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 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
339 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
340 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
341 static unsigned long required_kernelcore __initdata;
342 static unsigned long required_kernelcore_percent __initdata;
343 static unsigned long required_movablecore __initdata;
344 static unsigned long required_movablecore_percent __initdata;
345 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
346 static bool mirrored_kernelcore __meminitdata;
348 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
350 EXPORT_SYMBOL(movable_zone);
351 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
354 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
355 unsigned int nr_online_nodes __read_mostly = 1;
356 EXPORT_SYMBOL(nr_node_ids);
357 EXPORT_SYMBOL(nr_online_nodes);
360 int page_group_by_mobility_disabled __read_mostly;
362 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
364 * During boot we initialize deferred pages on-demand, as needed, but once
365 * page_alloc_init_late() has finished, the deferred pages are all initialized,
366 * and we can permanently disable that path.
368 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
371 * Calling kasan_free_pages() only after deferred memory initialization
372 * has completed. Poisoning pages during deferred memory init will greatly
373 * lengthen the process and cause problem in large memory systems as the
374 * deferred pages initialization is done with interrupt disabled.
376 * Assuming that there will be no reference to those newly initialized
377 * pages before they are ever allocated, this should have no effect on
378 * KASAN memory tracking as the poison will be properly inserted at page
379 * allocation time. The only corner case is when pages are allocated by
380 * on-demand allocation and then freed again before the deferred pages
381 * initialization is done, but this is not likely to happen.
383 static inline void kasan_free_nondeferred_pages(struct page *page, int order)
385 if (!static_branch_unlikely(&deferred_pages))
386 kasan_free_pages(page, order);
389 /* Returns true if the struct page for the pfn is uninitialised */
390 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
392 int nid = early_pfn_to_nid(pfn);
394 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
401 * Returns true when the remaining initialisation should be deferred until
402 * later in the boot cycle when it can be parallelised.
404 static bool __meminit
405 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
407 static unsigned long prev_end_pfn, nr_initialised;
410 * prev_end_pfn static that contains the end of previous zone
411 * No need to protect because called very early in boot before smp_init.
413 if (prev_end_pfn != end_pfn) {
414 prev_end_pfn = end_pfn;
418 /* Always populate low zones for address-constrained allocations */
419 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
423 * We start only with one section of pages, more pages are added as
424 * needed until the rest of deferred pages are initialized.
427 if ((nr_initialised > PAGES_PER_SECTION) &&
428 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
429 NODE_DATA(nid)->first_deferred_pfn = pfn;
435 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
437 static inline bool early_page_uninitialised(unsigned long pfn)
442 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
448 /* Return a pointer to the bitmap storing bits affecting a block of pages */
449 static inline unsigned long *get_pageblock_bitmap(struct page *page,
452 #ifdef CONFIG_SPARSEMEM
453 return section_to_usemap(__pfn_to_section(pfn));
455 return page_zone(page)->pageblock_flags;
456 #endif /* CONFIG_SPARSEMEM */
459 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
461 #ifdef CONFIG_SPARSEMEM
462 pfn &= (PAGES_PER_SECTION-1);
463 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
465 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
466 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
467 #endif /* CONFIG_SPARSEMEM */
471 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
472 * @page: The page within the block of interest
473 * @pfn: The target page frame number
474 * @end_bitidx: The last bit of interest to retrieve
475 * @mask: mask of bits that the caller is interested in
477 * Return: pageblock_bits flags
479 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
481 unsigned long end_bitidx,
484 unsigned long *bitmap;
485 unsigned long bitidx, word_bitidx;
488 bitmap = get_pageblock_bitmap(page, pfn);
489 bitidx = pfn_to_bitidx(page, pfn);
490 word_bitidx = bitidx / BITS_PER_LONG;
491 bitidx &= (BITS_PER_LONG-1);
493 word = bitmap[word_bitidx];
494 bitidx += end_bitidx;
495 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
498 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
499 unsigned long end_bitidx,
502 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
505 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
507 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
511 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
512 * @page: The page within the block of interest
513 * @flags: The flags to set
514 * @pfn: The target page frame number
515 * @end_bitidx: The last bit of interest
516 * @mask: mask of bits that the caller is interested in
518 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
520 unsigned long end_bitidx,
523 unsigned long *bitmap;
524 unsigned long bitidx, word_bitidx;
525 unsigned long old_word, word;
527 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
528 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
530 bitmap = get_pageblock_bitmap(page, pfn);
531 bitidx = pfn_to_bitidx(page, pfn);
532 word_bitidx = bitidx / BITS_PER_LONG;
533 bitidx &= (BITS_PER_LONG-1);
535 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
537 bitidx += end_bitidx;
538 mask <<= (BITS_PER_LONG - bitidx - 1);
539 flags <<= (BITS_PER_LONG - bitidx - 1);
541 word = READ_ONCE(bitmap[word_bitidx]);
543 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
544 if (word == old_word)
550 void set_pageblock_migratetype(struct page *page, int migratetype)
552 if (unlikely(page_group_by_mobility_disabled &&
553 migratetype < MIGRATE_PCPTYPES))
554 migratetype = MIGRATE_UNMOVABLE;
556 set_pageblock_flags_group(page, (unsigned long)migratetype,
557 PB_migrate, PB_migrate_end);
560 #ifdef CONFIG_DEBUG_VM
561 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
565 unsigned long pfn = page_to_pfn(page);
566 unsigned long sp, start_pfn;
569 seq = zone_span_seqbegin(zone);
570 start_pfn = zone->zone_start_pfn;
571 sp = zone->spanned_pages;
572 if (!zone_spans_pfn(zone, pfn))
574 } while (zone_span_seqretry(zone, seq));
577 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
578 pfn, zone_to_nid(zone), zone->name,
579 start_pfn, start_pfn + sp);
584 static int page_is_consistent(struct zone *zone, struct page *page)
586 if (!pfn_valid_within(page_to_pfn(page)))
588 if (zone != page_zone(page))
594 * Temporary debugging check for pages not lying within a given zone.
596 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
598 if (page_outside_zone_boundaries(zone, page))
600 if (!page_is_consistent(zone, page))
606 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
612 static void bad_page(struct page *page, const char *reason,
613 unsigned long bad_flags)
615 static unsigned long resume;
616 static unsigned long nr_shown;
617 static unsigned long nr_unshown;
620 * Allow a burst of 60 reports, then keep quiet for that minute;
621 * or allow a steady drip of one report per second.
623 if (nr_shown == 60) {
624 if (time_before(jiffies, resume)) {
630 "BUG: Bad page state: %lu messages suppressed\n",
637 resume = jiffies + 60 * HZ;
639 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
640 current->comm, page_to_pfn(page));
641 __dump_page(page, reason);
642 bad_flags &= page->flags;
644 pr_alert("bad because of flags: %#lx(%pGp)\n",
645 bad_flags, &bad_flags);
646 dump_page_owner(page);
651 /* Leave bad fields for debug, except PageBuddy could make trouble */
652 page_mapcount_reset(page); /* remove PageBuddy */
653 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
657 * Higher-order pages are called "compound pages". They are structured thusly:
659 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
661 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
662 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
664 * The first tail page's ->compound_dtor holds the offset in array of compound
665 * page destructors. See compound_page_dtors.
667 * The first tail page's ->compound_order holds the order of allocation.
668 * This usage means that zero-order pages may not be compound.
671 void free_compound_page(struct page *page)
673 mem_cgroup_uncharge(page);
674 __free_pages_ok(page, compound_order(page));
677 void prep_compound_page(struct page *page, unsigned int order)
680 int nr_pages = 1 << order;
682 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
683 set_compound_order(page, order);
685 for (i = 1; i < nr_pages; i++) {
686 struct page *p = page + i;
687 set_page_count(p, 0);
688 p->mapping = TAIL_MAPPING;
689 set_compound_head(p, page);
691 atomic_set(compound_mapcount_ptr(page), -1);
692 if (hpage_pincount_available(page))
693 atomic_set(compound_pincount_ptr(page), 0);
696 #ifdef CONFIG_DEBUG_PAGEALLOC
697 unsigned int _debug_guardpage_minorder;
699 bool _debug_pagealloc_enabled_early __read_mostly
700 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
701 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
702 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
703 EXPORT_SYMBOL(_debug_pagealloc_enabled);
705 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
707 static int __init early_debug_pagealloc(char *buf)
709 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
711 early_param("debug_pagealloc", early_debug_pagealloc);
713 void init_debug_pagealloc(void)
715 if (!debug_pagealloc_enabled())
718 static_branch_enable(&_debug_pagealloc_enabled);
720 if (!debug_guardpage_minorder())
723 static_branch_enable(&_debug_guardpage_enabled);
726 static int __init debug_guardpage_minorder_setup(char *buf)
730 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
731 pr_err("Bad debug_guardpage_minorder value\n");
734 _debug_guardpage_minorder = res;
735 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
738 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
740 static inline bool set_page_guard(struct zone *zone, struct page *page,
741 unsigned int order, int migratetype)
743 if (!debug_guardpage_enabled())
746 if (order >= debug_guardpage_minorder())
749 __SetPageGuard(page);
750 INIT_LIST_HEAD(&page->lru);
751 set_page_private(page, order);
752 /* Guard pages are not available for any usage */
753 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
758 static inline void clear_page_guard(struct zone *zone, struct page *page,
759 unsigned int order, int migratetype)
761 if (!debug_guardpage_enabled())
764 __ClearPageGuard(page);
766 set_page_private(page, 0);
767 if (!is_migrate_isolate(migratetype))
768 __mod_zone_freepage_state(zone, (1 << order), migratetype);
771 static inline bool set_page_guard(struct zone *zone, struct page *page,
772 unsigned int order, int migratetype) { return false; }
773 static inline void clear_page_guard(struct zone *zone, struct page *page,
774 unsigned int order, int migratetype) {}
777 static inline void set_page_order(struct page *page, unsigned int order)
779 set_page_private(page, order);
780 __SetPageBuddy(page);
784 * This function checks whether a page is free && is the buddy
785 * we can coalesce a page and its buddy if
786 * (a) the buddy is not in a hole (check before calling!) &&
787 * (b) the buddy is in the buddy system &&
788 * (c) a page and its buddy have the same order &&
789 * (d) a page and its buddy are in the same zone.
791 * For recording whether a page is in the buddy system, we set PageBuddy.
792 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
794 * For recording page's order, we use page_private(page).
796 static inline bool page_is_buddy(struct page *page, struct page *buddy,
799 if (!page_is_guard(buddy) && !PageBuddy(buddy))
802 if (page_order(buddy) != order)
806 * zone check is done late to avoid uselessly calculating
807 * zone/node ids for pages that could never merge.
809 if (page_zone_id(page) != page_zone_id(buddy))
812 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
817 #ifdef CONFIG_COMPACTION
818 static inline struct capture_control *task_capc(struct zone *zone)
820 struct capture_control *capc = current->capture_control;
823 !(current->flags & PF_KTHREAD) &&
825 capc->cc->zone == zone &&
826 capc->cc->direct_compaction ? capc : NULL;
830 compaction_capture(struct capture_control *capc, struct page *page,
831 int order, int migratetype)
833 if (!capc || order != capc->cc->order)
836 /* Do not accidentally pollute CMA or isolated regions*/
837 if (is_migrate_cma(migratetype) ||
838 is_migrate_isolate(migratetype))
842 * Do not let lower order allocations polluate a movable pageblock.
843 * This might let an unmovable request use a reclaimable pageblock
844 * and vice-versa but no more than normal fallback logic which can
845 * have trouble finding a high-order free page.
847 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
855 static inline struct capture_control *task_capc(struct zone *zone)
861 compaction_capture(struct capture_control *capc, struct page *page,
862 int order, int migratetype)
866 #endif /* CONFIG_COMPACTION */
868 /* Used for pages not on another list */
869 static inline void add_to_free_list(struct page *page, struct zone *zone,
870 unsigned int order, int migratetype)
872 struct free_area *area = &zone->free_area[order];
874 list_add(&page->lru, &area->free_list[migratetype]);
878 /* Used for pages not on another list */
879 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
880 unsigned int order, int migratetype)
882 struct free_area *area = &zone->free_area[order];
884 list_add_tail(&page->lru, &area->free_list[migratetype]);
888 /* Used for pages which are on another list */
889 static inline void move_to_free_list(struct page *page, struct zone *zone,
890 unsigned int order, int migratetype)
892 struct free_area *area = &zone->free_area[order];
894 list_move(&page->lru, &area->free_list[migratetype]);
897 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
900 /* clear reported state and update reported page count */
901 if (page_reported(page))
902 __ClearPageReported(page);
904 list_del(&page->lru);
905 __ClearPageBuddy(page);
906 set_page_private(page, 0);
907 zone->free_area[order].nr_free--;
911 * If this is not the largest possible page, check if the buddy
912 * of the next-highest order is free. If it is, it's possible
913 * that pages are being freed that will coalesce soon. In case,
914 * that is happening, add the free page to the tail of the list
915 * so it's less likely to be used soon and more likely to be merged
916 * as a higher order page
919 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
920 struct page *page, unsigned int order)
922 struct page *higher_page, *higher_buddy;
923 unsigned long combined_pfn;
925 if (order >= MAX_ORDER - 2)
928 if (!pfn_valid_within(buddy_pfn))
931 combined_pfn = buddy_pfn & pfn;
932 higher_page = page + (combined_pfn - pfn);
933 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
934 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
936 return pfn_valid_within(buddy_pfn) &&
937 page_is_buddy(higher_page, higher_buddy, order + 1);
941 * Freeing function for a buddy system allocator.
943 * The concept of a buddy system is to maintain direct-mapped table
944 * (containing bit values) for memory blocks of various "orders".
945 * The bottom level table contains the map for the smallest allocatable
946 * units of memory (here, pages), and each level above it describes
947 * pairs of units from the levels below, hence, "buddies".
948 * At a high level, all that happens here is marking the table entry
949 * at the bottom level available, and propagating the changes upward
950 * as necessary, plus some accounting needed to play nicely with other
951 * parts of the VM system.
952 * At each level, we keep a list of pages, which are heads of continuous
953 * free pages of length of (1 << order) and marked with PageBuddy.
954 * Page's order is recorded in page_private(page) field.
955 * So when we are allocating or freeing one, we can derive the state of the
956 * other. That is, if we allocate a small block, and both were
957 * free, the remainder of the region must be split into blocks.
958 * If a block is freed, and its buddy is also free, then this
959 * triggers coalescing into a block of larger size.
964 static inline void __free_one_page(struct page *page,
966 struct zone *zone, unsigned int order,
967 int migratetype, bool report)
969 struct capture_control *capc = task_capc(zone);
970 unsigned long uninitialized_var(buddy_pfn);
971 unsigned long combined_pfn;
972 unsigned int max_order;
976 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
978 VM_BUG_ON(!zone_is_initialized(zone));
979 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
981 VM_BUG_ON(migratetype == -1);
982 if (likely(!is_migrate_isolate(migratetype)))
983 __mod_zone_freepage_state(zone, 1 << order, migratetype);
985 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
986 VM_BUG_ON_PAGE(bad_range(zone, page), page);
989 while (order < max_order - 1) {
990 if (compaction_capture(capc, page, order, migratetype)) {
991 __mod_zone_freepage_state(zone, -(1 << order),
995 buddy_pfn = __find_buddy_pfn(pfn, order);
996 buddy = page + (buddy_pfn - pfn);
998 if (!pfn_valid_within(buddy_pfn))
1000 if (!page_is_buddy(page, buddy, order))
1003 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1004 * merge with it and move up one order.
1006 if (page_is_guard(buddy))
1007 clear_page_guard(zone, buddy, order, migratetype);
1009 del_page_from_free_list(buddy, zone, order);
1010 combined_pfn = buddy_pfn & pfn;
1011 page = page + (combined_pfn - pfn);
1015 if (max_order < MAX_ORDER) {
1016 /* If we are here, it means order is >= pageblock_order.
1017 * We want to prevent merge between freepages on isolate
1018 * pageblock and normal pageblock. Without this, pageblock
1019 * isolation could cause incorrect freepage or CMA accounting.
1021 * We don't want to hit this code for the more frequent
1022 * low-order merging.
1024 if (unlikely(has_isolate_pageblock(zone))) {
1027 buddy_pfn = __find_buddy_pfn(pfn, order);
1028 buddy = page + (buddy_pfn - pfn);
1029 buddy_mt = get_pageblock_migratetype(buddy);
1031 if (migratetype != buddy_mt
1032 && (is_migrate_isolate(migratetype) ||
1033 is_migrate_isolate(buddy_mt)))
1037 goto continue_merging;
1041 set_page_order(page, order);
1043 if (is_shuffle_order(order))
1044 to_tail = shuffle_pick_tail();
1046 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1049 add_to_free_list_tail(page, zone, order, migratetype);
1051 add_to_free_list(page, zone, order, migratetype);
1053 /* Notify page reporting subsystem of freed page */
1055 page_reporting_notify_free(order);
1059 * A bad page could be due to a number of fields. Instead of multiple branches,
1060 * try and check multiple fields with one check. The caller must do a detailed
1061 * check if necessary.
1063 static inline bool page_expected_state(struct page *page,
1064 unsigned long check_flags)
1066 if (unlikely(atomic_read(&page->_mapcount) != -1))
1069 if (unlikely((unsigned long)page->mapping |
1070 page_ref_count(page) |
1072 (unsigned long)page->mem_cgroup |
1074 (page->flags & check_flags)))
1080 static void free_pages_check_bad(struct page *page)
1082 const char *bad_reason;
1083 unsigned long bad_flags;
1088 if (unlikely(atomic_read(&page->_mapcount) != -1))
1089 bad_reason = "nonzero mapcount";
1090 if (unlikely(page->mapping != NULL))
1091 bad_reason = "non-NULL mapping";
1092 if (unlikely(page_ref_count(page) != 0))
1093 bad_reason = "nonzero _refcount";
1094 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
1095 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1096 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
1099 if (unlikely(page->mem_cgroup))
1100 bad_reason = "page still charged to cgroup";
1102 bad_page(page, bad_reason, bad_flags);
1105 static inline int free_pages_check(struct page *page)
1107 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1110 /* Something has gone sideways, find it */
1111 free_pages_check_bad(page);
1115 static int free_tail_pages_check(struct page *head_page, struct page *page)
1120 * We rely page->lru.next never has bit 0 set, unless the page
1121 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1123 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1125 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1129 switch (page - head_page) {
1131 /* the first tail page: ->mapping may be compound_mapcount() */
1132 if (unlikely(compound_mapcount(page))) {
1133 bad_page(page, "nonzero compound_mapcount", 0);
1139 * the second tail page: ->mapping is
1140 * deferred_list.next -- ignore value.
1144 if (page->mapping != TAIL_MAPPING) {
1145 bad_page(page, "corrupted mapping in tail page", 0);
1150 if (unlikely(!PageTail(page))) {
1151 bad_page(page, "PageTail not set", 0);
1154 if (unlikely(compound_head(page) != head_page)) {
1155 bad_page(page, "compound_head not consistent", 0);
1160 page->mapping = NULL;
1161 clear_compound_head(page);
1165 static void kernel_init_free_pages(struct page *page, int numpages)
1169 for (i = 0; i < numpages; i++)
1170 clear_highpage(page + i);
1173 static __always_inline bool free_pages_prepare(struct page *page,
1174 unsigned int order, bool check_free)
1178 VM_BUG_ON_PAGE(PageTail(page), page);
1180 trace_mm_page_free(page, order);
1183 * Check tail pages before head page information is cleared to
1184 * avoid checking PageCompound for order-0 pages.
1186 if (unlikely(order)) {
1187 bool compound = PageCompound(page);
1190 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1193 ClearPageDoubleMap(page);
1194 for (i = 1; i < (1 << order); i++) {
1196 bad += free_tail_pages_check(page, page + i);
1197 if (unlikely(free_pages_check(page + i))) {
1201 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1204 if (PageMappingFlags(page))
1205 page->mapping = NULL;
1206 if (memcg_kmem_enabled() && PageKmemcg(page))
1207 __memcg_kmem_uncharge_page(page, order);
1209 bad += free_pages_check(page);
1213 page_cpupid_reset_last(page);
1214 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1215 reset_page_owner(page, order);
1217 if (!PageHighMem(page)) {
1218 debug_check_no_locks_freed(page_address(page),
1219 PAGE_SIZE << order);
1220 debug_check_no_obj_freed(page_address(page),
1221 PAGE_SIZE << order);
1223 if (want_init_on_free())
1224 kernel_init_free_pages(page, 1 << order);
1226 kernel_poison_pages(page, 1 << order, 0);
1228 * arch_free_page() can make the page's contents inaccessible. s390
1229 * does this. So nothing which can access the page's contents should
1230 * happen after this.
1232 arch_free_page(page, order);
1234 if (debug_pagealloc_enabled_static())
1235 kernel_map_pages(page, 1 << order, 0);
1237 kasan_free_nondeferred_pages(page, order);
1242 #ifdef CONFIG_DEBUG_VM
1244 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1245 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1246 * moved from pcp lists to free lists.
1248 static bool free_pcp_prepare(struct page *page)
1250 return free_pages_prepare(page, 0, true);
1253 static bool bulkfree_pcp_prepare(struct page *page)
1255 if (debug_pagealloc_enabled_static())
1256 return free_pages_check(page);
1262 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1263 * moving from pcp lists to free list in order to reduce overhead. With
1264 * debug_pagealloc enabled, they are checked also immediately when being freed
1267 static bool free_pcp_prepare(struct page *page)
1269 if (debug_pagealloc_enabled_static())
1270 return free_pages_prepare(page, 0, true);
1272 return free_pages_prepare(page, 0, false);
1275 static bool bulkfree_pcp_prepare(struct page *page)
1277 return free_pages_check(page);
1279 #endif /* CONFIG_DEBUG_VM */
1281 static inline void prefetch_buddy(struct page *page)
1283 unsigned long pfn = page_to_pfn(page);
1284 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1285 struct page *buddy = page + (buddy_pfn - pfn);
1291 * Frees a number of pages from the PCP lists
1292 * Assumes all pages on list are in same zone, and of same order.
1293 * count is the number of pages to free.
1295 * If the zone was previously in an "all pages pinned" state then look to
1296 * see if this freeing clears that state.
1298 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1299 * pinned" detection logic.
1301 static void free_pcppages_bulk(struct zone *zone, int count,
1302 struct per_cpu_pages *pcp)
1304 int migratetype = 0;
1306 int prefetch_nr = 0;
1307 bool isolated_pageblocks;
1308 struct page *page, *tmp;
1312 struct list_head *list;
1315 * Remove pages from lists in a round-robin fashion. A
1316 * batch_free count is maintained that is incremented when an
1317 * empty list is encountered. This is so more pages are freed
1318 * off fuller lists instead of spinning excessively around empty
1323 if (++migratetype == MIGRATE_PCPTYPES)
1325 list = &pcp->lists[migratetype];
1326 } while (list_empty(list));
1328 /* This is the only non-empty list. Free them all. */
1329 if (batch_free == MIGRATE_PCPTYPES)
1333 page = list_last_entry(list, struct page, lru);
1334 /* must delete to avoid corrupting pcp list */
1335 list_del(&page->lru);
1338 if (bulkfree_pcp_prepare(page))
1341 list_add_tail(&page->lru, &head);
1344 * We are going to put the page back to the global
1345 * pool, prefetch its buddy to speed up later access
1346 * under zone->lock. It is believed the overhead of
1347 * an additional test and calculating buddy_pfn here
1348 * can be offset by reduced memory latency later. To
1349 * avoid excessive prefetching due to large count, only
1350 * prefetch buddy for the first pcp->batch nr of pages.
1352 if (prefetch_nr++ < pcp->batch)
1353 prefetch_buddy(page);
1354 } while (--count && --batch_free && !list_empty(list));
1357 spin_lock(&zone->lock);
1358 isolated_pageblocks = has_isolate_pageblock(zone);
1361 * Use safe version since after __free_one_page(),
1362 * page->lru.next will not point to original list.
1364 list_for_each_entry_safe(page, tmp, &head, lru) {
1365 int mt = get_pcppage_migratetype(page);
1366 /* MIGRATE_ISOLATE page should not go to pcplists */
1367 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1368 /* Pageblock could have been isolated meanwhile */
1369 if (unlikely(isolated_pageblocks))
1370 mt = get_pageblock_migratetype(page);
1372 __free_one_page(page, page_to_pfn(page), zone, 0, mt, true);
1373 trace_mm_page_pcpu_drain(page, 0, mt);
1375 spin_unlock(&zone->lock);
1378 static void free_one_page(struct zone *zone,
1379 struct page *page, unsigned long pfn,
1383 spin_lock(&zone->lock);
1384 if (unlikely(has_isolate_pageblock(zone) ||
1385 is_migrate_isolate(migratetype))) {
1386 migratetype = get_pfnblock_migratetype(page, pfn);
1388 __free_one_page(page, pfn, zone, order, migratetype, true);
1389 spin_unlock(&zone->lock);
1392 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1393 unsigned long zone, int nid)
1395 mm_zero_struct_page(page);
1396 set_page_links(page, zone, nid, pfn);
1397 init_page_count(page);
1398 page_mapcount_reset(page);
1399 page_cpupid_reset_last(page);
1400 page_kasan_tag_reset(page);
1402 INIT_LIST_HEAD(&page->lru);
1403 #ifdef WANT_PAGE_VIRTUAL
1404 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1405 if (!is_highmem_idx(zone))
1406 set_page_address(page, __va(pfn << PAGE_SHIFT));
1410 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1411 static void __meminit init_reserved_page(unsigned long pfn)
1416 if (!early_page_uninitialised(pfn))
1419 nid = early_pfn_to_nid(pfn);
1420 pgdat = NODE_DATA(nid);
1422 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1423 struct zone *zone = &pgdat->node_zones[zid];
1425 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1428 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1431 static inline void init_reserved_page(unsigned long pfn)
1434 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1437 * Initialised pages do not have PageReserved set. This function is
1438 * called for each range allocated by the bootmem allocator and
1439 * marks the pages PageReserved. The remaining valid pages are later
1440 * sent to the buddy page allocator.
1442 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1444 unsigned long start_pfn = PFN_DOWN(start);
1445 unsigned long end_pfn = PFN_UP(end);
1447 for (; start_pfn < end_pfn; start_pfn++) {
1448 if (pfn_valid(start_pfn)) {
1449 struct page *page = pfn_to_page(start_pfn);
1451 init_reserved_page(start_pfn);
1453 /* Avoid false-positive PageTail() */
1454 INIT_LIST_HEAD(&page->lru);
1457 * no need for atomic set_bit because the struct
1458 * page is not visible yet so nobody should
1461 __SetPageReserved(page);
1466 static void __free_pages_ok(struct page *page, unsigned int order)
1468 unsigned long flags;
1470 unsigned long pfn = page_to_pfn(page);
1472 if (!free_pages_prepare(page, order, true))
1475 migratetype = get_pfnblock_migratetype(page, pfn);
1476 local_irq_save(flags);
1477 __count_vm_events(PGFREE, 1 << order);
1478 free_one_page(page_zone(page), page, pfn, order, migratetype);
1479 local_irq_restore(flags);
1482 void __free_pages_core(struct page *page, unsigned int order)
1484 unsigned int nr_pages = 1 << order;
1485 struct page *p = page;
1489 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1491 __ClearPageReserved(p);
1492 set_page_count(p, 0);
1494 __ClearPageReserved(p);
1495 set_page_count(p, 0);
1497 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1498 set_page_refcounted(page);
1499 __free_pages(page, order);
1502 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1503 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1505 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1507 int __meminit early_pfn_to_nid(unsigned long pfn)
1509 static DEFINE_SPINLOCK(early_pfn_lock);
1512 spin_lock(&early_pfn_lock);
1513 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1515 nid = first_online_node;
1516 spin_unlock(&early_pfn_lock);
1522 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1523 /* Only safe to use early in boot when initialisation is single-threaded */
1524 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1528 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1529 if (nid >= 0 && nid != node)
1535 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1542 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1545 if (early_page_uninitialised(pfn))
1547 __free_pages_core(page, order);
1551 * Check that the whole (or subset of) a pageblock given by the interval of
1552 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1553 * with the migration of free compaction scanner. The scanners then need to
1554 * use only pfn_valid_within() check for arches that allow holes within
1557 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1559 * It's possible on some configurations to have a setup like node0 node1 node0
1560 * i.e. it's possible that all pages within a zones range of pages do not
1561 * belong to a single zone. We assume that a border between node0 and node1
1562 * can occur within a single pageblock, but not a node0 node1 node0
1563 * interleaving within a single pageblock. It is therefore sufficient to check
1564 * the first and last page of a pageblock and avoid checking each individual
1565 * page in a pageblock.
1567 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1568 unsigned long end_pfn, struct zone *zone)
1570 struct page *start_page;
1571 struct page *end_page;
1573 /* end_pfn is one past the range we are checking */
1576 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1579 start_page = pfn_to_online_page(start_pfn);
1583 if (page_zone(start_page) != zone)
1586 end_page = pfn_to_page(end_pfn);
1588 /* This gives a shorter code than deriving page_zone(end_page) */
1589 if (page_zone_id(start_page) != page_zone_id(end_page))
1595 void set_zone_contiguous(struct zone *zone)
1597 unsigned long block_start_pfn = zone->zone_start_pfn;
1598 unsigned long block_end_pfn;
1600 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1601 for (; block_start_pfn < zone_end_pfn(zone);
1602 block_start_pfn = block_end_pfn,
1603 block_end_pfn += pageblock_nr_pages) {
1605 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1607 if (!__pageblock_pfn_to_page(block_start_pfn,
1608 block_end_pfn, zone))
1613 /* We confirm that there is no hole */
1614 zone->contiguous = true;
1617 void clear_zone_contiguous(struct zone *zone)
1619 zone->contiguous = false;
1622 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1623 static void __init deferred_free_range(unsigned long pfn,
1624 unsigned long nr_pages)
1632 page = pfn_to_page(pfn);
1634 /* Free a large naturally-aligned chunk if possible */
1635 if (nr_pages == pageblock_nr_pages &&
1636 (pfn & (pageblock_nr_pages - 1)) == 0) {
1637 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1638 __free_pages_core(page, pageblock_order);
1642 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1643 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1644 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1645 __free_pages_core(page, 0);
1649 /* Completion tracking for deferred_init_memmap() threads */
1650 static atomic_t pgdat_init_n_undone __initdata;
1651 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1653 static inline void __init pgdat_init_report_one_done(void)
1655 if (atomic_dec_and_test(&pgdat_init_n_undone))
1656 complete(&pgdat_init_all_done_comp);
1660 * Returns true if page needs to be initialized or freed to buddy allocator.
1662 * First we check if pfn is valid on architectures where it is possible to have
1663 * holes within pageblock_nr_pages. On systems where it is not possible, this
1664 * function is optimized out.
1666 * Then, we check if a current large page is valid by only checking the validity
1669 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1671 if (!pfn_valid_within(pfn))
1673 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1679 * Free pages to buddy allocator. Try to free aligned pages in
1680 * pageblock_nr_pages sizes.
1682 static void __init deferred_free_pages(unsigned long pfn,
1683 unsigned long end_pfn)
1685 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1686 unsigned long nr_free = 0;
1688 for (; pfn < end_pfn; pfn++) {
1689 if (!deferred_pfn_valid(pfn)) {
1690 deferred_free_range(pfn - nr_free, nr_free);
1692 } else if (!(pfn & nr_pgmask)) {
1693 deferred_free_range(pfn - nr_free, nr_free);
1695 touch_nmi_watchdog();
1700 /* Free the last block of pages to allocator */
1701 deferred_free_range(pfn - nr_free, nr_free);
1705 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1706 * by performing it only once every pageblock_nr_pages.
1707 * Return number of pages initialized.
1709 static unsigned long __init deferred_init_pages(struct zone *zone,
1711 unsigned long end_pfn)
1713 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1714 int nid = zone_to_nid(zone);
1715 unsigned long nr_pages = 0;
1716 int zid = zone_idx(zone);
1717 struct page *page = NULL;
1719 for (; pfn < end_pfn; pfn++) {
1720 if (!deferred_pfn_valid(pfn)) {
1723 } else if (!page || !(pfn & nr_pgmask)) {
1724 page = pfn_to_page(pfn);
1725 touch_nmi_watchdog();
1729 __init_single_page(page, pfn, zid, nid);
1736 * This function is meant to pre-load the iterator for the zone init.
1737 * Specifically it walks through the ranges until we are caught up to the
1738 * first_init_pfn value and exits there. If we never encounter the value we
1739 * return false indicating there are no valid ranges left.
1742 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1743 unsigned long *spfn, unsigned long *epfn,
1744 unsigned long first_init_pfn)
1749 * Start out by walking through the ranges in this zone that have
1750 * already been initialized. We don't need to do anything with them
1751 * so we just need to flush them out of the system.
1753 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1754 if (*epfn <= first_init_pfn)
1756 if (*spfn < first_init_pfn)
1757 *spfn = first_init_pfn;
1766 * Initialize and free pages. We do it in two loops: first we initialize
1767 * struct page, then free to buddy allocator, because while we are
1768 * freeing pages we can access pages that are ahead (computing buddy
1769 * page in __free_one_page()).
1771 * In order to try and keep some memory in the cache we have the loop
1772 * broken along max page order boundaries. This way we will not cause
1773 * any issues with the buddy page computation.
1775 static unsigned long __init
1776 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1777 unsigned long *end_pfn)
1779 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1780 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1781 unsigned long nr_pages = 0;
1784 /* First we loop through and initialize the page values */
1785 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1788 if (mo_pfn <= *start_pfn)
1791 t = min(mo_pfn, *end_pfn);
1792 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1794 if (mo_pfn < *end_pfn) {
1795 *start_pfn = mo_pfn;
1800 /* Reset values and now loop through freeing pages as needed */
1803 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1809 t = min(mo_pfn, epfn);
1810 deferred_free_pages(spfn, t);
1819 /* Initialise remaining memory on a node */
1820 static int __init deferred_init_memmap(void *data)
1822 pg_data_t *pgdat = data;
1823 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1824 unsigned long spfn = 0, epfn = 0, nr_pages = 0;
1825 unsigned long first_init_pfn, flags;
1826 unsigned long start = jiffies;
1831 /* Bind memory initialisation thread to a local node if possible */
1832 if (!cpumask_empty(cpumask))
1833 set_cpus_allowed_ptr(current, cpumask);
1835 pgdat_resize_lock(pgdat, &flags);
1836 first_init_pfn = pgdat->first_deferred_pfn;
1837 if (first_init_pfn == ULONG_MAX) {
1838 pgdat_resize_unlock(pgdat, &flags);
1839 pgdat_init_report_one_done();
1843 /* Sanity check boundaries */
1844 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1845 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1846 pgdat->first_deferred_pfn = ULONG_MAX;
1848 /* Only the highest zone is deferred so find it */
1849 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1850 zone = pgdat->node_zones + zid;
1851 if (first_init_pfn < zone_end_pfn(zone))
1855 /* If the zone is empty somebody else may have cleared out the zone */
1856 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1861 * Initialize and free pages in MAX_ORDER sized increments so
1862 * that we can avoid introducing any issues with the buddy
1866 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1868 pgdat_resize_unlock(pgdat, &flags);
1870 /* Sanity check that the next zone really is unpopulated */
1871 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1873 pr_info("node %d initialised, %lu pages in %ums\n",
1874 pgdat->node_id, nr_pages, jiffies_to_msecs(jiffies - start));
1876 pgdat_init_report_one_done();
1881 * If this zone has deferred pages, try to grow it by initializing enough
1882 * deferred pages to satisfy the allocation specified by order, rounded up to
1883 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1884 * of SECTION_SIZE bytes by initializing struct pages in increments of
1885 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1887 * Return true when zone was grown, otherwise return false. We return true even
1888 * when we grow less than requested, to let the caller decide if there are
1889 * enough pages to satisfy the allocation.
1891 * Note: We use noinline because this function is needed only during boot, and
1892 * it is called from a __ref function _deferred_grow_zone. This way we are
1893 * making sure that it is not inlined into permanent text section.
1895 static noinline bool __init
1896 deferred_grow_zone(struct zone *zone, unsigned int order)
1898 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1899 pg_data_t *pgdat = zone->zone_pgdat;
1900 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1901 unsigned long spfn, epfn, flags;
1902 unsigned long nr_pages = 0;
1905 /* Only the last zone may have deferred pages */
1906 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1909 pgdat_resize_lock(pgdat, &flags);
1912 * If deferred pages have been initialized while we were waiting for
1913 * the lock, return true, as the zone was grown. The caller will retry
1914 * this zone. We won't return to this function since the caller also
1915 * has this static branch.
1917 if (!static_branch_unlikely(&deferred_pages)) {
1918 pgdat_resize_unlock(pgdat, &flags);
1923 * If someone grew this zone while we were waiting for spinlock, return
1924 * true, as there might be enough pages already.
1926 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1927 pgdat_resize_unlock(pgdat, &flags);
1931 /* If the zone is empty somebody else may have cleared out the zone */
1932 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1933 first_deferred_pfn)) {
1934 pgdat->first_deferred_pfn = ULONG_MAX;
1935 pgdat_resize_unlock(pgdat, &flags);
1936 /* Retry only once. */
1937 return first_deferred_pfn != ULONG_MAX;
1941 * Initialize and free pages in MAX_ORDER sized increments so
1942 * that we can avoid introducing any issues with the buddy
1945 while (spfn < epfn) {
1946 /* update our first deferred PFN for this section */
1947 first_deferred_pfn = spfn;
1949 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1951 /* We should only stop along section boundaries */
1952 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
1955 /* If our quota has been met we can stop here */
1956 if (nr_pages >= nr_pages_needed)
1960 pgdat->first_deferred_pfn = spfn;
1961 pgdat_resize_unlock(pgdat, &flags);
1963 return nr_pages > 0;
1967 * deferred_grow_zone() is __init, but it is called from
1968 * get_page_from_freelist() during early boot until deferred_pages permanently
1969 * disables this call. This is why we have refdata wrapper to avoid warning,
1970 * and to ensure that the function body gets unloaded.
1973 _deferred_grow_zone(struct zone *zone, unsigned int order)
1975 return deferred_grow_zone(zone, order);
1978 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1980 void __init page_alloc_init_late(void)
1985 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1987 /* There will be num_node_state(N_MEMORY) threads */
1988 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1989 for_each_node_state(nid, N_MEMORY) {
1990 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1993 /* Block until all are initialised */
1994 wait_for_completion(&pgdat_init_all_done_comp);
1997 * The number of managed pages has changed due to the initialisation
1998 * so the pcpu batch and high limits needs to be updated or the limits
1999 * will be artificially small.
2001 for_each_populated_zone(zone)
2002 zone_pcp_update(zone);
2005 * We initialized the rest of the deferred pages. Permanently disable
2006 * on-demand struct page initialization.
2008 static_branch_disable(&deferred_pages);
2010 /* Reinit limits that are based on free pages after the kernel is up */
2011 files_maxfiles_init();
2014 /* Discard memblock private memory */
2017 for_each_node_state(nid, N_MEMORY)
2018 shuffle_free_memory(NODE_DATA(nid));
2020 for_each_populated_zone(zone)
2021 set_zone_contiguous(zone);
2025 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2026 void __init init_cma_reserved_pageblock(struct page *page)
2028 unsigned i = pageblock_nr_pages;
2029 struct page *p = page;
2032 __ClearPageReserved(p);
2033 set_page_count(p, 0);
2036 set_pageblock_migratetype(page, MIGRATE_CMA);
2038 if (pageblock_order >= MAX_ORDER) {
2039 i = pageblock_nr_pages;
2042 set_page_refcounted(p);
2043 __free_pages(p, MAX_ORDER - 1);
2044 p += MAX_ORDER_NR_PAGES;
2045 } while (i -= MAX_ORDER_NR_PAGES);
2047 set_page_refcounted(page);
2048 __free_pages(page, pageblock_order);
2051 adjust_managed_page_count(page, pageblock_nr_pages);
2056 * The order of subdivision here is critical for the IO subsystem.
2057 * Please do not alter this order without good reasons and regression
2058 * testing. Specifically, as large blocks of memory are subdivided,
2059 * the order in which smaller blocks are delivered depends on the order
2060 * they're subdivided in this function. This is the primary factor
2061 * influencing the order in which pages are delivered to the IO
2062 * subsystem according to empirical testing, and this is also justified
2063 * by considering the behavior of a buddy system containing a single
2064 * large block of memory acted on by a series of small allocations.
2065 * This behavior is a critical factor in sglist merging's success.
2069 static inline void expand(struct zone *zone, struct page *page,
2070 int low, int high, int migratetype)
2072 unsigned long size = 1 << high;
2074 while (high > low) {
2077 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2080 * Mark as guard pages (or page), that will allow to
2081 * merge back to allocator when buddy will be freed.
2082 * Corresponding page table entries will not be touched,
2083 * pages will stay not present in virtual address space
2085 if (set_page_guard(zone, &page[size], high, migratetype))
2088 add_to_free_list(&page[size], zone, high, migratetype);
2089 set_page_order(&page[size], high);
2093 static void check_new_page_bad(struct page *page)
2095 const char *bad_reason = NULL;
2096 unsigned long bad_flags = 0;
2098 if (unlikely(atomic_read(&page->_mapcount) != -1))
2099 bad_reason = "nonzero mapcount";
2100 if (unlikely(page->mapping != NULL))
2101 bad_reason = "non-NULL mapping";
2102 if (unlikely(page_ref_count(page) != 0))
2103 bad_reason = "nonzero _refcount";
2104 if (unlikely(page->flags & __PG_HWPOISON)) {
2105 bad_reason = "HWPoisoned (hardware-corrupted)";
2106 bad_flags = __PG_HWPOISON;
2107 /* Don't complain about hwpoisoned pages */
2108 page_mapcount_reset(page); /* remove PageBuddy */
2111 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
2112 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
2113 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
2116 if (unlikely(page->mem_cgroup))
2117 bad_reason = "page still charged to cgroup";
2119 bad_page(page, bad_reason, bad_flags);
2123 * This page is about to be returned from the page allocator
2125 static inline int check_new_page(struct page *page)
2127 if (likely(page_expected_state(page,
2128 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2131 check_new_page_bad(page);
2135 static inline bool free_pages_prezeroed(void)
2137 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2138 page_poisoning_enabled()) || want_init_on_free();
2141 #ifdef CONFIG_DEBUG_VM
2143 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2144 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2145 * also checked when pcp lists are refilled from the free lists.
2147 static inline bool check_pcp_refill(struct page *page)
2149 if (debug_pagealloc_enabled_static())
2150 return check_new_page(page);
2155 static inline bool check_new_pcp(struct page *page)
2157 return check_new_page(page);
2161 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2162 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2163 * enabled, they are also checked when being allocated from the pcp lists.
2165 static inline bool check_pcp_refill(struct page *page)
2167 return check_new_page(page);
2169 static inline bool check_new_pcp(struct page *page)
2171 if (debug_pagealloc_enabled_static())
2172 return check_new_page(page);
2176 #endif /* CONFIG_DEBUG_VM */
2178 static bool check_new_pages(struct page *page, unsigned int order)
2181 for (i = 0; i < (1 << order); i++) {
2182 struct page *p = page + i;
2184 if (unlikely(check_new_page(p)))
2191 inline void post_alloc_hook(struct page *page, unsigned int order,
2194 set_page_private(page, 0);
2195 set_page_refcounted(page);
2197 arch_alloc_page(page, order);
2198 if (debug_pagealloc_enabled_static())
2199 kernel_map_pages(page, 1 << order, 1);
2200 kasan_alloc_pages(page, order);
2201 kernel_poison_pages(page, 1 << order, 1);
2202 set_page_owner(page, order, gfp_flags);
2205 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2206 unsigned int alloc_flags)
2208 post_alloc_hook(page, order, gfp_flags);
2210 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2211 kernel_init_free_pages(page, 1 << order);
2213 if (order && (gfp_flags & __GFP_COMP))
2214 prep_compound_page(page, order);
2217 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2218 * allocate the page. The expectation is that the caller is taking
2219 * steps that will free more memory. The caller should avoid the page
2220 * being used for !PFMEMALLOC purposes.
2222 if (alloc_flags & ALLOC_NO_WATERMARKS)
2223 set_page_pfmemalloc(page);
2225 clear_page_pfmemalloc(page);
2229 * Go through the free lists for the given migratetype and remove
2230 * the smallest available page from the freelists
2232 static __always_inline
2233 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2236 unsigned int current_order;
2237 struct free_area *area;
2240 /* Find a page of the appropriate size in the preferred list */
2241 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2242 area = &(zone->free_area[current_order]);
2243 page = get_page_from_free_area(area, migratetype);
2246 del_page_from_free_list(page, zone, current_order);
2247 expand(zone, page, order, current_order, migratetype);
2248 set_pcppage_migratetype(page, migratetype);
2257 * This array describes the order lists are fallen back to when
2258 * the free lists for the desirable migrate type are depleted
2260 static int fallbacks[MIGRATE_TYPES][4] = {
2261 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2262 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2263 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2265 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2267 #ifdef CONFIG_MEMORY_ISOLATION
2268 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2273 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2276 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2279 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2280 unsigned int order) { return NULL; }
2284 * Move the free pages in a range to the free lists of the requested type.
2285 * Note that start_page and end_pages are not aligned on a pageblock
2286 * boundary. If alignment is required, use move_freepages_block()
2288 static int move_freepages(struct zone *zone,
2289 struct page *start_page, struct page *end_page,
2290 int migratetype, int *num_movable)
2294 int pages_moved = 0;
2296 for (page = start_page; page <= end_page;) {
2297 if (!pfn_valid_within(page_to_pfn(page))) {
2302 if (!PageBuddy(page)) {
2304 * We assume that pages that could be isolated for
2305 * migration are movable. But we don't actually try
2306 * isolating, as that would be expensive.
2309 (PageLRU(page) || __PageMovable(page)))
2316 /* Make sure we are not inadvertently changing nodes */
2317 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2318 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2320 order = page_order(page);
2321 move_to_free_list(page, zone, order, migratetype);
2323 pages_moved += 1 << order;
2329 int move_freepages_block(struct zone *zone, struct page *page,
2330 int migratetype, int *num_movable)
2332 unsigned long start_pfn, end_pfn;
2333 struct page *start_page, *end_page;
2338 start_pfn = page_to_pfn(page);
2339 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2340 start_page = pfn_to_page(start_pfn);
2341 end_page = start_page + pageblock_nr_pages - 1;
2342 end_pfn = start_pfn + pageblock_nr_pages - 1;
2344 /* Do not cross zone boundaries */
2345 if (!zone_spans_pfn(zone, start_pfn))
2347 if (!zone_spans_pfn(zone, end_pfn))
2350 return move_freepages(zone, start_page, end_page, migratetype,
2354 static void change_pageblock_range(struct page *pageblock_page,
2355 int start_order, int migratetype)
2357 int nr_pageblocks = 1 << (start_order - pageblock_order);
2359 while (nr_pageblocks--) {
2360 set_pageblock_migratetype(pageblock_page, migratetype);
2361 pageblock_page += pageblock_nr_pages;
2366 * When we are falling back to another migratetype during allocation, try to
2367 * steal extra free pages from the same pageblocks to satisfy further
2368 * allocations, instead of polluting multiple pageblocks.
2370 * If we are stealing a relatively large buddy page, it is likely there will
2371 * be more free pages in the pageblock, so try to steal them all. For
2372 * reclaimable and unmovable allocations, we steal regardless of page size,
2373 * as fragmentation caused by those allocations polluting movable pageblocks
2374 * is worse than movable allocations stealing from unmovable and reclaimable
2377 static bool can_steal_fallback(unsigned int order, int start_mt)
2380 * Leaving this order check is intended, although there is
2381 * relaxed order check in next check. The reason is that
2382 * we can actually steal whole pageblock if this condition met,
2383 * but, below check doesn't guarantee it and that is just heuristic
2384 * so could be changed anytime.
2386 if (order >= pageblock_order)
2389 if (order >= pageblock_order / 2 ||
2390 start_mt == MIGRATE_RECLAIMABLE ||
2391 start_mt == MIGRATE_UNMOVABLE ||
2392 page_group_by_mobility_disabled)
2398 static inline void boost_watermark(struct zone *zone)
2400 unsigned long max_boost;
2402 if (!watermark_boost_factor)
2405 * Don't bother in zones that are unlikely to produce results.
2406 * On small machines, including kdump capture kernels running
2407 * in a small area, boosting the watermark can cause an out of
2408 * memory situation immediately.
2410 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2413 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2414 watermark_boost_factor, 10000);
2417 * high watermark may be uninitialised if fragmentation occurs
2418 * very early in boot so do not boost. We do not fall
2419 * through and boost by pageblock_nr_pages as failing
2420 * allocations that early means that reclaim is not going
2421 * to help and it may even be impossible to reclaim the
2422 * boosted watermark resulting in a hang.
2427 max_boost = max(pageblock_nr_pages, max_boost);
2429 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2434 * This function implements actual steal behaviour. If order is large enough,
2435 * we can steal whole pageblock. If not, we first move freepages in this
2436 * pageblock to our migratetype and determine how many already-allocated pages
2437 * are there in the pageblock with a compatible migratetype. If at least half
2438 * of pages are free or compatible, we can change migratetype of the pageblock
2439 * itself, so pages freed in the future will be put on the correct free list.
2441 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2442 unsigned int alloc_flags, int start_type, bool whole_block)
2444 unsigned int current_order = page_order(page);
2445 int free_pages, movable_pages, alike_pages;
2448 old_block_type = get_pageblock_migratetype(page);
2451 * This can happen due to races and we want to prevent broken
2452 * highatomic accounting.
2454 if (is_migrate_highatomic(old_block_type))
2457 /* Take ownership for orders >= pageblock_order */
2458 if (current_order >= pageblock_order) {
2459 change_pageblock_range(page, current_order, start_type);
2464 * Boost watermarks to increase reclaim pressure to reduce the
2465 * likelihood of future fallbacks. Wake kswapd now as the node
2466 * may be balanced overall and kswapd will not wake naturally.
2468 boost_watermark(zone);
2469 if (alloc_flags & ALLOC_KSWAPD)
2470 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2472 /* We are not allowed to try stealing from the whole block */
2476 free_pages = move_freepages_block(zone, page, start_type,
2479 * Determine how many pages are compatible with our allocation.
2480 * For movable allocation, it's the number of movable pages which
2481 * we just obtained. For other types it's a bit more tricky.
2483 if (start_type == MIGRATE_MOVABLE) {
2484 alike_pages = movable_pages;
2487 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2488 * to MOVABLE pageblock, consider all non-movable pages as
2489 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2490 * vice versa, be conservative since we can't distinguish the
2491 * exact migratetype of non-movable pages.
2493 if (old_block_type == MIGRATE_MOVABLE)
2494 alike_pages = pageblock_nr_pages
2495 - (free_pages + movable_pages);
2500 /* moving whole block can fail due to zone boundary conditions */
2505 * If a sufficient number of pages in the block are either free or of
2506 * comparable migratability as our allocation, claim the whole block.
2508 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2509 page_group_by_mobility_disabled)
2510 set_pageblock_migratetype(page, start_type);
2515 move_to_free_list(page, zone, current_order, start_type);
2519 * Check whether there is a suitable fallback freepage with requested order.
2520 * If only_stealable is true, this function returns fallback_mt only if
2521 * we can steal other freepages all together. This would help to reduce
2522 * fragmentation due to mixed migratetype pages in one pageblock.
2524 int find_suitable_fallback(struct free_area *area, unsigned int order,
2525 int migratetype, bool only_stealable, bool *can_steal)
2530 if (area->nr_free == 0)
2535 fallback_mt = fallbacks[migratetype][i];
2536 if (fallback_mt == MIGRATE_TYPES)
2539 if (free_area_empty(area, fallback_mt))
2542 if (can_steal_fallback(order, migratetype))
2545 if (!only_stealable)
2556 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2557 * there are no empty page blocks that contain a page with a suitable order
2559 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2560 unsigned int alloc_order)
2563 unsigned long max_managed, flags;
2566 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2567 * Check is race-prone but harmless.
2569 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2570 if (zone->nr_reserved_highatomic >= max_managed)
2573 spin_lock_irqsave(&zone->lock, flags);
2575 /* Recheck the nr_reserved_highatomic limit under the lock */
2576 if (zone->nr_reserved_highatomic >= max_managed)
2580 mt = get_pageblock_migratetype(page);
2581 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2582 && !is_migrate_cma(mt)) {
2583 zone->nr_reserved_highatomic += pageblock_nr_pages;
2584 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2585 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2589 spin_unlock_irqrestore(&zone->lock, flags);
2593 * Used when an allocation is about to fail under memory pressure. This
2594 * potentially hurts the reliability of high-order allocations when under
2595 * intense memory pressure but failed atomic allocations should be easier
2596 * to recover from than an OOM.
2598 * If @force is true, try to unreserve a pageblock even though highatomic
2599 * pageblock is exhausted.
2601 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2604 struct zonelist *zonelist = ac->zonelist;
2605 unsigned long flags;
2612 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2615 * Preserve at least one pageblock unless memory pressure
2618 if (!force && zone->nr_reserved_highatomic <=
2622 spin_lock_irqsave(&zone->lock, flags);
2623 for (order = 0; order < MAX_ORDER; order++) {
2624 struct free_area *area = &(zone->free_area[order]);
2626 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2631 * In page freeing path, migratetype change is racy so
2632 * we can counter several free pages in a pageblock
2633 * in this loop althoug we changed the pageblock type
2634 * from highatomic to ac->migratetype. So we should
2635 * adjust the count once.
2637 if (is_migrate_highatomic_page(page)) {
2639 * It should never happen but changes to
2640 * locking could inadvertently allow a per-cpu
2641 * drain to add pages to MIGRATE_HIGHATOMIC
2642 * while unreserving so be safe and watch for
2645 zone->nr_reserved_highatomic -= min(
2647 zone->nr_reserved_highatomic);
2651 * Convert to ac->migratetype and avoid the normal
2652 * pageblock stealing heuristics. Minimally, the caller
2653 * is doing the work and needs the pages. More
2654 * importantly, if the block was always converted to
2655 * MIGRATE_UNMOVABLE or another type then the number
2656 * of pageblocks that cannot be completely freed
2659 set_pageblock_migratetype(page, ac->migratetype);
2660 ret = move_freepages_block(zone, page, ac->migratetype,
2663 spin_unlock_irqrestore(&zone->lock, flags);
2667 spin_unlock_irqrestore(&zone->lock, flags);
2674 * Try finding a free buddy page on the fallback list and put it on the free
2675 * list of requested migratetype, possibly along with other pages from the same
2676 * block, depending on fragmentation avoidance heuristics. Returns true if
2677 * fallback was found so that __rmqueue_smallest() can grab it.
2679 * The use of signed ints for order and current_order is a deliberate
2680 * deviation from the rest of this file, to make the for loop
2681 * condition simpler.
2683 static __always_inline bool
2684 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2685 unsigned int alloc_flags)
2687 struct free_area *area;
2689 int min_order = order;
2695 * Do not steal pages from freelists belonging to other pageblocks
2696 * i.e. orders < pageblock_order. If there are no local zones free,
2697 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2699 if (alloc_flags & ALLOC_NOFRAGMENT)
2700 min_order = pageblock_order;
2703 * Find the largest available free page in the other list. This roughly
2704 * approximates finding the pageblock with the most free pages, which
2705 * would be too costly to do exactly.
2707 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2709 area = &(zone->free_area[current_order]);
2710 fallback_mt = find_suitable_fallback(area, current_order,
2711 start_migratetype, false, &can_steal);
2712 if (fallback_mt == -1)
2716 * We cannot steal all free pages from the pageblock and the
2717 * requested migratetype is movable. In that case it's better to
2718 * steal and split the smallest available page instead of the
2719 * largest available page, because even if the next movable
2720 * allocation falls back into a different pageblock than this
2721 * one, it won't cause permanent fragmentation.
2723 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2724 && current_order > order)
2733 for (current_order = order; current_order < MAX_ORDER;
2735 area = &(zone->free_area[current_order]);
2736 fallback_mt = find_suitable_fallback(area, current_order,
2737 start_migratetype, false, &can_steal);
2738 if (fallback_mt != -1)
2743 * This should not happen - we already found a suitable fallback
2744 * when looking for the largest page.
2746 VM_BUG_ON(current_order == MAX_ORDER);
2749 page = get_page_from_free_area(area, fallback_mt);
2751 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2754 trace_mm_page_alloc_extfrag(page, order, current_order,
2755 start_migratetype, fallback_mt);
2762 * Do the hard work of removing an element from the buddy allocator.
2763 * Call me with the zone->lock already held.
2765 static __always_inline struct page *
2766 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2767 unsigned int alloc_flags)
2772 page = __rmqueue_smallest(zone, order, migratetype);
2773 if (unlikely(!page)) {
2774 if (migratetype == MIGRATE_MOVABLE)
2775 page = __rmqueue_cma_fallback(zone, order);
2777 if (!page && __rmqueue_fallback(zone, order, migratetype,
2782 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2787 * Obtain a specified number of elements from the buddy allocator, all under
2788 * a single hold of the lock, for efficiency. Add them to the supplied list.
2789 * Returns the number of new pages which were placed at *list.
2791 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2792 unsigned long count, struct list_head *list,
2793 int migratetype, unsigned int alloc_flags)
2797 spin_lock(&zone->lock);
2798 for (i = 0; i < count; ++i) {
2799 struct page *page = __rmqueue(zone, order, migratetype,
2801 if (unlikely(page == NULL))
2804 if (unlikely(check_pcp_refill(page)))
2808 * Split buddy pages returned by expand() are received here in
2809 * physical page order. The page is added to the tail of
2810 * caller's list. From the callers perspective, the linked list
2811 * is ordered by page number under some conditions. This is
2812 * useful for IO devices that can forward direction from the
2813 * head, thus also in the physical page order. This is useful
2814 * for IO devices that can merge IO requests if the physical
2815 * pages are ordered properly.
2817 list_add_tail(&page->lru, list);
2819 if (is_migrate_cma(get_pcppage_migratetype(page)))
2820 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2825 * i pages were removed from the buddy list even if some leak due
2826 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2827 * on i. Do not confuse with 'alloced' which is the number of
2828 * pages added to the pcp list.
2830 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2831 spin_unlock(&zone->lock);
2837 * Called from the vmstat counter updater to drain pagesets of this
2838 * currently executing processor on remote nodes after they have
2841 * Note that this function must be called with the thread pinned to
2842 * a single processor.
2844 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2846 unsigned long flags;
2847 int to_drain, batch;
2849 local_irq_save(flags);
2850 batch = READ_ONCE(pcp->batch);
2851 to_drain = min(pcp->count, batch);
2853 free_pcppages_bulk(zone, to_drain, pcp);
2854 local_irq_restore(flags);
2859 * Drain pcplists of the indicated processor and zone.
2861 * The processor must either be the current processor and the
2862 * thread pinned to the current processor or a processor that
2865 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2867 unsigned long flags;
2868 struct per_cpu_pageset *pset;
2869 struct per_cpu_pages *pcp;
2871 local_irq_save(flags);
2872 pset = per_cpu_ptr(zone->pageset, cpu);
2876 free_pcppages_bulk(zone, pcp->count, pcp);
2877 local_irq_restore(flags);
2881 * Drain pcplists of all zones on the indicated processor.
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(unsigned int cpu)
2891 for_each_populated_zone(zone) {
2892 drain_pages_zone(cpu, zone);
2897 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2899 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2900 * the single zone's pages.
2902 void drain_local_pages(struct zone *zone)
2904 int cpu = smp_processor_id();
2907 drain_pages_zone(cpu, zone);
2912 static void drain_local_pages_wq(struct work_struct *work)
2914 struct pcpu_drain *drain;
2916 drain = container_of(work, struct pcpu_drain, work);
2919 * drain_all_pages doesn't use proper cpu hotplug protection so
2920 * we can race with cpu offline when the WQ can move this from
2921 * a cpu pinned worker to an unbound one. We can operate on a different
2922 * cpu which is allright but we also have to make sure to not move to
2926 drain_local_pages(drain->zone);
2931 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2933 * When zone parameter is non-NULL, spill just the single zone's pages.
2935 * Note that this can be extremely slow as the draining happens in a workqueue.
2937 void drain_all_pages(struct zone *zone)
2942 * Allocate in the BSS so we wont require allocation in
2943 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2945 static cpumask_t cpus_with_pcps;
2948 * Make sure nobody triggers this path before mm_percpu_wq is fully
2951 if (WARN_ON_ONCE(!mm_percpu_wq))
2955 * Do not drain if one is already in progress unless it's specific to
2956 * a zone. Such callers are primarily CMA and memory hotplug and need
2957 * the drain to be complete when the call returns.
2959 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2962 mutex_lock(&pcpu_drain_mutex);
2966 * We don't care about racing with CPU hotplug event
2967 * as offline notification will cause the notified
2968 * cpu to drain that CPU pcps and on_each_cpu_mask
2969 * disables preemption as part of its processing
2971 for_each_online_cpu(cpu) {
2972 struct per_cpu_pageset *pcp;
2974 bool has_pcps = false;
2977 pcp = per_cpu_ptr(zone->pageset, cpu);
2981 for_each_populated_zone(z) {
2982 pcp = per_cpu_ptr(z->pageset, cpu);
2983 if (pcp->pcp.count) {
2991 cpumask_set_cpu(cpu, &cpus_with_pcps);
2993 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2996 for_each_cpu(cpu, &cpus_with_pcps) {
2997 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3000 INIT_WORK(&drain->work, drain_local_pages_wq);
3001 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3003 for_each_cpu(cpu, &cpus_with_pcps)
3004 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3006 mutex_unlock(&pcpu_drain_mutex);
3009 #ifdef CONFIG_HIBERNATION
3012 * Touch the watchdog for every WD_PAGE_COUNT pages.
3014 #define WD_PAGE_COUNT (128*1024)
3016 void mark_free_pages(struct zone *zone)
3018 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3019 unsigned long flags;
3020 unsigned int order, t;
3023 if (zone_is_empty(zone))
3026 spin_lock_irqsave(&zone->lock, flags);
3028 max_zone_pfn = zone_end_pfn(zone);
3029 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3030 if (pfn_valid(pfn)) {
3031 page = pfn_to_page(pfn);
3033 if (!--page_count) {
3034 touch_nmi_watchdog();
3035 page_count = WD_PAGE_COUNT;
3038 if (page_zone(page) != zone)
3041 if (!swsusp_page_is_forbidden(page))
3042 swsusp_unset_page_free(page);
3045 for_each_migratetype_order(order, t) {
3046 list_for_each_entry(page,
3047 &zone->free_area[order].free_list[t], lru) {
3050 pfn = page_to_pfn(page);
3051 for (i = 0; i < (1UL << order); i++) {
3052 if (!--page_count) {
3053 touch_nmi_watchdog();
3054 page_count = WD_PAGE_COUNT;
3056 swsusp_set_page_free(pfn_to_page(pfn + i));
3060 spin_unlock_irqrestore(&zone->lock, flags);
3062 #endif /* CONFIG_PM */
3064 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3068 if (!free_pcp_prepare(page))
3071 migratetype = get_pfnblock_migratetype(page, pfn);
3072 set_pcppage_migratetype(page, migratetype);
3076 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3078 struct zone *zone = page_zone(page);
3079 struct per_cpu_pages *pcp;
3082 migratetype = get_pcppage_migratetype(page);
3083 __count_vm_event(PGFREE);
3086 * We only track unmovable, reclaimable and movable on pcp lists.
3087 * Free ISOLATE pages back to the allocator because they are being
3088 * offlined but treat HIGHATOMIC as movable pages so we can get those
3089 * areas back if necessary. Otherwise, we may have to free
3090 * excessively into the page allocator
3092 if (migratetype >= MIGRATE_PCPTYPES) {
3093 if (unlikely(is_migrate_isolate(migratetype))) {
3094 free_one_page(zone, page, pfn, 0, migratetype);
3097 migratetype = MIGRATE_MOVABLE;
3100 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3101 list_add(&page->lru, &pcp->lists[migratetype]);
3103 if (pcp->count >= pcp->high) {
3104 unsigned long batch = READ_ONCE(pcp->batch);
3105 free_pcppages_bulk(zone, batch, pcp);
3110 * Free a 0-order page
3112 void free_unref_page(struct page *page)
3114 unsigned long flags;
3115 unsigned long pfn = page_to_pfn(page);
3117 if (!free_unref_page_prepare(page, pfn))
3120 local_irq_save(flags);
3121 free_unref_page_commit(page, pfn);
3122 local_irq_restore(flags);
3126 * Free a list of 0-order pages
3128 void free_unref_page_list(struct list_head *list)
3130 struct page *page, *next;
3131 unsigned long flags, pfn;
3132 int batch_count = 0;
3134 /* Prepare pages for freeing */
3135 list_for_each_entry_safe(page, next, list, lru) {
3136 pfn = page_to_pfn(page);
3137 if (!free_unref_page_prepare(page, pfn))
3138 list_del(&page->lru);
3139 set_page_private(page, pfn);
3142 local_irq_save(flags);
3143 list_for_each_entry_safe(page, next, list, lru) {
3144 unsigned long pfn = page_private(page);
3146 set_page_private(page, 0);
3147 trace_mm_page_free_batched(page);
3148 free_unref_page_commit(page, pfn);
3151 * Guard against excessive IRQ disabled times when we get
3152 * a large list of pages to free.
3154 if (++batch_count == SWAP_CLUSTER_MAX) {
3155 local_irq_restore(flags);
3157 local_irq_save(flags);
3160 local_irq_restore(flags);
3164 * split_page takes a non-compound higher-order page, and splits it into
3165 * n (1<<order) sub-pages: page[0..n]
3166 * Each sub-page must be freed individually.
3168 * Note: this is probably too low level an operation for use in drivers.
3169 * Please consult with lkml before using this in your driver.
3171 void split_page(struct page *page, unsigned int order)
3175 VM_BUG_ON_PAGE(PageCompound(page), page);
3176 VM_BUG_ON_PAGE(!page_count(page), page);
3178 for (i = 1; i < (1 << order); i++)
3179 set_page_refcounted(page + i);
3180 split_page_owner(page, order);
3182 EXPORT_SYMBOL_GPL(split_page);
3184 int __isolate_free_page(struct page *page, unsigned int order)
3186 unsigned long watermark;
3190 BUG_ON(!PageBuddy(page));
3192 zone = page_zone(page);
3193 mt = get_pageblock_migratetype(page);
3195 if (!is_migrate_isolate(mt)) {
3197 * Obey watermarks as if the page was being allocated. We can
3198 * emulate a high-order watermark check with a raised order-0
3199 * watermark, because we already know our high-order page
3202 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3203 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3206 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3209 /* Remove page from free list */
3211 del_page_from_free_list(page, zone, order);
3214 * Set the pageblock if the isolated page is at least half of a
3217 if (order >= pageblock_order - 1) {
3218 struct page *endpage = page + (1 << order) - 1;
3219 for (; page < endpage; page += pageblock_nr_pages) {
3220 int mt = get_pageblock_migratetype(page);
3221 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3222 && !is_migrate_highatomic(mt))
3223 set_pageblock_migratetype(page,
3229 return 1UL << order;
3233 * __putback_isolated_page - Return a now-isolated page back where we got it
3234 * @page: Page that was isolated
3235 * @order: Order of the isolated page
3236 * @mt: The page's pageblock's migratetype
3238 * This function is meant to return a page pulled from the free lists via
3239 * __isolate_free_page back to the free lists they were pulled from.
3241 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3243 struct zone *zone = page_zone(page);
3245 /* zone lock should be held when this function is called */
3246 lockdep_assert_held(&zone->lock);
3248 /* Return isolated page to tail of freelist. */
3249 __free_one_page(page, page_to_pfn(page), zone, order, mt, false);
3253 * Update NUMA hit/miss statistics
3255 * Must be called with interrupts disabled.
3257 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3260 enum numa_stat_item local_stat = NUMA_LOCAL;
3262 /* skip numa counters update if numa stats is disabled */
3263 if (!static_branch_likely(&vm_numa_stat_key))
3266 if (zone_to_nid(z) != numa_node_id())
3267 local_stat = NUMA_OTHER;
3269 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3270 __inc_numa_state(z, NUMA_HIT);
3272 __inc_numa_state(z, NUMA_MISS);
3273 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3275 __inc_numa_state(z, local_stat);
3279 /* Remove page from the per-cpu list, caller must protect the list */
3280 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3281 unsigned int alloc_flags,
3282 struct per_cpu_pages *pcp,
3283 struct list_head *list)
3288 if (list_empty(list)) {
3289 pcp->count += rmqueue_bulk(zone, 0,
3291 migratetype, alloc_flags);
3292 if (unlikely(list_empty(list)))
3296 page = list_first_entry(list, struct page, lru);
3297 list_del(&page->lru);
3299 } while (check_new_pcp(page));
3304 /* Lock and remove page from the per-cpu list */
3305 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3306 struct zone *zone, gfp_t gfp_flags,
3307 int migratetype, unsigned int alloc_flags)
3309 struct per_cpu_pages *pcp;
3310 struct list_head *list;
3312 unsigned long flags;
3314 local_irq_save(flags);
3315 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3316 list = &pcp->lists[migratetype];
3317 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3319 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3320 zone_statistics(preferred_zone, zone);
3322 local_irq_restore(flags);
3327 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3330 struct page *rmqueue(struct zone *preferred_zone,
3331 struct zone *zone, unsigned int order,
3332 gfp_t gfp_flags, unsigned int alloc_flags,
3335 unsigned long flags;
3338 if (likely(order == 0)) {
3339 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3340 migratetype, alloc_flags);
3345 * We most definitely don't want callers attempting to
3346 * allocate greater than order-1 page units with __GFP_NOFAIL.
3348 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3349 spin_lock_irqsave(&zone->lock, flags);
3353 if (alloc_flags & ALLOC_HARDER) {
3354 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3356 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3359 page = __rmqueue(zone, order, migratetype, alloc_flags);
3360 } while (page && check_new_pages(page, order));
3361 spin_unlock(&zone->lock);
3364 __mod_zone_freepage_state(zone, -(1 << order),
3365 get_pcppage_migratetype(page));
3367 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3368 zone_statistics(preferred_zone, zone);
3369 local_irq_restore(flags);
3372 /* Separate test+clear to avoid unnecessary atomics */
3373 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3374 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3375 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3378 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3382 local_irq_restore(flags);
3386 #ifdef CONFIG_FAIL_PAGE_ALLOC
3389 struct fault_attr attr;
3391 bool ignore_gfp_highmem;
3392 bool ignore_gfp_reclaim;
3394 } fail_page_alloc = {
3395 .attr = FAULT_ATTR_INITIALIZER,
3396 .ignore_gfp_reclaim = true,
3397 .ignore_gfp_highmem = true,
3401 static int __init setup_fail_page_alloc(char *str)
3403 return setup_fault_attr(&fail_page_alloc.attr, str);
3405 __setup("fail_page_alloc=", setup_fail_page_alloc);
3407 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3409 if (order < fail_page_alloc.min_order)
3411 if (gfp_mask & __GFP_NOFAIL)
3413 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3415 if (fail_page_alloc.ignore_gfp_reclaim &&
3416 (gfp_mask & __GFP_DIRECT_RECLAIM))
3419 return should_fail(&fail_page_alloc.attr, 1 << order);
3422 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3424 static int __init fail_page_alloc_debugfs(void)
3426 umode_t mode = S_IFREG | 0600;
3429 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3430 &fail_page_alloc.attr);
3432 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3433 &fail_page_alloc.ignore_gfp_reclaim);
3434 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3435 &fail_page_alloc.ignore_gfp_highmem);
3436 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3441 late_initcall(fail_page_alloc_debugfs);
3443 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3445 #else /* CONFIG_FAIL_PAGE_ALLOC */
3447 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3452 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3454 static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3456 return __should_fail_alloc_page(gfp_mask, order);
3458 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3461 * Return true if free base pages are above 'mark'. For high-order checks it
3462 * will return true of the order-0 watermark is reached and there is at least
3463 * one free page of a suitable size. Checking now avoids taking the zone lock
3464 * to check in the allocation paths if no pages are free.
3466 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3467 int classzone_idx, unsigned int alloc_flags,
3472 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3474 /* free_pages may go negative - that's OK */
3475 free_pages -= (1 << order) - 1;
3477 if (alloc_flags & ALLOC_HIGH)
3481 * If the caller does not have rights to ALLOC_HARDER then subtract
3482 * the high-atomic reserves. This will over-estimate the size of the
3483 * atomic reserve but it avoids a search.
3485 if (likely(!alloc_harder)) {
3486 free_pages -= z->nr_reserved_highatomic;
3489 * OOM victims can try even harder than normal ALLOC_HARDER
3490 * users on the grounds that it's definitely going to be in
3491 * the exit path shortly and free memory. Any allocation it
3492 * makes during the free path will be small and short-lived.
3494 if (alloc_flags & ALLOC_OOM)
3502 /* If allocation can't use CMA areas don't use free CMA pages */
3503 if (!(alloc_flags & ALLOC_CMA))
3504 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3508 * Check watermarks for an order-0 allocation request. If these
3509 * are not met, then a high-order request also cannot go ahead
3510 * even if a suitable page happened to be free.
3512 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3515 /* If this is an order-0 request then the watermark is fine */
3519 /* For a high-order request, check at least one suitable page is free */
3520 for (o = order; o < MAX_ORDER; o++) {
3521 struct free_area *area = &z->free_area[o];
3527 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3528 if (!free_area_empty(area, mt))
3533 if ((alloc_flags & ALLOC_CMA) &&
3534 !free_area_empty(area, MIGRATE_CMA)) {
3538 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3544 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3545 int classzone_idx, unsigned int alloc_flags)
3547 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3548 zone_page_state(z, NR_FREE_PAGES));
3551 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3552 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3554 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3558 /* If allocation can't use CMA areas don't use free CMA pages */
3559 if (!(alloc_flags & ALLOC_CMA))
3560 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3564 * Fast check for order-0 only. If this fails then the reserves
3565 * need to be calculated. There is a corner case where the check
3566 * passes but only the high-order atomic reserve are free. If
3567 * the caller is !atomic then it'll uselessly search the free
3568 * list. That corner case is then slower but it is harmless.
3570 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3573 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3577 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3578 unsigned long mark, int classzone_idx)
3580 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3582 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3583 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3585 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3590 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3592 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3593 node_reclaim_distance;
3595 #else /* CONFIG_NUMA */
3596 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3600 #endif /* CONFIG_NUMA */
3603 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3604 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3605 * premature use of a lower zone may cause lowmem pressure problems that
3606 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3607 * probably too small. It only makes sense to spread allocations to avoid
3608 * fragmentation between the Normal and DMA32 zones.
3610 static inline unsigned int
3611 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3613 unsigned int alloc_flags;
3616 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3619 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3621 #ifdef CONFIG_ZONE_DMA32
3625 if (zone_idx(zone) != ZONE_NORMAL)
3629 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3630 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3631 * on UMA that if Normal is populated then so is DMA32.
3633 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3634 if (nr_online_nodes > 1 && !populated_zone(--zone))
3637 alloc_flags |= ALLOC_NOFRAGMENT;
3638 #endif /* CONFIG_ZONE_DMA32 */
3643 * get_page_from_freelist goes through the zonelist trying to allocate
3646 static struct page *
3647 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3648 const struct alloc_context *ac)
3652 struct pglist_data *last_pgdat_dirty_limit = NULL;
3657 * Scan zonelist, looking for a zone with enough free.
3658 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3660 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3661 z = ac->preferred_zoneref;
3662 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3667 if (cpusets_enabled() &&
3668 (alloc_flags & ALLOC_CPUSET) &&
3669 !__cpuset_zone_allowed(zone, gfp_mask))
3672 * When allocating a page cache page for writing, we
3673 * want to get it from a node that is within its dirty
3674 * limit, such that no single node holds more than its
3675 * proportional share of globally allowed dirty pages.
3676 * The dirty limits take into account the node's
3677 * lowmem reserves and high watermark so that kswapd
3678 * should be able to balance it without having to
3679 * write pages from its LRU list.
3681 * XXX: For now, allow allocations to potentially
3682 * exceed the per-node dirty limit in the slowpath
3683 * (spread_dirty_pages unset) before going into reclaim,
3684 * which is important when on a NUMA setup the allowed
3685 * nodes are together not big enough to reach the
3686 * global limit. The proper fix for these situations
3687 * will require awareness of nodes in the
3688 * dirty-throttling and the flusher threads.
3690 if (ac->spread_dirty_pages) {
3691 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3694 if (!node_dirty_ok(zone->zone_pgdat)) {
3695 last_pgdat_dirty_limit = zone->zone_pgdat;
3700 if (no_fallback && nr_online_nodes > 1 &&
3701 zone != ac->preferred_zoneref->zone) {
3705 * If moving to a remote node, retry but allow
3706 * fragmenting fallbacks. Locality is more important
3707 * than fragmentation avoidance.
3709 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3710 if (zone_to_nid(zone) != local_nid) {
3711 alloc_flags &= ~ALLOC_NOFRAGMENT;
3716 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3717 if (!zone_watermark_fast(zone, order, mark,
3718 ac_classzone_idx(ac), alloc_flags)) {
3721 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3723 * Watermark failed for this zone, but see if we can
3724 * grow this zone if it contains deferred pages.
3726 if (static_branch_unlikely(&deferred_pages)) {
3727 if (_deferred_grow_zone(zone, order))
3731 /* Checked here to keep the fast path fast */
3732 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3733 if (alloc_flags & ALLOC_NO_WATERMARKS)
3736 if (node_reclaim_mode == 0 ||
3737 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3740 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3742 case NODE_RECLAIM_NOSCAN:
3745 case NODE_RECLAIM_FULL:
3746 /* scanned but unreclaimable */
3749 /* did we reclaim enough */
3750 if (zone_watermark_ok(zone, order, mark,
3751 ac_classzone_idx(ac), alloc_flags))
3759 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3760 gfp_mask, alloc_flags, ac->migratetype);
3762 prep_new_page(page, order, gfp_mask, alloc_flags);
3765 * If this is a high-order atomic allocation then check
3766 * if the pageblock should be reserved for the future
3768 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3769 reserve_highatomic_pageblock(page, zone, order);
3773 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3774 /* Try again if zone has deferred pages */
3775 if (static_branch_unlikely(&deferred_pages)) {
3776 if (_deferred_grow_zone(zone, order))
3784 * It's possible on a UMA machine to get through all zones that are
3785 * fragmented. If avoiding fragmentation, reset and try again.
3788 alloc_flags &= ~ALLOC_NOFRAGMENT;
3795 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3797 unsigned int filter = SHOW_MEM_FILTER_NODES;
3800 * This documents exceptions given to allocations in certain
3801 * contexts that are allowed to allocate outside current's set
3804 if (!(gfp_mask & __GFP_NOMEMALLOC))
3805 if (tsk_is_oom_victim(current) ||
3806 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3807 filter &= ~SHOW_MEM_FILTER_NODES;
3808 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3809 filter &= ~SHOW_MEM_FILTER_NODES;
3811 show_mem(filter, nodemask);
3814 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3816 struct va_format vaf;
3818 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3820 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3823 va_start(args, fmt);
3826 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3827 current->comm, &vaf, gfp_mask, &gfp_mask,
3828 nodemask_pr_args(nodemask));
3831 cpuset_print_current_mems_allowed();
3834 warn_alloc_show_mem(gfp_mask, nodemask);
3837 static inline struct page *
3838 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3839 unsigned int alloc_flags,
3840 const struct alloc_context *ac)
3844 page = get_page_from_freelist(gfp_mask, order,
3845 alloc_flags|ALLOC_CPUSET, ac);
3847 * fallback to ignore cpuset restriction if our nodes
3851 page = get_page_from_freelist(gfp_mask, order,
3857 static inline struct page *
3858 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3859 const struct alloc_context *ac, unsigned long *did_some_progress)
3861 struct oom_control oc = {
3862 .zonelist = ac->zonelist,
3863 .nodemask = ac->nodemask,
3865 .gfp_mask = gfp_mask,
3870 *did_some_progress = 0;
3873 * Acquire the oom lock. If that fails, somebody else is
3874 * making progress for us.
3876 if (!mutex_trylock(&oom_lock)) {
3877 *did_some_progress = 1;
3878 schedule_timeout_uninterruptible(1);
3883 * Go through the zonelist yet one more time, keep very high watermark
3884 * here, this is only to catch a parallel oom killing, we must fail if
3885 * we're still under heavy pressure. But make sure that this reclaim
3886 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3887 * allocation which will never fail due to oom_lock already held.
3889 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3890 ~__GFP_DIRECT_RECLAIM, order,
3891 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3895 /* Coredumps can quickly deplete all memory reserves */
3896 if (current->flags & PF_DUMPCORE)
3898 /* The OOM killer will not help higher order allocs */
3899 if (order > PAGE_ALLOC_COSTLY_ORDER)
3902 * We have already exhausted all our reclaim opportunities without any
3903 * success so it is time to admit defeat. We will skip the OOM killer
3904 * because it is very likely that the caller has a more reasonable
3905 * fallback than shooting a random task.
3907 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3909 /* The OOM killer does not needlessly kill tasks for lowmem */
3910 if (ac->high_zoneidx < ZONE_NORMAL)
3912 if (pm_suspended_storage())
3915 * XXX: GFP_NOFS allocations should rather fail than rely on
3916 * other request to make a forward progress.
3917 * We are in an unfortunate situation where out_of_memory cannot
3918 * do much for this context but let's try it to at least get
3919 * access to memory reserved if the current task is killed (see
3920 * out_of_memory). Once filesystems are ready to handle allocation
3921 * failures more gracefully we should just bail out here.
3924 /* The OOM killer may not free memory on a specific node */
3925 if (gfp_mask & __GFP_THISNODE)
3928 /* Exhausted what can be done so it's blame time */
3929 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3930 *did_some_progress = 1;
3933 * Help non-failing allocations by giving them access to memory
3936 if (gfp_mask & __GFP_NOFAIL)
3937 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3938 ALLOC_NO_WATERMARKS, ac);
3941 mutex_unlock(&oom_lock);
3946 * Maximum number of compaction retries wit a progress before OOM
3947 * killer is consider as the only way to move forward.
3949 #define MAX_COMPACT_RETRIES 16
3951 #ifdef CONFIG_COMPACTION
3952 /* Try memory compaction for high-order allocations before reclaim */
3953 static struct page *
3954 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3955 unsigned int alloc_flags, const struct alloc_context *ac,
3956 enum compact_priority prio, enum compact_result *compact_result)
3958 struct page *page = NULL;
3959 unsigned long pflags;
3960 unsigned int noreclaim_flag;
3965 psi_memstall_enter(&pflags);
3966 noreclaim_flag = memalloc_noreclaim_save();
3968 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3971 memalloc_noreclaim_restore(noreclaim_flag);
3972 psi_memstall_leave(&pflags);
3975 * At least in one zone compaction wasn't deferred or skipped, so let's
3976 * count a compaction stall
3978 count_vm_event(COMPACTSTALL);
3980 /* Prep a captured page if available */
3982 prep_new_page(page, order, gfp_mask, alloc_flags);
3984 /* Try get a page from the freelist if available */
3986 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3989 struct zone *zone = page_zone(page);
3991 zone->compact_blockskip_flush = false;
3992 compaction_defer_reset(zone, order, true);
3993 count_vm_event(COMPACTSUCCESS);
3998 * It's bad if compaction run occurs and fails. The most likely reason
3999 * is that pages exist, but not enough to satisfy watermarks.
4001 count_vm_event(COMPACTFAIL);
4009 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4010 enum compact_result compact_result,
4011 enum compact_priority *compact_priority,
4012 int *compaction_retries)
4014 int max_retries = MAX_COMPACT_RETRIES;
4017 int retries = *compaction_retries;
4018 enum compact_priority priority = *compact_priority;
4023 if (compaction_made_progress(compact_result))
4024 (*compaction_retries)++;
4027 * compaction considers all the zone as desperately out of memory
4028 * so it doesn't really make much sense to retry except when the
4029 * failure could be caused by insufficient priority
4031 if (compaction_failed(compact_result))
4032 goto check_priority;
4035 * compaction was skipped because there are not enough order-0 pages
4036 * to work with, so we retry only if it looks like reclaim can help.
4038 if (compaction_needs_reclaim(compact_result)) {
4039 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4044 * make sure the compaction wasn't deferred or didn't bail out early
4045 * due to locks contention before we declare that we should give up.
4046 * But the next retry should use a higher priority if allowed, so
4047 * we don't just keep bailing out endlessly.
4049 if (compaction_withdrawn(compact_result)) {
4050 goto check_priority;
4054 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4055 * costly ones because they are de facto nofail and invoke OOM
4056 * killer to move on while costly can fail and users are ready
4057 * to cope with that. 1/4 retries is rather arbitrary but we
4058 * would need much more detailed feedback from compaction to
4059 * make a better decision.
4061 if (order > PAGE_ALLOC_COSTLY_ORDER)
4063 if (*compaction_retries <= max_retries) {
4069 * Make sure there are attempts at the highest priority if we exhausted
4070 * all retries or failed at the lower priorities.
4073 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4074 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4076 if (*compact_priority > min_priority) {
4077 (*compact_priority)--;
4078 *compaction_retries = 0;
4082 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4086 static inline struct page *
4087 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4088 unsigned int alloc_flags, const struct alloc_context *ac,
4089 enum compact_priority prio, enum compact_result *compact_result)
4091 *compact_result = COMPACT_SKIPPED;
4096 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4097 enum compact_result compact_result,
4098 enum compact_priority *compact_priority,
4099 int *compaction_retries)
4104 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4108 * There are setups with compaction disabled which would prefer to loop
4109 * inside the allocator rather than hit the oom killer prematurely.
4110 * Let's give them a good hope and keep retrying while the order-0
4111 * watermarks are OK.
4113 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4115 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4116 ac_classzone_idx(ac), alloc_flags))
4121 #endif /* CONFIG_COMPACTION */
4123 #ifdef CONFIG_LOCKDEP
4124 static struct lockdep_map __fs_reclaim_map =
4125 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4127 static bool __need_fs_reclaim(gfp_t gfp_mask)
4129 gfp_mask = current_gfp_context(gfp_mask);
4131 /* no reclaim without waiting on it */
4132 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4135 /* this guy won't enter reclaim */
4136 if (current->flags & PF_MEMALLOC)
4139 /* We're only interested __GFP_FS allocations for now */
4140 if (!(gfp_mask & __GFP_FS))
4143 if (gfp_mask & __GFP_NOLOCKDEP)
4149 void __fs_reclaim_acquire(void)
4151 lock_map_acquire(&__fs_reclaim_map);
4154 void __fs_reclaim_release(void)
4156 lock_map_release(&__fs_reclaim_map);
4159 void fs_reclaim_acquire(gfp_t gfp_mask)
4161 if (__need_fs_reclaim(gfp_mask))
4162 __fs_reclaim_acquire();
4164 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4166 void fs_reclaim_release(gfp_t gfp_mask)
4168 if (__need_fs_reclaim(gfp_mask))
4169 __fs_reclaim_release();
4171 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4174 /* Perform direct synchronous page reclaim */
4176 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4177 const struct alloc_context *ac)
4180 unsigned int noreclaim_flag;
4181 unsigned long pflags;
4185 /* We now go into synchronous reclaim */
4186 cpuset_memory_pressure_bump();
4187 psi_memstall_enter(&pflags);
4188 fs_reclaim_acquire(gfp_mask);
4189 noreclaim_flag = memalloc_noreclaim_save();
4191 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4194 memalloc_noreclaim_restore(noreclaim_flag);
4195 fs_reclaim_release(gfp_mask);
4196 psi_memstall_leave(&pflags);
4203 /* The really slow allocator path where we enter direct reclaim */
4204 static inline struct page *
4205 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4206 unsigned int alloc_flags, const struct alloc_context *ac,
4207 unsigned long *did_some_progress)
4209 struct page *page = NULL;
4210 bool drained = false;
4212 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4213 if (unlikely(!(*did_some_progress)))
4217 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4220 * If an allocation failed after direct reclaim, it could be because
4221 * pages are pinned on the per-cpu lists or in high alloc reserves.
4222 * Shrink them them and try again
4224 if (!page && !drained) {
4225 unreserve_highatomic_pageblock(ac, false);
4226 drain_all_pages(NULL);
4234 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4235 const struct alloc_context *ac)
4239 pg_data_t *last_pgdat = NULL;
4240 enum zone_type high_zoneidx = ac->high_zoneidx;
4242 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
4244 if (last_pgdat != zone->zone_pgdat)
4245 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
4246 last_pgdat = zone->zone_pgdat;
4250 static inline unsigned int
4251 gfp_to_alloc_flags(gfp_t gfp_mask)
4253 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4256 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4257 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4258 * to save two branches.
4260 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4261 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4264 * The caller may dip into page reserves a bit more if the caller
4265 * cannot run direct reclaim, or if the caller has realtime scheduling
4266 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4267 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4269 alloc_flags |= (__force int)
4270 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4272 if (gfp_mask & __GFP_ATOMIC) {
4274 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4275 * if it can't schedule.
4277 if (!(gfp_mask & __GFP_NOMEMALLOC))
4278 alloc_flags |= ALLOC_HARDER;
4280 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4281 * comment for __cpuset_node_allowed().
4283 alloc_flags &= ~ALLOC_CPUSET;
4284 } else if (unlikely(rt_task(current)) && !in_interrupt())
4285 alloc_flags |= ALLOC_HARDER;
4288 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4289 alloc_flags |= ALLOC_CMA;
4294 static bool oom_reserves_allowed(struct task_struct *tsk)
4296 if (!tsk_is_oom_victim(tsk))
4300 * !MMU doesn't have oom reaper so give access to memory reserves
4301 * only to the thread with TIF_MEMDIE set
4303 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4310 * Distinguish requests which really need access to full memory
4311 * reserves from oom victims which can live with a portion of it
4313 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4315 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4317 if (gfp_mask & __GFP_MEMALLOC)
4318 return ALLOC_NO_WATERMARKS;
4319 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4320 return ALLOC_NO_WATERMARKS;
4321 if (!in_interrupt()) {
4322 if (current->flags & PF_MEMALLOC)
4323 return ALLOC_NO_WATERMARKS;
4324 else if (oom_reserves_allowed(current))
4331 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4333 return !!__gfp_pfmemalloc_flags(gfp_mask);
4337 * Checks whether it makes sense to retry the reclaim to make a forward progress
4338 * for the given allocation request.
4340 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4341 * without success, or when we couldn't even meet the watermark if we
4342 * reclaimed all remaining pages on the LRU lists.
4344 * Returns true if a retry is viable or false to enter the oom path.
4347 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4348 struct alloc_context *ac, int alloc_flags,
4349 bool did_some_progress, int *no_progress_loops)
4356 * Costly allocations might have made a progress but this doesn't mean
4357 * their order will become available due to high fragmentation so
4358 * always increment the no progress counter for them
4360 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4361 *no_progress_loops = 0;
4363 (*no_progress_loops)++;
4366 * Make sure we converge to OOM if we cannot make any progress
4367 * several times in the row.
4369 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4370 /* Before OOM, exhaust highatomic_reserve */
4371 return unreserve_highatomic_pageblock(ac, true);
4375 * Keep reclaiming pages while there is a chance this will lead
4376 * somewhere. If none of the target zones can satisfy our allocation
4377 * request even if all reclaimable pages are considered then we are
4378 * screwed and have to go OOM.
4380 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4382 unsigned long available;
4383 unsigned long reclaimable;
4384 unsigned long min_wmark = min_wmark_pages(zone);
4387 available = reclaimable = zone_reclaimable_pages(zone);
4388 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4391 * Would the allocation succeed if we reclaimed all
4392 * reclaimable pages?
4394 wmark = __zone_watermark_ok(zone, order, min_wmark,
4395 ac_classzone_idx(ac), alloc_flags, available);
4396 trace_reclaim_retry_zone(z, order, reclaimable,
4397 available, min_wmark, *no_progress_loops, wmark);
4400 * If we didn't make any progress and have a lot of
4401 * dirty + writeback pages then we should wait for
4402 * an IO to complete to slow down the reclaim and
4403 * prevent from pre mature OOM
4405 if (!did_some_progress) {
4406 unsigned long write_pending;
4408 write_pending = zone_page_state_snapshot(zone,
4409 NR_ZONE_WRITE_PENDING);
4411 if (2 * write_pending > reclaimable) {
4412 congestion_wait(BLK_RW_ASYNC, HZ/10);
4424 * Memory allocation/reclaim might be called from a WQ context and the
4425 * current implementation of the WQ concurrency control doesn't
4426 * recognize that a particular WQ is congested if the worker thread is
4427 * looping without ever sleeping. Therefore we have to do a short sleep
4428 * here rather than calling cond_resched().
4430 if (current->flags & PF_WQ_WORKER)
4431 schedule_timeout_uninterruptible(1);
4438 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4441 * It's possible that cpuset's mems_allowed and the nodemask from
4442 * mempolicy don't intersect. This should be normally dealt with by
4443 * policy_nodemask(), but it's possible to race with cpuset update in
4444 * such a way the check therein was true, and then it became false
4445 * before we got our cpuset_mems_cookie here.
4446 * This assumes that for all allocations, ac->nodemask can come only
4447 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4448 * when it does not intersect with the cpuset restrictions) or the
4449 * caller can deal with a violated nodemask.
4451 if (cpusets_enabled() && ac->nodemask &&
4452 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4453 ac->nodemask = NULL;
4458 * When updating a task's mems_allowed or mempolicy nodemask, it is
4459 * possible to race with parallel threads in such a way that our
4460 * allocation can fail while the mask is being updated. If we are about
4461 * to fail, check if the cpuset changed during allocation and if so,
4464 if (read_mems_allowed_retry(cpuset_mems_cookie))
4470 static inline struct page *
4471 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4472 struct alloc_context *ac)
4474 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4475 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4476 struct page *page = NULL;
4477 unsigned int alloc_flags;
4478 unsigned long did_some_progress;
4479 enum compact_priority compact_priority;
4480 enum compact_result compact_result;
4481 int compaction_retries;
4482 int no_progress_loops;
4483 unsigned int cpuset_mems_cookie;
4487 * We also sanity check to catch abuse of atomic reserves being used by
4488 * callers that are not in atomic context.
4490 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4491 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4492 gfp_mask &= ~__GFP_ATOMIC;
4495 compaction_retries = 0;
4496 no_progress_loops = 0;
4497 compact_priority = DEF_COMPACT_PRIORITY;
4498 cpuset_mems_cookie = read_mems_allowed_begin();
4501 * The fast path uses conservative alloc_flags to succeed only until
4502 * kswapd needs to be woken up, and to avoid the cost of setting up
4503 * alloc_flags precisely. So we do that now.
4505 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4508 * We need to recalculate the starting point for the zonelist iterator
4509 * because we might have used different nodemask in the fast path, or
4510 * there was a cpuset modification and we are retrying - otherwise we
4511 * could end up iterating over non-eligible zones endlessly.
4513 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4514 ac->high_zoneidx, ac->nodemask);
4515 if (!ac->preferred_zoneref->zone)
4518 if (alloc_flags & ALLOC_KSWAPD)
4519 wake_all_kswapds(order, gfp_mask, ac);
4522 * The adjusted alloc_flags might result in immediate success, so try
4525 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4530 * For costly allocations, try direct compaction first, as it's likely
4531 * that we have enough base pages and don't need to reclaim. For non-
4532 * movable high-order allocations, do that as well, as compaction will
4533 * try prevent permanent fragmentation by migrating from blocks of the
4535 * Don't try this for allocations that are allowed to ignore
4536 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4538 if (can_direct_reclaim &&
4540 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4541 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4542 page = __alloc_pages_direct_compact(gfp_mask, order,
4544 INIT_COMPACT_PRIORITY,
4550 * Checks for costly allocations with __GFP_NORETRY, which
4551 * includes some THP page fault allocations
4553 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4555 * If allocating entire pageblock(s) and compaction
4556 * failed because all zones are below low watermarks
4557 * or is prohibited because it recently failed at this
4558 * order, fail immediately unless the allocator has
4559 * requested compaction and reclaim retry.
4562 * - potentially very expensive because zones are far
4563 * below their low watermarks or this is part of very
4564 * bursty high order allocations,
4565 * - not guaranteed to help because isolate_freepages()
4566 * may not iterate over freed pages as part of its
4568 * - unlikely to make entire pageblocks free on its
4571 if (compact_result == COMPACT_SKIPPED ||
4572 compact_result == COMPACT_DEFERRED)
4576 * Looks like reclaim/compaction is worth trying, but
4577 * sync compaction could be very expensive, so keep
4578 * using async compaction.
4580 compact_priority = INIT_COMPACT_PRIORITY;
4585 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4586 if (alloc_flags & ALLOC_KSWAPD)
4587 wake_all_kswapds(order, gfp_mask, ac);
4589 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4591 alloc_flags = reserve_flags;
4594 * Reset the nodemask and zonelist iterators if memory policies can be
4595 * ignored. These allocations are high priority and system rather than
4598 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4599 ac->nodemask = NULL;
4600 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4601 ac->high_zoneidx, ac->nodemask);
4604 /* Attempt with potentially adjusted zonelist and alloc_flags */
4605 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4609 /* Caller is not willing to reclaim, we can't balance anything */
4610 if (!can_direct_reclaim)
4613 /* Avoid recursion of direct reclaim */
4614 if (current->flags & PF_MEMALLOC)
4617 /* Try direct reclaim and then allocating */
4618 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4619 &did_some_progress);
4623 /* Try direct compaction and then allocating */
4624 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4625 compact_priority, &compact_result);
4629 /* Do not loop if specifically requested */
4630 if (gfp_mask & __GFP_NORETRY)
4634 * Do not retry costly high order allocations unless they are
4635 * __GFP_RETRY_MAYFAIL
4637 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4640 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4641 did_some_progress > 0, &no_progress_loops))
4645 * It doesn't make any sense to retry for the compaction if the order-0
4646 * reclaim is not able to make any progress because the current
4647 * implementation of the compaction depends on the sufficient amount
4648 * of free memory (see __compaction_suitable)
4650 if (did_some_progress > 0 &&
4651 should_compact_retry(ac, order, alloc_flags,
4652 compact_result, &compact_priority,
4653 &compaction_retries))
4657 /* Deal with possible cpuset update races before we start OOM killing */
4658 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4661 /* Reclaim has failed us, start killing things */
4662 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4666 /* Avoid allocations with no watermarks from looping endlessly */
4667 if (tsk_is_oom_victim(current) &&
4668 (alloc_flags == ALLOC_OOM ||
4669 (gfp_mask & __GFP_NOMEMALLOC)))
4672 /* Retry as long as the OOM killer is making progress */
4673 if (did_some_progress) {
4674 no_progress_loops = 0;
4679 /* Deal with possible cpuset update races before we fail */
4680 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4684 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4687 if (gfp_mask & __GFP_NOFAIL) {
4689 * All existing users of the __GFP_NOFAIL are blockable, so warn
4690 * of any new users that actually require GFP_NOWAIT
4692 if (WARN_ON_ONCE(!can_direct_reclaim))
4696 * PF_MEMALLOC request from this context is rather bizarre
4697 * because we cannot reclaim anything and only can loop waiting
4698 * for somebody to do a work for us
4700 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4703 * non failing costly orders are a hard requirement which we
4704 * are not prepared for much so let's warn about these users
4705 * so that we can identify them and convert them to something
4708 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4711 * Help non-failing allocations by giving them access to memory
4712 * reserves but do not use ALLOC_NO_WATERMARKS because this
4713 * could deplete whole memory reserves which would just make
4714 * the situation worse
4716 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4724 warn_alloc(gfp_mask, ac->nodemask,
4725 "page allocation failure: order:%u", order);
4730 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4731 int preferred_nid, nodemask_t *nodemask,
4732 struct alloc_context *ac, gfp_t *alloc_mask,
4733 unsigned int *alloc_flags)
4735 ac->high_zoneidx = gfp_zone(gfp_mask);
4736 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4737 ac->nodemask = nodemask;
4738 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4740 if (cpusets_enabled()) {
4741 *alloc_mask |= __GFP_HARDWALL;
4743 ac->nodemask = &cpuset_current_mems_allowed;
4745 *alloc_flags |= ALLOC_CPUSET;
4748 fs_reclaim_acquire(gfp_mask);
4749 fs_reclaim_release(gfp_mask);
4751 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4753 if (should_fail_alloc_page(gfp_mask, order))
4756 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4757 *alloc_flags |= ALLOC_CMA;
4762 /* Determine whether to spread dirty pages and what the first usable zone */
4763 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4765 /* Dirty zone balancing only done in the fast path */
4766 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4769 * The preferred zone is used for statistics but crucially it is
4770 * also used as the starting point for the zonelist iterator. It
4771 * may get reset for allocations that ignore memory policies.
4773 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4774 ac->high_zoneidx, ac->nodemask);
4778 * This is the 'heart' of the zoned buddy allocator.
4781 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4782 nodemask_t *nodemask)
4785 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4786 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4787 struct alloc_context ac = { };
4790 * There are several places where we assume that the order value is sane
4791 * so bail out early if the request is out of bound.
4793 if (unlikely(order >= MAX_ORDER)) {
4794 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4798 gfp_mask &= gfp_allowed_mask;
4799 alloc_mask = gfp_mask;
4800 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4803 finalise_ac(gfp_mask, &ac);
4806 * Forbid the first pass from falling back to types that fragment
4807 * memory until all local zones are considered.
4809 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4811 /* First allocation attempt */
4812 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4817 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4818 * resp. GFP_NOIO which has to be inherited for all allocation requests
4819 * from a particular context which has been marked by
4820 * memalloc_no{fs,io}_{save,restore}.
4822 alloc_mask = current_gfp_context(gfp_mask);
4823 ac.spread_dirty_pages = false;
4826 * Restore the original nodemask if it was potentially replaced with
4827 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4829 ac.nodemask = nodemask;
4831 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4834 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4835 unlikely(__memcg_kmem_charge_page(page, gfp_mask, order) != 0)) {
4836 __free_pages(page, order);
4840 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4844 EXPORT_SYMBOL(__alloc_pages_nodemask);
4847 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4848 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4849 * you need to access high mem.
4851 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4855 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4858 return (unsigned long) page_address(page);
4860 EXPORT_SYMBOL(__get_free_pages);
4862 unsigned long get_zeroed_page(gfp_t gfp_mask)
4864 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4866 EXPORT_SYMBOL(get_zeroed_page);
4868 static inline void free_the_page(struct page *page, unsigned int order)
4870 if (order == 0) /* Via pcp? */
4871 free_unref_page(page);
4873 __free_pages_ok(page, order);
4876 void __free_pages(struct page *page, unsigned int order)
4878 if (put_page_testzero(page))
4879 free_the_page(page, order);
4881 EXPORT_SYMBOL(__free_pages);
4883 void free_pages(unsigned long addr, unsigned int order)
4886 VM_BUG_ON(!virt_addr_valid((void *)addr));
4887 __free_pages(virt_to_page((void *)addr), order);
4891 EXPORT_SYMBOL(free_pages);
4895 * An arbitrary-length arbitrary-offset area of memory which resides
4896 * within a 0 or higher order page. Multiple fragments within that page
4897 * are individually refcounted, in the page's reference counter.
4899 * The page_frag functions below provide a simple allocation framework for
4900 * page fragments. This is used by the network stack and network device
4901 * drivers to provide a backing region of memory for use as either an
4902 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4904 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4907 struct page *page = NULL;
4908 gfp_t gfp = gfp_mask;
4910 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4911 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4913 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4914 PAGE_FRAG_CACHE_MAX_ORDER);
4915 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4917 if (unlikely(!page))
4918 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4920 nc->va = page ? page_address(page) : NULL;
4925 void __page_frag_cache_drain(struct page *page, unsigned int count)
4927 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4929 if (page_ref_sub_and_test(page, count))
4930 free_the_page(page, compound_order(page));
4932 EXPORT_SYMBOL(__page_frag_cache_drain);
4934 void *page_frag_alloc(struct page_frag_cache *nc,
4935 unsigned int fragsz, gfp_t gfp_mask)
4937 unsigned int size = PAGE_SIZE;
4941 if (unlikely(!nc->va)) {
4943 page = __page_frag_cache_refill(nc, gfp_mask);
4947 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4948 /* if size can vary use size else just use PAGE_SIZE */
4951 /* Even if we own the page, we do not use atomic_set().
4952 * This would break get_page_unless_zero() users.
4954 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4956 /* reset page count bias and offset to start of new frag */
4957 nc->pfmemalloc = page_is_pfmemalloc(page);
4958 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4962 offset = nc->offset - fragsz;
4963 if (unlikely(offset < 0)) {
4964 page = virt_to_page(nc->va);
4966 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4969 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4970 /* if size can vary use size else just use PAGE_SIZE */
4973 /* OK, page count is 0, we can safely set it */
4974 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4976 /* reset page count bias and offset to start of new frag */
4977 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4978 offset = size - fragsz;
4982 nc->offset = offset;
4984 return nc->va + offset;
4986 EXPORT_SYMBOL(page_frag_alloc);
4989 * Frees a page fragment allocated out of either a compound or order 0 page.
4991 void page_frag_free(void *addr)
4993 struct page *page = virt_to_head_page(addr);
4995 if (unlikely(put_page_testzero(page)))
4996 free_the_page(page, compound_order(page));
4998 EXPORT_SYMBOL(page_frag_free);
5000 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5004 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5005 unsigned long used = addr + PAGE_ALIGN(size);
5007 split_page(virt_to_page((void *)addr), order);
5008 while (used < alloc_end) {
5013 return (void *)addr;
5017 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5018 * @size: the number of bytes to allocate
5019 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5021 * This function is similar to alloc_pages(), except that it allocates the
5022 * minimum number of pages to satisfy the request. alloc_pages() can only
5023 * allocate memory in power-of-two pages.
5025 * This function is also limited by MAX_ORDER.
5027 * Memory allocated by this function must be released by free_pages_exact().
5029 * Return: pointer to the allocated area or %NULL in case of error.
5031 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5033 unsigned int order = get_order(size);
5036 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5037 gfp_mask &= ~__GFP_COMP;
5039 addr = __get_free_pages(gfp_mask, order);
5040 return make_alloc_exact(addr, order, size);
5042 EXPORT_SYMBOL(alloc_pages_exact);
5045 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5047 * @nid: the preferred node ID where memory should be allocated
5048 * @size: the number of bytes to allocate
5049 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5051 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5054 * Return: pointer to the allocated area or %NULL in case of error.
5056 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5058 unsigned int order = get_order(size);
5061 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5062 gfp_mask &= ~__GFP_COMP;
5064 p = alloc_pages_node(nid, gfp_mask, order);
5067 return make_alloc_exact((unsigned long)page_address(p), order, size);
5071 * free_pages_exact - release memory allocated via alloc_pages_exact()
5072 * @virt: the value returned by alloc_pages_exact.
5073 * @size: size of allocation, same value as passed to alloc_pages_exact().
5075 * Release the memory allocated by a previous call to alloc_pages_exact.
5077 void free_pages_exact(void *virt, size_t size)
5079 unsigned long addr = (unsigned long)virt;
5080 unsigned long end = addr + PAGE_ALIGN(size);
5082 while (addr < end) {
5087 EXPORT_SYMBOL(free_pages_exact);
5090 * nr_free_zone_pages - count number of pages beyond high watermark
5091 * @offset: The zone index of the highest zone
5093 * nr_free_zone_pages() counts the number of pages which are beyond the
5094 * high watermark within all zones at or below a given zone index. For each
5095 * zone, the number of pages is calculated as:
5097 * nr_free_zone_pages = managed_pages - high_pages
5099 * Return: number of pages beyond high watermark.
5101 static unsigned long nr_free_zone_pages(int offset)
5106 /* Just pick one node, since fallback list is circular */
5107 unsigned long sum = 0;
5109 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5111 for_each_zone_zonelist(zone, z, zonelist, offset) {
5112 unsigned long size = zone_managed_pages(zone);
5113 unsigned long high = high_wmark_pages(zone);
5122 * nr_free_buffer_pages - count number of pages beyond high watermark
5124 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5125 * watermark within ZONE_DMA and ZONE_NORMAL.
5127 * Return: number of pages beyond high watermark within ZONE_DMA and
5130 unsigned long nr_free_buffer_pages(void)
5132 return nr_free_zone_pages(gfp_zone(GFP_USER));
5134 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5137 * nr_free_pagecache_pages - count number of pages beyond high watermark
5139 * nr_free_pagecache_pages() counts the number of pages which are beyond the
5140 * high watermark within all zones.
5142 * Return: number of pages beyond high watermark within all zones.
5144 unsigned long nr_free_pagecache_pages(void)
5146 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5149 static inline void show_node(struct zone *zone)
5151 if (IS_ENABLED(CONFIG_NUMA))
5152 printk("Node %d ", zone_to_nid(zone));
5155 long si_mem_available(void)
5158 unsigned long pagecache;
5159 unsigned long wmark_low = 0;
5160 unsigned long pages[NR_LRU_LISTS];
5161 unsigned long reclaimable;
5165 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5166 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5169 wmark_low += low_wmark_pages(zone);
5172 * Estimate the amount of memory available for userspace allocations,
5173 * without causing swapping.
5175 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5178 * Not all the page cache can be freed, otherwise the system will
5179 * start swapping. Assume at least half of the page cache, or the
5180 * low watermark worth of cache, needs to stay.
5182 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5183 pagecache -= min(pagecache / 2, wmark_low);
5184 available += pagecache;
5187 * Part of the reclaimable slab and other kernel memory consists of
5188 * items that are in use, and cannot be freed. Cap this estimate at the
5191 reclaimable = global_node_page_state(NR_SLAB_RECLAIMABLE) +
5192 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5193 available += reclaimable - min(reclaimable / 2, wmark_low);
5199 EXPORT_SYMBOL_GPL(si_mem_available);
5201 void si_meminfo(struct sysinfo *val)
5203 val->totalram = totalram_pages();
5204 val->sharedram = global_node_page_state(NR_SHMEM);
5205 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5206 val->bufferram = nr_blockdev_pages();
5207 val->totalhigh = totalhigh_pages();
5208 val->freehigh = nr_free_highpages();
5209 val->mem_unit = PAGE_SIZE;
5212 EXPORT_SYMBOL(si_meminfo);
5215 void si_meminfo_node(struct sysinfo *val, int nid)
5217 int zone_type; /* needs to be signed */
5218 unsigned long managed_pages = 0;
5219 unsigned long managed_highpages = 0;
5220 unsigned long free_highpages = 0;
5221 pg_data_t *pgdat = NODE_DATA(nid);
5223 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5224 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5225 val->totalram = managed_pages;
5226 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5227 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5228 #ifdef CONFIG_HIGHMEM
5229 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5230 struct zone *zone = &pgdat->node_zones[zone_type];
5232 if (is_highmem(zone)) {
5233 managed_highpages += zone_managed_pages(zone);
5234 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5237 val->totalhigh = managed_highpages;
5238 val->freehigh = free_highpages;
5240 val->totalhigh = managed_highpages;
5241 val->freehigh = free_highpages;
5243 val->mem_unit = PAGE_SIZE;
5248 * Determine whether the node should be displayed or not, depending on whether
5249 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5251 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5253 if (!(flags & SHOW_MEM_FILTER_NODES))
5257 * no node mask - aka implicit memory numa policy. Do not bother with
5258 * the synchronization - read_mems_allowed_begin - because we do not
5259 * have to be precise here.
5262 nodemask = &cpuset_current_mems_allowed;
5264 return !node_isset(nid, *nodemask);
5267 #define K(x) ((x) << (PAGE_SHIFT-10))
5269 static void show_migration_types(unsigned char type)
5271 static const char types[MIGRATE_TYPES] = {
5272 [MIGRATE_UNMOVABLE] = 'U',
5273 [MIGRATE_MOVABLE] = 'M',
5274 [MIGRATE_RECLAIMABLE] = 'E',
5275 [MIGRATE_HIGHATOMIC] = 'H',
5277 [MIGRATE_CMA] = 'C',
5279 #ifdef CONFIG_MEMORY_ISOLATION
5280 [MIGRATE_ISOLATE] = 'I',
5283 char tmp[MIGRATE_TYPES + 1];
5287 for (i = 0; i < MIGRATE_TYPES; i++) {
5288 if (type & (1 << i))
5293 printk(KERN_CONT "(%s) ", tmp);
5297 * Show free area list (used inside shift_scroll-lock stuff)
5298 * We also calculate the percentage fragmentation. We do this by counting the
5299 * memory on each free list with the exception of the first item on the list.
5302 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5305 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5307 unsigned long free_pcp = 0;
5312 for_each_populated_zone(zone) {
5313 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5316 for_each_online_cpu(cpu)
5317 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5320 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5321 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5322 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
5323 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5324 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5325 " free:%lu free_pcp:%lu free_cma:%lu\n",
5326 global_node_page_state(NR_ACTIVE_ANON),
5327 global_node_page_state(NR_INACTIVE_ANON),
5328 global_node_page_state(NR_ISOLATED_ANON),
5329 global_node_page_state(NR_ACTIVE_FILE),
5330 global_node_page_state(NR_INACTIVE_FILE),
5331 global_node_page_state(NR_ISOLATED_FILE),
5332 global_node_page_state(NR_UNEVICTABLE),
5333 global_node_page_state(NR_FILE_DIRTY),
5334 global_node_page_state(NR_WRITEBACK),
5335 global_node_page_state(NR_UNSTABLE_NFS),
5336 global_node_page_state(NR_SLAB_RECLAIMABLE),
5337 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
5338 global_node_page_state(NR_FILE_MAPPED),
5339 global_node_page_state(NR_SHMEM),
5340 global_zone_page_state(NR_PAGETABLE),
5341 global_zone_page_state(NR_BOUNCE),
5342 global_zone_page_state(NR_FREE_PAGES),
5344 global_zone_page_state(NR_FREE_CMA_PAGES));
5346 for_each_online_pgdat(pgdat) {
5347 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5351 " active_anon:%lukB"
5352 " inactive_anon:%lukB"
5353 " active_file:%lukB"
5354 " inactive_file:%lukB"
5355 " unevictable:%lukB"
5356 " isolated(anon):%lukB"
5357 " isolated(file):%lukB"
5362 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5364 " shmem_pmdmapped: %lukB"
5367 " writeback_tmp:%lukB"
5369 " all_unreclaimable? %s"
5372 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5373 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5374 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5375 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5376 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5377 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5378 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5379 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5380 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5381 K(node_page_state(pgdat, NR_WRITEBACK)),
5382 K(node_page_state(pgdat, NR_SHMEM)),
5383 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5384 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5385 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5387 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5389 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5390 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
5391 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5395 for_each_populated_zone(zone) {
5398 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5402 for_each_online_cpu(cpu)
5403 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5412 " reserved_highatomic:%luKB"
5413 " active_anon:%lukB"
5414 " inactive_anon:%lukB"
5415 " active_file:%lukB"
5416 " inactive_file:%lukB"
5417 " unevictable:%lukB"
5418 " writepending:%lukB"
5422 " kernel_stack:%lukB"
5423 #ifdef CONFIG_SHADOW_CALL_STACK
5424 " shadow_call_stack:%lukB"
5433 K(zone_page_state(zone, NR_FREE_PAGES)),
5434 K(min_wmark_pages(zone)),
5435 K(low_wmark_pages(zone)),
5436 K(high_wmark_pages(zone)),
5437 K(zone->nr_reserved_highatomic),
5438 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5439 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5440 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5441 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5442 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5443 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5444 K(zone->present_pages),
5445 K(zone_managed_pages(zone)),
5446 K(zone_page_state(zone, NR_MLOCK)),
5447 zone_page_state(zone, NR_KERNEL_STACK_KB),
5448 #ifdef CONFIG_SHADOW_CALL_STACK
5449 zone_page_state(zone, NR_KERNEL_SCS_KB),
5451 K(zone_page_state(zone, NR_PAGETABLE)),
5452 K(zone_page_state(zone, NR_BOUNCE)),
5454 K(this_cpu_read(zone->pageset->pcp.count)),
5455 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5456 printk("lowmem_reserve[]:");
5457 for (i = 0; i < MAX_NR_ZONES; i++)
5458 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5459 printk(KERN_CONT "\n");
5462 for_each_populated_zone(zone) {
5464 unsigned long nr[MAX_ORDER], flags, total = 0;
5465 unsigned char types[MAX_ORDER];
5467 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5470 printk(KERN_CONT "%s: ", zone->name);
5472 spin_lock_irqsave(&zone->lock, flags);
5473 for (order = 0; order < MAX_ORDER; order++) {
5474 struct free_area *area = &zone->free_area[order];
5477 nr[order] = area->nr_free;
5478 total += nr[order] << order;
5481 for (type = 0; type < MIGRATE_TYPES; type++) {
5482 if (!free_area_empty(area, type))
5483 types[order] |= 1 << type;
5486 spin_unlock_irqrestore(&zone->lock, flags);
5487 for (order = 0; order < MAX_ORDER; order++) {
5488 printk(KERN_CONT "%lu*%lukB ",
5489 nr[order], K(1UL) << order);
5491 show_migration_types(types[order]);
5493 printk(KERN_CONT "= %lukB\n", K(total));
5496 hugetlb_show_meminfo();
5498 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5500 show_swap_cache_info();
5503 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5505 zoneref->zone = zone;
5506 zoneref->zone_idx = zone_idx(zone);
5510 * Builds allocation fallback zone lists.
5512 * Add all populated zones of a node to the zonelist.
5514 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5517 enum zone_type zone_type = MAX_NR_ZONES;
5522 zone = pgdat->node_zones + zone_type;
5523 if (managed_zone(zone)) {
5524 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5525 check_highest_zone(zone_type);
5527 } while (zone_type);
5534 static int __parse_numa_zonelist_order(char *s)
5537 * We used to support different zonlists modes but they turned
5538 * out to be just not useful. Let's keep the warning in place
5539 * if somebody still use the cmd line parameter so that we do
5540 * not fail it silently
5542 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5543 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5549 static __init int setup_numa_zonelist_order(char *s)
5554 return __parse_numa_zonelist_order(s);
5556 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5558 char numa_zonelist_order[] = "Node";
5561 * sysctl handler for numa_zonelist_order
5563 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5564 void __user *buffer, size_t *length,
5571 return proc_dostring(table, write, buffer, length, ppos);
5572 str = memdup_user_nul(buffer, 16);
5574 return PTR_ERR(str);
5576 ret = __parse_numa_zonelist_order(str);
5582 #define MAX_NODE_LOAD (nr_online_nodes)
5583 static int node_load[MAX_NUMNODES];
5586 * find_next_best_node - find the next node that should appear in a given node's fallback list
5587 * @node: node whose fallback list we're appending
5588 * @used_node_mask: nodemask_t of already used nodes
5590 * We use a number of factors to determine which is the next node that should
5591 * appear on a given node's fallback list. The node should not have appeared
5592 * already in @node's fallback list, and it should be the next closest node
5593 * according to the distance array (which contains arbitrary distance values
5594 * from each node to each node in the system), and should also prefer nodes
5595 * with no CPUs, since presumably they'll have very little allocation pressure
5596 * on them otherwise.
5598 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5600 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5603 int min_val = INT_MAX;
5604 int best_node = NUMA_NO_NODE;
5605 const struct cpumask *tmp = cpumask_of_node(0);
5607 /* Use the local node if we haven't already */
5608 if (!node_isset(node, *used_node_mask)) {
5609 node_set(node, *used_node_mask);
5613 for_each_node_state(n, N_MEMORY) {
5615 /* Don't want a node to appear more than once */
5616 if (node_isset(n, *used_node_mask))
5619 /* Use the distance array to find the distance */
5620 val = node_distance(node, n);
5622 /* Penalize nodes under us ("prefer the next node") */
5625 /* Give preference to headless and unused nodes */
5626 tmp = cpumask_of_node(n);
5627 if (!cpumask_empty(tmp))
5628 val += PENALTY_FOR_NODE_WITH_CPUS;
5630 /* Slight preference for less loaded node */
5631 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5632 val += node_load[n];
5634 if (val < min_val) {
5641 node_set(best_node, *used_node_mask);
5648 * Build zonelists ordered by node and zones within node.
5649 * This results in maximum locality--normal zone overflows into local
5650 * DMA zone, if any--but risks exhausting DMA zone.
5652 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5655 struct zoneref *zonerefs;
5658 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5660 for (i = 0; i < nr_nodes; i++) {
5663 pg_data_t *node = NODE_DATA(node_order[i]);
5665 nr_zones = build_zonerefs_node(node, zonerefs);
5666 zonerefs += nr_zones;
5668 zonerefs->zone = NULL;
5669 zonerefs->zone_idx = 0;
5673 * Build gfp_thisnode zonelists
5675 static void build_thisnode_zonelists(pg_data_t *pgdat)
5677 struct zoneref *zonerefs;
5680 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5681 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5682 zonerefs += nr_zones;
5683 zonerefs->zone = NULL;
5684 zonerefs->zone_idx = 0;
5688 * Build zonelists ordered by zone and nodes within zones.
5689 * This results in conserving DMA zone[s] until all Normal memory is
5690 * exhausted, but results in overflowing to remote node while memory
5691 * may still exist in local DMA zone.
5694 static void build_zonelists(pg_data_t *pgdat)
5696 static int node_order[MAX_NUMNODES];
5697 int node, load, nr_nodes = 0;
5698 nodemask_t used_mask;
5699 int local_node, prev_node;
5701 /* NUMA-aware ordering of nodes */
5702 local_node = pgdat->node_id;
5703 load = nr_online_nodes;
5704 prev_node = local_node;
5705 nodes_clear(used_mask);
5707 memset(node_order, 0, sizeof(node_order));
5708 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5710 * We don't want to pressure a particular node.
5711 * So adding penalty to the first node in same
5712 * distance group to make it round-robin.
5714 if (node_distance(local_node, node) !=
5715 node_distance(local_node, prev_node))
5716 node_load[node] = load;
5718 node_order[nr_nodes++] = node;
5723 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5724 build_thisnode_zonelists(pgdat);
5727 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5729 * Return node id of node used for "local" allocations.
5730 * I.e., first node id of first zone in arg node's generic zonelist.
5731 * Used for initializing percpu 'numa_mem', which is used primarily
5732 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5734 int local_memory_node(int node)
5738 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5739 gfp_zone(GFP_KERNEL),
5741 return zone_to_nid(z->zone);
5745 static void setup_min_unmapped_ratio(void);
5746 static void setup_min_slab_ratio(void);
5747 #else /* CONFIG_NUMA */
5749 static void build_zonelists(pg_data_t *pgdat)
5751 int node, local_node;
5752 struct zoneref *zonerefs;
5755 local_node = pgdat->node_id;
5757 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5758 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5759 zonerefs += nr_zones;
5762 * Now we build the zonelist so that it contains the zones
5763 * of all the other nodes.
5764 * We don't want to pressure a particular node, so when
5765 * building the zones for node N, we make sure that the
5766 * zones coming right after the local ones are those from
5767 * node N+1 (modulo N)
5769 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5770 if (!node_online(node))
5772 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5773 zonerefs += nr_zones;
5775 for (node = 0; node < local_node; node++) {
5776 if (!node_online(node))
5778 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5779 zonerefs += nr_zones;
5782 zonerefs->zone = NULL;
5783 zonerefs->zone_idx = 0;
5786 #endif /* CONFIG_NUMA */
5789 * Boot pageset table. One per cpu which is going to be used for all
5790 * zones and all nodes. The parameters will be set in such a way
5791 * that an item put on a list will immediately be handed over to
5792 * the buddy list. This is safe since pageset manipulation is done
5793 * with interrupts disabled.
5795 * The boot_pagesets must be kept even after bootup is complete for
5796 * unused processors and/or zones. They do play a role for bootstrapping
5797 * hotplugged processors.
5799 * zoneinfo_show() and maybe other functions do
5800 * not check if the processor is online before following the pageset pointer.
5801 * Other parts of the kernel may not check if the zone is available.
5803 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5804 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5805 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5807 static void __build_all_zonelists(void *data)
5810 int __maybe_unused cpu;
5811 pg_data_t *self = data;
5812 static DEFINE_SPINLOCK(lock);
5817 memset(node_load, 0, sizeof(node_load));
5821 * This node is hotadded and no memory is yet present. So just
5822 * building zonelists is fine - no need to touch other nodes.
5824 if (self && !node_online(self->node_id)) {
5825 build_zonelists(self);
5827 for_each_online_node(nid) {
5828 pg_data_t *pgdat = NODE_DATA(nid);
5830 build_zonelists(pgdat);
5833 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5835 * We now know the "local memory node" for each node--
5836 * i.e., the node of the first zone in the generic zonelist.
5837 * Set up numa_mem percpu variable for on-line cpus. During
5838 * boot, only the boot cpu should be on-line; we'll init the
5839 * secondary cpus' numa_mem as they come on-line. During
5840 * node/memory hotplug, we'll fixup all on-line cpus.
5842 for_each_online_cpu(cpu)
5843 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5850 static noinline void __init
5851 build_all_zonelists_init(void)
5855 __build_all_zonelists(NULL);
5858 * Initialize the boot_pagesets that are going to be used
5859 * for bootstrapping processors. The real pagesets for
5860 * each zone will be allocated later when the per cpu
5861 * allocator is available.
5863 * boot_pagesets are used also for bootstrapping offline
5864 * cpus if the system is already booted because the pagesets
5865 * are needed to initialize allocators on a specific cpu too.
5866 * F.e. the percpu allocator needs the page allocator which
5867 * needs the percpu allocator in order to allocate its pagesets
5868 * (a chicken-egg dilemma).
5870 for_each_possible_cpu(cpu)
5871 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5873 mminit_verify_zonelist();
5874 cpuset_init_current_mems_allowed();
5878 * unless system_state == SYSTEM_BOOTING.
5880 * __ref due to call of __init annotated helper build_all_zonelists_init
5881 * [protected by SYSTEM_BOOTING].
5883 void __ref build_all_zonelists(pg_data_t *pgdat)
5885 if (system_state == SYSTEM_BOOTING) {
5886 build_all_zonelists_init();
5888 __build_all_zonelists(pgdat);
5889 /* cpuset refresh routine should be here */
5891 vm_total_pages = nr_free_pagecache_pages();
5893 * Disable grouping by mobility if the number of pages in the
5894 * system is too low to allow the mechanism to work. It would be
5895 * more accurate, but expensive to check per-zone. This check is
5896 * made on memory-hotadd so a system can start with mobility
5897 * disabled and enable it later
5899 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5900 page_group_by_mobility_disabled = 1;
5902 page_group_by_mobility_disabled = 0;
5904 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5906 page_group_by_mobility_disabled ? "off" : "on",
5909 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5913 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5914 static bool __meminit
5915 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5917 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5918 static struct memblock_region *r;
5920 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5921 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5922 for_each_memblock(memory, r) {
5923 if (*pfn < memblock_region_memory_end_pfn(r))
5927 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5928 memblock_is_mirror(r)) {
5929 *pfn = memblock_region_memory_end_pfn(r);
5937 #ifdef CONFIG_SPARSEMEM
5938 /* Skip PFNs that belong to non-present sections */
5939 static inline __meminit unsigned long next_pfn(unsigned long pfn)
5941 const unsigned long section_nr = pfn_to_section_nr(++pfn);
5943 if (present_section_nr(section_nr))
5945 return section_nr_to_pfn(next_present_section_nr(section_nr));
5948 static inline __meminit unsigned long next_pfn(unsigned long pfn)
5955 * Initially all pages are reserved - free ones are freed
5956 * up by memblock_free_all() once the early boot process is
5957 * done. Non-atomic initialization, single-pass.
5959 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5960 unsigned long start_pfn, enum memmap_context context,
5961 struct vmem_altmap *altmap)
5963 unsigned long pfn, end_pfn = start_pfn + size;
5966 if (highest_memmap_pfn < end_pfn - 1)
5967 highest_memmap_pfn = end_pfn - 1;
5969 #ifdef CONFIG_ZONE_DEVICE
5971 * Honor reservation requested by the driver for this ZONE_DEVICE
5972 * memory. We limit the total number of pages to initialize to just
5973 * those that might contain the memory mapping. We will defer the
5974 * ZONE_DEVICE page initialization until after we have released
5977 if (zone == ZONE_DEVICE) {
5981 if (start_pfn == altmap->base_pfn)
5982 start_pfn += altmap->reserve;
5983 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5987 for (pfn = start_pfn; pfn < end_pfn; ) {
5989 * There can be holes in boot-time mem_map[]s handed to this
5990 * function. They do not exist on hotplugged memory.
5992 if (context == MEMMAP_EARLY) {
5993 if (!early_pfn_valid(pfn)) {
5994 pfn = next_pfn(pfn);
5997 if (!early_pfn_in_nid(pfn, nid)) {
6001 if (overlap_memmap_init(zone, &pfn))
6003 if (defer_init(nid, pfn, end_pfn))
6007 page = pfn_to_page(pfn);
6008 __init_single_page(page, pfn, zone, nid);
6009 if (context == MEMMAP_HOTPLUG)
6010 __SetPageReserved(page);
6013 * Mark the block movable so that blocks are reserved for
6014 * movable at startup. This will force kernel allocations
6015 * to reserve their blocks rather than leaking throughout
6016 * the address space during boot when many long-lived
6017 * kernel allocations are made.
6019 * bitmap is created for zone's valid pfn range. but memmap
6020 * can be created for invalid pages (for alignment)
6021 * check here not to call set_pageblock_migratetype() against
6024 if (!(pfn & (pageblock_nr_pages - 1))) {
6025 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6032 #ifdef CONFIG_ZONE_DEVICE
6033 void __ref memmap_init_zone_device(struct zone *zone,
6034 unsigned long start_pfn,
6035 unsigned long nr_pages,
6036 struct dev_pagemap *pgmap)
6038 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6039 struct pglist_data *pgdat = zone->zone_pgdat;
6040 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6041 unsigned long zone_idx = zone_idx(zone);
6042 unsigned long start = jiffies;
6043 int nid = pgdat->node_id;
6045 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6049 * The call to memmap_init_zone should have already taken care
6050 * of the pages reserved for the memmap, so we can just jump to
6051 * the end of that region and start processing the device pages.
6054 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6055 nr_pages = end_pfn - start_pfn;
6058 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6059 struct page *page = pfn_to_page(pfn);
6061 __init_single_page(page, pfn, zone_idx, nid);
6064 * Mark page reserved as it will need to wait for onlining
6065 * phase for it to be fully associated with a zone.
6067 * We can use the non-atomic __set_bit operation for setting
6068 * the flag as we are still initializing the pages.
6070 __SetPageReserved(page);
6073 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6074 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6075 * ever freed or placed on a driver-private list.
6077 page->pgmap = pgmap;
6078 page->zone_device_data = NULL;
6081 * Mark the block movable so that blocks are reserved for
6082 * movable at startup. This will force kernel allocations
6083 * to reserve their blocks rather than leaking throughout
6084 * the address space during boot when many long-lived
6085 * kernel allocations are made.
6087 * bitmap is created for zone's valid pfn range. but memmap
6088 * can be created for invalid pages (for alignment)
6089 * check here not to call set_pageblock_migratetype() against
6092 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
6093 * because this is done early in section_activate()
6095 if (!(pfn & (pageblock_nr_pages - 1))) {
6096 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6101 pr_info("%s initialised %lu pages in %ums\n", __func__,
6102 nr_pages, jiffies_to_msecs(jiffies - start));
6106 static void __meminit zone_init_free_lists(struct zone *zone)
6108 unsigned int order, t;
6109 for_each_migratetype_order(order, t) {
6110 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6111 zone->free_area[order].nr_free = 0;
6115 void __meminit __weak memmap_init(unsigned long size, int nid,
6116 unsigned long zone, unsigned long start_pfn)
6118 memmap_init_zone(size, nid, zone, start_pfn, MEMMAP_EARLY, NULL);
6121 static int zone_batchsize(struct zone *zone)
6127 * The per-cpu-pages pools are set to around 1000th of the
6130 batch = zone_managed_pages(zone) / 1024;
6131 /* But no more than a meg. */
6132 if (batch * PAGE_SIZE > 1024 * 1024)
6133 batch = (1024 * 1024) / PAGE_SIZE;
6134 batch /= 4; /* We effectively *= 4 below */
6139 * Clamp the batch to a 2^n - 1 value. Having a power
6140 * of 2 value was found to be more likely to have
6141 * suboptimal cache aliasing properties in some cases.
6143 * For example if 2 tasks are alternately allocating
6144 * batches of pages, one task can end up with a lot
6145 * of pages of one half of the possible page colors
6146 * and the other with pages of the other colors.
6148 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6153 /* The deferral and batching of frees should be suppressed under NOMMU
6156 * The problem is that NOMMU needs to be able to allocate large chunks
6157 * of contiguous memory as there's no hardware page translation to
6158 * assemble apparent contiguous memory from discontiguous pages.
6160 * Queueing large contiguous runs of pages for batching, however,
6161 * causes the pages to actually be freed in smaller chunks. As there
6162 * can be a significant delay between the individual batches being
6163 * recycled, this leads to the once large chunks of space being
6164 * fragmented and becoming unavailable for high-order allocations.
6171 * pcp->high and pcp->batch values are related and dependent on one another:
6172 * ->batch must never be higher then ->high.
6173 * The following function updates them in a safe manner without read side
6176 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6177 * those fields changing asynchronously (acording the the above rule).
6179 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6180 * outside of boot time (or some other assurance that no concurrent updaters
6183 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6184 unsigned long batch)
6186 /* start with a fail safe value for batch */
6190 /* Update high, then batch, in order */
6197 /* a companion to pageset_set_high() */
6198 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6200 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6203 static void pageset_init(struct per_cpu_pageset *p)
6205 struct per_cpu_pages *pcp;
6208 memset(p, 0, sizeof(*p));
6211 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6212 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6215 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6218 pageset_set_batch(p, batch);
6222 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6223 * to the value high for the pageset p.
6225 static void pageset_set_high(struct per_cpu_pageset *p,
6228 unsigned long batch = max(1UL, high / 4);
6229 if ((high / 4) > (PAGE_SHIFT * 8))
6230 batch = PAGE_SHIFT * 8;
6232 pageset_update(&p->pcp, high, batch);
6235 static void pageset_set_high_and_batch(struct zone *zone,
6236 struct per_cpu_pageset *pcp)
6238 if (percpu_pagelist_fraction)
6239 pageset_set_high(pcp,
6240 (zone_managed_pages(zone) /
6241 percpu_pagelist_fraction));
6243 pageset_set_batch(pcp, zone_batchsize(zone));
6246 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6248 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6251 pageset_set_high_and_batch(zone, pcp);
6254 void __meminit setup_zone_pageset(struct zone *zone)
6257 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6258 for_each_possible_cpu(cpu)
6259 zone_pageset_init(zone, cpu);
6263 * Allocate per cpu pagesets and initialize them.
6264 * Before this call only boot pagesets were available.
6266 void __init setup_per_cpu_pageset(void)
6268 struct pglist_data *pgdat;
6271 for_each_populated_zone(zone)
6272 setup_zone_pageset(zone);
6274 for_each_online_pgdat(pgdat)
6275 pgdat->per_cpu_nodestats =
6276 alloc_percpu(struct per_cpu_nodestat);
6279 static __meminit void zone_pcp_init(struct zone *zone)
6282 * per cpu subsystem is not up at this point. The following code
6283 * relies on the ability of the linker to provide the
6284 * offset of a (static) per cpu variable into the per cpu area.
6286 zone->pageset = &boot_pageset;
6288 if (populated_zone(zone))
6289 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6290 zone->name, zone->present_pages,
6291 zone_batchsize(zone));
6294 void __meminit init_currently_empty_zone(struct zone *zone,
6295 unsigned long zone_start_pfn,
6298 struct pglist_data *pgdat = zone->zone_pgdat;
6299 int zone_idx = zone_idx(zone) + 1;
6301 if (zone_idx > pgdat->nr_zones)
6302 pgdat->nr_zones = zone_idx;
6304 zone->zone_start_pfn = zone_start_pfn;
6306 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6307 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6309 (unsigned long)zone_idx(zone),
6310 zone_start_pfn, (zone_start_pfn + size));
6312 zone_init_free_lists(zone);
6313 zone->initialized = 1;
6316 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6317 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
6320 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
6322 int __meminit __early_pfn_to_nid(unsigned long pfn,
6323 struct mminit_pfnnid_cache *state)
6325 unsigned long start_pfn, end_pfn;
6328 if (state->last_start <= pfn && pfn < state->last_end)
6329 return state->last_nid;
6331 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
6332 if (nid != NUMA_NO_NODE) {
6333 state->last_start = start_pfn;
6334 state->last_end = end_pfn;
6335 state->last_nid = nid;
6340 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
6343 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
6344 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
6345 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
6347 * If an architecture guarantees that all ranges registered contain no holes
6348 * and may be freed, this this function may be used instead of calling
6349 * memblock_free_early_nid() manually.
6351 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
6353 unsigned long start_pfn, end_pfn;
6356 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
6357 start_pfn = min(start_pfn, max_low_pfn);
6358 end_pfn = min(end_pfn, max_low_pfn);
6360 if (start_pfn < end_pfn)
6361 memblock_free_early_nid(PFN_PHYS(start_pfn),
6362 (end_pfn - start_pfn) << PAGE_SHIFT,
6368 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6369 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6371 * If an architecture guarantees that all ranges registered contain no holes and may
6372 * be freed, this function may be used instead of calling memory_present() manually.
6374 void __init sparse_memory_present_with_active_regions(int nid)
6376 unsigned long start_pfn, end_pfn;
6379 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
6380 memory_present(this_nid, start_pfn, end_pfn);
6384 * get_pfn_range_for_nid - Return the start and end page frames for a node
6385 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6386 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6387 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6389 * It returns the start and end page frame of a node based on information
6390 * provided by memblock_set_node(). If called for a node
6391 * with no available memory, a warning is printed and the start and end
6394 void __init get_pfn_range_for_nid(unsigned int nid,
6395 unsigned long *start_pfn, unsigned long *end_pfn)
6397 unsigned long this_start_pfn, this_end_pfn;
6403 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6404 *start_pfn = min(*start_pfn, this_start_pfn);
6405 *end_pfn = max(*end_pfn, this_end_pfn);
6408 if (*start_pfn == -1UL)
6413 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6414 * assumption is made that zones within a node are ordered in monotonic
6415 * increasing memory addresses so that the "highest" populated zone is used
6417 static void __init find_usable_zone_for_movable(void)
6420 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6421 if (zone_index == ZONE_MOVABLE)
6424 if (arch_zone_highest_possible_pfn[zone_index] >
6425 arch_zone_lowest_possible_pfn[zone_index])
6429 VM_BUG_ON(zone_index == -1);
6430 movable_zone = zone_index;
6434 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6435 * because it is sized independent of architecture. Unlike the other zones,
6436 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6437 * in each node depending on the size of each node and how evenly kernelcore
6438 * is distributed. This helper function adjusts the zone ranges
6439 * provided by the architecture for a given node by using the end of the
6440 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6441 * zones within a node are in order of monotonic increases memory addresses
6443 static void __init adjust_zone_range_for_zone_movable(int nid,
6444 unsigned long zone_type,
6445 unsigned long node_start_pfn,
6446 unsigned long node_end_pfn,
6447 unsigned long *zone_start_pfn,
6448 unsigned long *zone_end_pfn)
6450 /* Only adjust if ZONE_MOVABLE is on this node */
6451 if (zone_movable_pfn[nid]) {
6452 /* Size ZONE_MOVABLE */
6453 if (zone_type == ZONE_MOVABLE) {
6454 *zone_start_pfn = zone_movable_pfn[nid];
6455 *zone_end_pfn = min(node_end_pfn,
6456 arch_zone_highest_possible_pfn[movable_zone]);
6458 /* Adjust for ZONE_MOVABLE starting within this range */
6459 } else if (!mirrored_kernelcore &&
6460 *zone_start_pfn < zone_movable_pfn[nid] &&
6461 *zone_end_pfn > zone_movable_pfn[nid]) {
6462 *zone_end_pfn = zone_movable_pfn[nid];
6464 /* Check if this whole range is within ZONE_MOVABLE */
6465 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6466 *zone_start_pfn = *zone_end_pfn;
6471 * Return the number of pages a zone spans in a node, including holes
6472 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6474 static unsigned long __init zone_spanned_pages_in_node(int nid,
6475 unsigned long zone_type,
6476 unsigned long node_start_pfn,
6477 unsigned long node_end_pfn,
6478 unsigned long *zone_start_pfn,
6479 unsigned long *zone_end_pfn,
6480 unsigned long *ignored)
6482 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6483 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6484 /* When hotadd a new node from cpu_up(), the node should be empty */
6485 if (!node_start_pfn && !node_end_pfn)
6488 /* Get the start and end of the zone */
6489 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6490 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6491 adjust_zone_range_for_zone_movable(nid, zone_type,
6492 node_start_pfn, node_end_pfn,
6493 zone_start_pfn, zone_end_pfn);
6495 /* Check that this node has pages within the zone's required range */
6496 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6499 /* Move the zone boundaries inside the node if necessary */
6500 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6501 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6503 /* Return the spanned pages */
6504 return *zone_end_pfn - *zone_start_pfn;
6508 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6509 * then all holes in the requested range will be accounted for.
6511 unsigned long __init __absent_pages_in_range(int nid,
6512 unsigned long range_start_pfn,
6513 unsigned long range_end_pfn)
6515 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6516 unsigned long start_pfn, end_pfn;
6519 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6520 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6521 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6522 nr_absent -= end_pfn - start_pfn;
6528 * absent_pages_in_range - Return number of page frames in holes within a range
6529 * @start_pfn: The start PFN to start searching for holes
6530 * @end_pfn: The end PFN to stop searching for holes
6532 * Return: the number of pages frames in memory holes within a range.
6534 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6535 unsigned long end_pfn)
6537 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6540 /* Return the number of page frames in holes in a zone on a node */
6541 static unsigned long __init zone_absent_pages_in_node(int nid,
6542 unsigned long zone_type,
6543 unsigned long node_start_pfn,
6544 unsigned long node_end_pfn,
6545 unsigned long *ignored)
6547 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6548 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6549 unsigned long zone_start_pfn, zone_end_pfn;
6550 unsigned long nr_absent;
6552 /* When hotadd a new node from cpu_up(), the node should be empty */
6553 if (!node_start_pfn && !node_end_pfn)
6556 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6557 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6559 adjust_zone_range_for_zone_movable(nid, zone_type,
6560 node_start_pfn, node_end_pfn,
6561 &zone_start_pfn, &zone_end_pfn);
6562 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6565 * ZONE_MOVABLE handling.
6566 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6569 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6570 unsigned long start_pfn, end_pfn;
6571 struct memblock_region *r;
6573 for_each_memblock(memory, r) {
6574 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6575 zone_start_pfn, zone_end_pfn);
6576 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6577 zone_start_pfn, zone_end_pfn);
6579 if (zone_type == ZONE_MOVABLE &&
6580 memblock_is_mirror(r))
6581 nr_absent += end_pfn - start_pfn;
6583 if (zone_type == ZONE_NORMAL &&
6584 !memblock_is_mirror(r))
6585 nr_absent += end_pfn - start_pfn;
6592 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6593 static inline unsigned long __init zone_spanned_pages_in_node(int nid,
6594 unsigned long zone_type,
6595 unsigned long node_start_pfn,
6596 unsigned long node_end_pfn,
6597 unsigned long *zone_start_pfn,
6598 unsigned long *zone_end_pfn,
6599 unsigned long *zones_size)
6603 *zone_start_pfn = node_start_pfn;
6604 for (zone = 0; zone < zone_type; zone++)
6605 *zone_start_pfn += zones_size[zone];
6607 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6609 return zones_size[zone_type];
6612 static inline unsigned long __init zone_absent_pages_in_node(int nid,
6613 unsigned long zone_type,
6614 unsigned long node_start_pfn,
6615 unsigned long node_end_pfn,
6616 unsigned long *zholes_size)
6621 return zholes_size[zone_type];
6624 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6626 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6627 unsigned long node_start_pfn,
6628 unsigned long node_end_pfn,
6629 unsigned long *zones_size,
6630 unsigned long *zholes_size)
6632 unsigned long realtotalpages = 0, totalpages = 0;
6635 for (i = 0; i < MAX_NR_ZONES; i++) {
6636 struct zone *zone = pgdat->node_zones + i;
6637 unsigned long zone_start_pfn, zone_end_pfn;
6638 unsigned long size, real_size;
6640 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6646 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6647 node_start_pfn, node_end_pfn,
6650 zone->zone_start_pfn = zone_start_pfn;
6652 zone->zone_start_pfn = 0;
6653 zone->spanned_pages = size;
6654 zone->present_pages = real_size;
6657 realtotalpages += real_size;
6660 pgdat->node_spanned_pages = totalpages;
6661 pgdat->node_present_pages = realtotalpages;
6662 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6666 #ifndef CONFIG_SPARSEMEM
6668 * Calculate the size of the zone->blockflags rounded to an unsigned long
6669 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6670 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6671 * round what is now in bits to nearest long in bits, then return it in
6674 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6676 unsigned long usemapsize;
6678 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6679 usemapsize = roundup(zonesize, pageblock_nr_pages);
6680 usemapsize = usemapsize >> pageblock_order;
6681 usemapsize *= NR_PAGEBLOCK_BITS;
6682 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6684 return usemapsize / 8;
6687 static void __ref setup_usemap(struct pglist_data *pgdat,
6689 unsigned long zone_start_pfn,
6690 unsigned long zonesize)
6692 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6693 zone->pageblock_flags = NULL;
6695 zone->pageblock_flags =
6696 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6698 if (!zone->pageblock_flags)
6699 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6700 usemapsize, zone->name, pgdat->node_id);
6704 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6705 unsigned long zone_start_pfn, unsigned long zonesize) {}
6706 #endif /* CONFIG_SPARSEMEM */
6708 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6710 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6711 void __init set_pageblock_order(void)
6715 /* Check that pageblock_nr_pages has not already been setup */
6716 if (pageblock_order)
6719 if (HPAGE_SHIFT > PAGE_SHIFT)
6720 order = HUGETLB_PAGE_ORDER;
6722 order = MAX_ORDER - 1;
6725 * Assume the largest contiguous order of interest is a huge page.
6726 * This value may be variable depending on boot parameters on IA64 and
6729 pageblock_order = order;
6731 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6734 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6735 * is unused as pageblock_order is set at compile-time. See
6736 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6739 void __init set_pageblock_order(void)
6743 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6745 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6746 unsigned long present_pages)
6748 unsigned long pages = spanned_pages;
6751 * Provide a more accurate estimation if there are holes within
6752 * the zone and SPARSEMEM is in use. If there are holes within the
6753 * zone, each populated memory region may cost us one or two extra
6754 * memmap pages due to alignment because memmap pages for each
6755 * populated regions may not be naturally aligned on page boundary.
6756 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6758 if (spanned_pages > present_pages + (present_pages >> 4) &&
6759 IS_ENABLED(CONFIG_SPARSEMEM))
6760 pages = present_pages;
6762 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6765 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6766 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6768 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6770 spin_lock_init(&ds_queue->split_queue_lock);
6771 INIT_LIST_HEAD(&ds_queue->split_queue);
6772 ds_queue->split_queue_len = 0;
6775 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6778 #ifdef CONFIG_COMPACTION
6779 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6781 init_waitqueue_head(&pgdat->kcompactd_wait);
6784 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6787 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6789 pgdat_resize_init(pgdat);
6791 pgdat_init_split_queue(pgdat);
6792 pgdat_init_kcompactd(pgdat);
6794 init_waitqueue_head(&pgdat->kswapd_wait);
6795 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6797 pgdat_page_ext_init(pgdat);
6798 spin_lock_init(&pgdat->lru_lock);
6799 lruvec_init(&pgdat->__lruvec);
6802 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6803 unsigned long remaining_pages)
6805 atomic_long_set(&zone->managed_pages, remaining_pages);
6806 zone_set_nid(zone, nid);
6807 zone->name = zone_names[idx];
6808 zone->zone_pgdat = NODE_DATA(nid);
6809 spin_lock_init(&zone->lock);
6810 zone_seqlock_init(zone);
6811 zone_pcp_init(zone);
6815 * Set up the zone data structures
6816 * - init pgdat internals
6817 * - init all zones belonging to this node
6819 * NOTE: this function is only called during memory hotplug
6821 #ifdef CONFIG_MEMORY_HOTPLUG
6822 void __ref free_area_init_core_hotplug(int nid)
6825 pg_data_t *pgdat = NODE_DATA(nid);
6827 pgdat_init_internals(pgdat);
6828 for (z = 0; z < MAX_NR_ZONES; z++)
6829 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6834 * Set up the zone data structures:
6835 * - mark all pages reserved
6836 * - mark all memory queues empty
6837 * - clear the memory bitmaps
6839 * NOTE: pgdat should get zeroed by caller.
6840 * NOTE: this function is only called during early init.
6842 static void __init free_area_init_core(struct pglist_data *pgdat)
6845 int nid = pgdat->node_id;
6847 pgdat_init_internals(pgdat);
6848 pgdat->per_cpu_nodestats = &boot_nodestats;
6850 for (j = 0; j < MAX_NR_ZONES; j++) {
6851 struct zone *zone = pgdat->node_zones + j;
6852 unsigned long size, freesize, memmap_pages;
6853 unsigned long zone_start_pfn = zone->zone_start_pfn;
6855 size = zone->spanned_pages;
6856 freesize = zone->present_pages;
6859 * Adjust freesize so that it accounts for how much memory
6860 * is used by this zone for memmap. This affects the watermark
6861 * and per-cpu initialisations
6863 memmap_pages = calc_memmap_size(size, freesize);
6864 if (!is_highmem_idx(j)) {
6865 if (freesize >= memmap_pages) {
6866 freesize -= memmap_pages;
6869 " %s zone: %lu pages used for memmap\n",
6870 zone_names[j], memmap_pages);
6872 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6873 zone_names[j], memmap_pages, freesize);
6876 /* Account for reserved pages */
6877 if (j == 0 && freesize > dma_reserve) {
6878 freesize -= dma_reserve;
6879 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6880 zone_names[0], dma_reserve);
6883 if (!is_highmem_idx(j))
6884 nr_kernel_pages += freesize;
6885 /* Charge for highmem memmap if there are enough kernel pages */
6886 else if (nr_kernel_pages > memmap_pages * 2)
6887 nr_kernel_pages -= memmap_pages;
6888 nr_all_pages += freesize;
6891 * Set an approximate value for lowmem here, it will be adjusted
6892 * when the bootmem allocator frees pages into the buddy system.
6893 * And all highmem pages will be managed by the buddy system.
6895 zone_init_internals(zone, j, nid, freesize);
6900 set_pageblock_order();
6901 setup_usemap(pgdat, zone, zone_start_pfn, size);
6902 init_currently_empty_zone(zone, zone_start_pfn, size);
6903 memmap_init(size, nid, j, zone_start_pfn);
6907 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6908 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6910 unsigned long __maybe_unused start = 0;
6911 unsigned long __maybe_unused offset = 0;
6913 /* Skip empty nodes */
6914 if (!pgdat->node_spanned_pages)
6917 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6918 offset = pgdat->node_start_pfn - start;
6919 /* ia64 gets its own node_mem_map, before this, without bootmem */
6920 if (!pgdat->node_mem_map) {
6921 unsigned long size, end;
6925 * The zone's endpoints aren't required to be MAX_ORDER
6926 * aligned but the node_mem_map endpoints must be in order
6927 * for the buddy allocator to function correctly.
6929 end = pgdat_end_pfn(pgdat);
6930 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6931 size = (end - start) * sizeof(struct page);
6932 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6935 panic("Failed to allocate %ld bytes for node %d memory map\n",
6936 size, pgdat->node_id);
6937 pgdat->node_mem_map = map + offset;
6939 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6940 __func__, pgdat->node_id, (unsigned long)pgdat,
6941 (unsigned long)pgdat->node_mem_map);
6942 #ifndef CONFIG_NEED_MULTIPLE_NODES
6944 * With no DISCONTIG, the global mem_map is just set as node 0's
6946 if (pgdat == NODE_DATA(0)) {
6947 mem_map = NODE_DATA(0)->node_mem_map;
6948 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6949 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6951 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6956 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6957 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6959 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6960 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6962 pgdat->first_deferred_pfn = ULONG_MAX;
6965 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6968 void __init free_area_init_node(int nid, unsigned long *zones_size,
6969 unsigned long node_start_pfn,
6970 unsigned long *zholes_size)
6972 pg_data_t *pgdat = NODE_DATA(nid);
6973 unsigned long start_pfn = 0;
6974 unsigned long end_pfn = 0;
6976 /* pg_data_t should be reset to zero when it's allocated */
6977 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6979 pgdat->node_id = nid;
6980 pgdat->node_start_pfn = node_start_pfn;
6981 pgdat->per_cpu_nodestats = NULL;
6982 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6983 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6984 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6985 (u64)start_pfn << PAGE_SHIFT,
6986 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6988 start_pfn = node_start_pfn;
6990 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6991 zones_size, zholes_size);
6993 alloc_node_mem_map(pgdat);
6994 pgdat_set_deferred_range(pgdat);
6996 free_area_init_core(pgdat);
6999 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
7001 * Initialize all valid struct pages in the range [spfn, epfn) and mark them
7002 * PageReserved(). Return the number of struct pages that were initialized.
7004 static u64 __init init_unavailable_range(unsigned long spfn, unsigned long epfn)
7009 for (pfn = spfn; pfn < epfn; pfn++) {
7010 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
7011 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
7012 + pageblock_nr_pages - 1;
7016 * Use a fake node/zone (0) for now. Some of these pages
7017 * (in memblock.reserved but not in memblock.memory) will
7018 * get re-initialized via reserve_bootmem_region() later.
7020 __init_single_page(pfn_to_page(pfn), pfn, 0, 0);
7021 __SetPageReserved(pfn_to_page(pfn));
7029 * Only struct pages that are backed by physical memory are zeroed and
7030 * initialized by going through __init_single_page(). But, there are some
7031 * struct pages which are reserved in memblock allocator and their fields
7032 * may be accessed (for example page_to_pfn() on some configuration accesses
7033 * flags). We must explicitly initialize those struct pages.
7035 * This function also addresses a similar issue where struct pages are left
7036 * uninitialized because the physical address range is not covered by
7037 * memblock.memory or memblock.reserved. That could happen when memblock
7038 * layout is manually configured via memmap=, or when the highest physical
7039 * address (max_pfn) does not end on a section boundary.
7041 static void __init init_unavailable_mem(void)
7043 phys_addr_t start, end;
7045 phys_addr_t next = 0;
7048 * Loop through unavailable ranges not covered by memblock.memory.
7051 for_each_mem_range(i, &memblock.memory, NULL,
7052 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
7054 pgcnt += init_unavailable_range(PFN_DOWN(next),
7060 * Early sections always have a fully populated memmap for the whole
7061 * section - see pfn_valid(). If the last section has holes at the
7062 * end and that section is marked "online", the memmap will be
7063 * considered initialized. Make sure that memmap has a well defined
7066 pgcnt += init_unavailable_range(PFN_DOWN(next),
7067 round_up(max_pfn, PAGES_PER_SECTION));
7070 * Struct pages that do not have backing memory. This could be because
7071 * firmware is using some of this memory, or for some other reasons.
7074 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
7077 static inline void __init init_unavailable_mem(void)
7080 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
7082 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
7084 #if MAX_NUMNODES > 1
7086 * Figure out the number of possible node ids.
7088 void __init setup_nr_node_ids(void)
7090 unsigned int highest;
7092 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7093 nr_node_ids = highest + 1;
7098 * node_map_pfn_alignment - determine the maximum internode alignment
7100 * This function should be called after node map is populated and sorted.
7101 * It calculates the maximum power of two alignment which can distinguish
7104 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7105 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7106 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7107 * shifted, 1GiB is enough and this function will indicate so.
7109 * This is used to test whether pfn -> nid mapping of the chosen memory
7110 * model has fine enough granularity to avoid incorrect mapping for the
7111 * populated node map.
7113 * Return: the determined alignment in pfn's. 0 if there is no alignment
7114 * requirement (single node).
7116 unsigned long __init node_map_pfn_alignment(void)
7118 unsigned long accl_mask = 0, last_end = 0;
7119 unsigned long start, end, mask;
7120 int last_nid = NUMA_NO_NODE;
7123 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7124 if (!start || last_nid < 0 || last_nid == nid) {
7131 * Start with a mask granular enough to pin-point to the
7132 * start pfn and tick off bits one-by-one until it becomes
7133 * too coarse to separate the current node from the last.
7135 mask = ~((1 << __ffs(start)) - 1);
7136 while (mask && last_end <= (start & (mask << 1)))
7139 /* accumulate all internode masks */
7143 /* convert mask to number of pages */
7144 return ~accl_mask + 1;
7147 /* Find the lowest pfn for a node */
7148 static unsigned long __init find_min_pfn_for_node(int nid)
7150 unsigned long min_pfn = ULONG_MAX;
7151 unsigned long start_pfn;
7154 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
7155 min_pfn = min(min_pfn, start_pfn);
7157 if (min_pfn == ULONG_MAX) {
7158 pr_warn("Could not find start_pfn for node %d\n", nid);
7166 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7168 * Return: the minimum PFN based on information provided via
7169 * memblock_set_node().
7171 unsigned long __init find_min_pfn_with_active_regions(void)
7173 return find_min_pfn_for_node(MAX_NUMNODES);
7177 * early_calculate_totalpages()
7178 * Sum pages in active regions for movable zone.
7179 * Populate N_MEMORY for calculating usable_nodes.
7181 static unsigned long __init early_calculate_totalpages(void)
7183 unsigned long totalpages = 0;
7184 unsigned long start_pfn, end_pfn;
7187 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7188 unsigned long pages = end_pfn - start_pfn;
7190 totalpages += pages;
7192 node_set_state(nid, N_MEMORY);
7198 * Find the PFN the Movable zone begins in each node. Kernel memory
7199 * is spread evenly between nodes as long as the nodes have enough
7200 * memory. When they don't, some nodes will have more kernelcore than
7203 static void __init find_zone_movable_pfns_for_nodes(void)
7206 unsigned long usable_startpfn;
7207 unsigned long kernelcore_node, kernelcore_remaining;
7208 /* save the state before borrow the nodemask */
7209 nodemask_t saved_node_state = node_states[N_MEMORY];
7210 unsigned long totalpages = early_calculate_totalpages();
7211 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7212 struct memblock_region *r;
7214 /* Need to find movable_zone earlier when movable_node is specified. */
7215 find_usable_zone_for_movable();
7218 * If movable_node is specified, ignore kernelcore and movablecore
7221 if (movable_node_is_enabled()) {
7222 for_each_memblock(memory, r) {
7223 if (!memblock_is_hotpluggable(r))
7228 usable_startpfn = PFN_DOWN(r->base);
7229 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7230 min(usable_startpfn, zone_movable_pfn[nid]) :
7238 * If kernelcore=mirror is specified, ignore movablecore option
7240 if (mirrored_kernelcore) {
7241 bool mem_below_4gb_not_mirrored = false;
7243 for_each_memblock(memory, r) {
7244 if (memblock_is_mirror(r))
7249 usable_startpfn = memblock_region_memory_base_pfn(r);
7251 if (usable_startpfn < 0x100000) {
7252 mem_below_4gb_not_mirrored = true;
7256 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7257 min(usable_startpfn, zone_movable_pfn[nid]) :
7261 if (mem_below_4gb_not_mirrored)
7262 pr_warn("This configuration results in unmirrored kernel memory.");
7268 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7269 * amount of necessary memory.
7271 if (required_kernelcore_percent)
7272 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7274 if (required_movablecore_percent)
7275 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7279 * If movablecore= was specified, calculate what size of
7280 * kernelcore that corresponds so that memory usable for
7281 * any allocation type is evenly spread. If both kernelcore
7282 * and movablecore are specified, then the value of kernelcore
7283 * will be used for required_kernelcore if it's greater than
7284 * what movablecore would have allowed.
7286 if (required_movablecore) {
7287 unsigned long corepages;
7290 * Round-up so that ZONE_MOVABLE is at least as large as what
7291 * was requested by the user
7293 required_movablecore =
7294 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7295 required_movablecore = min(totalpages, required_movablecore);
7296 corepages = totalpages - required_movablecore;
7298 required_kernelcore = max(required_kernelcore, corepages);
7302 * If kernelcore was not specified or kernelcore size is larger
7303 * than totalpages, there is no ZONE_MOVABLE.
7305 if (!required_kernelcore || required_kernelcore >= totalpages)
7308 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7309 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7312 /* Spread kernelcore memory as evenly as possible throughout nodes */
7313 kernelcore_node = required_kernelcore / usable_nodes;
7314 for_each_node_state(nid, N_MEMORY) {
7315 unsigned long start_pfn, end_pfn;
7318 * Recalculate kernelcore_node if the division per node
7319 * now exceeds what is necessary to satisfy the requested
7320 * amount of memory for the kernel
7322 if (required_kernelcore < kernelcore_node)
7323 kernelcore_node = required_kernelcore / usable_nodes;
7326 * As the map is walked, we track how much memory is usable
7327 * by the kernel using kernelcore_remaining. When it is
7328 * 0, the rest of the node is usable by ZONE_MOVABLE
7330 kernelcore_remaining = kernelcore_node;
7332 /* Go through each range of PFNs within this node */
7333 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7334 unsigned long size_pages;
7336 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7337 if (start_pfn >= end_pfn)
7340 /* Account for what is only usable for kernelcore */
7341 if (start_pfn < usable_startpfn) {
7342 unsigned long kernel_pages;
7343 kernel_pages = min(end_pfn, usable_startpfn)
7346 kernelcore_remaining -= min(kernel_pages,
7347 kernelcore_remaining);
7348 required_kernelcore -= min(kernel_pages,
7349 required_kernelcore);
7351 /* Continue if range is now fully accounted */
7352 if (end_pfn <= usable_startpfn) {
7355 * Push zone_movable_pfn to the end so
7356 * that if we have to rebalance
7357 * kernelcore across nodes, we will
7358 * not double account here
7360 zone_movable_pfn[nid] = end_pfn;
7363 start_pfn = usable_startpfn;
7367 * The usable PFN range for ZONE_MOVABLE is from
7368 * start_pfn->end_pfn. Calculate size_pages as the
7369 * number of pages used as kernelcore
7371 size_pages = end_pfn - start_pfn;
7372 if (size_pages > kernelcore_remaining)
7373 size_pages = kernelcore_remaining;
7374 zone_movable_pfn[nid] = start_pfn + size_pages;
7377 * Some kernelcore has been met, update counts and
7378 * break if the kernelcore for this node has been
7381 required_kernelcore -= min(required_kernelcore,
7383 kernelcore_remaining -= size_pages;
7384 if (!kernelcore_remaining)
7390 * If there is still required_kernelcore, we do another pass with one
7391 * less node in the count. This will push zone_movable_pfn[nid] further
7392 * along on the nodes that still have memory until kernelcore is
7396 if (usable_nodes && required_kernelcore > usable_nodes)
7400 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7401 for (nid = 0; nid < MAX_NUMNODES; nid++)
7402 zone_movable_pfn[nid] =
7403 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7406 /* restore the node_state */
7407 node_states[N_MEMORY] = saved_node_state;
7410 /* Any regular or high memory on that node ? */
7411 static void check_for_memory(pg_data_t *pgdat, int nid)
7413 enum zone_type zone_type;
7415 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7416 struct zone *zone = &pgdat->node_zones[zone_type];
7417 if (populated_zone(zone)) {
7418 if (IS_ENABLED(CONFIG_HIGHMEM))
7419 node_set_state(nid, N_HIGH_MEMORY);
7420 if (zone_type <= ZONE_NORMAL)
7421 node_set_state(nid, N_NORMAL_MEMORY);
7428 * free_area_init_nodes - Initialise all pg_data_t and zone data
7429 * @max_zone_pfn: an array of max PFNs for each zone
7431 * This will call free_area_init_node() for each active node in the system.
7432 * Using the page ranges provided by memblock_set_node(), the size of each
7433 * zone in each node and their holes is calculated. If the maximum PFN
7434 * between two adjacent zones match, it is assumed that the zone is empty.
7435 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7436 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7437 * starts where the previous one ended. For example, ZONE_DMA32 starts
7438 * at arch_max_dma_pfn.
7440 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
7442 unsigned long start_pfn, end_pfn;
7445 /* Record where the zone boundaries are */
7446 memset(arch_zone_lowest_possible_pfn, 0,
7447 sizeof(arch_zone_lowest_possible_pfn));
7448 memset(arch_zone_highest_possible_pfn, 0,
7449 sizeof(arch_zone_highest_possible_pfn));
7451 start_pfn = find_min_pfn_with_active_regions();
7453 for (i = 0; i < MAX_NR_ZONES; i++) {
7454 if (i == ZONE_MOVABLE)
7457 end_pfn = max(max_zone_pfn[i], start_pfn);
7458 arch_zone_lowest_possible_pfn[i] = start_pfn;
7459 arch_zone_highest_possible_pfn[i] = end_pfn;
7461 start_pfn = end_pfn;
7464 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7465 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7466 find_zone_movable_pfns_for_nodes();
7468 /* Print out the zone ranges */
7469 pr_info("Zone ranges:\n");
7470 for (i = 0; i < MAX_NR_ZONES; i++) {
7471 if (i == ZONE_MOVABLE)
7473 pr_info(" %-8s ", zone_names[i]);
7474 if (arch_zone_lowest_possible_pfn[i] ==
7475 arch_zone_highest_possible_pfn[i])
7478 pr_cont("[mem %#018Lx-%#018Lx]\n",
7479 (u64)arch_zone_lowest_possible_pfn[i]
7481 ((u64)arch_zone_highest_possible_pfn[i]
7482 << PAGE_SHIFT) - 1);
7485 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7486 pr_info("Movable zone start for each node\n");
7487 for (i = 0; i < MAX_NUMNODES; i++) {
7488 if (zone_movable_pfn[i])
7489 pr_info(" Node %d: %#018Lx\n", i,
7490 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7494 * Print out the early node map, and initialize the
7495 * subsection-map relative to active online memory ranges to
7496 * enable future "sub-section" extensions of the memory map.
7498 pr_info("Early memory node ranges\n");
7499 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7500 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7501 (u64)start_pfn << PAGE_SHIFT,
7502 ((u64)end_pfn << PAGE_SHIFT) - 1);
7503 subsection_map_init(start_pfn, end_pfn - start_pfn);
7506 /* Initialise every node */
7507 mminit_verify_pageflags_layout();
7508 setup_nr_node_ids();
7509 init_unavailable_mem();
7510 for_each_online_node(nid) {
7511 pg_data_t *pgdat = NODE_DATA(nid);
7512 free_area_init_node(nid, NULL,
7513 find_min_pfn_for_node(nid), NULL);
7515 /* Any memory on that node */
7516 if (pgdat->node_present_pages)
7517 node_set_state(nid, N_MEMORY);
7518 check_for_memory(pgdat, nid);
7522 static int __init cmdline_parse_core(char *p, unsigned long *core,
7523 unsigned long *percent)
7525 unsigned long long coremem;
7531 /* Value may be a percentage of total memory, otherwise bytes */
7532 coremem = simple_strtoull(p, &endptr, 0);
7533 if (*endptr == '%') {
7534 /* Paranoid check for percent values greater than 100 */
7535 WARN_ON(coremem > 100);
7539 coremem = memparse(p, &p);
7540 /* Paranoid check that UL is enough for the coremem value */
7541 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7543 *core = coremem >> PAGE_SHIFT;
7550 * kernelcore=size sets the amount of memory for use for allocations that
7551 * cannot be reclaimed or migrated.
7553 static int __init cmdline_parse_kernelcore(char *p)
7555 /* parse kernelcore=mirror */
7556 if (parse_option_str(p, "mirror")) {
7557 mirrored_kernelcore = true;
7561 return cmdline_parse_core(p, &required_kernelcore,
7562 &required_kernelcore_percent);
7566 * movablecore=size sets the amount of memory for use for allocations that
7567 * can be reclaimed or migrated.
7569 static int __init cmdline_parse_movablecore(char *p)
7571 return cmdline_parse_core(p, &required_movablecore,
7572 &required_movablecore_percent);
7575 early_param("kernelcore", cmdline_parse_kernelcore);
7576 early_param("movablecore", cmdline_parse_movablecore);
7578 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
7580 void adjust_managed_page_count(struct page *page, long count)
7582 atomic_long_add(count, &page_zone(page)->managed_pages);
7583 totalram_pages_add(count);
7584 #ifdef CONFIG_HIGHMEM
7585 if (PageHighMem(page))
7586 totalhigh_pages_add(count);
7589 EXPORT_SYMBOL(adjust_managed_page_count);
7591 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7594 unsigned long pages = 0;
7596 start = (void *)PAGE_ALIGN((unsigned long)start);
7597 end = (void *)((unsigned long)end & PAGE_MASK);
7598 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7599 struct page *page = virt_to_page(pos);
7600 void *direct_map_addr;
7603 * 'direct_map_addr' might be different from 'pos'
7604 * because some architectures' virt_to_page()
7605 * work with aliases. Getting the direct map
7606 * address ensures that we get a _writeable_
7607 * alias for the memset().
7609 direct_map_addr = page_address(page);
7610 if ((unsigned int)poison <= 0xFF)
7611 memset(direct_map_addr, poison, PAGE_SIZE);
7613 free_reserved_page(page);
7617 pr_info("Freeing %s memory: %ldK\n",
7618 s, pages << (PAGE_SHIFT - 10));
7623 #ifdef CONFIG_HIGHMEM
7624 void free_highmem_page(struct page *page)
7626 __free_reserved_page(page);
7627 totalram_pages_inc();
7628 atomic_long_inc(&page_zone(page)->managed_pages);
7629 totalhigh_pages_inc();
7634 void __init mem_init_print_info(const char *str)
7636 unsigned long physpages, codesize, datasize, rosize, bss_size;
7637 unsigned long init_code_size, init_data_size;
7639 physpages = get_num_physpages();
7640 codesize = _etext - _stext;
7641 datasize = _edata - _sdata;
7642 rosize = __end_rodata - __start_rodata;
7643 bss_size = __bss_stop - __bss_start;
7644 init_data_size = __init_end - __init_begin;
7645 init_code_size = _einittext - _sinittext;
7648 * Detect special cases and adjust section sizes accordingly:
7649 * 1) .init.* may be embedded into .data sections
7650 * 2) .init.text.* may be out of [__init_begin, __init_end],
7651 * please refer to arch/tile/kernel/vmlinux.lds.S.
7652 * 3) .rodata.* may be embedded into .text or .data sections.
7654 #define adj_init_size(start, end, size, pos, adj) \
7656 if (start <= pos && pos < end && size > adj) \
7660 adj_init_size(__init_begin, __init_end, init_data_size,
7661 _sinittext, init_code_size);
7662 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7663 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7664 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7665 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7667 #undef adj_init_size
7669 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7670 #ifdef CONFIG_HIGHMEM
7674 nr_free_pages() << (PAGE_SHIFT - 10),
7675 physpages << (PAGE_SHIFT - 10),
7676 codesize >> 10, datasize >> 10, rosize >> 10,
7677 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7678 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7679 totalcma_pages << (PAGE_SHIFT - 10),
7680 #ifdef CONFIG_HIGHMEM
7681 totalhigh_pages() << (PAGE_SHIFT - 10),
7683 str ? ", " : "", str ? str : "");
7687 * set_dma_reserve - set the specified number of pages reserved in the first zone
7688 * @new_dma_reserve: The number of pages to mark reserved
7690 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7691 * In the DMA zone, a significant percentage may be consumed by kernel image
7692 * and other unfreeable allocations which can skew the watermarks badly. This
7693 * function may optionally be used to account for unfreeable pages in the
7694 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7695 * smaller per-cpu batchsize.
7697 void __init set_dma_reserve(unsigned long new_dma_reserve)
7699 dma_reserve = new_dma_reserve;
7702 void __init free_area_init(unsigned long *zones_size)
7704 init_unavailable_mem();
7705 free_area_init_node(0, zones_size,
7706 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
7709 static int page_alloc_cpu_dead(unsigned int cpu)
7712 lru_add_drain_cpu(cpu);
7716 * Spill the event counters of the dead processor
7717 * into the current processors event counters.
7718 * This artificially elevates the count of the current
7721 vm_events_fold_cpu(cpu);
7724 * Zero the differential counters of the dead processor
7725 * so that the vm statistics are consistent.
7727 * This is only okay since the processor is dead and cannot
7728 * race with what we are doing.
7730 cpu_vm_stats_fold(cpu);
7735 int hashdist = HASHDIST_DEFAULT;
7737 static int __init set_hashdist(char *str)
7741 hashdist = simple_strtoul(str, &str, 0);
7744 __setup("hashdist=", set_hashdist);
7747 void __init page_alloc_init(void)
7752 if (num_node_state(N_MEMORY) == 1)
7756 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7757 "mm/page_alloc:dead", NULL,
7758 page_alloc_cpu_dead);
7763 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7764 * or min_free_kbytes changes.
7766 static void calculate_totalreserve_pages(void)
7768 struct pglist_data *pgdat;
7769 unsigned long reserve_pages = 0;
7770 enum zone_type i, j;
7772 for_each_online_pgdat(pgdat) {
7774 pgdat->totalreserve_pages = 0;
7776 for (i = 0; i < MAX_NR_ZONES; i++) {
7777 struct zone *zone = pgdat->node_zones + i;
7779 unsigned long managed_pages = zone_managed_pages(zone);
7781 /* Find valid and maximum lowmem_reserve in the zone */
7782 for (j = i; j < MAX_NR_ZONES; j++) {
7783 if (zone->lowmem_reserve[j] > max)
7784 max = zone->lowmem_reserve[j];
7787 /* we treat the high watermark as reserved pages. */
7788 max += high_wmark_pages(zone);
7790 if (max > managed_pages)
7791 max = managed_pages;
7793 pgdat->totalreserve_pages += max;
7795 reserve_pages += max;
7798 totalreserve_pages = reserve_pages;
7802 * setup_per_zone_lowmem_reserve - called whenever
7803 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7804 * has a correct pages reserved value, so an adequate number of
7805 * pages are left in the zone after a successful __alloc_pages().
7807 static void setup_per_zone_lowmem_reserve(void)
7809 struct pglist_data *pgdat;
7810 enum zone_type j, idx;
7812 for_each_online_pgdat(pgdat) {
7813 for (j = 0; j < MAX_NR_ZONES; j++) {
7814 struct zone *zone = pgdat->node_zones + j;
7815 unsigned long managed_pages = zone_managed_pages(zone);
7817 zone->lowmem_reserve[j] = 0;
7821 struct zone *lower_zone;
7824 lower_zone = pgdat->node_zones + idx;
7826 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7827 sysctl_lowmem_reserve_ratio[idx] = 0;
7828 lower_zone->lowmem_reserve[j] = 0;
7830 lower_zone->lowmem_reserve[j] =
7831 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7833 managed_pages += zone_managed_pages(lower_zone);
7838 /* update totalreserve_pages */
7839 calculate_totalreserve_pages();
7842 static void __setup_per_zone_wmarks(void)
7844 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7845 unsigned long lowmem_pages = 0;
7847 unsigned long flags;
7849 /* Calculate total number of !ZONE_HIGHMEM pages */
7850 for_each_zone(zone) {
7851 if (!is_highmem(zone))
7852 lowmem_pages += zone_managed_pages(zone);
7855 for_each_zone(zone) {
7858 spin_lock_irqsave(&zone->lock, flags);
7859 tmp = (u64)pages_min * zone_managed_pages(zone);
7860 do_div(tmp, lowmem_pages);
7861 if (is_highmem(zone)) {
7863 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7864 * need highmem pages, so cap pages_min to a small
7867 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7868 * deltas control async page reclaim, and so should
7869 * not be capped for highmem.
7871 unsigned long min_pages;
7873 min_pages = zone_managed_pages(zone) / 1024;
7874 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7875 zone->_watermark[WMARK_MIN] = min_pages;
7878 * If it's a lowmem zone, reserve a number of pages
7879 * proportionate to the zone's size.
7881 zone->_watermark[WMARK_MIN] = tmp;
7885 * Set the kswapd watermarks distance according to the
7886 * scale factor in proportion to available memory, but
7887 * ensure a minimum size on small systems.
7889 tmp = max_t(u64, tmp >> 2,
7890 mult_frac(zone_managed_pages(zone),
7891 watermark_scale_factor, 10000));
7893 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7894 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7895 zone->watermark_boost = 0;
7897 spin_unlock_irqrestore(&zone->lock, flags);
7900 /* update totalreserve_pages */
7901 calculate_totalreserve_pages();
7905 * setup_per_zone_wmarks - called when min_free_kbytes changes
7906 * or when memory is hot-{added|removed}
7908 * Ensures that the watermark[min,low,high] values for each zone are set
7909 * correctly with respect to min_free_kbytes.
7911 void setup_per_zone_wmarks(void)
7913 static DEFINE_SPINLOCK(lock);
7916 __setup_per_zone_wmarks();
7921 * Initialise min_free_kbytes.
7923 * For small machines we want it small (128k min). For large machines
7924 * we want it large (64MB max). But it is not linear, because network
7925 * bandwidth does not increase linearly with machine size. We use
7927 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7928 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7944 int __meminit init_per_zone_wmark_min(void)
7946 unsigned long lowmem_kbytes;
7947 int new_min_free_kbytes;
7949 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7950 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7952 if (new_min_free_kbytes > user_min_free_kbytes) {
7953 min_free_kbytes = new_min_free_kbytes;
7954 if (min_free_kbytes < 128)
7955 min_free_kbytes = 128;
7956 if (min_free_kbytes > 262144)
7957 min_free_kbytes = 262144;
7959 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7960 new_min_free_kbytes, user_min_free_kbytes);
7962 setup_per_zone_wmarks();
7963 refresh_zone_stat_thresholds();
7964 setup_per_zone_lowmem_reserve();
7967 setup_min_unmapped_ratio();
7968 setup_min_slab_ratio();
7973 core_initcall(init_per_zone_wmark_min)
7976 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7977 * that we can call two helper functions whenever min_free_kbytes
7980 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7981 void __user *buffer, size_t *length, loff_t *ppos)
7985 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7990 user_min_free_kbytes = min_free_kbytes;
7991 setup_per_zone_wmarks();
7996 int watermark_boost_factor_sysctl_handler(struct ctl_table *table, int write,
7997 void __user *buffer, size_t *length, loff_t *ppos)
8001 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8008 int watermark_scale_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);
8018 setup_per_zone_wmarks();
8024 static void setup_min_unmapped_ratio(void)
8029 for_each_online_pgdat(pgdat)
8030 pgdat->min_unmapped_pages = 0;
8033 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8034 sysctl_min_unmapped_ratio) / 100;
8038 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8039 void __user *buffer, size_t *length, loff_t *ppos)
8043 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8047 setup_min_unmapped_ratio();
8052 static void setup_min_slab_ratio(void)
8057 for_each_online_pgdat(pgdat)
8058 pgdat->min_slab_pages = 0;
8061 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8062 sysctl_min_slab_ratio) / 100;
8065 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8066 void __user *buffer, size_t *length, loff_t *ppos)
8070 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8074 setup_min_slab_ratio();
8081 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8082 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8083 * whenever sysctl_lowmem_reserve_ratio changes.
8085 * The reserve ratio obviously has absolutely no relation with the
8086 * minimum watermarks. The lowmem reserve ratio can only make sense
8087 * if in function of the boot time zone sizes.
8089 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8090 void __user *buffer, size_t *length, loff_t *ppos)
8092 proc_dointvec_minmax(table, write, buffer, length, ppos);
8093 setup_per_zone_lowmem_reserve();
8097 static void __zone_pcp_update(struct zone *zone)
8101 for_each_possible_cpu(cpu)
8102 pageset_set_high_and_batch(zone,
8103 per_cpu_ptr(zone->pageset, cpu));
8107 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8108 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8109 * pagelist can have before it gets flushed back to buddy allocator.
8111 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8112 void __user *buffer, size_t *length, loff_t *ppos)
8115 int old_percpu_pagelist_fraction;
8118 mutex_lock(&pcp_batch_high_lock);
8119 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8121 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8122 if (!write || ret < 0)
8125 /* Sanity checking to avoid pcp imbalance */
8126 if (percpu_pagelist_fraction &&
8127 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8128 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8134 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8137 for_each_populated_zone(zone)
8138 __zone_pcp_update(zone);
8140 mutex_unlock(&pcp_batch_high_lock);
8144 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8146 * Returns the number of pages that arch has reserved but
8147 * is not known to alloc_large_system_hash().
8149 static unsigned long __init arch_reserved_kernel_pages(void)
8156 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8157 * machines. As memory size is increased the scale is also increased but at
8158 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8159 * quadruples the scale is increased by one, which means the size of hash table
8160 * only doubles, instead of quadrupling as well.
8161 * Because 32-bit systems cannot have large physical memory, where this scaling
8162 * makes sense, it is disabled on such platforms.
8164 #if __BITS_PER_LONG > 32
8165 #define ADAPT_SCALE_BASE (64ul << 30)
8166 #define ADAPT_SCALE_SHIFT 2
8167 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8171 * allocate a large system hash table from bootmem
8172 * - it is assumed that the hash table must contain an exact power-of-2
8173 * quantity of entries
8174 * - limit is the number of hash buckets, not the total allocation size
8176 void *__init alloc_large_system_hash(const char *tablename,
8177 unsigned long bucketsize,
8178 unsigned long numentries,
8181 unsigned int *_hash_shift,
8182 unsigned int *_hash_mask,
8183 unsigned long low_limit,
8184 unsigned long high_limit)
8186 unsigned long long max = high_limit;
8187 unsigned long log2qty, size;
8192 /* allow the kernel cmdline to have a say */
8194 /* round applicable memory size up to nearest megabyte */
8195 numentries = nr_kernel_pages;
8196 numentries -= arch_reserved_kernel_pages();
8198 /* It isn't necessary when PAGE_SIZE >= 1MB */
8199 if (PAGE_SHIFT < 20)
8200 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8202 #if __BITS_PER_LONG > 32
8204 unsigned long adapt;
8206 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8207 adapt <<= ADAPT_SCALE_SHIFT)
8212 /* limit to 1 bucket per 2^scale bytes of low memory */
8213 if (scale > PAGE_SHIFT)
8214 numentries >>= (scale - PAGE_SHIFT);
8216 numentries <<= (PAGE_SHIFT - scale);
8218 /* Make sure we've got at least a 0-order allocation.. */
8219 if (unlikely(flags & HASH_SMALL)) {
8220 /* Makes no sense without HASH_EARLY */
8221 WARN_ON(!(flags & HASH_EARLY));
8222 if (!(numentries >> *_hash_shift)) {
8223 numentries = 1UL << *_hash_shift;
8224 BUG_ON(!numentries);
8226 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8227 numentries = PAGE_SIZE / bucketsize;
8229 numentries = roundup_pow_of_two(numentries);
8231 /* limit allocation size to 1/16 total memory by default */
8233 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8234 do_div(max, bucketsize);
8236 max = min(max, 0x80000000ULL);
8238 if (numentries < low_limit)
8239 numentries = low_limit;
8240 if (numentries > max)
8243 log2qty = ilog2(numentries);
8245 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8248 size = bucketsize << log2qty;
8249 if (flags & HASH_EARLY) {
8250 if (flags & HASH_ZERO)
8251 table = memblock_alloc(size, SMP_CACHE_BYTES);
8253 table = memblock_alloc_raw(size,
8255 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8256 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
8260 * If bucketsize is not a power-of-two, we may free
8261 * some pages at the end of hash table which
8262 * alloc_pages_exact() automatically does
8264 table = alloc_pages_exact(size, gfp_flags);
8265 kmemleak_alloc(table, size, 1, gfp_flags);
8267 } while (!table && size > PAGE_SIZE && --log2qty);
8270 panic("Failed to allocate %s hash table\n", tablename);
8272 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8273 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8274 virt ? "vmalloc" : "linear");
8277 *_hash_shift = log2qty;
8279 *_hash_mask = (1 << log2qty) - 1;
8285 * This function checks whether pageblock includes unmovable pages or not.
8287 * PageLRU check without isolation or lru_lock could race so that
8288 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8289 * check without lock_page also may miss some movable non-lru pages at
8290 * race condition. So you can't expect this function should be exact.
8292 * Returns a page without holding a reference. If the caller wants to
8293 * dereference that page (e.g., dumping), it has to make sure that that it
8294 * cannot get removed (e.g., via memory unplug) concurrently.
8297 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8298 int migratetype, int flags)
8300 unsigned long iter = 0;
8301 unsigned long pfn = page_to_pfn(page);
8304 * TODO we could make this much more efficient by not checking every
8305 * page in the range if we know all of them are in MOVABLE_ZONE and
8306 * that the movable zone guarantees that pages are migratable but
8307 * the later is not the case right now unfortunatelly. E.g. movablecore
8308 * can still lead to having bootmem allocations in zone_movable.
8311 if (is_migrate_cma_page(page)) {
8313 * CMA allocations (alloc_contig_range) really need to mark
8314 * isolate CMA pageblocks even when they are not movable in fact
8315 * so consider them movable here.
8317 if (is_migrate_cma(migratetype))
8323 for (; iter < pageblock_nr_pages; iter++) {
8324 if (!pfn_valid_within(pfn + iter))
8327 page = pfn_to_page(pfn + iter);
8329 if (PageReserved(page))
8333 * If the zone is movable and we have ruled out all reserved
8334 * pages then it should be reasonably safe to assume the rest
8337 if (zone_idx(zone) == ZONE_MOVABLE)
8341 * Hugepages are not in LRU lists, but they're movable.
8342 * THPs are on the LRU, but need to be counted as #small pages.
8343 * We need not scan over tail pages because we don't
8344 * handle each tail page individually in migration.
8346 if (PageHuge(page) || PageTransCompound(page)) {
8347 struct page *head = compound_head(page);
8348 unsigned int skip_pages;
8350 if (PageHuge(page)) {
8351 if (!hugepage_migration_supported(page_hstate(head)))
8353 } else if (!PageLRU(head) && !__PageMovable(head)) {
8357 skip_pages = compound_nr(head) - (page - head);
8358 iter += skip_pages - 1;
8363 * We can't use page_count without pin a page
8364 * because another CPU can free compound page.
8365 * This check already skips compound tails of THP
8366 * because their page->_refcount is zero at all time.
8368 if (!page_ref_count(page)) {
8369 if (PageBuddy(page))
8370 iter += (1 << page_order(page)) - 1;
8375 * The HWPoisoned page may be not in buddy system, and
8376 * page_count() is not 0.
8378 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8381 if (__PageMovable(page) || PageLRU(page))
8385 * If there are RECLAIMABLE pages, we need to check
8386 * it. But now, memory offline itself doesn't call
8387 * shrink_node_slabs() and it still to be fixed.
8390 * If the page is not RAM, page_count()should be 0.
8391 * we don't need more check. This is an _used_ not-movable page.
8393 * The problematic thing here is PG_reserved pages. PG_reserved
8394 * is set to both of a memory hole page and a _used_ kernel
8402 #ifdef CONFIG_CONTIG_ALLOC
8403 static unsigned long pfn_max_align_down(unsigned long pfn)
8405 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8406 pageblock_nr_pages) - 1);
8409 static unsigned long pfn_max_align_up(unsigned long pfn)
8411 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8412 pageblock_nr_pages));
8415 /* [start, end) must belong to a single zone. */
8416 static int __alloc_contig_migrate_range(struct compact_control *cc,
8417 unsigned long start, unsigned long end)
8419 /* This function is based on compact_zone() from compaction.c. */
8420 unsigned long nr_reclaimed;
8421 unsigned long pfn = start;
8422 unsigned int tries = 0;
8427 while (pfn < end || !list_empty(&cc->migratepages)) {
8428 if (fatal_signal_pending(current)) {
8433 if (list_empty(&cc->migratepages)) {
8434 cc->nr_migratepages = 0;
8435 pfn = isolate_migratepages_range(cc, pfn, end);
8441 } else if (++tries == 5) {
8442 ret = ret < 0 ? ret : -EBUSY;
8446 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8448 cc->nr_migratepages -= nr_reclaimed;
8450 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
8451 NULL, 0, cc->mode, MR_CONTIG_RANGE);
8454 putback_movable_pages(&cc->migratepages);
8461 * alloc_contig_range() -- tries to allocate given range of pages
8462 * @start: start PFN to allocate
8463 * @end: one-past-the-last PFN to allocate
8464 * @migratetype: migratetype of the underlaying pageblocks (either
8465 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8466 * in range must have the same migratetype and it must
8467 * be either of the two.
8468 * @gfp_mask: GFP mask to use during compaction
8470 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8471 * aligned. The PFN range must belong to a single zone.
8473 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8474 * pageblocks in the range. Once isolated, the pageblocks should not
8475 * be modified by others.
8477 * Return: zero on success or negative error code. On success all
8478 * pages which PFN is in [start, end) are allocated for the caller and
8479 * need to be freed with free_contig_range().
8481 int alloc_contig_range(unsigned long start, unsigned long end,
8482 unsigned migratetype, gfp_t gfp_mask)
8484 unsigned long outer_start, outer_end;
8488 struct compact_control cc = {
8489 .nr_migratepages = 0,
8491 .zone = page_zone(pfn_to_page(start)),
8492 .mode = MIGRATE_SYNC,
8493 .ignore_skip_hint = true,
8494 .no_set_skip_hint = true,
8495 .gfp_mask = current_gfp_context(gfp_mask),
8496 .alloc_contig = true,
8498 INIT_LIST_HEAD(&cc.migratepages);
8501 * What we do here is we mark all pageblocks in range as
8502 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8503 * have different sizes, and due to the way page allocator
8504 * work, we align the range to biggest of the two pages so
8505 * that page allocator won't try to merge buddies from
8506 * different pageblocks and change MIGRATE_ISOLATE to some
8507 * other migration type.
8509 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8510 * migrate the pages from an unaligned range (ie. pages that
8511 * we are interested in). This will put all the pages in
8512 * range back to page allocator as MIGRATE_ISOLATE.
8514 * When this is done, we take the pages in range from page
8515 * allocator removing them from the buddy system. This way
8516 * page allocator will never consider using them.
8518 * This lets us mark the pageblocks back as
8519 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8520 * aligned range but not in the unaligned, original range are
8521 * put back to page allocator so that buddy can use them.
8524 ret = start_isolate_page_range(pfn_max_align_down(start),
8525 pfn_max_align_up(end), migratetype, 0);
8530 * In case of -EBUSY, we'd like to know which page causes problem.
8531 * So, just fall through. test_pages_isolated() has a tracepoint
8532 * which will report the busy page.
8534 * It is possible that busy pages could become available before
8535 * the call to test_pages_isolated, and the range will actually be
8536 * allocated. So, if we fall through be sure to clear ret so that
8537 * -EBUSY is not accidentally used or returned to caller.
8539 ret = __alloc_contig_migrate_range(&cc, start, end);
8540 if (ret && ret != -EBUSY)
8545 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8546 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8547 * more, all pages in [start, end) are free in page allocator.
8548 * What we are going to do is to allocate all pages from
8549 * [start, end) (that is remove them from page allocator).
8551 * The only problem is that pages at the beginning and at the
8552 * end of interesting range may be not aligned with pages that
8553 * page allocator holds, ie. they can be part of higher order
8554 * pages. Because of this, we reserve the bigger range and
8555 * once this is done free the pages we are not interested in.
8557 * We don't have to hold zone->lock here because the pages are
8558 * isolated thus they won't get removed from buddy.
8561 lru_add_drain_all();
8564 outer_start = start;
8565 while (!PageBuddy(pfn_to_page(outer_start))) {
8566 if (++order >= MAX_ORDER) {
8567 outer_start = start;
8570 outer_start &= ~0UL << order;
8573 if (outer_start != start) {
8574 order = page_order(pfn_to_page(outer_start));
8577 * outer_start page could be small order buddy page and
8578 * it doesn't include start page. Adjust outer_start
8579 * in this case to report failed page properly
8580 * on tracepoint in test_pages_isolated()
8582 if (outer_start + (1UL << order) <= start)
8583 outer_start = start;
8586 /* Make sure the range is really isolated. */
8587 if (test_pages_isolated(outer_start, end, 0)) {
8588 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8589 __func__, outer_start, end);
8594 /* Grab isolated pages from freelists. */
8595 outer_end = isolate_freepages_range(&cc, outer_start, end);
8601 /* Free head and tail (if any) */
8602 if (start != outer_start)
8603 free_contig_range(outer_start, start - outer_start);
8604 if (end != outer_end)
8605 free_contig_range(end, outer_end - end);
8608 undo_isolate_page_range(pfn_max_align_down(start),
8609 pfn_max_align_up(end), migratetype);
8613 static int __alloc_contig_pages(unsigned long start_pfn,
8614 unsigned long nr_pages, gfp_t gfp_mask)
8616 unsigned long end_pfn = start_pfn + nr_pages;
8618 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8622 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8623 unsigned long nr_pages)
8625 unsigned long i, end_pfn = start_pfn + nr_pages;
8628 for (i = start_pfn; i < end_pfn; i++) {
8629 page = pfn_to_online_page(i);
8633 if (page_zone(page) != z)
8636 if (PageReserved(page))
8639 if (page_count(page) > 0)
8648 static bool zone_spans_last_pfn(const struct zone *zone,
8649 unsigned long start_pfn, unsigned long nr_pages)
8651 unsigned long last_pfn = start_pfn + nr_pages - 1;
8653 return zone_spans_pfn(zone, last_pfn);
8657 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8658 * @nr_pages: Number of contiguous pages to allocate
8659 * @gfp_mask: GFP mask to limit search and used during compaction
8661 * @nodemask: Mask for other possible nodes
8663 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8664 * on an applicable zonelist to find a contiguous pfn range which can then be
8665 * tried for allocation with alloc_contig_range(). This routine is intended
8666 * for allocation requests which can not be fulfilled with the buddy allocator.
8668 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8669 * power of two then the alignment is guaranteed to be to the given nr_pages
8670 * (e.g. 1GB request would be aligned to 1GB).
8672 * Allocated pages can be freed with free_contig_range() or by manually calling
8673 * __free_page() on each allocated page.
8675 * Return: pointer to contiguous pages on success, or NULL if not successful.
8677 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8678 int nid, nodemask_t *nodemask)
8680 unsigned long ret, pfn, flags;
8681 struct zonelist *zonelist;
8685 zonelist = node_zonelist(nid, gfp_mask);
8686 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8687 gfp_zone(gfp_mask), nodemask) {
8688 spin_lock_irqsave(&zone->lock, flags);
8690 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8691 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8692 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8694 * We release the zone lock here because
8695 * alloc_contig_range() will also lock the zone
8696 * at some point. If there's an allocation
8697 * spinning on this lock, it may win the race
8698 * and cause alloc_contig_range() to fail...
8700 spin_unlock_irqrestore(&zone->lock, flags);
8701 ret = __alloc_contig_pages(pfn, nr_pages,
8704 return pfn_to_page(pfn);
8705 spin_lock_irqsave(&zone->lock, flags);
8709 spin_unlock_irqrestore(&zone->lock, flags);
8713 #endif /* CONFIG_CONTIG_ALLOC */
8715 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8717 unsigned int count = 0;
8719 for (; nr_pages--; pfn++) {
8720 struct page *page = pfn_to_page(pfn);
8722 count += page_count(page) != 1;
8725 WARN(count != 0, "%d pages are still in use!\n", count);
8729 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8730 * page high values need to be recalulated.
8732 void __meminit zone_pcp_update(struct zone *zone)
8734 mutex_lock(&pcp_batch_high_lock);
8735 __zone_pcp_update(zone);
8736 mutex_unlock(&pcp_batch_high_lock);
8739 void zone_pcp_reset(struct zone *zone)
8741 unsigned long flags;
8743 struct per_cpu_pageset *pset;
8745 /* avoid races with drain_pages() */
8746 local_irq_save(flags);
8747 if (zone->pageset != &boot_pageset) {
8748 for_each_online_cpu(cpu) {
8749 pset = per_cpu_ptr(zone->pageset, cpu);
8750 drain_zonestat(zone, pset);
8752 free_percpu(zone->pageset);
8753 zone->pageset = &boot_pageset;
8755 local_irq_restore(flags);
8758 #ifdef CONFIG_MEMORY_HOTREMOVE
8760 * All pages in the range must be in a single zone and isolated
8761 * before calling this.
8764 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8770 unsigned long flags;
8771 unsigned long offlined_pages = 0;
8773 /* find the first valid pfn */
8774 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8778 return offlined_pages;
8780 offline_mem_sections(pfn, end_pfn);
8781 zone = page_zone(pfn_to_page(pfn));
8782 spin_lock_irqsave(&zone->lock, flags);
8784 while (pfn < end_pfn) {
8785 if (!pfn_valid(pfn)) {
8789 page = pfn_to_page(pfn);
8791 * The HWPoisoned page may be not in buddy system, and
8792 * page_count() is not 0.
8794 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8800 BUG_ON(page_count(page));
8801 BUG_ON(!PageBuddy(page));
8802 order = page_order(page);
8803 offlined_pages += 1 << order;
8804 del_page_from_free_list(page, zone, order);
8805 pfn += (1 << order);
8807 spin_unlock_irqrestore(&zone->lock, flags);
8809 return offlined_pages;
8813 bool is_free_buddy_page(struct page *page)
8815 struct zone *zone = page_zone(page);
8816 unsigned long pfn = page_to_pfn(page);
8817 unsigned long flags;
8820 spin_lock_irqsave(&zone->lock, flags);
8821 for (order = 0; order < MAX_ORDER; order++) {
8822 struct page *page_head = page - (pfn & ((1 << order) - 1));
8824 if (PageBuddy(page_head) && page_order(page_head) >= order)
8827 spin_unlock_irqrestore(&zone->lock, flags);
8829 return order < MAX_ORDER;
8832 #ifdef CONFIG_MEMORY_FAILURE
8834 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8835 * test is performed under the zone lock to prevent a race against page
8838 bool set_hwpoison_free_buddy_page(struct page *page)
8840 struct zone *zone = page_zone(page);
8841 unsigned long pfn = page_to_pfn(page);
8842 unsigned long flags;
8844 bool hwpoisoned = false;
8846 spin_lock_irqsave(&zone->lock, flags);
8847 for (order = 0; order < MAX_ORDER; order++) {
8848 struct page *page_head = page - (pfn & ((1 << order) - 1));
8850 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8851 if (!TestSetPageHWPoison(page))
8856 spin_unlock_irqrestore(&zone->lock, flags);