4682 if (__need_reclaim(gfp_mask)) {
4683 if (gfp_mask & __GFP_FS)
4684 __fs_reclaim_acquire(_RET_IP_);
4686 #ifdef CONFIG_MMU_NOTIFIER
4687 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4688 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4693 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4695 void fs_reclaim_release(gfp_t gfp_mask)
4697 gfp_mask = current_gfp_context(gfp_mask);
4699 if (__need_reclaim(gfp_mask)) {
4700 if (gfp_mask & __GFP_FS)
4701 __fs_reclaim_release(_RET_IP_);
4704 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4708 * Zonelists may change due to hotplug during allocation. Detect when zonelists
4709 * have been rebuilt so allocation retries. Reader side does not lock and
4710 * retries the allocation if zonelist changes. Writer side is protected by the
4711 * embedded spin_lock.
4713 static DEFINE_SEQLOCK(zonelist_update_seq);
4715 static unsigned int zonelist_iter_begin(void)
4717 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
4718 return read_seqbegin(&zonelist_update_seq);
4723 static unsigned int check_retry_zonelist(unsigned int seq)
4725 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
4726 return read_seqretry(&zonelist_update_seq, seq);
4731 /* Perform direct synchronous page reclaim */
4732 static unsigned long
4733 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4734 const struct alloc_context *ac)
4736 unsigned int noreclaim_flag;
4737 unsigned long progress;
4741 /* We now go into synchronous reclaim */
4742 cpuset_memory_pressure_bump();
4743 fs_reclaim_acquire(gfp_mask);
4744 noreclaim_flag = memalloc_noreclaim_save();
4746 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4749 memalloc_noreclaim_restore(noreclaim_flag);
4750 fs_reclaim_release(gfp_mask);
4757 /* The really slow allocator path where we enter direct reclaim */
4758 static inline struct page *
4759 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4760 unsigned int alloc_flags, const struct alloc_context *ac,
4761 unsigned long *did_some_progress)
4763 struct page *page = NULL;
4764 unsigned long pflags;
4765 bool drained = false;
4767 psi_memstall_enter(&pflags);
4768 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4769 if (unlikely(!(*did_some_progress)))
4773 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4776 * If an allocation failed after direct reclaim, it could be because
4777 * pages are pinned on the per-cpu lists or in high alloc reserves.
4778 * Shrink them and try again
4780 if (!page && !drained) {
4781 unreserve_highatomic_pageblock(ac, false);
4782 drain_all_pages(NULL);
4787 psi_memstall_leave(&pflags);
4792 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4793 const struct alloc_context *ac)
4797 pg_data_t *last_pgdat = NULL;
4798 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4800 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4802 if (!managed_zone(zone))
4804 if (last_pgdat != zone->zone_pgdat) {
4805 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4806 last_pgdat = zone->zone_pgdat;
4811 static inline unsigned int
4812 gfp_to_alloc_flags(gfp_t gfp_mask)
4814 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4817 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4818 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4819 * to save two branches.
4821 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4822 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4825 * The caller may dip into page reserves a bit more if the caller
4826 * cannot run direct reclaim, or if the caller has realtime scheduling
4827 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4828 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4830 alloc_flags |= (__force int)
4831 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4833 if (gfp_mask & __GFP_ATOMIC) {
4835 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4836 * if it can't schedule.
4838 if (!(gfp_mask & __GFP_NOMEMALLOC))
4839 alloc_flags |= ALLOC_HARDER;
4841 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4842 * comment for __cpuset_node_allowed().
4844 alloc_flags &= ~ALLOC_CPUSET;
4845 } else if (unlikely(rt_task(current)) && in_task())
4846 alloc_flags |= ALLOC_HARDER;
4848 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4853 static bool oom_reserves_allowed(struct task_struct *tsk)
4855 if (!tsk_is_oom_victim(tsk))
4859 * !MMU doesn't have oom reaper so give access to memory reserves
4860 * only to the thread with TIF_MEMDIE set
4862 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4869 * Distinguish requests which really need access to full memory
4870 * reserves from oom victims which can live with a portion of it
4872 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4874 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4876 if (gfp_mask & __GFP_MEMALLOC)
4877 return ALLOC_NO_WATERMARKS;
4878 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4879 return ALLOC_NO_WATERMARKS;
4880 if (!in_interrupt()) {
4881 if (current->flags & PF_MEMALLOC)
4882 return ALLOC_NO_WATERMARKS;
4883 else if (oom_reserves_allowed(current))
4890 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4892 return !!__gfp_pfmemalloc_flags(gfp_mask);
4896 * Checks whether it makes sense to retry the reclaim to make a forward progress
4897 * for the given allocation request.
4899 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4900 * without success, or when we couldn't even meet the watermark if we
4901 * reclaimed all remaining pages on the LRU lists.
4903 * Returns true if a retry is viable or false to enter the oom path.
4906 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4907 struct alloc_context *ac, int alloc_flags,
4908 bool did_some_progress, int *no_progress_loops)
4915 * Costly allocations might have made a progress but this doesn't mean
4916 * their order will become available due to high fragmentation so
4917 * always increment the no progress counter for them
4919 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4920 *no_progress_loops = 0;
4922 (*no_progress_loops)++;
4925 * Make sure we converge to OOM if we cannot make any progress
4926 * several times in the row.
4928 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4929 /* Before OOM, exhaust highatomic_reserve */
4930 return unreserve_highatomic_pageblock(ac, true);
4934 * Keep reclaiming pages while there is a chance this will lead
4935 * somewhere. If none of the target zones can satisfy our allocation
4936 * request even if all reclaimable pages are considered then we are
4937 * screwed and have to go OOM.
4939 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4940 ac->highest_zoneidx, ac->nodemask) {
4941 unsigned long available;
4942 unsigned long reclaimable;
4943 unsigned long min_wmark = min_wmark_pages(zone);
4946 available = reclaimable = zone_reclaimable_pages(zone);
4947 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4950 * Would the allocation succeed if we reclaimed all
4951 * reclaimable pages?
4953 wmark = __zone_watermark_ok(zone, order, min_wmark,
4954 ac->highest_zoneidx, alloc_flags, available);
4955 trace_reclaim_retry_zone(z, order, reclaimable,
4956 available, min_wmark, *no_progress_loops, wmark);
4964 * Memory allocation/reclaim might be called from a WQ context and the
4965 * current implementation of the WQ concurrency control doesn't
4966 * recognize that a particular WQ is congested if the worker thread is
4967 * looping without ever sleeping. Therefore we have to do a short sleep
4968 * here rather than calling cond_resched().
4970 if (current->flags & PF_WQ_WORKER)
4971 schedule_timeout_uninterruptible(1);
4978 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4981 * It's possible that cpuset's mems_allowed and the nodemask from
4982 * mempolicy don't intersect. This should be normally dealt with by
4983 * policy_nodemask(), but it's possible to race with cpuset update in
4984 * such a way the check therein was true, and then it became false
4985 * before we got our cpuset_mems_cookie here.
4986 * This assumes that for all allocations, ac->nodemask can come only
4987 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4988 * when it does not intersect with the cpuset restrictions) or the
4989 * caller can deal with a violated nodemask.
4991 if (cpusets_enabled() && ac->nodemask &&
4992 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4993 ac->nodemask = NULL;
4998 * When updating a task's mems_allowed or mempolicy nodemask, it is
4999 * possible to race with parallel threads in such a way that our
5000 * allocation can fail while the mask is being updated. If we are about
5001 * to fail, check if the cpuset changed during allocation and if so,
5004 if (read_mems_allowed_retry(cpuset_mems_cookie))
5010 static inline struct page *
5011 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
5012 struct alloc_context *ac)
5014 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
5015 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
5016 struct page *page = NULL;
5017 unsigned int alloc_flags;
5018 unsigned long did_some_progress;
5019 enum compact_priority compact_priority;
5020 enum compact_result compact_result;
5021 int compaction_retries;
5022 int no_progress_loops;
5023 unsigned int cpuset_mems_cookie;
5024 unsigned int zonelist_iter_cookie;
5028 * We also sanity check to catch abuse of atomic reserves being used by
5029 * callers that are not in atomic context.
5031 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
5032 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
5033 gfp_mask &= ~__GFP_ATOMIC;
5036 compaction_retries = 0;
5037 no_progress_loops = 0;
5038 compact_priority = DEF_COMPACT_PRIORITY;
5039 cpuset_mems_cookie = read_mems_allowed_begin();
5040 zonelist_iter_cookie = zonelist_iter_begin();
5043 * The fast path uses conservative alloc_flags to succeed only until
5044 * kswapd needs to be woken up, and to avoid the cost of setting up
5045 * alloc_flags precisely. So we do that now.
5047 alloc_flags = gfp_to_alloc_flags(gfp_mask);
5050 * We need to recalculate the starting point for the zonelist iterator
5051 * because we might have used different nodemask in the fast path, or
5052 * there was a cpuset modification and we are retrying - otherwise we
5053 * could end up iterating over non-eligible zones endlessly.
5055 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5056 ac->highest_zoneidx, ac->nodemask);
5057 if (!ac->preferred_zoneref->zone)
5061 * Check for insane configurations where the cpuset doesn't contain
5062 * any suitable zone to satisfy the request - e.g. non-movable
5063 * GFP_HIGHUSER allocations from MOVABLE nodes only.
5065 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
5066 struct zoneref *z = first_zones_zonelist(ac->zonelist,
5067 ac->highest_zoneidx,
5068 &cpuset_current_mems_allowed);
5073 if (alloc_flags & ALLOC_KSWAPD)
5074 wake_all_kswapds(order, gfp_mask, ac);
5077 * The adjusted alloc_flags might result in immediate success, so try
5080 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
5085 * For costly allocations, try direct compaction first, as it's likely
5086 * that we have enough base pages and don't need to reclaim. For non-
5087 * movable high-order allocations, do that as well, as compaction will
5088 * try prevent permanent fragmentation by migrating from blocks of the
5090 * Don't try this for allocations that are allowed to ignore
5091 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
5093 if (can_direct_reclaim &&
5095 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
5096 && !gfp_pfmemalloc_allowed(gfp_mask)) {
5097 page = __alloc_pages_direct_compact(gfp_mask, order,
5099 INIT_COMPACT_PRIORITY,
5105 * Checks for costly allocations with __GFP_NORETRY, which
5106 * includes some THP page fault allocations
5108 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
5110 * If allocating entire pageblock(s) and compaction
5111 * failed because all zones are below low watermarks
5112 * or is prohibited because it recently failed at this
5113 * order, fail immediately unless the allocator has
5114 * requested compaction and reclaim retry.
5117 * - potentially very expensive because zones are far
5118 * below their low watermarks or this is part of very
5119 * bursty high order allocations,
5120 * - not guaranteed to help because isolate_freepages()
5121 * may not iterate over freed pages as part of its
5123 * - unlikely to make entire pageblocks free on its
5126 if (compact_result == COMPACT_SKIPPED ||
5127 compact_result == COMPACT_DEFERRED)
5131 * Looks like reclaim/compaction is worth trying, but
5132 * sync compaction could be very expensive, so keep
5133 * using async compaction.
5135 compact_priority = INIT_COMPACT_PRIORITY;
5140 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
5141 if (alloc_flags & ALLOC_KSWAPD)
5142 wake_all_kswapds(order, gfp_mask, ac);
5144 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
5146 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) |
5147 (alloc_flags & ALLOC_KSWAPD);
5150 * Reset the nodemask and zonelist iterators if memory policies can be
5151 * ignored. These allocations are high priority and system rather than
5154 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
5155 ac->nodemask = NULL;
5156 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5157 ac->highest_zoneidx, ac->nodemask);
5160 /* Attempt with potentially adjusted zonelist and alloc_flags */
5161 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
5165 /* Caller is not willing to reclaim, we can't balance anything */
5166 if (!can_direct_reclaim)
5169 /* Avoid recursion of direct reclaim */
5170 if (current->flags & PF_MEMALLOC)
5173 /* Try direct reclaim and then allocating */
5174 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
5175 &did_some_progress);
5179 /* Try direct compaction and then allocating */
5180 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
5181 compact_priority, &compact_result);
5185 /* Do not loop if specifically requested */
5186 if (gfp_mask & __GFP_NORETRY)
5190 * Do not retry costly high order allocations unless they are
5191 * __GFP_RETRY_MAYFAIL
5193 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
5196 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
5197 did_some_progress > 0, &no_progress_loops))
5201 * It doesn't make any sense to retry for the compaction if the order-0
5202 * reclaim is not able to make any progress because the current
5203 * implementation of the compaction depends on the sufficient amount
5204 * of free memory (see __compaction_suitable)
5206 if (did_some_progress > 0 &&
5207 should_compact_retry(ac, order, alloc_flags,
5208 compact_result, &compact_priority,
5209 &compaction_retries))
5214 * Deal with possible cpuset update races or zonelist updates to avoid
5215 * a unnecessary OOM kill.
5217 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
5218 check_retry_zonelist(zonelist_iter_cookie))
5221 /* Reclaim has failed us, start killing things */
5222 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
5226 /* Avoid allocations with no watermarks from looping endlessly */
5227 if (tsk_is_oom_victim(current) &&
5228 (alloc_flags & ALLOC_OOM ||
5229 (gfp_mask & __GFP_NOMEMALLOC)))
5232 /* Retry as long as the OOM killer is making progress */
5233 if (did_some_progress) {
5234 no_progress_loops = 0;
5240 * Deal with possible cpuset update races or zonelist updates to avoid
5241 * a unnecessary OOM kill.
5243 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
5244 check_retry_zonelist(zonelist_iter_cookie))
5248 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
5251 if (gfp_mask & __GFP_NOFAIL) {
5253 * All existing users of the __GFP_NOFAIL are blockable, so warn
5254 * of any new users that actually require GFP_NOWAIT
5256 if (WARN_ON_ONCE_GFP(!can_direct_reclaim, gfp_mask))
5260 * PF_MEMALLOC request from this context is rather bizarre
5261 * because we cannot reclaim anything and only can loop waiting
5262 * for somebody to do a work for us
5264 WARN_ON_ONCE_GFP(current->flags & PF_MEMALLOC, gfp_mask);
5267 * non failing costly orders are a hard requirement which we
5268 * are not prepared for much so let's warn about these users
5269 * so that we can identify them and convert them to something
5272 WARN_ON_ONCE_GFP(costly_order, gfp_mask);
5275 * Help non-failing allocations by giving them access to memory
5276 * reserves but do not use ALLOC_NO_WATERMARKS because this
5277 * could deplete whole memory reserves which would just make
5278 * the situation worse
5280 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
5288 warn_alloc(gfp_mask, ac->nodemask,
5289 "page allocation failure: order:%u", order);
5294 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
5295 int preferred_nid, nodemask_t *nodemask,
5296 struct alloc_context *ac, gfp_t *alloc_gfp,
5297 unsigned int *alloc_flags)
5299 ac->highest_zoneidx = gfp_zone(gfp_mask);
5300 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
5301 ac->nodemask = nodemask;
5302 ac->migratetype = gfp_migratetype(gfp_mask);
5304 if (cpusets_enabled()) {
5305 *alloc_gfp |= __GFP_HARDWALL;
5307 * When we are in the interrupt context, it is irrelevant
5308 * to the current task context. It means that any node ok.
5310 if (in_task() && !ac->nodemask)
5311 ac->nodemask = &cpuset_current_mems_allowed;
5313 *alloc_flags |= ALLOC_CPUSET;
5316 might_alloc(gfp_mask);
5318 if (should_fail_alloc_page(gfp_mask, order))
5321 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
5323 /* Dirty zone balancing only done in the fast path */
5324 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5327 * The preferred zone is used for statistics but crucially it is
5328 * also used as the starting point for the zonelist iterator. It
5329 * may get reset for allocations that ignore memory policies.
5331 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5332 ac->highest_zoneidx, ac->nodemask);
5338 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
5339 * @gfp: GFP flags for the allocation
5340 * @preferred_nid: The preferred NUMA node ID to allocate from
5341 * @nodemask: Set of nodes to allocate from, may be NULL
5342 * @nr_pages: The number of pages desired on the list or array
5343 * @page_list: Optional list to store the allocated pages
5344 * @page_array: Optional array to store the pages
5346 * This is a batched version of the page allocator that attempts to
5347 * allocate nr_pages quickly. Pages are added to page_list if page_list
5348 * is not NULL, otherwise it is assumed that the page_array is valid.
5350 * For lists, nr_pages is the number of pages that should be allocated.
5352 * For arrays, only NULL elements are populated with pages and nr_pages
5353 * is the maximum number of pages that will be stored in the array.
5355 * Returns the number of pages on the list or array.
5357 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
5358 nodemask_t *nodemask, int nr_pages,
5359 struct list_head *page_list,
5360 struct page **page_array)
5363 unsigned long __maybe_unused UP_flags;
5366 struct per_cpu_pages *pcp;
5367 struct list_head *pcp_list;
5368 struct alloc_context ac;
5370 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5371 int nr_populated = 0, nr_account = 0;
5374 * Skip populated array elements to determine if any pages need
5375 * to be allocated before disabling IRQs.
5377 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
5380 /* No pages requested? */
5381 if (unlikely(nr_pages <= 0))
5384 /* Already populated array? */
5385 if (unlikely(page_array && nr_pages - nr_populated == 0))
5388 /* Bulk allocator does not support memcg accounting. */
5389 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT))
5392 /* Use the single page allocator for one page. */
5393 if (nr_pages - nr_populated == 1)
5396 #ifdef CONFIG_PAGE_OWNER
5398 * PAGE_OWNER may recurse into the allocator to allocate space to
5399 * save the stack with pagesets.lock held. Releasing/reacquiring
5400 * removes much of the performance benefit of bulk allocation so
5401 * force the caller to allocate one page at a time as it'll have
5402 * similar performance to added complexity to the bulk allocator.
5404 if (static_branch_unlikely(&page_owner_inited))
5408 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5409 gfp &= gfp_allowed_mask;
5411 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
5415 /* Find an allowed local zone that meets the low watermark. */
5416 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
5419 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5420 !__cpuset_zone_allowed(zone, gfp)) {
5424 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
5425 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
5429 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
5430 if (zone_watermark_fast(zone, 0, mark,
5431 zonelist_zone_idx(ac.preferred_zoneref),
5432 alloc_flags, gfp)) {
5438 * If there are no allowed local zones that meets the watermarks then
5439 * try to allocate a single page and reclaim if necessary.
5441 if (unlikely(!zone))
5444 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
5445 pcp_trylock_prepare(UP_flags);
5446 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
5450 /* Attempt the batch allocation */
5451 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
5452 while (nr_populated < nr_pages) {
5454 /* Skip existing pages */
5455 if (page_array && page_array[nr_populated]) {
5460 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
5462 if (unlikely(!page)) {
5463 /* Try and allocate at least one page */
5465 pcp_spin_unlock(pcp);
5472 prep_new_page(page, 0, gfp, 0);
5474 list_add(&page->lru, page_list);
5476 page_array[nr_populated] = page;
5480 pcp_spin_unlock(pcp);
5481 pcp_trylock_finish(UP_flags);
5483 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
5484 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
5487 return nr_populated;
5490 pcp_trylock_finish(UP_flags);
5493 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
5496 list_add(&page->lru, page_list);
5498 page_array[nr_populated] = page;
5504 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
5507 * This is the 'heart' of the zoned buddy allocator.
5509 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
5510 nodemask_t *nodemask)
5513 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5514 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5515 struct alloc_context ac = { };
5518 * There are several places where we assume that the order value is sane
5519 * so bail out early if the request is out of bound.
5521 if (WARN_ON_ONCE_GFP(order >= MAX_ORDER, gfp))
5524 gfp &= gfp_allowed_mask;
5526 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5527 * resp. GFP_NOIO which has to be inherited for all allocation requests
5528 * from a particular context which has been marked by
5529 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5530 * movable zones are not used during allocation.
5532 gfp = current_gfp_context(gfp);
5534 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5535 &alloc_gfp, &alloc_flags))
5539 * Forbid the first pass from falling back to types that fragment
5540 * memory until all local zones are considered.
5542 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5544 /* First allocation attempt */
5545 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5550 ac.spread_dirty_pages = false;
5553 * Restore the original nodemask if it was potentially replaced with
5554 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5556 ac.nodemask = nodemask;
5558 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5561 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5562 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5563 __free_pages(page, order);
5567 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5568 kmsan_alloc_page(page, order, alloc_gfp);
5572 EXPORT_SYMBOL(__alloc_pages);
5574 struct folio *__folio_alloc(gfp_t gfp, unsigned int order, int preferred_nid,
5575 nodemask_t *nodemask)
5577 struct page *page = __alloc_pages(gfp | __GFP_COMP, order,
5578 preferred_nid, nodemask);
5580 if (page && order > 1)
5581 prep_transhuge_page(page);
5582 return (struct folio *)page;
5584 EXPORT_SYMBOL(__folio_alloc);
5587 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5588 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5589 * you need to access high mem.
5591 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5595 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5598 return (unsigned long) page_address(page);
5600 EXPORT_SYMBOL(__get_free_pages);
5602 unsigned long get_zeroed_page(gfp_t gfp_mask)
5604 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5606 EXPORT_SYMBOL(get_zeroed_page);
5609 * __free_pages - Free pages allocated with alloc_pages().
5610 * @page: The page pointer returned from alloc_pages().
5611 * @order: The order of the allocation.
5613 * This function can free multi-page allocations that are not compound
5614 * pages. It does not check that the @order passed in matches that of
5615 * the allocation, so it is easy to leak memory. Freeing more memory
5616 * than was allocated will probably emit a warning.
5618 * If the last reference to this page is speculative, it will be released
5619 * by put_page() which only frees the first page of a non-compound
5620 * allocation. To prevent the remaining pages from being leaked, we free
5621 * the subsequent pages here. If you want to use the page's reference
5622 * count to decide when to free the allocation, you should allocate a
5623 * compound page, and use put_page() instead of __free_pages().
5625 * Context: May be called in interrupt context or while holding a normal
5626 * spinlock, but not in NMI context or while holding a raw spinlock.
5628 void __free_pages(struct page *page, unsigned int order)
5630 /* get PageHead before we drop reference */
5631 int head = PageHead(page);
5633 if (put_page_testzero(page))
5634 free_the_page(page, order);
5637 free_the_page(page + (1 << order), order);
5639 EXPORT_SYMBOL(__free_pages);
5641 void free_pages(unsigned long addr, unsigned int order)
5644 VM_BUG_ON(!virt_addr_valid((void *)addr));
5645 __free_pages(virt_to_page((void *)addr), order);
5649 EXPORT_SYMBOL(free_pages);
5653 * An arbitrary-length arbitrary-offset area of memory which resides
5654 * within a 0 or higher order page. Multiple fragments within that page
5655 * are individually refcounted, in the page's reference counter.
5657 * The page_frag functions below provide a simple allocation framework for
5658 * page fragments. This is used by the network stack and network device
5659 * drivers to provide a backing region of memory for use as either an
5660 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5662 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5665 struct page *page = NULL;
5666 gfp_t gfp = gfp_mask;
5668 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5669 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5671 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5672 PAGE_FRAG_CACHE_MAX_ORDER);
5673 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5675 if (unlikely(!page))
5676 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5678 nc->va = page ? page_address(page) : NULL;
5683 void __page_frag_cache_drain(struct page *page, unsigned int count)
5685 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5687 if (page_ref_sub_and_test(page, count))
5688 free_the_page(page, compound_order(page));
5690 EXPORT_SYMBOL(__page_frag_cache_drain);
5692 void *page_frag_alloc_align(struct page_frag_cache *nc,
5693 unsigned int fragsz, gfp_t gfp_mask,
5694 unsigned int align_mask)
5696 unsigned int size = PAGE_SIZE;
5700 if (unlikely(!nc->va)) {
5702 page = __page_frag_cache_refill(nc, gfp_mask);
5706 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5707 /* if size can vary use size else just use PAGE_SIZE */
5710 /* Even if we own the page, we do not use atomic_set().
5711 * This would break get_page_unless_zero() users.
5713 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5715 /* reset page count bias and offset to start of new frag */
5716 nc->pfmemalloc = page_is_pfmemalloc(page);
5717 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5721 offset = nc->offset - fragsz;
5722 if (unlikely(offset < 0)) {
5723 page = virt_to_page(nc->va);
5725 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5728 if (unlikely(nc->pfmemalloc)) {
5729 free_the_page(page, compound_order(page));
5733 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5734 /* if size can vary use size else just use PAGE_SIZE */
5737 /* OK, page count is 0, we can safely set it */
5738 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5740 /* reset page count bias and offset to start of new frag */
5741 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5742 offset = size - fragsz;
5743 if (unlikely(offset < 0)) {
5745 * The caller is trying to allocate a fragment
5746 * with fragsz > PAGE_SIZE but the cache isn't big
5747 * enough to satisfy the request, this may
5748 * happen in low memory conditions.
5749 * We don't release the cache page because
5750 * it could make memory pressure worse
5751 * so we simply return NULL here.
5758 offset &= align_mask;
5759 nc->offset = offset;
5761 return nc->va + offset;
5763 EXPORT_SYMBOL(page_frag_alloc_align);
5766 * Frees a page fragment allocated out of either a compound or order 0 page.
5768 void page_frag_free(void *addr)
5770 struct page *page = virt_to_head_page(addr);
5772 if (unlikely(put_page_testzero(page)))
5773 free_the_page(page, compound_order(page));
5775 EXPORT_SYMBOL(page_frag_free);
5777 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5781 unsigned long nr = DIV_ROUND_UP(size, PAGE_SIZE);
5782 struct page *page = virt_to_page((void *)addr);
5783 struct page *last = page + nr;
5785 split_page_owner(page, 1 << order);
5786 split_page_memcg(page, 1 << order);
5787 while (page < --last)
5788 set_page_refcounted(last);
5790 last = page + (1UL << order);
5791 for (page += nr; page < last; page++)
5792 __free_pages_ok(page, 0, FPI_TO_TAIL);
5794 return (void *)addr;
5798 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5799 * @size: the number of bytes to allocate
5800 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5802 * This function is similar to alloc_pages(), except that it allocates the
5803 * minimum number of pages to satisfy the request. alloc_pages() can only
5804 * allocate memory in power-of-two pages.
5806 * This function is also limited by MAX_ORDER.
5808 * Memory allocated by this function must be released by free_pages_exact().
5810 * Return: pointer to the allocated area or %NULL in case of error.
5812 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5814 unsigned int order = get_order(size);
5817 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5818 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5820 addr = __get_free_pages(gfp_mask, order);
5821 return make_alloc_exact(addr, order, size);
5823 EXPORT_SYMBOL(alloc_pages_exact);
5826 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5828 * @nid: the preferred node ID where memory should be allocated
5829 * @size: the number of bytes to allocate
5830 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5832 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5835 * Return: pointer to the allocated area or %NULL in case of error.
5837 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5839 unsigned int order = get_order(size);
5842 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5843 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5845 p = alloc_pages_node(nid, gfp_mask, order);
5848 return make_alloc_exact((unsigned long)page_address(p), order, size);
5852 * free_pages_exact - release memory allocated via alloc_pages_exact()
5853 * @virt: the value returned by alloc_pages_exact.
5854 * @size: size of allocation, same value as passed to alloc_pages_exact().
5856 * Release the memory allocated by a previous call to alloc_pages_exact.
5858 void free_pages_exact(void *virt, size_t size)
5860 unsigned long addr = (unsigned long)virt;
5861 unsigned long end = addr + PAGE_ALIGN(size);
5863 while (addr < end) {
5868 EXPORT_SYMBOL(free_pages_exact);
5871 * nr_free_zone_pages - count number of pages beyond high watermark
5872 * @offset: The zone index of the highest zone
5874 * nr_free_zone_pages() counts the number of pages which are beyond the
5875 * high watermark within all zones at or below a given zone index. For each
5876 * zone, the number of pages is calculated as:
5878 * nr_free_zone_pages = managed_pages - high_pages
5880 * Return: number of pages beyond high watermark.
5882 static unsigned long nr_free_zone_pages(int offset)
5887 /* Just pick one node, since fallback list is circular */
5888 unsigned long sum = 0;
5890 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5892 for_each_zone_zonelist(zone, z, zonelist, offset) {
5893 unsigned long size = zone_managed_pages(zone);
5894 unsigned long high = high_wmark_pages(zone);
5903 * nr_free_buffer_pages - count number of pages beyond high watermark
5905 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5906 * watermark within ZONE_DMA and ZONE_NORMAL.
5908 * Return: number of pages beyond high watermark within ZONE_DMA and
5911 unsigned long nr_free_buffer_pages(void)
5913 return nr_free_zone_pages(gfp_zone(GFP_USER));
5915 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5917 static inline void show_node(struct zone *zone)
5919 if (IS_ENABLED(CONFIG_NUMA))
5920 printk("Node %d ", zone_to_nid(zone));
5923 long si_mem_available(void)
5926 unsigned long pagecache;
5927 unsigned long wmark_low = 0;
5928 unsigned long pages[NR_LRU_LISTS];
5929 unsigned long reclaimable;
5933 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5934 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5937 wmark_low += low_wmark_pages(zone);
5940 * Estimate the amount of memory available for userspace allocations,
5941 * without causing swapping or OOM.
5943 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5946 * Not all the page cache can be freed, otherwise the system will
5947 * start swapping or thrashing. Assume at least half of the page
5948 * cache, or the low watermark worth of cache, needs to stay.
5950 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5951 pagecache -= min(pagecache / 2, wmark_low);
5952 available += pagecache;
5955 * Part of the reclaimable slab and other kernel memory consists of
5956 * items that are in use, and cannot be freed. Cap this estimate at the
5959 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5960 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5961 available += reclaimable - min(reclaimable / 2, wmark_low);
5967 EXPORT_SYMBOL_GPL(si_mem_available);
5969 void si_meminfo(struct sysinfo *val)
5971 val->totalram = totalram_pages();
5972 val->sharedram = global_node_page_state(NR_SHMEM);
5973 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5974 val->bufferram = nr_blockdev_pages();
5975 val->totalhigh = totalhigh_pages();
5976 val->freehigh = nr_free_highpages();
5977 val->mem_unit = PAGE_SIZE;
5980 EXPORT_SYMBOL(si_meminfo);
5983 void si_meminfo_node(struct sysinfo *val, int nid)
5985 int zone_type; /* needs to be signed */
5986 unsigned long managed_pages = 0;
5987 unsigned long managed_highpages = 0;
5988 unsigned long free_highpages = 0;
5989 pg_data_t *pgdat = NODE_DATA(nid);
5991 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5992 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5993 val->totalram = managed_pages;
5994 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5995 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5996 #ifdef CONFIG_HIGHMEM
5997 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5998 struct zone *zone = &pgdat->node_zones[zone_type];
6000 if (is_highmem(zone)) {
6001 managed_highpages += zone_managed_pages(zone);
6002 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
6005 val->totalhigh = managed_highpages;
6006 val->freehigh = free_highpages;
6008 val->totalhigh = managed_highpages;
6009 val->freehigh = free_highpages;
6011 val->mem_unit = PAGE_SIZE;
6016 * Determine whether the node should be displayed or not, depending on whether
6017 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
6019 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
6021 if (!(flags & SHOW_MEM_FILTER_NODES))
6025 * no node mask - aka implicit memory numa policy. Do not bother with
6026 * the synchronization - read_mems_allowed_begin - because we do not
6027 * have to be precise here.
6030 nodemask = &cpuset_current_mems_allowed;
6032 return !node_isset(nid, *nodemask);
6035 #define K(x) ((x) << (PAGE_SHIFT-10))
6037 static void show_migration_types(unsigned char type)
6039 static const char types[MIGRATE_TYPES] = {
6040 [MIGRATE_UNMOVABLE] = 'U',
6041 [MIGRATE_MOVABLE] = 'M',
6042 [MIGRATE_RECLAIMABLE] = 'E',
6043 [MIGRATE_HIGHATOMIC] = 'H',
6045 [MIGRATE_CMA] = 'C',
6047 #ifdef CONFIG_MEMORY_ISOLATION
6048 [MIGRATE_ISOLATE] = 'I',
6051 char tmp[MIGRATE_TYPES + 1];
6055 for (i = 0; i < MIGRATE_TYPES; i++) {
6056 if (type & (1 << i))
6061 printk(KERN_CONT "(%s) ", tmp);
6064 static bool node_has_managed_zones(pg_data_t *pgdat, int max_zone_idx)
6067 for (zone_idx = 0; zone_idx <= max_zone_idx; zone_idx++)
6068 if (zone_managed_pages(pgdat->node_zones + zone_idx))
6074 * Show free area list (used inside shift_scroll-lock stuff)
6075 * We also calculate the percentage fragmentation. We do this by counting the
6076 * memory on each free list with the exception of the first item on the list.
6079 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
6082 void __show_free_areas(unsigned int filter, nodemask_t *nodemask, int max_zone_idx)
6084 unsigned long free_pcp = 0;
6089 for_each_populated_zone(zone) {
6090 if (zone_idx(zone) > max_zone_idx)
6092 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6095 for_each_online_cpu(cpu)
6096 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
6099 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
6100 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
6101 " unevictable:%lu dirty:%lu writeback:%lu\n"
6102 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
6103 " mapped:%lu shmem:%lu pagetables:%lu\n"
6104 " sec_pagetables:%lu bounce:%lu\n"
6105 " kernel_misc_reclaimable:%lu\n"
6106 " free:%lu free_pcp:%lu free_cma:%lu\n",
6107 global_node_page_state(NR_ACTIVE_ANON),
6108 global_node_page_state(NR_INACTIVE_ANON),
6109 global_node_page_state(NR_ISOLATED_ANON),
6110 global_node_page_state(NR_ACTIVE_FILE),
6111 global_node_page_state(NR_INACTIVE_FILE),
6112 global_node_page_state(NR_ISOLATED_FILE),
6113 global_node_page_state(NR_UNEVICTABLE),
6114 global_node_page_state(NR_FILE_DIRTY),
6115 global_node_page_state(NR_WRITEBACK),
6116 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
6117 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
6118 global_node_page_state(NR_FILE_MAPPED),
6119 global_node_page_state(NR_SHMEM),
6120 global_node_page_state(NR_PAGETABLE),
6121 global_node_page_state(NR_SECONDARY_PAGETABLE),
6122 global_zone_page_state(NR_BOUNCE),
6123 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE),
6124 global_zone_page_state(NR_FREE_PAGES),
6126 global_zone_page_state(NR_FREE_CMA_PAGES));
6128 for_each_online_pgdat(pgdat) {
6129 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
6131 if (!node_has_managed_zones(pgdat, max_zone_idx))
6135 " active_anon:%lukB"
6136 " inactive_anon:%lukB"
6137 " active_file:%lukB"
6138 " inactive_file:%lukB"
6139 " unevictable:%lukB"
6140 " isolated(anon):%lukB"
6141 " isolated(file):%lukB"
6146 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6148 " shmem_pmdmapped: %lukB"
6151 " writeback_tmp:%lukB"
6152 " kernel_stack:%lukB"
6153 #ifdef CONFIG_SHADOW_CALL_STACK
6154 " shadow_call_stack:%lukB"
6157 " sec_pagetables:%lukB"
6158 " all_unreclaimable? %s"
6161 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
6162 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
6163 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
6164 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
6165 K(node_page_state(pgdat, NR_UNEVICTABLE)),
6166 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
6167 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
6168 K(node_page_state(pgdat, NR_FILE_MAPPED)),
6169 K(node_page_state(pgdat, NR_FILE_DIRTY)),
6170 K(node_page_state(pgdat, NR_WRITEBACK)),
6171 K(node_page_state(pgdat, NR_SHMEM)),
6172 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6173 K(node_page_state(pgdat, NR_SHMEM_THPS)),
6174 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
6175 K(node_page_state(pgdat, NR_ANON_THPS)),
6177 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
6178 node_page_state(pgdat, NR_KERNEL_STACK_KB),
6179 #ifdef CONFIG_SHADOW_CALL_STACK
6180 node_page_state(pgdat, NR_KERNEL_SCS_KB),
6182 K(node_page_state(pgdat, NR_PAGETABLE)),
6183 K(node_page_state(pgdat, NR_SECONDARY_PAGETABLE)),
6184 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
6188 for_each_populated_zone(zone) {
6191 if (zone_idx(zone) > max_zone_idx)
6193 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6197 for_each_online_cpu(cpu)
6198 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
6208 " reserved_highatomic:%luKB"
6209 " active_anon:%lukB"
6210 " inactive_anon:%lukB"
6211 " active_file:%lukB"
6212 " inactive_file:%lukB"
6213 " unevictable:%lukB"
6214 " writepending:%lukB"
6224 K(zone_page_state(zone, NR_FREE_PAGES)),
6225 K(zone->watermark_boost),
6226 K(min_wmark_pages(zone)),
6227 K(low_wmark_pages(zone)),
6228 K(high_wmark_pages(zone)),
6229 K(zone->nr_reserved_highatomic),
6230 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
6231 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
6232 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
6233 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
6234 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
6235 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
6236 K(zone->present_pages),
6237 K(zone_managed_pages(zone)),
6238 K(zone_page_state(zone, NR_MLOCK)),
6239 K(zone_page_state(zone, NR_BOUNCE)),
6241 K(this_cpu_read(zone->per_cpu_pageset->count)),
6242 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
6243 printk("lowmem_reserve[]:");
6244 for (i = 0; i < MAX_NR_ZONES; i++)
6245 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
6246 printk(KERN_CONT "\n");
6249 for_each_populated_zone(zone) {
6251 unsigned long nr[MAX_ORDER], flags, total = 0;
6252 unsigned char types[MAX_ORDER];
6254 if (zone_idx(zone) > max_zone_idx)
6256 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6259 printk(KERN_CONT "%s: ", zone->name);
6261 spin_lock_irqsave(&zone->lock, flags);
6262 for (order = 0; order < MAX_ORDER; order++) {
6263 struct free_area *area = &zone->free_area[order];
6266 nr[order] = area->nr_free;
6267 total += nr[order] << order;
6270 for (type = 0; type < MIGRATE_TYPES; type++) {
6271 if (!free_area_empty(area, type))
6272 types[order] |= 1 << type;
6275 spin_unlock_irqrestore(&zone->lock, flags);
6276 for (order = 0; order < MAX_ORDER; order++) {
6277 printk(KERN_CONT "%lu*%lukB ",
6278 nr[order], K(1UL) << order);
6280 show_migration_types(types[order]);
6282 printk(KERN_CONT "= %lukB\n", K(total));
6285 for_each_online_node(nid) {
6286 if (show_mem_node_skip(filter, nid, nodemask))
6288 hugetlb_show_meminfo_node(nid);
6291 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
6293 show_swap_cache_info();
6296 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
6298 zoneref->zone = zone;
6299 zoneref->zone_idx = zone_idx(zone);
6303 * Builds allocation fallback zone lists.
6305 * Add all populated zones of a node to the zonelist.
6307 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
6310 enum zone_type zone_type = MAX_NR_ZONES;
6315 zone = pgdat->node_zones + zone_type;
6316 if (populated_zone(zone)) {
6317 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
6318 check_highest_zone(zone_type);
6320 } while (zone_type);
6327 static int __parse_numa_zonelist_order(char *s)
6330 * We used to support different zonelists modes but they turned
6331 * out to be just not useful. Let's keep the warning in place
6332 * if somebody still use the cmd line parameter so that we do
6333 * not fail it silently
6335 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
6336 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
6342 char numa_zonelist_order[] = "Node";
6345 * sysctl handler for numa_zonelist_order
6347 int numa_zonelist_order_handler(struct ctl_table *table, int write,
6348 void *buffer, size_t *length, loff_t *ppos)
6351 return __parse_numa_zonelist_order(buffer);
6352 return proc_dostring(table, write, buffer, length, ppos);
6356 static int node_load[MAX_NUMNODES];
6359 * find_next_best_node - find the next node that should appear in a given node's fallback list
6360 * @node: node whose fallback list we're appending
6361 * @used_node_mask: nodemask_t of already used nodes
6363 * We use a number of factors to determine which is the next node that should
6364 * appear on a given node's fallback list. The node should not have appeared
6365 * already in @node's fallback list, and it should be the next closest node
6366 * according to the distance array (which contains arbitrary distance values
6367 * from each node to each node in the system), and should also prefer nodes
6368 * with no CPUs, since presumably they'll have very little allocation pressure
6369 * on them otherwise.
6371 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
6373 int find_next_best_node(int node, nodemask_t *used_node_mask)
6376 int min_val = INT_MAX;
6377 int best_node = NUMA_NO_NODE;
6379 /* Use the local node if we haven't already */
6380 if (!node_isset(node, *used_node_mask)) {
6381 node_set(node, *used_node_mask);
6385 for_each_node_state(n, N_MEMORY) {
6387 /* Don't want a node to appear more than once */
6388 if (node_isset(n, *used_node_mask))
6391 /* Use the distance array to find the distance */
6392 val = node_distance(node, n);
6394 /* Penalize nodes under us ("prefer the next node") */
6397 /* Give preference to headless and unused nodes */
6398 if (!cpumask_empty(cpumask_of_node(n)))
6399 val += PENALTY_FOR_NODE_WITH_CPUS;
6401 /* Slight preference for less loaded node */
6402 val *= MAX_NUMNODES;
6403 val += node_load[n];
6405 if (val < min_val) {
6412 node_set(best_node, *used_node_mask);
6419 * Build zonelists ordered by node and zones within node.
6420 * This results in maximum locality--normal zone overflows into local
6421 * DMA zone, if any--but risks exhausting DMA zone.
6423 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
6426 struct zoneref *zonerefs;
6429 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6431 for (i = 0; i < nr_nodes; i++) {
6434 pg_data_t *node = NODE_DATA(node_order[i]);
6436 nr_zones = build_zonerefs_node(node, zonerefs);
6437 zonerefs += nr_zones;
6439 zonerefs->zone = NULL;
6440 zonerefs->zone_idx = 0;
6444 * Build gfp_thisnode zonelists
6446 static void build_thisnode_zonelists(pg_data_t *pgdat)
6448 struct zoneref *zonerefs;
6451 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
6452 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6453 zonerefs += nr_zones;
6454 zonerefs->zone = NULL;
6455 zonerefs->zone_idx = 0;
6459 * Build zonelists ordered by zone and nodes within zones.
6460 * This results in conserving DMA zone[s] until all Normal memory is
6461 * exhausted, but results in overflowing to remote node while memory
6462 * may still exist in local DMA zone.
6465 static void build_zonelists(pg_data_t *pgdat)
6467 static int node_order[MAX_NUMNODES];
6468 int node, nr_nodes = 0;
6469 nodemask_t used_mask = NODE_MASK_NONE;
6470 int local_node, prev_node;
6472 /* NUMA-aware ordering of nodes */
6473 local_node = pgdat->node_id;
6474 prev_node = local_node;
6476 memset(node_order, 0, sizeof(node_order));
6477 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
6479 * We don't want to pressure a particular node.
6480 * So adding penalty to the first node in same
6481 * distance group to make it round-robin.
6483 if (node_distance(local_node, node) !=
6484 node_distance(local_node, prev_node))
6485 node_load[node] += 1;
6487 node_order[nr_nodes++] = node;
6491 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
6492 build_thisnode_zonelists(pgdat);
6493 pr_info("Fallback order for Node %d: ", local_node);
6494 for (node = 0; node < nr_nodes; node++)
6495 pr_cont("%d ", node_order[node]);
6499 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6501 * Return node id of node used for "local" allocations.
6502 * I.e., first node id of first zone in arg node's generic zonelist.
6503 * Used for initializing percpu 'numa_mem', which is used primarily
6504 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
6506 int local_memory_node(int node)
6510 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
6511 gfp_zone(GFP_KERNEL),
6513 return zone_to_nid(z->zone);
6517 static void setup_min_unmapped_ratio(void);
6518 static void setup_min_slab_ratio(void);
6519 #else /* CONFIG_NUMA */
6521 static void build_zonelists(pg_data_t *pgdat)
6523 int node, local_node;
6524 struct zoneref *zonerefs;
6527 local_node = pgdat->node_id;
6529 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6530 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6531 zonerefs += nr_zones;
6534 * Now we build the zonelist so that it contains the zones
6535 * of all the other nodes.
6536 * We don't want to pressure a particular node, so when
6537 * building the zones for node N, we make sure that the
6538 * zones coming right after the local ones are those from
6539 * node N+1 (modulo N)
6541 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
6542 if (!node_online(node))
6544 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6545 zonerefs += nr_zones;
6547 for (node = 0; node < local_node; node++) {
6548 if (!node_online(node))
6550 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6551 zonerefs += nr_zones;
6554 zonerefs->zone = NULL;
6555 zonerefs->zone_idx = 0;
6558 #endif /* CONFIG_NUMA */
6561 * Boot pageset table. One per cpu which is going to be used for all
6562 * zones and all nodes. The parameters will be set in such a way
6563 * that an item put on a list will immediately be handed over to
6564 * the buddy list. This is safe since pageset manipulation is done
6565 * with interrupts disabled.
6567 * The boot_pagesets must be kept even after bootup is complete for
6568 * unused processors and/or zones. They do play a role for bootstrapping
6569 * hotplugged processors.
6571 * zoneinfo_show() and maybe other functions do
6572 * not check if the processor is online before following the pageset pointer.
6573 * Other parts of the kernel may not check if the zone is available.
6575 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
6576 /* These effectively disable the pcplists in the boot pageset completely */
6577 #define BOOT_PAGESET_HIGH 0
6578 #define BOOT_PAGESET_BATCH 1
6579 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
6580 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
6581 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6583 static void __build_all_zonelists(void *data)
6586 int __maybe_unused cpu;
6587 pg_data_t *self = data;
6588 unsigned long flags;
6591 * Explicitly disable this CPU's interrupts before taking seqlock
6592 * to prevent any IRQ handler from calling into the page allocator
6593 * (e.g. GFP_ATOMIC) that could hit zonelist_iter_begin and livelock.
6595 local_irq_save(flags);
6597 * Explicitly disable this CPU's synchronous printk() before taking
6598 * seqlock to prevent any printk() from trying to hold port->lock, for
6599 * tty_insert_flip_string_and_push_buffer() on other CPU might be
6600 * calling kmalloc(GFP_ATOMIC | __GFP_NOWARN) with port->lock held.
6602 printk_deferred_enter();
6603 write_seqlock(&zonelist_update_seq);
6606 memset(node_load, 0, sizeof(node_load));
6610 * This node is hotadded and no memory is yet present. So just
6611 * building zonelists is fine - no need to touch other nodes.
6613 if (self && !node_online(self->node_id)) {
6614 build_zonelists(self);
6617 * All possible nodes have pgdat preallocated
6620 for_each_node(nid) {
6621 pg_data_t *pgdat = NODE_DATA(nid);
6623 build_zonelists(pgdat);
6626 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6628 * We now know the "local memory node" for each node--
6629 * i.e., the node of the first zone in the generic zonelist.
6630 * Set up numa_mem percpu variable for on-line cpus. During
6631 * boot, only the boot cpu should be on-line; we'll init the
6632 * secondary cpus' numa_mem as they come on-line. During
6633 * node/memory hotplug, we'll fixup all on-line cpus.
6635 for_each_online_cpu(cpu)
6636 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6640 write_sequnlock(&zonelist_update_seq);
6641 printk_deferred_exit();
6642 local_irq_restore(flags);
6645 static noinline void __init
6646 build_all_zonelists_init(void)
6650 __build_all_zonelists(NULL);
6653 * Initialize the boot_pagesets that are going to be used
6654 * for bootstrapping processors. The real pagesets for
6655 * each zone will be allocated later when the per cpu
6656 * allocator is available.
6658 * boot_pagesets are used also for bootstrapping offline
6659 * cpus if the system is already booted because the pagesets
6660 * are needed to initialize allocators on a specific cpu too.
6661 * F.e. the percpu allocator needs the page allocator which
6662 * needs the percpu allocator in order to allocate its pagesets
6663 * (a chicken-egg dilemma).
6665 for_each_possible_cpu(cpu)
6666 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
6668 mminit_verify_zonelist();
6669 cpuset_init_current_mems_allowed();
6673 * unless system_state == SYSTEM_BOOTING.
6675 * __ref due to call of __init annotated helper build_all_zonelists_init
6676 * [protected by SYSTEM_BOOTING].
6678 void __ref build_all_zonelists(pg_data_t *pgdat)
6680 unsigned long vm_total_pages;
6682 if (system_state == SYSTEM_BOOTING) {
6683 build_all_zonelists_init();
6685 __build_all_zonelists(pgdat);
6686 /* cpuset refresh routine should be here */
6688 /* Get the number of free pages beyond high watermark in all zones. */
6689 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6691 * Disable grouping by mobility if the number of pages in the
6692 * system is too low to allow the mechanism to work. It would be
6693 * more accurate, but expensive to check per-zone. This check is
6694 * made on memory-hotadd so a system can start with mobility
6695 * disabled and enable it later
6697 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6698 page_group_by_mobility_disabled = 1;
6700 page_group_by_mobility_disabled = 0;
6702 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6704 page_group_by_mobility_disabled ? "off" : "on",
6707 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6711 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6712 static bool __meminit
6713 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6715 static struct memblock_region *r;
6717 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6718 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6719 for_each_mem_region(r) {
6720 if (*pfn < memblock_region_memory_end_pfn(r))
6724 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6725 memblock_is_mirror(r)) {
6726 *pfn = memblock_region_memory_end_pfn(r);
6734 * Initially all pages are reserved - free ones are freed
6735 * up by memblock_free_all() once the early boot process is
6736 * done. Non-atomic initialization, single-pass.
6738 * All aligned pageblocks are initialized to the specified migratetype
6739 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6740 * zone stats (e.g., nr_isolate_pageblock) are touched.
6742 void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6743 unsigned long start_pfn, unsigned long zone_end_pfn,
6744 enum meminit_context context,
6745 struct vmem_altmap *altmap, int migratetype)
6747 unsigned long pfn, end_pfn = start_pfn + size;
6750 if (highest_memmap_pfn < end_pfn - 1)
6751 highest_memmap_pfn = end_pfn - 1;
6753 #ifdef CONFIG_ZONE_DEVICE
6755 * Honor reservation requested by the driver for this ZONE_DEVICE
6756 * memory. We limit the total number of pages to initialize to just
6757 * those that might contain the memory mapping. We will defer the
6758 * ZONE_DEVICE page initialization until after we have released
6761 if (zone == ZONE_DEVICE) {
6765 if (start_pfn == altmap->base_pfn)
6766 start_pfn += altmap->reserve;
6767 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6771 for (pfn = start_pfn; pfn < end_pfn; ) {
6773 * There can be holes in boot-time mem_map[]s handed to this
6774 * function. They do not exist on hotplugged memory.
6776 if (context == MEMINIT_EARLY) {
6777 if (overlap_memmap_init(zone, &pfn))
6779 if (defer_init(nid, pfn, zone_end_pfn))
6783 page = pfn_to_page(pfn);
6784 __init_single_page(page, pfn, zone, nid);
6785 if (context == MEMINIT_HOTPLUG)
6786 __SetPageReserved(page);
6789 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6790 * such that unmovable allocations won't be scattered all
6791 * over the place during system boot.
6793 if (pageblock_aligned(pfn)) {
6794 set_pageblock_migratetype(page, migratetype);
6801 #ifdef CONFIG_ZONE_DEVICE
6802 static void __ref __init_zone_device_page(struct page *page, unsigned long pfn,
6803 unsigned long zone_idx, int nid,
6804 struct dev_pagemap *pgmap)
6807 __init_single_page(page, pfn, zone_idx, nid);
6810 * Mark page reserved as it will need to wait for onlining
6811 * phase for it to be fully associated with a zone.
6813 * We can use the non-atomic __set_bit operation for setting
6814 * the flag as we are still initializing the pages.
6816 __SetPageReserved(page);
6819 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6820 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6821 * ever freed or placed on a driver-private list.
6823 page->pgmap = pgmap;
6824 page->zone_device_data = NULL;
6827 * Mark the block movable so that blocks are reserved for
6828 * movable at startup. This will force kernel allocations
6829 * to reserve their blocks rather than leaking throughout
6830 * the address space during boot when many long-lived
6831 * kernel allocations are made.
6833 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6834 * because this is done early in section_activate()
6836 if (pageblock_aligned(pfn)) {
6837 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6842 * ZONE_DEVICE pages are released directly to the driver page allocator
6843 * which will set the page count to 1 when allocating the page.
6845 if (pgmap->type == MEMORY_DEVICE_PRIVATE ||
6846 pgmap->type == MEMORY_DEVICE_COHERENT)
6847 set_page_count(page, 0);
6851 * With compound page geometry and when struct pages are stored in ram most
6852 * tail pages are reused. Consequently, the amount of unique struct pages to
6853 * initialize is a lot smaller that the total amount of struct pages being
6854 * mapped. This is a paired / mild layering violation with explicit knowledge
6855 * of how the sparse_vmemmap internals handle compound pages in the lack
6856 * of an altmap. See vmemmap_populate_compound_pages().
6858 static inline unsigned long compound_nr_pages(struct vmem_altmap *altmap,
6859 unsigned long nr_pages)
6861 return is_power_of_2(sizeof(struct page)) &&
6862 !altmap ? 2 * (PAGE_SIZE / sizeof(struct page)) : nr_pages;
6865 static void __ref memmap_init_compound(struct page *head,
6866 unsigned long head_pfn,
6867 unsigned long zone_idx, int nid,
6868 struct dev_pagemap *pgmap,
6869 unsigned long nr_pages)
6871 unsigned long pfn, end_pfn = head_pfn + nr_pages;
6872 unsigned int order = pgmap->vmemmap_shift;
6874 __SetPageHead(head);
6875 for (pfn = head_pfn + 1; pfn < end_pfn; pfn++) {
6876 struct page *page = pfn_to_page(pfn);
6878 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6879 prep_compound_tail(head, pfn - head_pfn);
6880 set_page_count(page, 0);
6883 * The first tail page stores compound_mapcount_ptr() and
6884 * compound_order() and the second tail page stores
6885 * compound_pincount_ptr(). Call prep_compound_head() after
6886 * the first and second tail pages have been initialized to
6887 * not have the data overwritten.
6889 if (pfn == head_pfn + 2)
6890 prep_compound_head(head, order);
6894 void __ref memmap_init_zone_device(struct zone *zone,
6895 unsigned long start_pfn,
6896 unsigned long nr_pages,
6897 struct dev_pagemap *pgmap)
6899 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6900 struct pglist_data *pgdat = zone->zone_pgdat;
6901 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6902 unsigned int pfns_per_compound = pgmap_vmemmap_nr(pgmap);
6903 unsigned long zone_idx = zone_idx(zone);
6904 unsigned long start = jiffies;
6905 int nid = pgdat->node_id;
6907 if (WARN_ON_ONCE(!pgmap || zone_idx != ZONE_DEVICE))
6911 * The call to memmap_init should have already taken care
6912 * of the pages reserved for the memmap, so we can just jump to
6913 * the end of that region and start processing the device pages.
6916 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6917 nr_pages = end_pfn - start_pfn;
6920 for (pfn = start_pfn; pfn < end_pfn; pfn += pfns_per_compound) {
6921 struct page *page = pfn_to_page(pfn);
6923 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6925 if (pfns_per_compound == 1)
6928 memmap_init_compound(page, pfn, zone_idx, nid, pgmap,
6929 compound_nr_pages(altmap, pfns_per_compound));
6932 pr_info("%s initialised %lu pages in %ums\n", __func__,
6933 nr_pages, jiffies_to_msecs(jiffies - start));
6937 static void __meminit zone_init_free_lists(struct zone *zone)
6939 unsigned int order, t;
6940 for_each_migratetype_order(order, t) {
6941 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6942 zone->free_area[order].nr_free = 0;
6947 * Only struct pages that correspond to ranges defined by memblock.memory
6948 * are zeroed and initialized by going through __init_single_page() during
6949 * memmap_init_zone_range().
6951 * But, there could be struct pages that correspond to holes in
6952 * memblock.memory. This can happen because of the following reasons:
6953 * - physical memory bank size is not necessarily the exact multiple of the
6954 * arbitrary section size
6955 * - early reserved memory may not be listed in memblock.memory
6956 * - memory layouts defined with memmap= kernel parameter may not align
6957 * nicely with memmap sections
6959 * Explicitly initialize those struct pages so that:
6960 * - PG_Reserved is set
6961 * - zone and node links point to zone and node that span the page if the
6962 * hole is in the middle of a zone
6963 * - zone and node links point to adjacent zone/node if the hole falls on
6964 * the zone boundary; the pages in such holes will be prepended to the
6965 * zone/node above the hole except for the trailing pages in the last
6966 * section that will be appended to the zone/node below.
6968 static void __init init_unavailable_range(unsigned long spfn,
6975 for (pfn = spfn; pfn < epfn; pfn++) {
6976 if (!pfn_valid(pageblock_start_pfn(pfn))) {
6977 pfn = pageblock_end_pfn(pfn) - 1;
6980 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6981 __SetPageReserved(pfn_to_page(pfn));
6986 pr_info("On node %d, zone %s: %lld pages in unavailable ranges",
6987 node, zone_names[zone], pgcnt);
6990 static void __init memmap_init_zone_range(struct zone *zone,
6991 unsigned long start_pfn,
6992 unsigned long end_pfn,
6993 unsigned long *hole_pfn)
6995 unsigned long zone_start_pfn = zone->zone_start_pfn;
6996 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6997 int nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6999 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
7000 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
7002 if (start_pfn >= end_pfn)
7005 memmap_init_range(end_pfn - start_pfn, nid, zone_id, start_pfn,
7006 zone_end_pfn, MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
7008 if (*hole_pfn < start_pfn)
7009 init_unavailable_range(*hole_pfn, start_pfn, zone_id, nid);
7011 *hole_pfn = end_pfn;
7014 static void __init memmap_init(void)
7016 unsigned long start_pfn, end_pfn;
7017 unsigned long hole_pfn = 0;
7018 int i, j, zone_id = 0, nid;
7020 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7021 struct pglist_data *node = NODE_DATA(nid);
7023 for (j = 0; j < MAX_NR_ZONES; j++) {
7024 struct zone *zone = node->node_zones + j;
7026 if (!populated_zone(zone))
7029 memmap_init_zone_range(zone, start_pfn, end_pfn,
7035 #ifdef CONFIG_SPARSEMEM
7037 * Initialize the memory map for hole in the range [memory_end,
7039 * Append the pages in this hole to the highest zone in the last
7041 * The call to init_unavailable_range() is outside the ifdef to
7042 * silence the compiler warining about zone_id set but not used;
7043 * for FLATMEM it is a nop anyway
7045 end_pfn = round_up(end_pfn, PAGES_PER_SECTION);
7046 if (hole_pfn < end_pfn)
7048 init_unavailable_range(hole_pfn, end_pfn, zone_id, nid);
7051 void __init *memmap_alloc(phys_addr_t size, phys_addr_t align,
7052 phys_addr_t min_addr, int nid, bool exact_nid)
7057 ptr = memblock_alloc_exact_nid_raw(size, align, min_addr,
7058 MEMBLOCK_ALLOC_ACCESSIBLE,
7061 ptr = memblock_alloc_try_nid_raw(size, align, min_addr,
7062 MEMBLOCK_ALLOC_ACCESSIBLE,
7065 if (ptr && size > 0)
7066 page_init_poison(ptr, size);
7071 static int zone_batchsize(struct zone *zone)
7077 * The number of pages to batch allocate is either ~0.1%
7078 * of the zone or 1MB, whichever is smaller. The batch
7079 * size is striking a balance between allocation latency
7080 * and zone lock contention.
7082 batch = min(zone_managed_pages(zone) >> 10, SZ_1M / PAGE_SIZE);
7083 batch /= 4; /* We effectively *= 4 below */
7088 * Clamp the batch to a 2^n - 1 value. Having a power
7089 * of 2 value was found to be more likely to have
7090 * suboptimal cache aliasing properties in some cases.
7092 * For example if 2 tasks are alternately allocating
7093 * batches of pages, one task can end up with a lot
7094 * of pages of one half of the possible page colors
7095 * and the other with pages of the other colors.
7097 batch = rounddown_pow_of_two(batch + batch/2) - 1;
7102 /* The deferral and batching of frees should be suppressed under NOMMU
7105 * The problem is that NOMMU needs to be able to allocate large chunks
7106 * of contiguous memory as there's no hardware page translation to
7107 * assemble apparent contiguous memory from discontiguous pages.
7109 * Queueing large contiguous runs of pages for batching, however,
7110 * causes the pages to actually be freed in smaller chunks. As there
7111 * can be a significant delay between the individual batches being
7112 * recycled, this leads to the once large chunks of space being
7113 * fragmented and becoming unavailable for high-order allocations.
7119 static int zone_highsize(struct zone *zone, int batch, int cpu_online)
7124 unsigned long total_pages;
7126 if (!percpu_pagelist_high_fraction) {
7128 * By default, the high value of the pcp is based on the zone
7129 * low watermark so that if they are full then background
7130 * reclaim will not be started prematurely.
7132 total_pages = low_wmark_pages(zone);
7135 * If percpu_pagelist_high_fraction is configured, the high
7136 * value is based on a fraction of the managed pages in the
7139 total_pages = zone_managed_pages(zone) / percpu_pagelist_high_fraction;
7143 * Split the high value across all online CPUs local to the zone. Note
7144 * that early in boot that CPUs may not be online yet and that during
7145 * CPU hotplug that the cpumask is not yet updated when a CPU is being
7146 * onlined. For memory nodes that have no CPUs, split pcp->high across
7147 * all online CPUs to mitigate the risk that reclaim is triggered
7148 * prematurely due to pages stored on pcp lists.
7150 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
7152 nr_split_cpus = num_online_cpus();
7153 high = total_pages / nr_split_cpus;
7156 * Ensure high is at least batch*4. The multiple is based on the
7157 * historical relationship between high and batch.
7159 high = max(high, batch << 2);
7168 * pcp->high and pcp->batch values are related and generally batch is lower
7169 * than high. They are also related to pcp->count such that count is lower
7170 * than high, and as soon as it reaches high, the pcplist is flushed.
7172 * However, guaranteeing these relations at all times would require e.g. write
7173 * barriers here but also careful usage of read barriers at the read side, and
7174 * thus be prone to error and bad for performance. Thus the update only prevents
7175 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
7176 * can cope with those fields changing asynchronously, and fully trust only the
7177 * pcp->count field on the local CPU with interrupts disabled.
7179 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
7180 * outside of boot time (or some other assurance that no concurrent updaters
7183 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
7184 unsigned long batch)
7186 WRITE_ONCE(pcp->batch, batch);
7187 WRITE_ONCE(pcp->high, high);
7190 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
7194 memset(pcp, 0, sizeof(*pcp));
7195 memset(pzstats, 0, sizeof(*pzstats));
7197 spin_lock_init(&pcp->lock);
7198 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
7199 INIT_LIST_HEAD(&pcp->lists[pindex]);
7202 * Set batch and high values safe for a boot pageset. A true percpu
7203 * pageset's initialization will update them subsequently. Here we don't
7204 * need to be as careful as pageset_update() as nobody can access the
7207 pcp->high = BOOT_PAGESET_HIGH;
7208 pcp->batch = BOOT_PAGESET_BATCH;
7209 pcp->free_factor = 0;
7212 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
7213 unsigned long batch)
7215 struct per_cpu_pages *pcp;
7218 for_each_possible_cpu(cpu) {
7219 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
7220 pageset_update(pcp, high, batch);
7225 * Calculate and set new high and batch values for all per-cpu pagesets of a
7226 * zone based on the zone's size.
7228 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
7230 int new_high, new_batch;
7232 new_batch = max(1, zone_batchsize(zone));
7233 new_high = zone_highsize(zone, new_batch, cpu_online);
7235 if (zone->pageset_high == new_high &&
7236 zone->pageset_batch == new_batch)
7239 zone->pageset_high = new_high;
7240 zone->pageset_batch = new_batch;
7242 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
7245 void __meminit setup_zone_pageset(struct zone *zone)
7249 /* Size may be 0 on !SMP && !NUMA */
7250 if (sizeof(struct per_cpu_zonestat) > 0)
7251 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
7253 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
7254 for_each_possible_cpu(cpu) {
7255 struct per_cpu_pages *pcp;
7256 struct per_cpu_zonestat *pzstats;
7258 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
7259 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
7260 per_cpu_pages_init(pcp, pzstats);
7263 zone_set_pageset_high_and_batch(zone, 0);
7267 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7268 * page high values need to be recalculated.
7270 static void zone_pcp_update(struct zone *zone, int cpu_online)
7272 mutex_lock(&pcp_batch_high_lock);
7273 zone_set_pageset_high_and_batch(zone, cpu_online);
7274 mutex_unlock(&pcp_batch_high_lock);
7278 * Allocate per cpu pagesets and initialize them.
7279 * Before this call only boot pagesets were available.
7281 void __init setup_per_cpu_pageset(void)
7283 struct pglist_data *pgdat;
7285 int __maybe_unused cpu;
7287 for_each_populated_zone(zone)
7288 setup_zone_pageset(zone);
7292 * Unpopulated zones continue using the boot pagesets.
7293 * The numa stats for these pagesets need to be reset.
7294 * Otherwise, they will end up skewing the stats of
7295 * the nodes these zones are associated with.
7297 for_each_possible_cpu(cpu) {
7298 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
7299 memset(pzstats->vm_numa_event, 0,
7300 sizeof(pzstats->vm_numa_event));
7304 for_each_online_pgdat(pgdat)
7305 pgdat->per_cpu_nodestats =
7306 alloc_percpu(struct per_cpu_nodestat);
7309 static __meminit void zone_pcp_init(struct zone *zone)
7312 * per cpu subsystem is not up at this point. The following code
7313 * relies on the ability of the linker to provide the
7314 * offset of a (static) per cpu variable into the per cpu area.
7316 zone->per_cpu_pageset = &boot_pageset;
7317 zone->per_cpu_zonestats = &boot_zonestats;
7318 zone->pageset_high = BOOT_PAGESET_HIGH;
7319 zone->pageset_batch = BOOT_PAGESET_BATCH;
7321 if (populated_zone(zone))
7322 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
7323 zone->present_pages, zone_batchsize(zone));
7326 void __meminit init_currently_empty_zone(struct zone *zone,
7327 unsigned long zone_start_pfn,
7330 struct pglist_data *pgdat = zone->zone_pgdat;
7331 int zone_idx = zone_idx(zone) + 1;
7333 if (zone_idx > pgdat->nr_zones)
7334 pgdat->nr_zones = zone_idx;
7336 zone->zone_start_pfn = zone_start_pfn;
7338 mminit_dprintk(MMINIT_TRACE, "memmap_init",
7339 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
7341 (unsigned long)zone_idx(zone),
7342 zone_start_pfn, (zone_start_pfn + size));
7344 zone_init_free_lists(zone);
7345 zone->initialized = 1;
7349 * get_pfn_range_for_nid - Return the start and end page frames for a node
7350 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
7351 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
7352 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
7354 * It returns the start and end page frame of a node based on information
7355 * provided by memblock_set_node(). If called for a node
7356 * with no available memory, a warning is printed and the start and end
7359 void __init get_pfn_range_for_nid(unsigned int nid,
7360 unsigned long *start_pfn, unsigned long *end_pfn)
7362 unsigned long this_start_pfn, this_end_pfn;
7368 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
7369 *start_pfn = min(*start_pfn, this_start_pfn);
7370 *end_pfn = max(*end_pfn, this_end_pfn);
7373 if (*start_pfn == -1UL)
7378 * This finds a zone that can be used for ZONE_MOVABLE pages. The
7379 * assumption is made that zones within a node are ordered in monotonic
7380 * increasing memory addresses so that the "highest" populated zone is used
7382 static void __init find_usable_zone_for_movable(void)
7385 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
7386 if (zone_index == ZONE_MOVABLE)
7389 if (arch_zone_highest_possible_pfn[zone_index] >
7390 arch_zone_lowest_possible_pfn[zone_index])
7394 VM_BUG_ON(zone_index == -1);
7395 movable_zone = zone_index;
7399 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
7400 * because it is sized independent of architecture. Unlike the other zones,
7401 * the starting point for ZONE_MOVABLE is not fixed. It may be different
7402 * in each node depending on the size of each node and how evenly kernelcore
7403 * is distributed. This helper function adjusts the zone ranges
7404 * provided by the architecture for a given node by using the end of the
7405 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
7406 * zones within a node are in order of monotonic increases memory addresses
7408 static void __init adjust_zone_range_for_zone_movable(int nid,
7409 unsigned long zone_type,
7410 unsigned long node_start_pfn,
7411 unsigned long node_end_pfn,
7412 unsigned long *zone_start_pfn,
7413 unsigned long *zone_end_pfn)
7415 /* Only adjust if ZONE_MOVABLE is on this node */
7416 if (zone_movable_pfn[nid]) {
7417 /* Size ZONE_MOVABLE */
7418 if (zone_type == ZONE_MOVABLE) {
7419 *zone_start_pfn = zone_movable_pfn[nid];
7420 *zone_end_pfn = min(node_end_pfn,
7421 arch_zone_highest_possible_pfn[movable_zone]);
7423 /* Adjust for ZONE_MOVABLE starting within this range */
7424 } else if (!mirrored_kernelcore &&
7425 *zone_start_pfn < zone_movable_pfn[nid] &&
7426 *zone_end_pfn > zone_movable_pfn[nid]) {
7427 *zone_end_pfn = zone_movable_pfn[nid];
7429 /* Check if this whole range is within ZONE_MOVABLE */
7430 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
7431 *zone_start_pfn = *zone_end_pfn;
7436 * Return the number of pages a zone spans in a node, including holes
7437 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
7439 static unsigned long __init zone_spanned_pages_in_node(int nid,
7440 unsigned long zone_type,
7441 unsigned long node_start_pfn,
7442 unsigned long node_end_pfn,
7443 unsigned long *zone_start_pfn,
7444 unsigned long *zone_end_pfn)
7446 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7447 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7448 /* When hotadd a new node from cpu_up(), the node should be empty */
7449 if (!node_start_pfn && !node_end_pfn)
7452 /* Get the start and end of the zone */
7453 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7454 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7455 adjust_zone_range_for_zone_movable(nid, zone_type,
7456 node_start_pfn, node_end_pfn,
7457 zone_start_pfn, zone_end_pfn);
7459 /* Check that this node has pages within the zone's required range */
7460 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
7463 /* Move the zone boundaries inside the node if necessary */
7464 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
7465 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
7467 /* Return the spanned pages */
7468 return *zone_end_pfn - *zone_start_pfn;
7472 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
7473 * then all holes in the requested range will be accounted for.
7475 unsigned long __init __absent_pages_in_range(int nid,
7476 unsigned long range_start_pfn,
7477 unsigned long range_end_pfn)
7479 unsigned long nr_absent = range_end_pfn - range_start_pfn;
7480 unsigned long start_pfn, end_pfn;
7483 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7484 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
7485 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
7486 nr_absent -= end_pfn - start_pfn;
7492 * absent_pages_in_range - Return number of page frames in holes within a range
7493 * @start_pfn: The start PFN to start searching for holes
7494 * @end_pfn: The end PFN to stop searching for holes
7496 * Return: the number of pages frames in memory holes within a range.
7498 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
7499 unsigned long end_pfn)
7501 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
7504 /* Return the number of page frames in holes in a zone on a node */
7505 static unsigned long __init zone_absent_pages_in_node(int nid,
7506 unsigned long zone_type,
7507 unsigned long node_start_pfn,
7508 unsigned long node_end_pfn)
7510 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7511 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7512 unsigned long zone_start_pfn, zone_end_pfn;
7513 unsigned long nr_absent;
7515 /* When hotadd a new node from cpu_up(), the node should be empty */
7516 if (!node_start_pfn && !node_end_pfn)
7519 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7520 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7522 adjust_zone_range_for_zone_movable(nid, zone_type,
7523 node_start_pfn, node_end_pfn,
7524 &zone_start_pfn, &zone_end_pfn);
7525 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
7528 * ZONE_MOVABLE handling.
7529 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
7532 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
7533 unsigned long start_pfn, end_pfn;
7534 struct memblock_region *r;
7536 for_each_mem_region(r) {
7537 start_pfn = clamp(memblock_region_memory_base_pfn(r),
7538 zone_start_pfn, zone_end_pfn);
7539 end_pfn = clamp(memblock_region_memory_end_pfn(r),
7540 zone_start_pfn, zone_end_pfn);
7542 if (zone_type == ZONE_MOVABLE &&
7543 memblock_is_mirror(r))
7544 nr_absent += end_pfn - start_pfn;
7546 if (zone_type == ZONE_NORMAL &&
7547 !memblock_is_mirror(r))
7548 nr_absent += end_pfn - start_pfn;
7555 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
7556 unsigned long node_start_pfn,
7557 unsigned long node_end_pfn)
7559 unsigned long realtotalpages = 0, totalpages = 0;
7562 for (i = 0; i < MAX_NR_ZONES; i++) {
7563 struct zone *zone = pgdat->node_zones + i;
7564 unsigned long zone_start_pfn, zone_end_pfn;
7565 unsigned long spanned, absent;
7566 unsigned long size, real_size;
7568 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
7573 absent = zone_absent_pages_in_node(pgdat->node_id, i,
7578 real_size = size - absent;
7581 zone->zone_start_pfn = zone_start_pfn;
7583 zone->zone_start_pfn = 0;
7584 zone->spanned_pages = size;
7585 zone->present_pages = real_size;
7586 #if defined(CONFIG_MEMORY_HOTPLUG)
7587 zone->present_early_pages = real_size;
7591 realtotalpages += real_size;
7594 pgdat->node_spanned_pages = totalpages;
7595 pgdat->node_present_pages = realtotalpages;
7596 pr_debug("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
7599 #ifndef CONFIG_SPARSEMEM
7601 * Calculate the size of the zone->blockflags rounded to an unsigned long
7602 * Start by making sure zonesize is a multiple of pageblock_order by rounding
7603 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
7604 * round what is now in bits to nearest long in bits, then return it in
7607 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
7609 unsigned long usemapsize;
7611 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
7612 usemapsize = roundup(zonesize, pageblock_nr_pages);
7613 usemapsize = usemapsize >> pageblock_order;
7614 usemapsize *= NR_PAGEBLOCK_BITS;
7615 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
7617 return usemapsize / 8;
7620 static void __ref setup_usemap(struct zone *zone)
7622 unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
7623 zone->spanned_pages);
7624 zone->pageblock_flags = NULL;
7626 zone->pageblock_flags =
7627 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
7629 if (!zone->pageblock_flags)
7630 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
7631 usemapsize, zone->name, zone_to_nid(zone));
7635 static inline void setup_usemap(struct zone *zone) {}
7636 #endif /* CONFIG_SPARSEMEM */
7638 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
7640 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
7641 void __init set_pageblock_order(void)
7643 unsigned int order = MAX_ORDER - 1;
7645 /* Check that pageblock_nr_pages has not already been setup */
7646 if (pageblock_order)
7649 /* Don't let pageblocks exceed the maximum allocation granularity. */
7650 if (HPAGE_SHIFT > PAGE_SHIFT && HUGETLB_PAGE_ORDER < order)
7651 order = HUGETLB_PAGE_ORDER;
7654 * Assume the largest contiguous order of interest is a huge page.
7655 * This value may be variable depending on boot parameters on IA64 and
7658 pageblock_order = order;
7660 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7663 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
7664 * is unused as pageblock_order is set at compile-time. See
7665 * include/linux/pageblock-flags.h for the values of pageblock_order based on
7668 void __init set_pageblock_order(void)
7672 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7674 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
7675 unsigned long present_pages)
7677 unsigned long pages = spanned_pages;
7680 * Provide a more accurate estimation if there are holes within
7681 * the zone and SPARSEMEM is in use. If there are holes within the
7682 * zone, each populated memory region may cost us one or two extra
7683 * memmap pages due to alignment because memmap pages for each
7684 * populated regions may not be naturally aligned on page boundary.
7685 * So the (present_pages >> 4) heuristic is a tradeoff for that.
7687 if (spanned_pages > present_pages + (present_pages >> 4) &&
7688 IS_ENABLED(CONFIG_SPARSEMEM))
7689 pages = present_pages;
7691 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
7694 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
7695 static void pgdat_init_split_queue(struct pglist_data *pgdat)
7697 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
7699 spin_lock_init(&ds_queue->split_queue_lock);
7700 INIT_LIST_HEAD(&ds_queue->split_queue);
7701 ds_queue->split_queue_len = 0;
7704 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
7707 #ifdef CONFIG_COMPACTION
7708 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
7710 init_waitqueue_head(&pgdat->kcompactd_wait);
7713 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
7716 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
7720 pgdat_resize_init(pgdat);
7721 pgdat_kswapd_lock_init(pgdat);
7723 pgdat_init_split_queue(pgdat);
7724 pgdat_init_kcompactd(pgdat);
7726 init_waitqueue_head(&pgdat->kswapd_wait);
7727 init_waitqueue_head(&pgdat->pfmemalloc_wait);
7729 for (i = 0; i < NR_VMSCAN_THROTTLE; i++)
7730 init_waitqueue_head(&pgdat->reclaim_wait[i]);
7732 pgdat_page_ext_init(pgdat);
7733 lruvec_init(&pgdat->__lruvec);
7736 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7737 unsigned long remaining_pages)
7739 atomic_long_set(&zone->managed_pages, remaining_pages);
7740 zone_set_nid(zone, nid);
7741 zone->name = zone_names[idx];
7742 zone->zone_pgdat = NODE_DATA(nid);
7743 spin_lock_init(&zone->lock);
7744 zone_seqlock_init(zone);
7745 zone_pcp_init(zone);
7749 * Set up the zone data structures
7750 * - init pgdat internals
7751 * - init all zones belonging to this node
7753 * NOTE: this function is only called during memory hotplug
7755 #ifdef CONFIG_MEMORY_HOTPLUG
7756 void __ref free_area_init_core_hotplug(struct pglist_data *pgdat)
7758 int nid = pgdat->node_id;
7762 pgdat_init_internals(pgdat);
7764 if (pgdat->per_cpu_nodestats == &boot_nodestats)
7765 pgdat->per_cpu_nodestats = alloc_percpu(struct per_cpu_nodestat);
7768 * Reset the nr_zones, order and highest_zoneidx before reuse.
7769 * Note that kswapd will init kswapd_highest_zoneidx properly
7770 * when it starts in the near future.
7772 pgdat->nr_zones = 0;
7773 pgdat->kswapd_order = 0;
7774 pgdat->kswapd_highest_zoneidx = 0;
7775 pgdat->node_start_pfn = 0;
7776 for_each_online_cpu(cpu) {
7777 struct per_cpu_nodestat *p;
7779 p = per_cpu_ptr(pgdat->per_cpu_nodestats, cpu);
7780 memset(p, 0, sizeof(*p));
7783 for (z = 0; z < MAX_NR_ZONES; z++)
7784 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7789 * Set up the zone data structures:
7790 * - mark all pages reserved
7791 * - mark all memory queues empty
7792 * - clear the memory bitmaps
7794 * NOTE: pgdat should get zeroed by caller.
7795 * NOTE: this function is only called during early init.
7797 static void __init free_area_init_core(struct pglist_data *pgdat)
7800 int nid = pgdat->node_id;
7802 pgdat_init_internals(pgdat);
7803 pgdat->per_cpu_nodestats = &boot_nodestats;
7805 for (j = 0; j < MAX_NR_ZONES; j++) {
7806 struct zone *zone = pgdat->node_zones + j;
7807 unsigned long size, freesize, memmap_pages;
7809 size = zone->spanned_pages;
7810 freesize = zone->present_pages;
7813 * Adjust freesize so that it accounts for how much memory
7814 * is used by this zone for memmap. This affects the watermark
7815 * and per-cpu initialisations
7817 memmap_pages = calc_memmap_size(size, freesize);
7818 if (!is_highmem_idx(j)) {
7819 if (freesize >= memmap_pages) {
7820 freesize -= memmap_pages;
7822 pr_debug(" %s zone: %lu pages used for memmap\n",
7823 zone_names[j], memmap_pages);
7825 pr_warn(" %s zone: %lu memmap pages exceeds freesize %lu\n",
7826 zone_names[j], memmap_pages, freesize);
7829 /* Account for reserved pages */
7830 if (j == 0 && freesize > dma_reserve) {
7831 freesize -= dma_reserve;
7832 pr_debug(" %s zone: %lu pages reserved\n", zone_names[0], dma_reserve);
7835 if (!is_highmem_idx(j))
7836 nr_kernel_pages += freesize;
7837 /* Charge for highmem memmap if there are enough kernel pages */
7838 else if (nr_kernel_pages > memmap_pages * 2)
7839 nr_kernel_pages -= memmap_pages;
7840 nr_all_pages += freesize;
7843 * Set an approximate value for lowmem here, it will be adjusted
7844 * when the bootmem allocator frees pages into the buddy system.
7845 * And all highmem pages will be managed by the buddy system.
7847 zone_init_internals(zone, j, nid, freesize);
7852 set_pageblock_order();
7854 init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7858 #ifdef CONFIG_FLATMEM
7859 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
7861 unsigned long __maybe_unused start = 0;
7862 unsigned long __maybe_unused offset = 0;
7864 /* Skip empty nodes */
7865 if (!pgdat->node_spanned_pages)
7868 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7869 offset = pgdat->node_start_pfn - start;
7870 /* ia64 gets its own node_mem_map, before this, without bootmem */
7871 if (!pgdat->node_mem_map) {
7872 unsigned long size, end;
7876 * The zone's endpoints aren't required to be MAX_ORDER
7877 * aligned but the node_mem_map endpoints must be in order
7878 * for the buddy allocator to function correctly.
7880 end = pgdat_end_pfn(pgdat);
7881 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7882 size = (end - start) * sizeof(struct page);
7883 map = memmap_alloc(size, SMP_CACHE_BYTES, MEMBLOCK_LOW_LIMIT,
7884 pgdat->node_id, false);
7886 panic("Failed to allocate %ld bytes for node %d memory map\n",
7887 size, pgdat->node_id);
7888 pgdat->node_mem_map = map + offset;
7890 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7891 __func__, pgdat->node_id, (unsigned long)pgdat,
7892 (unsigned long)pgdat->node_mem_map);
7895 * With no DISCONTIG, the global mem_map is just set as node 0's
7897 if (pgdat == NODE_DATA(0)) {
7898 mem_map = NODE_DATA(0)->node_mem_map;
7899 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7905 static inline void alloc_node_mem_map(struct pglist_data *pgdat) { }
7906 #endif /* CONFIG_FLATMEM */
7908 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7909 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7911 pgdat->first_deferred_pfn = ULONG_MAX;
7914 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7917 static void __init free_area_init_node(int nid)
7919 pg_data_t *pgdat = NODE_DATA(nid);
7920 unsigned long start_pfn = 0;
7921 unsigned long end_pfn = 0;
7923 /* pg_data_t should be reset to zero when it's allocated */
7924 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7926 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7928 pgdat->node_id = nid;
7929 pgdat->node_start_pfn = start_pfn;
7930 pgdat->per_cpu_nodestats = NULL;
7932 if (start_pfn != end_pfn) {
7933 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7934 (u64)start_pfn << PAGE_SHIFT,
7935 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7937 pr_info("Initmem setup node %d as memoryless\n", nid);
7940 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7942 alloc_node_mem_map(pgdat);
7943 pgdat_set_deferred_range(pgdat);
7945 free_area_init_core(pgdat);
7948 static void __init free_area_init_memoryless_node(int nid)
7950 free_area_init_node(nid);
7953 #if MAX_NUMNODES > 1
7955 * Figure out the number of possible node ids.
7957 void __init setup_nr_node_ids(void)
7959 unsigned int highest;
7961 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7962 nr_node_ids = highest + 1;
7967 * node_map_pfn_alignment - determine the maximum internode alignment
7969 * This function should be called after node map is populated and sorted.
7970 * It calculates the maximum power of two alignment which can distinguish
7973 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7974 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7975 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7976 * shifted, 1GiB is enough and this function will indicate so.
7978 * This is used to test whether pfn -> nid mapping of the chosen memory
7979 * model has fine enough granularity to avoid incorrect mapping for the
7980 * populated node map.
7982 * Return: the determined alignment in pfn's. 0 if there is no alignment
7983 * requirement (single node).
7985 unsigned long __init node_map_pfn_alignment(void)
7987 unsigned long accl_mask = 0, last_end = 0;
7988 unsigned long start, end, mask;
7989 int last_nid = NUMA_NO_NODE;
7992 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7993 if (!start || last_nid < 0 || last_nid == nid) {
8000 * Start with a mask granular enough to pin-point to the
8001 * start pfn and tick off bits one-by-one until it becomes
8002 * too coarse to separate the current node from the last.
8004 mask = ~((1 << __ffs(start)) - 1);
8005 while (mask && last_end <= (start & (mask << 1)))
8008 /* accumulate all internode masks */
8012 /* convert mask to number of pages */
8013 return ~accl_mask + 1;
8017 * early_calculate_totalpages()
8018 * Sum pages in active regions for movable zone.
8019 * Populate N_MEMORY for calculating usable_nodes.
8021 static unsigned long __init early_calculate_totalpages(void)
8023 unsigned long totalpages = 0;
8024 unsigned long start_pfn, end_pfn;
8027 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
8028 unsigned long pages = end_pfn - start_pfn;
8030 totalpages += pages;
8032 node_set_state(nid, N_MEMORY);
8038 * Find the PFN the Movable zone begins in each node. Kernel memory
8039 * is spread evenly between nodes as long as the nodes have enough
8040 * memory. When they don't, some nodes will have more kernelcore than
8043 static void __init find_zone_movable_pfns_for_nodes(void)
8046 unsigned long usable_startpfn;
8047 unsigned long kernelcore_node, kernelcore_remaining;
8048 /* save the state before borrow the nodemask */
8049 nodemask_t saved_node_state = node_states[N_MEMORY];
8050 unsigned long totalpages = early_calculate_totalpages();
8051 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
8052 struct memblock_region *r;
8054 /* Need to find movable_zone earlier when movable_node is specified. */
8055 find_usable_zone_for_movable();
8058 * If movable_node is specified, ignore kernelcore and movablecore
8061 if (movable_node_is_enabled()) {
8062 for_each_mem_region(r) {
8063 if (!memblock_is_hotpluggable(r))
8066 nid = memblock_get_region_node(r);
8068 usable_startpfn = PFN_DOWN(r->base);
8069 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
8070 min(usable_startpfn, zone_movable_pfn[nid]) :
8078 * If kernelcore=mirror is specified, ignore movablecore option
8080 if (mirrored_kernelcore) {
8081 bool mem_below_4gb_not_mirrored = false;
8083 for_each_mem_region(r) {
8084 if (memblock_is_mirror(r))
8087 nid = memblock_get_region_node(r);
8089 usable_startpfn = memblock_region_memory_base_pfn(r);
8091 if (usable_startpfn < PHYS_PFN(SZ_4G)) {
8092 mem_below_4gb_not_mirrored = true;
8096 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
8097 min(usable_startpfn, zone_movable_pfn[nid]) :
8101 if (mem_below_4gb_not_mirrored)
8102 pr_warn("This configuration results in unmirrored kernel memory.\n");
8108 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
8109 * amount of necessary memory.
8111 if (required_kernelcore_percent)
8112 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
8114 if (required_movablecore_percent)
8115 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
8119 * If movablecore= was specified, calculate what size of
8120 * kernelcore that corresponds so that memory usable for
8121 * any allocation type is evenly spread. If both kernelcore
8122 * and movablecore are specified, then the value of kernelcore
8123 * will be used for required_kernelcore if it's greater than
8124 * what movablecore would have allowed.
8126 if (required_movablecore) {
8127 unsigned long corepages;
8130 * Round-up so that ZONE_MOVABLE is at least as large as what
8131 * was requested by the user
8133 required_movablecore =
8134 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
8135 required_movablecore = min(totalpages, required_movablecore);
8136 corepages = totalpages - required_movablecore;
8138 required_kernelcore = max(required_kernelcore, corepages);
8142 * If kernelcore was not specified or kernelcore size is larger
8143 * than totalpages, there is no ZONE_MOVABLE.
8145 if (!required_kernelcore || required_kernelcore >= totalpages)
8148 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
8149 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
8152 /* Spread kernelcore memory as evenly as possible throughout nodes */
8153 kernelcore_node = required_kernelcore / usable_nodes;
8154 for_each_node_state(nid, N_MEMORY) {
8155 unsigned long start_pfn, end_pfn;
8158 * Recalculate kernelcore_node if the division per node
8159 * now exceeds what is necessary to satisfy the requested
8160 * amount of memory for the kernel
8162 if (required_kernelcore < kernelcore_node)
8163 kernelcore_node = required_kernelcore / usable_nodes;
8166 * As the map is walked, we track how much memory is usable
8167 * by the kernel using kernelcore_remaining. When it is
8168 * 0, the rest of the node is usable by ZONE_MOVABLE
8170 kernelcore_remaining = kernelcore_node;
8172 /* Go through each range of PFNs within this node */
8173 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
8174 unsigned long size_pages;
8176 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
8177 if (start_pfn >= end_pfn)
8180 /* Account for what is only usable for kernelcore */
8181 if (start_pfn < usable_startpfn) {
8182 unsigned long kernel_pages;
8183 kernel_pages = min(end_pfn, usable_startpfn)
8186 kernelcore_remaining -= min(kernel_pages,
8187 kernelcore_remaining);
8188 required_kernelcore -= min(kernel_pages,
8189 required_kernelcore);
8191 /* Continue if range is now fully accounted */
8192 if (end_pfn <= usable_startpfn) {
8195 * Push zone_movable_pfn to the end so
8196 * that if we have to rebalance
8197 * kernelcore across nodes, we will
8198 * not double account here
8200 zone_movable_pfn[nid] = end_pfn;
8203 start_pfn = usable_startpfn;
8207 * The usable PFN range for ZONE_MOVABLE is from
8208 * start_pfn->end_pfn. Calculate size_pages as the
8209 * number of pages used as kernelcore
8211 size_pages = end_pfn - start_pfn;
8212 if (size_pages > kernelcore_remaining)
8213 size_pages = kernelcore_remaining;
8214 zone_movable_pfn[nid] = start_pfn + size_pages;
8217 * Some kernelcore has been met, update counts and
8218 * break if the kernelcore for this node has been
8221 required_kernelcore -= min(required_kernelcore,
8223 kernelcore_remaining -= size_pages;
8224 if (!kernelcore_remaining)
8230 * If there is still required_kernelcore, we do another pass with one
8231 * less node in the count. This will push zone_movable_pfn[nid] further
8232 * along on the nodes that still have memory until kernelcore is
8236 if (usable_nodes && required_kernelcore > usable_nodes)
8240 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
8241 for (nid = 0; nid < MAX_NUMNODES; nid++) {
8242 unsigned long start_pfn, end_pfn;
8244 zone_movable_pfn[nid] =
8245 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
8247 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
8248 if (zone_movable_pfn[nid] >= end_pfn)
8249 zone_movable_pfn[nid] = 0;
8253 /* restore the node_state */
8254 node_states[N_MEMORY] = saved_node_state;
8257 /* Any regular or high memory on that node ? */
8258 static void check_for_memory(pg_data_t *pgdat, int nid)
8260 enum zone_type zone_type;
8262 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
8263 struct zone *zone = &pgdat->node_zones[zone_type];
8264 if (populated_zone(zone)) {
8265 if (IS_ENABLED(CONFIG_HIGHMEM))
8266 node_set_state(nid, N_HIGH_MEMORY);
8267 if (zone_type <= ZONE_NORMAL)
8268 node_set_state(nid, N_NORMAL_MEMORY);
8275 * Some architectures, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
8276 * such cases we allow max_zone_pfn sorted in the descending order
8278 bool __weak arch_has_descending_max_zone_pfns(void)
8284 * free_area_init - Initialise all pg_data_t and zone data
8285 * @max_zone_pfn: an array of max PFNs for each zone
8287 * This will call free_area_init_node() for each active node in the system.
8288 * Using the page ranges provided by memblock_set_node(), the size of each
8289 * zone in each node and their holes is calculated. If the maximum PFN
8290 * between two adjacent zones match, it is assumed that the zone is empty.
8291 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
8292 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
8293 * starts where the previous one ended. For example, ZONE_DMA32 starts
8294 * at arch_max_dma_pfn.
8296 void __init free_area_init(unsigned long *max_zone_pfn)
8298 unsigned long start_pfn, end_pfn;
8302 /* Record where the zone boundaries are */
8303 memset(arch_zone_lowest_possible_pfn, 0,
8304 sizeof(arch_zone_lowest_possible_pfn));
8305 memset(arch_zone_highest_possible_pfn, 0,
8306 sizeof(arch_zone_highest_possible_pfn));
8308 start_pfn = PHYS_PFN(memblock_start_of_DRAM());
8309 descending = arch_has_descending_max_zone_pfns();
8311 for (i = 0; i < MAX_NR_ZONES; i++) {
8313 zone = MAX_NR_ZONES - i - 1;
8317 if (zone == ZONE_MOVABLE)
8320 end_pfn = max(max_zone_pfn[zone], start_pfn);
8321 arch_zone_lowest_possible_pfn[zone] = start_pfn;
8322 arch_zone_highest_possible_pfn[zone] = end_pfn;
8324 start_pfn = end_pfn;
8327 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
8328 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
8329 find_zone_movable_pfns_for_nodes();
8331 /* Print out the zone ranges */
8332 pr_info("Zone ranges:\n");
8333 for (i = 0; i < MAX_NR_ZONES; i++) {
8334 if (i == ZONE_MOVABLE)
8336 pr_info(" %-8s ", zone_names[i]);
8337 if (arch_zone_lowest_possible_pfn[i] ==
8338 arch_zone_highest_possible_pfn[i])
8341 pr_cont("[mem %#018Lx-%#018Lx]\n",
8342 (u64)arch_zone_lowest_possible_pfn[i]
8344 ((u64)arch_zone_highest_possible_pfn[i]
8345 << PAGE_SHIFT) - 1);
8348 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
8349 pr_info("Movable zone start for each node\n");
8350 for (i = 0; i < MAX_NUMNODES; i++) {
8351 if (zone_movable_pfn[i])
8352 pr_info(" Node %d: %#018Lx\n", i,
8353 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
8357 * Print out the early node map, and initialize the
8358 * subsection-map relative to active online memory ranges to
8359 * enable future "sub-section" extensions of the memory map.
8361 pr_info("Early memory node ranges\n");
8362 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
8363 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
8364 (u64)start_pfn << PAGE_SHIFT,
8365 ((u64)end_pfn << PAGE_SHIFT) - 1);
8366 subsection_map_init(start_pfn, end_pfn - start_pfn);
8369 /* Initialise every node */
8370 mminit_verify_pageflags_layout();
8371 setup_nr_node_ids();
8372 for_each_node(nid) {
8375 if (!node_online(nid)) {
8376 pr_info("Initializing node %d as memoryless\n", nid);
8378 /* Allocator not initialized yet */
8379 pgdat = arch_alloc_nodedata(nid);
8381 pr_err("Cannot allocate %zuB for node %d.\n",
8382 sizeof(*pgdat), nid);
8385 arch_refresh_nodedata(nid, pgdat);
8386 free_area_init_memoryless_node(nid);
8389 * We do not want to confuse userspace by sysfs
8390 * files/directories for node without any memory
8391 * attached to it, so this node is not marked as
8392 * N_MEMORY and not marked online so that no sysfs
8393 * hierarchy will be created via register_one_node for
8394 * it. The pgdat will get fully initialized by
8395 * hotadd_init_pgdat() when memory is hotplugged into
8401 pgdat = NODE_DATA(nid);
8402 free_area_init_node(nid);
8404 /* Any memory on that node */
8405 if (pgdat->node_present_pages)
8406 node_set_state(nid, N_MEMORY);
8407 check_for_memory(pgdat, nid);
8413 static int __init cmdline_parse_core(char *p, unsigned long *core,
8414 unsigned long *percent)
8416 unsigned long long coremem;
8422 /* Value may be a percentage of total memory, otherwise bytes */
8423 coremem = simple_strtoull(p, &endptr, 0);
8424 if (*endptr == '%') {
8425 /* Paranoid check for percent values greater than 100 */
8426 WARN_ON(coremem > 100);
8430 coremem = memparse(p, &p);
8431 /* Paranoid check that UL is enough for the coremem value */
8432 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
8434 *core = coremem >> PAGE_SHIFT;
8441 * kernelcore=size sets the amount of memory for use for allocations that
8442 * cannot be reclaimed or migrated.
8444 static int __init cmdline_parse_kernelcore(char *p)
8446 /* parse kernelcore=mirror */
8447 if (parse_option_str(p, "mirror")) {
8448 mirrored_kernelcore = true;
8452 return cmdline_parse_core(p, &required_kernelcore,
8453 &required_kernelcore_percent);
8457 * movablecore=size sets the amount of memory for use for allocations that
8458 * can be reclaimed or migrated.
8460 static int __init cmdline_parse_movablecore(char *p)
8462 return cmdline_parse_core(p, &required_movablecore,
8463 &required_movablecore_percent);
8466 early_param("kernelcore", cmdline_parse_kernelcore);
8467 early_param("movablecore", cmdline_parse_movablecore);
8469 void adjust_managed_page_count(struct page *page, long count)
8471 atomic_long_add(count, &page_zone(page)->managed_pages);
8472 totalram_pages_add(count);
8473 #ifdef CONFIG_HIGHMEM
8474 if (PageHighMem(page))
8475 totalhigh_pages_add(count);
8478 EXPORT_SYMBOL(adjust_managed_page_count);
8480 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
8483 unsigned long pages = 0;
8485 start = (void *)PAGE_ALIGN((unsigned long)start);
8486 end = (void *)((unsigned long)end & PAGE_MASK);
8487 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
8488 struct page *page = virt_to_page(pos);
8489 void *direct_map_addr;
8492 * 'direct_map_addr' might be different from 'pos'
8493 * because some architectures' virt_to_page()
8494 * work with aliases. Getting the direct map
8495 * address ensures that we get a _writeable_
8496 * alias for the memset().
8498 direct_map_addr = page_address(page);
8500 * Perform a kasan-unchecked memset() since this memory
8501 * has not been initialized.
8503 direct_map_addr = kasan_reset_tag(direct_map_addr);
8504 if ((unsigned int)poison <= 0xFF)
8505 memset(direct_map_addr, poison, PAGE_SIZE);
8507 free_reserved_page(page);
8511 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
8516 void __init mem_init_print_info(void)
8518 unsigned long physpages, codesize, datasize, rosize, bss_size;
8519 unsigned long init_code_size, init_data_size;
8521 physpages = get_num_physpages();
8522 codesize = _etext - _stext;
8523 datasize = _edata - _sdata;
8524 rosize = __end_rodata - __start_rodata;
8525 bss_size = __bss_stop - __bss_start;
8526 init_data_size = __init_end - __init_begin;
8527 init_code_size = _einittext - _sinittext;
8530 * Detect special cases and adjust section sizes accordingly:
8531 * 1) .init.* may be embedded into .data sections
8532 * 2) .init.text.* may be out of [__init_begin, __init_end],
8533 * please refer to arch/tile/kernel/vmlinux.lds.S.
8534 * 3) .rodata.* may be embedded into .text or .data sections.
8536 #define adj_init_size(start, end, size, pos, adj) \
8538 if (&start[0] <= &pos[0] && &pos[0] < &end[0] && size > adj) \
8542 adj_init_size(__init_begin, __init_end, init_data_size,
8543 _sinittext, init_code_size);
8544 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
8545 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
8546 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
8547 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
8549 #undef adj_init_size
8551 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
8552 #ifdef CONFIG_HIGHMEM
8556 K(nr_free_pages()), K(physpages),
8557 codesize / SZ_1K, datasize / SZ_1K, rosize / SZ_1K,
8558 (init_data_size + init_code_size) / SZ_1K, bss_size / SZ_1K,
8559 K(physpages - totalram_pages() - totalcma_pages),
8561 #ifdef CONFIG_HIGHMEM
8562 , K(totalhigh_pages())
8568 * set_dma_reserve - set the specified number of pages reserved in the first zone
8569 * @new_dma_reserve: The number of pages to mark reserved
8571 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
8572 * In the DMA zone, a significant percentage may be consumed by kernel image
8573 * and other unfreeable allocations which can skew the watermarks badly. This
8574 * function may optionally be used to account for unfreeable pages in the
8575 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
8576 * smaller per-cpu batchsize.
8578 void __init set_dma_reserve(unsigned long new_dma_reserve)
8580 dma_reserve = new_dma_reserve;
8583 static int page_alloc_cpu_dead(unsigned int cpu)
8587 lru_add_drain_cpu(cpu);
8588 mlock_page_drain_remote(cpu);
8592 * Spill the event counters of the dead processor
8593 * into the current processors event counters.
8594 * This artificially elevates the count of the current
8597 vm_events_fold_cpu(cpu);
8600 * Zero the differential counters of the dead processor
8601 * so that the vm statistics are consistent.
8603 * This is only okay since the processor is dead and cannot
8604 * race with what we are doing.
8606 cpu_vm_stats_fold(cpu);
8608 for_each_populated_zone(zone)
8609 zone_pcp_update(zone, 0);
8614 static int page_alloc_cpu_online(unsigned int cpu)
8618 for_each_populated_zone(zone)
8619 zone_pcp_update(zone, 1);
8624 int hashdist = HASHDIST_DEFAULT;
8626 static int __init set_hashdist(char *str)
8630 hashdist = simple_strtoul(str, &str, 0);
8633 __setup("hashdist=", set_hashdist);
8636 void __init page_alloc_init(void)
8641 if (num_node_state(N_MEMORY) == 1)
8645 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
8646 "mm/page_alloc:pcp",
8647 page_alloc_cpu_online,
8648 page_alloc_cpu_dead);
8653 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
8654 * or min_free_kbytes changes.
8656 static void calculate_totalreserve_pages(void)
8658 struct pglist_data *pgdat;
8659 unsigned long reserve_pages = 0;
8660 enum zone_type i, j;
8662 for_each_online_pgdat(pgdat) {
8664 pgdat->totalreserve_pages = 0;
8666 for (i = 0; i < MAX_NR_ZONES; i++) {
8667 struct zone *zone = pgdat->node_zones + i;
8669 unsigned long managed_pages = zone_managed_pages(zone);
8671 /* Find valid and maximum lowmem_reserve in the zone */
8672 for (j = i; j < MAX_NR_ZONES; j++) {
8673 if (zone->lowmem_reserve[j] > max)
8674 max = zone->lowmem_reserve[j];
8677 /* we treat the high watermark as reserved pages. */
8678 max += high_wmark_pages(zone);
8680 if (max > managed_pages)
8681 max = managed_pages;
8683 pgdat->totalreserve_pages += max;
8685 reserve_pages += max;
8688 totalreserve_pages = reserve_pages;
8692 * setup_per_zone_lowmem_reserve - called whenever
8693 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
8694 * has a correct pages reserved value, so an adequate number of
8695 * pages are left in the zone after a successful __alloc_pages().
8697 static void setup_per_zone_lowmem_reserve(void)
8699 struct pglist_data *pgdat;
8700 enum zone_type i, j;
8702 for_each_online_pgdat(pgdat) {
8703 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
8704 struct zone *zone = &pgdat->node_zones[i];
8705 int ratio = sysctl_lowmem_reserve_ratio[i];
8706 bool clear = !ratio || !zone_managed_pages(zone);
8707 unsigned long managed_pages = 0;
8709 for (j = i + 1; j < MAX_NR_ZONES; j++) {
8710 struct zone *upper_zone = &pgdat->node_zones[j];
8712 managed_pages += zone_managed_pages(upper_zone);
8715 zone->lowmem_reserve[j] = 0;
8717 zone->lowmem_reserve[j] = managed_pages / ratio;
8722 /* update totalreserve_pages */
8723 calculate_totalreserve_pages();
8726 static void __setup_per_zone_wmarks(void)
8728 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
8729 unsigned long lowmem_pages = 0;
8731 unsigned long flags;
8733 /* Calculate total number of !ZONE_HIGHMEM pages */
8734 for_each_zone(zone) {
8735 if (!is_highmem(zone))
8736 lowmem_pages += zone_managed_pages(zone);
8739 for_each_zone(zone) {
8742 spin_lock_irqsave(&zone->lock, flags);
8743 tmp = (u64)pages_min * zone_managed_pages(zone);
8744 do_div(tmp, lowmem_pages);
8745 if (is_highmem(zone)) {
8747 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
8748 * need highmem pages, so cap pages_min to a small
8751 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
8752 * deltas control async page reclaim, and so should
8753 * not be capped for highmem.
8755 unsigned long min_pages;
8757 min_pages = zone_managed_pages(zone) / 1024;
8758 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
8759 zone->_watermark[WMARK_MIN] = min_pages;
8762 * If it's a lowmem zone, reserve a number of pages
8763 * proportionate to the zone's size.
8765 zone->_watermark[WMARK_MIN] = tmp;
8769 * Set the kswapd watermarks distance according to the
8770 * scale factor in proportion to available memory, but
8771 * ensure a minimum size on small systems.
8773 tmp = max_t(u64, tmp >> 2,
8774 mult_frac(zone_managed_pages(zone),
8775 watermark_scale_factor, 10000));
8777 zone->watermark_boost = 0;
8778 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
8779 zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
8780 zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
8782 spin_unlock_irqrestore(&zone->lock, flags);
8785 /* update totalreserve_pages */
8786 calculate_totalreserve_pages();
8790 * setup_per_zone_wmarks - called when min_free_kbytes changes
8791 * or when memory is hot-{added|removed}
8793 * Ensures that the watermark[min,low,high] values for each zone are set
8794 * correctly with respect to min_free_kbytes.
8796 void setup_per_zone_wmarks(void)
8799 static DEFINE_SPINLOCK(lock);
8802 __setup_per_zone_wmarks();
8806 * The watermark size have changed so update the pcpu batch
8807 * and high limits or the limits may be inappropriate.
8810 zone_pcp_update(zone, 0);
8814 * Initialise min_free_kbytes.
8816 * For small machines we want it small (128k min). For large machines
8817 * we want it large (256MB max). But it is not linear, because network
8818 * bandwidth does not increase linearly with machine size. We use
8820 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8821 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8837 void calculate_min_free_kbytes(void)
8839 unsigned long lowmem_kbytes;
8840 int new_min_free_kbytes;
8842 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8843 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8845 if (new_min_free_kbytes > user_min_free_kbytes)
8846 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
8848 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8849 new_min_free_kbytes, user_min_free_kbytes);
8853 int __meminit init_per_zone_wmark_min(void)
8855 calculate_min_free_kbytes();
8856 setup_per_zone_wmarks();
8857 refresh_zone_stat_thresholds();
8858 setup_per_zone_lowmem_reserve();
8861 setup_min_unmapped_ratio();
8862 setup_min_slab_ratio();
8865 khugepaged_min_free_kbytes_update();
8869 postcore_initcall(init_per_zone_wmark_min)
8872 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8873 * that we can call two helper functions whenever min_free_kbytes
8876 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8877 void *buffer, size_t *length, loff_t *ppos)
8881 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8886 user_min_free_kbytes = min_free_kbytes;
8887 setup_per_zone_wmarks();
8892 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8893 void *buffer, size_t *length, loff_t *ppos)
8897 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8902 setup_per_zone_wmarks();
8908 static void setup_min_unmapped_ratio(void)
8913 for_each_online_pgdat(pgdat)
8914 pgdat->min_unmapped_pages = 0;
8917 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8918 sysctl_min_unmapped_ratio) / 100;
8922 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8923 void *buffer, size_t *length, loff_t *ppos)
8927 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8931 setup_min_unmapped_ratio();
8936 static void setup_min_slab_ratio(void)
8941 for_each_online_pgdat(pgdat)
8942 pgdat->min_slab_pages = 0;
8945 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8946 sysctl_min_slab_ratio) / 100;
8949 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8950 void *buffer, size_t *length, loff_t *ppos)
8954 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8958 setup_min_slab_ratio();
8965 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8966 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8967 * whenever sysctl_lowmem_reserve_ratio changes.
8969 * The reserve ratio obviously has absolutely no relation with the
8970 * minimum watermarks. The lowmem reserve ratio can only make sense
8971 * if in function of the boot time zone sizes.
8973 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8974 void *buffer, size_t *length, loff_t *ppos)
8978 proc_dointvec_minmax(table, write, buffer, length, ppos);
8980 for (i = 0; i < MAX_NR_ZONES; i++) {
8981 if (sysctl_lowmem_reserve_ratio[i] < 1)
8982 sysctl_lowmem_reserve_ratio[i] = 0;
8985 setup_per_zone_lowmem_reserve();
8990 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
8991 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8992 * pagelist can have before it gets flushed back to buddy allocator.
8994 int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
8995 int write, void *buffer, size_t *length, loff_t *ppos)
8998 int old_percpu_pagelist_high_fraction;
9001 mutex_lock(&pcp_batch_high_lock);
9002 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
9004 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
9005 if (!write || ret < 0)
9008 /* Sanity checking to avoid pcp imbalance */
9009 if (percpu_pagelist_high_fraction &&
9010 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
9011 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
9017 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
9020 for_each_populated_zone(zone)
9021 zone_set_pageset_high_and_batch(zone, 0);
9023 mutex_unlock(&pcp_batch_high_lock);
9027 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
9029 * Returns the number of pages that arch has reserved but
9030 * is not known to alloc_large_system_hash().
9032 static unsigned long __init arch_reserved_kernel_pages(void)
9039 * Adaptive scale is meant to reduce sizes of hash tables on large memory
9040 * machines. As memory size is increased the scale is also increased but at
9041 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
9042 * quadruples the scale is increased by one, which means the size of hash table
9043 * only doubles, instead of quadrupling as well.
9044 * Because 32-bit systems cannot have large physical memory, where this scaling
9045 * makes sense, it is disabled on such platforms.
9047 #if __BITS_PER_LONG > 32
9048 #define ADAPT_SCALE_BASE (64ul << 30)
9049 #define ADAPT_SCALE_SHIFT 2
9050 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
9054 * allocate a large system hash table from bootmem
9055 * - it is assumed that the hash table must contain an exact power-of-2
9056 * quantity of entries
9057 * - limit is the number of hash buckets, not the total allocation size
9059 void *__init alloc_large_system_hash(const char *tablename,
9060 unsigned long bucketsize,
9061 unsigned long numentries,
9064 unsigned int *_hash_shift,
9065 unsigned int *_hash_mask,
9066 unsigned long low_limit,
9067 unsigned long high_limit)
9069 unsigned long long max = high_limit;
9070 unsigned long log2qty, size;
9076 /* allow the kernel cmdline to have a say */
9078 /* round applicable memory size up to nearest megabyte */
9079 numentries = nr_kernel_pages;
9080 numentries -= arch_reserved_kernel_pages();
9082 /* It isn't necessary when PAGE_SIZE >= 1MB */
9083 if (PAGE_SIZE < SZ_1M)
9084 numentries = round_up(numentries, SZ_1M / PAGE_SIZE);
9086 #if __BITS_PER_LONG > 32
9088 unsigned long adapt;
9090 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
9091 adapt <<= ADAPT_SCALE_SHIFT)
9096 /* limit to 1 bucket per 2^scale bytes of low memory */
9097 if (scale > PAGE_SHIFT)
9098 numentries >>= (scale - PAGE_SHIFT);
9100 numentries <<= (PAGE_SHIFT - scale);
9102 /* Make sure we've got at least a 0-order allocation.. */
9103 if (unlikely(flags & HASH_SMALL)) {
9104 /* Makes no sense without HASH_EARLY */
9105 WARN_ON(!(flags & HASH_EARLY));
9106 if (!(numentries >> *_hash_shift)) {
9107 numentries = 1UL << *_hash_shift;
9108 BUG_ON(!numentries);
9110 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
9111 numentries = PAGE_SIZE / bucketsize;
9113 numentries = roundup_pow_of_two(numentries);
9115 /* limit allocation size to 1/16 total memory by default */
9117 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
9118 do_div(max, bucketsize);
9120 max = min(max, 0x80000000ULL);
9122 if (numentries < low_limit)
9123 numentries = low_limit;
9124 if (numentries > max)
9127 log2qty = ilog2(numentries);
9129 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
9132 size = bucketsize << log2qty;
9133 if (flags & HASH_EARLY) {
9134 if (flags & HASH_ZERO)
9135 table = memblock_alloc(size, SMP_CACHE_BYTES);
9137 table = memblock_alloc_raw(size,
9139 } else if (get_order(size) >= MAX_ORDER || hashdist) {
9140 table = vmalloc_huge(size, gfp_flags);
9143 huge = is_vm_area_hugepages(table);
9146 * If bucketsize is not a power-of-two, we may free
9147 * some pages at the end of hash table which
9148 * alloc_pages_exact() automatically does
9150 table = alloc_pages_exact(size, gfp_flags);
9151 kmemleak_alloc(table, size, 1, gfp_flags);
9153 } while (!table && size > PAGE_SIZE && --log2qty);
9156 panic("Failed to allocate %s hash table\n", tablename);
9158 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
9159 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
9160 virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
9163 *_hash_shift = log2qty;
9165 *_hash_mask = (1 << log2qty) - 1;
9170 #ifdef CONFIG_CONTIG_ALLOC
9171 #if defined(CONFIG_DYNAMIC_DEBUG) || \
9172 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
9173 /* Usage: See admin-guide/dynamic-debug-howto.rst */
9174 static void alloc_contig_dump_pages(struct list_head *page_list)
9176 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
9178 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
9182 list_for_each_entry(page, page_list, lru)
9183 dump_page(page, "migration failure");
9187 static inline void alloc_contig_dump_pages(struct list_head *page_list)
9192 /* [start, end) must belong to a single zone. */
9193 int __alloc_contig_migrate_range(struct compact_control *cc,
9194 unsigned long start, unsigned long end)
9196 /* This function is based on compact_zone() from compaction.c. */
9197 unsigned int nr_reclaimed;
9198 unsigned long pfn = start;
9199 unsigned int tries = 0;
9201 struct migration_target_control mtc = {
9202 .nid = zone_to_nid(cc->zone),
9203 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
9206 lru_cache_disable();
9208 while (pfn < end || !list_empty(&cc->migratepages)) {
9209 if (fatal_signal_pending(current)) {
9214 if (list_empty(&cc->migratepages)) {
9215 cc->nr_migratepages = 0;
9216 ret = isolate_migratepages_range(cc, pfn, end);
9217 if (ret && ret != -EAGAIN)
9219 pfn = cc->migrate_pfn;
9221 } else if (++tries == 5) {
9226 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
9228 cc->nr_migratepages -= nr_reclaimed;
9230 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
9231 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
9234 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
9235 * to retry again over this error, so do the same here.
9243 if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY)
9244 alloc_contig_dump_pages(&cc->migratepages);
9245 putback_movable_pages(&cc->migratepages);
9252 * alloc_contig_range() -- tries to allocate given range of pages
9253 * @start: start PFN to allocate
9254 * @end: one-past-the-last PFN to allocate
9255 * @migratetype: migratetype of the underlying pageblocks (either
9256 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
9257 * in range must have the same migratetype and it must
9258 * be either of the two.
9259 * @gfp_mask: GFP mask to use during compaction
9261 * The PFN range does not have to be pageblock aligned. The PFN range must
9262 * belong to a single zone.
9264 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
9265 * pageblocks in the range. Once isolated, the pageblocks should not
9266 * be modified by others.
9268 * Return: zero on success or negative error code. On success all
9269 * pages which PFN is in [start, end) are allocated for the caller and
9270 * need to be freed with free_contig_range().
9272 int alloc_contig_range(unsigned long start, unsigned long end,
9273 unsigned migratetype, gfp_t gfp_mask)
9275 unsigned long outer_start, outer_end;
9279 struct compact_control cc = {
9280 .nr_migratepages = 0,
9282 .zone = page_zone(pfn_to_page(start)),
9283 .mode = MIGRATE_SYNC,
9284 .ignore_skip_hint = true,
9285 .no_set_skip_hint = true,
9286 .gfp_mask = current_gfp_context(gfp_mask),
9287 .alloc_contig = true,
9289 INIT_LIST_HEAD(&cc.migratepages);
9292 * What we do here is we mark all pageblocks in range as
9293 * MIGRATE_ISOLATE. Because pageblock and max order pages may
9294 * have different sizes, and due to the way page allocator
9295 * work, start_isolate_page_range() has special handlings for this.
9297 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
9298 * migrate the pages from an unaligned range (ie. pages that
9299 * we are interested in). This will put all the pages in
9300 * range back to page allocator as MIGRATE_ISOLATE.
9302 * When this is done, we take the pages in range from page
9303 * allocator removing them from the buddy system. This way
9304 * page allocator will never consider using them.
9306 * This lets us mark the pageblocks back as
9307 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
9308 * aligned range but not in the unaligned, original range are
9309 * put back to page allocator so that buddy can use them.
9312 ret = start_isolate_page_range(start, end, migratetype, 0, gfp_mask);
9316 drain_all_pages(cc.zone);
9319 * In case of -EBUSY, we'd like to know which page causes problem.
9320 * So, just fall through. test_pages_isolated() has a tracepoint
9321 * which will report the busy page.
9323 * It is possible that busy pages could become available before
9324 * the call to test_pages_isolated, and the range will actually be
9325 * allocated. So, if we fall through be sure to clear ret so that
9326 * -EBUSY is not accidentally used or returned to caller.
9328 ret = __alloc_contig_migrate_range(&cc, start, end);
9329 if (ret && ret != -EBUSY)
9334 * Pages from [start, end) are within a pageblock_nr_pages
9335 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
9336 * more, all pages in [start, end) are free in page allocator.
9337 * What we are going to do is to allocate all pages from
9338 * [start, end) (that is remove them from page allocator).
9340 * The only problem is that pages at the beginning and at the
9341 * end of interesting range may be not aligned with pages that
9342 * page allocator holds, ie. they can be part of higher order
9343 * pages. Because of this, we reserve the bigger range and
9344 * once this is done free the pages we are not interested in.
9346 * We don't have to hold zone->lock here because the pages are
9347 * isolated thus they won't get removed from buddy.
9351 outer_start = start;
9352 while (!PageBuddy(pfn_to_page(outer_start))) {
9353 if (++order >= MAX_ORDER) {
9354 outer_start = start;
9357 outer_start &= ~0UL << order;
9360 if (outer_start != start) {
9361 order = buddy_order(pfn_to_page(outer_start));
9364 * outer_start page could be small order buddy page and
9365 * it doesn't include start page. Adjust outer_start
9366 * in this case to report failed page properly
9367 * on tracepoint in test_pages_isolated()
9369 if (outer_start + (1UL << order) <= start)
9370 outer_start = start;
9373 /* Make sure the range is really isolated. */
9374 if (test_pages_isolated(outer_start, end, 0)) {
9379 /* Grab isolated pages from freelists. */
9380 outer_end = isolate_freepages_range(&cc, outer_start, end);
9386 /* Free head and tail (if any) */
9387 if (start != outer_start)
9388 free_contig_range(outer_start, start - outer_start);
9389 if (end != outer_end)
9390 free_contig_range(end, outer_end - end);
9393 undo_isolate_page_range(start, end, migratetype);
9396 EXPORT_SYMBOL(alloc_contig_range);
9398 static int __alloc_contig_pages(unsigned long start_pfn,
9399 unsigned long nr_pages, gfp_t gfp_mask)
9401 unsigned long end_pfn = start_pfn + nr_pages;
9403 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
9407 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
9408 unsigned long nr_pages)
9410 unsigned long i, end_pfn = start_pfn + nr_pages;
9413 for (i = start_pfn; i < end_pfn; i++) {
9414 page = pfn_to_online_page(i);
9418 if (page_zone(page) != z)
9421 if (PageReserved(page))
9430 static bool zone_spans_last_pfn(const struct zone *zone,
9431 unsigned long start_pfn, unsigned long nr_pages)
9433 unsigned long last_pfn = start_pfn + nr_pages - 1;
9435 return zone_spans_pfn(zone, last_pfn);
9439 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
9440 * @nr_pages: Number of contiguous pages to allocate
9441 * @gfp_mask: GFP mask to limit search and used during compaction
9443 * @nodemask: Mask for other possible nodes
9445 * This routine is a wrapper around alloc_contig_range(). It scans over zones
9446 * on an applicable zonelist to find a contiguous pfn range which can then be
9447 * tried for allocation with alloc_contig_range(). This routine is intended
9448 * for allocation requests which can not be fulfilled with the buddy allocator.
9450 * The allocated memory is always aligned to a page boundary. If nr_pages is a
9451 * power of two, then allocated range is also guaranteed to be aligned to same
9452 * nr_pages (e.g. 1GB request would be aligned to 1GB).
9454 * Allocated pages can be freed with free_contig_range() or by manually calling
9455 * __free_page() on each allocated page.
9457 * Return: pointer to contiguous pages on success, or NULL if not successful.
9459 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
9460 int nid, nodemask_t *nodemask)
9462 unsigned long ret, pfn, flags;
9463 struct zonelist *zonelist;
9467 zonelist = node_zonelist(nid, gfp_mask);
9468 for_each_zone_zonelist_nodemask(zone, z, zonelist,
9469 gfp_zone(gfp_mask), nodemask) {
9470 spin_lock_irqsave(&zone->lock, flags);
9472 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
9473 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
9474 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
9476 * We release the zone lock here because
9477 * alloc_contig_range() will also lock the zone
9478 * at some point. If there's an allocation
9479 * spinning on this lock, it may win the race
9480 * and cause alloc_contig_range() to fail...
9482 spin_unlock_irqrestore(&zone->lock, flags);
9483 ret = __alloc_contig_pages(pfn, nr_pages,
9486 return pfn_to_page(pfn);
9487 spin_lock_irqsave(&zone->lock, flags);
9491 spin_unlock_irqrestore(&zone->lock, flags);
9495 #endif /* CONFIG_CONTIG_ALLOC */
9497 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
9499 unsigned long count = 0;
9501 for (; nr_pages--; pfn++) {
9502 struct page *page = pfn_to_page(pfn);
9504 count += page_count(page) != 1;
9507 WARN(count != 0, "%lu pages are still in use!\n", count);
9509 EXPORT_SYMBOL(free_contig_range);
9512 * Effectively disable pcplists for the zone by setting the high limit to 0
9513 * and draining all cpus. A concurrent page freeing on another CPU that's about
9514 * to put the page on pcplist will either finish before the drain and the page
9515 * will be drained, or observe the new high limit and skip the pcplist.
9517 * Must be paired with a call to zone_pcp_enable().
9519 void zone_pcp_disable(struct zone *zone)
9521 mutex_lock(&pcp_batch_high_lock);
9522 __zone_set_pageset_high_and_batch(zone, 0, 1);
9523 __drain_all_pages(zone, true);
9526 void zone_pcp_enable(struct zone *zone)
9528 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
9529 mutex_unlock(&pcp_batch_high_lock);
9532 void zone_pcp_reset(struct zone *zone)
9535 struct per_cpu_zonestat *pzstats;
9537 if (zone->per_cpu_pageset != &boot_pageset) {
9538 for_each_online_cpu(cpu) {
9539 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
9540 drain_zonestat(zone, pzstats);
9542 free_percpu(zone->per_cpu_pageset);
9543 zone->per_cpu_pageset = &boot_pageset;
9544 if (zone->per_cpu_zonestats != &boot_zonestats) {
9545 free_percpu(zone->per_cpu_zonestats);
9546 zone->per_cpu_zonestats = &boot_zonestats;
9551 #ifdef CONFIG_MEMORY_HOTREMOVE
9553 * All pages in the range must be in a single zone, must not contain holes,
9554 * must span full sections, and must be isolated before calling this function.
9556 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
9558 unsigned long pfn = start_pfn;
9562 unsigned long flags;
9564 offline_mem_sections(pfn, end_pfn);
9565 zone = page_zone(pfn_to_page(pfn));
9566 spin_lock_irqsave(&zone->lock, flags);
9567 while (pfn < end_pfn) {
9568 page = pfn_to_page(pfn);
9570 * The HWPoisoned page may be not in buddy system, and
9571 * page_count() is not 0.
9573 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
9578 * At this point all remaining PageOffline() pages have a
9579 * reference count of 0 and can simply be skipped.
9581 if (PageOffline(page)) {
9582 BUG_ON(page_count(page));
9583 BUG_ON(PageBuddy(page));
9588 BUG_ON(page_count(page));
9589 BUG_ON(!PageBuddy(page));
9590 order = buddy_order(page);
9591 del_page_from_free_list(page, zone, order);
9592 pfn += (1 << order);
9594 spin_unlock_irqrestore(&zone->lock, flags);
9599 * This function returns a stable result only if called under zone lock.
9601 bool is_free_buddy_page(struct page *page)
9603 unsigned long pfn = page_to_pfn(page);
9606 for (order = 0; order < MAX_ORDER; order++) {
9607 struct page *page_head = page - (pfn & ((1 << order) - 1));
9609 if (PageBuddy(page_head) &&
9610 buddy_order_unsafe(page_head) >= order)
9614 return order < MAX_ORDER;
9616 EXPORT_SYMBOL(is_free_buddy_page);
9618 #ifdef CONFIG_MEMORY_FAILURE
9620 * Break down a higher-order page in sub-pages, and keep our target out of
9623 static void break_down_buddy_pages(struct zone *zone, struct page *page,
9624 struct page *target, int low, int high,
9627 unsigned long size = 1 << high;
9628 struct page *current_buddy, *next_page;
9630 while (high > low) {
9634 if (target >= &page[size]) {
9635 next_page = page + size;
9636 current_buddy = page;
9639 current_buddy = page + size;
9642 if (set_page_guard(zone, current_buddy, high, migratetype))
9645 if (current_buddy != target) {
9646 add_to_free_list(current_buddy, zone, high, migratetype);
9647 set_buddy_order(current_buddy, high);
9654 * Take a page that will be marked as poisoned off the buddy allocator.
9656 bool take_page_off_buddy(struct page *page)
9658 struct zone *zone = page_zone(page);
9659 unsigned long pfn = page_to_pfn(page);
9660 unsigned long flags;
9664 spin_lock_irqsave(&zone->lock, flags);
9665 for (order = 0; order < MAX_ORDER; order++) {
9666 struct page *page_head = page - (pfn & ((1 << order) - 1));
9667 int page_order = buddy_order(page_head);
9669 if (PageBuddy(page_head) && page_order >= order) {
9670 unsigned long pfn_head = page_to_pfn(page_head);
9671 int migratetype = get_pfnblock_migratetype(page_head,
9674 del_page_from_free_list(page_head, zone, page_order);
9675 break_down_buddy_pages(zone, page_head, page, 0,
9676 page_order, migratetype);
9677 SetPageHWPoisonTakenOff(page);
9678 if (!is_migrate_isolate(migratetype))
9679 __mod_zone_freepage_state(zone, -1, migratetype);
9683 if (page_count(page_head) > 0)
9686 spin_unlock_irqrestore(&zone->lock, flags);
9691 * Cancel takeoff done by take_page_off_buddy().
9693 bool put_page_back_buddy(struct page *page)
9695 struct zone *zone = page_zone(page);
9696 unsigned long pfn = page_to_pfn(page);
9697 unsigned long flags;
9698 int migratetype = get_pfnblock_migratetype(page, pfn);
9701 spin_lock_irqsave(&zone->lock, flags);
9702 if (put_page_testzero(page)) {
9703 ClearPageHWPoisonTakenOff(page);
9704 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
9705 if (TestClearPageHWPoison(page)) {
9709 spin_unlock_irqrestore(&zone->lock, flags);
9715 #ifdef CONFIG_ZONE_DMA
9716 bool has_managed_dma(void)
9718 struct pglist_data *pgdat;
9720 for_each_online_pgdat(pgdat) {
9721 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
9723 if (managed_zone(zone))
9728 #endif /* CONFIG_ZONE_DMA */