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/interrupt.h>
23 #include <linux/pagemap.h>
24 #include <linux/jiffies.h>
25 #include <linux/memblock.h>
26 #include <linux/compiler.h>
27 #include <linux/kernel.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/memremap.h>
46 #include <linux/stop_machine.h>
47 #include <linux/random.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <trace/events/oom.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/sched/mm.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
66 #include <linux/memcontrol.h>
67 #include <linux/ftrace.h>
68 #include <linux/lockdep.h>
69 #include <linux/nmi.h>
70 #include <linux/psi.h>
72 #include <asm/sections.h>
73 #include <asm/tlbflush.h>
74 #include <asm/div64.h>
77 #include "page_reporting.h"
79 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
80 static DEFINE_MUTEX(pcp_batch_high_lock);
81 #define MIN_PERCPU_PAGELIST_FRACTION (8)
83 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
84 DEFINE_PER_CPU(int, numa_node);
85 EXPORT_PER_CPU_SYMBOL(numa_node);
88 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
90 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
92 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
93 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
94 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
95 * defined in <linux/topology.h>.
97 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
98 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
101 /* work_structs for global per-cpu drains */
104 struct work_struct work;
106 static DEFINE_MUTEX(pcpu_drain_mutex);
107 static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
109 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
110 volatile unsigned long latent_entropy __latent_entropy;
111 EXPORT_SYMBOL(latent_entropy);
115 * Array of node states.
117 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
118 [N_POSSIBLE] = NODE_MASK_ALL,
119 [N_ONLINE] = { { [0] = 1UL } },
121 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
122 #ifdef CONFIG_HIGHMEM
123 [N_HIGH_MEMORY] = { { [0] = 1UL } },
125 [N_MEMORY] = { { [0] = 1UL } },
126 [N_CPU] = { { [0] = 1UL } },
129 EXPORT_SYMBOL(node_states);
131 atomic_long_t _totalram_pages __read_mostly;
132 EXPORT_SYMBOL(_totalram_pages);
133 unsigned long totalreserve_pages __read_mostly;
134 unsigned long totalcma_pages __read_mostly;
136 int percpu_pagelist_fraction;
137 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
138 #ifdef CONFIG_INIT_ON_ALLOC_DEFAULT_ON
139 DEFINE_STATIC_KEY_TRUE(init_on_alloc);
141 DEFINE_STATIC_KEY_FALSE(init_on_alloc);
143 EXPORT_SYMBOL(init_on_alloc);
145 #ifdef CONFIG_INIT_ON_FREE_DEFAULT_ON
146 DEFINE_STATIC_KEY_TRUE(init_on_free);
148 DEFINE_STATIC_KEY_FALSE(init_on_free);
150 EXPORT_SYMBOL(init_on_free);
152 static int __init early_init_on_alloc(char *buf)
159 ret = kstrtobool(buf, &bool_result);
160 if (bool_result && page_poisoning_enabled())
161 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_alloc\n");
163 static_branch_enable(&init_on_alloc);
165 static_branch_disable(&init_on_alloc);
168 early_param("init_on_alloc", early_init_on_alloc);
170 static int __init early_init_on_free(char *buf)
177 ret = kstrtobool(buf, &bool_result);
178 if (bool_result && page_poisoning_enabled())
179 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_free\n");
181 static_branch_enable(&init_on_free);
183 static_branch_disable(&init_on_free);
186 early_param("init_on_free", early_init_on_free);
189 * A cached value of the page's pageblock's migratetype, used when the page is
190 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
191 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
192 * Also the migratetype set in the page does not necessarily match the pcplist
193 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
194 * other index - this ensures that it will be put on the correct CMA freelist.
196 static inline int get_pcppage_migratetype(struct page *page)
201 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
203 page->index = migratetype;
206 #ifdef CONFIG_PM_SLEEP
208 * The following functions are used by the suspend/hibernate code to temporarily
209 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
210 * while devices are suspended. To avoid races with the suspend/hibernate code,
211 * they should always be called with system_transition_mutex held
212 * (gfp_allowed_mask also should only be modified with system_transition_mutex
213 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
214 * with that modification).
217 static gfp_t saved_gfp_mask;
219 void pm_restore_gfp_mask(void)
221 WARN_ON(!mutex_is_locked(&system_transition_mutex));
222 if (saved_gfp_mask) {
223 gfp_allowed_mask = saved_gfp_mask;
228 void pm_restrict_gfp_mask(void)
230 WARN_ON(!mutex_is_locked(&system_transition_mutex));
231 WARN_ON(saved_gfp_mask);
232 saved_gfp_mask = gfp_allowed_mask;
233 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
236 bool pm_suspended_storage(void)
238 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
242 #endif /* CONFIG_PM_SLEEP */
244 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
245 unsigned int pageblock_order __read_mostly;
248 static void __free_pages_ok(struct page *page, unsigned int order);
251 * results with 256, 32 in the lowmem_reserve sysctl:
252 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
253 * 1G machine -> (16M dma, 784M normal, 224M high)
254 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
255 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
256 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
258 * TBD: should special case ZONE_DMA32 machines here - in those we normally
259 * don't need any ZONE_NORMAL reservation
261 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
262 #ifdef CONFIG_ZONE_DMA
265 #ifdef CONFIG_ZONE_DMA32
269 #ifdef CONFIG_HIGHMEM
275 static char * const zone_names[MAX_NR_ZONES] = {
276 #ifdef CONFIG_ZONE_DMA
279 #ifdef CONFIG_ZONE_DMA32
283 #ifdef CONFIG_HIGHMEM
287 #ifdef CONFIG_ZONE_DEVICE
292 const char * const migratetype_names[MIGRATE_TYPES] = {
300 #ifdef CONFIG_MEMORY_ISOLATION
305 compound_page_dtor * const compound_page_dtors[] = {
308 #ifdef CONFIG_HUGETLB_PAGE
311 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
316 int min_free_kbytes = 1024;
317 int user_min_free_kbytes = -1;
318 #ifdef CONFIG_DISCONTIGMEM
320 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
321 * are not on separate NUMA nodes. Functionally this works but with
322 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
323 * quite small. By default, do not boost watermarks on discontigmem as in
324 * many cases very high-order allocations like THP are likely to be
325 * unsupported and the premature reclaim offsets the advantage of long-term
326 * fragmentation avoidance.
328 int watermark_boost_factor __read_mostly;
330 int watermark_boost_factor __read_mostly = 15000;
332 int watermark_scale_factor = 10;
334 static unsigned long nr_kernel_pages __initdata;
335 static unsigned long nr_all_pages __initdata;
336 static unsigned long dma_reserve __initdata;
338 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
339 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
340 static unsigned long required_kernelcore __initdata;
341 static unsigned long required_kernelcore_percent __initdata;
342 static unsigned long required_movablecore __initdata;
343 static unsigned long required_movablecore_percent __initdata;
344 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
345 static bool mirrored_kernelcore __meminitdata;
347 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
349 EXPORT_SYMBOL(movable_zone);
352 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
353 unsigned int nr_online_nodes __read_mostly = 1;
354 EXPORT_SYMBOL(nr_node_ids);
355 EXPORT_SYMBOL(nr_online_nodes);
358 int page_group_by_mobility_disabled __read_mostly;
360 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
362 * During boot we initialize deferred pages on-demand, as needed, but once
363 * page_alloc_init_late() has finished, the deferred pages are all initialized,
364 * and we can permanently disable that path.
366 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
369 * Calling kasan_free_pages() only after deferred memory initialization
370 * has completed. Poisoning pages during deferred memory init will greatly
371 * lengthen the process and cause problem in large memory systems as the
372 * deferred pages initialization is done with interrupt disabled.
374 * Assuming that there will be no reference to those newly initialized
375 * pages before they are ever allocated, this should have no effect on
376 * KASAN memory tracking as the poison will be properly inserted at page
377 * allocation time. The only corner case is when pages are allocated by
378 * on-demand allocation and then freed again before the deferred pages
379 * initialization is done, but this is not likely to happen.
381 static inline void kasan_free_nondeferred_pages(struct page *page, int order)
383 if (!static_branch_unlikely(&deferred_pages))
384 kasan_free_pages(page, order);
387 /* Returns true if the struct page for the pfn is uninitialised */
388 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
390 int nid = early_pfn_to_nid(pfn);
392 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
399 * Returns true when the remaining initialisation should be deferred until
400 * later in the boot cycle when it can be parallelised.
402 static bool __meminit
403 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
405 static unsigned long prev_end_pfn, nr_initialised;
408 * prev_end_pfn static that contains the end of previous zone
409 * No need to protect because called very early in boot before smp_init.
411 if (prev_end_pfn != end_pfn) {
412 prev_end_pfn = end_pfn;
416 /* Always populate low zones for address-constrained allocations */
417 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
421 * We start only with one section of pages, more pages are added as
422 * needed until the rest of deferred pages are initialized.
425 if ((nr_initialised > PAGES_PER_SECTION) &&
426 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
427 NODE_DATA(nid)->first_deferred_pfn = pfn;
433 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
435 static inline bool early_page_uninitialised(unsigned long pfn)
440 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
446 /* Return a pointer to the bitmap storing bits affecting a block of pages */
447 static inline unsigned long *get_pageblock_bitmap(struct page *page,
450 #ifdef CONFIG_SPARSEMEM
451 return section_to_usemap(__pfn_to_section(pfn));
453 return page_zone(page)->pageblock_flags;
454 #endif /* CONFIG_SPARSEMEM */
457 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
459 #ifdef CONFIG_SPARSEMEM
460 pfn &= (PAGES_PER_SECTION-1);
461 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
463 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
464 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
465 #endif /* CONFIG_SPARSEMEM */
469 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
470 * @page: The page within the block of interest
471 * @pfn: The target page frame number
472 * @end_bitidx: The last bit of interest to retrieve
473 * @mask: mask of bits that the caller is interested in
475 * Return: pageblock_bits flags
477 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
479 unsigned long end_bitidx,
482 unsigned long *bitmap;
483 unsigned long bitidx, word_bitidx;
486 bitmap = get_pageblock_bitmap(page, pfn);
487 bitidx = pfn_to_bitidx(page, pfn);
488 word_bitidx = bitidx / BITS_PER_LONG;
489 bitidx &= (BITS_PER_LONG-1);
491 word = bitmap[word_bitidx];
492 bitidx += end_bitidx;
493 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
496 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
497 unsigned long end_bitidx,
500 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
503 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
505 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
509 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
510 * @page: The page within the block of interest
511 * @flags: The flags to set
512 * @pfn: The target page frame number
513 * @end_bitidx: The last bit of interest
514 * @mask: mask of bits that the caller is interested in
516 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
518 unsigned long end_bitidx,
521 unsigned long *bitmap;
522 unsigned long bitidx, word_bitidx;
523 unsigned long old_word, word;
525 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
526 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
528 bitmap = get_pageblock_bitmap(page, pfn);
529 bitidx = pfn_to_bitidx(page, pfn);
530 word_bitidx = bitidx / BITS_PER_LONG;
531 bitidx &= (BITS_PER_LONG-1);
533 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
535 bitidx += end_bitidx;
536 mask <<= (BITS_PER_LONG - bitidx - 1);
537 flags <<= (BITS_PER_LONG - bitidx - 1);
539 word = READ_ONCE(bitmap[word_bitidx]);
541 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
542 if (word == old_word)
548 void set_pageblock_migratetype(struct page *page, int migratetype)
550 if (unlikely(page_group_by_mobility_disabled &&
551 migratetype < MIGRATE_PCPTYPES))
552 migratetype = MIGRATE_UNMOVABLE;
554 set_pageblock_flags_group(page, (unsigned long)migratetype,
555 PB_migrate, PB_migrate_end);
558 #ifdef CONFIG_DEBUG_VM
559 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
563 unsigned long pfn = page_to_pfn(page);
564 unsigned long sp, start_pfn;
567 seq = zone_span_seqbegin(zone);
568 start_pfn = zone->zone_start_pfn;
569 sp = zone->spanned_pages;
570 if (!zone_spans_pfn(zone, pfn))
572 } while (zone_span_seqretry(zone, seq));
575 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
576 pfn, zone_to_nid(zone), zone->name,
577 start_pfn, start_pfn + sp);
582 static int page_is_consistent(struct zone *zone, struct page *page)
584 if (!pfn_valid_within(page_to_pfn(page)))
586 if (zone != page_zone(page))
592 * Temporary debugging check for pages not lying within a given zone.
594 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
596 if (page_outside_zone_boundaries(zone, page))
598 if (!page_is_consistent(zone, page))
604 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
610 static void bad_page(struct page *page, const char *reason,
611 unsigned long bad_flags)
613 static unsigned long resume;
614 static unsigned long nr_shown;
615 static unsigned long nr_unshown;
618 * Allow a burst of 60 reports, then keep quiet for that minute;
619 * or allow a steady drip of one report per second.
621 if (nr_shown == 60) {
622 if (time_before(jiffies, resume)) {
628 "BUG: Bad page state: %lu messages suppressed\n",
635 resume = jiffies + 60 * HZ;
637 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
638 current->comm, page_to_pfn(page));
639 __dump_page(page, reason);
640 bad_flags &= page->flags;
642 pr_alert("bad because of flags: %#lx(%pGp)\n",
643 bad_flags, &bad_flags);
644 dump_page_owner(page);
649 /* Leave bad fields for debug, except PageBuddy could make trouble */
650 page_mapcount_reset(page); /* remove PageBuddy */
651 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
655 * Higher-order pages are called "compound pages". They are structured thusly:
657 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
659 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
660 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
662 * The first tail page's ->compound_dtor holds the offset in array of compound
663 * page destructors. See compound_page_dtors.
665 * The first tail page's ->compound_order holds the order of allocation.
666 * This usage means that zero-order pages may not be compound.
669 void free_compound_page(struct page *page)
671 mem_cgroup_uncharge(page);
672 __free_pages_ok(page, compound_order(page));
675 void prep_compound_page(struct page *page, unsigned int order)
678 int nr_pages = 1 << order;
680 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
681 set_compound_order(page, order);
683 for (i = 1; i < nr_pages; i++) {
684 struct page *p = page + i;
685 set_page_count(p, 0);
686 p->mapping = TAIL_MAPPING;
687 set_compound_head(p, page);
689 atomic_set(compound_mapcount_ptr(page), -1);
690 if (hpage_pincount_available(page))
691 atomic_set(compound_pincount_ptr(page), 0);
694 #ifdef CONFIG_DEBUG_PAGEALLOC
695 unsigned int _debug_guardpage_minorder;
697 bool _debug_pagealloc_enabled_early __read_mostly
698 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
699 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
700 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
701 EXPORT_SYMBOL(_debug_pagealloc_enabled);
703 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
705 static int __init early_debug_pagealloc(char *buf)
707 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
709 early_param("debug_pagealloc", early_debug_pagealloc);
711 void init_debug_pagealloc(void)
713 if (!debug_pagealloc_enabled())
716 static_branch_enable(&_debug_pagealloc_enabled);
718 if (!debug_guardpage_minorder())
721 static_branch_enable(&_debug_guardpage_enabled);
724 static int __init debug_guardpage_minorder_setup(char *buf)
728 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
729 pr_err("Bad debug_guardpage_minorder value\n");
732 _debug_guardpage_minorder = res;
733 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
736 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
738 static inline bool set_page_guard(struct zone *zone, struct page *page,
739 unsigned int order, int migratetype)
741 if (!debug_guardpage_enabled())
744 if (order >= debug_guardpage_minorder())
747 __SetPageGuard(page);
748 INIT_LIST_HEAD(&page->lru);
749 set_page_private(page, order);
750 /* Guard pages are not available for any usage */
751 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
756 static inline void clear_page_guard(struct zone *zone, struct page *page,
757 unsigned int order, int migratetype)
759 if (!debug_guardpage_enabled())
762 __ClearPageGuard(page);
764 set_page_private(page, 0);
765 if (!is_migrate_isolate(migratetype))
766 __mod_zone_freepage_state(zone, (1 << order), migratetype);
769 static inline bool set_page_guard(struct zone *zone, struct page *page,
770 unsigned int order, int migratetype) { return false; }
771 static inline void clear_page_guard(struct zone *zone, struct page *page,
772 unsigned int order, int migratetype) {}
775 static inline void set_page_order(struct page *page, unsigned int order)
777 set_page_private(page, order);
778 __SetPageBuddy(page);
782 * This function checks whether a page is free && is the buddy
783 * we can coalesce a page and its buddy if
784 * (a) the buddy is not in a hole (check before calling!) &&
785 * (b) the buddy is in the buddy system &&
786 * (c) a page and its buddy have the same order &&
787 * (d) a page and its buddy are in the same zone.
789 * For recording whether a page is in the buddy system, we set PageBuddy.
790 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
792 * For recording page's order, we use page_private(page).
794 static inline bool page_is_buddy(struct page *page, struct page *buddy,
797 if (!page_is_guard(buddy) && !PageBuddy(buddy))
800 if (page_order(buddy) != order)
804 * zone check is done late to avoid uselessly calculating
805 * zone/node ids for pages that could never merge.
807 if (page_zone_id(page) != page_zone_id(buddy))
810 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
815 #ifdef CONFIG_COMPACTION
816 static inline struct capture_control *task_capc(struct zone *zone)
818 struct capture_control *capc = current->capture_control;
821 !(current->flags & PF_KTHREAD) &&
823 capc->cc->zone == zone &&
824 capc->cc->direct_compaction ? capc : NULL;
828 compaction_capture(struct capture_control *capc, struct page *page,
829 int order, int migratetype)
831 if (!capc || order != capc->cc->order)
834 /* Do not accidentally pollute CMA or isolated regions*/
835 if (is_migrate_cma(migratetype) ||
836 is_migrate_isolate(migratetype))
840 * Do not let lower order allocations polluate a movable pageblock.
841 * This might let an unmovable request use a reclaimable pageblock
842 * and vice-versa but no more than normal fallback logic which can
843 * have trouble finding a high-order free page.
845 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
853 static inline struct capture_control *task_capc(struct zone *zone)
859 compaction_capture(struct capture_control *capc, struct page *page,
860 int order, int migratetype)
864 #endif /* CONFIG_COMPACTION */
866 /* Used for pages not on another list */
867 static inline void add_to_free_list(struct page *page, struct zone *zone,
868 unsigned int order, int migratetype)
870 struct free_area *area = &zone->free_area[order];
872 list_add(&page->lru, &area->free_list[migratetype]);
876 /* Used for pages not on another list */
877 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
878 unsigned int order, int migratetype)
880 struct free_area *area = &zone->free_area[order];
882 list_add_tail(&page->lru, &area->free_list[migratetype]);
886 /* Used for pages which are on another list */
887 static inline void move_to_free_list(struct page *page, struct zone *zone,
888 unsigned int order, int migratetype)
890 struct free_area *area = &zone->free_area[order];
892 list_move(&page->lru, &area->free_list[migratetype]);
895 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
898 /* clear reported state and update reported page count */
899 if (page_reported(page))
900 __ClearPageReported(page);
902 list_del(&page->lru);
903 __ClearPageBuddy(page);
904 set_page_private(page, 0);
905 zone->free_area[order].nr_free--;
909 * If this is not the largest possible page, check if the buddy
910 * of the next-highest order is free. If it is, it's possible
911 * that pages are being freed that will coalesce soon. In case,
912 * that is happening, add the free page to the tail of the list
913 * so it's less likely to be used soon and more likely to be merged
914 * as a higher order page
917 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
918 struct page *page, unsigned int order)
920 struct page *higher_page, *higher_buddy;
921 unsigned long combined_pfn;
923 if (order >= MAX_ORDER - 2)
926 if (!pfn_valid_within(buddy_pfn))
929 combined_pfn = buddy_pfn & pfn;
930 higher_page = page + (combined_pfn - pfn);
931 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
932 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
934 return pfn_valid_within(buddy_pfn) &&
935 page_is_buddy(higher_page, higher_buddy, order + 1);
939 * Freeing function for a buddy system allocator.
941 * The concept of a buddy system is to maintain direct-mapped table
942 * (containing bit values) for memory blocks of various "orders".
943 * The bottom level table contains the map for the smallest allocatable
944 * units of memory (here, pages), and each level above it describes
945 * pairs of units from the levels below, hence, "buddies".
946 * At a high level, all that happens here is marking the table entry
947 * at the bottom level available, and propagating the changes upward
948 * as necessary, plus some accounting needed to play nicely with other
949 * parts of the VM system.
950 * At each level, we keep a list of pages, which are heads of continuous
951 * free pages of length of (1 << order) and marked with PageBuddy.
952 * Page's order is recorded in page_private(page) field.
953 * So when we are allocating or freeing one, we can derive the state of the
954 * other. That is, if we allocate a small block, and both were
955 * free, the remainder of the region must be split into blocks.
956 * If a block is freed, and its buddy is also free, then this
957 * triggers coalescing into a block of larger size.
962 static inline void __free_one_page(struct page *page,
964 struct zone *zone, unsigned int order,
965 int migratetype, bool report)
967 struct capture_control *capc = task_capc(zone);
968 unsigned long uninitialized_var(buddy_pfn);
969 unsigned long combined_pfn;
970 unsigned int max_order;
974 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
976 VM_BUG_ON(!zone_is_initialized(zone));
977 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
979 VM_BUG_ON(migratetype == -1);
980 if (likely(!is_migrate_isolate(migratetype)))
981 __mod_zone_freepage_state(zone, 1 << order, migratetype);
983 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
984 VM_BUG_ON_PAGE(bad_range(zone, page), page);
987 while (order < max_order - 1) {
988 if (compaction_capture(capc, page, order, migratetype)) {
989 __mod_zone_freepage_state(zone, -(1 << order),
993 buddy_pfn = __find_buddy_pfn(pfn, order);
994 buddy = page + (buddy_pfn - pfn);
996 if (!pfn_valid_within(buddy_pfn))
998 if (!page_is_buddy(page, buddy, order))
1001 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1002 * merge with it and move up one order.
1004 if (page_is_guard(buddy))
1005 clear_page_guard(zone, buddy, order, migratetype);
1007 del_page_from_free_list(buddy, zone, order);
1008 combined_pfn = buddy_pfn & pfn;
1009 page = page + (combined_pfn - pfn);
1013 if (max_order < MAX_ORDER) {
1014 /* If we are here, it means order is >= pageblock_order.
1015 * We want to prevent merge between freepages on isolate
1016 * pageblock and normal pageblock. Without this, pageblock
1017 * isolation could cause incorrect freepage or CMA accounting.
1019 * We don't want to hit this code for the more frequent
1020 * low-order merging.
1022 if (unlikely(has_isolate_pageblock(zone))) {
1025 buddy_pfn = __find_buddy_pfn(pfn, order);
1026 buddy = page + (buddy_pfn - pfn);
1027 buddy_mt = get_pageblock_migratetype(buddy);
1029 if (migratetype != buddy_mt
1030 && (is_migrate_isolate(migratetype) ||
1031 is_migrate_isolate(buddy_mt)))
1035 goto continue_merging;
1039 set_page_order(page, order);
1041 if (is_shuffle_order(order))
1042 to_tail = shuffle_pick_tail();
1044 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1047 add_to_free_list_tail(page, zone, order, migratetype);
1049 add_to_free_list(page, zone, order, migratetype);
1051 /* Notify page reporting subsystem of freed page */
1053 page_reporting_notify_free(order);
1057 * A bad page could be due to a number of fields. Instead of multiple branches,
1058 * try and check multiple fields with one check. The caller must do a detailed
1059 * check if necessary.
1061 static inline bool page_expected_state(struct page *page,
1062 unsigned long check_flags)
1064 if (unlikely(atomic_read(&page->_mapcount) != -1))
1067 if (unlikely((unsigned long)page->mapping |
1068 page_ref_count(page) |
1070 (unsigned long)page->mem_cgroup |
1072 (page->flags & check_flags)))
1078 static void free_pages_check_bad(struct page *page)
1080 const char *bad_reason;
1081 unsigned long bad_flags;
1086 if (unlikely(atomic_read(&page->_mapcount) != -1))
1087 bad_reason = "nonzero mapcount";
1088 if (unlikely(page->mapping != NULL))
1089 bad_reason = "non-NULL mapping";
1090 if (unlikely(page_ref_count(page) != 0))
1091 bad_reason = "nonzero _refcount";
1092 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
1093 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1094 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
1097 if (unlikely(page->mem_cgroup))
1098 bad_reason = "page still charged to cgroup";
1100 bad_page(page, bad_reason, bad_flags);
1103 static inline int free_pages_check(struct page *page)
1105 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1108 /* Something has gone sideways, find it */
1109 free_pages_check_bad(page);
1113 static int free_tail_pages_check(struct page *head_page, struct page *page)
1118 * We rely page->lru.next never has bit 0 set, unless the page
1119 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1121 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1123 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1127 switch (page - head_page) {
1129 /* the first tail page: ->mapping may be compound_mapcount() */
1130 if (unlikely(compound_mapcount(page))) {
1131 bad_page(page, "nonzero compound_mapcount", 0);
1137 * the second tail page: ->mapping is
1138 * deferred_list.next -- ignore value.
1142 if (page->mapping != TAIL_MAPPING) {
1143 bad_page(page, "corrupted mapping in tail page", 0);
1148 if (unlikely(!PageTail(page))) {
1149 bad_page(page, "PageTail not set", 0);
1152 if (unlikely(compound_head(page) != head_page)) {
1153 bad_page(page, "compound_head not consistent", 0);
1158 page->mapping = NULL;
1159 clear_compound_head(page);
1163 static void kernel_init_free_pages(struct page *page, int numpages)
1167 for (i = 0; i < numpages; i++)
1168 clear_highpage(page + i);
1171 static __always_inline bool free_pages_prepare(struct page *page,
1172 unsigned int order, bool check_free)
1176 VM_BUG_ON_PAGE(PageTail(page), page);
1178 trace_mm_page_free(page, order);
1181 * Check tail pages before head page information is cleared to
1182 * avoid checking PageCompound for order-0 pages.
1184 if (unlikely(order)) {
1185 bool compound = PageCompound(page);
1188 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1191 ClearPageDoubleMap(page);
1192 for (i = 1; i < (1 << order); i++) {
1194 bad += free_tail_pages_check(page, page + i);
1195 if (unlikely(free_pages_check(page + i))) {
1199 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1202 if (PageMappingFlags(page))
1203 page->mapping = NULL;
1204 if (memcg_kmem_enabled() && PageKmemcg(page))
1205 __memcg_kmem_uncharge_page(page, order);
1207 bad += free_pages_check(page);
1211 page_cpupid_reset_last(page);
1212 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1213 reset_page_owner(page, order);
1215 if (!PageHighMem(page)) {
1216 debug_check_no_locks_freed(page_address(page),
1217 PAGE_SIZE << order);
1218 debug_check_no_obj_freed(page_address(page),
1219 PAGE_SIZE << order);
1221 if (want_init_on_free())
1222 kernel_init_free_pages(page, 1 << order);
1224 kernel_poison_pages(page, 1 << order, 0);
1226 * arch_free_page() can make the page's contents inaccessible. s390
1227 * does this. So nothing which can access the page's contents should
1228 * happen after this.
1230 arch_free_page(page, order);
1232 if (debug_pagealloc_enabled_static())
1233 kernel_map_pages(page, 1 << order, 0);
1235 kasan_free_nondeferred_pages(page, order);
1240 #ifdef CONFIG_DEBUG_VM
1242 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1243 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1244 * moved from pcp lists to free lists.
1246 static bool free_pcp_prepare(struct page *page)
1248 return free_pages_prepare(page, 0, true);
1251 static bool bulkfree_pcp_prepare(struct page *page)
1253 if (debug_pagealloc_enabled_static())
1254 return free_pages_check(page);
1260 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1261 * moving from pcp lists to free list in order to reduce overhead. With
1262 * debug_pagealloc enabled, they are checked also immediately when being freed
1265 static bool free_pcp_prepare(struct page *page)
1267 if (debug_pagealloc_enabled_static())
1268 return free_pages_prepare(page, 0, true);
1270 return free_pages_prepare(page, 0, false);
1273 static bool bulkfree_pcp_prepare(struct page *page)
1275 return free_pages_check(page);
1277 #endif /* CONFIG_DEBUG_VM */
1279 static inline void prefetch_buddy(struct page *page)
1281 unsigned long pfn = page_to_pfn(page);
1282 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1283 struct page *buddy = page + (buddy_pfn - pfn);
1289 * Frees a number of pages from the PCP lists
1290 * Assumes all pages on list are in same zone, and of same order.
1291 * count is the number of pages to free.
1293 * If the zone was previously in an "all pages pinned" state then look to
1294 * see if this freeing clears that state.
1296 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1297 * pinned" detection logic.
1299 static void free_pcppages_bulk(struct zone *zone, int count,
1300 struct per_cpu_pages *pcp)
1302 int migratetype = 0;
1304 int prefetch_nr = 0;
1305 bool isolated_pageblocks;
1306 struct page *page, *tmp;
1310 struct list_head *list;
1313 * Remove pages from lists in a round-robin fashion. A
1314 * batch_free count is maintained that is incremented when an
1315 * empty list is encountered. This is so more pages are freed
1316 * off fuller lists instead of spinning excessively around empty
1321 if (++migratetype == MIGRATE_PCPTYPES)
1323 list = &pcp->lists[migratetype];
1324 } while (list_empty(list));
1326 /* This is the only non-empty list. Free them all. */
1327 if (batch_free == MIGRATE_PCPTYPES)
1331 page = list_last_entry(list, struct page, lru);
1332 /* must delete to avoid corrupting pcp list */
1333 list_del(&page->lru);
1336 if (bulkfree_pcp_prepare(page))
1339 list_add_tail(&page->lru, &head);
1342 * We are going to put the page back to the global
1343 * pool, prefetch its buddy to speed up later access
1344 * under zone->lock. It is believed the overhead of
1345 * an additional test and calculating buddy_pfn here
1346 * can be offset by reduced memory latency later. To
1347 * avoid excessive prefetching due to large count, only
1348 * prefetch buddy for the first pcp->batch nr of pages.
1350 if (prefetch_nr++ < pcp->batch)
1351 prefetch_buddy(page);
1352 } while (--count && --batch_free && !list_empty(list));
1355 spin_lock(&zone->lock);
1356 isolated_pageblocks = has_isolate_pageblock(zone);
1359 * Use safe version since after __free_one_page(),
1360 * page->lru.next will not point to original list.
1362 list_for_each_entry_safe(page, tmp, &head, lru) {
1363 int mt = get_pcppage_migratetype(page);
1364 /* MIGRATE_ISOLATE page should not go to pcplists */
1365 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1366 /* Pageblock could have been isolated meanwhile */
1367 if (unlikely(isolated_pageblocks))
1368 mt = get_pageblock_migratetype(page);
1370 __free_one_page(page, page_to_pfn(page), zone, 0, mt, true);
1371 trace_mm_page_pcpu_drain(page, 0, mt);
1373 spin_unlock(&zone->lock);
1376 static void free_one_page(struct zone *zone,
1377 struct page *page, unsigned long pfn,
1381 spin_lock(&zone->lock);
1382 if (unlikely(has_isolate_pageblock(zone) ||
1383 is_migrate_isolate(migratetype))) {
1384 migratetype = get_pfnblock_migratetype(page, pfn);
1386 __free_one_page(page, pfn, zone, order, migratetype, true);
1387 spin_unlock(&zone->lock);
1390 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1391 unsigned long zone, int nid)
1393 mm_zero_struct_page(page);
1394 set_page_links(page, zone, nid, pfn);
1395 init_page_count(page);
1396 page_mapcount_reset(page);
1397 page_cpupid_reset_last(page);
1398 page_kasan_tag_reset(page);
1400 INIT_LIST_HEAD(&page->lru);
1401 #ifdef WANT_PAGE_VIRTUAL
1402 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1403 if (!is_highmem_idx(zone))
1404 set_page_address(page, __va(pfn << PAGE_SHIFT));
1408 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1409 static void __meminit init_reserved_page(unsigned long pfn)
1414 if (!early_page_uninitialised(pfn))
1417 nid = early_pfn_to_nid(pfn);
1418 pgdat = NODE_DATA(nid);
1420 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1421 struct zone *zone = &pgdat->node_zones[zid];
1423 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1426 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1429 static inline void init_reserved_page(unsigned long pfn)
1432 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1435 * Initialised pages do not have PageReserved set. This function is
1436 * called for each range allocated by the bootmem allocator and
1437 * marks the pages PageReserved. The remaining valid pages are later
1438 * sent to the buddy page allocator.
1440 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1442 unsigned long start_pfn = PFN_DOWN(start);
1443 unsigned long end_pfn = PFN_UP(end);
1445 for (; start_pfn < end_pfn; start_pfn++) {
1446 if (pfn_valid(start_pfn)) {
1447 struct page *page = pfn_to_page(start_pfn);
1449 init_reserved_page(start_pfn);
1451 /* Avoid false-positive PageTail() */
1452 INIT_LIST_HEAD(&page->lru);
1455 * no need for atomic set_bit because the struct
1456 * page is not visible yet so nobody should
1459 __SetPageReserved(page);
1464 static void __free_pages_ok(struct page *page, unsigned int order)
1466 unsigned long flags;
1468 unsigned long pfn = page_to_pfn(page);
1470 if (!free_pages_prepare(page, order, true))
1473 migratetype = get_pfnblock_migratetype(page, pfn);
1474 local_irq_save(flags);
1475 __count_vm_events(PGFREE, 1 << order);
1476 free_one_page(page_zone(page), page, pfn, order, migratetype);
1477 local_irq_restore(flags);
1480 void __free_pages_core(struct page *page, unsigned int order)
1482 unsigned int nr_pages = 1 << order;
1483 struct page *p = page;
1487 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1489 __ClearPageReserved(p);
1490 set_page_count(p, 0);
1492 __ClearPageReserved(p);
1493 set_page_count(p, 0);
1495 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1496 set_page_refcounted(page);
1497 __free_pages(page, order);
1500 #ifdef CONFIG_NEED_MULTIPLE_NODES
1502 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1504 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
1507 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1509 int __meminit __early_pfn_to_nid(unsigned long pfn,
1510 struct mminit_pfnnid_cache *state)
1512 unsigned long start_pfn, end_pfn;
1515 if (state->last_start <= pfn && pfn < state->last_end)
1516 return state->last_nid;
1518 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1519 if (nid != NUMA_NO_NODE) {
1520 state->last_start = start_pfn;
1521 state->last_end = end_pfn;
1522 state->last_nid = nid;
1527 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
1529 int __meminit early_pfn_to_nid(unsigned long pfn)
1531 static DEFINE_SPINLOCK(early_pfn_lock);
1534 spin_lock(&early_pfn_lock);
1535 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1537 nid = first_online_node;
1538 spin_unlock(&early_pfn_lock);
1542 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1544 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1545 /* Only safe to use early in boot when initialisation is single-threaded */
1546 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1550 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1551 if (nid >= 0 && nid != node)
1557 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1564 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1567 if (early_page_uninitialised(pfn))
1569 __free_pages_core(page, order);
1573 * Check that the whole (or subset of) a pageblock given by the interval of
1574 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1575 * with the migration of free compaction scanner. The scanners then need to
1576 * use only pfn_valid_within() check for arches that allow holes within
1579 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1581 * It's possible on some configurations to have a setup like node0 node1 node0
1582 * i.e. it's possible that all pages within a zones range of pages do not
1583 * belong to a single zone. We assume that a border between node0 and node1
1584 * can occur within a single pageblock, but not a node0 node1 node0
1585 * interleaving within a single pageblock. It is therefore sufficient to check
1586 * the first and last page of a pageblock and avoid checking each individual
1587 * page in a pageblock.
1589 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1590 unsigned long end_pfn, struct zone *zone)
1592 struct page *start_page;
1593 struct page *end_page;
1595 /* end_pfn is one past the range we are checking */
1598 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1601 start_page = pfn_to_online_page(start_pfn);
1605 if (page_zone(start_page) != zone)
1608 end_page = pfn_to_page(end_pfn);
1610 /* This gives a shorter code than deriving page_zone(end_page) */
1611 if (page_zone_id(start_page) != page_zone_id(end_page))
1617 void set_zone_contiguous(struct zone *zone)
1619 unsigned long block_start_pfn = zone->zone_start_pfn;
1620 unsigned long block_end_pfn;
1622 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1623 for (; block_start_pfn < zone_end_pfn(zone);
1624 block_start_pfn = block_end_pfn,
1625 block_end_pfn += pageblock_nr_pages) {
1627 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1629 if (!__pageblock_pfn_to_page(block_start_pfn,
1630 block_end_pfn, zone))
1635 /* We confirm that there is no hole */
1636 zone->contiguous = true;
1639 void clear_zone_contiguous(struct zone *zone)
1641 zone->contiguous = false;
1644 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1645 static void __init deferred_free_range(unsigned long pfn,
1646 unsigned long nr_pages)
1654 page = pfn_to_page(pfn);
1656 /* Free a large naturally-aligned chunk if possible */
1657 if (nr_pages == pageblock_nr_pages &&
1658 (pfn & (pageblock_nr_pages - 1)) == 0) {
1659 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1660 __free_pages_core(page, pageblock_order);
1664 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1665 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1666 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1667 __free_pages_core(page, 0);
1671 /* Completion tracking for deferred_init_memmap() threads */
1672 static atomic_t pgdat_init_n_undone __initdata;
1673 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1675 static inline void __init pgdat_init_report_one_done(void)
1677 if (atomic_dec_and_test(&pgdat_init_n_undone))
1678 complete(&pgdat_init_all_done_comp);
1682 * Returns true if page needs to be initialized or freed to buddy allocator.
1684 * First we check if pfn is valid on architectures where it is possible to have
1685 * holes within pageblock_nr_pages. On systems where it is not possible, this
1686 * function is optimized out.
1688 * Then, we check if a current large page is valid by only checking the validity
1691 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1693 if (!pfn_valid_within(pfn))
1695 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1701 * Free pages to buddy allocator. Try to free aligned pages in
1702 * pageblock_nr_pages sizes.
1704 static void __init deferred_free_pages(unsigned long pfn,
1705 unsigned long end_pfn)
1707 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1708 unsigned long nr_free = 0;
1710 for (; pfn < end_pfn; pfn++) {
1711 if (!deferred_pfn_valid(pfn)) {
1712 deferred_free_range(pfn - nr_free, nr_free);
1714 } else if (!(pfn & nr_pgmask)) {
1715 deferred_free_range(pfn - nr_free, nr_free);
1717 touch_nmi_watchdog();
1722 /* Free the last block of pages to allocator */
1723 deferred_free_range(pfn - nr_free, nr_free);
1727 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1728 * by performing it only once every pageblock_nr_pages.
1729 * Return number of pages initialized.
1731 static unsigned long __init deferred_init_pages(struct zone *zone,
1733 unsigned long end_pfn)
1735 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1736 int nid = zone_to_nid(zone);
1737 unsigned long nr_pages = 0;
1738 int zid = zone_idx(zone);
1739 struct page *page = NULL;
1741 for (; pfn < end_pfn; pfn++) {
1742 if (!deferred_pfn_valid(pfn)) {
1745 } else if (!page || !(pfn & nr_pgmask)) {
1746 page = pfn_to_page(pfn);
1747 touch_nmi_watchdog();
1751 __init_single_page(page, pfn, zid, nid);
1758 * This function is meant to pre-load the iterator for the zone init.
1759 * Specifically it walks through the ranges until we are caught up to the
1760 * first_init_pfn value and exits there. If we never encounter the value we
1761 * return false indicating there are no valid ranges left.
1764 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1765 unsigned long *spfn, unsigned long *epfn,
1766 unsigned long first_init_pfn)
1771 * Start out by walking through the ranges in this zone that have
1772 * already been initialized. We don't need to do anything with them
1773 * so we just need to flush them out of the system.
1775 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1776 if (*epfn <= first_init_pfn)
1778 if (*spfn < first_init_pfn)
1779 *spfn = first_init_pfn;
1788 * Initialize and free pages. We do it in two loops: first we initialize
1789 * struct page, then free to buddy allocator, because while we are
1790 * freeing pages we can access pages that are ahead (computing buddy
1791 * page in __free_one_page()).
1793 * In order to try and keep some memory in the cache we have the loop
1794 * broken along max page order boundaries. This way we will not cause
1795 * any issues with the buddy page computation.
1797 static unsigned long __init
1798 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1799 unsigned long *end_pfn)
1801 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1802 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1803 unsigned long nr_pages = 0;
1806 /* First we loop through and initialize the page values */
1807 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1810 if (mo_pfn <= *start_pfn)
1813 t = min(mo_pfn, *end_pfn);
1814 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1816 if (mo_pfn < *end_pfn) {
1817 *start_pfn = mo_pfn;
1822 /* Reset values and now loop through freeing pages as needed */
1825 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1831 t = min(mo_pfn, epfn);
1832 deferred_free_pages(spfn, t);
1841 /* Initialise remaining memory on a node */
1842 static int __init deferred_init_memmap(void *data)
1844 pg_data_t *pgdat = data;
1845 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1846 unsigned long spfn = 0, epfn = 0, nr_pages = 0;
1847 unsigned long first_init_pfn, flags;
1848 unsigned long start = jiffies;
1853 /* Bind memory initialisation thread to a local node if possible */
1854 if (!cpumask_empty(cpumask))
1855 set_cpus_allowed_ptr(current, cpumask);
1857 pgdat_resize_lock(pgdat, &flags);
1858 first_init_pfn = pgdat->first_deferred_pfn;
1859 if (first_init_pfn == ULONG_MAX) {
1860 pgdat_resize_unlock(pgdat, &flags);
1861 pgdat_init_report_one_done();
1865 /* Sanity check boundaries */
1866 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1867 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1868 pgdat->first_deferred_pfn = ULONG_MAX;
1870 /* Only the highest zone is deferred so find it */
1871 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1872 zone = pgdat->node_zones + zid;
1873 if (first_init_pfn < zone_end_pfn(zone))
1877 /* If the zone is empty somebody else may have cleared out the zone */
1878 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1883 * Initialize and free pages in MAX_ORDER sized increments so
1884 * that we can avoid introducing any issues with the buddy
1888 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1890 pgdat_resize_unlock(pgdat, &flags);
1892 /* Sanity check that the next zone really is unpopulated */
1893 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1895 pr_info("node %d initialised, %lu pages in %ums\n",
1896 pgdat->node_id, nr_pages, jiffies_to_msecs(jiffies - start));
1898 pgdat_init_report_one_done();
1903 * If this zone has deferred pages, try to grow it by initializing enough
1904 * deferred pages to satisfy the allocation specified by order, rounded up to
1905 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1906 * of SECTION_SIZE bytes by initializing struct pages in increments of
1907 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1909 * Return true when zone was grown, otherwise return false. We return true even
1910 * when we grow less than requested, to let the caller decide if there are
1911 * enough pages to satisfy the allocation.
1913 * Note: We use noinline because this function is needed only during boot, and
1914 * it is called from a __ref function _deferred_grow_zone. This way we are
1915 * making sure that it is not inlined into permanent text section.
1917 static noinline bool __init
1918 deferred_grow_zone(struct zone *zone, unsigned int order)
1920 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1921 pg_data_t *pgdat = zone->zone_pgdat;
1922 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1923 unsigned long spfn, epfn, flags;
1924 unsigned long nr_pages = 0;
1927 /* Only the last zone may have deferred pages */
1928 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1931 pgdat_resize_lock(pgdat, &flags);
1934 * If deferred pages have been initialized while we were waiting for
1935 * the lock, return true, as the zone was grown. The caller will retry
1936 * this zone. We won't return to this function since the caller also
1937 * has this static branch.
1939 if (!static_branch_unlikely(&deferred_pages)) {
1940 pgdat_resize_unlock(pgdat, &flags);
1945 * If someone grew this zone while we were waiting for spinlock, return
1946 * true, as there might be enough pages already.
1948 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1949 pgdat_resize_unlock(pgdat, &flags);
1953 /* If the zone is empty somebody else may have cleared out the zone */
1954 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1955 first_deferred_pfn)) {
1956 pgdat->first_deferred_pfn = ULONG_MAX;
1957 pgdat_resize_unlock(pgdat, &flags);
1958 /* Retry only once. */
1959 return first_deferred_pfn != ULONG_MAX;
1963 * Initialize and free pages in MAX_ORDER sized increments so
1964 * that we can avoid introducing any issues with the buddy
1967 while (spfn < epfn) {
1968 /* update our first deferred PFN for this section */
1969 first_deferred_pfn = spfn;
1971 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1973 /* We should only stop along section boundaries */
1974 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
1977 /* If our quota has been met we can stop here */
1978 if (nr_pages >= nr_pages_needed)
1982 pgdat->first_deferred_pfn = spfn;
1983 pgdat_resize_unlock(pgdat, &flags);
1985 return nr_pages > 0;
1989 * deferred_grow_zone() is __init, but it is called from
1990 * get_page_from_freelist() during early boot until deferred_pages permanently
1991 * disables this call. This is why we have refdata wrapper to avoid warning,
1992 * and to ensure that the function body gets unloaded.
1995 _deferred_grow_zone(struct zone *zone, unsigned int order)
1997 return deferred_grow_zone(zone, order);
2000 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2002 void __init page_alloc_init_late(void)
2007 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2009 /* There will be num_node_state(N_MEMORY) threads */
2010 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2011 for_each_node_state(nid, N_MEMORY) {
2012 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2015 /* Block until all are initialised */
2016 wait_for_completion(&pgdat_init_all_done_comp);
2019 * The number of managed pages has changed due to the initialisation
2020 * so the pcpu batch and high limits needs to be updated or the limits
2021 * will be artificially small.
2023 for_each_populated_zone(zone)
2024 zone_pcp_update(zone);
2027 * We initialized the rest of the deferred pages. Permanently disable
2028 * on-demand struct page initialization.
2030 static_branch_disable(&deferred_pages);
2032 /* Reinit limits that are based on free pages after the kernel is up */
2033 files_maxfiles_init();
2036 /* Discard memblock private memory */
2039 for_each_node_state(nid, N_MEMORY)
2040 shuffle_free_memory(NODE_DATA(nid));
2042 for_each_populated_zone(zone)
2043 set_zone_contiguous(zone);
2047 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2048 void __init init_cma_reserved_pageblock(struct page *page)
2050 unsigned i = pageblock_nr_pages;
2051 struct page *p = page;
2054 __ClearPageReserved(p);
2055 set_page_count(p, 0);
2058 set_pageblock_migratetype(page, MIGRATE_CMA);
2060 if (pageblock_order >= MAX_ORDER) {
2061 i = pageblock_nr_pages;
2064 set_page_refcounted(p);
2065 __free_pages(p, MAX_ORDER - 1);
2066 p += MAX_ORDER_NR_PAGES;
2067 } while (i -= MAX_ORDER_NR_PAGES);
2069 set_page_refcounted(page);
2070 __free_pages(page, pageblock_order);
2073 adjust_managed_page_count(page, pageblock_nr_pages);
2078 * The order of subdivision here is critical for the IO subsystem.
2079 * Please do not alter this order without good reasons and regression
2080 * testing. Specifically, as large blocks of memory are subdivided,
2081 * the order in which smaller blocks are delivered depends on the order
2082 * they're subdivided in this function. This is the primary factor
2083 * influencing the order in which pages are delivered to the IO
2084 * subsystem according to empirical testing, and this is also justified
2085 * by considering the behavior of a buddy system containing a single
2086 * large block of memory acted on by a series of small allocations.
2087 * This behavior is a critical factor in sglist merging's success.
2091 static inline void expand(struct zone *zone, struct page *page,
2092 int low, int high, int migratetype)
2094 unsigned long size = 1 << high;
2096 while (high > low) {
2099 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2102 * Mark as guard pages (or page), that will allow to
2103 * merge back to allocator when buddy will be freed.
2104 * Corresponding page table entries will not be touched,
2105 * pages will stay not present in virtual address space
2107 if (set_page_guard(zone, &page[size], high, migratetype))
2110 add_to_free_list(&page[size], zone, high, migratetype);
2111 set_page_order(&page[size], high);
2115 static void check_new_page_bad(struct page *page)
2117 const char *bad_reason = NULL;
2118 unsigned long bad_flags = 0;
2120 if (unlikely(atomic_read(&page->_mapcount) != -1))
2121 bad_reason = "nonzero mapcount";
2122 if (unlikely(page->mapping != NULL))
2123 bad_reason = "non-NULL mapping";
2124 if (unlikely(page_ref_count(page) != 0))
2125 bad_reason = "nonzero _refcount";
2126 if (unlikely(page->flags & __PG_HWPOISON)) {
2127 bad_reason = "HWPoisoned (hardware-corrupted)";
2128 bad_flags = __PG_HWPOISON;
2129 /* Don't complain about hwpoisoned pages */
2130 page_mapcount_reset(page); /* remove PageBuddy */
2133 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
2134 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
2135 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
2138 if (unlikely(page->mem_cgroup))
2139 bad_reason = "page still charged to cgroup";
2141 bad_page(page, bad_reason, bad_flags);
2145 * This page is about to be returned from the page allocator
2147 static inline int check_new_page(struct page *page)
2149 if (likely(page_expected_state(page,
2150 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2153 check_new_page_bad(page);
2157 static inline bool free_pages_prezeroed(void)
2159 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2160 page_poisoning_enabled()) || want_init_on_free();
2163 #ifdef CONFIG_DEBUG_VM
2165 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2166 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2167 * also checked when pcp lists are refilled from the free lists.
2169 static inline bool check_pcp_refill(struct page *page)
2171 if (debug_pagealloc_enabled_static())
2172 return check_new_page(page);
2177 static inline bool check_new_pcp(struct page *page)
2179 return check_new_page(page);
2183 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2184 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2185 * enabled, they are also checked when being allocated from the pcp lists.
2187 static inline bool check_pcp_refill(struct page *page)
2189 return check_new_page(page);
2191 static inline bool check_new_pcp(struct page *page)
2193 if (debug_pagealloc_enabled_static())
2194 return check_new_page(page);
2198 #endif /* CONFIG_DEBUG_VM */
2200 static bool check_new_pages(struct page *page, unsigned int order)
2203 for (i = 0; i < (1 << order); i++) {
2204 struct page *p = page + i;
2206 if (unlikely(check_new_page(p)))
2213 inline void post_alloc_hook(struct page *page, unsigned int order,
2216 set_page_private(page, 0);
2217 set_page_refcounted(page);
2219 arch_alloc_page(page, order);
2220 if (debug_pagealloc_enabled_static())
2221 kernel_map_pages(page, 1 << order, 1);
2222 kasan_alloc_pages(page, order);
2223 kernel_poison_pages(page, 1 << order, 1);
2224 set_page_owner(page, order, gfp_flags);
2227 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2228 unsigned int alloc_flags)
2230 post_alloc_hook(page, order, gfp_flags);
2232 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2233 kernel_init_free_pages(page, 1 << order);
2235 if (order && (gfp_flags & __GFP_COMP))
2236 prep_compound_page(page, order);
2239 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2240 * allocate the page. The expectation is that the caller is taking
2241 * steps that will free more memory. The caller should avoid the page
2242 * being used for !PFMEMALLOC purposes.
2244 if (alloc_flags & ALLOC_NO_WATERMARKS)
2245 set_page_pfmemalloc(page);
2247 clear_page_pfmemalloc(page);
2251 * Go through the free lists for the given migratetype and remove
2252 * the smallest available page from the freelists
2254 static __always_inline
2255 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2258 unsigned int current_order;
2259 struct free_area *area;
2262 /* Find a page of the appropriate size in the preferred list */
2263 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2264 area = &(zone->free_area[current_order]);
2265 page = get_page_from_free_area(area, migratetype);
2268 del_page_from_free_list(page, zone, current_order);
2269 expand(zone, page, order, current_order, migratetype);
2270 set_pcppage_migratetype(page, migratetype);
2279 * This array describes the order lists are fallen back to when
2280 * the free lists for the desirable migrate type are depleted
2282 static int fallbacks[MIGRATE_TYPES][4] = {
2283 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2284 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2285 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2287 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2289 #ifdef CONFIG_MEMORY_ISOLATION
2290 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2295 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2298 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2301 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2302 unsigned int order) { return NULL; }
2306 * Move the free pages in a range to the free lists of the requested type.
2307 * Note that start_page and end_pages are not aligned on a pageblock
2308 * boundary. If alignment is required, use move_freepages_block()
2310 static int move_freepages(struct zone *zone,
2311 struct page *start_page, struct page *end_page,
2312 int migratetype, int *num_movable)
2316 int pages_moved = 0;
2318 for (page = start_page; page <= end_page;) {
2319 if (!pfn_valid_within(page_to_pfn(page))) {
2324 if (!PageBuddy(page)) {
2326 * We assume that pages that could be isolated for
2327 * migration are movable. But we don't actually try
2328 * isolating, as that would be expensive.
2331 (PageLRU(page) || __PageMovable(page)))
2338 /* Make sure we are not inadvertently changing nodes */
2339 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2340 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2342 order = page_order(page);
2343 move_to_free_list(page, zone, order, migratetype);
2345 pages_moved += 1 << order;
2351 int move_freepages_block(struct zone *zone, struct page *page,
2352 int migratetype, int *num_movable)
2354 unsigned long start_pfn, end_pfn;
2355 struct page *start_page, *end_page;
2360 start_pfn = page_to_pfn(page);
2361 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2362 start_page = pfn_to_page(start_pfn);
2363 end_page = start_page + pageblock_nr_pages - 1;
2364 end_pfn = start_pfn + pageblock_nr_pages - 1;
2366 /* Do not cross zone boundaries */
2367 if (!zone_spans_pfn(zone, start_pfn))
2369 if (!zone_spans_pfn(zone, end_pfn))
2372 return move_freepages(zone, start_page, end_page, migratetype,
2376 static void change_pageblock_range(struct page *pageblock_page,
2377 int start_order, int migratetype)
2379 int nr_pageblocks = 1 << (start_order - pageblock_order);
2381 while (nr_pageblocks--) {
2382 set_pageblock_migratetype(pageblock_page, migratetype);
2383 pageblock_page += pageblock_nr_pages;
2388 * When we are falling back to another migratetype during allocation, try to
2389 * steal extra free pages from the same pageblocks to satisfy further
2390 * allocations, instead of polluting multiple pageblocks.
2392 * If we are stealing a relatively large buddy page, it is likely there will
2393 * be more free pages in the pageblock, so try to steal them all. For
2394 * reclaimable and unmovable allocations, we steal regardless of page size,
2395 * as fragmentation caused by those allocations polluting movable pageblocks
2396 * is worse than movable allocations stealing from unmovable and reclaimable
2399 static bool can_steal_fallback(unsigned int order, int start_mt)
2402 * Leaving this order check is intended, although there is
2403 * relaxed order check in next check. The reason is that
2404 * we can actually steal whole pageblock if this condition met,
2405 * but, below check doesn't guarantee it and that is just heuristic
2406 * so could be changed anytime.
2408 if (order >= pageblock_order)
2411 if (order >= pageblock_order / 2 ||
2412 start_mt == MIGRATE_RECLAIMABLE ||
2413 start_mt == MIGRATE_UNMOVABLE ||
2414 page_group_by_mobility_disabled)
2420 static inline void boost_watermark(struct zone *zone)
2422 unsigned long max_boost;
2424 if (!watermark_boost_factor)
2427 * Don't bother in zones that are unlikely to produce results.
2428 * On small machines, including kdump capture kernels running
2429 * in a small area, boosting the watermark can cause an out of
2430 * memory situation immediately.
2432 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2435 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2436 watermark_boost_factor, 10000);
2439 * high watermark may be uninitialised if fragmentation occurs
2440 * very early in boot so do not boost. We do not fall
2441 * through and boost by pageblock_nr_pages as failing
2442 * allocations that early means that reclaim is not going
2443 * to help and it may even be impossible to reclaim the
2444 * boosted watermark resulting in a hang.
2449 max_boost = max(pageblock_nr_pages, max_boost);
2451 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2456 * This function implements actual steal behaviour. If order is large enough,
2457 * we can steal whole pageblock. If not, we first move freepages in this
2458 * pageblock to our migratetype and determine how many already-allocated pages
2459 * are there in the pageblock with a compatible migratetype. If at least half
2460 * of pages are free or compatible, we can change migratetype of the pageblock
2461 * itself, so pages freed in the future will be put on the correct free list.
2463 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2464 unsigned int alloc_flags, int start_type, bool whole_block)
2466 unsigned int current_order = page_order(page);
2467 int free_pages, movable_pages, alike_pages;
2470 old_block_type = get_pageblock_migratetype(page);
2473 * This can happen due to races and we want to prevent broken
2474 * highatomic accounting.
2476 if (is_migrate_highatomic(old_block_type))
2479 /* Take ownership for orders >= pageblock_order */
2480 if (current_order >= pageblock_order) {
2481 change_pageblock_range(page, current_order, start_type);
2486 * Boost watermarks to increase reclaim pressure to reduce the
2487 * likelihood of future fallbacks. Wake kswapd now as the node
2488 * may be balanced overall and kswapd will not wake naturally.
2490 boost_watermark(zone);
2491 if (alloc_flags & ALLOC_KSWAPD)
2492 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2494 /* We are not allowed to try stealing from the whole block */
2498 free_pages = move_freepages_block(zone, page, start_type,
2501 * Determine how many pages are compatible with our allocation.
2502 * For movable allocation, it's the number of movable pages which
2503 * we just obtained. For other types it's a bit more tricky.
2505 if (start_type == MIGRATE_MOVABLE) {
2506 alike_pages = movable_pages;
2509 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2510 * to MOVABLE pageblock, consider all non-movable pages as
2511 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2512 * vice versa, be conservative since we can't distinguish the
2513 * exact migratetype of non-movable pages.
2515 if (old_block_type == MIGRATE_MOVABLE)
2516 alike_pages = pageblock_nr_pages
2517 - (free_pages + movable_pages);
2522 /* moving whole block can fail due to zone boundary conditions */
2527 * If a sufficient number of pages in the block are either free or of
2528 * comparable migratability as our allocation, claim the whole block.
2530 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2531 page_group_by_mobility_disabled)
2532 set_pageblock_migratetype(page, start_type);
2537 move_to_free_list(page, zone, current_order, start_type);
2541 * Check whether there is a suitable fallback freepage with requested order.
2542 * If only_stealable is true, this function returns fallback_mt only if
2543 * we can steal other freepages all together. This would help to reduce
2544 * fragmentation due to mixed migratetype pages in one pageblock.
2546 int find_suitable_fallback(struct free_area *area, unsigned int order,
2547 int migratetype, bool only_stealable, bool *can_steal)
2552 if (area->nr_free == 0)
2557 fallback_mt = fallbacks[migratetype][i];
2558 if (fallback_mt == MIGRATE_TYPES)
2561 if (free_area_empty(area, fallback_mt))
2564 if (can_steal_fallback(order, migratetype))
2567 if (!only_stealable)
2578 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2579 * there are no empty page blocks that contain a page with a suitable order
2581 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2582 unsigned int alloc_order)
2585 unsigned long max_managed, flags;
2588 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2589 * Check is race-prone but harmless.
2591 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2592 if (zone->nr_reserved_highatomic >= max_managed)
2595 spin_lock_irqsave(&zone->lock, flags);
2597 /* Recheck the nr_reserved_highatomic limit under the lock */
2598 if (zone->nr_reserved_highatomic >= max_managed)
2602 mt = get_pageblock_migratetype(page);
2603 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2604 && !is_migrate_cma(mt)) {
2605 zone->nr_reserved_highatomic += pageblock_nr_pages;
2606 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2607 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2611 spin_unlock_irqrestore(&zone->lock, flags);
2615 * Used when an allocation is about to fail under memory pressure. This
2616 * potentially hurts the reliability of high-order allocations when under
2617 * intense memory pressure but failed atomic allocations should be easier
2618 * to recover from than an OOM.
2620 * If @force is true, try to unreserve a pageblock even though highatomic
2621 * pageblock is exhausted.
2623 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2626 struct zonelist *zonelist = ac->zonelist;
2627 unsigned long flags;
2634 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2637 * Preserve at least one pageblock unless memory pressure
2640 if (!force && zone->nr_reserved_highatomic <=
2644 spin_lock_irqsave(&zone->lock, flags);
2645 for (order = 0; order < MAX_ORDER; order++) {
2646 struct free_area *area = &(zone->free_area[order]);
2648 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2653 * In page freeing path, migratetype change is racy so
2654 * we can counter several free pages in a pageblock
2655 * in this loop althoug we changed the pageblock type
2656 * from highatomic to ac->migratetype. So we should
2657 * adjust the count once.
2659 if (is_migrate_highatomic_page(page)) {
2661 * It should never happen but changes to
2662 * locking could inadvertently allow a per-cpu
2663 * drain to add pages to MIGRATE_HIGHATOMIC
2664 * while unreserving so be safe and watch for
2667 zone->nr_reserved_highatomic -= min(
2669 zone->nr_reserved_highatomic);
2673 * Convert to ac->migratetype and avoid the normal
2674 * pageblock stealing heuristics. Minimally, the caller
2675 * is doing the work and needs the pages. More
2676 * importantly, if the block was always converted to
2677 * MIGRATE_UNMOVABLE or another type then the number
2678 * of pageblocks that cannot be completely freed
2681 set_pageblock_migratetype(page, ac->migratetype);
2682 ret = move_freepages_block(zone, page, ac->migratetype,
2685 spin_unlock_irqrestore(&zone->lock, flags);
2689 spin_unlock_irqrestore(&zone->lock, flags);
2696 * Try finding a free buddy page on the fallback list and put it on the free
2697 * list of requested migratetype, possibly along with other pages from the same
2698 * block, depending on fragmentation avoidance heuristics. Returns true if
2699 * fallback was found so that __rmqueue_smallest() can grab it.
2701 * The use of signed ints for order and current_order is a deliberate
2702 * deviation from the rest of this file, to make the for loop
2703 * condition simpler.
2705 static __always_inline bool
2706 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2707 unsigned int alloc_flags)
2709 struct free_area *area;
2711 int min_order = order;
2717 * Do not steal pages from freelists belonging to other pageblocks
2718 * i.e. orders < pageblock_order. If there are no local zones free,
2719 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2721 if (alloc_flags & ALLOC_NOFRAGMENT)
2722 min_order = pageblock_order;
2725 * Find the largest available free page in the other list. This roughly
2726 * approximates finding the pageblock with the most free pages, which
2727 * would be too costly to do exactly.
2729 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2731 area = &(zone->free_area[current_order]);
2732 fallback_mt = find_suitable_fallback(area, current_order,
2733 start_migratetype, false, &can_steal);
2734 if (fallback_mt == -1)
2738 * We cannot steal all free pages from the pageblock and the
2739 * requested migratetype is movable. In that case it's better to
2740 * steal and split the smallest available page instead of the
2741 * largest available page, because even if the next movable
2742 * allocation falls back into a different pageblock than this
2743 * one, it won't cause permanent fragmentation.
2745 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2746 && current_order > order)
2755 for (current_order = order; current_order < MAX_ORDER;
2757 area = &(zone->free_area[current_order]);
2758 fallback_mt = find_suitable_fallback(area, current_order,
2759 start_migratetype, false, &can_steal);
2760 if (fallback_mt != -1)
2765 * This should not happen - we already found a suitable fallback
2766 * when looking for the largest page.
2768 VM_BUG_ON(current_order == MAX_ORDER);
2771 page = get_page_from_free_area(area, fallback_mt);
2773 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2776 trace_mm_page_alloc_extfrag(page, order, current_order,
2777 start_migratetype, fallback_mt);
2784 * Do the hard work of removing an element from the buddy allocator.
2785 * Call me with the zone->lock already held.
2787 static __always_inline struct page *
2788 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2789 unsigned int alloc_flags)
2794 page = __rmqueue_smallest(zone, order, migratetype);
2795 if (unlikely(!page)) {
2796 if (migratetype == MIGRATE_MOVABLE)
2797 page = __rmqueue_cma_fallback(zone, order);
2799 if (!page && __rmqueue_fallback(zone, order, migratetype,
2804 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2809 * Obtain a specified number of elements from the buddy allocator, all under
2810 * a single hold of the lock, for efficiency. Add them to the supplied list.
2811 * Returns the number of new pages which were placed at *list.
2813 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2814 unsigned long count, struct list_head *list,
2815 int migratetype, unsigned int alloc_flags)
2819 spin_lock(&zone->lock);
2820 for (i = 0; i < count; ++i) {
2821 struct page *page = __rmqueue(zone, order, migratetype,
2823 if (unlikely(page == NULL))
2826 if (unlikely(check_pcp_refill(page)))
2830 * Split buddy pages returned by expand() are received here in
2831 * physical page order. The page is added to the tail of
2832 * caller's list. From the callers perspective, the linked list
2833 * is ordered by page number under some conditions. This is
2834 * useful for IO devices that can forward direction from the
2835 * head, thus also in the physical page order. This is useful
2836 * for IO devices that can merge IO requests if the physical
2837 * pages are ordered properly.
2839 list_add_tail(&page->lru, list);
2841 if (is_migrate_cma(get_pcppage_migratetype(page)))
2842 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2847 * i pages were removed from the buddy list even if some leak due
2848 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2849 * on i. Do not confuse with 'alloced' which is the number of
2850 * pages added to the pcp list.
2852 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2853 spin_unlock(&zone->lock);
2859 * Called from the vmstat counter updater to drain pagesets of this
2860 * currently executing processor on remote nodes after they have
2863 * Note that this function must be called with the thread pinned to
2864 * a single processor.
2866 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2868 unsigned long flags;
2869 int to_drain, batch;
2871 local_irq_save(flags);
2872 batch = READ_ONCE(pcp->batch);
2873 to_drain = min(pcp->count, batch);
2875 free_pcppages_bulk(zone, to_drain, pcp);
2876 local_irq_restore(flags);
2881 * Drain pcplists of the indicated processor and zone.
2883 * The processor must either be the current processor and the
2884 * thread pinned to the current processor or a processor that
2887 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2889 unsigned long flags;
2890 struct per_cpu_pageset *pset;
2891 struct per_cpu_pages *pcp;
2893 local_irq_save(flags);
2894 pset = per_cpu_ptr(zone->pageset, cpu);
2898 free_pcppages_bulk(zone, pcp->count, pcp);
2899 local_irq_restore(flags);
2903 * Drain pcplists of all zones on the indicated processor.
2905 * The processor must either be the current processor and the
2906 * thread pinned to the current processor or a processor that
2909 static void drain_pages(unsigned int cpu)
2913 for_each_populated_zone(zone) {
2914 drain_pages_zone(cpu, zone);
2919 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2921 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2922 * the single zone's pages.
2924 void drain_local_pages(struct zone *zone)
2926 int cpu = smp_processor_id();
2929 drain_pages_zone(cpu, zone);
2934 static void drain_local_pages_wq(struct work_struct *work)
2936 struct pcpu_drain *drain;
2938 drain = container_of(work, struct pcpu_drain, work);
2941 * drain_all_pages doesn't use proper cpu hotplug protection so
2942 * we can race with cpu offline when the WQ can move this from
2943 * a cpu pinned worker to an unbound one. We can operate on a different
2944 * cpu which is allright but we also have to make sure to not move to
2948 drain_local_pages(drain->zone);
2953 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2955 * When zone parameter is non-NULL, spill just the single zone's pages.
2957 * Note that this can be extremely slow as the draining happens in a workqueue.
2959 void drain_all_pages(struct zone *zone)
2964 * Allocate in the BSS so we wont require allocation in
2965 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2967 static cpumask_t cpus_with_pcps;
2970 * Make sure nobody triggers this path before mm_percpu_wq is fully
2973 if (WARN_ON_ONCE(!mm_percpu_wq))
2977 * Do not drain if one is already in progress unless it's specific to
2978 * a zone. Such callers are primarily CMA and memory hotplug and need
2979 * the drain to be complete when the call returns.
2981 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2984 mutex_lock(&pcpu_drain_mutex);
2988 * We don't care about racing with CPU hotplug event
2989 * as offline notification will cause the notified
2990 * cpu to drain that CPU pcps and on_each_cpu_mask
2991 * disables preemption as part of its processing
2993 for_each_online_cpu(cpu) {
2994 struct per_cpu_pageset *pcp;
2996 bool has_pcps = false;
2999 pcp = per_cpu_ptr(zone->pageset, cpu);
3003 for_each_populated_zone(z) {
3004 pcp = per_cpu_ptr(z->pageset, cpu);
3005 if (pcp->pcp.count) {
3013 cpumask_set_cpu(cpu, &cpus_with_pcps);
3015 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3018 for_each_cpu(cpu, &cpus_with_pcps) {
3019 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3022 INIT_WORK(&drain->work, drain_local_pages_wq);
3023 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3025 for_each_cpu(cpu, &cpus_with_pcps)
3026 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3028 mutex_unlock(&pcpu_drain_mutex);
3031 #ifdef CONFIG_HIBERNATION
3034 * Touch the watchdog for every WD_PAGE_COUNT pages.
3036 #define WD_PAGE_COUNT (128*1024)
3038 void mark_free_pages(struct zone *zone)
3040 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3041 unsigned long flags;
3042 unsigned int order, t;
3045 if (zone_is_empty(zone))
3048 spin_lock_irqsave(&zone->lock, flags);
3050 max_zone_pfn = zone_end_pfn(zone);
3051 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3052 if (pfn_valid(pfn)) {
3053 page = pfn_to_page(pfn);
3055 if (!--page_count) {
3056 touch_nmi_watchdog();
3057 page_count = WD_PAGE_COUNT;
3060 if (page_zone(page) != zone)
3063 if (!swsusp_page_is_forbidden(page))
3064 swsusp_unset_page_free(page);
3067 for_each_migratetype_order(order, t) {
3068 list_for_each_entry(page,
3069 &zone->free_area[order].free_list[t], lru) {
3072 pfn = page_to_pfn(page);
3073 for (i = 0; i < (1UL << order); i++) {
3074 if (!--page_count) {
3075 touch_nmi_watchdog();
3076 page_count = WD_PAGE_COUNT;
3078 swsusp_set_page_free(pfn_to_page(pfn + i));
3082 spin_unlock_irqrestore(&zone->lock, flags);
3084 #endif /* CONFIG_PM */
3086 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3090 if (!free_pcp_prepare(page))
3093 migratetype = get_pfnblock_migratetype(page, pfn);
3094 set_pcppage_migratetype(page, migratetype);
3098 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3100 struct zone *zone = page_zone(page);
3101 struct per_cpu_pages *pcp;
3104 migratetype = get_pcppage_migratetype(page);
3105 __count_vm_event(PGFREE);
3108 * We only track unmovable, reclaimable and movable on pcp lists.
3109 * Free ISOLATE pages back to the allocator because they are being
3110 * offlined but treat HIGHATOMIC as movable pages so we can get those
3111 * areas back if necessary. Otherwise, we may have to free
3112 * excessively into the page allocator
3114 if (migratetype >= MIGRATE_PCPTYPES) {
3115 if (unlikely(is_migrate_isolate(migratetype))) {
3116 free_one_page(zone, page, pfn, 0, migratetype);
3119 migratetype = MIGRATE_MOVABLE;
3122 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3123 list_add(&page->lru, &pcp->lists[migratetype]);
3125 if (pcp->count >= pcp->high) {
3126 unsigned long batch = READ_ONCE(pcp->batch);
3127 free_pcppages_bulk(zone, batch, pcp);
3132 * Free a 0-order page
3134 void free_unref_page(struct page *page)
3136 unsigned long flags;
3137 unsigned long pfn = page_to_pfn(page);
3139 if (!free_unref_page_prepare(page, pfn))
3142 local_irq_save(flags);
3143 free_unref_page_commit(page, pfn);
3144 local_irq_restore(flags);
3148 * Free a list of 0-order pages
3150 void free_unref_page_list(struct list_head *list)
3152 struct page *page, *next;
3153 unsigned long flags, pfn;
3154 int batch_count = 0;
3156 /* Prepare pages for freeing */
3157 list_for_each_entry_safe(page, next, list, lru) {
3158 pfn = page_to_pfn(page);
3159 if (!free_unref_page_prepare(page, pfn))
3160 list_del(&page->lru);
3161 set_page_private(page, pfn);
3164 local_irq_save(flags);
3165 list_for_each_entry_safe(page, next, list, lru) {
3166 unsigned long pfn = page_private(page);
3168 set_page_private(page, 0);
3169 trace_mm_page_free_batched(page);
3170 free_unref_page_commit(page, pfn);
3173 * Guard against excessive IRQ disabled times when we get
3174 * a large list of pages to free.
3176 if (++batch_count == SWAP_CLUSTER_MAX) {
3177 local_irq_restore(flags);
3179 local_irq_save(flags);
3182 local_irq_restore(flags);
3186 * split_page takes a non-compound higher-order page, and splits it into
3187 * n (1<<order) sub-pages: page[0..n]
3188 * Each sub-page must be freed individually.
3190 * Note: this is probably too low level an operation for use in drivers.
3191 * Please consult with lkml before using this in your driver.
3193 void split_page(struct page *page, unsigned int order)
3197 VM_BUG_ON_PAGE(PageCompound(page), page);
3198 VM_BUG_ON_PAGE(!page_count(page), page);
3200 for (i = 1; i < (1 << order); i++)
3201 set_page_refcounted(page + i);
3202 split_page_owner(page, order);
3204 EXPORT_SYMBOL_GPL(split_page);
3206 int __isolate_free_page(struct page *page, unsigned int order)
3208 unsigned long watermark;
3212 BUG_ON(!PageBuddy(page));
3214 zone = page_zone(page);
3215 mt = get_pageblock_migratetype(page);
3217 if (!is_migrate_isolate(mt)) {
3219 * Obey watermarks as if the page was being allocated. We can
3220 * emulate a high-order watermark check with a raised order-0
3221 * watermark, because we already know our high-order page
3224 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3225 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3228 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3231 /* Remove page from free list */
3233 del_page_from_free_list(page, zone, order);
3236 * Set the pageblock if the isolated page is at least half of a
3239 if (order >= pageblock_order - 1) {
3240 struct page *endpage = page + (1 << order) - 1;
3241 for (; page < endpage; page += pageblock_nr_pages) {
3242 int mt = get_pageblock_migratetype(page);
3243 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3244 && !is_migrate_highatomic(mt))
3245 set_pageblock_migratetype(page,
3251 return 1UL << order;
3255 * __putback_isolated_page - Return a now-isolated page back where we got it
3256 * @page: Page that was isolated
3257 * @order: Order of the isolated page
3258 * @mt: The page's pageblock's migratetype
3260 * This function is meant to return a page pulled from the free lists via
3261 * __isolate_free_page back to the free lists they were pulled from.
3263 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3265 struct zone *zone = page_zone(page);
3267 /* zone lock should be held when this function is called */
3268 lockdep_assert_held(&zone->lock);
3270 /* Return isolated page to tail of freelist. */
3271 __free_one_page(page, page_to_pfn(page), zone, order, mt, false);
3275 * Update NUMA hit/miss statistics
3277 * Must be called with interrupts disabled.
3279 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3282 enum numa_stat_item local_stat = NUMA_LOCAL;
3284 /* skip numa counters update if numa stats is disabled */
3285 if (!static_branch_likely(&vm_numa_stat_key))
3288 if (zone_to_nid(z) != numa_node_id())
3289 local_stat = NUMA_OTHER;
3291 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3292 __inc_numa_state(z, NUMA_HIT);
3294 __inc_numa_state(z, NUMA_MISS);
3295 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3297 __inc_numa_state(z, local_stat);
3301 /* Remove page from the per-cpu list, caller must protect the list */
3302 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3303 unsigned int alloc_flags,
3304 struct per_cpu_pages *pcp,
3305 struct list_head *list)
3310 if (list_empty(list)) {
3311 pcp->count += rmqueue_bulk(zone, 0,
3313 migratetype, alloc_flags);
3314 if (unlikely(list_empty(list)))
3318 page = list_first_entry(list, struct page, lru);
3319 list_del(&page->lru);
3321 } while (check_new_pcp(page));
3326 /* Lock and remove page from the per-cpu list */
3327 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3328 struct zone *zone, gfp_t gfp_flags,
3329 int migratetype, unsigned int alloc_flags)
3331 struct per_cpu_pages *pcp;
3332 struct list_head *list;
3334 unsigned long flags;
3336 local_irq_save(flags);
3337 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3338 list = &pcp->lists[migratetype];
3339 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3341 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3342 zone_statistics(preferred_zone, zone);
3344 local_irq_restore(flags);
3349 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3352 struct page *rmqueue(struct zone *preferred_zone,
3353 struct zone *zone, unsigned int order,
3354 gfp_t gfp_flags, unsigned int alloc_flags,
3357 unsigned long flags;
3360 if (likely(order == 0)) {
3361 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3362 migratetype, alloc_flags);
3367 * We most definitely don't want callers attempting to
3368 * allocate greater than order-1 page units with __GFP_NOFAIL.
3370 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3371 spin_lock_irqsave(&zone->lock, flags);
3375 if (alloc_flags & ALLOC_HARDER) {
3376 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3378 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3381 page = __rmqueue(zone, order, migratetype, alloc_flags);
3382 } while (page && check_new_pages(page, order));
3383 spin_unlock(&zone->lock);
3386 __mod_zone_freepage_state(zone, -(1 << order),
3387 get_pcppage_migratetype(page));
3389 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3390 zone_statistics(preferred_zone, zone);
3391 local_irq_restore(flags);
3394 /* Separate test+clear to avoid unnecessary atomics */
3395 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3396 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3397 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3400 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3404 local_irq_restore(flags);
3408 #ifdef CONFIG_FAIL_PAGE_ALLOC
3411 struct fault_attr attr;
3413 bool ignore_gfp_highmem;
3414 bool ignore_gfp_reclaim;
3416 } fail_page_alloc = {
3417 .attr = FAULT_ATTR_INITIALIZER,
3418 .ignore_gfp_reclaim = true,
3419 .ignore_gfp_highmem = true,
3423 static int __init setup_fail_page_alloc(char *str)
3425 return setup_fault_attr(&fail_page_alloc.attr, str);
3427 __setup("fail_page_alloc=", setup_fail_page_alloc);
3429 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3431 if (order < fail_page_alloc.min_order)
3433 if (gfp_mask & __GFP_NOFAIL)
3435 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3437 if (fail_page_alloc.ignore_gfp_reclaim &&
3438 (gfp_mask & __GFP_DIRECT_RECLAIM))
3441 return should_fail(&fail_page_alloc.attr, 1 << order);
3444 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3446 static int __init fail_page_alloc_debugfs(void)
3448 umode_t mode = S_IFREG | 0600;
3451 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3452 &fail_page_alloc.attr);
3454 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3455 &fail_page_alloc.ignore_gfp_reclaim);
3456 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3457 &fail_page_alloc.ignore_gfp_highmem);
3458 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3463 late_initcall(fail_page_alloc_debugfs);
3465 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3467 #else /* CONFIG_FAIL_PAGE_ALLOC */
3469 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3474 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3476 static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3478 return __should_fail_alloc_page(gfp_mask, order);
3480 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3483 * Return true if free base pages are above 'mark'. For high-order checks it
3484 * will return true of the order-0 watermark is reached and there is at least
3485 * one free page of a suitable size. Checking now avoids taking the zone lock
3486 * to check in the allocation paths if no pages are free.
3488 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3489 int classzone_idx, unsigned int alloc_flags,
3494 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3496 /* free_pages may go negative - that's OK */
3497 free_pages -= (1 << order) - 1;
3499 if (alloc_flags & ALLOC_HIGH)
3503 * If the caller does not have rights to ALLOC_HARDER then subtract
3504 * the high-atomic reserves. This will over-estimate the size of the
3505 * atomic reserve but it avoids a search.
3507 if (likely(!alloc_harder)) {
3508 free_pages -= z->nr_reserved_highatomic;
3511 * OOM victims can try even harder than normal ALLOC_HARDER
3512 * users on the grounds that it's definitely going to be in
3513 * the exit path shortly and free memory. Any allocation it
3514 * makes during the free path will be small and short-lived.
3516 if (alloc_flags & ALLOC_OOM)
3524 /* If allocation can't use CMA areas don't use free CMA pages */
3525 if (!(alloc_flags & ALLOC_CMA))
3526 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3530 * Check watermarks for an order-0 allocation request. If these
3531 * are not met, then a high-order request also cannot go ahead
3532 * even if a suitable page happened to be free.
3534 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3537 /* If this is an order-0 request then the watermark is fine */
3541 /* For a high-order request, check at least one suitable page is free */
3542 for (o = order; o < MAX_ORDER; o++) {
3543 struct free_area *area = &z->free_area[o];
3549 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3550 if (!free_area_empty(area, mt))
3555 if ((alloc_flags & ALLOC_CMA) &&
3556 !free_area_empty(area, MIGRATE_CMA)) {
3560 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3566 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3567 int classzone_idx, unsigned int alloc_flags)
3569 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3570 zone_page_state(z, NR_FREE_PAGES));
3573 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3574 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3576 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3580 /* If allocation can't use CMA areas don't use free CMA pages */
3581 if (!(alloc_flags & ALLOC_CMA))
3582 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3586 * Fast check for order-0 only. If this fails then the reserves
3587 * need to be calculated. There is a corner case where the check
3588 * passes but only the high-order atomic reserve are free. If
3589 * the caller is !atomic then it'll uselessly search the free
3590 * list. That corner case is then slower but it is harmless.
3592 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3595 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3599 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3600 unsigned long mark, int classzone_idx)
3602 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3604 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3605 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3607 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3612 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3614 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3615 node_reclaim_distance;
3617 #else /* CONFIG_NUMA */
3618 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3622 #endif /* CONFIG_NUMA */
3625 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3626 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3627 * premature use of a lower zone may cause lowmem pressure problems that
3628 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3629 * probably too small. It only makes sense to spread allocations to avoid
3630 * fragmentation between the Normal and DMA32 zones.
3632 static inline unsigned int
3633 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3635 unsigned int alloc_flags;
3638 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3641 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3643 #ifdef CONFIG_ZONE_DMA32
3647 if (zone_idx(zone) != ZONE_NORMAL)
3651 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3652 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3653 * on UMA that if Normal is populated then so is DMA32.
3655 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3656 if (nr_online_nodes > 1 && !populated_zone(--zone))
3659 alloc_flags |= ALLOC_NOFRAGMENT;
3660 #endif /* CONFIG_ZONE_DMA32 */
3665 * get_page_from_freelist goes through the zonelist trying to allocate
3668 static struct page *
3669 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3670 const struct alloc_context *ac)
3674 struct pglist_data *last_pgdat_dirty_limit = NULL;
3679 * Scan zonelist, looking for a zone with enough free.
3680 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3682 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3683 z = ac->preferred_zoneref;
3684 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3689 if (cpusets_enabled() &&
3690 (alloc_flags & ALLOC_CPUSET) &&
3691 !__cpuset_zone_allowed(zone, gfp_mask))
3694 * When allocating a page cache page for writing, we
3695 * want to get it from a node that is within its dirty
3696 * limit, such that no single node holds more than its
3697 * proportional share of globally allowed dirty pages.
3698 * The dirty limits take into account the node's
3699 * lowmem reserves and high watermark so that kswapd
3700 * should be able to balance it without having to
3701 * write pages from its LRU list.
3703 * XXX: For now, allow allocations to potentially
3704 * exceed the per-node dirty limit in the slowpath
3705 * (spread_dirty_pages unset) before going into reclaim,
3706 * which is important when on a NUMA setup the allowed
3707 * nodes are together not big enough to reach the
3708 * global limit. The proper fix for these situations
3709 * will require awareness of nodes in the
3710 * dirty-throttling and the flusher threads.
3712 if (ac->spread_dirty_pages) {
3713 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3716 if (!node_dirty_ok(zone->zone_pgdat)) {
3717 last_pgdat_dirty_limit = zone->zone_pgdat;
3722 if (no_fallback && nr_online_nodes > 1 &&
3723 zone != ac->preferred_zoneref->zone) {
3727 * If moving to a remote node, retry but allow
3728 * fragmenting fallbacks. Locality is more important
3729 * than fragmentation avoidance.
3731 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3732 if (zone_to_nid(zone) != local_nid) {
3733 alloc_flags &= ~ALLOC_NOFRAGMENT;
3738 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3739 if (!zone_watermark_fast(zone, order, mark,
3740 ac_classzone_idx(ac), alloc_flags)) {
3743 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3745 * Watermark failed for this zone, but see if we can
3746 * grow this zone if it contains deferred pages.
3748 if (static_branch_unlikely(&deferred_pages)) {
3749 if (_deferred_grow_zone(zone, order))
3753 /* Checked here to keep the fast path fast */
3754 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3755 if (alloc_flags & ALLOC_NO_WATERMARKS)
3758 if (node_reclaim_mode == 0 ||
3759 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3762 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3764 case NODE_RECLAIM_NOSCAN:
3767 case NODE_RECLAIM_FULL:
3768 /* scanned but unreclaimable */
3771 /* did we reclaim enough */
3772 if (zone_watermark_ok(zone, order, mark,
3773 ac_classzone_idx(ac), alloc_flags))
3781 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3782 gfp_mask, alloc_flags, ac->migratetype);
3784 prep_new_page(page, order, gfp_mask, alloc_flags);
3787 * If this is a high-order atomic allocation then check
3788 * if the pageblock should be reserved for the future
3790 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3791 reserve_highatomic_pageblock(page, zone, order);
3795 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3796 /* Try again if zone has deferred pages */
3797 if (static_branch_unlikely(&deferred_pages)) {
3798 if (_deferred_grow_zone(zone, order))
3806 * It's possible on a UMA machine to get through all zones that are
3807 * fragmented. If avoiding fragmentation, reset and try again.
3810 alloc_flags &= ~ALLOC_NOFRAGMENT;
3817 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3819 unsigned int filter = SHOW_MEM_FILTER_NODES;
3822 * This documents exceptions given to allocations in certain
3823 * contexts that are allowed to allocate outside current's set
3826 if (!(gfp_mask & __GFP_NOMEMALLOC))
3827 if (tsk_is_oom_victim(current) ||
3828 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3829 filter &= ~SHOW_MEM_FILTER_NODES;
3830 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3831 filter &= ~SHOW_MEM_FILTER_NODES;
3833 show_mem(filter, nodemask);
3836 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3838 struct va_format vaf;
3840 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3842 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3845 va_start(args, fmt);
3848 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3849 current->comm, &vaf, gfp_mask, &gfp_mask,
3850 nodemask_pr_args(nodemask));
3853 cpuset_print_current_mems_allowed();
3856 warn_alloc_show_mem(gfp_mask, nodemask);
3859 static inline struct page *
3860 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3861 unsigned int alloc_flags,
3862 const struct alloc_context *ac)
3866 page = get_page_from_freelist(gfp_mask, order,
3867 alloc_flags|ALLOC_CPUSET, ac);
3869 * fallback to ignore cpuset restriction if our nodes
3873 page = get_page_from_freelist(gfp_mask, order,
3879 static inline struct page *
3880 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3881 const struct alloc_context *ac, unsigned long *did_some_progress)
3883 struct oom_control oc = {
3884 .zonelist = ac->zonelist,
3885 .nodemask = ac->nodemask,
3887 .gfp_mask = gfp_mask,
3892 *did_some_progress = 0;
3895 * Acquire the oom lock. If that fails, somebody else is
3896 * making progress for us.
3898 if (!mutex_trylock(&oom_lock)) {
3899 *did_some_progress = 1;
3900 schedule_timeout_uninterruptible(1);
3905 * Go through the zonelist yet one more time, keep very high watermark
3906 * here, this is only to catch a parallel oom killing, we must fail if
3907 * we're still under heavy pressure. But make sure that this reclaim
3908 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3909 * allocation which will never fail due to oom_lock already held.
3911 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3912 ~__GFP_DIRECT_RECLAIM, order,
3913 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3917 /* Coredumps can quickly deplete all memory reserves */
3918 if (current->flags & PF_DUMPCORE)
3920 /* The OOM killer will not help higher order allocs */
3921 if (order > PAGE_ALLOC_COSTLY_ORDER)
3924 * We have already exhausted all our reclaim opportunities without any
3925 * success so it is time to admit defeat. We will skip the OOM killer
3926 * because it is very likely that the caller has a more reasonable
3927 * fallback than shooting a random task.
3929 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3931 /* The OOM killer does not needlessly kill tasks for lowmem */
3932 if (ac->high_zoneidx < ZONE_NORMAL)
3934 if (pm_suspended_storage())
3937 * XXX: GFP_NOFS allocations should rather fail than rely on
3938 * other request to make a forward progress.
3939 * We are in an unfortunate situation where out_of_memory cannot
3940 * do much for this context but let's try it to at least get
3941 * access to memory reserved if the current task is killed (see
3942 * out_of_memory). Once filesystems are ready to handle allocation
3943 * failures more gracefully we should just bail out here.
3946 /* The OOM killer may not free memory on a specific node */
3947 if (gfp_mask & __GFP_THISNODE)
3950 /* Exhausted what can be done so it's blame time */
3951 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3952 *did_some_progress = 1;
3955 * Help non-failing allocations by giving them access to memory
3958 if (gfp_mask & __GFP_NOFAIL)
3959 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3960 ALLOC_NO_WATERMARKS, ac);
3963 mutex_unlock(&oom_lock);
3968 * Maximum number of compaction retries wit a progress before OOM
3969 * killer is consider as the only way to move forward.
3971 #define MAX_COMPACT_RETRIES 16
3973 #ifdef CONFIG_COMPACTION
3974 /* Try memory compaction for high-order allocations before reclaim */
3975 static struct page *
3976 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3977 unsigned int alloc_flags, const struct alloc_context *ac,
3978 enum compact_priority prio, enum compact_result *compact_result)
3980 struct page *page = NULL;
3981 unsigned long pflags;
3982 unsigned int noreclaim_flag;
3987 psi_memstall_enter(&pflags);
3988 noreclaim_flag = memalloc_noreclaim_save();
3990 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3993 memalloc_noreclaim_restore(noreclaim_flag);
3994 psi_memstall_leave(&pflags);
3997 * At least in one zone compaction wasn't deferred or skipped, so let's
3998 * count a compaction stall
4000 count_vm_event(COMPACTSTALL);
4002 /* Prep a captured page if available */
4004 prep_new_page(page, order, gfp_mask, alloc_flags);
4006 /* Try get a page from the freelist if available */
4008 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4011 struct zone *zone = page_zone(page);
4013 zone->compact_blockskip_flush = false;
4014 compaction_defer_reset(zone, order, true);
4015 count_vm_event(COMPACTSUCCESS);
4020 * It's bad if compaction run occurs and fails. The most likely reason
4021 * is that pages exist, but not enough to satisfy watermarks.
4023 count_vm_event(COMPACTFAIL);
4031 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4032 enum compact_result compact_result,
4033 enum compact_priority *compact_priority,
4034 int *compaction_retries)
4036 int max_retries = MAX_COMPACT_RETRIES;
4039 int retries = *compaction_retries;
4040 enum compact_priority priority = *compact_priority;
4045 if (compaction_made_progress(compact_result))
4046 (*compaction_retries)++;
4049 * compaction considers all the zone as desperately out of memory
4050 * so it doesn't really make much sense to retry except when the
4051 * failure could be caused by insufficient priority
4053 if (compaction_failed(compact_result))
4054 goto check_priority;
4057 * compaction was skipped because there are not enough order-0 pages
4058 * to work with, so we retry only if it looks like reclaim can help.
4060 if (compaction_needs_reclaim(compact_result)) {
4061 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4066 * make sure the compaction wasn't deferred or didn't bail out early
4067 * due to locks contention before we declare that we should give up.
4068 * But the next retry should use a higher priority if allowed, so
4069 * we don't just keep bailing out endlessly.
4071 if (compaction_withdrawn(compact_result)) {
4072 goto check_priority;
4076 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4077 * costly ones because they are de facto nofail and invoke OOM
4078 * killer to move on while costly can fail and users are ready
4079 * to cope with that. 1/4 retries is rather arbitrary but we
4080 * would need much more detailed feedback from compaction to
4081 * make a better decision.
4083 if (order > PAGE_ALLOC_COSTLY_ORDER)
4085 if (*compaction_retries <= max_retries) {
4091 * Make sure there are attempts at the highest priority if we exhausted
4092 * all retries or failed at the lower priorities.
4095 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4096 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4098 if (*compact_priority > min_priority) {
4099 (*compact_priority)--;
4100 *compaction_retries = 0;
4104 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4108 static inline struct page *
4109 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4110 unsigned int alloc_flags, const struct alloc_context *ac,
4111 enum compact_priority prio, enum compact_result *compact_result)
4113 *compact_result = COMPACT_SKIPPED;
4118 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4119 enum compact_result compact_result,
4120 enum compact_priority *compact_priority,
4121 int *compaction_retries)
4126 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4130 * There are setups with compaction disabled which would prefer to loop
4131 * inside the allocator rather than hit the oom killer prematurely.
4132 * Let's give them a good hope and keep retrying while the order-0
4133 * watermarks are OK.
4135 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4137 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4138 ac_classzone_idx(ac), alloc_flags))
4143 #endif /* CONFIG_COMPACTION */
4145 #ifdef CONFIG_LOCKDEP
4146 static struct lockdep_map __fs_reclaim_map =
4147 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4149 static bool __need_fs_reclaim(gfp_t gfp_mask)
4151 gfp_mask = current_gfp_context(gfp_mask);
4153 /* no reclaim without waiting on it */
4154 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4157 /* this guy won't enter reclaim */
4158 if (current->flags & PF_MEMALLOC)
4161 /* We're only interested __GFP_FS allocations for now */
4162 if (!(gfp_mask & __GFP_FS))
4165 if (gfp_mask & __GFP_NOLOCKDEP)
4171 void __fs_reclaim_acquire(void)
4173 lock_map_acquire(&__fs_reclaim_map);
4176 void __fs_reclaim_release(void)
4178 lock_map_release(&__fs_reclaim_map);
4181 void fs_reclaim_acquire(gfp_t gfp_mask)
4183 if (__need_fs_reclaim(gfp_mask))
4184 __fs_reclaim_acquire();
4186 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4188 void fs_reclaim_release(gfp_t gfp_mask)
4190 if (__need_fs_reclaim(gfp_mask))
4191 __fs_reclaim_release();
4193 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4196 /* Perform direct synchronous page reclaim */
4198 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4199 const struct alloc_context *ac)
4202 unsigned int noreclaim_flag;
4203 unsigned long pflags;
4207 /* We now go into synchronous reclaim */
4208 cpuset_memory_pressure_bump();
4209 psi_memstall_enter(&pflags);
4210 fs_reclaim_acquire(gfp_mask);
4211 noreclaim_flag = memalloc_noreclaim_save();
4213 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4216 memalloc_noreclaim_restore(noreclaim_flag);
4217 fs_reclaim_release(gfp_mask);
4218 psi_memstall_leave(&pflags);
4225 /* The really slow allocator path where we enter direct reclaim */
4226 static inline struct page *
4227 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4228 unsigned int alloc_flags, const struct alloc_context *ac,
4229 unsigned long *did_some_progress)
4231 struct page *page = NULL;
4232 bool drained = false;
4234 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4235 if (unlikely(!(*did_some_progress)))
4239 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4242 * If an allocation failed after direct reclaim, it could be because
4243 * pages are pinned on the per-cpu lists or in high alloc reserves.
4244 * Shrink them them and try again
4246 if (!page && !drained) {
4247 unreserve_highatomic_pageblock(ac, false);
4248 drain_all_pages(NULL);
4256 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4257 const struct alloc_context *ac)
4261 pg_data_t *last_pgdat = NULL;
4262 enum zone_type high_zoneidx = ac->high_zoneidx;
4264 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
4266 if (last_pgdat != zone->zone_pgdat)
4267 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
4268 last_pgdat = zone->zone_pgdat;
4272 static inline unsigned int
4273 gfp_to_alloc_flags(gfp_t gfp_mask)
4275 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4278 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4279 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4280 * to save two branches.
4282 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4283 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4286 * The caller may dip into page reserves a bit more if the caller
4287 * cannot run direct reclaim, or if the caller has realtime scheduling
4288 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4289 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4291 alloc_flags |= (__force int)
4292 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4294 if (gfp_mask & __GFP_ATOMIC) {
4296 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4297 * if it can't schedule.
4299 if (!(gfp_mask & __GFP_NOMEMALLOC))
4300 alloc_flags |= ALLOC_HARDER;
4302 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4303 * comment for __cpuset_node_allowed().
4305 alloc_flags &= ~ALLOC_CPUSET;
4306 } else if (unlikely(rt_task(current)) && !in_interrupt())
4307 alloc_flags |= ALLOC_HARDER;
4310 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4311 alloc_flags |= ALLOC_CMA;
4316 static bool oom_reserves_allowed(struct task_struct *tsk)
4318 if (!tsk_is_oom_victim(tsk))
4322 * !MMU doesn't have oom reaper so give access to memory reserves
4323 * only to the thread with TIF_MEMDIE set
4325 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4332 * Distinguish requests which really need access to full memory
4333 * reserves from oom victims which can live with a portion of it
4335 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4337 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4339 if (gfp_mask & __GFP_MEMALLOC)
4340 return ALLOC_NO_WATERMARKS;
4341 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4342 return ALLOC_NO_WATERMARKS;
4343 if (!in_interrupt()) {
4344 if (current->flags & PF_MEMALLOC)
4345 return ALLOC_NO_WATERMARKS;
4346 else if (oom_reserves_allowed(current))
4353 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4355 return !!__gfp_pfmemalloc_flags(gfp_mask);
4359 * Checks whether it makes sense to retry the reclaim to make a forward progress
4360 * for the given allocation request.
4362 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4363 * without success, or when we couldn't even meet the watermark if we
4364 * reclaimed all remaining pages on the LRU lists.
4366 * Returns true if a retry is viable or false to enter the oom path.
4369 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4370 struct alloc_context *ac, int alloc_flags,
4371 bool did_some_progress, int *no_progress_loops)
4378 * Costly allocations might have made a progress but this doesn't mean
4379 * their order will become available due to high fragmentation so
4380 * always increment the no progress counter for them
4382 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4383 *no_progress_loops = 0;
4385 (*no_progress_loops)++;
4388 * Make sure we converge to OOM if we cannot make any progress
4389 * several times in the row.
4391 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4392 /* Before OOM, exhaust highatomic_reserve */
4393 return unreserve_highatomic_pageblock(ac, true);
4397 * Keep reclaiming pages while there is a chance this will lead
4398 * somewhere. If none of the target zones can satisfy our allocation
4399 * request even if all reclaimable pages are considered then we are
4400 * screwed and have to go OOM.
4402 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4404 unsigned long available;
4405 unsigned long reclaimable;
4406 unsigned long min_wmark = min_wmark_pages(zone);
4409 available = reclaimable = zone_reclaimable_pages(zone);
4410 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4413 * Would the allocation succeed if we reclaimed all
4414 * reclaimable pages?
4416 wmark = __zone_watermark_ok(zone, order, min_wmark,
4417 ac_classzone_idx(ac), alloc_flags, available);
4418 trace_reclaim_retry_zone(z, order, reclaimable,
4419 available, min_wmark, *no_progress_loops, wmark);
4422 * If we didn't make any progress and have a lot of
4423 * dirty + writeback pages then we should wait for
4424 * an IO to complete to slow down the reclaim and
4425 * prevent from pre mature OOM
4427 if (!did_some_progress) {
4428 unsigned long write_pending;
4430 write_pending = zone_page_state_snapshot(zone,
4431 NR_ZONE_WRITE_PENDING);
4433 if (2 * write_pending > reclaimable) {
4434 congestion_wait(BLK_RW_ASYNC, HZ/10);
4446 * Memory allocation/reclaim might be called from a WQ context and the
4447 * current implementation of the WQ concurrency control doesn't
4448 * recognize that a particular WQ is congested if the worker thread is
4449 * looping without ever sleeping. Therefore we have to do a short sleep
4450 * here rather than calling cond_resched().
4452 if (current->flags & PF_WQ_WORKER)
4453 schedule_timeout_uninterruptible(1);
4460 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4463 * It's possible that cpuset's mems_allowed and the nodemask from
4464 * mempolicy don't intersect. This should be normally dealt with by
4465 * policy_nodemask(), but it's possible to race with cpuset update in
4466 * such a way the check therein was true, and then it became false
4467 * before we got our cpuset_mems_cookie here.
4468 * This assumes that for all allocations, ac->nodemask can come only
4469 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4470 * when it does not intersect with the cpuset restrictions) or the
4471 * caller can deal with a violated nodemask.
4473 if (cpusets_enabled() && ac->nodemask &&
4474 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4475 ac->nodemask = NULL;
4480 * When updating a task's mems_allowed or mempolicy nodemask, it is
4481 * possible to race with parallel threads in such a way that our
4482 * allocation can fail while the mask is being updated. If we are about
4483 * to fail, check if the cpuset changed during allocation and if so,
4486 if (read_mems_allowed_retry(cpuset_mems_cookie))
4492 static inline struct page *
4493 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4494 struct alloc_context *ac)
4496 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4497 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4498 struct page *page = NULL;
4499 unsigned int alloc_flags;
4500 unsigned long did_some_progress;
4501 enum compact_priority compact_priority;
4502 enum compact_result compact_result;
4503 int compaction_retries;
4504 int no_progress_loops;
4505 unsigned int cpuset_mems_cookie;
4509 * We also sanity check to catch abuse of atomic reserves being used by
4510 * callers that are not in atomic context.
4512 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4513 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4514 gfp_mask &= ~__GFP_ATOMIC;
4517 compaction_retries = 0;
4518 no_progress_loops = 0;
4519 compact_priority = DEF_COMPACT_PRIORITY;
4520 cpuset_mems_cookie = read_mems_allowed_begin();
4523 * The fast path uses conservative alloc_flags to succeed only until
4524 * kswapd needs to be woken up, and to avoid the cost of setting up
4525 * alloc_flags precisely. So we do that now.
4527 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4530 * We need to recalculate the starting point for the zonelist iterator
4531 * because we might have used different nodemask in the fast path, or
4532 * there was a cpuset modification and we are retrying - otherwise we
4533 * could end up iterating over non-eligible zones endlessly.
4535 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4536 ac->high_zoneidx, ac->nodemask);
4537 if (!ac->preferred_zoneref->zone)
4540 if (alloc_flags & ALLOC_KSWAPD)
4541 wake_all_kswapds(order, gfp_mask, ac);
4544 * The adjusted alloc_flags might result in immediate success, so try
4547 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4552 * For costly allocations, try direct compaction first, as it's likely
4553 * that we have enough base pages and don't need to reclaim. For non-
4554 * movable high-order allocations, do that as well, as compaction will
4555 * try prevent permanent fragmentation by migrating from blocks of the
4557 * Don't try this for allocations that are allowed to ignore
4558 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4560 if (can_direct_reclaim &&
4562 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4563 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4564 page = __alloc_pages_direct_compact(gfp_mask, order,
4566 INIT_COMPACT_PRIORITY,
4572 * Checks for costly allocations with __GFP_NORETRY, which
4573 * includes some THP page fault allocations
4575 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4577 * If allocating entire pageblock(s) and compaction
4578 * failed because all zones are below low watermarks
4579 * or is prohibited because it recently failed at this
4580 * order, fail immediately unless the allocator has
4581 * requested compaction and reclaim retry.
4584 * - potentially very expensive because zones are far
4585 * below their low watermarks or this is part of very
4586 * bursty high order allocations,
4587 * - not guaranteed to help because isolate_freepages()
4588 * may not iterate over freed pages as part of its
4590 * - unlikely to make entire pageblocks free on its
4593 if (compact_result == COMPACT_SKIPPED ||
4594 compact_result == COMPACT_DEFERRED)
4598 * Looks like reclaim/compaction is worth trying, but
4599 * sync compaction could be very expensive, so keep
4600 * using async compaction.
4602 compact_priority = INIT_COMPACT_PRIORITY;
4607 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4608 if (alloc_flags & ALLOC_KSWAPD)
4609 wake_all_kswapds(order, gfp_mask, ac);
4611 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4613 alloc_flags = reserve_flags;
4616 * Reset the nodemask and zonelist iterators if memory policies can be
4617 * ignored. These allocations are high priority and system rather than
4620 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4621 ac->nodemask = NULL;
4622 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4623 ac->high_zoneidx, ac->nodemask);
4626 /* Attempt with potentially adjusted zonelist and alloc_flags */
4627 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4631 /* Caller is not willing to reclaim, we can't balance anything */
4632 if (!can_direct_reclaim)
4635 /* Avoid recursion of direct reclaim */
4636 if (current->flags & PF_MEMALLOC)
4639 /* Try direct reclaim and then allocating */
4640 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4641 &did_some_progress);
4645 /* Try direct compaction and then allocating */
4646 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4647 compact_priority, &compact_result);
4651 /* Do not loop if specifically requested */
4652 if (gfp_mask & __GFP_NORETRY)
4656 * Do not retry costly high order allocations unless they are
4657 * __GFP_RETRY_MAYFAIL
4659 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4662 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4663 did_some_progress > 0, &no_progress_loops))
4667 * It doesn't make any sense to retry for the compaction if the order-0
4668 * reclaim is not able to make any progress because the current
4669 * implementation of the compaction depends on the sufficient amount
4670 * of free memory (see __compaction_suitable)
4672 if (did_some_progress > 0 &&
4673 should_compact_retry(ac, order, alloc_flags,
4674 compact_result, &compact_priority,
4675 &compaction_retries))
4679 /* Deal with possible cpuset update races before we start OOM killing */
4680 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4683 /* Reclaim has failed us, start killing things */
4684 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4688 /* Avoid allocations with no watermarks from looping endlessly */
4689 if (tsk_is_oom_victim(current) &&
4690 (alloc_flags == ALLOC_OOM ||
4691 (gfp_mask & __GFP_NOMEMALLOC)))
4694 /* Retry as long as the OOM killer is making progress */
4695 if (did_some_progress) {
4696 no_progress_loops = 0;
4701 /* Deal with possible cpuset update races before we fail */
4702 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4706 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4709 if (gfp_mask & __GFP_NOFAIL) {
4711 * All existing users of the __GFP_NOFAIL are blockable, so warn
4712 * of any new users that actually require GFP_NOWAIT
4714 if (WARN_ON_ONCE(!can_direct_reclaim))
4718 * PF_MEMALLOC request from this context is rather bizarre
4719 * because we cannot reclaim anything and only can loop waiting
4720 * for somebody to do a work for us
4722 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4725 * non failing costly orders are a hard requirement which we
4726 * are not prepared for much so let's warn about these users
4727 * so that we can identify them and convert them to something
4730 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4733 * Help non-failing allocations by giving them access to memory
4734 * reserves but do not use ALLOC_NO_WATERMARKS because this
4735 * could deplete whole memory reserves which would just make
4736 * the situation worse
4738 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4746 warn_alloc(gfp_mask, ac->nodemask,
4747 "page allocation failure: order:%u", order);
4752 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4753 int preferred_nid, nodemask_t *nodemask,
4754 struct alloc_context *ac, gfp_t *alloc_mask,
4755 unsigned int *alloc_flags)
4757 ac->high_zoneidx = gfp_zone(gfp_mask);
4758 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4759 ac->nodemask = nodemask;
4760 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4762 if (cpusets_enabled()) {
4763 *alloc_mask |= __GFP_HARDWALL;
4765 ac->nodemask = &cpuset_current_mems_allowed;
4767 *alloc_flags |= ALLOC_CPUSET;
4770 fs_reclaim_acquire(gfp_mask);
4771 fs_reclaim_release(gfp_mask);
4773 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4775 if (should_fail_alloc_page(gfp_mask, order))
4778 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4779 *alloc_flags |= ALLOC_CMA;
4784 /* Determine whether to spread dirty pages and what the first usable zone */
4785 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4787 /* Dirty zone balancing only done in the fast path */
4788 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4791 * The preferred zone is used for statistics but crucially it is
4792 * also used as the starting point for the zonelist iterator. It
4793 * may get reset for allocations that ignore memory policies.
4795 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4796 ac->high_zoneidx, ac->nodemask);
4800 * This is the 'heart' of the zoned buddy allocator.
4803 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4804 nodemask_t *nodemask)
4807 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4808 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4809 struct alloc_context ac = { };
4812 * There are several places where we assume that the order value is sane
4813 * so bail out early if the request is out of bound.
4815 if (unlikely(order >= MAX_ORDER)) {
4816 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4820 gfp_mask &= gfp_allowed_mask;
4821 alloc_mask = gfp_mask;
4822 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4825 finalise_ac(gfp_mask, &ac);
4828 * Forbid the first pass from falling back to types that fragment
4829 * memory until all local zones are considered.
4831 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4833 /* First allocation attempt */
4834 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4839 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4840 * resp. GFP_NOIO which has to be inherited for all allocation requests
4841 * from a particular context which has been marked by
4842 * memalloc_no{fs,io}_{save,restore}.
4844 alloc_mask = current_gfp_context(gfp_mask);
4845 ac.spread_dirty_pages = false;
4848 * Restore the original nodemask if it was potentially replaced with
4849 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4851 ac.nodemask = nodemask;
4853 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4856 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4857 unlikely(__memcg_kmem_charge_page(page, gfp_mask, order) != 0)) {
4858 __free_pages(page, order);
4862 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4866 EXPORT_SYMBOL(__alloc_pages_nodemask);
4869 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4870 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4871 * you need to access high mem.
4873 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4877 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4880 return (unsigned long) page_address(page);
4882 EXPORT_SYMBOL(__get_free_pages);
4884 unsigned long get_zeroed_page(gfp_t gfp_mask)
4886 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4888 EXPORT_SYMBOL(get_zeroed_page);
4890 static inline void free_the_page(struct page *page, unsigned int order)
4892 if (order == 0) /* Via pcp? */
4893 free_unref_page(page);
4895 __free_pages_ok(page, order);
4898 void __free_pages(struct page *page, unsigned int order)
4900 if (put_page_testzero(page))
4901 free_the_page(page, order);
4903 EXPORT_SYMBOL(__free_pages);
4905 void free_pages(unsigned long addr, unsigned int order)
4908 VM_BUG_ON(!virt_addr_valid((void *)addr));
4909 __free_pages(virt_to_page((void *)addr), order);
4913 EXPORT_SYMBOL(free_pages);
4917 * An arbitrary-length arbitrary-offset area of memory which resides
4918 * within a 0 or higher order page. Multiple fragments within that page
4919 * are individually refcounted, in the page's reference counter.
4921 * The page_frag functions below provide a simple allocation framework for
4922 * page fragments. This is used by the network stack and network device
4923 * drivers to provide a backing region of memory for use as either an
4924 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4926 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4929 struct page *page = NULL;
4930 gfp_t gfp = gfp_mask;
4932 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4933 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4935 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4936 PAGE_FRAG_CACHE_MAX_ORDER);
4937 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4939 if (unlikely(!page))
4940 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4942 nc->va = page ? page_address(page) : NULL;
4947 void __page_frag_cache_drain(struct page *page, unsigned int count)
4949 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4951 if (page_ref_sub_and_test(page, count))
4952 free_the_page(page, compound_order(page));
4954 EXPORT_SYMBOL(__page_frag_cache_drain);
4956 void *page_frag_alloc(struct page_frag_cache *nc,
4957 unsigned int fragsz, gfp_t gfp_mask)
4959 unsigned int size = PAGE_SIZE;
4963 if (unlikely(!nc->va)) {
4965 page = __page_frag_cache_refill(nc, gfp_mask);
4969 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4970 /* if size can vary use size else just use PAGE_SIZE */
4973 /* Even if we own the page, we do not use atomic_set().
4974 * This would break get_page_unless_zero() users.
4976 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4978 /* reset page count bias and offset to start of new frag */
4979 nc->pfmemalloc = page_is_pfmemalloc(page);
4980 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4984 offset = nc->offset - fragsz;
4985 if (unlikely(offset < 0)) {
4986 page = virt_to_page(nc->va);
4988 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4991 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4992 /* if size can vary use size else just use PAGE_SIZE */
4995 /* OK, page count is 0, we can safely set it */
4996 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4998 /* reset page count bias and offset to start of new frag */
4999 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5000 offset = size - fragsz;
5004 nc->offset = offset;
5006 return nc->va + offset;
5008 EXPORT_SYMBOL(page_frag_alloc);
5011 * Frees a page fragment allocated out of either a compound or order 0 page.
5013 void page_frag_free(void *addr)
5015 struct page *page = virt_to_head_page(addr);
5017 if (unlikely(put_page_testzero(page)))
5018 free_the_page(page, compound_order(page));
5020 EXPORT_SYMBOL(page_frag_free);
5022 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5026 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5027 unsigned long used = addr + PAGE_ALIGN(size);
5029 split_page(virt_to_page((void *)addr), order);
5030 while (used < alloc_end) {
5035 return (void *)addr;
5039 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5040 * @size: the number of bytes to allocate
5041 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5043 * This function is similar to alloc_pages(), except that it allocates the
5044 * minimum number of pages to satisfy the request. alloc_pages() can only
5045 * allocate memory in power-of-two pages.
5047 * This function is also limited by MAX_ORDER.
5049 * Memory allocated by this function must be released by free_pages_exact().
5051 * Return: pointer to the allocated area or %NULL in case of error.
5053 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5055 unsigned int order = get_order(size);
5058 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5059 gfp_mask &= ~__GFP_COMP;
5061 addr = __get_free_pages(gfp_mask, order);
5062 return make_alloc_exact(addr, order, size);
5064 EXPORT_SYMBOL(alloc_pages_exact);
5067 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5069 * @nid: the preferred node ID where memory should be allocated
5070 * @size: the number of bytes to allocate
5071 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5073 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5076 * Return: pointer to the allocated area or %NULL in case of error.
5078 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5080 unsigned int order = get_order(size);
5083 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5084 gfp_mask &= ~__GFP_COMP;
5086 p = alloc_pages_node(nid, gfp_mask, order);
5089 return make_alloc_exact((unsigned long)page_address(p), order, size);
5093 * free_pages_exact - release memory allocated via alloc_pages_exact()
5094 * @virt: the value returned by alloc_pages_exact.
5095 * @size: size of allocation, same value as passed to alloc_pages_exact().
5097 * Release the memory allocated by a previous call to alloc_pages_exact.
5099 void free_pages_exact(void *virt, size_t size)
5101 unsigned long addr = (unsigned long)virt;
5102 unsigned long end = addr + PAGE_ALIGN(size);
5104 while (addr < end) {
5109 EXPORT_SYMBOL(free_pages_exact);
5112 * nr_free_zone_pages - count number of pages beyond high watermark
5113 * @offset: The zone index of the highest zone
5115 * nr_free_zone_pages() counts the number of pages which are beyond the
5116 * high watermark within all zones at or below a given zone index. For each
5117 * zone, the number of pages is calculated as:
5119 * nr_free_zone_pages = managed_pages - high_pages
5121 * Return: number of pages beyond high watermark.
5123 static unsigned long nr_free_zone_pages(int offset)
5128 /* Just pick one node, since fallback list is circular */
5129 unsigned long sum = 0;
5131 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5133 for_each_zone_zonelist(zone, z, zonelist, offset) {
5134 unsigned long size = zone_managed_pages(zone);
5135 unsigned long high = high_wmark_pages(zone);
5144 * nr_free_buffer_pages - count number of pages beyond high watermark
5146 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5147 * watermark within ZONE_DMA and ZONE_NORMAL.
5149 * Return: number of pages beyond high watermark within ZONE_DMA and
5152 unsigned long nr_free_buffer_pages(void)
5154 return nr_free_zone_pages(gfp_zone(GFP_USER));
5156 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5159 * nr_free_pagecache_pages - count number of pages beyond high watermark
5161 * nr_free_pagecache_pages() counts the number of pages which are beyond the
5162 * high watermark within all zones.
5164 * Return: number of pages beyond high watermark within all zones.
5166 unsigned long nr_free_pagecache_pages(void)
5168 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5171 static inline void show_node(struct zone *zone)
5173 if (IS_ENABLED(CONFIG_NUMA))
5174 printk("Node %d ", zone_to_nid(zone));
5177 long si_mem_available(void)
5180 unsigned long pagecache;
5181 unsigned long wmark_low = 0;
5182 unsigned long pages[NR_LRU_LISTS];
5183 unsigned long reclaimable;
5187 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5188 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5191 wmark_low += low_wmark_pages(zone);
5194 * Estimate the amount of memory available for userspace allocations,
5195 * without causing swapping.
5197 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5200 * Not all the page cache can be freed, otherwise the system will
5201 * start swapping. Assume at least half of the page cache, or the
5202 * low watermark worth of cache, needs to stay.
5204 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5205 pagecache -= min(pagecache / 2, wmark_low);
5206 available += pagecache;
5209 * Part of the reclaimable slab and other kernel memory consists of
5210 * items that are in use, and cannot be freed. Cap this estimate at the
5213 reclaimable = global_node_page_state(NR_SLAB_RECLAIMABLE) +
5214 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5215 available += reclaimable - min(reclaimable / 2, wmark_low);
5221 EXPORT_SYMBOL_GPL(si_mem_available);
5223 void si_meminfo(struct sysinfo *val)
5225 val->totalram = totalram_pages();
5226 val->sharedram = global_node_page_state(NR_SHMEM);
5227 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5228 val->bufferram = nr_blockdev_pages();
5229 val->totalhigh = totalhigh_pages();
5230 val->freehigh = nr_free_highpages();
5231 val->mem_unit = PAGE_SIZE;
5234 EXPORT_SYMBOL(si_meminfo);
5237 void si_meminfo_node(struct sysinfo *val, int nid)
5239 int zone_type; /* needs to be signed */
5240 unsigned long managed_pages = 0;
5241 unsigned long managed_highpages = 0;
5242 unsigned long free_highpages = 0;
5243 pg_data_t *pgdat = NODE_DATA(nid);
5245 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5246 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5247 val->totalram = managed_pages;
5248 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5249 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5250 #ifdef CONFIG_HIGHMEM
5251 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5252 struct zone *zone = &pgdat->node_zones[zone_type];
5254 if (is_highmem(zone)) {
5255 managed_highpages += zone_managed_pages(zone);
5256 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5259 val->totalhigh = managed_highpages;
5260 val->freehigh = free_highpages;
5262 val->totalhigh = managed_highpages;
5263 val->freehigh = free_highpages;
5265 val->mem_unit = PAGE_SIZE;
5270 * Determine whether the node should be displayed or not, depending on whether
5271 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5273 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5275 if (!(flags & SHOW_MEM_FILTER_NODES))
5279 * no node mask - aka implicit memory numa policy. Do not bother with
5280 * the synchronization - read_mems_allowed_begin - because we do not
5281 * have to be precise here.
5284 nodemask = &cpuset_current_mems_allowed;
5286 return !node_isset(nid, *nodemask);
5289 #define K(x) ((x) << (PAGE_SHIFT-10))
5291 static void show_migration_types(unsigned char type)
5293 static const char types[MIGRATE_TYPES] = {
5294 [MIGRATE_UNMOVABLE] = 'U',
5295 [MIGRATE_MOVABLE] = 'M',
5296 [MIGRATE_RECLAIMABLE] = 'E',
5297 [MIGRATE_HIGHATOMIC] = 'H',
5299 [MIGRATE_CMA] = 'C',
5301 #ifdef CONFIG_MEMORY_ISOLATION
5302 [MIGRATE_ISOLATE] = 'I',
5305 char tmp[MIGRATE_TYPES + 1];
5309 for (i = 0; i < MIGRATE_TYPES; i++) {
5310 if (type & (1 << i))
5315 printk(KERN_CONT "(%s) ", tmp);
5319 * Show free area list (used inside shift_scroll-lock stuff)
5320 * We also calculate the percentage fragmentation. We do this by counting the
5321 * memory on each free list with the exception of the first item on the list.
5324 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5327 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5329 unsigned long free_pcp = 0;
5334 for_each_populated_zone(zone) {
5335 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5338 for_each_online_cpu(cpu)
5339 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5342 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5343 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5344 " unevictable:%lu dirty:%lu writeback:%lu\n"
5345 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5346 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5347 " free:%lu free_pcp:%lu free_cma:%lu\n",
5348 global_node_page_state(NR_ACTIVE_ANON),
5349 global_node_page_state(NR_INACTIVE_ANON),
5350 global_node_page_state(NR_ISOLATED_ANON),
5351 global_node_page_state(NR_ACTIVE_FILE),
5352 global_node_page_state(NR_INACTIVE_FILE),
5353 global_node_page_state(NR_ISOLATED_FILE),
5354 global_node_page_state(NR_UNEVICTABLE),
5355 global_node_page_state(NR_FILE_DIRTY),
5356 global_node_page_state(NR_WRITEBACK),
5357 global_node_page_state(NR_SLAB_RECLAIMABLE),
5358 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
5359 global_node_page_state(NR_FILE_MAPPED),
5360 global_node_page_state(NR_SHMEM),
5361 global_zone_page_state(NR_PAGETABLE),
5362 global_zone_page_state(NR_BOUNCE),
5363 global_zone_page_state(NR_FREE_PAGES),
5365 global_zone_page_state(NR_FREE_CMA_PAGES));
5367 for_each_online_pgdat(pgdat) {
5368 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5372 " active_anon:%lukB"
5373 " inactive_anon:%lukB"
5374 " active_file:%lukB"
5375 " inactive_file:%lukB"
5376 " unevictable:%lukB"
5377 " isolated(anon):%lukB"
5378 " isolated(file):%lukB"
5383 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5385 " shmem_pmdmapped: %lukB"
5388 " writeback_tmp:%lukB"
5389 " all_unreclaimable? %s"
5392 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5393 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5394 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5395 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5396 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5397 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5398 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5399 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5400 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5401 K(node_page_state(pgdat, NR_WRITEBACK)),
5402 K(node_page_state(pgdat, NR_SHMEM)),
5403 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5404 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5405 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5407 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5409 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5410 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5414 for_each_populated_zone(zone) {
5417 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5421 for_each_online_cpu(cpu)
5422 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5431 " reserved_highatomic:%luKB"
5432 " active_anon:%lukB"
5433 " inactive_anon:%lukB"
5434 " active_file:%lukB"
5435 " inactive_file:%lukB"
5436 " unevictable:%lukB"
5437 " writepending:%lukB"
5441 " kernel_stack:%lukB"
5442 #ifdef CONFIG_SHADOW_CALL_STACK
5443 " shadow_call_stack:%lukB"
5452 K(zone_page_state(zone, NR_FREE_PAGES)),
5453 K(min_wmark_pages(zone)),
5454 K(low_wmark_pages(zone)),
5455 K(high_wmark_pages(zone)),
5456 K(zone->nr_reserved_highatomic),
5457 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5458 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5459 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5460 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5461 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5462 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5463 K(zone->present_pages),
5464 K(zone_managed_pages(zone)),
5465 K(zone_page_state(zone, NR_MLOCK)),
5466 zone_page_state(zone, NR_KERNEL_STACK_KB),
5467 #ifdef CONFIG_SHADOW_CALL_STACK
5468 zone_page_state(zone, NR_KERNEL_SCS_KB),
5470 K(zone_page_state(zone, NR_PAGETABLE)),
5471 K(zone_page_state(zone, NR_BOUNCE)),
5473 K(this_cpu_read(zone->pageset->pcp.count)),
5474 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5475 printk("lowmem_reserve[]:");
5476 for (i = 0; i < MAX_NR_ZONES; i++)
5477 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5478 printk(KERN_CONT "\n");
5481 for_each_populated_zone(zone) {
5483 unsigned long nr[MAX_ORDER], flags, total = 0;
5484 unsigned char types[MAX_ORDER];
5486 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5489 printk(KERN_CONT "%s: ", zone->name);
5491 spin_lock_irqsave(&zone->lock, flags);
5492 for (order = 0; order < MAX_ORDER; order++) {
5493 struct free_area *area = &zone->free_area[order];
5496 nr[order] = area->nr_free;
5497 total += nr[order] << order;
5500 for (type = 0; type < MIGRATE_TYPES; type++) {
5501 if (!free_area_empty(area, type))
5502 types[order] |= 1 << type;
5505 spin_unlock_irqrestore(&zone->lock, flags);
5506 for (order = 0; order < MAX_ORDER; order++) {
5507 printk(KERN_CONT "%lu*%lukB ",
5508 nr[order], K(1UL) << order);
5510 show_migration_types(types[order]);
5512 printk(KERN_CONT "= %lukB\n", K(total));
5515 hugetlb_show_meminfo();
5517 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5519 show_swap_cache_info();
5522 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5524 zoneref->zone = zone;
5525 zoneref->zone_idx = zone_idx(zone);
5529 * Builds allocation fallback zone lists.
5531 * Add all populated zones of a node to the zonelist.
5533 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5536 enum zone_type zone_type = MAX_NR_ZONES;
5541 zone = pgdat->node_zones + zone_type;
5542 if (managed_zone(zone)) {
5543 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5544 check_highest_zone(zone_type);
5546 } while (zone_type);
5553 static int __parse_numa_zonelist_order(char *s)
5556 * We used to support different zonlists modes but they turned
5557 * out to be just not useful. Let's keep the warning in place
5558 * if somebody still use the cmd line parameter so that we do
5559 * not fail it silently
5561 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5562 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5568 static __init int setup_numa_zonelist_order(char *s)
5573 return __parse_numa_zonelist_order(s);
5575 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5577 char numa_zonelist_order[] = "Node";
5580 * sysctl handler for numa_zonelist_order
5582 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5583 void __user *buffer, size_t *length,
5590 return proc_dostring(table, write, buffer, length, ppos);
5591 str = memdup_user_nul(buffer, 16);
5593 return PTR_ERR(str);
5595 ret = __parse_numa_zonelist_order(str);
5601 #define MAX_NODE_LOAD (nr_online_nodes)
5602 static int node_load[MAX_NUMNODES];
5605 * find_next_best_node - find the next node that should appear in a given node's fallback list
5606 * @node: node whose fallback list we're appending
5607 * @used_node_mask: nodemask_t of already used nodes
5609 * We use a number of factors to determine which is the next node that should
5610 * appear on a given node's fallback list. The node should not have appeared
5611 * already in @node's fallback list, and it should be the next closest node
5612 * according to the distance array (which contains arbitrary distance values
5613 * from each node to each node in the system), and should also prefer nodes
5614 * with no CPUs, since presumably they'll have very little allocation pressure
5615 * on them otherwise.
5617 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5619 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5622 int min_val = INT_MAX;
5623 int best_node = NUMA_NO_NODE;
5624 const struct cpumask *tmp = cpumask_of_node(0);
5626 /* Use the local node if we haven't already */
5627 if (!node_isset(node, *used_node_mask)) {
5628 node_set(node, *used_node_mask);
5632 for_each_node_state(n, N_MEMORY) {
5634 /* Don't want a node to appear more than once */
5635 if (node_isset(n, *used_node_mask))
5638 /* Use the distance array to find the distance */
5639 val = node_distance(node, n);
5641 /* Penalize nodes under us ("prefer the next node") */
5644 /* Give preference to headless and unused nodes */
5645 tmp = cpumask_of_node(n);
5646 if (!cpumask_empty(tmp))
5647 val += PENALTY_FOR_NODE_WITH_CPUS;
5649 /* Slight preference for less loaded node */
5650 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5651 val += node_load[n];
5653 if (val < min_val) {
5660 node_set(best_node, *used_node_mask);
5667 * Build zonelists ordered by node and zones within node.
5668 * This results in maximum locality--normal zone overflows into local
5669 * DMA zone, if any--but risks exhausting DMA zone.
5671 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5674 struct zoneref *zonerefs;
5677 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5679 for (i = 0; i < nr_nodes; i++) {
5682 pg_data_t *node = NODE_DATA(node_order[i]);
5684 nr_zones = build_zonerefs_node(node, zonerefs);
5685 zonerefs += nr_zones;
5687 zonerefs->zone = NULL;
5688 zonerefs->zone_idx = 0;
5692 * Build gfp_thisnode zonelists
5694 static void build_thisnode_zonelists(pg_data_t *pgdat)
5696 struct zoneref *zonerefs;
5699 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5700 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5701 zonerefs += nr_zones;
5702 zonerefs->zone = NULL;
5703 zonerefs->zone_idx = 0;
5707 * Build zonelists ordered by zone and nodes within zones.
5708 * This results in conserving DMA zone[s] until all Normal memory is
5709 * exhausted, but results in overflowing to remote node while memory
5710 * may still exist in local DMA zone.
5713 static void build_zonelists(pg_data_t *pgdat)
5715 static int node_order[MAX_NUMNODES];
5716 int node, load, nr_nodes = 0;
5717 nodemask_t used_mask;
5718 int local_node, prev_node;
5720 /* NUMA-aware ordering of nodes */
5721 local_node = pgdat->node_id;
5722 load = nr_online_nodes;
5723 prev_node = local_node;
5724 nodes_clear(used_mask);
5726 memset(node_order, 0, sizeof(node_order));
5727 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5729 * We don't want to pressure a particular node.
5730 * So adding penalty to the first node in same
5731 * distance group to make it round-robin.
5733 if (node_distance(local_node, node) !=
5734 node_distance(local_node, prev_node))
5735 node_load[node] = load;
5737 node_order[nr_nodes++] = node;
5742 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5743 build_thisnode_zonelists(pgdat);
5746 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5748 * Return node id of node used for "local" allocations.
5749 * I.e., first node id of first zone in arg node's generic zonelist.
5750 * Used for initializing percpu 'numa_mem', which is used primarily
5751 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5753 int local_memory_node(int node)
5757 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5758 gfp_zone(GFP_KERNEL),
5760 return zone_to_nid(z->zone);
5764 static void setup_min_unmapped_ratio(void);
5765 static void setup_min_slab_ratio(void);
5766 #else /* CONFIG_NUMA */
5768 static void build_zonelists(pg_data_t *pgdat)
5770 int node, local_node;
5771 struct zoneref *zonerefs;
5774 local_node = pgdat->node_id;
5776 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5777 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5778 zonerefs += nr_zones;
5781 * Now we build the zonelist so that it contains the zones
5782 * of all the other nodes.
5783 * We don't want to pressure a particular node, so when
5784 * building the zones for node N, we make sure that the
5785 * zones coming right after the local ones are those from
5786 * node N+1 (modulo N)
5788 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5789 if (!node_online(node))
5791 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5792 zonerefs += nr_zones;
5794 for (node = 0; node < local_node; node++) {
5795 if (!node_online(node))
5797 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5798 zonerefs += nr_zones;
5801 zonerefs->zone = NULL;
5802 zonerefs->zone_idx = 0;
5805 #endif /* CONFIG_NUMA */
5808 * Boot pageset table. One per cpu which is going to be used for all
5809 * zones and all nodes. The parameters will be set in such a way
5810 * that an item put on a list will immediately be handed over to
5811 * the buddy list. This is safe since pageset manipulation is done
5812 * with interrupts disabled.
5814 * The boot_pagesets must be kept even after bootup is complete for
5815 * unused processors and/or zones. They do play a role for bootstrapping
5816 * hotplugged processors.
5818 * zoneinfo_show() and maybe other functions do
5819 * not check if the processor is online before following the pageset pointer.
5820 * Other parts of the kernel may not check if the zone is available.
5822 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5823 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5824 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5826 static void __build_all_zonelists(void *data)
5829 int __maybe_unused cpu;
5830 pg_data_t *self = data;
5831 static DEFINE_SPINLOCK(lock);
5836 memset(node_load, 0, sizeof(node_load));
5840 * This node is hotadded and no memory is yet present. So just
5841 * building zonelists is fine - no need to touch other nodes.
5843 if (self && !node_online(self->node_id)) {
5844 build_zonelists(self);
5846 for_each_online_node(nid) {
5847 pg_data_t *pgdat = NODE_DATA(nid);
5849 build_zonelists(pgdat);
5852 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5854 * We now know the "local memory node" for each node--
5855 * i.e., the node of the first zone in the generic zonelist.
5856 * Set up numa_mem percpu variable for on-line cpus. During
5857 * boot, only the boot cpu should be on-line; we'll init the
5858 * secondary cpus' numa_mem as they come on-line. During
5859 * node/memory hotplug, we'll fixup all on-line cpus.
5861 for_each_online_cpu(cpu)
5862 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5869 static noinline void __init
5870 build_all_zonelists_init(void)
5874 __build_all_zonelists(NULL);
5877 * Initialize the boot_pagesets that are going to be used
5878 * for bootstrapping processors. The real pagesets for
5879 * each zone will be allocated later when the per cpu
5880 * allocator is available.
5882 * boot_pagesets are used also for bootstrapping offline
5883 * cpus if the system is already booted because the pagesets
5884 * are needed to initialize allocators on a specific cpu too.
5885 * F.e. the percpu allocator needs the page allocator which
5886 * needs the percpu allocator in order to allocate its pagesets
5887 * (a chicken-egg dilemma).
5889 for_each_possible_cpu(cpu)
5890 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5892 mminit_verify_zonelist();
5893 cpuset_init_current_mems_allowed();
5897 * unless system_state == SYSTEM_BOOTING.
5899 * __ref due to call of __init annotated helper build_all_zonelists_init
5900 * [protected by SYSTEM_BOOTING].
5902 void __ref build_all_zonelists(pg_data_t *pgdat)
5904 if (system_state == SYSTEM_BOOTING) {
5905 build_all_zonelists_init();
5907 __build_all_zonelists(pgdat);
5908 /* cpuset refresh routine should be here */
5910 vm_total_pages = nr_free_pagecache_pages();
5912 * Disable grouping by mobility if the number of pages in the
5913 * system is too low to allow the mechanism to work. It would be
5914 * more accurate, but expensive to check per-zone. This check is
5915 * made on memory-hotadd so a system can start with mobility
5916 * disabled and enable it later
5918 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5919 page_group_by_mobility_disabled = 1;
5921 page_group_by_mobility_disabled = 0;
5923 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5925 page_group_by_mobility_disabled ? "off" : "on",
5928 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5932 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5933 static bool __meminit
5934 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5936 static struct memblock_region *r;
5938 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5939 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5940 for_each_memblock(memory, r) {
5941 if (*pfn < memblock_region_memory_end_pfn(r))
5945 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5946 memblock_is_mirror(r)) {
5947 *pfn = memblock_region_memory_end_pfn(r);
5954 #ifdef CONFIG_SPARSEMEM
5955 /* Skip PFNs that belong to non-present sections */
5956 static inline __meminit unsigned long next_pfn(unsigned long pfn)
5958 const unsigned long section_nr = pfn_to_section_nr(++pfn);
5960 if (present_section_nr(section_nr))
5962 return section_nr_to_pfn(next_present_section_nr(section_nr));
5965 static inline __meminit unsigned long next_pfn(unsigned long pfn)
5972 * Initially all pages are reserved - free ones are freed
5973 * up by memblock_free_all() once the early boot process is
5974 * done. Non-atomic initialization, single-pass.
5976 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5977 unsigned long start_pfn, enum memmap_context context,
5978 struct vmem_altmap *altmap)
5980 unsigned long pfn, end_pfn = start_pfn + size;
5983 if (highest_memmap_pfn < end_pfn - 1)
5984 highest_memmap_pfn = end_pfn - 1;
5986 #ifdef CONFIG_ZONE_DEVICE
5988 * Honor reservation requested by the driver for this ZONE_DEVICE
5989 * memory. We limit the total number of pages to initialize to just
5990 * those that might contain the memory mapping. We will defer the
5991 * ZONE_DEVICE page initialization until after we have released
5994 if (zone == ZONE_DEVICE) {
5998 if (start_pfn == altmap->base_pfn)
5999 start_pfn += altmap->reserve;
6000 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6004 for (pfn = start_pfn; pfn < end_pfn; ) {
6006 * There can be holes in boot-time mem_map[]s handed to this
6007 * function. They do not exist on hotplugged memory.
6009 if (context == MEMMAP_EARLY) {
6010 if (!early_pfn_valid(pfn)) {
6011 pfn = next_pfn(pfn);
6014 if (!early_pfn_in_nid(pfn, nid)) {
6018 if (overlap_memmap_init(zone, &pfn))
6020 if (defer_init(nid, pfn, end_pfn))
6024 page = pfn_to_page(pfn);
6025 __init_single_page(page, pfn, zone, nid);
6026 if (context == MEMMAP_HOTPLUG)
6027 __SetPageReserved(page);
6030 * Mark the block movable so that blocks are reserved for
6031 * movable at startup. This will force kernel allocations
6032 * to reserve their blocks rather than leaking throughout
6033 * the address space during boot when many long-lived
6034 * kernel allocations are made.
6036 * bitmap is created for zone's valid pfn range. but memmap
6037 * can be created for invalid pages (for alignment)
6038 * check here not to call set_pageblock_migratetype() against
6041 if (!(pfn & (pageblock_nr_pages - 1))) {
6042 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6049 #ifdef CONFIG_ZONE_DEVICE
6050 void __ref memmap_init_zone_device(struct zone *zone,
6051 unsigned long start_pfn,
6052 unsigned long nr_pages,
6053 struct dev_pagemap *pgmap)
6055 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6056 struct pglist_data *pgdat = zone->zone_pgdat;
6057 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6058 unsigned long zone_idx = zone_idx(zone);
6059 unsigned long start = jiffies;
6060 int nid = pgdat->node_id;
6062 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6066 * The call to memmap_init_zone should have already taken care
6067 * of the pages reserved for the memmap, so we can just jump to
6068 * the end of that region and start processing the device pages.
6071 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6072 nr_pages = end_pfn - start_pfn;
6075 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6076 struct page *page = pfn_to_page(pfn);
6078 __init_single_page(page, pfn, zone_idx, nid);
6081 * Mark page reserved as it will need to wait for onlining
6082 * phase for it to be fully associated with a zone.
6084 * We can use the non-atomic __set_bit operation for setting
6085 * the flag as we are still initializing the pages.
6087 __SetPageReserved(page);
6090 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6091 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6092 * ever freed or placed on a driver-private list.
6094 page->pgmap = pgmap;
6095 page->zone_device_data = NULL;
6098 * Mark the block movable so that blocks are reserved for
6099 * movable at startup. This will force kernel allocations
6100 * to reserve their blocks rather than leaking throughout
6101 * the address space during boot when many long-lived
6102 * kernel allocations are made.
6104 * bitmap is created for zone's valid pfn range. but memmap
6105 * can be created for invalid pages (for alignment)
6106 * check here not to call set_pageblock_migratetype() against
6109 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
6110 * because this is done early in section_activate()
6112 if (!(pfn & (pageblock_nr_pages - 1))) {
6113 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6118 pr_info("%s initialised %lu pages in %ums\n", __func__,
6119 nr_pages, jiffies_to_msecs(jiffies - start));
6123 static void __meminit zone_init_free_lists(struct zone *zone)
6125 unsigned int order, t;
6126 for_each_migratetype_order(order, t) {
6127 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6128 zone->free_area[order].nr_free = 0;
6132 void __meminit __weak memmap_init(unsigned long size, int nid,
6133 unsigned long zone, unsigned long start_pfn)
6135 memmap_init_zone(size, nid, zone, start_pfn, MEMMAP_EARLY, NULL);
6138 static int zone_batchsize(struct zone *zone)
6144 * The per-cpu-pages pools are set to around 1000th of the
6147 batch = zone_managed_pages(zone) / 1024;
6148 /* But no more than a meg. */
6149 if (batch * PAGE_SIZE > 1024 * 1024)
6150 batch = (1024 * 1024) / PAGE_SIZE;
6151 batch /= 4; /* We effectively *= 4 below */
6156 * Clamp the batch to a 2^n - 1 value. Having a power
6157 * of 2 value was found to be more likely to have
6158 * suboptimal cache aliasing properties in some cases.
6160 * For example if 2 tasks are alternately allocating
6161 * batches of pages, one task can end up with a lot
6162 * of pages of one half of the possible page colors
6163 * and the other with pages of the other colors.
6165 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6170 /* The deferral and batching of frees should be suppressed under NOMMU
6173 * The problem is that NOMMU needs to be able to allocate large chunks
6174 * of contiguous memory as there's no hardware page translation to
6175 * assemble apparent contiguous memory from discontiguous pages.
6177 * Queueing large contiguous runs of pages for batching, however,
6178 * causes the pages to actually be freed in smaller chunks. As there
6179 * can be a significant delay between the individual batches being
6180 * recycled, this leads to the once large chunks of space being
6181 * fragmented and becoming unavailable for high-order allocations.
6188 * pcp->high and pcp->batch values are related and dependent on one another:
6189 * ->batch must never be higher then ->high.
6190 * The following function updates them in a safe manner without read side
6193 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6194 * those fields changing asynchronously (acording the the above rule).
6196 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6197 * outside of boot time (or some other assurance that no concurrent updaters
6200 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6201 unsigned long batch)
6203 /* start with a fail safe value for batch */
6207 /* Update high, then batch, in order */
6214 /* a companion to pageset_set_high() */
6215 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6217 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6220 static void pageset_init(struct per_cpu_pageset *p)
6222 struct per_cpu_pages *pcp;
6225 memset(p, 0, sizeof(*p));
6228 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6229 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6232 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6235 pageset_set_batch(p, batch);
6239 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6240 * to the value high for the pageset p.
6242 static void pageset_set_high(struct per_cpu_pageset *p,
6245 unsigned long batch = max(1UL, high / 4);
6246 if ((high / 4) > (PAGE_SHIFT * 8))
6247 batch = PAGE_SHIFT * 8;
6249 pageset_update(&p->pcp, high, batch);
6252 static void pageset_set_high_and_batch(struct zone *zone,
6253 struct per_cpu_pageset *pcp)
6255 if (percpu_pagelist_fraction)
6256 pageset_set_high(pcp,
6257 (zone_managed_pages(zone) /
6258 percpu_pagelist_fraction));
6260 pageset_set_batch(pcp, zone_batchsize(zone));
6263 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6265 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6268 pageset_set_high_and_batch(zone, pcp);
6271 void __meminit setup_zone_pageset(struct zone *zone)
6274 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6275 for_each_possible_cpu(cpu)
6276 zone_pageset_init(zone, cpu);
6280 * Allocate per cpu pagesets and initialize them.
6281 * Before this call only boot pagesets were available.
6283 void __init setup_per_cpu_pageset(void)
6285 struct pglist_data *pgdat;
6288 for_each_populated_zone(zone)
6289 setup_zone_pageset(zone);
6291 for_each_online_pgdat(pgdat)
6292 pgdat->per_cpu_nodestats =
6293 alloc_percpu(struct per_cpu_nodestat);
6296 static __meminit void zone_pcp_init(struct zone *zone)
6299 * per cpu subsystem is not up at this point. The following code
6300 * relies on the ability of the linker to provide the
6301 * offset of a (static) per cpu variable into the per cpu area.
6303 zone->pageset = &boot_pageset;
6305 if (populated_zone(zone))
6306 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6307 zone->name, zone->present_pages,
6308 zone_batchsize(zone));
6311 void __meminit init_currently_empty_zone(struct zone *zone,
6312 unsigned long zone_start_pfn,
6315 struct pglist_data *pgdat = zone->zone_pgdat;
6316 int zone_idx = zone_idx(zone) + 1;
6318 if (zone_idx > pgdat->nr_zones)
6319 pgdat->nr_zones = zone_idx;
6321 zone->zone_start_pfn = zone_start_pfn;
6323 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6324 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6326 (unsigned long)zone_idx(zone),
6327 zone_start_pfn, (zone_start_pfn + size));
6329 zone_init_free_lists(zone);
6330 zone->initialized = 1;
6334 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
6335 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
6336 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
6338 * If an architecture guarantees that all ranges registered contain no holes
6339 * and may be freed, this this function may be used instead of calling
6340 * memblock_free_early_nid() manually.
6342 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
6344 unsigned long start_pfn, end_pfn;
6347 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
6348 start_pfn = min(start_pfn, max_low_pfn);
6349 end_pfn = min(end_pfn, max_low_pfn);
6351 if (start_pfn < end_pfn)
6352 memblock_free_early_nid(PFN_PHYS(start_pfn),
6353 (end_pfn - start_pfn) << PAGE_SHIFT,
6359 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6360 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6362 * If an architecture guarantees that all ranges registered contain no holes and may
6363 * be freed, this function may be used instead of calling memory_present() manually.
6365 void __init sparse_memory_present_with_active_regions(int nid)
6367 unsigned long start_pfn, end_pfn;
6370 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
6371 memory_present(this_nid, start_pfn, end_pfn);
6375 * get_pfn_range_for_nid - Return the start and end page frames for a node
6376 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6377 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6378 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6380 * It returns the start and end page frame of a node based on information
6381 * provided by memblock_set_node(). If called for a node
6382 * with no available memory, a warning is printed and the start and end
6385 void __init get_pfn_range_for_nid(unsigned int nid,
6386 unsigned long *start_pfn, unsigned long *end_pfn)
6388 unsigned long this_start_pfn, this_end_pfn;
6394 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6395 *start_pfn = min(*start_pfn, this_start_pfn);
6396 *end_pfn = max(*end_pfn, this_end_pfn);
6399 if (*start_pfn == -1UL)
6404 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6405 * assumption is made that zones within a node are ordered in monotonic
6406 * increasing memory addresses so that the "highest" populated zone is used
6408 static void __init find_usable_zone_for_movable(void)
6411 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6412 if (zone_index == ZONE_MOVABLE)
6415 if (arch_zone_highest_possible_pfn[zone_index] >
6416 arch_zone_lowest_possible_pfn[zone_index])
6420 VM_BUG_ON(zone_index == -1);
6421 movable_zone = zone_index;
6425 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6426 * because it is sized independent of architecture. Unlike the other zones,
6427 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6428 * in each node depending on the size of each node and how evenly kernelcore
6429 * is distributed. This helper function adjusts the zone ranges
6430 * provided by the architecture for a given node by using the end of the
6431 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6432 * zones within a node are in order of monotonic increases memory addresses
6434 static void __init adjust_zone_range_for_zone_movable(int nid,
6435 unsigned long zone_type,
6436 unsigned long node_start_pfn,
6437 unsigned long node_end_pfn,
6438 unsigned long *zone_start_pfn,
6439 unsigned long *zone_end_pfn)
6441 /* Only adjust if ZONE_MOVABLE is on this node */
6442 if (zone_movable_pfn[nid]) {
6443 /* Size ZONE_MOVABLE */
6444 if (zone_type == ZONE_MOVABLE) {
6445 *zone_start_pfn = zone_movable_pfn[nid];
6446 *zone_end_pfn = min(node_end_pfn,
6447 arch_zone_highest_possible_pfn[movable_zone]);
6449 /* Adjust for ZONE_MOVABLE starting within this range */
6450 } else if (!mirrored_kernelcore &&
6451 *zone_start_pfn < zone_movable_pfn[nid] &&
6452 *zone_end_pfn > zone_movable_pfn[nid]) {
6453 *zone_end_pfn = zone_movable_pfn[nid];
6455 /* Check if this whole range is within ZONE_MOVABLE */
6456 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6457 *zone_start_pfn = *zone_end_pfn;
6462 * Return the number of pages a zone spans in a node, including holes
6463 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6465 static unsigned long __init zone_spanned_pages_in_node(int nid,
6466 unsigned long zone_type,
6467 unsigned long node_start_pfn,
6468 unsigned long node_end_pfn,
6469 unsigned long *zone_start_pfn,
6470 unsigned long *zone_end_pfn,
6471 unsigned long *ignored)
6473 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6474 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6475 /* When hotadd a new node from cpu_up(), the node should be empty */
6476 if (!node_start_pfn && !node_end_pfn)
6479 /* Get the start and end of the zone */
6480 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6481 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6482 adjust_zone_range_for_zone_movable(nid, zone_type,
6483 node_start_pfn, node_end_pfn,
6484 zone_start_pfn, zone_end_pfn);
6486 /* Check that this node has pages within the zone's required range */
6487 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6490 /* Move the zone boundaries inside the node if necessary */
6491 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6492 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6494 /* Return the spanned pages */
6495 return *zone_end_pfn - *zone_start_pfn;
6499 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6500 * then all holes in the requested range will be accounted for.
6502 unsigned long __init __absent_pages_in_range(int nid,
6503 unsigned long range_start_pfn,
6504 unsigned long range_end_pfn)
6506 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6507 unsigned long start_pfn, end_pfn;
6510 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6511 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6512 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6513 nr_absent -= end_pfn - start_pfn;
6519 * absent_pages_in_range - Return number of page frames in holes within a range
6520 * @start_pfn: The start PFN to start searching for holes
6521 * @end_pfn: The end PFN to stop searching for holes
6523 * Return: the number of pages frames in memory holes within a range.
6525 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6526 unsigned long end_pfn)
6528 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6531 /* Return the number of page frames in holes in a zone on a node */
6532 static unsigned long __init zone_absent_pages_in_node(int nid,
6533 unsigned long zone_type,
6534 unsigned long node_start_pfn,
6535 unsigned long node_end_pfn,
6536 unsigned long *ignored)
6538 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6539 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6540 unsigned long zone_start_pfn, zone_end_pfn;
6541 unsigned long nr_absent;
6543 /* When hotadd a new node from cpu_up(), the node should be empty */
6544 if (!node_start_pfn && !node_end_pfn)
6547 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6548 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6550 adjust_zone_range_for_zone_movable(nid, zone_type,
6551 node_start_pfn, node_end_pfn,
6552 &zone_start_pfn, &zone_end_pfn);
6553 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6556 * ZONE_MOVABLE handling.
6557 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6560 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6561 unsigned long start_pfn, end_pfn;
6562 struct memblock_region *r;
6564 for_each_memblock(memory, r) {
6565 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6566 zone_start_pfn, zone_end_pfn);
6567 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6568 zone_start_pfn, zone_end_pfn);
6570 if (zone_type == ZONE_MOVABLE &&
6571 memblock_is_mirror(r))
6572 nr_absent += end_pfn - start_pfn;
6574 if (zone_type == ZONE_NORMAL &&
6575 !memblock_is_mirror(r))
6576 nr_absent += end_pfn - start_pfn;
6583 static inline unsigned long __init compat_zone_spanned_pages_in_node(int nid,
6584 unsigned long zone_type,
6585 unsigned long node_start_pfn,
6586 unsigned long node_end_pfn,
6587 unsigned long *zone_start_pfn,
6588 unsigned long *zone_end_pfn,
6589 unsigned long *zones_size)
6593 *zone_start_pfn = node_start_pfn;
6594 for (zone = 0; zone < zone_type; zone++)
6595 *zone_start_pfn += zones_size[zone];
6597 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6599 return zones_size[zone_type];
6602 static inline unsigned long __init compat_zone_absent_pages_in_node(int nid,
6603 unsigned long zone_type,
6604 unsigned long node_start_pfn,
6605 unsigned long node_end_pfn,
6606 unsigned long *zholes_size)
6611 return zholes_size[zone_type];
6614 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6615 unsigned long node_start_pfn,
6616 unsigned long node_end_pfn,
6617 unsigned long *zones_size,
6618 unsigned long *zholes_size,
6621 unsigned long realtotalpages = 0, totalpages = 0;
6624 for (i = 0; i < MAX_NR_ZONES; i++) {
6625 struct zone *zone = pgdat->node_zones + i;
6626 unsigned long zone_start_pfn, zone_end_pfn;
6627 unsigned long spanned, absent;
6628 unsigned long size, real_size;
6631 spanned = compat_zone_spanned_pages_in_node(
6638 absent = compat_zone_absent_pages_in_node(
6644 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
6650 absent = zone_absent_pages_in_node(pgdat->node_id, i,
6657 real_size = size - absent;
6660 zone->zone_start_pfn = zone_start_pfn;
6662 zone->zone_start_pfn = 0;
6663 zone->spanned_pages = size;
6664 zone->present_pages = real_size;
6667 realtotalpages += real_size;
6670 pgdat->node_spanned_pages = totalpages;
6671 pgdat->node_present_pages = realtotalpages;
6672 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6676 #ifndef CONFIG_SPARSEMEM
6678 * Calculate the size of the zone->blockflags rounded to an unsigned long
6679 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6680 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6681 * round what is now in bits to nearest long in bits, then return it in
6684 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6686 unsigned long usemapsize;
6688 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6689 usemapsize = roundup(zonesize, pageblock_nr_pages);
6690 usemapsize = usemapsize >> pageblock_order;
6691 usemapsize *= NR_PAGEBLOCK_BITS;
6692 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6694 return usemapsize / 8;
6697 static void __ref setup_usemap(struct pglist_data *pgdat,
6699 unsigned long zone_start_pfn,
6700 unsigned long zonesize)
6702 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6703 zone->pageblock_flags = NULL;
6705 zone->pageblock_flags =
6706 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6708 if (!zone->pageblock_flags)
6709 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6710 usemapsize, zone->name, pgdat->node_id);
6714 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6715 unsigned long zone_start_pfn, unsigned long zonesize) {}
6716 #endif /* CONFIG_SPARSEMEM */
6718 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6720 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6721 void __init set_pageblock_order(void)
6725 /* Check that pageblock_nr_pages has not already been setup */
6726 if (pageblock_order)
6729 if (HPAGE_SHIFT > PAGE_SHIFT)
6730 order = HUGETLB_PAGE_ORDER;
6732 order = MAX_ORDER - 1;
6735 * Assume the largest contiguous order of interest is a huge page.
6736 * This value may be variable depending on boot parameters on IA64 and
6739 pageblock_order = order;
6741 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6744 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6745 * is unused as pageblock_order is set at compile-time. See
6746 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6749 void __init set_pageblock_order(void)
6753 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6755 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6756 unsigned long present_pages)
6758 unsigned long pages = spanned_pages;
6761 * Provide a more accurate estimation if there are holes within
6762 * the zone and SPARSEMEM is in use. If there are holes within the
6763 * zone, each populated memory region may cost us one or two extra
6764 * memmap pages due to alignment because memmap pages for each
6765 * populated regions may not be naturally aligned on page boundary.
6766 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6768 if (spanned_pages > present_pages + (present_pages >> 4) &&
6769 IS_ENABLED(CONFIG_SPARSEMEM))
6770 pages = present_pages;
6772 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6775 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6776 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6778 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6780 spin_lock_init(&ds_queue->split_queue_lock);
6781 INIT_LIST_HEAD(&ds_queue->split_queue);
6782 ds_queue->split_queue_len = 0;
6785 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6788 #ifdef CONFIG_COMPACTION
6789 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6791 init_waitqueue_head(&pgdat->kcompactd_wait);
6794 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6797 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6799 pgdat_resize_init(pgdat);
6801 pgdat_init_split_queue(pgdat);
6802 pgdat_init_kcompactd(pgdat);
6804 init_waitqueue_head(&pgdat->kswapd_wait);
6805 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6807 pgdat_page_ext_init(pgdat);
6808 spin_lock_init(&pgdat->lru_lock);
6809 lruvec_init(&pgdat->__lruvec);
6812 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6813 unsigned long remaining_pages)
6815 atomic_long_set(&zone->managed_pages, remaining_pages);
6816 zone_set_nid(zone, nid);
6817 zone->name = zone_names[idx];
6818 zone->zone_pgdat = NODE_DATA(nid);
6819 spin_lock_init(&zone->lock);
6820 zone_seqlock_init(zone);
6821 zone_pcp_init(zone);
6825 * Set up the zone data structures
6826 * - init pgdat internals
6827 * - init all zones belonging to this node
6829 * NOTE: this function is only called during memory hotplug
6831 #ifdef CONFIG_MEMORY_HOTPLUG
6832 void __ref free_area_init_core_hotplug(int nid)
6835 pg_data_t *pgdat = NODE_DATA(nid);
6837 pgdat_init_internals(pgdat);
6838 for (z = 0; z < MAX_NR_ZONES; z++)
6839 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6844 * Set up the zone data structures:
6845 * - mark all pages reserved
6846 * - mark all memory queues empty
6847 * - clear the memory bitmaps
6849 * NOTE: pgdat should get zeroed by caller.
6850 * NOTE: this function is only called during early init.
6852 static void __init free_area_init_core(struct pglist_data *pgdat)
6855 int nid = pgdat->node_id;
6857 pgdat_init_internals(pgdat);
6858 pgdat->per_cpu_nodestats = &boot_nodestats;
6860 for (j = 0; j < MAX_NR_ZONES; j++) {
6861 struct zone *zone = pgdat->node_zones + j;
6862 unsigned long size, freesize, memmap_pages;
6863 unsigned long zone_start_pfn = zone->zone_start_pfn;
6865 size = zone->spanned_pages;
6866 freesize = zone->present_pages;
6869 * Adjust freesize so that it accounts for how much memory
6870 * is used by this zone for memmap. This affects the watermark
6871 * and per-cpu initialisations
6873 memmap_pages = calc_memmap_size(size, freesize);
6874 if (!is_highmem_idx(j)) {
6875 if (freesize >= memmap_pages) {
6876 freesize -= memmap_pages;
6879 " %s zone: %lu pages used for memmap\n",
6880 zone_names[j], memmap_pages);
6882 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6883 zone_names[j], memmap_pages, freesize);
6886 /* Account for reserved pages */
6887 if (j == 0 && freesize > dma_reserve) {
6888 freesize -= dma_reserve;
6889 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6890 zone_names[0], dma_reserve);
6893 if (!is_highmem_idx(j))
6894 nr_kernel_pages += freesize;
6895 /* Charge for highmem memmap if there are enough kernel pages */
6896 else if (nr_kernel_pages > memmap_pages * 2)
6897 nr_kernel_pages -= memmap_pages;
6898 nr_all_pages += freesize;
6901 * Set an approximate value for lowmem here, it will be adjusted
6902 * when the bootmem allocator frees pages into the buddy system.
6903 * And all highmem pages will be managed by the buddy system.
6905 zone_init_internals(zone, j, nid, freesize);
6910 set_pageblock_order();
6911 setup_usemap(pgdat, zone, zone_start_pfn, size);
6912 init_currently_empty_zone(zone, zone_start_pfn, size);
6913 memmap_init(size, nid, j, zone_start_pfn);
6917 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6918 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6920 unsigned long __maybe_unused start = 0;
6921 unsigned long __maybe_unused offset = 0;
6923 /* Skip empty nodes */
6924 if (!pgdat->node_spanned_pages)
6927 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6928 offset = pgdat->node_start_pfn - start;
6929 /* ia64 gets its own node_mem_map, before this, without bootmem */
6930 if (!pgdat->node_mem_map) {
6931 unsigned long size, end;
6935 * The zone's endpoints aren't required to be MAX_ORDER
6936 * aligned but the node_mem_map endpoints must be in order
6937 * for the buddy allocator to function correctly.
6939 end = pgdat_end_pfn(pgdat);
6940 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6941 size = (end - start) * sizeof(struct page);
6942 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6945 panic("Failed to allocate %ld bytes for node %d memory map\n",
6946 size, pgdat->node_id);
6947 pgdat->node_mem_map = map + offset;
6949 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6950 __func__, pgdat->node_id, (unsigned long)pgdat,
6951 (unsigned long)pgdat->node_mem_map);
6952 #ifndef CONFIG_NEED_MULTIPLE_NODES
6954 * With no DISCONTIG, the global mem_map is just set as node 0's
6956 if (pgdat == NODE_DATA(0)) {
6957 mem_map = NODE_DATA(0)->node_mem_map;
6958 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6964 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6965 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6967 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6968 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6970 pgdat->first_deferred_pfn = ULONG_MAX;
6973 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6976 static void __init __free_area_init_node(int nid, unsigned long *zones_size,
6977 unsigned long node_start_pfn,
6978 unsigned long *zholes_size,
6981 pg_data_t *pgdat = NODE_DATA(nid);
6982 unsigned long start_pfn = 0;
6983 unsigned long end_pfn = 0;
6985 /* pg_data_t should be reset to zero when it's allocated */
6986 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6988 pgdat->node_id = nid;
6989 pgdat->node_start_pfn = node_start_pfn;
6990 pgdat->per_cpu_nodestats = NULL;
6992 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6993 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6994 (u64)start_pfn << PAGE_SHIFT,
6995 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6997 start_pfn = node_start_pfn;
6999 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
7000 zones_size, zholes_size, compat);
7002 alloc_node_mem_map(pgdat);
7003 pgdat_set_deferred_range(pgdat);
7005 free_area_init_core(pgdat);
7008 void __init free_area_init_node(int nid, unsigned long *zones_size,
7009 unsigned long node_start_pfn,
7010 unsigned long *zholes_size)
7012 __free_area_init_node(nid, zones_size, node_start_pfn, zholes_size,
7016 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
7018 * Initialize all valid struct pages in the range [spfn, epfn) and mark them
7019 * PageReserved(). Return the number of struct pages that were initialized.
7021 static u64 __init init_unavailable_range(unsigned long spfn, unsigned long epfn)
7026 for (pfn = spfn; pfn < epfn; pfn++) {
7027 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
7028 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
7029 + pageblock_nr_pages - 1;
7033 * Use a fake node/zone (0) for now. Some of these pages
7034 * (in memblock.reserved but not in memblock.memory) will
7035 * get re-initialized via reserve_bootmem_region() later.
7037 __init_single_page(pfn_to_page(pfn), pfn, 0, 0);
7038 __SetPageReserved(pfn_to_page(pfn));
7046 * Only struct pages that are backed by physical memory are zeroed and
7047 * initialized by going through __init_single_page(). But, there are some
7048 * struct pages which are reserved in memblock allocator and their fields
7049 * may be accessed (for example page_to_pfn() on some configuration accesses
7050 * flags). We must explicitly initialize those struct pages.
7052 * This function also addresses a similar issue where struct pages are left
7053 * uninitialized because the physical address range is not covered by
7054 * memblock.memory or memblock.reserved. That could happen when memblock
7055 * layout is manually configured via memmap=, or when the highest physical
7056 * address (max_pfn) does not end on a section boundary.
7058 static void __init init_unavailable_mem(void)
7060 phys_addr_t start, end;
7062 phys_addr_t next = 0;
7065 * Loop through unavailable ranges not covered by memblock.memory.
7068 for_each_mem_range(i, &memblock.memory, NULL,
7069 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
7071 pgcnt += init_unavailable_range(PFN_DOWN(next),
7077 * Early sections always have a fully populated memmap for the whole
7078 * section - see pfn_valid(). If the last section has holes at the
7079 * end and that section is marked "online", the memmap will be
7080 * considered initialized. Make sure that memmap has a well defined
7083 pgcnt += init_unavailable_range(PFN_DOWN(next),
7084 round_up(max_pfn, PAGES_PER_SECTION));
7087 * Struct pages that do not have backing memory. This could be because
7088 * firmware is using some of this memory, or for some other reasons.
7091 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
7094 static inline void __init init_unavailable_mem(void)
7097 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
7099 #if MAX_NUMNODES > 1
7101 * Figure out the number of possible node ids.
7103 void __init setup_nr_node_ids(void)
7105 unsigned int highest;
7107 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7108 nr_node_ids = highest + 1;
7113 * node_map_pfn_alignment - determine the maximum internode alignment
7115 * This function should be called after node map is populated and sorted.
7116 * It calculates the maximum power of two alignment which can distinguish
7119 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7120 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7121 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7122 * shifted, 1GiB is enough and this function will indicate so.
7124 * This is used to test whether pfn -> nid mapping of the chosen memory
7125 * model has fine enough granularity to avoid incorrect mapping for the
7126 * populated node map.
7128 * Return: the determined alignment in pfn's. 0 if there is no alignment
7129 * requirement (single node).
7131 unsigned long __init node_map_pfn_alignment(void)
7133 unsigned long accl_mask = 0, last_end = 0;
7134 unsigned long start, end, mask;
7135 int last_nid = NUMA_NO_NODE;
7138 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7139 if (!start || last_nid < 0 || last_nid == nid) {
7146 * Start with a mask granular enough to pin-point to the
7147 * start pfn and tick off bits one-by-one until it becomes
7148 * too coarse to separate the current node from the last.
7150 mask = ~((1 << __ffs(start)) - 1);
7151 while (mask && last_end <= (start & (mask << 1)))
7154 /* accumulate all internode masks */
7158 /* convert mask to number of pages */
7159 return ~accl_mask + 1;
7162 /* Find the lowest pfn for a node */
7163 static unsigned long __init find_min_pfn_for_node(int nid)
7165 unsigned long min_pfn = ULONG_MAX;
7166 unsigned long start_pfn;
7169 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
7170 min_pfn = min(min_pfn, start_pfn);
7172 if (min_pfn == ULONG_MAX) {
7173 pr_warn("Could not find start_pfn for node %d\n", nid);
7181 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7183 * Return: the minimum PFN based on information provided via
7184 * memblock_set_node().
7186 unsigned long __init find_min_pfn_with_active_regions(void)
7188 return find_min_pfn_for_node(MAX_NUMNODES);
7192 * early_calculate_totalpages()
7193 * Sum pages in active regions for movable zone.
7194 * Populate N_MEMORY for calculating usable_nodes.
7196 static unsigned long __init early_calculate_totalpages(void)
7198 unsigned long totalpages = 0;
7199 unsigned long start_pfn, end_pfn;
7202 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7203 unsigned long pages = end_pfn - start_pfn;
7205 totalpages += pages;
7207 node_set_state(nid, N_MEMORY);
7213 * Find the PFN the Movable zone begins in each node. Kernel memory
7214 * is spread evenly between nodes as long as the nodes have enough
7215 * memory. When they don't, some nodes will have more kernelcore than
7218 static void __init find_zone_movable_pfns_for_nodes(void)
7221 unsigned long usable_startpfn;
7222 unsigned long kernelcore_node, kernelcore_remaining;
7223 /* save the state before borrow the nodemask */
7224 nodemask_t saved_node_state = node_states[N_MEMORY];
7225 unsigned long totalpages = early_calculate_totalpages();
7226 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7227 struct memblock_region *r;
7229 /* Need to find movable_zone earlier when movable_node is specified. */
7230 find_usable_zone_for_movable();
7233 * If movable_node is specified, ignore kernelcore and movablecore
7236 if (movable_node_is_enabled()) {
7237 for_each_memblock(memory, r) {
7238 if (!memblock_is_hotpluggable(r))
7241 nid = memblock_get_region_node(r);
7243 usable_startpfn = PFN_DOWN(r->base);
7244 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7245 min(usable_startpfn, zone_movable_pfn[nid]) :
7253 * If kernelcore=mirror is specified, ignore movablecore option
7255 if (mirrored_kernelcore) {
7256 bool mem_below_4gb_not_mirrored = false;
7258 for_each_memblock(memory, r) {
7259 if (memblock_is_mirror(r))
7262 nid = memblock_get_region_node(r);
7264 usable_startpfn = memblock_region_memory_base_pfn(r);
7266 if (usable_startpfn < 0x100000) {
7267 mem_below_4gb_not_mirrored = true;
7271 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7272 min(usable_startpfn, zone_movable_pfn[nid]) :
7276 if (mem_below_4gb_not_mirrored)
7277 pr_warn("This configuration results in unmirrored kernel memory.");
7283 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7284 * amount of necessary memory.
7286 if (required_kernelcore_percent)
7287 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7289 if (required_movablecore_percent)
7290 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7294 * If movablecore= was specified, calculate what size of
7295 * kernelcore that corresponds so that memory usable for
7296 * any allocation type is evenly spread. If both kernelcore
7297 * and movablecore are specified, then the value of kernelcore
7298 * will be used for required_kernelcore if it's greater than
7299 * what movablecore would have allowed.
7301 if (required_movablecore) {
7302 unsigned long corepages;
7305 * Round-up so that ZONE_MOVABLE is at least as large as what
7306 * was requested by the user
7308 required_movablecore =
7309 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7310 required_movablecore = min(totalpages, required_movablecore);
7311 corepages = totalpages - required_movablecore;
7313 required_kernelcore = max(required_kernelcore, corepages);
7317 * If kernelcore was not specified or kernelcore size is larger
7318 * than totalpages, there is no ZONE_MOVABLE.
7320 if (!required_kernelcore || required_kernelcore >= totalpages)
7323 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7324 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7327 /* Spread kernelcore memory as evenly as possible throughout nodes */
7328 kernelcore_node = required_kernelcore / usable_nodes;
7329 for_each_node_state(nid, N_MEMORY) {
7330 unsigned long start_pfn, end_pfn;
7333 * Recalculate kernelcore_node if the division per node
7334 * now exceeds what is necessary to satisfy the requested
7335 * amount of memory for the kernel
7337 if (required_kernelcore < kernelcore_node)
7338 kernelcore_node = required_kernelcore / usable_nodes;
7341 * As the map is walked, we track how much memory is usable
7342 * by the kernel using kernelcore_remaining. When it is
7343 * 0, the rest of the node is usable by ZONE_MOVABLE
7345 kernelcore_remaining = kernelcore_node;
7347 /* Go through each range of PFNs within this node */
7348 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7349 unsigned long size_pages;
7351 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7352 if (start_pfn >= end_pfn)
7355 /* Account for what is only usable for kernelcore */
7356 if (start_pfn < usable_startpfn) {
7357 unsigned long kernel_pages;
7358 kernel_pages = min(end_pfn, usable_startpfn)
7361 kernelcore_remaining -= min(kernel_pages,
7362 kernelcore_remaining);
7363 required_kernelcore -= min(kernel_pages,
7364 required_kernelcore);
7366 /* Continue if range is now fully accounted */
7367 if (end_pfn <= usable_startpfn) {
7370 * Push zone_movable_pfn to the end so
7371 * that if we have to rebalance
7372 * kernelcore across nodes, we will
7373 * not double account here
7375 zone_movable_pfn[nid] = end_pfn;
7378 start_pfn = usable_startpfn;
7382 * The usable PFN range for ZONE_MOVABLE is from
7383 * start_pfn->end_pfn. Calculate size_pages as the
7384 * number of pages used as kernelcore
7386 size_pages = end_pfn - start_pfn;
7387 if (size_pages > kernelcore_remaining)
7388 size_pages = kernelcore_remaining;
7389 zone_movable_pfn[nid] = start_pfn + size_pages;
7392 * Some kernelcore has been met, update counts and
7393 * break if the kernelcore for this node has been
7396 required_kernelcore -= min(required_kernelcore,
7398 kernelcore_remaining -= size_pages;
7399 if (!kernelcore_remaining)
7405 * If there is still required_kernelcore, we do another pass with one
7406 * less node in the count. This will push zone_movable_pfn[nid] further
7407 * along on the nodes that still have memory until kernelcore is
7411 if (usable_nodes && required_kernelcore > usable_nodes)
7415 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7416 for (nid = 0; nid < MAX_NUMNODES; nid++)
7417 zone_movable_pfn[nid] =
7418 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7421 /* restore the node_state */
7422 node_states[N_MEMORY] = saved_node_state;
7425 /* Any regular or high memory on that node ? */
7426 static void check_for_memory(pg_data_t *pgdat, int nid)
7428 enum zone_type zone_type;
7430 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7431 struct zone *zone = &pgdat->node_zones[zone_type];
7432 if (populated_zone(zone)) {
7433 if (IS_ENABLED(CONFIG_HIGHMEM))
7434 node_set_state(nid, N_HIGH_MEMORY);
7435 if (zone_type <= ZONE_NORMAL)
7436 node_set_state(nid, N_NORMAL_MEMORY);
7443 * free_area_init_nodes - Initialise all pg_data_t and zone data
7444 * @max_zone_pfn: an array of max PFNs for each zone
7446 * This will call free_area_init_node() for each active node in the system.
7447 * Using the page ranges provided by memblock_set_node(), the size of each
7448 * zone in each node and their holes is calculated. If the maximum PFN
7449 * between two adjacent zones match, it is assumed that the zone is empty.
7450 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7451 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7452 * starts where the previous one ended. For example, ZONE_DMA32 starts
7453 * at arch_max_dma_pfn.
7455 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
7457 unsigned long start_pfn, end_pfn;
7460 /* Record where the zone boundaries are */
7461 memset(arch_zone_lowest_possible_pfn, 0,
7462 sizeof(arch_zone_lowest_possible_pfn));
7463 memset(arch_zone_highest_possible_pfn, 0,
7464 sizeof(arch_zone_highest_possible_pfn));
7466 start_pfn = find_min_pfn_with_active_regions();
7468 for (i = 0; i < MAX_NR_ZONES; i++) {
7469 if (i == ZONE_MOVABLE)
7472 end_pfn = max(max_zone_pfn[i], start_pfn);
7473 arch_zone_lowest_possible_pfn[i] = start_pfn;
7474 arch_zone_highest_possible_pfn[i] = end_pfn;
7476 start_pfn = end_pfn;
7479 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7480 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7481 find_zone_movable_pfns_for_nodes();
7483 /* Print out the zone ranges */
7484 pr_info("Zone ranges:\n");
7485 for (i = 0; i < MAX_NR_ZONES; i++) {
7486 if (i == ZONE_MOVABLE)
7488 pr_info(" %-8s ", zone_names[i]);
7489 if (arch_zone_lowest_possible_pfn[i] ==
7490 arch_zone_highest_possible_pfn[i])
7493 pr_cont("[mem %#018Lx-%#018Lx]\n",
7494 (u64)arch_zone_lowest_possible_pfn[i]
7496 ((u64)arch_zone_highest_possible_pfn[i]
7497 << PAGE_SHIFT) - 1);
7500 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7501 pr_info("Movable zone start for each node\n");
7502 for (i = 0; i < MAX_NUMNODES; i++) {
7503 if (zone_movable_pfn[i])
7504 pr_info(" Node %d: %#018Lx\n", i,
7505 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7509 * Print out the early node map, and initialize the
7510 * subsection-map relative to active online memory ranges to
7511 * enable future "sub-section" extensions of the memory map.
7513 pr_info("Early memory node ranges\n");
7514 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7515 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7516 (u64)start_pfn << PAGE_SHIFT,
7517 ((u64)end_pfn << PAGE_SHIFT) - 1);
7518 subsection_map_init(start_pfn, end_pfn - start_pfn);
7521 /* Initialise every node */
7522 mminit_verify_pageflags_layout();
7523 setup_nr_node_ids();
7524 init_unavailable_mem();
7525 for_each_online_node(nid) {
7526 pg_data_t *pgdat = NODE_DATA(nid);
7527 __free_area_init_node(nid, NULL,
7528 find_min_pfn_for_node(nid), NULL, false);
7530 /* Any memory on that node */
7531 if (pgdat->node_present_pages)
7532 node_set_state(nid, N_MEMORY);
7533 check_for_memory(pgdat, nid);
7537 static int __init cmdline_parse_core(char *p, unsigned long *core,
7538 unsigned long *percent)
7540 unsigned long long coremem;
7546 /* Value may be a percentage of total memory, otherwise bytes */
7547 coremem = simple_strtoull(p, &endptr, 0);
7548 if (*endptr == '%') {
7549 /* Paranoid check for percent values greater than 100 */
7550 WARN_ON(coremem > 100);
7554 coremem = memparse(p, &p);
7555 /* Paranoid check that UL is enough for the coremem value */
7556 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7558 *core = coremem >> PAGE_SHIFT;
7565 * kernelcore=size sets the amount of memory for use for allocations that
7566 * cannot be reclaimed or migrated.
7568 static int __init cmdline_parse_kernelcore(char *p)
7570 /* parse kernelcore=mirror */
7571 if (parse_option_str(p, "mirror")) {
7572 mirrored_kernelcore = true;
7576 return cmdline_parse_core(p, &required_kernelcore,
7577 &required_kernelcore_percent);
7581 * movablecore=size sets the amount of memory for use for allocations that
7582 * can be reclaimed or migrated.
7584 static int __init cmdline_parse_movablecore(char *p)
7586 return cmdline_parse_core(p, &required_movablecore,
7587 &required_movablecore_percent);
7590 early_param("kernelcore", cmdline_parse_kernelcore);
7591 early_param("movablecore", cmdline_parse_movablecore);
7593 void adjust_managed_page_count(struct page *page, long count)
7595 atomic_long_add(count, &page_zone(page)->managed_pages);
7596 totalram_pages_add(count);
7597 #ifdef CONFIG_HIGHMEM
7598 if (PageHighMem(page))
7599 totalhigh_pages_add(count);
7602 EXPORT_SYMBOL(adjust_managed_page_count);
7604 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7607 unsigned long pages = 0;
7609 start = (void *)PAGE_ALIGN((unsigned long)start);
7610 end = (void *)((unsigned long)end & PAGE_MASK);
7611 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7612 struct page *page = virt_to_page(pos);
7613 void *direct_map_addr;
7616 * 'direct_map_addr' might be different from 'pos'
7617 * because some architectures' virt_to_page()
7618 * work with aliases. Getting the direct map
7619 * address ensures that we get a _writeable_
7620 * alias for the memset().
7622 direct_map_addr = page_address(page);
7623 if ((unsigned int)poison <= 0xFF)
7624 memset(direct_map_addr, poison, PAGE_SIZE);
7626 free_reserved_page(page);
7630 pr_info("Freeing %s memory: %ldK\n",
7631 s, pages << (PAGE_SHIFT - 10));
7636 #ifdef CONFIG_HIGHMEM
7637 void free_highmem_page(struct page *page)
7639 __free_reserved_page(page);
7640 totalram_pages_inc();
7641 atomic_long_inc(&page_zone(page)->managed_pages);
7642 totalhigh_pages_inc();
7647 void __init mem_init_print_info(const char *str)
7649 unsigned long physpages, codesize, datasize, rosize, bss_size;
7650 unsigned long init_code_size, init_data_size;
7652 physpages = get_num_physpages();
7653 codesize = _etext - _stext;
7654 datasize = _edata - _sdata;
7655 rosize = __end_rodata - __start_rodata;
7656 bss_size = __bss_stop - __bss_start;
7657 init_data_size = __init_end - __init_begin;
7658 init_code_size = _einittext - _sinittext;
7661 * Detect special cases and adjust section sizes accordingly:
7662 * 1) .init.* may be embedded into .data sections
7663 * 2) .init.text.* may be out of [__init_begin, __init_end],
7664 * please refer to arch/tile/kernel/vmlinux.lds.S.
7665 * 3) .rodata.* may be embedded into .text or .data sections.
7667 #define adj_init_size(start, end, size, pos, adj) \
7669 if (start <= pos && pos < end && size > adj) \
7673 adj_init_size(__init_begin, __init_end, init_data_size,
7674 _sinittext, init_code_size);
7675 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7676 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7677 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7678 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7680 #undef adj_init_size
7682 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7683 #ifdef CONFIG_HIGHMEM
7687 nr_free_pages() << (PAGE_SHIFT - 10),
7688 physpages << (PAGE_SHIFT - 10),
7689 codesize >> 10, datasize >> 10, rosize >> 10,
7690 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7691 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7692 totalcma_pages << (PAGE_SHIFT - 10),
7693 #ifdef CONFIG_HIGHMEM
7694 totalhigh_pages() << (PAGE_SHIFT - 10),
7696 str ? ", " : "", str ? str : "");
7700 * set_dma_reserve - set the specified number of pages reserved in the first zone
7701 * @new_dma_reserve: The number of pages to mark reserved
7703 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7704 * In the DMA zone, a significant percentage may be consumed by kernel image
7705 * and other unfreeable allocations which can skew the watermarks badly. This
7706 * function may optionally be used to account for unfreeable pages in the
7707 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7708 * smaller per-cpu batchsize.
7710 void __init set_dma_reserve(unsigned long new_dma_reserve)
7712 dma_reserve = new_dma_reserve;
7715 void __init free_area_init(unsigned long *max_zone_pfn)
7717 init_unavailable_mem();
7718 free_area_init_nodes(max_zone_pfn);
7721 static int page_alloc_cpu_dead(unsigned int cpu)
7724 lru_add_drain_cpu(cpu);
7728 * Spill the event counters of the dead processor
7729 * into the current processors event counters.
7730 * This artificially elevates the count of the current
7733 vm_events_fold_cpu(cpu);
7736 * Zero the differential counters of the dead processor
7737 * so that the vm statistics are consistent.
7739 * This is only okay since the processor is dead and cannot
7740 * race with what we are doing.
7742 cpu_vm_stats_fold(cpu);
7747 int hashdist = HASHDIST_DEFAULT;
7749 static int __init set_hashdist(char *str)
7753 hashdist = simple_strtoul(str, &str, 0);
7756 __setup("hashdist=", set_hashdist);
7759 void __init page_alloc_init(void)
7764 if (num_node_state(N_MEMORY) == 1)
7768 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7769 "mm/page_alloc:dead", NULL,
7770 page_alloc_cpu_dead);
7775 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7776 * or min_free_kbytes changes.
7778 static void calculate_totalreserve_pages(void)
7780 struct pglist_data *pgdat;
7781 unsigned long reserve_pages = 0;
7782 enum zone_type i, j;
7784 for_each_online_pgdat(pgdat) {
7786 pgdat->totalreserve_pages = 0;
7788 for (i = 0; i < MAX_NR_ZONES; i++) {
7789 struct zone *zone = pgdat->node_zones + i;
7791 unsigned long managed_pages = zone_managed_pages(zone);
7793 /* Find valid and maximum lowmem_reserve in the zone */
7794 for (j = i; j < MAX_NR_ZONES; j++) {
7795 if (zone->lowmem_reserve[j] > max)
7796 max = zone->lowmem_reserve[j];
7799 /* we treat the high watermark as reserved pages. */
7800 max += high_wmark_pages(zone);
7802 if (max > managed_pages)
7803 max = managed_pages;
7805 pgdat->totalreserve_pages += max;
7807 reserve_pages += max;
7810 totalreserve_pages = reserve_pages;
7814 * setup_per_zone_lowmem_reserve - called whenever
7815 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7816 * has a correct pages reserved value, so an adequate number of
7817 * pages are left in the zone after a successful __alloc_pages().
7819 static void setup_per_zone_lowmem_reserve(void)
7821 struct pglist_data *pgdat;
7822 enum zone_type j, idx;
7824 for_each_online_pgdat(pgdat) {
7825 for (j = 0; j < MAX_NR_ZONES; j++) {
7826 struct zone *zone = pgdat->node_zones + j;
7827 unsigned long managed_pages = zone_managed_pages(zone);
7829 zone->lowmem_reserve[j] = 0;
7833 struct zone *lower_zone;
7836 lower_zone = pgdat->node_zones + idx;
7838 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7839 sysctl_lowmem_reserve_ratio[idx] = 0;
7840 lower_zone->lowmem_reserve[j] = 0;
7842 lower_zone->lowmem_reserve[j] =
7843 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7845 managed_pages += zone_managed_pages(lower_zone);
7850 /* update totalreserve_pages */
7851 calculate_totalreserve_pages();
7854 static void __setup_per_zone_wmarks(void)
7856 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7857 unsigned long lowmem_pages = 0;
7859 unsigned long flags;
7861 /* Calculate total number of !ZONE_HIGHMEM pages */
7862 for_each_zone(zone) {
7863 if (!is_highmem(zone))
7864 lowmem_pages += zone_managed_pages(zone);
7867 for_each_zone(zone) {
7870 spin_lock_irqsave(&zone->lock, flags);
7871 tmp = (u64)pages_min * zone_managed_pages(zone);
7872 do_div(tmp, lowmem_pages);
7873 if (is_highmem(zone)) {
7875 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7876 * need highmem pages, so cap pages_min to a small
7879 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7880 * deltas control async page reclaim, and so should
7881 * not be capped for highmem.
7883 unsigned long min_pages;
7885 min_pages = zone_managed_pages(zone) / 1024;
7886 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7887 zone->_watermark[WMARK_MIN] = min_pages;
7890 * If it's a lowmem zone, reserve a number of pages
7891 * proportionate to the zone's size.
7893 zone->_watermark[WMARK_MIN] = tmp;
7897 * Set the kswapd watermarks distance according to the
7898 * scale factor in proportion to available memory, but
7899 * ensure a minimum size on small systems.
7901 tmp = max_t(u64, tmp >> 2,
7902 mult_frac(zone_managed_pages(zone),
7903 watermark_scale_factor, 10000));
7905 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7906 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7907 zone->watermark_boost = 0;
7909 spin_unlock_irqrestore(&zone->lock, flags);
7912 /* update totalreserve_pages */
7913 calculate_totalreserve_pages();
7917 * setup_per_zone_wmarks - called when min_free_kbytes changes
7918 * or when memory is hot-{added|removed}
7920 * Ensures that the watermark[min,low,high] values for each zone are set
7921 * correctly with respect to min_free_kbytes.
7923 void setup_per_zone_wmarks(void)
7925 static DEFINE_SPINLOCK(lock);
7928 __setup_per_zone_wmarks();
7933 * Initialise min_free_kbytes.
7935 * For small machines we want it small (128k min). For large machines
7936 * we want it large (64MB max). But it is not linear, because network
7937 * bandwidth does not increase linearly with machine size. We use
7939 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7940 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7956 int __meminit init_per_zone_wmark_min(void)
7958 unsigned long lowmem_kbytes;
7959 int new_min_free_kbytes;
7961 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7962 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7964 if (new_min_free_kbytes > user_min_free_kbytes) {
7965 min_free_kbytes = new_min_free_kbytes;
7966 if (min_free_kbytes < 128)
7967 min_free_kbytes = 128;
7968 if (min_free_kbytes > 262144)
7969 min_free_kbytes = 262144;
7971 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7972 new_min_free_kbytes, user_min_free_kbytes);
7974 setup_per_zone_wmarks();
7975 refresh_zone_stat_thresholds();
7976 setup_per_zone_lowmem_reserve();
7979 setup_min_unmapped_ratio();
7980 setup_min_slab_ratio();
7985 core_initcall(init_per_zone_wmark_min)
7988 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7989 * that we can call two helper functions whenever min_free_kbytes
7992 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7993 void __user *buffer, size_t *length, loff_t *ppos)
7997 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8002 user_min_free_kbytes = min_free_kbytes;
8003 setup_per_zone_wmarks();
8008 int watermark_boost_factor_sysctl_handler(struct ctl_table *table, int write,
8009 void __user *buffer, size_t *length, loff_t *ppos)
8013 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8020 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8021 void __user *buffer, size_t *length, loff_t *ppos)
8025 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8030 setup_per_zone_wmarks();
8036 static void setup_min_unmapped_ratio(void)
8041 for_each_online_pgdat(pgdat)
8042 pgdat->min_unmapped_pages = 0;
8045 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8046 sysctl_min_unmapped_ratio) / 100;
8050 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8051 void __user *buffer, size_t *length, loff_t *ppos)
8055 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8059 setup_min_unmapped_ratio();
8064 static void setup_min_slab_ratio(void)
8069 for_each_online_pgdat(pgdat)
8070 pgdat->min_slab_pages = 0;
8073 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8074 sysctl_min_slab_ratio) / 100;
8077 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8078 void __user *buffer, size_t *length, loff_t *ppos)
8082 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8086 setup_min_slab_ratio();
8093 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8094 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8095 * whenever sysctl_lowmem_reserve_ratio changes.
8097 * The reserve ratio obviously has absolutely no relation with the
8098 * minimum watermarks. The lowmem reserve ratio can only make sense
8099 * if in function of the boot time zone sizes.
8101 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8102 void __user *buffer, size_t *length, loff_t *ppos)
8104 proc_dointvec_minmax(table, write, buffer, length, ppos);
8105 setup_per_zone_lowmem_reserve();
8109 static void __zone_pcp_update(struct zone *zone)
8113 for_each_possible_cpu(cpu)
8114 pageset_set_high_and_batch(zone,
8115 per_cpu_ptr(zone->pageset, cpu));
8119 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8120 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8121 * pagelist can have before it gets flushed back to buddy allocator.
8123 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8124 void __user *buffer, size_t *length, loff_t *ppos)
8127 int old_percpu_pagelist_fraction;
8130 mutex_lock(&pcp_batch_high_lock);
8131 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8133 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8134 if (!write || ret < 0)
8137 /* Sanity checking to avoid pcp imbalance */
8138 if (percpu_pagelist_fraction &&
8139 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8140 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8146 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8149 for_each_populated_zone(zone)
8150 __zone_pcp_update(zone);
8152 mutex_unlock(&pcp_batch_high_lock);
8156 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8158 * Returns the number of pages that arch has reserved but
8159 * is not known to alloc_large_system_hash().
8161 static unsigned long __init arch_reserved_kernel_pages(void)
8168 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8169 * machines. As memory size is increased the scale is also increased but at
8170 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8171 * quadruples the scale is increased by one, which means the size of hash table
8172 * only doubles, instead of quadrupling as well.
8173 * Because 32-bit systems cannot have large physical memory, where this scaling
8174 * makes sense, it is disabled on such platforms.
8176 #if __BITS_PER_LONG > 32
8177 #define ADAPT_SCALE_BASE (64ul << 30)
8178 #define ADAPT_SCALE_SHIFT 2
8179 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8183 * allocate a large system hash table from bootmem
8184 * - it is assumed that the hash table must contain an exact power-of-2
8185 * quantity of entries
8186 * - limit is the number of hash buckets, not the total allocation size
8188 void *__init alloc_large_system_hash(const char *tablename,
8189 unsigned long bucketsize,
8190 unsigned long numentries,
8193 unsigned int *_hash_shift,
8194 unsigned int *_hash_mask,
8195 unsigned long low_limit,
8196 unsigned long high_limit)
8198 unsigned long long max = high_limit;
8199 unsigned long log2qty, size;
8204 /* allow the kernel cmdline to have a say */
8206 /* round applicable memory size up to nearest megabyte */
8207 numentries = nr_kernel_pages;
8208 numentries -= arch_reserved_kernel_pages();
8210 /* It isn't necessary when PAGE_SIZE >= 1MB */
8211 if (PAGE_SHIFT < 20)
8212 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8214 #if __BITS_PER_LONG > 32
8216 unsigned long adapt;
8218 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8219 adapt <<= ADAPT_SCALE_SHIFT)
8224 /* limit to 1 bucket per 2^scale bytes of low memory */
8225 if (scale > PAGE_SHIFT)
8226 numentries >>= (scale - PAGE_SHIFT);
8228 numentries <<= (PAGE_SHIFT - scale);
8230 /* Make sure we've got at least a 0-order allocation.. */
8231 if (unlikely(flags & HASH_SMALL)) {
8232 /* Makes no sense without HASH_EARLY */
8233 WARN_ON(!(flags & HASH_EARLY));
8234 if (!(numentries >> *_hash_shift)) {
8235 numentries = 1UL << *_hash_shift;
8236 BUG_ON(!numentries);
8238 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8239 numentries = PAGE_SIZE / bucketsize;
8241 numentries = roundup_pow_of_two(numentries);
8243 /* limit allocation size to 1/16 total memory by default */
8245 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8246 do_div(max, bucketsize);
8248 max = min(max, 0x80000000ULL);
8250 if (numentries < low_limit)
8251 numentries = low_limit;
8252 if (numentries > max)
8255 log2qty = ilog2(numentries);
8257 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8260 size = bucketsize << log2qty;
8261 if (flags & HASH_EARLY) {
8262 if (flags & HASH_ZERO)
8263 table = memblock_alloc(size, SMP_CACHE_BYTES);
8265 table = memblock_alloc_raw(size,
8267 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8268 table = __vmalloc(size, gfp_flags);
8272 * If bucketsize is not a power-of-two, we may free
8273 * some pages at the end of hash table which
8274 * alloc_pages_exact() automatically does
8276 table = alloc_pages_exact(size, gfp_flags);
8277 kmemleak_alloc(table, size, 1, gfp_flags);
8279 } while (!table && size > PAGE_SIZE && --log2qty);
8282 panic("Failed to allocate %s hash table\n", tablename);
8284 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8285 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8286 virt ? "vmalloc" : "linear");
8289 *_hash_shift = log2qty;
8291 *_hash_mask = (1 << log2qty) - 1;
8297 * This function checks whether pageblock includes unmovable pages or not.
8299 * PageLRU check without isolation or lru_lock could race so that
8300 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8301 * check without lock_page also may miss some movable non-lru pages at
8302 * race condition. So you can't expect this function should be exact.
8304 * Returns a page without holding a reference. If the caller wants to
8305 * dereference that page (e.g., dumping), it has to make sure that that it
8306 * cannot get removed (e.g., via memory unplug) concurrently.
8309 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8310 int migratetype, int flags)
8312 unsigned long iter = 0;
8313 unsigned long pfn = page_to_pfn(page);
8316 * TODO we could make this much more efficient by not checking every
8317 * page in the range if we know all of them are in MOVABLE_ZONE and
8318 * that the movable zone guarantees that pages are migratable but
8319 * the later is not the case right now unfortunatelly. E.g. movablecore
8320 * can still lead to having bootmem allocations in zone_movable.
8323 if (is_migrate_cma_page(page)) {
8325 * CMA allocations (alloc_contig_range) really need to mark
8326 * isolate CMA pageblocks even when they are not movable in fact
8327 * so consider them movable here.
8329 if (is_migrate_cma(migratetype))
8335 for (; iter < pageblock_nr_pages; iter++) {
8336 if (!pfn_valid_within(pfn + iter))
8339 page = pfn_to_page(pfn + iter);
8341 if (PageReserved(page))
8345 * If the zone is movable and we have ruled out all reserved
8346 * pages then it should be reasonably safe to assume the rest
8349 if (zone_idx(zone) == ZONE_MOVABLE)
8353 * Hugepages are not in LRU lists, but they're movable.
8354 * THPs are on the LRU, but need to be counted as #small pages.
8355 * We need not scan over tail pages because we don't
8356 * handle each tail page individually in migration.
8358 if (PageHuge(page) || PageTransCompound(page)) {
8359 struct page *head = compound_head(page);
8360 unsigned int skip_pages;
8362 if (PageHuge(page)) {
8363 if (!hugepage_migration_supported(page_hstate(head)))
8365 } else if (!PageLRU(head) && !__PageMovable(head)) {
8369 skip_pages = compound_nr(head) - (page - head);
8370 iter += skip_pages - 1;
8375 * We can't use page_count without pin a page
8376 * because another CPU can free compound page.
8377 * This check already skips compound tails of THP
8378 * because their page->_refcount is zero at all time.
8380 if (!page_ref_count(page)) {
8381 if (PageBuddy(page))
8382 iter += (1 << page_order(page)) - 1;
8387 * The HWPoisoned page may be not in buddy system, and
8388 * page_count() is not 0.
8390 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8393 if (__PageMovable(page) || PageLRU(page))
8397 * If there are RECLAIMABLE pages, we need to check
8398 * it. But now, memory offline itself doesn't call
8399 * shrink_node_slabs() and it still to be fixed.
8402 * If the page is not RAM, page_count()should be 0.
8403 * we don't need more check. This is an _used_ not-movable page.
8405 * The problematic thing here is PG_reserved pages. PG_reserved
8406 * is set to both of a memory hole page and a _used_ kernel
8414 #ifdef CONFIG_CONTIG_ALLOC
8415 static unsigned long pfn_max_align_down(unsigned long pfn)
8417 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8418 pageblock_nr_pages) - 1);
8421 static unsigned long pfn_max_align_up(unsigned long pfn)
8423 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8424 pageblock_nr_pages));
8427 /* [start, end) must belong to a single zone. */
8428 static int __alloc_contig_migrate_range(struct compact_control *cc,
8429 unsigned long start, unsigned long end)
8431 /* This function is based on compact_zone() from compaction.c. */
8432 unsigned long nr_reclaimed;
8433 unsigned long pfn = start;
8434 unsigned int tries = 0;
8439 while (pfn < end || !list_empty(&cc->migratepages)) {
8440 if (fatal_signal_pending(current)) {
8445 if (list_empty(&cc->migratepages)) {
8446 cc->nr_migratepages = 0;
8447 pfn = isolate_migratepages_range(cc, pfn, end);
8453 } else if (++tries == 5) {
8454 ret = ret < 0 ? ret : -EBUSY;
8458 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8460 cc->nr_migratepages -= nr_reclaimed;
8462 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
8463 NULL, 0, cc->mode, MR_CONTIG_RANGE);
8466 putback_movable_pages(&cc->migratepages);
8473 * alloc_contig_range() -- tries to allocate given range of pages
8474 * @start: start PFN to allocate
8475 * @end: one-past-the-last PFN to allocate
8476 * @migratetype: migratetype of the underlaying pageblocks (either
8477 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8478 * in range must have the same migratetype and it must
8479 * be either of the two.
8480 * @gfp_mask: GFP mask to use during compaction
8482 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8483 * aligned. The PFN range must belong to a single zone.
8485 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8486 * pageblocks in the range. Once isolated, the pageblocks should not
8487 * be modified by others.
8489 * Return: zero on success or negative error code. On success all
8490 * pages which PFN is in [start, end) are allocated for the caller and
8491 * need to be freed with free_contig_range().
8493 int alloc_contig_range(unsigned long start, unsigned long end,
8494 unsigned migratetype, gfp_t gfp_mask)
8496 unsigned long outer_start, outer_end;
8500 struct compact_control cc = {
8501 .nr_migratepages = 0,
8503 .zone = page_zone(pfn_to_page(start)),
8504 .mode = MIGRATE_SYNC,
8505 .ignore_skip_hint = true,
8506 .no_set_skip_hint = true,
8507 .gfp_mask = current_gfp_context(gfp_mask),
8508 .alloc_contig = true,
8510 INIT_LIST_HEAD(&cc.migratepages);
8513 * What we do here is we mark all pageblocks in range as
8514 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8515 * have different sizes, and due to the way page allocator
8516 * work, we align the range to biggest of the two pages so
8517 * that page allocator won't try to merge buddies from
8518 * different pageblocks and change MIGRATE_ISOLATE to some
8519 * other migration type.
8521 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8522 * migrate the pages from an unaligned range (ie. pages that
8523 * we are interested in). This will put all the pages in
8524 * range back to page allocator as MIGRATE_ISOLATE.
8526 * When this is done, we take the pages in range from page
8527 * allocator removing them from the buddy system. This way
8528 * page allocator will never consider using them.
8530 * This lets us mark the pageblocks back as
8531 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8532 * aligned range but not in the unaligned, original range are
8533 * put back to page allocator so that buddy can use them.
8536 ret = start_isolate_page_range(pfn_max_align_down(start),
8537 pfn_max_align_up(end), migratetype, 0);
8542 * In case of -EBUSY, we'd like to know which page causes problem.
8543 * So, just fall through. test_pages_isolated() has a tracepoint
8544 * which will report the busy page.
8546 * It is possible that busy pages could become available before
8547 * the call to test_pages_isolated, and the range will actually be
8548 * allocated. So, if we fall through be sure to clear ret so that
8549 * -EBUSY is not accidentally used or returned to caller.
8551 ret = __alloc_contig_migrate_range(&cc, start, end);
8552 if (ret && ret != -EBUSY)
8557 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8558 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8559 * more, all pages in [start, end) are free in page allocator.
8560 * What we are going to do is to allocate all pages from
8561 * [start, end) (that is remove them from page allocator).
8563 * The only problem is that pages at the beginning and at the
8564 * end of interesting range may be not aligned with pages that
8565 * page allocator holds, ie. they can be part of higher order
8566 * pages. Because of this, we reserve the bigger range and
8567 * once this is done free the pages we are not interested in.
8569 * We don't have to hold zone->lock here because the pages are
8570 * isolated thus they won't get removed from buddy.
8573 lru_add_drain_all();
8576 outer_start = start;
8577 while (!PageBuddy(pfn_to_page(outer_start))) {
8578 if (++order >= MAX_ORDER) {
8579 outer_start = start;
8582 outer_start &= ~0UL << order;
8585 if (outer_start != start) {
8586 order = page_order(pfn_to_page(outer_start));
8589 * outer_start page could be small order buddy page and
8590 * it doesn't include start page. Adjust outer_start
8591 * in this case to report failed page properly
8592 * on tracepoint in test_pages_isolated()
8594 if (outer_start + (1UL << order) <= start)
8595 outer_start = start;
8598 /* Make sure the range is really isolated. */
8599 if (test_pages_isolated(outer_start, end, 0)) {
8600 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8601 __func__, outer_start, end);
8606 /* Grab isolated pages from freelists. */
8607 outer_end = isolate_freepages_range(&cc, outer_start, end);
8613 /* Free head and tail (if any) */
8614 if (start != outer_start)
8615 free_contig_range(outer_start, start - outer_start);
8616 if (end != outer_end)
8617 free_contig_range(end, outer_end - end);
8620 undo_isolate_page_range(pfn_max_align_down(start),
8621 pfn_max_align_up(end), migratetype);
8625 static int __alloc_contig_pages(unsigned long start_pfn,
8626 unsigned long nr_pages, gfp_t gfp_mask)
8628 unsigned long end_pfn = start_pfn + nr_pages;
8630 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8634 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8635 unsigned long nr_pages)
8637 unsigned long i, end_pfn = start_pfn + nr_pages;
8640 for (i = start_pfn; i < end_pfn; i++) {
8641 page = pfn_to_online_page(i);
8645 if (page_zone(page) != z)
8648 if (PageReserved(page))
8651 if (page_count(page) > 0)
8660 static bool zone_spans_last_pfn(const struct zone *zone,
8661 unsigned long start_pfn, unsigned long nr_pages)
8663 unsigned long last_pfn = start_pfn + nr_pages - 1;
8665 return zone_spans_pfn(zone, last_pfn);
8669 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8670 * @nr_pages: Number of contiguous pages to allocate
8671 * @gfp_mask: GFP mask to limit search and used during compaction
8673 * @nodemask: Mask for other possible nodes
8675 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8676 * on an applicable zonelist to find a contiguous pfn range which can then be
8677 * tried for allocation with alloc_contig_range(). This routine is intended
8678 * for allocation requests which can not be fulfilled with the buddy allocator.
8680 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8681 * power of two then the alignment is guaranteed to be to the given nr_pages
8682 * (e.g. 1GB request would be aligned to 1GB).
8684 * Allocated pages can be freed with free_contig_range() or by manually calling
8685 * __free_page() on each allocated page.
8687 * Return: pointer to contiguous pages on success, or NULL if not successful.
8689 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8690 int nid, nodemask_t *nodemask)
8692 unsigned long ret, pfn, flags;
8693 struct zonelist *zonelist;
8697 zonelist = node_zonelist(nid, gfp_mask);
8698 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8699 gfp_zone(gfp_mask), nodemask) {
8700 spin_lock_irqsave(&zone->lock, flags);
8702 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8703 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8704 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8706 * We release the zone lock here because
8707 * alloc_contig_range() will also lock the zone
8708 * at some point. If there's an allocation
8709 * spinning on this lock, it may win the race
8710 * and cause alloc_contig_range() to fail...
8712 spin_unlock_irqrestore(&zone->lock, flags);
8713 ret = __alloc_contig_pages(pfn, nr_pages,
8716 return pfn_to_page(pfn);
8717 spin_lock_irqsave(&zone->lock, flags);
8721 spin_unlock_irqrestore(&zone->lock, flags);
8725 #endif /* CONFIG_CONTIG_ALLOC */
8727 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8729 unsigned int count = 0;
8731 for (; nr_pages--; pfn++) {
8732 struct page *page = pfn_to_page(pfn);
8734 count += page_count(page) != 1;
8737 WARN(count != 0, "%d pages are still in use!\n", count);
8741 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8742 * page high values need to be recalulated.
8744 void __meminit zone_pcp_update(struct zone *zone)
8746 mutex_lock(&pcp_batch_high_lock);
8747 __zone_pcp_update(zone);
8748 mutex_unlock(&pcp_batch_high_lock);
8751 void zone_pcp_reset(struct zone *zone)
8753 unsigned long flags;
8755 struct per_cpu_pageset *pset;
8757 /* avoid races with drain_pages() */
8758 local_irq_save(flags);
8759 if (zone->pageset != &boot_pageset) {
8760 for_each_online_cpu(cpu) {
8761 pset = per_cpu_ptr(zone->pageset, cpu);
8762 drain_zonestat(zone, pset);
8764 free_percpu(zone->pageset);
8765 zone->pageset = &boot_pageset;
8767 local_irq_restore(flags);
8770 #ifdef CONFIG_MEMORY_HOTREMOVE
8772 * All pages in the range must be in a single zone and isolated
8773 * before calling this.
8776 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8782 unsigned long flags;
8783 unsigned long offlined_pages = 0;
8785 /* find the first valid pfn */
8786 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8790 return offlined_pages;
8792 offline_mem_sections(pfn, end_pfn);
8793 zone = page_zone(pfn_to_page(pfn));
8794 spin_lock_irqsave(&zone->lock, flags);
8796 while (pfn < end_pfn) {
8797 if (!pfn_valid(pfn)) {
8801 page = pfn_to_page(pfn);
8803 * The HWPoisoned page may be not in buddy system, and
8804 * page_count() is not 0.
8806 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8812 BUG_ON(page_count(page));
8813 BUG_ON(!PageBuddy(page));
8814 order = page_order(page);
8815 offlined_pages += 1 << order;
8816 del_page_from_free_list(page, zone, order);
8817 pfn += (1 << order);
8819 spin_unlock_irqrestore(&zone->lock, flags);
8821 return offlined_pages;
8825 bool is_free_buddy_page(struct page *page)
8827 struct zone *zone = page_zone(page);
8828 unsigned long pfn = page_to_pfn(page);
8829 unsigned long flags;
8832 spin_lock_irqsave(&zone->lock, flags);
8833 for (order = 0; order < MAX_ORDER; order++) {
8834 struct page *page_head = page - (pfn & ((1 << order) - 1));
8836 if (PageBuddy(page_head) && page_order(page_head) >= order)
8839 spin_unlock_irqrestore(&zone->lock, flags);
8841 return order < MAX_ORDER;
8844 #ifdef CONFIG_MEMORY_FAILURE
8846 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8847 * test is performed under the zone lock to prevent a race against page
8850 bool set_hwpoison_free_buddy_page(struct page *page)
8852 struct zone *zone = page_zone(page);
8853 unsigned long pfn = page_to_pfn(page);
8854 unsigned long flags;
8856 bool hwpoisoned = false;
8858 spin_lock_irqsave(&zone->lock, flags);
8859 for (order = 0; order < MAX_ORDER; order++) {
8860 struct page *page_head = page - (pfn & ((1 << order) - 1));
8862 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8863 if (!TestSetPageHWPoison(page))
8868 spin_unlock_irqrestore(&zone->lock, flags);