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 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
1510 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1512 int __meminit __early_pfn_to_nid(unsigned long pfn,
1513 struct mminit_pfnnid_cache *state)
1515 unsigned long start_pfn, end_pfn;
1518 if (state->last_start <= pfn && pfn < state->last_end)
1519 return state->last_nid;
1521 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1522 if (nid != NUMA_NO_NODE) {
1523 state->last_start = start_pfn;
1524 state->last_end = end_pfn;
1525 state->last_nid = nid;
1530 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
1532 int __meminit early_pfn_to_nid(unsigned long pfn)
1534 static DEFINE_SPINLOCK(early_pfn_lock);
1537 spin_lock(&early_pfn_lock);
1538 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1540 nid = first_online_node;
1541 spin_unlock(&early_pfn_lock);
1547 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1548 /* Only safe to use early in boot when initialisation is single-threaded */
1549 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1553 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1554 if (nid >= 0 && nid != node)
1560 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1567 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1570 if (early_page_uninitialised(pfn))
1572 __free_pages_core(page, order);
1576 * Check that the whole (or subset of) a pageblock given by the interval of
1577 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1578 * with the migration of free compaction scanner. The scanners then need to
1579 * use only pfn_valid_within() check for arches that allow holes within
1582 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1584 * It's possible on some configurations to have a setup like node0 node1 node0
1585 * i.e. it's possible that all pages within a zones range of pages do not
1586 * belong to a single zone. We assume that a border between node0 and node1
1587 * can occur within a single pageblock, but not a node0 node1 node0
1588 * interleaving within a single pageblock. It is therefore sufficient to check
1589 * the first and last page of a pageblock and avoid checking each individual
1590 * page in a pageblock.
1592 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1593 unsigned long end_pfn, struct zone *zone)
1595 struct page *start_page;
1596 struct page *end_page;
1598 /* end_pfn is one past the range we are checking */
1601 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1604 start_page = pfn_to_online_page(start_pfn);
1608 if (page_zone(start_page) != zone)
1611 end_page = pfn_to_page(end_pfn);
1613 /* This gives a shorter code than deriving page_zone(end_page) */
1614 if (page_zone_id(start_page) != page_zone_id(end_page))
1620 void set_zone_contiguous(struct zone *zone)
1622 unsigned long block_start_pfn = zone->zone_start_pfn;
1623 unsigned long block_end_pfn;
1625 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1626 for (; block_start_pfn < zone_end_pfn(zone);
1627 block_start_pfn = block_end_pfn,
1628 block_end_pfn += pageblock_nr_pages) {
1630 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1632 if (!__pageblock_pfn_to_page(block_start_pfn,
1633 block_end_pfn, zone))
1638 /* We confirm that there is no hole */
1639 zone->contiguous = true;
1642 void clear_zone_contiguous(struct zone *zone)
1644 zone->contiguous = false;
1647 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1648 static void __init deferred_free_range(unsigned long pfn,
1649 unsigned long nr_pages)
1657 page = pfn_to_page(pfn);
1659 /* Free a large naturally-aligned chunk if possible */
1660 if (nr_pages == pageblock_nr_pages &&
1661 (pfn & (pageblock_nr_pages - 1)) == 0) {
1662 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1663 __free_pages_core(page, pageblock_order);
1667 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1668 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1669 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1670 __free_pages_core(page, 0);
1674 /* Completion tracking for deferred_init_memmap() threads */
1675 static atomic_t pgdat_init_n_undone __initdata;
1676 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1678 static inline void __init pgdat_init_report_one_done(void)
1680 if (atomic_dec_and_test(&pgdat_init_n_undone))
1681 complete(&pgdat_init_all_done_comp);
1685 * Returns true if page needs to be initialized or freed to buddy allocator.
1687 * First we check if pfn is valid on architectures where it is possible to have
1688 * holes within pageblock_nr_pages. On systems where it is not possible, this
1689 * function is optimized out.
1691 * Then, we check if a current large page is valid by only checking the validity
1694 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1696 if (!pfn_valid_within(pfn))
1698 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1704 * Free pages to buddy allocator. Try to free aligned pages in
1705 * pageblock_nr_pages sizes.
1707 static void __init deferred_free_pages(unsigned long pfn,
1708 unsigned long end_pfn)
1710 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1711 unsigned long nr_free = 0;
1713 for (; pfn < end_pfn; pfn++) {
1714 if (!deferred_pfn_valid(pfn)) {
1715 deferred_free_range(pfn - nr_free, nr_free);
1717 } else if (!(pfn & nr_pgmask)) {
1718 deferred_free_range(pfn - nr_free, nr_free);
1720 touch_nmi_watchdog();
1725 /* Free the last block of pages to allocator */
1726 deferred_free_range(pfn - nr_free, nr_free);
1730 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1731 * by performing it only once every pageblock_nr_pages.
1732 * Return number of pages initialized.
1734 static unsigned long __init deferred_init_pages(struct zone *zone,
1736 unsigned long end_pfn)
1738 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1739 int nid = zone_to_nid(zone);
1740 unsigned long nr_pages = 0;
1741 int zid = zone_idx(zone);
1742 struct page *page = NULL;
1744 for (; pfn < end_pfn; pfn++) {
1745 if (!deferred_pfn_valid(pfn)) {
1748 } else if (!page || !(pfn & nr_pgmask)) {
1749 page = pfn_to_page(pfn);
1750 touch_nmi_watchdog();
1754 __init_single_page(page, pfn, zid, nid);
1761 * This function is meant to pre-load the iterator for the zone init.
1762 * Specifically it walks through the ranges until we are caught up to the
1763 * first_init_pfn value and exits there. If we never encounter the value we
1764 * return false indicating there are no valid ranges left.
1767 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1768 unsigned long *spfn, unsigned long *epfn,
1769 unsigned long first_init_pfn)
1774 * Start out by walking through the ranges in this zone that have
1775 * already been initialized. We don't need to do anything with them
1776 * so we just need to flush them out of the system.
1778 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1779 if (*epfn <= first_init_pfn)
1781 if (*spfn < first_init_pfn)
1782 *spfn = first_init_pfn;
1791 * Initialize and free pages. We do it in two loops: first we initialize
1792 * struct page, then free to buddy allocator, because while we are
1793 * freeing pages we can access pages that are ahead (computing buddy
1794 * page in __free_one_page()).
1796 * In order to try and keep some memory in the cache we have the loop
1797 * broken along max page order boundaries. This way we will not cause
1798 * any issues with the buddy page computation.
1800 static unsigned long __init
1801 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1802 unsigned long *end_pfn)
1804 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1805 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1806 unsigned long nr_pages = 0;
1809 /* First we loop through and initialize the page values */
1810 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1813 if (mo_pfn <= *start_pfn)
1816 t = min(mo_pfn, *end_pfn);
1817 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1819 if (mo_pfn < *end_pfn) {
1820 *start_pfn = mo_pfn;
1825 /* Reset values and now loop through freeing pages as needed */
1828 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1834 t = min(mo_pfn, epfn);
1835 deferred_free_pages(spfn, t);
1844 /* Initialise remaining memory on a node */
1845 static int __init deferred_init_memmap(void *data)
1847 pg_data_t *pgdat = data;
1848 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1849 unsigned long spfn = 0, epfn = 0, nr_pages = 0;
1850 unsigned long first_init_pfn, flags;
1851 unsigned long start = jiffies;
1856 /* Bind memory initialisation thread to a local node if possible */
1857 if (!cpumask_empty(cpumask))
1858 set_cpus_allowed_ptr(current, cpumask);
1860 pgdat_resize_lock(pgdat, &flags);
1861 first_init_pfn = pgdat->first_deferred_pfn;
1862 if (first_init_pfn == ULONG_MAX) {
1863 pgdat_resize_unlock(pgdat, &flags);
1864 pgdat_init_report_one_done();
1868 /* Sanity check boundaries */
1869 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1870 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1871 pgdat->first_deferred_pfn = ULONG_MAX;
1873 /* Only the highest zone is deferred so find it */
1874 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1875 zone = pgdat->node_zones + zid;
1876 if (first_init_pfn < zone_end_pfn(zone))
1880 /* If the zone is empty somebody else may have cleared out the zone */
1881 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1886 * Initialize and free pages in MAX_ORDER sized increments so
1887 * that we can avoid introducing any issues with the buddy
1891 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1893 pgdat_resize_unlock(pgdat, &flags);
1895 /* Sanity check that the next zone really is unpopulated */
1896 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1898 pr_info("node %d initialised, %lu pages in %ums\n",
1899 pgdat->node_id, nr_pages, jiffies_to_msecs(jiffies - start));
1901 pgdat_init_report_one_done();
1906 * If this zone has deferred pages, try to grow it by initializing enough
1907 * deferred pages to satisfy the allocation specified by order, rounded up to
1908 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1909 * of SECTION_SIZE bytes by initializing struct pages in increments of
1910 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1912 * Return true when zone was grown, otherwise return false. We return true even
1913 * when we grow less than requested, to let the caller decide if there are
1914 * enough pages to satisfy the allocation.
1916 * Note: We use noinline because this function is needed only during boot, and
1917 * it is called from a __ref function _deferred_grow_zone. This way we are
1918 * making sure that it is not inlined into permanent text section.
1920 static noinline bool __init
1921 deferred_grow_zone(struct zone *zone, unsigned int order)
1923 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1924 pg_data_t *pgdat = zone->zone_pgdat;
1925 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1926 unsigned long spfn, epfn, flags;
1927 unsigned long nr_pages = 0;
1930 /* Only the last zone may have deferred pages */
1931 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1934 pgdat_resize_lock(pgdat, &flags);
1937 * If deferred pages have been initialized while we were waiting for
1938 * the lock, return true, as the zone was grown. The caller will retry
1939 * this zone. We won't return to this function since the caller also
1940 * has this static branch.
1942 if (!static_branch_unlikely(&deferred_pages)) {
1943 pgdat_resize_unlock(pgdat, &flags);
1948 * If someone grew this zone while we were waiting for spinlock, return
1949 * true, as there might be enough pages already.
1951 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1952 pgdat_resize_unlock(pgdat, &flags);
1956 /* If the zone is empty somebody else may have cleared out the zone */
1957 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1958 first_deferred_pfn)) {
1959 pgdat->first_deferred_pfn = ULONG_MAX;
1960 pgdat_resize_unlock(pgdat, &flags);
1961 /* Retry only once. */
1962 return first_deferred_pfn != ULONG_MAX;
1966 * Initialize and free pages in MAX_ORDER sized increments so
1967 * that we can avoid introducing any issues with the buddy
1970 while (spfn < epfn) {
1971 /* update our first deferred PFN for this section */
1972 first_deferred_pfn = spfn;
1974 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1976 /* We should only stop along section boundaries */
1977 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
1980 /* If our quota has been met we can stop here */
1981 if (nr_pages >= nr_pages_needed)
1985 pgdat->first_deferred_pfn = spfn;
1986 pgdat_resize_unlock(pgdat, &flags);
1988 return nr_pages > 0;
1992 * deferred_grow_zone() is __init, but it is called from
1993 * get_page_from_freelist() during early boot until deferred_pages permanently
1994 * disables this call. This is why we have refdata wrapper to avoid warning,
1995 * and to ensure that the function body gets unloaded.
1998 _deferred_grow_zone(struct zone *zone, unsigned int order)
2000 return deferred_grow_zone(zone, order);
2003 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2005 void __init page_alloc_init_late(void)
2010 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2012 /* There will be num_node_state(N_MEMORY) threads */
2013 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2014 for_each_node_state(nid, N_MEMORY) {
2015 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2018 /* Block until all are initialised */
2019 wait_for_completion(&pgdat_init_all_done_comp);
2022 * The number of managed pages has changed due to the initialisation
2023 * so the pcpu batch and high limits needs to be updated or the limits
2024 * will be artificially small.
2026 for_each_populated_zone(zone)
2027 zone_pcp_update(zone);
2030 * We initialized the rest of the deferred pages. Permanently disable
2031 * on-demand struct page initialization.
2033 static_branch_disable(&deferred_pages);
2035 /* Reinit limits that are based on free pages after the kernel is up */
2036 files_maxfiles_init();
2039 /* Discard memblock private memory */
2042 for_each_node_state(nid, N_MEMORY)
2043 shuffle_free_memory(NODE_DATA(nid));
2045 for_each_populated_zone(zone)
2046 set_zone_contiguous(zone);
2050 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2051 void __init init_cma_reserved_pageblock(struct page *page)
2053 unsigned i = pageblock_nr_pages;
2054 struct page *p = page;
2057 __ClearPageReserved(p);
2058 set_page_count(p, 0);
2061 set_pageblock_migratetype(page, MIGRATE_CMA);
2063 if (pageblock_order >= MAX_ORDER) {
2064 i = pageblock_nr_pages;
2067 set_page_refcounted(p);
2068 __free_pages(p, MAX_ORDER - 1);
2069 p += MAX_ORDER_NR_PAGES;
2070 } while (i -= MAX_ORDER_NR_PAGES);
2072 set_page_refcounted(page);
2073 __free_pages(page, pageblock_order);
2076 adjust_managed_page_count(page, pageblock_nr_pages);
2081 * The order of subdivision here is critical for the IO subsystem.
2082 * Please do not alter this order without good reasons and regression
2083 * testing. Specifically, as large blocks of memory are subdivided,
2084 * the order in which smaller blocks are delivered depends on the order
2085 * they're subdivided in this function. This is the primary factor
2086 * influencing the order in which pages are delivered to the IO
2087 * subsystem according to empirical testing, and this is also justified
2088 * by considering the behavior of a buddy system containing a single
2089 * large block of memory acted on by a series of small allocations.
2090 * This behavior is a critical factor in sglist merging's success.
2094 static inline void expand(struct zone *zone, struct page *page,
2095 int low, int high, int migratetype)
2097 unsigned long size = 1 << high;
2099 while (high > low) {
2102 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2105 * Mark as guard pages (or page), that will allow to
2106 * merge back to allocator when buddy will be freed.
2107 * Corresponding page table entries will not be touched,
2108 * pages will stay not present in virtual address space
2110 if (set_page_guard(zone, &page[size], high, migratetype))
2113 add_to_free_list(&page[size], zone, high, migratetype);
2114 set_page_order(&page[size], high);
2118 static void check_new_page_bad(struct page *page)
2120 const char *bad_reason = NULL;
2121 unsigned long bad_flags = 0;
2123 if (unlikely(atomic_read(&page->_mapcount) != -1))
2124 bad_reason = "nonzero mapcount";
2125 if (unlikely(page->mapping != NULL))
2126 bad_reason = "non-NULL mapping";
2127 if (unlikely(page_ref_count(page) != 0))
2128 bad_reason = "nonzero _refcount";
2129 if (unlikely(page->flags & __PG_HWPOISON)) {
2130 bad_reason = "HWPoisoned (hardware-corrupted)";
2131 bad_flags = __PG_HWPOISON;
2132 /* Don't complain about hwpoisoned pages */
2133 page_mapcount_reset(page); /* remove PageBuddy */
2136 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
2137 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
2138 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
2141 if (unlikely(page->mem_cgroup))
2142 bad_reason = "page still charged to cgroup";
2144 bad_page(page, bad_reason, bad_flags);
2148 * This page is about to be returned from the page allocator
2150 static inline int check_new_page(struct page *page)
2152 if (likely(page_expected_state(page,
2153 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2156 check_new_page_bad(page);
2160 static inline bool free_pages_prezeroed(void)
2162 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2163 page_poisoning_enabled()) || want_init_on_free();
2166 #ifdef CONFIG_DEBUG_VM
2168 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2169 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2170 * also checked when pcp lists are refilled from the free lists.
2172 static inline bool check_pcp_refill(struct page *page)
2174 if (debug_pagealloc_enabled_static())
2175 return check_new_page(page);
2180 static inline bool check_new_pcp(struct page *page)
2182 return check_new_page(page);
2186 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2187 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2188 * enabled, they are also checked when being allocated from the pcp lists.
2190 static inline bool check_pcp_refill(struct page *page)
2192 return check_new_page(page);
2194 static inline bool check_new_pcp(struct page *page)
2196 if (debug_pagealloc_enabled_static())
2197 return check_new_page(page);
2201 #endif /* CONFIG_DEBUG_VM */
2203 static bool check_new_pages(struct page *page, unsigned int order)
2206 for (i = 0; i < (1 << order); i++) {
2207 struct page *p = page + i;
2209 if (unlikely(check_new_page(p)))
2216 inline void post_alloc_hook(struct page *page, unsigned int order,
2219 set_page_private(page, 0);
2220 set_page_refcounted(page);
2222 arch_alloc_page(page, order);
2223 if (debug_pagealloc_enabled_static())
2224 kernel_map_pages(page, 1 << order, 1);
2225 kasan_alloc_pages(page, order);
2226 kernel_poison_pages(page, 1 << order, 1);
2227 set_page_owner(page, order, gfp_flags);
2230 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2231 unsigned int alloc_flags)
2233 post_alloc_hook(page, order, gfp_flags);
2235 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2236 kernel_init_free_pages(page, 1 << order);
2238 if (order && (gfp_flags & __GFP_COMP))
2239 prep_compound_page(page, order);
2242 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2243 * allocate the page. The expectation is that the caller is taking
2244 * steps that will free more memory. The caller should avoid the page
2245 * being used for !PFMEMALLOC purposes.
2247 if (alloc_flags & ALLOC_NO_WATERMARKS)
2248 set_page_pfmemalloc(page);
2250 clear_page_pfmemalloc(page);
2254 * Go through the free lists for the given migratetype and remove
2255 * the smallest available page from the freelists
2257 static __always_inline
2258 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2261 unsigned int current_order;
2262 struct free_area *area;
2265 /* Find a page of the appropriate size in the preferred list */
2266 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2267 area = &(zone->free_area[current_order]);
2268 page = get_page_from_free_area(area, migratetype);
2271 del_page_from_free_list(page, zone, current_order);
2272 expand(zone, page, order, current_order, migratetype);
2273 set_pcppage_migratetype(page, migratetype);
2282 * This array describes the order lists are fallen back to when
2283 * the free lists for the desirable migrate type are depleted
2285 static int fallbacks[MIGRATE_TYPES][4] = {
2286 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2287 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2288 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2290 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2292 #ifdef CONFIG_MEMORY_ISOLATION
2293 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2298 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2301 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2304 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2305 unsigned int order) { return NULL; }
2309 * Move the free pages in a range to the free lists of the requested type.
2310 * Note that start_page and end_pages are not aligned on a pageblock
2311 * boundary. If alignment is required, use move_freepages_block()
2313 static int move_freepages(struct zone *zone,
2314 struct page *start_page, struct page *end_page,
2315 int migratetype, int *num_movable)
2319 int pages_moved = 0;
2321 for (page = start_page; page <= end_page;) {
2322 if (!pfn_valid_within(page_to_pfn(page))) {
2327 if (!PageBuddy(page)) {
2329 * We assume that pages that could be isolated for
2330 * migration are movable. But we don't actually try
2331 * isolating, as that would be expensive.
2334 (PageLRU(page) || __PageMovable(page)))
2341 /* Make sure we are not inadvertently changing nodes */
2342 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2343 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2345 order = page_order(page);
2346 move_to_free_list(page, zone, order, migratetype);
2348 pages_moved += 1 << order;
2354 int move_freepages_block(struct zone *zone, struct page *page,
2355 int migratetype, int *num_movable)
2357 unsigned long start_pfn, end_pfn;
2358 struct page *start_page, *end_page;
2363 start_pfn = page_to_pfn(page);
2364 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2365 start_page = pfn_to_page(start_pfn);
2366 end_page = start_page + pageblock_nr_pages - 1;
2367 end_pfn = start_pfn + pageblock_nr_pages - 1;
2369 /* Do not cross zone boundaries */
2370 if (!zone_spans_pfn(zone, start_pfn))
2372 if (!zone_spans_pfn(zone, end_pfn))
2375 return move_freepages(zone, start_page, end_page, migratetype,
2379 static void change_pageblock_range(struct page *pageblock_page,
2380 int start_order, int migratetype)
2382 int nr_pageblocks = 1 << (start_order - pageblock_order);
2384 while (nr_pageblocks--) {
2385 set_pageblock_migratetype(pageblock_page, migratetype);
2386 pageblock_page += pageblock_nr_pages;
2391 * When we are falling back to another migratetype during allocation, try to
2392 * steal extra free pages from the same pageblocks to satisfy further
2393 * allocations, instead of polluting multiple pageblocks.
2395 * If we are stealing a relatively large buddy page, it is likely there will
2396 * be more free pages in the pageblock, so try to steal them all. For
2397 * reclaimable and unmovable allocations, we steal regardless of page size,
2398 * as fragmentation caused by those allocations polluting movable pageblocks
2399 * is worse than movable allocations stealing from unmovable and reclaimable
2402 static bool can_steal_fallback(unsigned int order, int start_mt)
2405 * Leaving this order check is intended, although there is
2406 * relaxed order check in next check. The reason is that
2407 * we can actually steal whole pageblock if this condition met,
2408 * but, below check doesn't guarantee it and that is just heuristic
2409 * so could be changed anytime.
2411 if (order >= pageblock_order)
2414 if (order >= pageblock_order / 2 ||
2415 start_mt == MIGRATE_RECLAIMABLE ||
2416 start_mt == MIGRATE_UNMOVABLE ||
2417 page_group_by_mobility_disabled)
2423 static inline void boost_watermark(struct zone *zone)
2425 unsigned long max_boost;
2427 if (!watermark_boost_factor)
2430 * Don't bother in zones that are unlikely to produce results.
2431 * On small machines, including kdump capture kernels running
2432 * in a small area, boosting the watermark can cause an out of
2433 * memory situation immediately.
2435 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2438 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2439 watermark_boost_factor, 10000);
2442 * high watermark may be uninitialised if fragmentation occurs
2443 * very early in boot so do not boost. We do not fall
2444 * through and boost by pageblock_nr_pages as failing
2445 * allocations that early means that reclaim is not going
2446 * to help and it may even be impossible to reclaim the
2447 * boosted watermark resulting in a hang.
2452 max_boost = max(pageblock_nr_pages, max_boost);
2454 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2459 * This function implements actual steal behaviour. If order is large enough,
2460 * we can steal whole pageblock. If not, we first move freepages in this
2461 * pageblock to our migratetype and determine how many already-allocated pages
2462 * are there in the pageblock with a compatible migratetype. If at least half
2463 * of pages are free or compatible, we can change migratetype of the pageblock
2464 * itself, so pages freed in the future will be put on the correct free list.
2466 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2467 unsigned int alloc_flags, int start_type, bool whole_block)
2469 unsigned int current_order = page_order(page);
2470 int free_pages, movable_pages, alike_pages;
2473 old_block_type = get_pageblock_migratetype(page);
2476 * This can happen due to races and we want to prevent broken
2477 * highatomic accounting.
2479 if (is_migrate_highatomic(old_block_type))
2482 /* Take ownership for orders >= pageblock_order */
2483 if (current_order >= pageblock_order) {
2484 change_pageblock_range(page, current_order, start_type);
2489 * Boost watermarks to increase reclaim pressure to reduce the
2490 * likelihood of future fallbacks. Wake kswapd now as the node
2491 * may be balanced overall and kswapd will not wake naturally.
2493 boost_watermark(zone);
2494 if (alloc_flags & ALLOC_KSWAPD)
2495 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2497 /* We are not allowed to try stealing from the whole block */
2501 free_pages = move_freepages_block(zone, page, start_type,
2504 * Determine how many pages are compatible with our allocation.
2505 * For movable allocation, it's the number of movable pages which
2506 * we just obtained. For other types it's a bit more tricky.
2508 if (start_type == MIGRATE_MOVABLE) {
2509 alike_pages = movable_pages;
2512 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2513 * to MOVABLE pageblock, consider all non-movable pages as
2514 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2515 * vice versa, be conservative since we can't distinguish the
2516 * exact migratetype of non-movable pages.
2518 if (old_block_type == MIGRATE_MOVABLE)
2519 alike_pages = pageblock_nr_pages
2520 - (free_pages + movable_pages);
2525 /* moving whole block can fail due to zone boundary conditions */
2530 * If a sufficient number of pages in the block are either free or of
2531 * comparable migratability as our allocation, claim the whole block.
2533 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2534 page_group_by_mobility_disabled)
2535 set_pageblock_migratetype(page, start_type);
2540 move_to_free_list(page, zone, current_order, start_type);
2544 * Check whether there is a suitable fallback freepage with requested order.
2545 * If only_stealable is true, this function returns fallback_mt only if
2546 * we can steal other freepages all together. This would help to reduce
2547 * fragmentation due to mixed migratetype pages in one pageblock.
2549 int find_suitable_fallback(struct free_area *area, unsigned int order,
2550 int migratetype, bool only_stealable, bool *can_steal)
2555 if (area->nr_free == 0)
2560 fallback_mt = fallbacks[migratetype][i];
2561 if (fallback_mt == MIGRATE_TYPES)
2564 if (free_area_empty(area, fallback_mt))
2567 if (can_steal_fallback(order, migratetype))
2570 if (!only_stealable)
2581 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2582 * there are no empty page blocks that contain a page with a suitable order
2584 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2585 unsigned int alloc_order)
2588 unsigned long max_managed, flags;
2591 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2592 * Check is race-prone but harmless.
2594 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2595 if (zone->nr_reserved_highatomic >= max_managed)
2598 spin_lock_irqsave(&zone->lock, flags);
2600 /* Recheck the nr_reserved_highatomic limit under the lock */
2601 if (zone->nr_reserved_highatomic >= max_managed)
2605 mt = get_pageblock_migratetype(page);
2606 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2607 && !is_migrate_cma(mt)) {
2608 zone->nr_reserved_highatomic += pageblock_nr_pages;
2609 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2610 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2614 spin_unlock_irqrestore(&zone->lock, flags);
2618 * Used when an allocation is about to fail under memory pressure. This
2619 * potentially hurts the reliability of high-order allocations when under
2620 * intense memory pressure but failed atomic allocations should be easier
2621 * to recover from than an OOM.
2623 * If @force is true, try to unreserve a pageblock even though highatomic
2624 * pageblock is exhausted.
2626 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2629 struct zonelist *zonelist = ac->zonelist;
2630 unsigned long flags;
2637 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2640 * Preserve at least one pageblock unless memory pressure
2643 if (!force && zone->nr_reserved_highatomic <=
2647 spin_lock_irqsave(&zone->lock, flags);
2648 for (order = 0; order < MAX_ORDER; order++) {
2649 struct free_area *area = &(zone->free_area[order]);
2651 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2656 * In page freeing path, migratetype change is racy so
2657 * we can counter several free pages in a pageblock
2658 * in this loop althoug we changed the pageblock type
2659 * from highatomic to ac->migratetype. So we should
2660 * adjust the count once.
2662 if (is_migrate_highatomic_page(page)) {
2664 * It should never happen but changes to
2665 * locking could inadvertently allow a per-cpu
2666 * drain to add pages to MIGRATE_HIGHATOMIC
2667 * while unreserving so be safe and watch for
2670 zone->nr_reserved_highatomic -= min(
2672 zone->nr_reserved_highatomic);
2676 * Convert to ac->migratetype and avoid the normal
2677 * pageblock stealing heuristics. Minimally, the caller
2678 * is doing the work and needs the pages. More
2679 * importantly, if the block was always converted to
2680 * MIGRATE_UNMOVABLE or another type then the number
2681 * of pageblocks that cannot be completely freed
2684 set_pageblock_migratetype(page, ac->migratetype);
2685 ret = move_freepages_block(zone, page, ac->migratetype,
2688 spin_unlock_irqrestore(&zone->lock, flags);
2692 spin_unlock_irqrestore(&zone->lock, flags);
2699 * Try finding a free buddy page on the fallback list and put it on the free
2700 * list of requested migratetype, possibly along with other pages from the same
2701 * block, depending on fragmentation avoidance heuristics. Returns true if
2702 * fallback was found so that __rmqueue_smallest() can grab it.
2704 * The use of signed ints for order and current_order is a deliberate
2705 * deviation from the rest of this file, to make the for loop
2706 * condition simpler.
2708 static __always_inline bool
2709 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2710 unsigned int alloc_flags)
2712 struct free_area *area;
2714 int min_order = order;
2720 * Do not steal pages from freelists belonging to other pageblocks
2721 * i.e. orders < pageblock_order. If there are no local zones free,
2722 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2724 if (alloc_flags & ALLOC_NOFRAGMENT)
2725 min_order = pageblock_order;
2728 * Find the largest available free page in the other list. This roughly
2729 * approximates finding the pageblock with the most free pages, which
2730 * would be too costly to do exactly.
2732 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2734 area = &(zone->free_area[current_order]);
2735 fallback_mt = find_suitable_fallback(area, current_order,
2736 start_migratetype, false, &can_steal);
2737 if (fallback_mt == -1)
2741 * We cannot steal all free pages from the pageblock and the
2742 * requested migratetype is movable. In that case it's better to
2743 * steal and split the smallest available page instead of the
2744 * largest available page, because even if the next movable
2745 * allocation falls back into a different pageblock than this
2746 * one, it won't cause permanent fragmentation.
2748 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2749 && current_order > order)
2758 for (current_order = order; current_order < MAX_ORDER;
2760 area = &(zone->free_area[current_order]);
2761 fallback_mt = find_suitable_fallback(area, current_order,
2762 start_migratetype, false, &can_steal);
2763 if (fallback_mt != -1)
2768 * This should not happen - we already found a suitable fallback
2769 * when looking for the largest page.
2771 VM_BUG_ON(current_order == MAX_ORDER);
2774 page = get_page_from_free_area(area, fallback_mt);
2776 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2779 trace_mm_page_alloc_extfrag(page, order, current_order,
2780 start_migratetype, fallback_mt);
2787 * Do the hard work of removing an element from the buddy allocator.
2788 * Call me with the zone->lock already held.
2790 static __always_inline struct page *
2791 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2792 unsigned int alloc_flags)
2797 page = __rmqueue_smallest(zone, order, migratetype);
2798 if (unlikely(!page)) {
2799 if (migratetype == MIGRATE_MOVABLE)
2800 page = __rmqueue_cma_fallback(zone, order);
2802 if (!page && __rmqueue_fallback(zone, order, migratetype,
2807 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2812 * Obtain a specified number of elements from the buddy allocator, all under
2813 * a single hold of the lock, for efficiency. Add them to the supplied list.
2814 * Returns the number of new pages which were placed at *list.
2816 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2817 unsigned long count, struct list_head *list,
2818 int migratetype, unsigned int alloc_flags)
2822 spin_lock(&zone->lock);
2823 for (i = 0; i < count; ++i) {
2824 struct page *page = __rmqueue(zone, order, migratetype,
2826 if (unlikely(page == NULL))
2829 if (unlikely(check_pcp_refill(page)))
2833 * Split buddy pages returned by expand() are received here in
2834 * physical page order. The page is added to the tail of
2835 * caller's list. From the callers perspective, the linked list
2836 * is ordered by page number under some conditions. This is
2837 * useful for IO devices that can forward direction from the
2838 * head, thus also in the physical page order. This is useful
2839 * for IO devices that can merge IO requests if the physical
2840 * pages are ordered properly.
2842 list_add_tail(&page->lru, list);
2844 if (is_migrate_cma(get_pcppage_migratetype(page)))
2845 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2850 * i pages were removed from the buddy list even if some leak due
2851 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2852 * on i. Do not confuse with 'alloced' which is the number of
2853 * pages added to the pcp list.
2855 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2856 spin_unlock(&zone->lock);
2862 * Called from the vmstat counter updater to drain pagesets of this
2863 * currently executing processor on remote nodes after they have
2866 * Note that this function must be called with the thread pinned to
2867 * a single processor.
2869 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2871 unsigned long flags;
2872 int to_drain, batch;
2874 local_irq_save(flags);
2875 batch = READ_ONCE(pcp->batch);
2876 to_drain = min(pcp->count, batch);
2878 free_pcppages_bulk(zone, to_drain, pcp);
2879 local_irq_restore(flags);
2884 * Drain pcplists of the indicated processor and zone.
2886 * The processor must either be the current processor and the
2887 * thread pinned to the current processor or a processor that
2890 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2892 unsigned long flags;
2893 struct per_cpu_pageset *pset;
2894 struct per_cpu_pages *pcp;
2896 local_irq_save(flags);
2897 pset = per_cpu_ptr(zone->pageset, cpu);
2901 free_pcppages_bulk(zone, pcp->count, pcp);
2902 local_irq_restore(flags);
2906 * Drain pcplists of all zones on the indicated processor.
2908 * The processor must either be the current processor and the
2909 * thread pinned to the current processor or a processor that
2912 static void drain_pages(unsigned int cpu)
2916 for_each_populated_zone(zone) {
2917 drain_pages_zone(cpu, zone);
2922 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2924 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2925 * the single zone's pages.
2927 void drain_local_pages(struct zone *zone)
2929 int cpu = smp_processor_id();
2932 drain_pages_zone(cpu, zone);
2937 static void drain_local_pages_wq(struct work_struct *work)
2939 struct pcpu_drain *drain;
2941 drain = container_of(work, struct pcpu_drain, work);
2944 * drain_all_pages doesn't use proper cpu hotplug protection so
2945 * we can race with cpu offline when the WQ can move this from
2946 * a cpu pinned worker to an unbound one. We can operate on a different
2947 * cpu which is allright but we also have to make sure to not move to
2951 drain_local_pages(drain->zone);
2956 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2958 * When zone parameter is non-NULL, spill just the single zone's pages.
2960 * Note that this can be extremely slow as the draining happens in a workqueue.
2962 void drain_all_pages(struct zone *zone)
2967 * Allocate in the BSS so we wont require allocation in
2968 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2970 static cpumask_t cpus_with_pcps;
2973 * Make sure nobody triggers this path before mm_percpu_wq is fully
2976 if (WARN_ON_ONCE(!mm_percpu_wq))
2980 * Do not drain if one is already in progress unless it's specific to
2981 * a zone. Such callers are primarily CMA and memory hotplug and need
2982 * the drain to be complete when the call returns.
2984 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2987 mutex_lock(&pcpu_drain_mutex);
2991 * We don't care about racing with CPU hotplug event
2992 * as offline notification will cause the notified
2993 * cpu to drain that CPU pcps and on_each_cpu_mask
2994 * disables preemption as part of its processing
2996 for_each_online_cpu(cpu) {
2997 struct per_cpu_pageset *pcp;
2999 bool has_pcps = false;
3002 pcp = per_cpu_ptr(zone->pageset, cpu);
3006 for_each_populated_zone(z) {
3007 pcp = per_cpu_ptr(z->pageset, cpu);
3008 if (pcp->pcp.count) {
3016 cpumask_set_cpu(cpu, &cpus_with_pcps);
3018 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3021 for_each_cpu(cpu, &cpus_with_pcps) {
3022 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3025 INIT_WORK(&drain->work, drain_local_pages_wq);
3026 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3028 for_each_cpu(cpu, &cpus_with_pcps)
3029 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3031 mutex_unlock(&pcpu_drain_mutex);
3034 #ifdef CONFIG_HIBERNATION
3037 * Touch the watchdog for every WD_PAGE_COUNT pages.
3039 #define WD_PAGE_COUNT (128*1024)
3041 void mark_free_pages(struct zone *zone)
3043 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3044 unsigned long flags;
3045 unsigned int order, t;
3048 if (zone_is_empty(zone))
3051 spin_lock_irqsave(&zone->lock, flags);
3053 max_zone_pfn = zone_end_pfn(zone);
3054 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3055 if (pfn_valid(pfn)) {
3056 page = pfn_to_page(pfn);
3058 if (!--page_count) {
3059 touch_nmi_watchdog();
3060 page_count = WD_PAGE_COUNT;
3063 if (page_zone(page) != zone)
3066 if (!swsusp_page_is_forbidden(page))
3067 swsusp_unset_page_free(page);
3070 for_each_migratetype_order(order, t) {
3071 list_for_each_entry(page,
3072 &zone->free_area[order].free_list[t], lru) {
3075 pfn = page_to_pfn(page);
3076 for (i = 0; i < (1UL << order); i++) {
3077 if (!--page_count) {
3078 touch_nmi_watchdog();
3079 page_count = WD_PAGE_COUNT;
3081 swsusp_set_page_free(pfn_to_page(pfn + i));
3085 spin_unlock_irqrestore(&zone->lock, flags);
3087 #endif /* CONFIG_PM */
3089 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3093 if (!free_pcp_prepare(page))
3096 migratetype = get_pfnblock_migratetype(page, pfn);
3097 set_pcppage_migratetype(page, migratetype);
3101 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3103 struct zone *zone = page_zone(page);
3104 struct per_cpu_pages *pcp;
3107 migratetype = get_pcppage_migratetype(page);
3108 __count_vm_event(PGFREE);
3111 * We only track unmovable, reclaimable and movable on pcp lists.
3112 * Free ISOLATE pages back to the allocator because they are being
3113 * offlined but treat HIGHATOMIC as movable pages so we can get those
3114 * areas back if necessary. Otherwise, we may have to free
3115 * excessively into the page allocator
3117 if (migratetype >= MIGRATE_PCPTYPES) {
3118 if (unlikely(is_migrate_isolate(migratetype))) {
3119 free_one_page(zone, page, pfn, 0, migratetype);
3122 migratetype = MIGRATE_MOVABLE;
3125 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3126 list_add(&page->lru, &pcp->lists[migratetype]);
3128 if (pcp->count >= pcp->high) {
3129 unsigned long batch = READ_ONCE(pcp->batch);
3130 free_pcppages_bulk(zone, batch, pcp);
3135 * Free a 0-order page
3137 void free_unref_page(struct page *page)
3139 unsigned long flags;
3140 unsigned long pfn = page_to_pfn(page);
3142 if (!free_unref_page_prepare(page, pfn))
3145 local_irq_save(flags);
3146 free_unref_page_commit(page, pfn);
3147 local_irq_restore(flags);
3151 * Free a list of 0-order pages
3153 void free_unref_page_list(struct list_head *list)
3155 struct page *page, *next;
3156 unsigned long flags, pfn;
3157 int batch_count = 0;
3159 /* Prepare pages for freeing */
3160 list_for_each_entry_safe(page, next, list, lru) {
3161 pfn = page_to_pfn(page);
3162 if (!free_unref_page_prepare(page, pfn))
3163 list_del(&page->lru);
3164 set_page_private(page, pfn);
3167 local_irq_save(flags);
3168 list_for_each_entry_safe(page, next, list, lru) {
3169 unsigned long pfn = page_private(page);
3171 set_page_private(page, 0);
3172 trace_mm_page_free_batched(page);
3173 free_unref_page_commit(page, pfn);
3176 * Guard against excessive IRQ disabled times when we get
3177 * a large list of pages to free.
3179 if (++batch_count == SWAP_CLUSTER_MAX) {
3180 local_irq_restore(flags);
3182 local_irq_save(flags);
3185 local_irq_restore(flags);
3189 * split_page takes a non-compound higher-order page, and splits it into
3190 * n (1<<order) sub-pages: page[0..n]
3191 * Each sub-page must be freed individually.
3193 * Note: this is probably too low level an operation for use in drivers.
3194 * Please consult with lkml before using this in your driver.
3196 void split_page(struct page *page, unsigned int order)
3200 VM_BUG_ON_PAGE(PageCompound(page), page);
3201 VM_BUG_ON_PAGE(!page_count(page), page);
3203 for (i = 1; i < (1 << order); i++)
3204 set_page_refcounted(page + i);
3205 split_page_owner(page, order);
3207 EXPORT_SYMBOL_GPL(split_page);
3209 int __isolate_free_page(struct page *page, unsigned int order)
3211 unsigned long watermark;
3215 BUG_ON(!PageBuddy(page));
3217 zone = page_zone(page);
3218 mt = get_pageblock_migratetype(page);
3220 if (!is_migrate_isolate(mt)) {
3222 * Obey watermarks as if the page was being allocated. We can
3223 * emulate a high-order watermark check with a raised order-0
3224 * watermark, because we already know our high-order page
3227 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3228 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3231 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3234 /* Remove page from free list */
3236 del_page_from_free_list(page, zone, order);
3239 * Set the pageblock if the isolated page is at least half of a
3242 if (order >= pageblock_order - 1) {
3243 struct page *endpage = page + (1 << order) - 1;
3244 for (; page < endpage; page += pageblock_nr_pages) {
3245 int mt = get_pageblock_migratetype(page);
3246 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3247 && !is_migrate_highatomic(mt))
3248 set_pageblock_migratetype(page,
3254 return 1UL << order;
3258 * __putback_isolated_page - Return a now-isolated page back where we got it
3259 * @page: Page that was isolated
3260 * @order: Order of the isolated page
3261 * @mt: The page's pageblock's migratetype
3263 * This function is meant to return a page pulled from the free lists via
3264 * __isolate_free_page back to the free lists they were pulled from.
3266 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3268 struct zone *zone = page_zone(page);
3270 /* zone lock should be held when this function is called */
3271 lockdep_assert_held(&zone->lock);
3273 /* Return isolated page to tail of freelist. */
3274 __free_one_page(page, page_to_pfn(page), zone, order, mt, false);
3278 * Update NUMA hit/miss statistics
3280 * Must be called with interrupts disabled.
3282 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3285 enum numa_stat_item local_stat = NUMA_LOCAL;
3287 /* skip numa counters update if numa stats is disabled */
3288 if (!static_branch_likely(&vm_numa_stat_key))
3291 if (zone_to_nid(z) != numa_node_id())
3292 local_stat = NUMA_OTHER;
3294 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3295 __inc_numa_state(z, NUMA_HIT);
3297 __inc_numa_state(z, NUMA_MISS);
3298 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3300 __inc_numa_state(z, local_stat);
3304 /* Remove page from the per-cpu list, caller must protect the list */
3305 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3306 unsigned int alloc_flags,
3307 struct per_cpu_pages *pcp,
3308 struct list_head *list)
3313 if (list_empty(list)) {
3314 pcp->count += rmqueue_bulk(zone, 0,
3316 migratetype, alloc_flags);
3317 if (unlikely(list_empty(list)))
3321 page = list_first_entry(list, struct page, lru);
3322 list_del(&page->lru);
3324 } while (check_new_pcp(page));
3329 /* Lock and remove page from the per-cpu list */
3330 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3331 struct zone *zone, gfp_t gfp_flags,
3332 int migratetype, unsigned int alloc_flags)
3334 struct per_cpu_pages *pcp;
3335 struct list_head *list;
3337 unsigned long flags;
3339 local_irq_save(flags);
3340 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3341 list = &pcp->lists[migratetype];
3342 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3344 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3345 zone_statistics(preferred_zone, zone);
3347 local_irq_restore(flags);
3352 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3355 struct page *rmqueue(struct zone *preferred_zone,
3356 struct zone *zone, unsigned int order,
3357 gfp_t gfp_flags, unsigned int alloc_flags,
3360 unsigned long flags;
3363 if (likely(order == 0)) {
3364 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3365 migratetype, alloc_flags);
3370 * We most definitely don't want callers attempting to
3371 * allocate greater than order-1 page units with __GFP_NOFAIL.
3373 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3374 spin_lock_irqsave(&zone->lock, flags);
3378 if (alloc_flags & ALLOC_HARDER) {
3379 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3381 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3384 page = __rmqueue(zone, order, migratetype, alloc_flags);
3385 } while (page && check_new_pages(page, order));
3386 spin_unlock(&zone->lock);
3389 __mod_zone_freepage_state(zone, -(1 << order),
3390 get_pcppage_migratetype(page));
3392 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3393 zone_statistics(preferred_zone, zone);
3394 local_irq_restore(flags);
3397 /* Separate test+clear to avoid unnecessary atomics */
3398 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3399 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3400 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3403 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3407 local_irq_restore(flags);
3411 #ifdef CONFIG_FAIL_PAGE_ALLOC
3414 struct fault_attr attr;
3416 bool ignore_gfp_highmem;
3417 bool ignore_gfp_reclaim;
3419 } fail_page_alloc = {
3420 .attr = FAULT_ATTR_INITIALIZER,
3421 .ignore_gfp_reclaim = true,
3422 .ignore_gfp_highmem = true,
3426 static int __init setup_fail_page_alloc(char *str)
3428 return setup_fault_attr(&fail_page_alloc.attr, str);
3430 __setup("fail_page_alloc=", setup_fail_page_alloc);
3432 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3434 if (order < fail_page_alloc.min_order)
3436 if (gfp_mask & __GFP_NOFAIL)
3438 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3440 if (fail_page_alloc.ignore_gfp_reclaim &&
3441 (gfp_mask & __GFP_DIRECT_RECLAIM))
3444 return should_fail(&fail_page_alloc.attr, 1 << order);
3447 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3449 static int __init fail_page_alloc_debugfs(void)
3451 umode_t mode = S_IFREG | 0600;
3454 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3455 &fail_page_alloc.attr);
3457 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3458 &fail_page_alloc.ignore_gfp_reclaim);
3459 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3460 &fail_page_alloc.ignore_gfp_highmem);
3461 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3466 late_initcall(fail_page_alloc_debugfs);
3468 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3470 #else /* CONFIG_FAIL_PAGE_ALLOC */
3472 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3477 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3479 static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3481 return __should_fail_alloc_page(gfp_mask, order);
3483 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3486 * Return true if free base pages are above 'mark'. For high-order checks it
3487 * will return true of the order-0 watermark is reached and there is at least
3488 * one free page of a suitable size. Checking now avoids taking the zone lock
3489 * to check in the allocation paths if no pages are free.
3491 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3492 int classzone_idx, unsigned int alloc_flags,
3497 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3499 /* free_pages may go negative - that's OK */
3500 free_pages -= (1 << order) - 1;
3502 if (alloc_flags & ALLOC_HIGH)
3506 * If the caller does not have rights to ALLOC_HARDER then subtract
3507 * the high-atomic reserves. This will over-estimate the size of the
3508 * atomic reserve but it avoids a search.
3510 if (likely(!alloc_harder)) {
3511 free_pages -= z->nr_reserved_highatomic;
3514 * OOM victims can try even harder than normal ALLOC_HARDER
3515 * users on the grounds that it's definitely going to be in
3516 * the exit path shortly and free memory. Any allocation it
3517 * makes during the free path will be small and short-lived.
3519 if (alloc_flags & ALLOC_OOM)
3527 /* If allocation can't use CMA areas don't use free CMA pages */
3528 if (!(alloc_flags & ALLOC_CMA))
3529 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3533 * Check watermarks for an order-0 allocation request. If these
3534 * are not met, then a high-order request also cannot go ahead
3535 * even if a suitable page happened to be free.
3537 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3540 /* If this is an order-0 request then the watermark is fine */
3544 /* For a high-order request, check at least one suitable page is free */
3545 for (o = order; o < MAX_ORDER; o++) {
3546 struct free_area *area = &z->free_area[o];
3552 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3553 if (!free_area_empty(area, mt))
3558 if ((alloc_flags & ALLOC_CMA) &&
3559 !free_area_empty(area, MIGRATE_CMA)) {
3563 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3569 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3570 int classzone_idx, unsigned int alloc_flags)
3572 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3573 zone_page_state(z, NR_FREE_PAGES));
3576 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3577 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3579 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3583 /* If allocation can't use CMA areas don't use free CMA pages */
3584 if (!(alloc_flags & ALLOC_CMA))
3585 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3589 * Fast check for order-0 only. If this fails then the reserves
3590 * need to be calculated. There is a corner case where the check
3591 * passes but only the high-order atomic reserve are free. If
3592 * the caller is !atomic then it'll uselessly search the free
3593 * list. That corner case is then slower but it is harmless.
3595 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3598 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3602 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3603 unsigned long mark, int classzone_idx)
3605 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3607 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3608 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3610 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3615 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3617 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3618 node_reclaim_distance;
3620 #else /* CONFIG_NUMA */
3621 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3625 #endif /* CONFIG_NUMA */
3628 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3629 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3630 * premature use of a lower zone may cause lowmem pressure problems that
3631 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3632 * probably too small. It only makes sense to spread allocations to avoid
3633 * fragmentation between the Normal and DMA32 zones.
3635 static inline unsigned int
3636 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3638 unsigned int alloc_flags;
3641 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3644 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3646 #ifdef CONFIG_ZONE_DMA32
3650 if (zone_idx(zone) != ZONE_NORMAL)
3654 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3655 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3656 * on UMA that if Normal is populated then so is DMA32.
3658 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3659 if (nr_online_nodes > 1 && !populated_zone(--zone))
3662 alloc_flags |= ALLOC_NOFRAGMENT;
3663 #endif /* CONFIG_ZONE_DMA32 */
3668 * get_page_from_freelist goes through the zonelist trying to allocate
3671 static struct page *
3672 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3673 const struct alloc_context *ac)
3677 struct pglist_data *last_pgdat_dirty_limit = NULL;
3682 * Scan zonelist, looking for a zone with enough free.
3683 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3685 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3686 z = ac->preferred_zoneref;
3687 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3692 if (cpusets_enabled() &&
3693 (alloc_flags & ALLOC_CPUSET) &&
3694 !__cpuset_zone_allowed(zone, gfp_mask))
3697 * When allocating a page cache page for writing, we
3698 * want to get it from a node that is within its dirty
3699 * limit, such that no single node holds more than its
3700 * proportional share of globally allowed dirty pages.
3701 * The dirty limits take into account the node's
3702 * lowmem reserves and high watermark so that kswapd
3703 * should be able to balance it without having to
3704 * write pages from its LRU list.
3706 * XXX: For now, allow allocations to potentially
3707 * exceed the per-node dirty limit in the slowpath
3708 * (spread_dirty_pages unset) before going into reclaim,
3709 * which is important when on a NUMA setup the allowed
3710 * nodes are together not big enough to reach the
3711 * global limit. The proper fix for these situations
3712 * will require awareness of nodes in the
3713 * dirty-throttling and the flusher threads.
3715 if (ac->spread_dirty_pages) {
3716 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3719 if (!node_dirty_ok(zone->zone_pgdat)) {
3720 last_pgdat_dirty_limit = zone->zone_pgdat;
3725 if (no_fallback && nr_online_nodes > 1 &&
3726 zone != ac->preferred_zoneref->zone) {
3730 * If moving to a remote node, retry but allow
3731 * fragmenting fallbacks. Locality is more important
3732 * than fragmentation avoidance.
3734 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3735 if (zone_to_nid(zone) != local_nid) {
3736 alloc_flags &= ~ALLOC_NOFRAGMENT;
3741 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3742 if (!zone_watermark_fast(zone, order, mark,
3743 ac_classzone_idx(ac), alloc_flags)) {
3746 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3748 * Watermark failed for this zone, but see if we can
3749 * grow this zone if it contains deferred pages.
3751 if (static_branch_unlikely(&deferred_pages)) {
3752 if (_deferred_grow_zone(zone, order))
3756 /* Checked here to keep the fast path fast */
3757 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3758 if (alloc_flags & ALLOC_NO_WATERMARKS)
3761 if (node_reclaim_mode == 0 ||
3762 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3765 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3767 case NODE_RECLAIM_NOSCAN:
3770 case NODE_RECLAIM_FULL:
3771 /* scanned but unreclaimable */
3774 /* did we reclaim enough */
3775 if (zone_watermark_ok(zone, order, mark,
3776 ac_classzone_idx(ac), alloc_flags))
3784 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3785 gfp_mask, alloc_flags, ac->migratetype);
3787 prep_new_page(page, order, gfp_mask, alloc_flags);
3790 * If this is a high-order atomic allocation then check
3791 * if the pageblock should be reserved for the future
3793 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3794 reserve_highatomic_pageblock(page, zone, order);
3798 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3799 /* Try again if zone has deferred pages */
3800 if (static_branch_unlikely(&deferred_pages)) {
3801 if (_deferred_grow_zone(zone, order))
3809 * It's possible on a UMA machine to get through all zones that are
3810 * fragmented. If avoiding fragmentation, reset and try again.
3813 alloc_flags &= ~ALLOC_NOFRAGMENT;
3820 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3822 unsigned int filter = SHOW_MEM_FILTER_NODES;
3825 * This documents exceptions given to allocations in certain
3826 * contexts that are allowed to allocate outside current's set
3829 if (!(gfp_mask & __GFP_NOMEMALLOC))
3830 if (tsk_is_oom_victim(current) ||
3831 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3832 filter &= ~SHOW_MEM_FILTER_NODES;
3833 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3834 filter &= ~SHOW_MEM_FILTER_NODES;
3836 show_mem(filter, nodemask);
3839 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3841 struct va_format vaf;
3843 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3845 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3848 va_start(args, fmt);
3851 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3852 current->comm, &vaf, gfp_mask, &gfp_mask,
3853 nodemask_pr_args(nodemask));
3856 cpuset_print_current_mems_allowed();
3859 warn_alloc_show_mem(gfp_mask, nodemask);
3862 static inline struct page *
3863 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3864 unsigned int alloc_flags,
3865 const struct alloc_context *ac)
3869 page = get_page_from_freelist(gfp_mask, order,
3870 alloc_flags|ALLOC_CPUSET, ac);
3872 * fallback to ignore cpuset restriction if our nodes
3876 page = get_page_from_freelist(gfp_mask, order,
3882 static inline struct page *
3883 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3884 const struct alloc_context *ac, unsigned long *did_some_progress)
3886 struct oom_control oc = {
3887 .zonelist = ac->zonelist,
3888 .nodemask = ac->nodemask,
3890 .gfp_mask = gfp_mask,
3895 *did_some_progress = 0;
3898 * Acquire the oom lock. If that fails, somebody else is
3899 * making progress for us.
3901 if (!mutex_trylock(&oom_lock)) {
3902 *did_some_progress = 1;
3903 schedule_timeout_uninterruptible(1);
3908 * Go through the zonelist yet one more time, keep very high watermark
3909 * here, this is only to catch a parallel oom killing, we must fail if
3910 * we're still under heavy pressure. But make sure that this reclaim
3911 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3912 * allocation which will never fail due to oom_lock already held.
3914 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3915 ~__GFP_DIRECT_RECLAIM, order,
3916 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3920 /* Coredumps can quickly deplete all memory reserves */
3921 if (current->flags & PF_DUMPCORE)
3923 /* The OOM killer will not help higher order allocs */
3924 if (order > PAGE_ALLOC_COSTLY_ORDER)
3927 * We have already exhausted all our reclaim opportunities without any
3928 * success so it is time to admit defeat. We will skip the OOM killer
3929 * because it is very likely that the caller has a more reasonable
3930 * fallback than shooting a random task.
3932 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3934 /* The OOM killer does not needlessly kill tasks for lowmem */
3935 if (ac->high_zoneidx < ZONE_NORMAL)
3937 if (pm_suspended_storage())
3940 * XXX: GFP_NOFS allocations should rather fail than rely on
3941 * other request to make a forward progress.
3942 * We are in an unfortunate situation where out_of_memory cannot
3943 * do much for this context but let's try it to at least get
3944 * access to memory reserved if the current task is killed (see
3945 * out_of_memory). Once filesystems are ready to handle allocation
3946 * failures more gracefully we should just bail out here.
3949 /* The OOM killer may not free memory on a specific node */
3950 if (gfp_mask & __GFP_THISNODE)
3953 /* Exhausted what can be done so it's blame time */
3954 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3955 *did_some_progress = 1;
3958 * Help non-failing allocations by giving them access to memory
3961 if (gfp_mask & __GFP_NOFAIL)
3962 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3963 ALLOC_NO_WATERMARKS, ac);
3966 mutex_unlock(&oom_lock);
3971 * Maximum number of compaction retries wit a progress before OOM
3972 * killer is consider as the only way to move forward.
3974 #define MAX_COMPACT_RETRIES 16
3976 #ifdef CONFIG_COMPACTION
3977 /* Try memory compaction for high-order allocations before reclaim */
3978 static struct page *
3979 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3980 unsigned int alloc_flags, const struct alloc_context *ac,
3981 enum compact_priority prio, enum compact_result *compact_result)
3983 struct page *page = NULL;
3984 unsigned long pflags;
3985 unsigned int noreclaim_flag;
3990 psi_memstall_enter(&pflags);
3991 noreclaim_flag = memalloc_noreclaim_save();
3993 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3996 memalloc_noreclaim_restore(noreclaim_flag);
3997 psi_memstall_leave(&pflags);
4000 * At least in one zone compaction wasn't deferred or skipped, so let's
4001 * count a compaction stall
4003 count_vm_event(COMPACTSTALL);
4005 /* Prep a captured page if available */
4007 prep_new_page(page, order, gfp_mask, alloc_flags);
4009 /* Try get a page from the freelist if available */
4011 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4014 struct zone *zone = page_zone(page);
4016 zone->compact_blockskip_flush = false;
4017 compaction_defer_reset(zone, order, true);
4018 count_vm_event(COMPACTSUCCESS);
4023 * It's bad if compaction run occurs and fails. The most likely reason
4024 * is that pages exist, but not enough to satisfy watermarks.
4026 count_vm_event(COMPACTFAIL);
4034 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4035 enum compact_result compact_result,
4036 enum compact_priority *compact_priority,
4037 int *compaction_retries)
4039 int max_retries = MAX_COMPACT_RETRIES;
4042 int retries = *compaction_retries;
4043 enum compact_priority priority = *compact_priority;
4048 if (compaction_made_progress(compact_result))
4049 (*compaction_retries)++;
4052 * compaction considers all the zone as desperately out of memory
4053 * so it doesn't really make much sense to retry except when the
4054 * failure could be caused by insufficient priority
4056 if (compaction_failed(compact_result))
4057 goto check_priority;
4060 * compaction was skipped because there are not enough order-0 pages
4061 * to work with, so we retry only if it looks like reclaim can help.
4063 if (compaction_needs_reclaim(compact_result)) {
4064 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4069 * make sure the compaction wasn't deferred or didn't bail out early
4070 * due to locks contention before we declare that we should give up.
4071 * But the next retry should use a higher priority if allowed, so
4072 * we don't just keep bailing out endlessly.
4074 if (compaction_withdrawn(compact_result)) {
4075 goto check_priority;
4079 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4080 * costly ones because they are de facto nofail and invoke OOM
4081 * killer to move on while costly can fail and users are ready
4082 * to cope with that. 1/4 retries is rather arbitrary but we
4083 * would need much more detailed feedback from compaction to
4084 * make a better decision.
4086 if (order > PAGE_ALLOC_COSTLY_ORDER)
4088 if (*compaction_retries <= max_retries) {
4094 * Make sure there are attempts at the highest priority if we exhausted
4095 * all retries or failed at the lower priorities.
4098 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4099 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4101 if (*compact_priority > min_priority) {
4102 (*compact_priority)--;
4103 *compaction_retries = 0;
4107 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4111 static inline struct page *
4112 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4113 unsigned int alloc_flags, const struct alloc_context *ac,
4114 enum compact_priority prio, enum compact_result *compact_result)
4116 *compact_result = COMPACT_SKIPPED;
4121 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4122 enum compact_result compact_result,
4123 enum compact_priority *compact_priority,
4124 int *compaction_retries)
4129 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4133 * There are setups with compaction disabled which would prefer to loop
4134 * inside the allocator rather than hit the oom killer prematurely.
4135 * Let's give them a good hope and keep retrying while the order-0
4136 * watermarks are OK.
4138 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4140 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4141 ac_classzone_idx(ac), alloc_flags))
4146 #endif /* CONFIG_COMPACTION */
4148 #ifdef CONFIG_LOCKDEP
4149 static struct lockdep_map __fs_reclaim_map =
4150 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4152 static bool __need_fs_reclaim(gfp_t gfp_mask)
4154 gfp_mask = current_gfp_context(gfp_mask);
4156 /* no reclaim without waiting on it */
4157 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4160 /* this guy won't enter reclaim */
4161 if (current->flags & PF_MEMALLOC)
4164 /* We're only interested __GFP_FS allocations for now */
4165 if (!(gfp_mask & __GFP_FS))
4168 if (gfp_mask & __GFP_NOLOCKDEP)
4174 void __fs_reclaim_acquire(void)
4176 lock_map_acquire(&__fs_reclaim_map);
4179 void __fs_reclaim_release(void)
4181 lock_map_release(&__fs_reclaim_map);
4184 void fs_reclaim_acquire(gfp_t gfp_mask)
4186 if (__need_fs_reclaim(gfp_mask))
4187 __fs_reclaim_acquire();
4189 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4191 void fs_reclaim_release(gfp_t gfp_mask)
4193 if (__need_fs_reclaim(gfp_mask))
4194 __fs_reclaim_release();
4196 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4199 /* Perform direct synchronous page reclaim */
4201 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4202 const struct alloc_context *ac)
4205 unsigned int noreclaim_flag;
4206 unsigned long pflags;
4210 /* We now go into synchronous reclaim */
4211 cpuset_memory_pressure_bump();
4212 psi_memstall_enter(&pflags);
4213 fs_reclaim_acquire(gfp_mask);
4214 noreclaim_flag = memalloc_noreclaim_save();
4216 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4219 memalloc_noreclaim_restore(noreclaim_flag);
4220 fs_reclaim_release(gfp_mask);
4221 psi_memstall_leave(&pflags);
4228 /* The really slow allocator path where we enter direct reclaim */
4229 static inline struct page *
4230 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4231 unsigned int alloc_flags, const struct alloc_context *ac,
4232 unsigned long *did_some_progress)
4234 struct page *page = NULL;
4235 bool drained = false;
4237 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4238 if (unlikely(!(*did_some_progress)))
4242 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4245 * If an allocation failed after direct reclaim, it could be because
4246 * pages are pinned on the per-cpu lists or in high alloc reserves.
4247 * Shrink them them and try again
4249 if (!page && !drained) {
4250 unreserve_highatomic_pageblock(ac, false);
4251 drain_all_pages(NULL);
4259 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4260 const struct alloc_context *ac)
4264 pg_data_t *last_pgdat = NULL;
4265 enum zone_type high_zoneidx = ac->high_zoneidx;
4267 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
4269 if (last_pgdat != zone->zone_pgdat)
4270 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
4271 last_pgdat = zone->zone_pgdat;
4275 static inline unsigned int
4276 gfp_to_alloc_flags(gfp_t gfp_mask)
4278 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4281 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4282 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4283 * to save two branches.
4285 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4286 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4289 * The caller may dip into page reserves a bit more if the caller
4290 * cannot run direct reclaim, or if the caller has realtime scheduling
4291 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4292 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4294 alloc_flags |= (__force int)
4295 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4297 if (gfp_mask & __GFP_ATOMIC) {
4299 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4300 * if it can't schedule.
4302 if (!(gfp_mask & __GFP_NOMEMALLOC))
4303 alloc_flags |= ALLOC_HARDER;
4305 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4306 * comment for __cpuset_node_allowed().
4308 alloc_flags &= ~ALLOC_CPUSET;
4309 } else if (unlikely(rt_task(current)) && !in_interrupt())
4310 alloc_flags |= ALLOC_HARDER;
4313 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4314 alloc_flags |= ALLOC_CMA;
4319 static bool oom_reserves_allowed(struct task_struct *tsk)
4321 if (!tsk_is_oom_victim(tsk))
4325 * !MMU doesn't have oom reaper so give access to memory reserves
4326 * only to the thread with TIF_MEMDIE set
4328 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4335 * Distinguish requests which really need access to full memory
4336 * reserves from oom victims which can live with a portion of it
4338 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4340 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4342 if (gfp_mask & __GFP_MEMALLOC)
4343 return ALLOC_NO_WATERMARKS;
4344 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4345 return ALLOC_NO_WATERMARKS;
4346 if (!in_interrupt()) {
4347 if (current->flags & PF_MEMALLOC)
4348 return ALLOC_NO_WATERMARKS;
4349 else if (oom_reserves_allowed(current))
4356 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4358 return !!__gfp_pfmemalloc_flags(gfp_mask);
4362 * Checks whether it makes sense to retry the reclaim to make a forward progress
4363 * for the given allocation request.
4365 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4366 * without success, or when we couldn't even meet the watermark if we
4367 * reclaimed all remaining pages on the LRU lists.
4369 * Returns true if a retry is viable or false to enter the oom path.
4372 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4373 struct alloc_context *ac, int alloc_flags,
4374 bool did_some_progress, int *no_progress_loops)
4381 * Costly allocations might have made a progress but this doesn't mean
4382 * their order will become available due to high fragmentation so
4383 * always increment the no progress counter for them
4385 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4386 *no_progress_loops = 0;
4388 (*no_progress_loops)++;
4391 * Make sure we converge to OOM if we cannot make any progress
4392 * several times in the row.
4394 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4395 /* Before OOM, exhaust highatomic_reserve */
4396 return unreserve_highatomic_pageblock(ac, true);
4400 * Keep reclaiming pages while there is a chance this will lead
4401 * somewhere. If none of the target zones can satisfy our allocation
4402 * request even if all reclaimable pages are considered then we are
4403 * screwed and have to go OOM.
4405 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4407 unsigned long available;
4408 unsigned long reclaimable;
4409 unsigned long min_wmark = min_wmark_pages(zone);
4412 available = reclaimable = zone_reclaimable_pages(zone);
4413 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4416 * Would the allocation succeed if we reclaimed all
4417 * reclaimable pages?
4419 wmark = __zone_watermark_ok(zone, order, min_wmark,
4420 ac_classzone_idx(ac), alloc_flags, available);
4421 trace_reclaim_retry_zone(z, order, reclaimable,
4422 available, min_wmark, *no_progress_loops, wmark);
4425 * If we didn't make any progress and have a lot of
4426 * dirty + writeback pages then we should wait for
4427 * an IO to complete to slow down the reclaim and
4428 * prevent from pre mature OOM
4430 if (!did_some_progress) {
4431 unsigned long write_pending;
4433 write_pending = zone_page_state_snapshot(zone,
4434 NR_ZONE_WRITE_PENDING);
4436 if (2 * write_pending > reclaimable) {
4437 congestion_wait(BLK_RW_ASYNC, HZ/10);
4449 * Memory allocation/reclaim might be called from a WQ context and the
4450 * current implementation of the WQ concurrency control doesn't
4451 * recognize that a particular WQ is congested if the worker thread is
4452 * looping without ever sleeping. Therefore we have to do a short sleep
4453 * here rather than calling cond_resched().
4455 if (current->flags & PF_WQ_WORKER)
4456 schedule_timeout_uninterruptible(1);
4463 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4466 * It's possible that cpuset's mems_allowed and the nodemask from
4467 * mempolicy don't intersect. This should be normally dealt with by
4468 * policy_nodemask(), but it's possible to race with cpuset update in
4469 * such a way the check therein was true, and then it became false
4470 * before we got our cpuset_mems_cookie here.
4471 * This assumes that for all allocations, ac->nodemask can come only
4472 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4473 * when it does not intersect with the cpuset restrictions) or the
4474 * caller can deal with a violated nodemask.
4476 if (cpusets_enabled() && ac->nodemask &&
4477 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4478 ac->nodemask = NULL;
4483 * When updating a task's mems_allowed or mempolicy nodemask, it is
4484 * possible to race with parallel threads in such a way that our
4485 * allocation can fail while the mask is being updated. If we are about
4486 * to fail, check if the cpuset changed during allocation and if so,
4489 if (read_mems_allowed_retry(cpuset_mems_cookie))
4495 static inline struct page *
4496 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4497 struct alloc_context *ac)
4499 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4500 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4501 struct page *page = NULL;
4502 unsigned int alloc_flags;
4503 unsigned long did_some_progress;
4504 enum compact_priority compact_priority;
4505 enum compact_result compact_result;
4506 int compaction_retries;
4507 int no_progress_loops;
4508 unsigned int cpuset_mems_cookie;
4512 * We also sanity check to catch abuse of atomic reserves being used by
4513 * callers that are not in atomic context.
4515 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4516 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4517 gfp_mask &= ~__GFP_ATOMIC;
4520 compaction_retries = 0;
4521 no_progress_loops = 0;
4522 compact_priority = DEF_COMPACT_PRIORITY;
4523 cpuset_mems_cookie = read_mems_allowed_begin();
4526 * The fast path uses conservative alloc_flags to succeed only until
4527 * kswapd needs to be woken up, and to avoid the cost of setting up
4528 * alloc_flags precisely. So we do that now.
4530 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4533 * We need to recalculate the starting point for the zonelist iterator
4534 * because we might have used different nodemask in the fast path, or
4535 * there was a cpuset modification and we are retrying - otherwise we
4536 * could end up iterating over non-eligible zones endlessly.
4538 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4539 ac->high_zoneidx, ac->nodemask);
4540 if (!ac->preferred_zoneref->zone)
4543 if (alloc_flags & ALLOC_KSWAPD)
4544 wake_all_kswapds(order, gfp_mask, ac);
4547 * The adjusted alloc_flags might result in immediate success, so try
4550 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4555 * For costly allocations, try direct compaction first, as it's likely
4556 * that we have enough base pages and don't need to reclaim. For non-
4557 * movable high-order allocations, do that as well, as compaction will
4558 * try prevent permanent fragmentation by migrating from blocks of the
4560 * Don't try this for allocations that are allowed to ignore
4561 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4563 if (can_direct_reclaim &&
4565 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4566 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4567 page = __alloc_pages_direct_compact(gfp_mask, order,
4569 INIT_COMPACT_PRIORITY,
4575 * Checks for costly allocations with __GFP_NORETRY, which
4576 * includes some THP page fault allocations
4578 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4580 * If allocating entire pageblock(s) and compaction
4581 * failed because all zones are below low watermarks
4582 * or is prohibited because it recently failed at this
4583 * order, fail immediately unless the allocator has
4584 * requested compaction and reclaim retry.
4587 * - potentially very expensive because zones are far
4588 * below their low watermarks or this is part of very
4589 * bursty high order allocations,
4590 * - not guaranteed to help because isolate_freepages()
4591 * may not iterate over freed pages as part of its
4593 * - unlikely to make entire pageblocks free on its
4596 if (compact_result == COMPACT_SKIPPED ||
4597 compact_result == COMPACT_DEFERRED)
4601 * Looks like reclaim/compaction is worth trying, but
4602 * sync compaction could be very expensive, so keep
4603 * using async compaction.
4605 compact_priority = INIT_COMPACT_PRIORITY;
4610 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4611 if (alloc_flags & ALLOC_KSWAPD)
4612 wake_all_kswapds(order, gfp_mask, ac);
4614 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4616 alloc_flags = reserve_flags;
4619 * Reset the nodemask and zonelist iterators if memory policies can be
4620 * ignored. These allocations are high priority and system rather than
4623 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4624 ac->nodemask = NULL;
4625 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4626 ac->high_zoneidx, ac->nodemask);
4629 /* Attempt with potentially adjusted zonelist and alloc_flags */
4630 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4634 /* Caller is not willing to reclaim, we can't balance anything */
4635 if (!can_direct_reclaim)
4638 /* Avoid recursion of direct reclaim */
4639 if (current->flags & PF_MEMALLOC)
4642 /* Try direct reclaim and then allocating */
4643 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4644 &did_some_progress);
4648 /* Try direct compaction and then allocating */
4649 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4650 compact_priority, &compact_result);
4654 /* Do not loop if specifically requested */
4655 if (gfp_mask & __GFP_NORETRY)
4659 * Do not retry costly high order allocations unless they are
4660 * __GFP_RETRY_MAYFAIL
4662 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4665 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4666 did_some_progress > 0, &no_progress_loops))
4670 * It doesn't make any sense to retry for the compaction if the order-0
4671 * reclaim is not able to make any progress because the current
4672 * implementation of the compaction depends on the sufficient amount
4673 * of free memory (see __compaction_suitable)
4675 if (did_some_progress > 0 &&
4676 should_compact_retry(ac, order, alloc_flags,
4677 compact_result, &compact_priority,
4678 &compaction_retries))
4682 /* Deal with possible cpuset update races before we start OOM killing */
4683 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4686 /* Reclaim has failed us, start killing things */
4687 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4691 /* Avoid allocations with no watermarks from looping endlessly */
4692 if (tsk_is_oom_victim(current) &&
4693 (alloc_flags == ALLOC_OOM ||
4694 (gfp_mask & __GFP_NOMEMALLOC)))
4697 /* Retry as long as the OOM killer is making progress */
4698 if (did_some_progress) {
4699 no_progress_loops = 0;
4704 /* Deal with possible cpuset update races before we fail */
4705 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4709 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4712 if (gfp_mask & __GFP_NOFAIL) {
4714 * All existing users of the __GFP_NOFAIL are blockable, so warn
4715 * of any new users that actually require GFP_NOWAIT
4717 if (WARN_ON_ONCE(!can_direct_reclaim))
4721 * PF_MEMALLOC request from this context is rather bizarre
4722 * because we cannot reclaim anything and only can loop waiting
4723 * for somebody to do a work for us
4725 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4728 * non failing costly orders are a hard requirement which we
4729 * are not prepared for much so let's warn about these users
4730 * so that we can identify them and convert them to something
4733 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4736 * Help non-failing allocations by giving them access to memory
4737 * reserves but do not use ALLOC_NO_WATERMARKS because this
4738 * could deplete whole memory reserves which would just make
4739 * the situation worse
4741 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4749 warn_alloc(gfp_mask, ac->nodemask,
4750 "page allocation failure: order:%u", order);
4755 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4756 int preferred_nid, nodemask_t *nodemask,
4757 struct alloc_context *ac, gfp_t *alloc_mask,
4758 unsigned int *alloc_flags)
4760 ac->high_zoneidx = gfp_zone(gfp_mask);
4761 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4762 ac->nodemask = nodemask;
4763 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4765 if (cpusets_enabled()) {
4766 *alloc_mask |= __GFP_HARDWALL;
4768 ac->nodemask = &cpuset_current_mems_allowed;
4770 *alloc_flags |= ALLOC_CPUSET;
4773 fs_reclaim_acquire(gfp_mask);
4774 fs_reclaim_release(gfp_mask);
4776 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4778 if (should_fail_alloc_page(gfp_mask, order))
4781 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4782 *alloc_flags |= ALLOC_CMA;
4787 /* Determine whether to spread dirty pages and what the first usable zone */
4788 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4790 /* Dirty zone balancing only done in the fast path */
4791 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4794 * The preferred zone is used for statistics but crucially it is
4795 * also used as the starting point for the zonelist iterator. It
4796 * may get reset for allocations that ignore memory policies.
4798 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4799 ac->high_zoneidx, ac->nodemask);
4803 * This is the 'heart' of the zoned buddy allocator.
4806 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4807 nodemask_t *nodemask)
4810 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4811 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4812 struct alloc_context ac = { };
4815 * There are several places where we assume that the order value is sane
4816 * so bail out early if the request is out of bound.
4818 if (unlikely(order >= MAX_ORDER)) {
4819 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4823 gfp_mask &= gfp_allowed_mask;
4824 alloc_mask = gfp_mask;
4825 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4828 finalise_ac(gfp_mask, &ac);
4831 * Forbid the first pass from falling back to types that fragment
4832 * memory until all local zones are considered.
4834 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4836 /* First allocation attempt */
4837 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4842 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4843 * resp. GFP_NOIO which has to be inherited for all allocation requests
4844 * from a particular context which has been marked by
4845 * memalloc_no{fs,io}_{save,restore}.
4847 alloc_mask = current_gfp_context(gfp_mask);
4848 ac.spread_dirty_pages = false;
4851 * Restore the original nodemask if it was potentially replaced with
4852 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4854 ac.nodemask = nodemask;
4856 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4859 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4860 unlikely(__memcg_kmem_charge_page(page, gfp_mask, order) != 0)) {
4861 __free_pages(page, order);
4865 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4869 EXPORT_SYMBOL(__alloc_pages_nodemask);
4872 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4873 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4874 * you need to access high mem.
4876 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4880 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4883 return (unsigned long) page_address(page);
4885 EXPORT_SYMBOL(__get_free_pages);
4887 unsigned long get_zeroed_page(gfp_t gfp_mask)
4889 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4891 EXPORT_SYMBOL(get_zeroed_page);
4893 static inline void free_the_page(struct page *page, unsigned int order)
4895 if (order == 0) /* Via pcp? */
4896 free_unref_page(page);
4898 __free_pages_ok(page, order);
4901 void __free_pages(struct page *page, unsigned int order)
4903 if (put_page_testzero(page))
4904 free_the_page(page, order);
4906 EXPORT_SYMBOL(__free_pages);
4908 void free_pages(unsigned long addr, unsigned int order)
4911 VM_BUG_ON(!virt_addr_valid((void *)addr));
4912 __free_pages(virt_to_page((void *)addr), order);
4916 EXPORT_SYMBOL(free_pages);
4920 * An arbitrary-length arbitrary-offset area of memory which resides
4921 * within a 0 or higher order page. Multiple fragments within that page
4922 * are individually refcounted, in the page's reference counter.
4924 * The page_frag functions below provide a simple allocation framework for
4925 * page fragments. This is used by the network stack and network device
4926 * drivers to provide a backing region of memory for use as either an
4927 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4929 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4932 struct page *page = NULL;
4933 gfp_t gfp = gfp_mask;
4935 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4936 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4938 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4939 PAGE_FRAG_CACHE_MAX_ORDER);
4940 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4942 if (unlikely(!page))
4943 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4945 nc->va = page ? page_address(page) : NULL;
4950 void __page_frag_cache_drain(struct page *page, unsigned int count)
4952 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4954 if (page_ref_sub_and_test(page, count))
4955 free_the_page(page, compound_order(page));
4957 EXPORT_SYMBOL(__page_frag_cache_drain);
4959 void *page_frag_alloc(struct page_frag_cache *nc,
4960 unsigned int fragsz, gfp_t gfp_mask)
4962 unsigned int size = PAGE_SIZE;
4966 if (unlikely(!nc->va)) {
4968 page = __page_frag_cache_refill(nc, gfp_mask);
4972 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4973 /* if size can vary use size else just use PAGE_SIZE */
4976 /* Even if we own the page, we do not use atomic_set().
4977 * This would break get_page_unless_zero() users.
4979 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4981 /* reset page count bias and offset to start of new frag */
4982 nc->pfmemalloc = page_is_pfmemalloc(page);
4983 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4987 offset = nc->offset - fragsz;
4988 if (unlikely(offset < 0)) {
4989 page = virt_to_page(nc->va);
4991 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4994 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4995 /* if size can vary use size else just use PAGE_SIZE */
4998 /* OK, page count is 0, we can safely set it */
4999 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5001 /* reset page count bias and offset to start of new frag */
5002 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5003 offset = size - fragsz;
5007 nc->offset = offset;
5009 return nc->va + offset;
5011 EXPORT_SYMBOL(page_frag_alloc);
5014 * Frees a page fragment allocated out of either a compound or order 0 page.
5016 void page_frag_free(void *addr)
5018 struct page *page = virt_to_head_page(addr);
5020 if (unlikely(put_page_testzero(page)))
5021 free_the_page(page, compound_order(page));
5023 EXPORT_SYMBOL(page_frag_free);
5025 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5029 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5030 unsigned long used = addr + PAGE_ALIGN(size);
5032 split_page(virt_to_page((void *)addr), order);
5033 while (used < alloc_end) {
5038 return (void *)addr;
5042 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5043 * @size: the number of bytes to allocate
5044 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5046 * This function is similar to alloc_pages(), except that it allocates the
5047 * minimum number of pages to satisfy the request. alloc_pages() can only
5048 * allocate memory in power-of-two pages.
5050 * This function is also limited by MAX_ORDER.
5052 * Memory allocated by this function must be released by free_pages_exact().
5054 * Return: pointer to the allocated area or %NULL in case of error.
5056 void *alloc_pages_exact(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 addr = __get_free_pages(gfp_mask, order);
5065 return make_alloc_exact(addr, order, size);
5067 EXPORT_SYMBOL(alloc_pages_exact);
5070 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5072 * @nid: the preferred node ID where memory should be allocated
5073 * @size: the number of bytes to allocate
5074 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5076 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5079 * Return: pointer to the allocated area or %NULL in case of error.
5081 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5083 unsigned int order = get_order(size);
5086 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5087 gfp_mask &= ~__GFP_COMP;
5089 p = alloc_pages_node(nid, gfp_mask, order);
5092 return make_alloc_exact((unsigned long)page_address(p), order, size);
5096 * free_pages_exact - release memory allocated via alloc_pages_exact()
5097 * @virt: the value returned by alloc_pages_exact.
5098 * @size: size of allocation, same value as passed to alloc_pages_exact().
5100 * Release the memory allocated by a previous call to alloc_pages_exact.
5102 void free_pages_exact(void *virt, size_t size)
5104 unsigned long addr = (unsigned long)virt;
5105 unsigned long end = addr + PAGE_ALIGN(size);
5107 while (addr < end) {
5112 EXPORT_SYMBOL(free_pages_exact);
5115 * nr_free_zone_pages - count number of pages beyond high watermark
5116 * @offset: The zone index of the highest zone
5118 * nr_free_zone_pages() counts the number of pages which are beyond the
5119 * high watermark within all zones at or below a given zone index. For each
5120 * zone, the number of pages is calculated as:
5122 * nr_free_zone_pages = managed_pages - high_pages
5124 * Return: number of pages beyond high watermark.
5126 static unsigned long nr_free_zone_pages(int offset)
5131 /* Just pick one node, since fallback list is circular */
5132 unsigned long sum = 0;
5134 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5136 for_each_zone_zonelist(zone, z, zonelist, offset) {
5137 unsigned long size = zone_managed_pages(zone);
5138 unsigned long high = high_wmark_pages(zone);
5147 * nr_free_buffer_pages - count number of pages beyond high watermark
5149 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5150 * watermark within ZONE_DMA and ZONE_NORMAL.
5152 * Return: number of pages beyond high watermark within ZONE_DMA and
5155 unsigned long nr_free_buffer_pages(void)
5157 return nr_free_zone_pages(gfp_zone(GFP_USER));
5159 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5162 * nr_free_pagecache_pages - count number of pages beyond high watermark
5164 * nr_free_pagecache_pages() counts the number of pages which are beyond the
5165 * high watermark within all zones.
5167 * Return: number of pages beyond high watermark within all zones.
5169 unsigned long nr_free_pagecache_pages(void)
5171 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5174 static inline void show_node(struct zone *zone)
5176 if (IS_ENABLED(CONFIG_NUMA))
5177 printk("Node %d ", zone_to_nid(zone));
5180 long si_mem_available(void)
5183 unsigned long pagecache;
5184 unsigned long wmark_low = 0;
5185 unsigned long pages[NR_LRU_LISTS];
5186 unsigned long reclaimable;
5190 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5191 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5194 wmark_low += low_wmark_pages(zone);
5197 * Estimate the amount of memory available for userspace allocations,
5198 * without causing swapping.
5200 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5203 * Not all the page cache can be freed, otherwise the system will
5204 * start swapping. Assume at least half of the page cache, or the
5205 * low watermark worth of cache, needs to stay.
5207 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5208 pagecache -= min(pagecache / 2, wmark_low);
5209 available += pagecache;
5212 * Part of the reclaimable slab and other kernel memory consists of
5213 * items that are in use, and cannot be freed. Cap this estimate at the
5216 reclaimable = global_node_page_state(NR_SLAB_RECLAIMABLE) +
5217 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5218 available += reclaimable - min(reclaimable / 2, wmark_low);
5224 EXPORT_SYMBOL_GPL(si_mem_available);
5226 void si_meminfo(struct sysinfo *val)
5228 val->totalram = totalram_pages();
5229 val->sharedram = global_node_page_state(NR_SHMEM);
5230 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5231 val->bufferram = nr_blockdev_pages();
5232 val->totalhigh = totalhigh_pages();
5233 val->freehigh = nr_free_highpages();
5234 val->mem_unit = PAGE_SIZE;
5237 EXPORT_SYMBOL(si_meminfo);
5240 void si_meminfo_node(struct sysinfo *val, int nid)
5242 int zone_type; /* needs to be signed */
5243 unsigned long managed_pages = 0;
5244 unsigned long managed_highpages = 0;
5245 unsigned long free_highpages = 0;
5246 pg_data_t *pgdat = NODE_DATA(nid);
5248 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5249 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5250 val->totalram = managed_pages;
5251 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5252 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5253 #ifdef CONFIG_HIGHMEM
5254 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5255 struct zone *zone = &pgdat->node_zones[zone_type];
5257 if (is_highmem(zone)) {
5258 managed_highpages += zone_managed_pages(zone);
5259 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5262 val->totalhigh = managed_highpages;
5263 val->freehigh = free_highpages;
5265 val->totalhigh = managed_highpages;
5266 val->freehigh = free_highpages;
5268 val->mem_unit = PAGE_SIZE;
5273 * Determine whether the node should be displayed or not, depending on whether
5274 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5276 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5278 if (!(flags & SHOW_MEM_FILTER_NODES))
5282 * no node mask - aka implicit memory numa policy. Do not bother with
5283 * the synchronization - read_mems_allowed_begin - because we do not
5284 * have to be precise here.
5287 nodemask = &cpuset_current_mems_allowed;
5289 return !node_isset(nid, *nodemask);
5292 #define K(x) ((x) << (PAGE_SHIFT-10))
5294 static void show_migration_types(unsigned char type)
5296 static const char types[MIGRATE_TYPES] = {
5297 [MIGRATE_UNMOVABLE] = 'U',
5298 [MIGRATE_MOVABLE] = 'M',
5299 [MIGRATE_RECLAIMABLE] = 'E',
5300 [MIGRATE_HIGHATOMIC] = 'H',
5302 [MIGRATE_CMA] = 'C',
5304 #ifdef CONFIG_MEMORY_ISOLATION
5305 [MIGRATE_ISOLATE] = 'I',
5308 char tmp[MIGRATE_TYPES + 1];
5312 for (i = 0; i < MIGRATE_TYPES; i++) {
5313 if (type & (1 << i))
5318 printk(KERN_CONT "(%s) ", tmp);
5322 * Show free area list (used inside shift_scroll-lock stuff)
5323 * We also calculate the percentage fragmentation. We do this by counting the
5324 * memory on each free list with the exception of the first item on the list.
5327 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5330 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5332 unsigned long free_pcp = 0;
5337 for_each_populated_zone(zone) {
5338 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5341 for_each_online_cpu(cpu)
5342 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5345 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5346 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5347 " unevictable:%lu dirty:%lu writeback:%lu\n"
5348 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5349 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5350 " free:%lu free_pcp:%lu free_cma:%lu\n",
5351 global_node_page_state(NR_ACTIVE_ANON),
5352 global_node_page_state(NR_INACTIVE_ANON),
5353 global_node_page_state(NR_ISOLATED_ANON),
5354 global_node_page_state(NR_ACTIVE_FILE),
5355 global_node_page_state(NR_INACTIVE_FILE),
5356 global_node_page_state(NR_ISOLATED_FILE),
5357 global_node_page_state(NR_UNEVICTABLE),
5358 global_node_page_state(NR_FILE_DIRTY),
5359 global_node_page_state(NR_WRITEBACK),
5360 global_node_page_state(NR_SLAB_RECLAIMABLE),
5361 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
5362 global_node_page_state(NR_FILE_MAPPED),
5363 global_node_page_state(NR_SHMEM),
5364 global_zone_page_state(NR_PAGETABLE),
5365 global_zone_page_state(NR_BOUNCE),
5366 global_zone_page_state(NR_FREE_PAGES),
5368 global_zone_page_state(NR_FREE_CMA_PAGES));
5370 for_each_online_pgdat(pgdat) {
5371 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5375 " active_anon:%lukB"
5376 " inactive_anon:%lukB"
5377 " active_file:%lukB"
5378 " inactive_file:%lukB"
5379 " unevictable:%lukB"
5380 " isolated(anon):%lukB"
5381 " isolated(file):%lukB"
5386 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5388 " shmem_pmdmapped: %lukB"
5391 " writeback_tmp:%lukB"
5392 " all_unreclaimable? %s"
5395 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5396 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5397 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5398 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5399 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5400 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5401 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5402 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5403 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5404 K(node_page_state(pgdat, NR_WRITEBACK)),
5405 K(node_page_state(pgdat, NR_SHMEM)),
5406 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5407 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5408 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5410 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5412 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5413 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5417 for_each_populated_zone(zone) {
5420 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5424 for_each_online_cpu(cpu)
5425 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5434 " reserved_highatomic:%luKB"
5435 " active_anon:%lukB"
5436 " inactive_anon:%lukB"
5437 " active_file:%lukB"
5438 " inactive_file:%lukB"
5439 " unevictable:%lukB"
5440 " writepending:%lukB"
5444 " kernel_stack:%lukB"
5445 #ifdef CONFIG_SHADOW_CALL_STACK
5446 " shadow_call_stack:%lukB"
5455 K(zone_page_state(zone, NR_FREE_PAGES)),
5456 K(min_wmark_pages(zone)),
5457 K(low_wmark_pages(zone)),
5458 K(high_wmark_pages(zone)),
5459 K(zone->nr_reserved_highatomic),
5460 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5461 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5462 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5463 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5464 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5465 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5466 K(zone->present_pages),
5467 K(zone_managed_pages(zone)),
5468 K(zone_page_state(zone, NR_MLOCK)),
5469 zone_page_state(zone, NR_KERNEL_STACK_KB),
5470 #ifdef CONFIG_SHADOW_CALL_STACK
5471 zone_page_state(zone, NR_KERNEL_SCS_KB),
5473 K(zone_page_state(zone, NR_PAGETABLE)),
5474 K(zone_page_state(zone, NR_BOUNCE)),
5476 K(this_cpu_read(zone->pageset->pcp.count)),
5477 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5478 printk("lowmem_reserve[]:");
5479 for (i = 0; i < MAX_NR_ZONES; i++)
5480 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5481 printk(KERN_CONT "\n");
5484 for_each_populated_zone(zone) {
5486 unsigned long nr[MAX_ORDER], flags, total = 0;
5487 unsigned char types[MAX_ORDER];
5489 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5492 printk(KERN_CONT "%s: ", zone->name);
5494 spin_lock_irqsave(&zone->lock, flags);
5495 for (order = 0; order < MAX_ORDER; order++) {
5496 struct free_area *area = &zone->free_area[order];
5499 nr[order] = area->nr_free;
5500 total += nr[order] << order;
5503 for (type = 0; type < MIGRATE_TYPES; type++) {
5504 if (!free_area_empty(area, type))
5505 types[order] |= 1 << type;
5508 spin_unlock_irqrestore(&zone->lock, flags);
5509 for (order = 0; order < MAX_ORDER; order++) {
5510 printk(KERN_CONT "%lu*%lukB ",
5511 nr[order], K(1UL) << order);
5513 show_migration_types(types[order]);
5515 printk(KERN_CONT "= %lukB\n", K(total));
5518 hugetlb_show_meminfo();
5520 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5522 show_swap_cache_info();
5525 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5527 zoneref->zone = zone;
5528 zoneref->zone_idx = zone_idx(zone);
5532 * Builds allocation fallback zone lists.
5534 * Add all populated zones of a node to the zonelist.
5536 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5539 enum zone_type zone_type = MAX_NR_ZONES;
5544 zone = pgdat->node_zones + zone_type;
5545 if (managed_zone(zone)) {
5546 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5547 check_highest_zone(zone_type);
5549 } while (zone_type);
5556 static int __parse_numa_zonelist_order(char *s)
5559 * We used to support different zonlists modes but they turned
5560 * out to be just not useful. Let's keep the warning in place
5561 * if somebody still use the cmd line parameter so that we do
5562 * not fail it silently
5564 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5565 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5571 static __init int setup_numa_zonelist_order(char *s)
5576 return __parse_numa_zonelist_order(s);
5578 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5580 char numa_zonelist_order[] = "Node";
5583 * sysctl handler for numa_zonelist_order
5585 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5586 void __user *buffer, size_t *length,
5593 return proc_dostring(table, write, buffer, length, ppos);
5594 str = memdup_user_nul(buffer, 16);
5596 return PTR_ERR(str);
5598 ret = __parse_numa_zonelist_order(str);
5604 #define MAX_NODE_LOAD (nr_online_nodes)
5605 static int node_load[MAX_NUMNODES];
5608 * find_next_best_node - find the next node that should appear in a given node's fallback list
5609 * @node: node whose fallback list we're appending
5610 * @used_node_mask: nodemask_t of already used nodes
5612 * We use a number of factors to determine which is the next node that should
5613 * appear on a given node's fallback list. The node should not have appeared
5614 * already in @node's fallback list, and it should be the next closest node
5615 * according to the distance array (which contains arbitrary distance values
5616 * from each node to each node in the system), and should also prefer nodes
5617 * with no CPUs, since presumably they'll have very little allocation pressure
5618 * on them otherwise.
5620 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5622 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5625 int min_val = INT_MAX;
5626 int best_node = NUMA_NO_NODE;
5627 const struct cpumask *tmp = cpumask_of_node(0);
5629 /* Use the local node if we haven't already */
5630 if (!node_isset(node, *used_node_mask)) {
5631 node_set(node, *used_node_mask);
5635 for_each_node_state(n, N_MEMORY) {
5637 /* Don't want a node to appear more than once */
5638 if (node_isset(n, *used_node_mask))
5641 /* Use the distance array to find the distance */
5642 val = node_distance(node, n);
5644 /* Penalize nodes under us ("prefer the next node") */
5647 /* Give preference to headless and unused nodes */
5648 tmp = cpumask_of_node(n);
5649 if (!cpumask_empty(tmp))
5650 val += PENALTY_FOR_NODE_WITH_CPUS;
5652 /* Slight preference for less loaded node */
5653 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5654 val += node_load[n];
5656 if (val < min_val) {
5663 node_set(best_node, *used_node_mask);
5670 * Build zonelists ordered by node and zones within node.
5671 * This results in maximum locality--normal zone overflows into local
5672 * DMA zone, if any--but risks exhausting DMA zone.
5674 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5677 struct zoneref *zonerefs;
5680 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5682 for (i = 0; i < nr_nodes; i++) {
5685 pg_data_t *node = NODE_DATA(node_order[i]);
5687 nr_zones = build_zonerefs_node(node, zonerefs);
5688 zonerefs += nr_zones;
5690 zonerefs->zone = NULL;
5691 zonerefs->zone_idx = 0;
5695 * Build gfp_thisnode zonelists
5697 static void build_thisnode_zonelists(pg_data_t *pgdat)
5699 struct zoneref *zonerefs;
5702 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5703 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5704 zonerefs += nr_zones;
5705 zonerefs->zone = NULL;
5706 zonerefs->zone_idx = 0;
5710 * Build zonelists ordered by zone and nodes within zones.
5711 * This results in conserving DMA zone[s] until all Normal memory is
5712 * exhausted, but results in overflowing to remote node while memory
5713 * may still exist in local DMA zone.
5716 static void build_zonelists(pg_data_t *pgdat)
5718 static int node_order[MAX_NUMNODES];
5719 int node, load, nr_nodes = 0;
5720 nodemask_t used_mask;
5721 int local_node, prev_node;
5723 /* NUMA-aware ordering of nodes */
5724 local_node = pgdat->node_id;
5725 load = nr_online_nodes;
5726 prev_node = local_node;
5727 nodes_clear(used_mask);
5729 memset(node_order, 0, sizeof(node_order));
5730 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5732 * We don't want to pressure a particular node.
5733 * So adding penalty to the first node in same
5734 * distance group to make it round-robin.
5736 if (node_distance(local_node, node) !=
5737 node_distance(local_node, prev_node))
5738 node_load[node] = load;
5740 node_order[nr_nodes++] = node;
5745 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5746 build_thisnode_zonelists(pgdat);
5749 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5751 * Return node id of node used for "local" allocations.
5752 * I.e., first node id of first zone in arg node's generic zonelist.
5753 * Used for initializing percpu 'numa_mem', which is used primarily
5754 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5756 int local_memory_node(int node)
5760 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5761 gfp_zone(GFP_KERNEL),
5763 return zone_to_nid(z->zone);
5767 static void setup_min_unmapped_ratio(void);
5768 static void setup_min_slab_ratio(void);
5769 #else /* CONFIG_NUMA */
5771 static void build_zonelists(pg_data_t *pgdat)
5773 int node, local_node;
5774 struct zoneref *zonerefs;
5777 local_node = pgdat->node_id;
5779 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5780 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5781 zonerefs += nr_zones;
5784 * Now we build the zonelist so that it contains the zones
5785 * of all the other nodes.
5786 * We don't want to pressure a particular node, so when
5787 * building the zones for node N, we make sure that the
5788 * zones coming right after the local ones are those from
5789 * node N+1 (modulo N)
5791 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5792 if (!node_online(node))
5794 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5795 zonerefs += nr_zones;
5797 for (node = 0; node < local_node; node++) {
5798 if (!node_online(node))
5800 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5801 zonerefs += nr_zones;
5804 zonerefs->zone = NULL;
5805 zonerefs->zone_idx = 0;
5808 #endif /* CONFIG_NUMA */
5811 * Boot pageset table. One per cpu which is going to be used for all
5812 * zones and all nodes. The parameters will be set in such a way
5813 * that an item put on a list will immediately be handed over to
5814 * the buddy list. This is safe since pageset manipulation is done
5815 * with interrupts disabled.
5817 * The boot_pagesets must be kept even after bootup is complete for
5818 * unused processors and/or zones. They do play a role for bootstrapping
5819 * hotplugged processors.
5821 * zoneinfo_show() and maybe other functions do
5822 * not check if the processor is online before following the pageset pointer.
5823 * Other parts of the kernel may not check if the zone is available.
5825 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5826 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5827 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5829 static void __build_all_zonelists(void *data)
5832 int __maybe_unused cpu;
5833 pg_data_t *self = data;
5834 static DEFINE_SPINLOCK(lock);
5839 memset(node_load, 0, sizeof(node_load));
5843 * This node is hotadded and no memory is yet present. So just
5844 * building zonelists is fine - no need to touch other nodes.
5846 if (self && !node_online(self->node_id)) {
5847 build_zonelists(self);
5849 for_each_online_node(nid) {
5850 pg_data_t *pgdat = NODE_DATA(nid);
5852 build_zonelists(pgdat);
5855 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5857 * We now know the "local memory node" for each node--
5858 * i.e., the node of the first zone in the generic zonelist.
5859 * Set up numa_mem percpu variable for on-line cpus. During
5860 * boot, only the boot cpu should be on-line; we'll init the
5861 * secondary cpus' numa_mem as they come on-line. During
5862 * node/memory hotplug, we'll fixup all on-line cpus.
5864 for_each_online_cpu(cpu)
5865 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5872 static noinline void __init
5873 build_all_zonelists_init(void)
5877 __build_all_zonelists(NULL);
5880 * Initialize the boot_pagesets that are going to be used
5881 * for bootstrapping processors. The real pagesets for
5882 * each zone will be allocated later when the per cpu
5883 * allocator is available.
5885 * boot_pagesets are used also for bootstrapping offline
5886 * cpus if the system is already booted because the pagesets
5887 * are needed to initialize allocators on a specific cpu too.
5888 * F.e. the percpu allocator needs the page allocator which
5889 * needs the percpu allocator in order to allocate its pagesets
5890 * (a chicken-egg dilemma).
5892 for_each_possible_cpu(cpu)
5893 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5895 mminit_verify_zonelist();
5896 cpuset_init_current_mems_allowed();
5900 * unless system_state == SYSTEM_BOOTING.
5902 * __ref due to call of __init annotated helper build_all_zonelists_init
5903 * [protected by SYSTEM_BOOTING].
5905 void __ref build_all_zonelists(pg_data_t *pgdat)
5907 if (system_state == SYSTEM_BOOTING) {
5908 build_all_zonelists_init();
5910 __build_all_zonelists(pgdat);
5911 /* cpuset refresh routine should be here */
5913 vm_total_pages = nr_free_pagecache_pages();
5915 * Disable grouping by mobility if the number of pages in the
5916 * system is too low to allow the mechanism to work. It would be
5917 * more accurate, but expensive to check per-zone. This check is
5918 * made on memory-hotadd so a system can start with mobility
5919 * disabled and enable it later
5921 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5922 page_group_by_mobility_disabled = 1;
5924 page_group_by_mobility_disabled = 0;
5926 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5928 page_group_by_mobility_disabled ? "off" : "on",
5931 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5935 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5936 static bool __meminit
5937 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5939 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5940 static struct memblock_region *r;
5942 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5943 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5944 for_each_memblock(memory, r) {
5945 if (*pfn < memblock_region_memory_end_pfn(r))
5949 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5950 memblock_is_mirror(r)) {
5951 *pfn = memblock_region_memory_end_pfn(r);
5959 #ifdef CONFIG_SPARSEMEM
5960 /* Skip PFNs that belong to non-present sections */
5961 static inline __meminit unsigned long next_pfn(unsigned long pfn)
5963 const unsigned long section_nr = pfn_to_section_nr(++pfn);
5965 if (present_section_nr(section_nr))
5967 return section_nr_to_pfn(next_present_section_nr(section_nr));
5970 static inline __meminit unsigned long next_pfn(unsigned long pfn)
5977 * Initially all pages are reserved - free ones are freed
5978 * up by memblock_free_all() once the early boot process is
5979 * done. Non-atomic initialization, single-pass.
5981 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5982 unsigned long start_pfn, enum memmap_context context,
5983 struct vmem_altmap *altmap)
5985 unsigned long pfn, end_pfn = start_pfn + size;
5988 if (highest_memmap_pfn < end_pfn - 1)
5989 highest_memmap_pfn = end_pfn - 1;
5991 #ifdef CONFIG_ZONE_DEVICE
5993 * Honor reservation requested by the driver for this ZONE_DEVICE
5994 * memory. We limit the total number of pages to initialize to just
5995 * those that might contain the memory mapping. We will defer the
5996 * ZONE_DEVICE page initialization until after we have released
5999 if (zone == ZONE_DEVICE) {
6003 if (start_pfn == altmap->base_pfn)
6004 start_pfn += altmap->reserve;
6005 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6009 for (pfn = start_pfn; pfn < end_pfn; ) {
6011 * There can be holes in boot-time mem_map[]s handed to this
6012 * function. They do not exist on hotplugged memory.
6014 if (context == MEMMAP_EARLY) {
6015 if (!early_pfn_valid(pfn)) {
6016 pfn = next_pfn(pfn);
6019 if (!early_pfn_in_nid(pfn, nid)) {
6023 if (overlap_memmap_init(zone, &pfn))
6025 if (defer_init(nid, pfn, end_pfn))
6029 page = pfn_to_page(pfn);
6030 __init_single_page(page, pfn, zone, nid);
6031 if (context == MEMMAP_HOTPLUG)
6032 __SetPageReserved(page);
6035 * Mark the block movable so that blocks are reserved for
6036 * movable at startup. This will force kernel allocations
6037 * to reserve their blocks rather than leaking throughout
6038 * the address space during boot when many long-lived
6039 * kernel allocations are made.
6041 * bitmap is created for zone's valid pfn range. but memmap
6042 * can be created for invalid pages (for alignment)
6043 * check here not to call set_pageblock_migratetype() against
6046 if (!(pfn & (pageblock_nr_pages - 1))) {
6047 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6054 #ifdef CONFIG_ZONE_DEVICE
6055 void __ref memmap_init_zone_device(struct zone *zone,
6056 unsigned long start_pfn,
6057 unsigned long nr_pages,
6058 struct dev_pagemap *pgmap)
6060 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6061 struct pglist_data *pgdat = zone->zone_pgdat;
6062 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6063 unsigned long zone_idx = zone_idx(zone);
6064 unsigned long start = jiffies;
6065 int nid = pgdat->node_id;
6067 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6071 * The call to memmap_init_zone should have already taken care
6072 * of the pages reserved for the memmap, so we can just jump to
6073 * the end of that region and start processing the device pages.
6076 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6077 nr_pages = end_pfn - start_pfn;
6080 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6081 struct page *page = pfn_to_page(pfn);
6083 __init_single_page(page, pfn, zone_idx, nid);
6086 * Mark page reserved as it will need to wait for onlining
6087 * phase for it to be fully associated with a zone.
6089 * We can use the non-atomic __set_bit operation for setting
6090 * the flag as we are still initializing the pages.
6092 __SetPageReserved(page);
6095 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6096 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6097 * ever freed or placed on a driver-private list.
6099 page->pgmap = pgmap;
6100 page->zone_device_data = NULL;
6103 * Mark the block movable so that blocks are reserved for
6104 * movable at startup. This will force kernel allocations
6105 * to reserve their blocks rather than leaking throughout
6106 * the address space during boot when many long-lived
6107 * kernel allocations are made.
6109 * bitmap is created for zone's valid pfn range. but memmap
6110 * can be created for invalid pages (for alignment)
6111 * check here not to call set_pageblock_migratetype() against
6114 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
6115 * because this is done early in section_activate()
6117 if (!(pfn & (pageblock_nr_pages - 1))) {
6118 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6123 pr_info("%s initialised %lu pages in %ums\n", __func__,
6124 nr_pages, jiffies_to_msecs(jiffies - start));
6128 static void __meminit zone_init_free_lists(struct zone *zone)
6130 unsigned int order, t;
6131 for_each_migratetype_order(order, t) {
6132 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6133 zone->free_area[order].nr_free = 0;
6137 void __meminit __weak memmap_init(unsigned long size, int nid,
6138 unsigned long zone, unsigned long start_pfn)
6140 memmap_init_zone(size, nid, zone, start_pfn, MEMMAP_EARLY, NULL);
6143 static int zone_batchsize(struct zone *zone)
6149 * The per-cpu-pages pools are set to around 1000th of the
6152 batch = zone_managed_pages(zone) / 1024;
6153 /* But no more than a meg. */
6154 if (batch * PAGE_SIZE > 1024 * 1024)
6155 batch = (1024 * 1024) / PAGE_SIZE;
6156 batch /= 4; /* We effectively *= 4 below */
6161 * Clamp the batch to a 2^n - 1 value. Having a power
6162 * of 2 value was found to be more likely to have
6163 * suboptimal cache aliasing properties in some cases.
6165 * For example if 2 tasks are alternately allocating
6166 * batches of pages, one task can end up with a lot
6167 * of pages of one half of the possible page colors
6168 * and the other with pages of the other colors.
6170 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6175 /* The deferral and batching of frees should be suppressed under NOMMU
6178 * The problem is that NOMMU needs to be able to allocate large chunks
6179 * of contiguous memory as there's no hardware page translation to
6180 * assemble apparent contiguous memory from discontiguous pages.
6182 * Queueing large contiguous runs of pages for batching, however,
6183 * causes the pages to actually be freed in smaller chunks. As there
6184 * can be a significant delay between the individual batches being
6185 * recycled, this leads to the once large chunks of space being
6186 * fragmented and becoming unavailable for high-order allocations.
6193 * pcp->high and pcp->batch values are related and dependent on one another:
6194 * ->batch must never be higher then ->high.
6195 * The following function updates them in a safe manner without read side
6198 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6199 * those fields changing asynchronously (acording the the above rule).
6201 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6202 * outside of boot time (or some other assurance that no concurrent updaters
6205 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6206 unsigned long batch)
6208 /* start with a fail safe value for batch */
6212 /* Update high, then batch, in order */
6219 /* a companion to pageset_set_high() */
6220 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6222 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6225 static void pageset_init(struct per_cpu_pageset *p)
6227 struct per_cpu_pages *pcp;
6230 memset(p, 0, sizeof(*p));
6233 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6234 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6237 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6240 pageset_set_batch(p, batch);
6244 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6245 * to the value high for the pageset p.
6247 static void pageset_set_high(struct per_cpu_pageset *p,
6250 unsigned long batch = max(1UL, high / 4);
6251 if ((high / 4) > (PAGE_SHIFT * 8))
6252 batch = PAGE_SHIFT * 8;
6254 pageset_update(&p->pcp, high, batch);
6257 static void pageset_set_high_and_batch(struct zone *zone,
6258 struct per_cpu_pageset *pcp)
6260 if (percpu_pagelist_fraction)
6261 pageset_set_high(pcp,
6262 (zone_managed_pages(zone) /
6263 percpu_pagelist_fraction));
6265 pageset_set_batch(pcp, zone_batchsize(zone));
6268 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6270 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6273 pageset_set_high_and_batch(zone, pcp);
6276 void __meminit setup_zone_pageset(struct zone *zone)
6279 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6280 for_each_possible_cpu(cpu)
6281 zone_pageset_init(zone, cpu);
6285 * Allocate per cpu pagesets and initialize them.
6286 * Before this call only boot pagesets were available.
6288 void __init setup_per_cpu_pageset(void)
6290 struct pglist_data *pgdat;
6293 for_each_populated_zone(zone)
6294 setup_zone_pageset(zone);
6296 for_each_online_pgdat(pgdat)
6297 pgdat->per_cpu_nodestats =
6298 alloc_percpu(struct per_cpu_nodestat);
6301 static __meminit void zone_pcp_init(struct zone *zone)
6304 * per cpu subsystem is not up at this point. The following code
6305 * relies on the ability of the linker to provide the
6306 * offset of a (static) per cpu variable into the per cpu area.
6308 zone->pageset = &boot_pageset;
6310 if (populated_zone(zone))
6311 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6312 zone->name, zone->present_pages,
6313 zone_batchsize(zone));
6316 void __meminit init_currently_empty_zone(struct zone *zone,
6317 unsigned long zone_start_pfn,
6320 struct pglist_data *pgdat = zone->zone_pgdat;
6321 int zone_idx = zone_idx(zone) + 1;
6323 if (zone_idx > pgdat->nr_zones)
6324 pgdat->nr_zones = zone_idx;
6326 zone->zone_start_pfn = zone_start_pfn;
6328 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6329 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6331 (unsigned long)zone_idx(zone),
6332 zone_start_pfn, (zone_start_pfn + size));
6334 zone_init_free_lists(zone);
6335 zone->initialized = 1;
6339 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
6340 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
6341 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
6343 * If an architecture guarantees that all ranges registered contain no holes
6344 * and may be freed, this this function may be used instead of calling
6345 * memblock_free_early_nid() manually.
6347 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
6349 unsigned long start_pfn, end_pfn;
6352 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
6353 start_pfn = min(start_pfn, max_low_pfn);
6354 end_pfn = min(end_pfn, max_low_pfn);
6356 if (start_pfn < end_pfn)
6357 memblock_free_early_nid(PFN_PHYS(start_pfn),
6358 (end_pfn - start_pfn) << PAGE_SHIFT,
6364 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6365 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6367 * If an architecture guarantees that all ranges registered contain no holes and may
6368 * be freed, this function may be used instead of calling memory_present() manually.
6370 void __init sparse_memory_present_with_active_regions(int nid)
6372 unsigned long start_pfn, end_pfn;
6375 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
6376 memory_present(this_nid, start_pfn, end_pfn);
6380 * get_pfn_range_for_nid - Return the start and end page frames for a node
6381 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6382 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6383 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6385 * It returns the start and end page frame of a node based on information
6386 * provided by memblock_set_node(). If called for a node
6387 * with no available memory, a warning is printed and the start and end
6390 void __init get_pfn_range_for_nid(unsigned int nid,
6391 unsigned long *start_pfn, unsigned long *end_pfn)
6393 unsigned long this_start_pfn, this_end_pfn;
6399 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6400 *start_pfn = min(*start_pfn, this_start_pfn);
6401 *end_pfn = max(*end_pfn, this_end_pfn);
6404 if (*start_pfn == -1UL)
6409 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6410 * assumption is made that zones within a node are ordered in monotonic
6411 * increasing memory addresses so that the "highest" populated zone is used
6413 static void __init find_usable_zone_for_movable(void)
6416 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6417 if (zone_index == ZONE_MOVABLE)
6420 if (arch_zone_highest_possible_pfn[zone_index] >
6421 arch_zone_lowest_possible_pfn[zone_index])
6425 VM_BUG_ON(zone_index == -1);
6426 movable_zone = zone_index;
6430 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6431 * because it is sized independent of architecture. Unlike the other zones,
6432 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6433 * in each node depending on the size of each node and how evenly kernelcore
6434 * is distributed. This helper function adjusts the zone ranges
6435 * provided by the architecture for a given node by using the end of the
6436 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6437 * zones within a node are in order of monotonic increases memory addresses
6439 static void __init adjust_zone_range_for_zone_movable(int nid,
6440 unsigned long zone_type,
6441 unsigned long node_start_pfn,
6442 unsigned long node_end_pfn,
6443 unsigned long *zone_start_pfn,
6444 unsigned long *zone_end_pfn)
6446 /* Only adjust if ZONE_MOVABLE is on this node */
6447 if (zone_movable_pfn[nid]) {
6448 /* Size ZONE_MOVABLE */
6449 if (zone_type == ZONE_MOVABLE) {
6450 *zone_start_pfn = zone_movable_pfn[nid];
6451 *zone_end_pfn = min(node_end_pfn,
6452 arch_zone_highest_possible_pfn[movable_zone]);
6454 /* Adjust for ZONE_MOVABLE starting within this range */
6455 } else if (!mirrored_kernelcore &&
6456 *zone_start_pfn < zone_movable_pfn[nid] &&
6457 *zone_end_pfn > zone_movable_pfn[nid]) {
6458 *zone_end_pfn = zone_movable_pfn[nid];
6460 /* Check if this whole range is within ZONE_MOVABLE */
6461 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6462 *zone_start_pfn = *zone_end_pfn;
6467 * Return the number of pages a zone spans in a node, including holes
6468 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6470 static unsigned long __init zone_spanned_pages_in_node(int nid,
6471 unsigned long zone_type,
6472 unsigned long node_start_pfn,
6473 unsigned long node_end_pfn,
6474 unsigned long *zone_start_pfn,
6475 unsigned long *zone_end_pfn,
6476 unsigned long *ignored)
6478 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6479 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6480 /* When hotadd a new node from cpu_up(), the node should be empty */
6481 if (!node_start_pfn && !node_end_pfn)
6484 /* Get the start and end of the zone */
6485 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6486 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6487 adjust_zone_range_for_zone_movable(nid, zone_type,
6488 node_start_pfn, node_end_pfn,
6489 zone_start_pfn, zone_end_pfn);
6491 /* Check that this node has pages within the zone's required range */
6492 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6495 /* Move the zone boundaries inside the node if necessary */
6496 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6497 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6499 /* Return the spanned pages */
6500 return *zone_end_pfn - *zone_start_pfn;
6504 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6505 * then all holes in the requested range will be accounted for.
6507 unsigned long __init __absent_pages_in_range(int nid,
6508 unsigned long range_start_pfn,
6509 unsigned long range_end_pfn)
6511 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6512 unsigned long start_pfn, end_pfn;
6515 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6516 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6517 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6518 nr_absent -= end_pfn - start_pfn;
6524 * absent_pages_in_range - Return number of page frames in holes within a range
6525 * @start_pfn: The start PFN to start searching for holes
6526 * @end_pfn: The end PFN to stop searching for holes
6528 * Return: the number of pages frames in memory holes within a range.
6530 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6531 unsigned long end_pfn)
6533 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6536 /* Return the number of page frames in holes in a zone on a node */
6537 static unsigned long __init zone_absent_pages_in_node(int nid,
6538 unsigned long zone_type,
6539 unsigned long node_start_pfn,
6540 unsigned long node_end_pfn,
6541 unsigned long *ignored)
6543 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6544 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6545 unsigned long zone_start_pfn, zone_end_pfn;
6546 unsigned long nr_absent;
6548 /* When hotadd a new node from cpu_up(), the node should be empty */
6549 if (!node_start_pfn && !node_end_pfn)
6552 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6553 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6555 adjust_zone_range_for_zone_movable(nid, zone_type,
6556 node_start_pfn, node_end_pfn,
6557 &zone_start_pfn, &zone_end_pfn);
6558 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6561 * ZONE_MOVABLE handling.
6562 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6565 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6566 unsigned long start_pfn, end_pfn;
6567 struct memblock_region *r;
6569 for_each_memblock(memory, r) {
6570 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6571 zone_start_pfn, zone_end_pfn);
6572 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6573 zone_start_pfn, zone_end_pfn);
6575 if (zone_type == ZONE_MOVABLE &&
6576 memblock_is_mirror(r))
6577 nr_absent += end_pfn - start_pfn;
6579 if (zone_type == ZONE_NORMAL &&
6580 !memblock_is_mirror(r))
6581 nr_absent += end_pfn - start_pfn;
6588 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6589 static inline unsigned long __init zone_spanned_pages_in_node(int nid,
6590 unsigned long zone_type,
6591 unsigned long node_start_pfn,
6592 unsigned long node_end_pfn,
6593 unsigned long *zone_start_pfn,
6594 unsigned long *zone_end_pfn,
6595 unsigned long *zones_size)
6599 *zone_start_pfn = node_start_pfn;
6600 for (zone = 0; zone < zone_type; zone++)
6601 *zone_start_pfn += zones_size[zone];
6603 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6605 return zones_size[zone_type];
6608 static inline unsigned long __init zone_absent_pages_in_node(int nid,
6609 unsigned long zone_type,
6610 unsigned long node_start_pfn,
6611 unsigned long node_end_pfn,
6612 unsigned long *zholes_size)
6617 return zholes_size[zone_type];
6620 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6622 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6623 unsigned long node_start_pfn,
6624 unsigned long node_end_pfn,
6625 unsigned long *zones_size,
6626 unsigned long *zholes_size)
6628 unsigned long realtotalpages = 0, totalpages = 0;
6631 for (i = 0; i < MAX_NR_ZONES; i++) {
6632 struct zone *zone = pgdat->node_zones + i;
6633 unsigned long zone_start_pfn, zone_end_pfn;
6634 unsigned long size, real_size;
6636 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6642 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6643 node_start_pfn, node_end_pfn,
6646 zone->zone_start_pfn = zone_start_pfn;
6648 zone->zone_start_pfn = 0;
6649 zone->spanned_pages = size;
6650 zone->present_pages = real_size;
6653 realtotalpages += real_size;
6656 pgdat->node_spanned_pages = totalpages;
6657 pgdat->node_present_pages = realtotalpages;
6658 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6662 #ifndef CONFIG_SPARSEMEM
6664 * Calculate the size of the zone->blockflags rounded to an unsigned long
6665 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6666 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6667 * round what is now in bits to nearest long in bits, then return it in
6670 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6672 unsigned long usemapsize;
6674 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6675 usemapsize = roundup(zonesize, pageblock_nr_pages);
6676 usemapsize = usemapsize >> pageblock_order;
6677 usemapsize *= NR_PAGEBLOCK_BITS;
6678 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6680 return usemapsize / 8;
6683 static void __ref setup_usemap(struct pglist_data *pgdat,
6685 unsigned long zone_start_pfn,
6686 unsigned long zonesize)
6688 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6689 zone->pageblock_flags = NULL;
6691 zone->pageblock_flags =
6692 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6694 if (!zone->pageblock_flags)
6695 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6696 usemapsize, zone->name, pgdat->node_id);
6700 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6701 unsigned long zone_start_pfn, unsigned long zonesize) {}
6702 #endif /* CONFIG_SPARSEMEM */
6704 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6706 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6707 void __init set_pageblock_order(void)
6711 /* Check that pageblock_nr_pages has not already been setup */
6712 if (pageblock_order)
6715 if (HPAGE_SHIFT > PAGE_SHIFT)
6716 order = HUGETLB_PAGE_ORDER;
6718 order = MAX_ORDER - 1;
6721 * Assume the largest contiguous order of interest is a huge page.
6722 * This value may be variable depending on boot parameters on IA64 and
6725 pageblock_order = order;
6727 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6730 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6731 * is unused as pageblock_order is set at compile-time. See
6732 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6735 void __init set_pageblock_order(void)
6739 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6741 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6742 unsigned long present_pages)
6744 unsigned long pages = spanned_pages;
6747 * Provide a more accurate estimation if there are holes within
6748 * the zone and SPARSEMEM is in use. If there are holes within the
6749 * zone, each populated memory region may cost us one or two extra
6750 * memmap pages due to alignment because memmap pages for each
6751 * populated regions may not be naturally aligned on page boundary.
6752 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6754 if (spanned_pages > present_pages + (present_pages >> 4) &&
6755 IS_ENABLED(CONFIG_SPARSEMEM))
6756 pages = present_pages;
6758 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6761 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6762 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6764 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6766 spin_lock_init(&ds_queue->split_queue_lock);
6767 INIT_LIST_HEAD(&ds_queue->split_queue);
6768 ds_queue->split_queue_len = 0;
6771 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6774 #ifdef CONFIG_COMPACTION
6775 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6777 init_waitqueue_head(&pgdat->kcompactd_wait);
6780 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6783 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6785 pgdat_resize_init(pgdat);
6787 pgdat_init_split_queue(pgdat);
6788 pgdat_init_kcompactd(pgdat);
6790 init_waitqueue_head(&pgdat->kswapd_wait);
6791 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6793 pgdat_page_ext_init(pgdat);
6794 spin_lock_init(&pgdat->lru_lock);
6795 lruvec_init(&pgdat->__lruvec);
6798 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6799 unsigned long remaining_pages)
6801 atomic_long_set(&zone->managed_pages, remaining_pages);
6802 zone_set_nid(zone, nid);
6803 zone->name = zone_names[idx];
6804 zone->zone_pgdat = NODE_DATA(nid);
6805 spin_lock_init(&zone->lock);
6806 zone_seqlock_init(zone);
6807 zone_pcp_init(zone);
6811 * Set up the zone data structures
6812 * - init pgdat internals
6813 * - init all zones belonging to this node
6815 * NOTE: this function is only called during memory hotplug
6817 #ifdef CONFIG_MEMORY_HOTPLUG
6818 void __ref free_area_init_core_hotplug(int nid)
6821 pg_data_t *pgdat = NODE_DATA(nid);
6823 pgdat_init_internals(pgdat);
6824 for (z = 0; z < MAX_NR_ZONES; z++)
6825 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6830 * Set up the zone data structures:
6831 * - mark all pages reserved
6832 * - mark all memory queues empty
6833 * - clear the memory bitmaps
6835 * NOTE: pgdat should get zeroed by caller.
6836 * NOTE: this function is only called during early init.
6838 static void __init free_area_init_core(struct pglist_data *pgdat)
6841 int nid = pgdat->node_id;
6843 pgdat_init_internals(pgdat);
6844 pgdat->per_cpu_nodestats = &boot_nodestats;
6846 for (j = 0; j < MAX_NR_ZONES; j++) {
6847 struct zone *zone = pgdat->node_zones + j;
6848 unsigned long size, freesize, memmap_pages;
6849 unsigned long zone_start_pfn = zone->zone_start_pfn;
6851 size = zone->spanned_pages;
6852 freesize = zone->present_pages;
6855 * Adjust freesize so that it accounts for how much memory
6856 * is used by this zone for memmap. This affects the watermark
6857 * and per-cpu initialisations
6859 memmap_pages = calc_memmap_size(size, freesize);
6860 if (!is_highmem_idx(j)) {
6861 if (freesize >= memmap_pages) {
6862 freesize -= memmap_pages;
6865 " %s zone: %lu pages used for memmap\n",
6866 zone_names[j], memmap_pages);
6868 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6869 zone_names[j], memmap_pages, freesize);
6872 /* Account for reserved pages */
6873 if (j == 0 && freesize > dma_reserve) {
6874 freesize -= dma_reserve;
6875 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6876 zone_names[0], dma_reserve);
6879 if (!is_highmem_idx(j))
6880 nr_kernel_pages += freesize;
6881 /* Charge for highmem memmap if there are enough kernel pages */
6882 else if (nr_kernel_pages > memmap_pages * 2)
6883 nr_kernel_pages -= memmap_pages;
6884 nr_all_pages += freesize;
6887 * Set an approximate value for lowmem here, it will be adjusted
6888 * when the bootmem allocator frees pages into the buddy system.
6889 * And all highmem pages will be managed by the buddy system.
6891 zone_init_internals(zone, j, nid, freesize);
6896 set_pageblock_order();
6897 setup_usemap(pgdat, zone, zone_start_pfn, size);
6898 init_currently_empty_zone(zone, zone_start_pfn, size);
6899 memmap_init(size, nid, j, zone_start_pfn);
6903 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6904 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6906 unsigned long __maybe_unused start = 0;
6907 unsigned long __maybe_unused offset = 0;
6909 /* Skip empty nodes */
6910 if (!pgdat->node_spanned_pages)
6913 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6914 offset = pgdat->node_start_pfn - start;
6915 /* ia64 gets its own node_mem_map, before this, without bootmem */
6916 if (!pgdat->node_mem_map) {
6917 unsigned long size, end;
6921 * The zone's endpoints aren't required to be MAX_ORDER
6922 * aligned but the node_mem_map endpoints must be in order
6923 * for the buddy allocator to function correctly.
6925 end = pgdat_end_pfn(pgdat);
6926 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6927 size = (end - start) * sizeof(struct page);
6928 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6931 panic("Failed to allocate %ld bytes for node %d memory map\n",
6932 size, pgdat->node_id);
6933 pgdat->node_mem_map = map + offset;
6935 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6936 __func__, pgdat->node_id, (unsigned long)pgdat,
6937 (unsigned long)pgdat->node_mem_map);
6938 #ifndef CONFIG_NEED_MULTIPLE_NODES
6940 * With no DISCONTIG, the global mem_map is just set as node 0's
6942 if (pgdat == NODE_DATA(0)) {
6943 mem_map = NODE_DATA(0)->node_mem_map;
6944 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6945 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6947 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6952 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6953 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6955 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6956 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6958 pgdat->first_deferred_pfn = ULONG_MAX;
6961 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6964 void __init free_area_init_node(int nid, unsigned long *zones_size,
6965 unsigned long node_start_pfn,
6966 unsigned long *zholes_size)
6968 pg_data_t *pgdat = NODE_DATA(nid);
6969 unsigned long start_pfn = 0;
6970 unsigned long end_pfn = 0;
6972 /* pg_data_t should be reset to zero when it's allocated */
6973 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6975 pgdat->node_id = nid;
6976 pgdat->node_start_pfn = node_start_pfn;
6977 pgdat->per_cpu_nodestats = NULL;
6978 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6979 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6980 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6981 (u64)start_pfn << PAGE_SHIFT,
6982 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6984 start_pfn = node_start_pfn;
6986 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6987 zones_size, zholes_size);
6989 alloc_node_mem_map(pgdat);
6990 pgdat_set_deferred_range(pgdat);
6992 free_area_init_core(pgdat);
6995 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6997 * Initialize all valid struct pages in the range [spfn, epfn) and mark them
6998 * PageReserved(). Return the number of struct pages that were initialized.
7000 static u64 __init init_unavailable_range(unsigned long spfn, unsigned long epfn)
7005 for (pfn = spfn; pfn < epfn; pfn++) {
7006 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
7007 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
7008 + pageblock_nr_pages - 1;
7012 * Use a fake node/zone (0) for now. Some of these pages
7013 * (in memblock.reserved but not in memblock.memory) will
7014 * get re-initialized via reserve_bootmem_region() later.
7016 __init_single_page(pfn_to_page(pfn), pfn, 0, 0);
7017 __SetPageReserved(pfn_to_page(pfn));
7025 * Only struct pages that are backed by physical memory are zeroed and
7026 * initialized by going through __init_single_page(). But, there are some
7027 * struct pages which are reserved in memblock allocator and their fields
7028 * may be accessed (for example page_to_pfn() on some configuration accesses
7029 * flags). We must explicitly initialize those struct pages.
7031 * This function also addresses a similar issue where struct pages are left
7032 * uninitialized because the physical address range is not covered by
7033 * memblock.memory or memblock.reserved. That could happen when memblock
7034 * layout is manually configured via memmap=, or when the highest physical
7035 * address (max_pfn) does not end on a section boundary.
7037 static void __init init_unavailable_mem(void)
7039 phys_addr_t start, end;
7041 phys_addr_t next = 0;
7044 * Loop through unavailable ranges not covered by memblock.memory.
7047 for_each_mem_range(i, &memblock.memory, NULL,
7048 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
7050 pgcnt += init_unavailable_range(PFN_DOWN(next),
7056 * Early sections always have a fully populated memmap for the whole
7057 * section - see pfn_valid(). If the last section has holes at the
7058 * end and that section is marked "online", the memmap will be
7059 * considered initialized. Make sure that memmap has a well defined
7062 pgcnt += init_unavailable_range(PFN_DOWN(next),
7063 round_up(max_pfn, PAGES_PER_SECTION));
7066 * Struct pages that do not have backing memory. This could be because
7067 * firmware is using some of this memory, or for some other reasons.
7070 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
7073 static inline void __init init_unavailable_mem(void)
7076 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
7078 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
7080 #if MAX_NUMNODES > 1
7082 * Figure out the number of possible node ids.
7084 void __init setup_nr_node_ids(void)
7086 unsigned int highest;
7088 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7089 nr_node_ids = highest + 1;
7094 * node_map_pfn_alignment - determine the maximum internode alignment
7096 * This function should be called after node map is populated and sorted.
7097 * It calculates the maximum power of two alignment which can distinguish
7100 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7101 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7102 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7103 * shifted, 1GiB is enough and this function will indicate so.
7105 * This is used to test whether pfn -> nid mapping of the chosen memory
7106 * model has fine enough granularity to avoid incorrect mapping for the
7107 * populated node map.
7109 * Return: the determined alignment in pfn's. 0 if there is no alignment
7110 * requirement (single node).
7112 unsigned long __init node_map_pfn_alignment(void)
7114 unsigned long accl_mask = 0, last_end = 0;
7115 unsigned long start, end, mask;
7116 int last_nid = NUMA_NO_NODE;
7119 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7120 if (!start || last_nid < 0 || last_nid == nid) {
7127 * Start with a mask granular enough to pin-point to the
7128 * start pfn and tick off bits one-by-one until it becomes
7129 * too coarse to separate the current node from the last.
7131 mask = ~((1 << __ffs(start)) - 1);
7132 while (mask && last_end <= (start & (mask << 1)))
7135 /* accumulate all internode masks */
7139 /* convert mask to number of pages */
7140 return ~accl_mask + 1;
7143 /* Find the lowest pfn for a node */
7144 static unsigned long __init find_min_pfn_for_node(int nid)
7146 unsigned long min_pfn = ULONG_MAX;
7147 unsigned long start_pfn;
7150 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
7151 min_pfn = min(min_pfn, start_pfn);
7153 if (min_pfn == ULONG_MAX) {
7154 pr_warn("Could not find start_pfn for node %d\n", nid);
7162 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7164 * Return: the minimum PFN based on information provided via
7165 * memblock_set_node().
7167 unsigned long __init find_min_pfn_with_active_regions(void)
7169 return find_min_pfn_for_node(MAX_NUMNODES);
7173 * early_calculate_totalpages()
7174 * Sum pages in active regions for movable zone.
7175 * Populate N_MEMORY for calculating usable_nodes.
7177 static unsigned long __init early_calculate_totalpages(void)
7179 unsigned long totalpages = 0;
7180 unsigned long start_pfn, end_pfn;
7183 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7184 unsigned long pages = end_pfn - start_pfn;
7186 totalpages += pages;
7188 node_set_state(nid, N_MEMORY);
7194 * Find the PFN the Movable zone begins in each node. Kernel memory
7195 * is spread evenly between nodes as long as the nodes have enough
7196 * memory. When they don't, some nodes will have more kernelcore than
7199 static void __init find_zone_movable_pfns_for_nodes(void)
7202 unsigned long usable_startpfn;
7203 unsigned long kernelcore_node, kernelcore_remaining;
7204 /* save the state before borrow the nodemask */
7205 nodemask_t saved_node_state = node_states[N_MEMORY];
7206 unsigned long totalpages = early_calculate_totalpages();
7207 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7208 struct memblock_region *r;
7210 /* Need to find movable_zone earlier when movable_node is specified. */
7211 find_usable_zone_for_movable();
7214 * If movable_node is specified, ignore kernelcore and movablecore
7217 if (movable_node_is_enabled()) {
7218 for_each_memblock(memory, r) {
7219 if (!memblock_is_hotpluggable(r))
7222 nid = memblock_get_region_node(r);
7224 usable_startpfn = PFN_DOWN(r->base);
7225 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7226 min(usable_startpfn, zone_movable_pfn[nid]) :
7234 * If kernelcore=mirror is specified, ignore movablecore option
7236 if (mirrored_kernelcore) {
7237 bool mem_below_4gb_not_mirrored = false;
7239 for_each_memblock(memory, r) {
7240 if (memblock_is_mirror(r))
7243 nid = memblock_get_region_node(r);
7245 usable_startpfn = memblock_region_memory_base_pfn(r);
7247 if (usable_startpfn < 0x100000) {
7248 mem_below_4gb_not_mirrored = true;
7252 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7253 min(usable_startpfn, zone_movable_pfn[nid]) :
7257 if (mem_below_4gb_not_mirrored)
7258 pr_warn("This configuration results in unmirrored kernel memory.");
7264 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7265 * amount of necessary memory.
7267 if (required_kernelcore_percent)
7268 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7270 if (required_movablecore_percent)
7271 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7275 * If movablecore= was specified, calculate what size of
7276 * kernelcore that corresponds so that memory usable for
7277 * any allocation type is evenly spread. If both kernelcore
7278 * and movablecore are specified, then the value of kernelcore
7279 * will be used for required_kernelcore if it's greater than
7280 * what movablecore would have allowed.
7282 if (required_movablecore) {
7283 unsigned long corepages;
7286 * Round-up so that ZONE_MOVABLE is at least as large as what
7287 * was requested by the user
7289 required_movablecore =
7290 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7291 required_movablecore = min(totalpages, required_movablecore);
7292 corepages = totalpages - required_movablecore;
7294 required_kernelcore = max(required_kernelcore, corepages);
7298 * If kernelcore was not specified or kernelcore size is larger
7299 * than totalpages, there is no ZONE_MOVABLE.
7301 if (!required_kernelcore || required_kernelcore >= totalpages)
7304 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7305 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7308 /* Spread kernelcore memory as evenly as possible throughout nodes */
7309 kernelcore_node = required_kernelcore / usable_nodes;
7310 for_each_node_state(nid, N_MEMORY) {
7311 unsigned long start_pfn, end_pfn;
7314 * Recalculate kernelcore_node if the division per node
7315 * now exceeds what is necessary to satisfy the requested
7316 * amount of memory for the kernel
7318 if (required_kernelcore < kernelcore_node)
7319 kernelcore_node = required_kernelcore / usable_nodes;
7322 * As the map is walked, we track how much memory is usable
7323 * by the kernel using kernelcore_remaining. When it is
7324 * 0, the rest of the node is usable by ZONE_MOVABLE
7326 kernelcore_remaining = kernelcore_node;
7328 /* Go through each range of PFNs within this node */
7329 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7330 unsigned long size_pages;
7332 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7333 if (start_pfn >= end_pfn)
7336 /* Account for what is only usable for kernelcore */
7337 if (start_pfn < usable_startpfn) {
7338 unsigned long kernel_pages;
7339 kernel_pages = min(end_pfn, usable_startpfn)
7342 kernelcore_remaining -= min(kernel_pages,
7343 kernelcore_remaining);
7344 required_kernelcore -= min(kernel_pages,
7345 required_kernelcore);
7347 /* Continue if range is now fully accounted */
7348 if (end_pfn <= usable_startpfn) {
7351 * Push zone_movable_pfn to the end so
7352 * that if we have to rebalance
7353 * kernelcore across nodes, we will
7354 * not double account here
7356 zone_movable_pfn[nid] = end_pfn;
7359 start_pfn = usable_startpfn;
7363 * The usable PFN range for ZONE_MOVABLE is from
7364 * start_pfn->end_pfn. Calculate size_pages as the
7365 * number of pages used as kernelcore
7367 size_pages = end_pfn - start_pfn;
7368 if (size_pages > kernelcore_remaining)
7369 size_pages = kernelcore_remaining;
7370 zone_movable_pfn[nid] = start_pfn + size_pages;
7373 * Some kernelcore has been met, update counts and
7374 * break if the kernelcore for this node has been
7377 required_kernelcore -= min(required_kernelcore,
7379 kernelcore_remaining -= size_pages;
7380 if (!kernelcore_remaining)
7386 * If there is still required_kernelcore, we do another pass with one
7387 * less node in the count. This will push zone_movable_pfn[nid] further
7388 * along on the nodes that still have memory until kernelcore is
7392 if (usable_nodes && required_kernelcore > usable_nodes)
7396 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7397 for (nid = 0; nid < MAX_NUMNODES; nid++)
7398 zone_movable_pfn[nid] =
7399 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7402 /* restore the node_state */
7403 node_states[N_MEMORY] = saved_node_state;
7406 /* Any regular or high memory on that node ? */
7407 static void check_for_memory(pg_data_t *pgdat, int nid)
7409 enum zone_type zone_type;
7411 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7412 struct zone *zone = &pgdat->node_zones[zone_type];
7413 if (populated_zone(zone)) {
7414 if (IS_ENABLED(CONFIG_HIGHMEM))
7415 node_set_state(nid, N_HIGH_MEMORY);
7416 if (zone_type <= ZONE_NORMAL)
7417 node_set_state(nid, N_NORMAL_MEMORY);
7424 * free_area_init_nodes - Initialise all pg_data_t and zone data
7425 * @max_zone_pfn: an array of max PFNs for each zone
7427 * This will call free_area_init_node() for each active node in the system.
7428 * Using the page ranges provided by memblock_set_node(), the size of each
7429 * zone in each node and their holes is calculated. If the maximum PFN
7430 * between two adjacent zones match, it is assumed that the zone is empty.
7431 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7432 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7433 * starts where the previous one ended. For example, ZONE_DMA32 starts
7434 * at arch_max_dma_pfn.
7436 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
7438 unsigned long start_pfn, end_pfn;
7441 /* Record where the zone boundaries are */
7442 memset(arch_zone_lowest_possible_pfn, 0,
7443 sizeof(arch_zone_lowest_possible_pfn));
7444 memset(arch_zone_highest_possible_pfn, 0,
7445 sizeof(arch_zone_highest_possible_pfn));
7447 start_pfn = find_min_pfn_with_active_regions();
7449 for (i = 0; i < MAX_NR_ZONES; i++) {
7450 if (i == ZONE_MOVABLE)
7453 end_pfn = max(max_zone_pfn[i], start_pfn);
7454 arch_zone_lowest_possible_pfn[i] = start_pfn;
7455 arch_zone_highest_possible_pfn[i] = end_pfn;
7457 start_pfn = end_pfn;
7460 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7461 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7462 find_zone_movable_pfns_for_nodes();
7464 /* Print out the zone ranges */
7465 pr_info("Zone ranges:\n");
7466 for (i = 0; i < MAX_NR_ZONES; i++) {
7467 if (i == ZONE_MOVABLE)
7469 pr_info(" %-8s ", zone_names[i]);
7470 if (arch_zone_lowest_possible_pfn[i] ==
7471 arch_zone_highest_possible_pfn[i])
7474 pr_cont("[mem %#018Lx-%#018Lx]\n",
7475 (u64)arch_zone_lowest_possible_pfn[i]
7477 ((u64)arch_zone_highest_possible_pfn[i]
7478 << PAGE_SHIFT) - 1);
7481 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7482 pr_info("Movable zone start for each node\n");
7483 for (i = 0; i < MAX_NUMNODES; i++) {
7484 if (zone_movable_pfn[i])
7485 pr_info(" Node %d: %#018Lx\n", i,
7486 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7490 * Print out the early node map, and initialize the
7491 * subsection-map relative to active online memory ranges to
7492 * enable future "sub-section" extensions of the memory map.
7494 pr_info("Early memory node ranges\n");
7495 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7496 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7497 (u64)start_pfn << PAGE_SHIFT,
7498 ((u64)end_pfn << PAGE_SHIFT) - 1);
7499 subsection_map_init(start_pfn, end_pfn - start_pfn);
7502 /* Initialise every node */
7503 mminit_verify_pageflags_layout();
7504 setup_nr_node_ids();
7505 init_unavailable_mem();
7506 for_each_online_node(nid) {
7507 pg_data_t *pgdat = NODE_DATA(nid);
7508 free_area_init_node(nid, NULL,
7509 find_min_pfn_for_node(nid), NULL);
7511 /* Any memory on that node */
7512 if (pgdat->node_present_pages)
7513 node_set_state(nid, N_MEMORY);
7514 check_for_memory(pgdat, nid);
7518 static int __init cmdline_parse_core(char *p, unsigned long *core,
7519 unsigned long *percent)
7521 unsigned long long coremem;
7527 /* Value may be a percentage of total memory, otherwise bytes */
7528 coremem = simple_strtoull(p, &endptr, 0);
7529 if (*endptr == '%') {
7530 /* Paranoid check for percent values greater than 100 */
7531 WARN_ON(coremem > 100);
7535 coremem = memparse(p, &p);
7536 /* Paranoid check that UL is enough for the coremem value */
7537 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7539 *core = coremem >> PAGE_SHIFT;
7546 * kernelcore=size sets the amount of memory for use for allocations that
7547 * cannot be reclaimed or migrated.
7549 static int __init cmdline_parse_kernelcore(char *p)
7551 /* parse kernelcore=mirror */
7552 if (parse_option_str(p, "mirror")) {
7553 mirrored_kernelcore = true;
7557 return cmdline_parse_core(p, &required_kernelcore,
7558 &required_kernelcore_percent);
7562 * movablecore=size sets the amount of memory for use for allocations that
7563 * can be reclaimed or migrated.
7565 static int __init cmdline_parse_movablecore(char *p)
7567 return cmdline_parse_core(p, &required_movablecore,
7568 &required_movablecore_percent);
7571 early_param("kernelcore", cmdline_parse_kernelcore);
7572 early_param("movablecore", cmdline_parse_movablecore);
7574 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
7576 void adjust_managed_page_count(struct page *page, long count)
7578 atomic_long_add(count, &page_zone(page)->managed_pages);
7579 totalram_pages_add(count);
7580 #ifdef CONFIG_HIGHMEM
7581 if (PageHighMem(page))
7582 totalhigh_pages_add(count);
7585 EXPORT_SYMBOL(adjust_managed_page_count);
7587 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7590 unsigned long pages = 0;
7592 start = (void *)PAGE_ALIGN((unsigned long)start);
7593 end = (void *)((unsigned long)end & PAGE_MASK);
7594 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7595 struct page *page = virt_to_page(pos);
7596 void *direct_map_addr;
7599 * 'direct_map_addr' might be different from 'pos'
7600 * because some architectures' virt_to_page()
7601 * work with aliases. Getting the direct map
7602 * address ensures that we get a _writeable_
7603 * alias for the memset().
7605 direct_map_addr = page_address(page);
7606 if ((unsigned int)poison <= 0xFF)
7607 memset(direct_map_addr, poison, PAGE_SIZE);
7609 free_reserved_page(page);
7613 pr_info("Freeing %s memory: %ldK\n",
7614 s, pages << (PAGE_SHIFT - 10));
7619 #ifdef CONFIG_HIGHMEM
7620 void free_highmem_page(struct page *page)
7622 __free_reserved_page(page);
7623 totalram_pages_inc();
7624 atomic_long_inc(&page_zone(page)->managed_pages);
7625 totalhigh_pages_inc();
7630 void __init mem_init_print_info(const char *str)
7632 unsigned long physpages, codesize, datasize, rosize, bss_size;
7633 unsigned long init_code_size, init_data_size;
7635 physpages = get_num_physpages();
7636 codesize = _etext - _stext;
7637 datasize = _edata - _sdata;
7638 rosize = __end_rodata - __start_rodata;
7639 bss_size = __bss_stop - __bss_start;
7640 init_data_size = __init_end - __init_begin;
7641 init_code_size = _einittext - _sinittext;
7644 * Detect special cases and adjust section sizes accordingly:
7645 * 1) .init.* may be embedded into .data sections
7646 * 2) .init.text.* may be out of [__init_begin, __init_end],
7647 * please refer to arch/tile/kernel/vmlinux.lds.S.
7648 * 3) .rodata.* may be embedded into .text or .data sections.
7650 #define adj_init_size(start, end, size, pos, adj) \
7652 if (start <= pos && pos < end && size > adj) \
7656 adj_init_size(__init_begin, __init_end, init_data_size,
7657 _sinittext, init_code_size);
7658 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7659 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7660 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7661 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7663 #undef adj_init_size
7665 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7666 #ifdef CONFIG_HIGHMEM
7670 nr_free_pages() << (PAGE_SHIFT - 10),
7671 physpages << (PAGE_SHIFT - 10),
7672 codesize >> 10, datasize >> 10, rosize >> 10,
7673 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7674 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7675 totalcma_pages << (PAGE_SHIFT - 10),
7676 #ifdef CONFIG_HIGHMEM
7677 totalhigh_pages() << (PAGE_SHIFT - 10),
7679 str ? ", " : "", str ? str : "");
7683 * set_dma_reserve - set the specified number of pages reserved in the first zone
7684 * @new_dma_reserve: The number of pages to mark reserved
7686 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7687 * In the DMA zone, a significant percentage may be consumed by kernel image
7688 * and other unfreeable allocations which can skew the watermarks badly. This
7689 * function may optionally be used to account for unfreeable pages in the
7690 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7691 * smaller per-cpu batchsize.
7693 void __init set_dma_reserve(unsigned long new_dma_reserve)
7695 dma_reserve = new_dma_reserve;
7698 void __init free_area_init(unsigned long *zones_size)
7700 init_unavailable_mem();
7701 free_area_init_node(0, zones_size,
7702 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
7705 static int page_alloc_cpu_dead(unsigned int cpu)
7708 lru_add_drain_cpu(cpu);
7712 * Spill the event counters of the dead processor
7713 * into the current processors event counters.
7714 * This artificially elevates the count of the current
7717 vm_events_fold_cpu(cpu);
7720 * Zero the differential counters of the dead processor
7721 * so that the vm statistics are consistent.
7723 * This is only okay since the processor is dead and cannot
7724 * race with what we are doing.
7726 cpu_vm_stats_fold(cpu);
7731 int hashdist = HASHDIST_DEFAULT;
7733 static int __init set_hashdist(char *str)
7737 hashdist = simple_strtoul(str, &str, 0);
7740 __setup("hashdist=", set_hashdist);
7743 void __init page_alloc_init(void)
7748 if (num_node_state(N_MEMORY) == 1)
7752 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7753 "mm/page_alloc:dead", NULL,
7754 page_alloc_cpu_dead);
7759 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7760 * or min_free_kbytes changes.
7762 static void calculate_totalreserve_pages(void)
7764 struct pglist_data *pgdat;
7765 unsigned long reserve_pages = 0;
7766 enum zone_type i, j;
7768 for_each_online_pgdat(pgdat) {
7770 pgdat->totalreserve_pages = 0;
7772 for (i = 0; i < MAX_NR_ZONES; i++) {
7773 struct zone *zone = pgdat->node_zones + i;
7775 unsigned long managed_pages = zone_managed_pages(zone);
7777 /* Find valid and maximum lowmem_reserve in the zone */
7778 for (j = i; j < MAX_NR_ZONES; j++) {
7779 if (zone->lowmem_reserve[j] > max)
7780 max = zone->lowmem_reserve[j];
7783 /* we treat the high watermark as reserved pages. */
7784 max += high_wmark_pages(zone);
7786 if (max > managed_pages)
7787 max = managed_pages;
7789 pgdat->totalreserve_pages += max;
7791 reserve_pages += max;
7794 totalreserve_pages = reserve_pages;
7798 * setup_per_zone_lowmem_reserve - called whenever
7799 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7800 * has a correct pages reserved value, so an adequate number of
7801 * pages are left in the zone after a successful __alloc_pages().
7803 static void setup_per_zone_lowmem_reserve(void)
7805 struct pglist_data *pgdat;
7806 enum zone_type j, idx;
7808 for_each_online_pgdat(pgdat) {
7809 for (j = 0; j < MAX_NR_ZONES; j++) {
7810 struct zone *zone = pgdat->node_zones + j;
7811 unsigned long managed_pages = zone_managed_pages(zone);
7813 zone->lowmem_reserve[j] = 0;
7817 struct zone *lower_zone;
7820 lower_zone = pgdat->node_zones + idx;
7822 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7823 sysctl_lowmem_reserve_ratio[idx] = 0;
7824 lower_zone->lowmem_reserve[j] = 0;
7826 lower_zone->lowmem_reserve[j] =
7827 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7829 managed_pages += zone_managed_pages(lower_zone);
7834 /* update totalreserve_pages */
7835 calculate_totalreserve_pages();
7838 static void __setup_per_zone_wmarks(void)
7840 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7841 unsigned long lowmem_pages = 0;
7843 unsigned long flags;
7845 /* Calculate total number of !ZONE_HIGHMEM pages */
7846 for_each_zone(zone) {
7847 if (!is_highmem(zone))
7848 lowmem_pages += zone_managed_pages(zone);
7851 for_each_zone(zone) {
7854 spin_lock_irqsave(&zone->lock, flags);
7855 tmp = (u64)pages_min * zone_managed_pages(zone);
7856 do_div(tmp, lowmem_pages);
7857 if (is_highmem(zone)) {
7859 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7860 * need highmem pages, so cap pages_min to a small
7863 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7864 * deltas control async page reclaim, and so should
7865 * not be capped for highmem.
7867 unsigned long min_pages;
7869 min_pages = zone_managed_pages(zone) / 1024;
7870 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7871 zone->_watermark[WMARK_MIN] = min_pages;
7874 * If it's a lowmem zone, reserve a number of pages
7875 * proportionate to the zone's size.
7877 zone->_watermark[WMARK_MIN] = tmp;
7881 * Set the kswapd watermarks distance according to the
7882 * scale factor in proportion to available memory, but
7883 * ensure a minimum size on small systems.
7885 tmp = max_t(u64, tmp >> 2,
7886 mult_frac(zone_managed_pages(zone),
7887 watermark_scale_factor, 10000));
7889 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7890 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7891 zone->watermark_boost = 0;
7893 spin_unlock_irqrestore(&zone->lock, flags);
7896 /* update totalreserve_pages */
7897 calculate_totalreserve_pages();
7901 * setup_per_zone_wmarks - called when min_free_kbytes changes
7902 * or when memory is hot-{added|removed}
7904 * Ensures that the watermark[min,low,high] values for each zone are set
7905 * correctly with respect to min_free_kbytes.
7907 void setup_per_zone_wmarks(void)
7909 static DEFINE_SPINLOCK(lock);
7912 __setup_per_zone_wmarks();
7917 * Initialise min_free_kbytes.
7919 * For small machines we want it small (128k min). For large machines
7920 * we want it large (64MB max). But it is not linear, because network
7921 * bandwidth does not increase linearly with machine size. We use
7923 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7924 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7940 int __meminit init_per_zone_wmark_min(void)
7942 unsigned long lowmem_kbytes;
7943 int new_min_free_kbytes;
7945 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7946 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7948 if (new_min_free_kbytes > user_min_free_kbytes) {
7949 min_free_kbytes = new_min_free_kbytes;
7950 if (min_free_kbytes < 128)
7951 min_free_kbytes = 128;
7952 if (min_free_kbytes > 262144)
7953 min_free_kbytes = 262144;
7955 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7956 new_min_free_kbytes, user_min_free_kbytes);
7958 setup_per_zone_wmarks();
7959 refresh_zone_stat_thresholds();
7960 setup_per_zone_lowmem_reserve();
7963 setup_min_unmapped_ratio();
7964 setup_min_slab_ratio();
7969 core_initcall(init_per_zone_wmark_min)
7972 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7973 * that we can call two helper functions whenever min_free_kbytes
7976 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7977 void __user *buffer, size_t *length, loff_t *ppos)
7981 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7986 user_min_free_kbytes = min_free_kbytes;
7987 setup_per_zone_wmarks();
7992 int watermark_boost_factor_sysctl_handler(struct ctl_table *table, int write,
7993 void __user *buffer, size_t *length, loff_t *ppos)
7997 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8004 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8005 void __user *buffer, size_t *length, loff_t *ppos)
8009 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8014 setup_per_zone_wmarks();
8020 static void setup_min_unmapped_ratio(void)
8025 for_each_online_pgdat(pgdat)
8026 pgdat->min_unmapped_pages = 0;
8029 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8030 sysctl_min_unmapped_ratio) / 100;
8034 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8035 void __user *buffer, size_t *length, loff_t *ppos)
8039 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8043 setup_min_unmapped_ratio();
8048 static void setup_min_slab_ratio(void)
8053 for_each_online_pgdat(pgdat)
8054 pgdat->min_slab_pages = 0;
8057 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8058 sysctl_min_slab_ratio) / 100;
8061 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8062 void __user *buffer, size_t *length, loff_t *ppos)
8066 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8070 setup_min_slab_ratio();
8077 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8078 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8079 * whenever sysctl_lowmem_reserve_ratio changes.
8081 * The reserve ratio obviously has absolutely no relation with the
8082 * minimum watermarks. The lowmem reserve ratio can only make sense
8083 * if in function of the boot time zone sizes.
8085 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8086 void __user *buffer, size_t *length, loff_t *ppos)
8088 proc_dointvec_minmax(table, write, buffer, length, ppos);
8089 setup_per_zone_lowmem_reserve();
8093 static void __zone_pcp_update(struct zone *zone)
8097 for_each_possible_cpu(cpu)
8098 pageset_set_high_and_batch(zone,
8099 per_cpu_ptr(zone->pageset, cpu));
8103 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8104 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8105 * pagelist can have before it gets flushed back to buddy allocator.
8107 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8108 void __user *buffer, size_t *length, loff_t *ppos)
8111 int old_percpu_pagelist_fraction;
8114 mutex_lock(&pcp_batch_high_lock);
8115 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8117 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8118 if (!write || ret < 0)
8121 /* Sanity checking to avoid pcp imbalance */
8122 if (percpu_pagelist_fraction &&
8123 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8124 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8130 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8133 for_each_populated_zone(zone)
8134 __zone_pcp_update(zone);
8136 mutex_unlock(&pcp_batch_high_lock);
8140 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8142 * Returns the number of pages that arch has reserved but
8143 * is not known to alloc_large_system_hash().
8145 static unsigned long __init arch_reserved_kernel_pages(void)
8152 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8153 * machines. As memory size is increased the scale is also increased but at
8154 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8155 * quadruples the scale is increased by one, which means the size of hash table
8156 * only doubles, instead of quadrupling as well.
8157 * Because 32-bit systems cannot have large physical memory, where this scaling
8158 * makes sense, it is disabled on such platforms.
8160 #if __BITS_PER_LONG > 32
8161 #define ADAPT_SCALE_BASE (64ul << 30)
8162 #define ADAPT_SCALE_SHIFT 2
8163 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8167 * allocate a large system hash table from bootmem
8168 * - it is assumed that the hash table must contain an exact power-of-2
8169 * quantity of entries
8170 * - limit is the number of hash buckets, not the total allocation size
8172 void *__init alloc_large_system_hash(const char *tablename,
8173 unsigned long bucketsize,
8174 unsigned long numentries,
8177 unsigned int *_hash_shift,
8178 unsigned int *_hash_mask,
8179 unsigned long low_limit,
8180 unsigned long high_limit)
8182 unsigned long long max = high_limit;
8183 unsigned long log2qty, size;
8188 /* allow the kernel cmdline to have a say */
8190 /* round applicable memory size up to nearest megabyte */
8191 numentries = nr_kernel_pages;
8192 numentries -= arch_reserved_kernel_pages();
8194 /* It isn't necessary when PAGE_SIZE >= 1MB */
8195 if (PAGE_SHIFT < 20)
8196 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8198 #if __BITS_PER_LONG > 32
8200 unsigned long adapt;
8202 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8203 adapt <<= ADAPT_SCALE_SHIFT)
8208 /* limit to 1 bucket per 2^scale bytes of low memory */
8209 if (scale > PAGE_SHIFT)
8210 numentries >>= (scale - PAGE_SHIFT);
8212 numentries <<= (PAGE_SHIFT - scale);
8214 /* Make sure we've got at least a 0-order allocation.. */
8215 if (unlikely(flags & HASH_SMALL)) {
8216 /* Makes no sense without HASH_EARLY */
8217 WARN_ON(!(flags & HASH_EARLY));
8218 if (!(numentries >> *_hash_shift)) {
8219 numentries = 1UL << *_hash_shift;
8220 BUG_ON(!numentries);
8222 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8223 numentries = PAGE_SIZE / bucketsize;
8225 numentries = roundup_pow_of_two(numentries);
8227 /* limit allocation size to 1/16 total memory by default */
8229 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8230 do_div(max, bucketsize);
8232 max = min(max, 0x80000000ULL);
8234 if (numentries < low_limit)
8235 numentries = low_limit;
8236 if (numentries > max)
8239 log2qty = ilog2(numentries);
8241 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8244 size = bucketsize << log2qty;
8245 if (flags & HASH_EARLY) {
8246 if (flags & HASH_ZERO)
8247 table = memblock_alloc(size, SMP_CACHE_BYTES);
8249 table = memblock_alloc_raw(size,
8251 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8252 table = __vmalloc(size, gfp_flags);
8256 * If bucketsize is not a power-of-two, we may free
8257 * some pages at the end of hash table which
8258 * alloc_pages_exact() automatically does
8260 table = alloc_pages_exact(size, gfp_flags);
8261 kmemleak_alloc(table, size, 1, gfp_flags);
8263 } while (!table && size > PAGE_SIZE && --log2qty);
8266 panic("Failed to allocate %s hash table\n", tablename);
8268 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8269 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8270 virt ? "vmalloc" : "linear");
8273 *_hash_shift = log2qty;
8275 *_hash_mask = (1 << log2qty) - 1;
8281 * This function checks whether pageblock includes unmovable pages or not.
8283 * PageLRU check without isolation or lru_lock could race so that
8284 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8285 * check without lock_page also may miss some movable non-lru pages at
8286 * race condition. So you can't expect this function should be exact.
8288 * Returns a page without holding a reference. If the caller wants to
8289 * dereference that page (e.g., dumping), it has to make sure that that it
8290 * cannot get removed (e.g., via memory unplug) concurrently.
8293 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8294 int migratetype, int flags)
8296 unsigned long iter = 0;
8297 unsigned long pfn = page_to_pfn(page);
8300 * TODO we could make this much more efficient by not checking every
8301 * page in the range if we know all of them are in MOVABLE_ZONE and
8302 * that the movable zone guarantees that pages are migratable but
8303 * the later is not the case right now unfortunatelly. E.g. movablecore
8304 * can still lead to having bootmem allocations in zone_movable.
8307 if (is_migrate_cma_page(page)) {
8309 * CMA allocations (alloc_contig_range) really need to mark
8310 * isolate CMA pageblocks even when they are not movable in fact
8311 * so consider them movable here.
8313 if (is_migrate_cma(migratetype))
8319 for (; iter < pageblock_nr_pages; iter++) {
8320 if (!pfn_valid_within(pfn + iter))
8323 page = pfn_to_page(pfn + iter);
8325 if (PageReserved(page))
8329 * If the zone is movable and we have ruled out all reserved
8330 * pages then it should be reasonably safe to assume the rest
8333 if (zone_idx(zone) == ZONE_MOVABLE)
8337 * Hugepages are not in LRU lists, but they're movable.
8338 * THPs are on the LRU, but need to be counted as #small pages.
8339 * We need not scan over tail pages because we don't
8340 * handle each tail page individually in migration.
8342 if (PageHuge(page) || PageTransCompound(page)) {
8343 struct page *head = compound_head(page);
8344 unsigned int skip_pages;
8346 if (PageHuge(page)) {
8347 if (!hugepage_migration_supported(page_hstate(head)))
8349 } else if (!PageLRU(head) && !__PageMovable(head)) {
8353 skip_pages = compound_nr(head) - (page - head);
8354 iter += skip_pages - 1;
8359 * We can't use page_count without pin a page
8360 * because another CPU can free compound page.
8361 * This check already skips compound tails of THP
8362 * because their page->_refcount is zero at all time.
8364 if (!page_ref_count(page)) {
8365 if (PageBuddy(page))
8366 iter += (1 << page_order(page)) - 1;
8371 * The HWPoisoned page may be not in buddy system, and
8372 * page_count() is not 0.
8374 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8377 if (__PageMovable(page) || PageLRU(page))
8381 * If there are RECLAIMABLE pages, we need to check
8382 * it. But now, memory offline itself doesn't call
8383 * shrink_node_slabs() and it still to be fixed.
8386 * If the page is not RAM, page_count()should be 0.
8387 * we don't need more check. This is an _used_ not-movable page.
8389 * The problematic thing here is PG_reserved pages. PG_reserved
8390 * is set to both of a memory hole page and a _used_ kernel
8398 #ifdef CONFIG_CONTIG_ALLOC
8399 static unsigned long pfn_max_align_down(unsigned long pfn)
8401 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8402 pageblock_nr_pages) - 1);
8405 static unsigned long pfn_max_align_up(unsigned long pfn)
8407 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8408 pageblock_nr_pages));
8411 /* [start, end) must belong to a single zone. */
8412 static int __alloc_contig_migrate_range(struct compact_control *cc,
8413 unsigned long start, unsigned long end)
8415 /* This function is based on compact_zone() from compaction.c. */
8416 unsigned long nr_reclaimed;
8417 unsigned long pfn = start;
8418 unsigned int tries = 0;
8423 while (pfn < end || !list_empty(&cc->migratepages)) {
8424 if (fatal_signal_pending(current)) {
8429 if (list_empty(&cc->migratepages)) {
8430 cc->nr_migratepages = 0;
8431 pfn = isolate_migratepages_range(cc, pfn, end);
8437 } else if (++tries == 5) {
8438 ret = ret < 0 ? ret : -EBUSY;
8442 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8444 cc->nr_migratepages -= nr_reclaimed;
8446 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
8447 NULL, 0, cc->mode, MR_CONTIG_RANGE);
8450 putback_movable_pages(&cc->migratepages);
8457 * alloc_contig_range() -- tries to allocate given range of pages
8458 * @start: start PFN to allocate
8459 * @end: one-past-the-last PFN to allocate
8460 * @migratetype: migratetype of the underlaying pageblocks (either
8461 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8462 * in range must have the same migratetype and it must
8463 * be either of the two.
8464 * @gfp_mask: GFP mask to use during compaction
8466 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8467 * aligned. The PFN range must belong to a single zone.
8469 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8470 * pageblocks in the range. Once isolated, the pageblocks should not
8471 * be modified by others.
8473 * Return: zero on success or negative error code. On success all
8474 * pages which PFN is in [start, end) are allocated for the caller and
8475 * need to be freed with free_contig_range().
8477 int alloc_contig_range(unsigned long start, unsigned long end,
8478 unsigned migratetype, gfp_t gfp_mask)
8480 unsigned long outer_start, outer_end;
8484 struct compact_control cc = {
8485 .nr_migratepages = 0,
8487 .zone = page_zone(pfn_to_page(start)),
8488 .mode = MIGRATE_SYNC,
8489 .ignore_skip_hint = true,
8490 .no_set_skip_hint = true,
8491 .gfp_mask = current_gfp_context(gfp_mask),
8492 .alloc_contig = true,
8494 INIT_LIST_HEAD(&cc.migratepages);
8497 * What we do here is we mark all pageblocks in range as
8498 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8499 * have different sizes, and due to the way page allocator
8500 * work, we align the range to biggest of the two pages so
8501 * that page allocator won't try to merge buddies from
8502 * different pageblocks and change MIGRATE_ISOLATE to some
8503 * other migration type.
8505 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8506 * migrate the pages from an unaligned range (ie. pages that
8507 * we are interested in). This will put all the pages in
8508 * range back to page allocator as MIGRATE_ISOLATE.
8510 * When this is done, we take the pages in range from page
8511 * allocator removing them from the buddy system. This way
8512 * page allocator will never consider using them.
8514 * This lets us mark the pageblocks back as
8515 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8516 * aligned range but not in the unaligned, original range are
8517 * put back to page allocator so that buddy can use them.
8520 ret = start_isolate_page_range(pfn_max_align_down(start),
8521 pfn_max_align_up(end), migratetype, 0);
8526 * In case of -EBUSY, we'd like to know which page causes problem.
8527 * So, just fall through. test_pages_isolated() has a tracepoint
8528 * which will report the busy page.
8530 * It is possible that busy pages could become available before
8531 * the call to test_pages_isolated, and the range will actually be
8532 * allocated. So, if we fall through be sure to clear ret so that
8533 * -EBUSY is not accidentally used or returned to caller.
8535 ret = __alloc_contig_migrate_range(&cc, start, end);
8536 if (ret && ret != -EBUSY)
8541 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8542 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8543 * more, all pages in [start, end) are free in page allocator.
8544 * What we are going to do is to allocate all pages from
8545 * [start, end) (that is remove them from page allocator).
8547 * The only problem is that pages at the beginning and at the
8548 * end of interesting range may be not aligned with pages that
8549 * page allocator holds, ie. they can be part of higher order
8550 * pages. Because of this, we reserve the bigger range and
8551 * once this is done free the pages we are not interested in.
8553 * We don't have to hold zone->lock here because the pages are
8554 * isolated thus they won't get removed from buddy.
8557 lru_add_drain_all();
8560 outer_start = start;
8561 while (!PageBuddy(pfn_to_page(outer_start))) {
8562 if (++order >= MAX_ORDER) {
8563 outer_start = start;
8566 outer_start &= ~0UL << order;
8569 if (outer_start != start) {
8570 order = page_order(pfn_to_page(outer_start));
8573 * outer_start page could be small order buddy page and
8574 * it doesn't include start page. Adjust outer_start
8575 * in this case to report failed page properly
8576 * on tracepoint in test_pages_isolated()
8578 if (outer_start + (1UL << order) <= start)
8579 outer_start = start;
8582 /* Make sure the range is really isolated. */
8583 if (test_pages_isolated(outer_start, end, 0)) {
8584 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8585 __func__, outer_start, end);
8590 /* Grab isolated pages from freelists. */
8591 outer_end = isolate_freepages_range(&cc, outer_start, end);
8597 /* Free head and tail (if any) */
8598 if (start != outer_start)
8599 free_contig_range(outer_start, start - outer_start);
8600 if (end != outer_end)
8601 free_contig_range(end, outer_end - end);
8604 undo_isolate_page_range(pfn_max_align_down(start),
8605 pfn_max_align_up(end), migratetype);
8609 static int __alloc_contig_pages(unsigned long start_pfn,
8610 unsigned long nr_pages, gfp_t gfp_mask)
8612 unsigned long end_pfn = start_pfn + nr_pages;
8614 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8618 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8619 unsigned long nr_pages)
8621 unsigned long i, end_pfn = start_pfn + nr_pages;
8624 for (i = start_pfn; i < end_pfn; i++) {
8625 page = pfn_to_online_page(i);
8629 if (page_zone(page) != z)
8632 if (PageReserved(page))
8635 if (page_count(page) > 0)
8644 static bool zone_spans_last_pfn(const struct zone *zone,
8645 unsigned long start_pfn, unsigned long nr_pages)
8647 unsigned long last_pfn = start_pfn + nr_pages - 1;
8649 return zone_spans_pfn(zone, last_pfn);
8653 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8654 * @nr_pages: Number of contiguous pages to allocate
8655 * @gfp_mask: GFP mask to limit search and used during compaction
8657 * @nodemask: Mask for other possible nodes
8659 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8660 * on an applicable zonelist to find a contiguous pfn range which can then be
8661 * tried for allocation with alloc_contig_range(). This routine is intended
8662 * for allocation requests which can not be fulfilled with the buddy allocator.
8664 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8665 * power of two then the alignment is guaranteed to be to the given nr_pages
8666 * (e.g. 1GB request would be aligned to 1GB).
8668 * Allocated pages can be freed with free_contig_range() or by manually calling
8669 * __free_page() on each allocated page.
8671 * Return: pointer to contiguous pages on success, or NULL if not successful.
8673 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8674 int nid, nodemask_t *nodemask)
8676 unsigned long ret, pfn, flags;
8677 struct zonelist *zonelist;
8681 zonelist = node_zonelist(nid, gfp_mask);
8682 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8683 gfp_zone(gfp_mask), nodemask) {
8684 spin_lock_irqsave(&zone->lock, flags);
8686 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8687 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8688 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8690 * We release the zone lock here because
8691 * alloc_contig_range() will also lock the zone
8692 * at some point. If there's an allocation
8693 * spinning on this lock, it may win the race
8694 * and cause alloc_contig_range() to fail...
8696 spin_unlock_irqrestore(&zone->lock, flags);
8697 ret = __alloc_contig_pages(pfn, nr_pages,
8700 return pfn_to_page(pfn);
8701 spin_lock_irqsave(&zone->lock, flags);
8705 spin_unlock_irqrestore(&zone->lock, flags);
8709 #endif /* CONFIG_CONTIG_ALLOC */
8711 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8713 unsigned int count = 0;
8715 for (; nr_pages--; pfn++) {
8716 struct page *page = pfn_to_page(pfn);
8718 count += page_count(page) != 1;
8721 WARN(count != 0, "%d pages are still in use!\n", count);
8725 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8726 * page high values need to be recalulated.
8728 void __meminit zone_pcp_update(struct zone *zone)
8730 mutex_lock(&pcp_batch_high_lock);
8731 __zone_pcp_update(zone);
8732 mutex_unlock(&pcp_batch_high_lock);
8735 void zone_pcp_reset(struct zone *zone)
8737 unsigned long flags;
8739 struct per_cpu_pageset *pset;
8741 /* avoid races with drain_pages() */
8742 local_irq_save(flags);
8743 if (zone->pageset != &boot_pageset) {
8744 for_each_online_cpu(cpu) {
8745 pset = per_cpu_ptr(zone->pageset, cpu);
8746 drain_zonestat(zone, pset);
8748 free_percpu(zone->pageset);
8749 zone->pageset = &boot_pageset;
8751 local_irq_restore(flags);
8754 #ifdef CONFIG_MEMORY_HOTREMOVE
8756 * All pages in the range must be in a single zone and isolated
8757 * before calling this.
8760 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8766 unsigned long flags;
8767 unsigned long offlined_pages = 0;
8769 /* find the first valid pfn */
8770 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8774 return offlined_pages;
8776 offline_mem_sections(pfn, end_pfn);
8777 zone = page_zone(pfn_to_page(pfn));
8778 spin_lock_irqsave(&zone->lock, flags);
8780 while (pfn < end_pfn) {
8781 if (!pfn_valid(pfn)) {
8785 page = pfn_to_page(pfn);
8787 * The HWPoisoned page may be not in buddy system, and
8788 * page_count() is not 0.
8790 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8796 BUG_ON(page_count(page));
8797 BUG_ON(!PageBuddy(page));
8798 order = page_order(page);
8799 offlined_pages += 1 << order;
8800 del_page_from_free_list(page, zone, order);
8801 pfn += (1 << order);
8803 spin_unlock_irqrestore(&zone->lock, flags);
8805 return offlined_pages;
8809 bool is_free_buddy_page(struct page *page)
8811 struct zone *zone = page_zone(page);
8812 unsigned long pfn = page_to_pfn(page);
8813 unsigned long flags;
8816 spin_lock_irqsave(&zone->lock, flags);
8817 for (order = 0; order < MAX_ORDER; order++) {
8818 struct page *page_head = page - (pfn & ((1 << order) - 1));
8820 if (PageBuddy(page_head) && page_order(page_head) >= order)
8823 spin_unlock_irqrestore(&zone->lock, flags);
8825 return order < MAX_ORDER;
8828 #ifdef CONFIG_MEMORY_FAILURE
8830 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8831 * test is performed under the zone lock to prevent a race against page
8834 bool set_hwpoison_free_buddy_page(struct page *page)
8836 struct zone *zone = page_zone(page);
8837 unsigned long pfn = page_to_pfn(page);
8838 unsigned long flags;
8840 bool hwpoisoned = false;
8842 spin_lock_irqsave(&zone->lock, flags);
8843 for (order = 0; order < MAX_ORDER; order++) {
8844 struct page *page_head = page - (pfn & ((1 << order) - 1));
8846 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8847 if (!TestSetPageHWPoison(page))
8852 spin_unlock_irqrestore(&zone->lock, flags);