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>
78 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
79 static DEFINE_MUTEX(pcp_batch_high_lock);
80 #define MIN_PERCPU_PAGELIST_FRACTION (8)
82 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
83 DEFINE_PER_CPU(int, numa_node);
84 EXPORT_PER_CPU_SYMBOL(numa_node);
87 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
89 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
91 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
92 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
93 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
94 * defined in <linux/topology.h>.
96 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
97 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
98 int _node_numa_mem_[MAX_NUMNODES];
101 /* work_structs for global per-cpu drains */
104 struct work_struct work;
106 DEFINE_MUTEX(pcpu_drain_mutex);
107 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);
694 #ifdef CONFIG_DEBUG_PAGEALLOC
695 unsigned int _debug_guardpage_minorder;
697 bool _debug_pagealloc_enabled_early __read_mostly
698 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
699 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
700 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
701 EXPORT_SYMBOL(_debug_pagealloc_enabled);
703 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
705 static int __init early_debug_pagealloc(char *buf)
707 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
709 early_param("debug_pagealloc", early_debug_pagealloc);
711 void init_debug_pagealloc(void)
713 if (!debug_pagealloc_enabled())
716 static_branch_enable(&_debug_pagealloc_enabled);
718 if (!debug_guardpage_minorder())
721 static_branch_enable(&_debug_guardpage_enabled);
724 static int __init debug_guardpage_minorder_setup(char *buf)
728 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
729 pr_err("Bad debug_guardpage_minorder value\n");
732 _debug_guardpage_minorder = res;
733 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
736 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
738 static inline bool set_page_guard(struct zone *zone, struct page *page,
739 unsigned int order, int migratetype)
741 if (!debug_guardpage_enabled())
744 if (order >= debug_guardpage_minorder())
747 __SetPageGuard(page);
748 INIT_LIST_HEAD(&page->lru);
749 set_page_private(page, order);
750 /* Guard pages are not available for any usage */
751 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
756 static inline void clear_page_guard(struct zone *zone, struct page *page,
757 unsigned int order, int migratetype)
759 if (!debug_guardpage_enabled())
762 __ClearPageGuard(page);
764 set_page_private(page, 0);
765 if (!is_migrate_isolate(migratetype))
766 __mod_zone_freepage_state(zone, (1 << order), migratetype);
769 static inline bool set_page_guard(struct zone *zone, struct page *page,
770 unsigned int order, int migratetype) { return false; }
771 static inline void clear_page_guard(struct zone *zone, struct page *page,
772 unsigned int order, int migratetype) {}
775 static inline void set_page_order(struct page *page, unsigned int order)
777 set_page_private(page, order);
778 __SetPageBuddy(page);
782 * This function checks whether a page is free && is the buddy
783 * we can coalesce a page and its buddy if
784 * (a) the buddy is not in a hole (check before calling!) &&
785 * (b) the buddy is in the buddy system &&
786 * (c) a page and its buddy have the same order &&
787 * (d) a page and its buddy are in the same zone.
789 * For recording whether a page is in the buddy system, we set PageBuddy.
790 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
792 * For recording page's order, we use page_private(page).
794 static inline int page_is_buddy(struct page *page, struct page *buddy,
797 if (page_is_guard(buddy) && page_order(buddy) == order) {
798 if (page_zone_id(page) != page_zone_id(buddy))
801 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
806 if (PageBuddy(buddy) && page_order(buddy) == order) {
808 * zone check is done late to avoid uselessly
809 * calculating zone/node ids for pages that could
812 if (page_zone_id(page) != page_zone_id(buddy))
815 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
822 #ifdef CONFIG_COMPACTION
823 static inline struct capture_control *task_capc(struct zone *zone)
825 struct capture_control *capc = current->capture_control;
828 !(current->flags & PF_KTHREAD) &&
830 capc->cc->zone == zone &&
831 capc->cc->direct_compaction ? capc : NULL;
835 compaction_capture(struct capture_control *capc, struct page *page,
836 int order, int migratetype)
838 if (!capc || order != capc->cc->order)
841 /* Do not accidentally pollute CMA or isolated regions*/
842 if (is_migrate_cma(migratetype) ||
843 is_migrate_isolate(migratetype))
847 * Do not let lower order allocations polluate a movable pageblock.
848 * This might let an unmovable request use a reclaimable pageblock
849 * and vice-versa but no more than normal fallback logic which can
850 * have trouble finding a high-order free page.
852 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
860 static inline struct capture_control *task_capc(struct zone *zone)
866 compaction_capture(struct capture_control *capc, struct page *page,
867 int order, int migratetype)
871 #endif /* CONFIG_COMPACTION */
874 * Freeing function for a buddy system allocator.
876 * The concept of a buddy system is to maintain direct-mapped table
877 * (containing bit values) for memory blocks of various "orders".
878 * The bottom level table contains the map for the smallest allocatable
879 * units of memory (here, pages), and each level above it describes
880 * pairs of units from the levels below, hence, "buddies".
881 * At a high level, all that happens here is marking the table entry
882 * at the bottom level available, and propagating the changes upward
883 * as necessary, plus some accounting needed to play nicely with other
884 * parts of the VM system.
885 * At each level, we keep a list of pages, which are heads of continuous
886 * free pages of length of (1 << order) and marked with PageBuddy.
887 * Page's order is recorded in page_private(page) field.
888 * So when we are allocating or freeing one, we can derive the state of the
889 * other. That is, if we allocate a small block, and both were
890 * free, the remainder of the region must be split into blocks.
891 * If a block is freed, and its buddy is also free, then this
892 * triggers coalescing into a block of larger size.
897 static inline void __free_one_page(struct page *page,
899 struct zone *zone, unsigned int order,
902 unsigned long combined_pfn;
903 unsigned long uninitialized_var(buddy_pfn);
905 unsigned int max_order;
906 struct capture_control *capc = task_capc(zone);
908 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
910 VM_BUG_ON(!zone_is_initialized(zone));
911 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
913 VM_BUG_ON(migratetype == -1);
914 if (likely(!is_migrate_isolate(migratetype)))
915 __mod_zone_freepage_state(zone, 1 << order, migratetype);
917 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
918 VM_BUG_ON_PAGE(bad_range(zone, page), page);
921 while (order < max_order - 1) {
922 if (compaction_capture(capc, page, order, migratetype)) {
923 __mod_zone_freepage_state(zone, -(1 << order),
927 buddy_pfn = __find_buddy_pfn(pfn, order);
928 buddy = page + (buddy_pfn - pfn);
930 if (!pfn_valid_within(buddy_pfn))
932 if (!page_is_buddy(page, buddy, order))
935 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
936 * merge with it and move up one order.
938 if (page_is_guard(buddy))
939 clear_page_guard(zone, buddy, order, migratetype);
941 del_page_from_free_area(buddy, &zone->free_area[order]);
942 combined_pfn = buddy_pfn & pfn;
943 page = page + (combined_pfn - pfn);
947 if (max_order < MAX_ORDER) {
948 /* If we are here, it means order is >= pageblock_order.
949 * We want to prevent merge between freepages on isolate
950 * pageblock and normal pageblock. Without this, pageblock
951 * isolation could cause incorrect freepage or CMA accounting.
953 * We don't want to hit this code for the more frequent
956 if (unlikely(has_isolate_pageblock(zone))) {
959 buddy_pfn = __find_buddy_pfn(pfn, order);
960 buddy = page + (buddy_pfn - pfn);
961 buddy_mt = get_pageblock_migratetype(buddy);
963 if (migratetype != buddy_mt
964 && (is_migrate_isolate(migratetype) ||
965 is_migrate_isolate(buddy_mt)))
969 goto continue_merging;
973 set_page_order(page, order);
976 * If this is not the largest possible page, check if the buddy
977 * of the next-highest order is free. If it is, it's possible
978 * that pages are being freed that will coalesce soon. In case,
979 * that is happening, add the free page to the tail of the list
980 * so it's less likely to be used soon and more likely to be merged
981 * as a higher order page
983 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)
984 && !is_shuffle_order(order)) {
985 struct page *higher_page, *higher_buddy;
986 combined_pfn = buddy_pfn & pfn;
987 higher_page = page + (combined_pfn - pfn);
988 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
989 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
990 if (pfn_valid_within(buddy_pfn) &&
991 page_is_buddy(higher_page, higher_buddy, order + 1)) {
992 add_to_free_area_tail(page, &zone->free_area[order],
998 if (is_shuffle_order(order))
999 add_to_free_area_random(page, &zone->free_area[order],
1002 add_to_free_area(page, &zone->free_area[order], migratetype);
1007 * A bad page could be due to a number of fields. Instead of multiple branches,
1008 * try and check multiple fields with one check. The caller must do a detailed
1009 * check if necessary.
1011 static inline bool page_expected_state(struct page *page,
1012 unsigned long check_flags)
1014 if (unlikely(atomic_read(&page->_mapcount) != -1))
1017 if (unlikely((unsigned long)page->mapping |
1018 page_ref_count(page) |
1020 (unsigned long)page->mem_cgroup |
1022 (page->flags & check_flags)))
1028 static void free_pages_check_bad(struct page *page)
1030 const char *bad_reason;
1031 unsigned long bad_flags;
1036 if (unlikely(atomic_read(&page->_mapcount) != -1))
1037 bad_reason = "nonzero mapcount";
1038 if (unlikely(page->mapping != NULL))
1039 bad_reason = "non-NULL mapping";
1040 if (unlikely(page_ref_count(page) != 0))
1041 bad_reason = "nonzero _refcount";
1042 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
1043 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1044 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
1047 if (unlikely(page->mem_cgroup))
1048 bad_reason = "page still charged to cgroup";
1050 bad_page(page, bad_reason, bad_flags);
1053 static inline int free_pages_check(struct page *page)
1055 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1058 /* Something has gone sideways, find it */
1059 free_pages_check_bad(page);
1063 static int free_tail_pages_check(struct page *head_page, struct page *page)
1068 * We rely page->lru.next never has bit 0 set, unless the page
1069 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1071 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1073 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1077 switch (page - head_page) {
1079 /* the first tail page: ->mapping may be compound_mapcount() */
1080 if (unlikely(compound_mapcount(page))) {
1081 bad_page(page, "nonzero compound_mapcount", 0);
1087 * the second tail page: ->mapping is
1088 * deferred_list.next -- ignore value.
1092 if (page->mapping != TAIL_MAPPING) {
1093 bad_page(page, "corrupted mapping in tail page", 0);
1098 if (unlikely(!PageTail(page))) {
1099 bad_page(page, "PageTail not set", 0);
1102 if (unlikely(compound_head(page) != head_page)) {
1103 bad_page(page, "compound_head not consistent", 0);
1108 page->mapping = NULL;
1109 clear_compound_head(page);
1113 static void kernel_init_free_pages(struct page *page, int numpages)
1117 for (i = 0; i < numpages; i++)
1118 clear_highpage(page + i);
1121 static __always_inline bool free_pages_prepare(struct page *page,
1122 unsigned int order, bool check_free)
1126 VM_BUG_ON_PAGE(PageTail(page), page);
1128 trace_mm_page_free(page, order);
1131 * Check tail pages before head page information is cleared to
1132 * avoid checking PageCompound for order-0 pages.
1134 if (unlikely(order)) {
1135 bool compound = PageCompound(page);
1138 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1141 ClearPageDoubleMap(page);
1142 for (i = 1; i < (1 << order); i++) {
1144 bad += free_tail_pages_check(page, page + i);
1145 if (unlikely(free_pages_check(page + i))) {
1149 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1152 if (PageMappingFlags(page))
1153 page->mapping = NULL;
1154 if (memcg_kmem_enabled() && PageKmemcg(page))
1155 __memcg_kmem_uncharge(page, order);
1157 bad += free_pages_check(page);
1161 page_cpupid_reset_last(page);
1162 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1163 reset_page_owner(page, order);
1165 if (!PageHighMem(page)) {
1166 debug_check_no_locks_freed(page_address(page),
1167 PAGE_SIZE << order);
1168 debug_check_no_obj_freed(page_address(page),
1169 PAGE_SIZE << order);
1171 if (want_init_on_free())
1172 kernel_init_free_pages(page, 1 << order);
1174 kernel_poison_pages(page, 1 << order, 0);
1176 * arch_free_page() can make the page's contents inaccessible. s390
1177 * does this. So nothing which can access the page's contents should
1178 * happen after this.
1180 arch_free_page(page, order);
1182 if (debug_pagealloc_enabled_static())
1183 kernel_map_pages(page, 1 << order, 0);
1185 kasan_free_nondeferred_pages(page, order);
1190 #ifdef CONFIG_DEBUG_VM
1192 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1193 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1194 * moved from pcp lists to free lists.
1196 static bool free_pcp_prepare(struct page *page)
1198 return free_pages_prepare(page, 0, true);
1201 static bool bulkfree_pcp_prepare(struct page *page)
1203 if (debug_pagealloc_enabled_static())
1204 return free_pages_check(page);
1210 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1211 * moving from pcp lists to free list in order to reduce overhead. With
1212 * debug_pagealloc enabled, they are checked also immediately when being freed
1215 static bool free_pcp_prepare(struct page *page)
1217 if (debug_pagealloc_enabled_static())
1218 return free_pages_prepare(page, 0, true);
1220 return free_pages_prepare(page, 0, false);
1223 static bool bulkfree_pcp_prepare(struct page *page)
1225 return free_pages_check(page);
1227 #endif /* CONFIG_DEBUG_VM */
1229 static inline void prefetch_buddy(struct page *page)
1231 unsigned long pfn = page_to_pfn(page);
1232 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1233 struct page *buddy = page + (buddy_pfn - pfn);
1239 * Frees a number of pages from the PCP lists
1240 * Assumes all pages on list are in same zone, and of same order.
1241 * count is the number of pages to free.
1243 * If the zone was previously in an "all pages pinned" state then look to
1244 * see if this freeing clears that state.
1246 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1247 * pinned" detection logic.
1249 static void free_pcppages_bulk(struct zone *zone, int count,
1250 struct per_cpu_pages *pcp)
1252 int migratetype = 0;
1254 int prefetch_nr = 0;
1255 bool isolated_pageblocks;
1256 struct page *page, *tmp;
1260 struct list_head *list;
1263 * Remove pages from lists in a round-robin fashion. A
1264 * batch_free count is maintained that is incremented when an
1265 * empty list is encountered. This is so more pages are freed
1266 * off fuller lists instead of spinning excessively around empty
1271 if (++migratetype == MIGRATE_PCPTYPES)
1273 list = &pcp->lists[migratetype];
1274 } while (list_empty(list));
1276 /* This is the only non-empty list. Free them all. */
1277 if (batch_free == MIGRATE_PCPTYPES)
1281 page = list_last_entry(list, struct page, lru);
1282 /* must delete to avoid corrupting pcp list */
1283 list_del(&page->lru);
1286 if (bulkfree_pcp_prepare(page))
1289 list_add_tail(&page->lru, &head);
1292 * We are going to put the page back to the global
1293 * pool, prefetch its buddy to speed up later access
1294 * under zone->lock. It is believed the overhead of
1295 * an additional test and calculating buddy_pfn here
1296 * can be offset by reduced memory latency later. To
1297 * avoid excessive prefetching due to large count, only
1298 * prefetch buddy for the first pcp->batch nr of pages.
1300 if (prefetch_nr++ < pcp->batch)
1301 prefetch_buddy(page);
1302 } while (--count && --batch_free && !list_empty(list));
1305 spin_lock(&zone->lock);
1306 isolated_pageblocks = has_isolate_pageblock(zone);
1309 * Use safe version since after __free_one_page(),
1310 * page->lru.next will not point to original list.
1312 list_for_each_entry_safe(page, tmp, &head, lru) {
1313 int mt = get_pcppage_migratetype(page);
1314 /* MIGRATE_ISOLATE page should not go to pcplists */
1315 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1316 /* Pageblock could have been isolated meanwhile */
1317 if (unlikely(isolated_pageblocks))
1318 mt = get_pageblock_migratetype(page);
1320 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1321 trace_mm_page_pcpu_drain(page, 0, mt);
1323 spin_unlock(&zone->lock);
1326 static void free_one_page(struct zone *zone,
1327 struct page *page, unsigned long pfn,
1331 spin_lock(&zone->lock);
1332 if (unlikely(has_isolate_pageblock(zone) ||
1333 is_migrate_isolate(migratetype))) {
1334 migratetype = get_pfnblock_migratetype(page, pfn);
1336 __free_one_page(page, pfn, zone, order, migratetype);
1337 spin_unlock(&zone->lock);
1340 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1341 unsigned long zone, int nid)
1343 mm_zero_struct_page(page);
1344 set_page_links(page, zone, nid, pfn);
1345 init_page_count(page);
1346 page_mapcount_reset(page);
1347 page_cpupid_reset_last(page);
1348 page_kasan_tag_reset(page);
1350 INIT_LIST_HEAD(&page->lru);
1351 #ifdef WANT_PAGE_VIRTUAL
1352 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1353 if (!is_highmem_idx(zone))
1354 set_page_address(page, __va(pfn << PAGE_SHIFT));
1358 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1359 static void __meminit init_reserved_page(unsigned long pfn)
1364 if (!early_page_uninitialised(pfn))
1367 nid = early_pfn_to_nid(pfn);
1368 pgdat = NODE_DATA(nid);
1370 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1371 struct zone *zone = &pgdat->node_zones[zid];
1373 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1376 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1379 static inline void init_reserved_page(unsigned long pfn)
1382 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1385 * Initialised pages do not have PageReserved set. This function is
1386 * called for each range allocated by the bootmem allocator and
1387 * marks the pages PageReserved. The remaining valid pages are later
1388 * sent to the buddy page allocator.
1390 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1392 unsigned long start_pfn = PFN_DOWN(start);
1393 unsigned long end_pfn = PFN_UP(end);
1395 for (; start_pfn < end_pfn; start_pfn++) {
1396 if (pfn_valid(start_pfn)) {
1397 struct page *page = pfn_to_page(start_pfn);
1399 init_reserved_page(start_pfn);
1401 /* Avoid false-positive PageTail() */
1402 INIT_LIST_HEAD(&page->lru);
1405 * no need for atomic set_bit because the struct
1406 * page is not visible yet so nobody should
1409 __SetPageReserved(page);
1414 static void __free_pages_ok(struct page *page, unsigned int order)
1416 unsigned long flags;
1418 unsigned long pfn = page_to_pfn(page);
1420 if (!free_pages_prepare(page, order, true))
1423 migratetype = get_pfnblock_migratetype(page, pfn);
1424 local_irq_save(flags);
1425 __count_vm_events(PGFREE, 1 << order);
1426 free_one_page(page_zone(page), page, pfn, order, migratetype);
1427 local_irq_restore(flags);
1430 void __free_pages_core(struct page *page, unsigned int order)
1432 unsigned int nr_pages = 1 << order;
1433 struct page *p = page;
1437 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1439 __ClearPageReserved(p);
1440 set_page_count(p, 0);
1442 __ClearPageReserved(p);
1443 set_page_count(p, 0);
1445 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1446 set_page_refcounted(page);
1447 __free_pages(page, order);
1450 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1451 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1453 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1455 int __meminit early_pfn_to_nid(unsigned long pfn)
1457 static DEFINE_SPINLOCK(early_pfn_lock);
1460 spin_lock(&early_pfn_lock);
1461 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1463 nid = first_online_node;
1464 spin_unlock(&early_pfn_lock);
1470 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1471 /* Only safe to use early in boot when initialisation is single-threaded */
1472 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1476 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1477 if (nid >= 0 && nid != node)
1483 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1490 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1493 if (early_page_uninitialised(pfn))
1495 __free_pages_core(page, order);
1499 * Check that the whole (or subset of) a pageblock given by the interval of
1500 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1501 * with the migration of free compaction scanner. The scanners then need to
1502 * use only pfn_valid_within() check for arches that allow holes within
1505 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1507 * It's possible on some configurations to have a setup like node0 node1 node0
1508 * i.e. it's possible that all pages within a zones range of pages do not
1509 * belong to a single zone. We assume that a border between node0 and node1
1510 * can occur within a single pageblock, but not a node0 node1 node0
1511 * interleaving within a single pageblock. It is therefore sufficient to check
1512 * the first and last page of a pageblock and avoid checking each individual
1513 * page in a pageblock.
1515 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1516 unsigned long end_pfn, struct zone *zone)
1518 struct page *start_page;
1519 struct page *end_page;
1521 /* end_pfn is one past the range we are checking */
1524 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1527 start_page = pfn_to_online_page(start_pfn);
1531 if (page_zone(start_page) != zone)
1534 end_page = pfn_to_page(end_pfn);
1536 /* This gives a shorter code than deriving page_zone(end_page) */
1537 if (page_zone_id(start_page) != page_zone_id(end_page))
1543 void set_zone_contiguous(struct zone *zone)
1545 unsigned long block_start_pfn = zone->zone_start_pfn;
1546 unsigned long block_end_pfn;
1548 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1549 for (; block_start_pfn < zone_end_pfn(zone);
1550 block_start_pfn = block_end_pfn,
1551 block_end_pfn += pageblock_nr_pages) {
1553 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1555 if (!__pageblock_pfn_to_page(block_start_pfn,
1556 block_end_pfn, zone))
1561 /* We confirm that there is no hole */
1562 zone->contiguous = true;
1565 void clear_zone_contiguous(struct zone *zone)
1567 zone->contiguous = false;
1570 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1571 static void __init deferred_free_range(unsigned long pfn,
1572 unsigned long nr_pages)
1580 page = pfn_to_page(pfn);
1582 /* Free a large naturally-aligned chunk if possible */
1583 if (nr_pages == pageblock_nr_pages &&
1584 (pfn & (pageblock_nr_pages - 1)) == 0) {
1585 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1586 __free_pages_core(page, pageblock_order);
1590 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1591 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1592 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1593 __free_pages_core(page, 0);
1597 /* Completion tracking for deferred_init_memmap() threads */
1598 static atomic_t pgdat_init_n_undone __initdata;
1599 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1601 static inline void __init pgdat_init_report_one_done(void)
1603 if (atomic_dec_and_test(&pgdat_init_n_undone))
1604 complete(&pgdat_init_all_done_comp);
1608 * Returns true if page needs to be initialized or freed to buddy allocator.
1610 * First we check if pfn is valid on architectures where it is possible to have
1611 * holes within pageblock_nr_pages. On systems where it is not possible, this
1612 * function is optimized out.
1614 * Then, we check if a current large page is valid by only checking the validity
1617 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1619 if (!pfn_valid_within(pfn))
1621 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1627 * Free pages to buddy allocator. Try to free aligned pages in
1628 * pageblock_nr_pages sizes.
1630 static void __init deferred_free_pages(unsigned long pfn,
1631 unsigned long end_pfn)
1633 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1634 unsigned long nr_free = 0;
1636 for (; pfn < end_pfn; pfn++) {
1637 if (!deferred_pfn_valid(pfn)) {
1638 deferred_free_range(pfn - nr_free, nr_free);
1640 } else if (!(pfn & nr_pgmask)) {
1641 deferred_free_range(pfn - nr_free, nr_free);
1647 /* Free the last block of pages to allocator */
1648 deferred_free_range(pfn - nr_free, nr_free);
1652 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1653 * by performing it only once every pageblock_nr_pages.
1654 * Return number of pages initialized.
1656 static unsigned long __init deferred_init_pages(struct zone *zone,
1658 unsigned long end_pfn)
1660 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1661 int nid = zone_to_nid(zone);
1662 unsigned long nr_pages = 0;
1663 int zid = zone_idx(zone);
1664 struct page *page = NULL;
1666 for (; pfn < end_pfn; pfn++) {
1667 if (!deferred_pfn_valid(pfn)) {
1670 } else if (!page || !(pfn & nr_pgmask)) {
1671 page = pfn_to_page(pfn);
1675 __init_single_page(page, pfn, zid, nid);
1682 * This function is meant to pre-load the iterator for the zone init.
1683 * Specifically it walks through the ranges until we are caught up to the
1684 * first_init_pfn value and exits there. If we never encounter the value we
1685 * return false indicating there are no valid ranges left.
1688 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1689 unsigned long *spfn, unsigned long *epfn,
1690 unsigned long first_init_pfn)
1695 * Start out by walking through the ranges in this zone that have
1696 * already been initialized. We don't need to do anything with them
1697 * so we just need to flush them out of the system.
1699 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1700 if (*epfn <= first_init_pfn)
1702 if (*spfn < first_init_pfn)
1703 *spfn = first_init_pfn;
1712 * Initialize and free pages. We do it in two loops: first we initialize
1713 * struct page, then free to buddy allocator, because while we are
1714 * freeing pages we can access pages that are ahead (computing buddy
1715 * page in __free_one_page()).
1717 * In order to try and keep some memory in the cache we have the loop
1718 * broken along max page order boundaries. This way we will not cause
1719 * any issues with the buddy page computation.
1721 static unsigned long __init
1722 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1723 unsigned long *end_pfn)
1725 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1726 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1727 unsigned long nr_pages = 0;
1730 /* First we loop through and initialize the page values */
1731 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1734 if (mo_pfn <= *start_pfn)
1737 t = min(mo_pfn, *end_pfn);
1738 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1740 if (mo_pfn < *end_pfn) {
1741 *start_pfn = mo_pfn;
1746 /* Reset values and now loop through freeing pages as needed */
1749 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1755 t = min(mo_pfn, epfn);
1756 deferred_free_pages(spfn, t);
1765 /* Initialise remaining memory on a node */
1766 static int __init deferred_init_memmap(void *data)
1768 pg_data_t *pgdat = data;
1769 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1770 unsigned long spfn = 0, epfn = 0, nr_pages = 0;
1771 unsigned long first_init_pfn, flags;
1772 unsigned long start = jiffies;
1777 /* Bind memory initialisation thread to a local node if possible */
1778 if (!cpumask_empty(cpumask))
1779 set_cpus_allowed_ptr(current, cpumask);
1781 pgdat_resize_lock(pgdat, &flags);
1782 first_init_pfn = pgdat->first_deferred_pfn;
1783 if (first_init_pfn == ULONG_MAX) {
1784 pgdat_resize_unlock(pgdat, &flags);
1785 pgdat_init_report_one_done();
1789 /* Sanity check boundaries */
1790 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1791 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1792 pgdat->first_deferred_pfn = ULONG_MAX;
1795 * Once we unlock here, the zone cannot be grown anymore, thus if an
1796 * interrupt thread must allocate this early in boot, zone must be
1797 * pre-grown prior to start of deferred page initialization.
1799 pgdat_resize_unlock(pgdat, &flags);
1801 /* Only the highest zone is deferred so find it */
1802 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1803 zone = pgdat->node_zones + zid;
1804 if (first_init_pfn < zone_end_pfn(zone))
1808 /* If the zone is empty somebody else may have cleared out the zone */
1809 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1814 * Initialize and free pages in MAX_ORDER sized increments so
1815 * that we can avoid introducing any issues with the buddy
1818 while (spfn < epfn) {
1819 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1820 touch_nmi_watchdog();
1823 /* Sanity check that the next zone really is unpopulated */
1824 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1826 pr_info("node %d initialised, %lu pages in %ums\n",
1827 pgdat->node_id, nr_pages, jiffies_to_msecs(jiffies - start));
1829 pgdat_init_report_one_done();
1834 * If this zone has deferred pages, try to grow it by initializing enough
1835 * deferred pages to satisfy the allocation specified by order, rounded up to
1836 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1837 * of SECTION_SIZE bytes by initializing struct pages in increments of
1838 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1840 * Return true when zone was grown, otherwise return false. We return true even
1841 * when we grow less than requested, to let the caller decide if there are
1842 * enough pages to satisfy the allocation.
1844 * Note: We use noinline because this function is needed only during boot, and
1845 * it is called from a __ref function _deferred_grow_zone. This way we are
1846 * making sure that it is not inlined into permanent text section.
1848 static noinline bool __init
1849 deferred_grow_zone(struct zone *zone, unsigned int order)
1851 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1852 pg_data_t *pgdat = zone->zone_pgdat;
1853 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1854 unsigned long spfn, epfn, flags;
1855 unsigned long nr_pages = 0;
1858 /* Only the last zone may have deferred pages */
1859 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1862 pgdat_resize_lock(pgdat, &flags);
1865 * If someone grew this zone while we were waiting for spinlock, return
1866 * true, as there might be enough pages already.
1868 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1869 pgdat_resize_unlock(pgdat, &flags);
1873 /* If the zone is empty somebody else may have cleared out the zone */
1874 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1875 first_deferred_pfn)) {
1876 pgdat->first_deferred_pfn = ULONG_MAX;
1877 pgdat_resize_unlock(pgdat, &flags);
1878 /* Retry only once. */
1879 return first_deferred_pfn != ULONG_MAX;
1883 * Initialize and free pages in MAX_ORDER sized increments so
1884 * that we can avoid introducing any issues with the buddy
1887 while (spfn < epfn) {
1888 /* update our first deferred PFN for this section */
1889 first_deferred_pfn = spfn;
1891 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1892 touch_nmi_watchdog();
1894 /* We should only stop along section boundaries */
1895 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
1898 /* If our quota has been met we can stop here */
1899 if (nr_pages >= nr_pages_needed)
1903 pgdat->first_deferred_pfn = spfn;
1904 pgdat_resize_unlock(pgdat, &flags);
1906 return nr_pages > 0;
1910 * deferred_grow_zone() is __init, but it is called from
1911 * get_page_from_freelist() during early boot until deferred_pages permanently
1912 * disables this call. This is why we have refdata wrapper to avoid warning,
1913 * and to ensure that the function body gets unloaded.
1916 _deferred_grow_zone(struct zone *zone, unsigned int order)
1918 return deferred_grow_zone(zone, order);
1921 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1923 void __init page_alloc_init_late(void)
1928 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1930 /* There will be num_node_state(N_MEMORY) threads */
1931 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1932 for_each_node_state(nid, N_MEMORY) {
1933 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1936 /* Block until all are initialised */
1937 wait_for_completion(&pgdat_init_all_done_comp);
1940 * The number of managed pages has changed due to the initialisation
1941 * so the pcpu batch and high limits needs to be updated or the limits
1942 * will be artificially small.
1944 for_each_populated_zone(zone)
1945 zone_pcp_update(zone);
1948 * We initialized the rest of the deferred pages. Permanently disable
1949 * on-demand struct page initialization.
1951 static_branch_disable(&deferred_pages);
1953 /* Reinit limits that are based on free pages after the kernel is up */
1954 files_maxfiles_init();
1957 /* Discard memblock private memory */
1960 for_each_node_state(nid, N_MEMORY)
1961 shuffle_free_memory(NODE_DATA(nid));
1963 for_each_populated_zone(zone)
1964 set_zone_contiguous(zone);
1968 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1969 void __init init_cma_reserved_pageblock(struct page *page)
1971 unsigned i = pageblock_nr_pages;
1972 struct page *p = page;
1975 __ClearPageReserved(p);
1976 set_page_count(p, 0);
1979 set_pageblock_migratetype(page, MIGRATE_CMA);
1981 if (pageblock_order >= MAX_ORDER) {
1982 i = pageblock_nr_pages;
1985 set_page_refcounted(p);
1986 __free_pages(p, MAX_ORDER - 1);
1987 p += MAX_ORDER_NR_PAGES;
1988 } while (i -= MAX_ORDER_NR_PAGES);
1990 set_page_refcounted(page);
1991 __free_pages(page, pageblock_order);
1994 adjust_managed_page_count(page, pageblock_nr_pages);
1999 * The order of subdivision here is critical for the IO subsystem.
2000 * Please do not alter this order without good reasons and regression
2001 * testing. Specifically, as large blocks of memory are subdivided,
2002 * the order in which smaller blocks are delivered depends on the order
2003 * they're subdivided in this function. This is the primary factor
2004 * influencing the order in which pages are delivered to the IO
2005 * subsystem according to empirical testing, and this is also justified
2006 * by considering the behavior of a buddy system containing a single
2007 * large block of memory acted on by a series of small allocations.
2008 * This behavior is a critical factor in sglist merging's success.
2012 static inline void expand(struct zone *zone, struct page *page,
2013 int low, int high, struct free_area *area,
2016 unsigned long size = 1 << high;
2018 while (high > low) {
2022 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2025 * Mark as guard pages (or page), that will allow to
2026 * merge back to allocator when buddy will be freed.
2027 * Corresponding page table entries will not be touched,
2028 * pages will stay not present in virtual address space
2030 if (set_page_guard(zone, &page[size], high, migratetype))
2033 add_to_free_area(&page[size], area, migratetype);
2034 set_page_order(&page[size], high);
2038 static void check_new_page_bad(struct page *page)
2040 const char *bad_reason = NULL;
2041 unsigned long bad_flags = 0;
2043 if (unlikely(atomic_read(&page->_mapcount) != -1))
2044 bad_reason = "nonzero mapcount";
2045 if (unlikely(page->mapping != NULL))
2046 bad_reason = "non-NULL mapping";
2047 if (unlikely(page_ref_count(page) != 0))
2048 bad_reason = "nonzero _refcount";
2049 if (unlikely(page->flags & __PG_HWPOISON)) {
2050 bad_reason = "HWPoisoned (hardware-corrupted)";
2051 bad_flags = __PG_HWPOISON;
2052 /* Don't complain about hwpoisoned pages */
2053 page_mapcount_reset(page); /* remove PageBuddy */
2056 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
2057 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
2058 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
2061 if (unlikely(page->mem_cgroup))
2062 bad_reason = "page still charged to cgroup";
2064 bad_page(page, bad_reason, bad_flags);
2068 * This page is about to be returned from the page allocator
2070 static inline int check_new_page(struct page *page)
2072 if (likely(page_expected_state(page,
2073 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2076 check_new_page_bad(page);
2080 static inline bool free_pages_prezeroed(void)
2082 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2083 page_poisoning_enabled()) || want_init_on_free();
2086 #ifdef CONFIG_DEBUG_VM
2088 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2089 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2090 * also checked when pcp lists are refilled from the free lists.
2092 static inline bool check_pcp_refill(struct page *page)
2094 if (debug_pagealloc_enabled_static())
2095 return check_new_page(page);
2100 static inline bool check_new_pcp(struct page *page)
2102 return check_new_page(page);
2106 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2107 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2108 * enabled, they are also checked when being allocated from the pcp lists.
2110 static inline bool check_pcp_refill(struct page *page)
2112 return check_new_page(page);
2114 static inline bool check_new_pcp(struct page *page)
2116 if (debug_pagealloc_enabled_static())
2117 return check_new_page(page);
2121 #endif /* CONFIG_DEBUG_VM */
2123 static bool check_new_pages(struct page *page, unsigned int order)
2126 for (i = 0; i < (1 << order); i++) {
2127 struct page *p = page + i;
2129 if (unlikely(check_new_page(p)))
2136 inline void post_alloc_hook(struct page *page, unsigned int order,
2139 set_page_private(page, 0);
2140 set_page_refcounted(page);
2142 arch_alloc_page(page, order);
2143 if (debug_pagealloc_enabled_static())
2144 kernel_map_pages(page, 1 << order, 1);
2145 kasan_alloc_pages(page, order);
2146 kernel_poison_pages(page, 1 << order, 1);
2147 set_page_owner(page, order, gfp_flags);
2150 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2151 unsigned int alloc_flags)
2153 post_alloc_hook(page, order, gfp_flags);
2155 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2156 kernel_init_free_pages(page, 1 << order);
2158 if (order && (gfp_flags & __GFP_COMP))
2159 prep_compound_page(page, order);
2162 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2163 * allocate the page. The expectation is that the caller is taking
2164 * steps that will free more memory. The caller should avoid the page
2165 * being used for !PFMEMALLOC purposes.
2167 if (alloc_flags & ALLOC_NO_WATERMARKS)
2168 set_page_pfmemalloc(page);
2170 clear_page_pfmemalloc(page);
2174 * Go through the free lists for the given migratetype and remove
2175 * the smallest available page from the freelists
2177 static __always_inline
2178 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2181 unsigned int current_order;
2182 struct free_area *area;
2185 /* Find a page of the appropriate size in the preferred list */
2186 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2187 area = &(zone->free_area[current_order]);
2188 page = get_page_from_free_area(area, migratetype);
2191 del_page_from_free_area(page, area);
2192 expand(zone, page, order, current_order, area, migratetype);
2193 set_pcppage_migratetype(page, migratetype);
2202 * This array describes the order lists are fallen back to when
2203 * the free lists for the desirable migrate type are depleted
2205 static int fallbacks[MIGRATE_TYPES][4] = {
2206 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2207 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2208 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2210 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2212 #ifdef CONFIG_MEMORY_ISOLATION
2213 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2218 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2221 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2224 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2225 unsigned int order) { return NULL; }
2229 * Move the free pages in a range to the free lists of the requested type.
2230 * Note that start_page and end_pages are not aligned on a pageblock
2231 * boundary. If alignment is required, use move_freepages_block()
2233 static int move_freepages(struct zone *zone,
2234 struct page *start_page, struct page *end_page,
2235 int migratetype, int *num_movable)
2239 int pages_moved = 0;
2241 for (page = start_page; page <= end_page;) {
2242 if (!pfn_valid_within(page_to_pfn(page))) {
2247 if (!PageBuddy(page)) {
2249 * We assume that pages that could be isolated for
2250 * migration are movable. But we don't actually try
2251 * isolating, as that would be expensive.
2254 (PageLRU(page) || __PageMovable(page)))
2261 /* Make sure we are not inadvertently changing nodes */
2262 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2263 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2265 order = page_order(page);
2266 move_to_free_area(page, &zone->free_area[order], migratetype);
2268 pages_moved += 1 << order;
2274 int move_freepages_block(struct zone *zone, struct page *page,
2275 int migratetype, int *num_movable)
2277 unsigned long start_pfn, end_pfn;
2278 struct page *start_page, *end_page;
2283 start_pfn = page_to_pfn(page);
2284 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2285 start_page = pfn_to_page(start_pfn);
2286 end_page = start_page + pageblock_nr_pages - 1;
2287 end_pfn = start_pfn + pageblock_nr_pages - 1;
2289 /* Do not cross zone boundaries */
2290 if (!zone_spans_pfn(zone, start_pfn))
2292 if (!zone_spans_pfn(zone, end_pfn))
2295 return move_freepages(zone, start_page, end_page, migratetype,
2299 static void change_pageblock_range(struct page *pageblock_page,
2300 int start_order, int migratetype)
2302 int nr_pageblocks = 1 << (start_order - pageblock_order);
2304 while (nr_pageblocks--) {
2305 set_pageblock_migratetype(pageblock_page, migratetype);
2306 pageblock_page += pageblock_nr_pages;
2311 * When we are falling back to another migratetype during allocation, try to
2312 * steal extra free pages from the same pageblocks to satisfy further
2313 * allocations, instead of polluting multiple pageblocks.
2315 * If we are stealing a relatively large buddy page, it is likely there will
2316 * be more free pages in the pageblock, so try to steal them all. For
2317 * reclaimable and unmovable allocations, we steal regardless of page size,
2318 * as fragmentation caused by those allocations polluting movable pageblocks
2319 * is worse than movable allocations stealing from unmovable and reclaimable
2322 static bool can_steal_fallback(unsigned int order, int start_mt)
2325 * Leaving this order check is intended, although there is
2326 * relaxed order check in next check. The reason is that
2327 * we can actually steal whole pageblock if this condition met,
2328 * but, below check doesn't guarantee it and that is just heuristic
2329 * so could be changed anytime.
2331 if (order >= pageblock_order)
2334 if (order >= pageblock_order / 2 ||
2335 start_mt == MIGRATE_RECLAIMABLE ||
2336 start_mt == MIGRATE_UNMOVABLE ||
2337 page_group_by_mobility_disabled)
2343 static inline void boost_watermark(struct zone *zone)
2345 unsigned long max_boost;
2347 if (!watermark_boost_factor)
2350 * Don't bother in zones that are unlikely to produce results.
2351 * On small machines, including kdump capture kernels running
2352 * in a small area, boosting the watermark can cause an out of
2353 * memory situation immediately.
2355 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2358 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2359 watermark_boost_factor, 10000);
2362 * high watermark may be uninitialised if fragmentation occurs
2363 * very early in boot so do not boost. We do not fall
2364 * through and boost by pageblock_nr_pages as failing
2365 * allocations that early means that reclaim is not going
2366 * to help and it may even be impossible to reclaim the
2367 * boosted watermark resulting in a hang.
2372 max_boost = max(pageblock_nr_pages, max_boost);
2374 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2379 * This function implements actual steal behaviour. If order is large enough,
2380 * we can steal whole pageblock. If not, we first move freepages in this
2381 * pageblock to our migratetype and determine how many already-allocated pages
2382 * are there in the pageblock with a compatible migratetype. If at least half
2383 * of pages are free or compatible, we can change migratetype of the pageblock
2384 * itself, so pages freed in the future will be put on the correct free list.
2386 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2387 unsigned int alloc_flags, int start_type, bool whole_block)
2389 unsigned int current_order = page_order(page);
2390 struct free_area *area;
2391 int free_pages, movable_pages, alike_pages;
2394 old_block_type = get_pageblock_migratetype(page);
2397 * This can happen due to races and we want to prevent broken
2398 * highatomic accounting.
2400 if (is_migrate_highatomic(old_block_type))
2403 /* Take ownership for orders >= pageblock_order */
2404 if (current_order >= pageblock_order) {
2405 change_pageblock_range(page, current_order, start_type);
2410 * Boost watermarks to increase reclaim pressure to reduce the
2411 * likelihood of future fallbacks. Wake kswapd now as the node
2412 * may be balanced overall and kswapd will not wake naturally.
2414 boost_watermark(zone);
2415 if (alloc_flags & ALLOC_KSWAPD)
2416 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2418 /* We are not allowed to try stealing from the whole block */
2422 free_pages = move_freepages_block(zone, page, start_type,
2425 * Determine how many pages are compatible with our allocation.
2426 * For movable allocation, it's the number of movable pages which
2427 * we just obtained. For other types it's a bit more tricky.
2429 if (start_type == MIGRATE_MOVABLE) {
2430 alike_pages = movable_pages;
2433 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2434 * to MOVABLE pageblock, consider all non-movable pages as
2435 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2436 * vice versa, be conservative since we can't distinguish the
2437 * exact migratetype of non-movable pages.
2439 if (old_block_type == MIGRATE_MOVABLE)
2440 alike_pages = pageblock_nr_pages
2441 - (free_pages + movable_pages);
2446 /* moving whole block can fail due to zone boundary conditions */
2451 * If a sufficient number of pages in the block are either free or of
2452 * comparable migratability as our allocation, claim the whole block.
2454 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2455 page_group_by_mobility_disabled)
2456 set_pageblock_migratetype(page, start_type);
2461 area = &zone->free_area[current_order];
2462 move_to_free_area(page, area, start_type);
2466 * Check whether there is a suitable fallback freepage with requested order.
2467 * If only_stealable is true, this function returns fallback_mt only if
2468 * we can steal other freepages all together. This would help to reduce
2469 * fragmentation due to mixed migratetype pages in one pageblock.
2471 int find_suitable_fallback(struct free_area *area, unsigned int order,
2472 int migratetype, bool only_stealable, bool *can_steal)
2477 if (area->nr_free == 0)
2482 fallback_mt = fallbacks[migratetype][i];
2483 if (fallback_mt == MIGRATE_TYPES)
2486 if (free_area_empty(area, fallback_mt))
2489 if (can_steal_fallback(order, migratetype))
2492 if (!only_stealable)
2503 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2504 * there are no empty page blocks that contain a page with a suitable order
2506 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2507 unsigned int alloc_order)
2510 unsigned long max_managed, flags;
2513 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2514 * Check is race-prone but harmless.
2516 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2517 if (zone->nr_reserved_highatomic >= max_managed)
2520 spin_lock_irqsave(&zone->lock, flags);
2522 /* Recheck the nr_reserved_highatomic limit under the lock */
2523 if (zone->nr_reserved_highatomic >= max_managed)
2527 mt = get_pageblock_migratetype(page);
2528 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2529 && !is_migrate_cma(mt)) {
2530 zone->nr_reserved_highatomic += pageblock_nr_pages;
2531 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2532 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2536 spin_unlock_irqrestore(&zone->lock, flags);
2540 * Used when an allocation is about to fail under memory pressure. This
2541 * potentially hurts the reliability of high-order allocations when under
2542 * intense memory pressure but failed atomic allocations should be easier
2543 * to recover from than an OOM.
2545 * If @force is true, try to unreserve a pageblock even though highatomic
2546 * pageblock is exhausted.
2548 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2551 struct zonelist *zonelist = ac->zonelist;
2552 unsigned long flags;
2559 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2562 * Preserve at least one pageblock unless memory pressure
2565 if (!force && zone->nr_reserved_highatomic <=
2569 spin_lock_irqsave(&zone->lock, flags);
2570 for (order = 0; order < MAX_ORDER; order++) {
2571 struct free_area *area = &(zone->free_area[order]);
2573 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2578 * In page freeing path, migratetype change is racy so
2579 * we can counter several free pages in a pageblock
2580 * in this loop althoug we changed the pageblock type
2581 * from highatomic to ac->migratetype. So we should
2582 * adjust the count once.
2584 if (is_migrate_highatomic_page(page)) {
2586 * It should never happen but changes to
2587 * locking could inadvertently allow a per-cpu
2588 * drain to add pages to MIGRATE_HIGHATOMIC
2589 * while unreserving so be safe and watch for
2592 zone->nr_reserved_highatomic -= min(
2594 zone->nr_reserved_highatomic);
2598 * Convert to ac->migratetype and avoid the normal
2599 * pageblock stealing heuristics. Minimally, the caller
2600 * is doing the work and needs the pages. More
2601 * importantly, if the block was always converted to
2602 * MIGRATE_UNMOVABLE or another type then the number
2603 * of pageblocks that cannot be completely freed
2606 set_pageblock_migratetype(page, ac->migratetype);
2607 ret = move_freepages_block(zone, page, ac->migratetype,
2610 spin_unlock_irqrestore(&zone->lock, flags);
2614 spin_unlock_irqrestore(&zone->lock, flags);
2621 * Try finding a free buddy page on the fallback list and put it on the free
2622 * list of requested migratetype, possibly along with other pages from the same
2623 * block, depending on fragmentation avoidance heuristics. Returns true if
2624 * fallback was found so that __rmqueue_smallest() can grab it.
2626 * The use of signed ints for order and current_order is a deliberate
2627 * deviation from the rest of this file, to make the for loop
2628 * condition simpler.
2630 static __always_inline bool
2631 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2632 unsigned int alloc_flags)
2634 struct free_area *area;
2636 int min_order = order;
2642 * Do not steal pages from freelists belonging to other pageblocks
2643 * i.e. orders < pageblock_order. If there are no local zones free,
2644 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2646 if (alloc_flags & ALLOC_NOFRAGMENT)
2647 min_order = pageblock_order;
2650 * Find the largest available free page in the other list. This roughly
2651 * approximates finding the pageblock with the most free pages, which
2652 * would be too costly to do exactly.
2654 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2656 area = &(zone->free_area[current_order]);
2657 fallback_mt = find_suitable_fallback(area, current_order,
2658 start_migratetype, false, &can_steal);
2659 if (fallback_mt == -1)
2663 * We cannot steal all free pages from the pageblock and the
2664 * requested migratetype is movable. In that case it's better to
2665 * steal and split the smallest available page instead of the
2666 * largest available page, because even if the next movable
2667 * allocation falls back into a different pageblock than this
2668 * one, it won't cause permanent fragmentation.
2670 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2671 && current_order > order)
2680 for (current_order = order; current_order < MAX_ORDER;
2682 area = &(zone->free_area[current_order]);
2683 fallback_mt = find_suitable_fallback(area, current_order,
2684 start_migratetype, false, &can_steal);
2685 if (fallback_mt != -1)
2690 * This should not happen - we already found a suitable fallback
2691 * when looking for the largest page.
2693 VM_BUG_ON(current_order == MAX_ORDER);
2696 page = get_page_from_free_area(area, fallback_mt);
2698 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2701 trace_mm_page_alloc_extfrag(page, order, current_order,
2702 start_migratetype, fallback_mt);
2709 * Do the hard work of removing an element from the buddy allocator.
2710 * Call me with the zone->lock already held.
2712 static __always_inline struct page *
2713 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2714 unsigned int alloc_flags)
2719 page = __rmqueue_smallest(zone, order, migratetype);
2720 if (unlikely(!page)) {
2721 if (migratetype == MIGRATE_MOVABLE)
2722 page = __rmqueue_cma_fallback(zone, order);
2724 if (!page && __rmqueue_fallback(zone, order, migratetype,
2729 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2734 * Obtain a specified number of elements from the buddy allocator, all under
2735 * a single hold of the lock, for efficiency. Add them to the supplied list.
2736 * Returns the number of new pages which were placed at *list.
2738 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2739 unsigned long count, struct list_head *list,
2740 int migratetype, unsigned int alloc_flags)
2744 spin_lock(&zone->lock);
2745 for (i = 0; i < count; ++i) {
2746 struct page *page = __rmqueue(zone, order, migratetype,
2748 if (unlikely(page == NULL))
2751 if (unlikely(check_pcp_refill(page)))
2755 * Split buddy pages returned by expand() are received here in
2756 * physical page order. The page is added to the tail of
2757 * caller's list. From the callers perspective, the linked list
2758 * is ordered by page number under some conditions. This is
2759 * useful for IO devices that can forward direction from the
2760 * head, thus also in the physical page order. This is useful
2761 * for IO devices that can merge IO requests if the physical
2762 * pages are ordered properly.
2764 list_add_tail(&page->lru, list);
2766 if (is_migrate_cma(get_pcppage_migratetype(page)))
2767 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2772 * i pages were removed from the buddy list even if some leak due
2773 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2774 * on i. Do not confuse with 'alloced' which is the number of
2775 * pages added to the pcp list.
2777 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2778 spin_unlock(&zone->lock);
2784 * Called from the vmstat counter updater to drain pagesets of this
2785 * currently executing processor on remote nodes after they have
2788 * Note that this function must be called with the thread pinned to
2789 * a single processor.
2791 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2793 unsigned long flags;
2794 int to_drain, batch;
2796 local_irq_save(flags);
2797 batch = READ_ONCE(pcp->batch);
2798 to_drain = min(pcp->count, batch);
2800 free_pcppages_bulk(zone, to_drain, pcp);
2801 local_irq_restore(flags);
2806 * Drain pcplists of the indicated processor and zone.
2808 * The processor must either be the current processor and the
2809 * thread pinned to the current processor or a processor that
2812 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2814 unsigned long flags;
2815 struct per_cpu_pageset *pset;
2816 struct per_cpu_pages *pcp;
2818 local_irq_save(flags);
2819 pset = per_cpu_ptr(zone->pageset, cpu);
2823 free_pcppages_bulk(zone, pcp->count, pcp);
2824 local_irq_restore(flags);
2828 * Drain pcplists of all zones on the indicated processor.
2830 * The processor must either be the current processor and the
2831 * thread pinned to the current processor or a processor that
2834 static void drain_pages(unsigned int cpu)
2838 for_each_populated_zone(zone) {
2839 drain_pages_zone(cpu, zone);
2844 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2846 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2847 * the single zone's pages.
2849 void drain_local_pages(struct zone *zone)
2851 int cpu = smp_processor_id();
2854 drain_pages_zone(cpu, zone);
2859 static void drain_local_pages_wq(struct work_struct *work)
2861 struct pcpu_drain *drain;
2863 drain = container_of(work, struct pcpu_drain, work);
2866 * drain_all_pages doesn't use proper cpu hotplug protection so
2867 * we can race with cpu offline when the WQ can move this from
2868 * a cpu pinned worker to an unbound one. We can operate on a different
2869 * cpu which is allright but we also have to make sure to not move to
2873 drain_local_pages(drain->zone);
2878 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2880 * When zone parameter is non-NULL, spill just the single zone's pages.
2882 * Note that this can be extremely slow as the draining happens in a workqueue.
2884 void drain_all_pages(struct zone *zone)
2889 * Allocate in the BSS so we wont require allocation in
2890 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2892 static cpumask_t cpus_with_pcps;
2895 * Make sure nobody triggers this path before mm_percpu_wq is fully
2898 if (WARN_ON_ONCE(!mm_percpu_wq))
2902 * Do not drain if one is already in progress unless it's specific to
2903 * a zone. Such callers are primarily CMA and memory hotplug and need
2904 * the drain to be complete when the call returns.
2906 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2909 mutex_lock(&pcpu_drain_mutex);
2913 * We don't care about racing with CPU hotplug event
2914 * as offline notification will cause the notified
2915 * cpu to drain that CPU pcps and on_each_cpu_mask
2916 * disables preemption as part of its processing
2918 for_each_online_cpu(cpu) {
2919 struct per_cpu_pageset *pcp;
2921 bool has_pcps = false;
2924 pcp = per_cpu_ptr(zone->pageset, cpu);
2928 for_each_populated_zone(z) {
2929 pcp = per_cpu_ptr(z->pageset, cpu);
2930 if (pcp->pcp.count) {
2938 cpumask_set_cpu(cpu, &cpus_with_pcps);
2940 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2943 for_each_cpu(cpu, &cpus_with_pcps) {
2944 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
2947 INIT_WORK(&drain->work, drain_local_pages_wq);
2948 queue_work_on(cpu, mm_percpu_wq, &drain->work);
2950 for_each_cpu(cpu, &cpus_with_pcps)
2951 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
2953 mutex_unlock(&pcpu_drain_mutex);
2956 #ifdef CONFIG_HIBERNATION
2959 * Touch the watchdog for every WD_PAGE_COUNT pages.
2961 #define WD_PAGE_COUNT (128*1024)
2963 void mark_free_pages(struct zone *zone)
2965 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2966 unsigned long flags;
2967 unsigned int order, t;
2970 if (zone_is_empty(zone))
2973 spin_lock_irqsave(&zone->lock, flags);
2975 max_zone_pfn = zone_end_pfn(zone);
2976 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2977 if (pfn_valid(pfn)) {
2978 page = pfn_to_page(pfn);
2980 if (!--page_count) {
2981 touch_nmi_watchdog();
2982 page_count = WD_PAGE_COUNT;
2985 if (page_zone(page) != zone)
2988 if (!swsusp_page_is_forbidden(page))
2989 swsusp_unset_page_free(page);
2992 for_each_migratetype_order(order, t) {
2993 list_for_each_entry(page,
2994 &zone->free_area[order].free_list[t], lru) {
2997 pfn = page_to_pfn(page);
2998 for (i = 0; i < (1UL << order); i++) {
2999 if (!--page_count) {
3000 touch_nmi_watchdog();
3001 page_count = WD_PAGE_COUNT;
3003 swsusp_set_page_free(pfn_to_page(pfn + i));
3007 spin_unlock_irqrestore(&zone->lock, flags);
3009 #endif /* CONFIG_PM */
3011 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3015 if (!free_pcp_prepare(page))
3018 migratetype = get_pfnblock_migratetype(page, pfn);
3019 set_pcppage_migratetype(page, migratetype);
3023 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3025 struct zone *zone = page_zone(page);
3026 struct per_cpu_pages *pcp;
3029 migratetype = get_pcppage_migratetype(page);
3030 __count_vm_event(PGFREE);
3033 * We only track unmovable, reclaimable and movable on pcp lists.
3034 * Free ISOLATE pages back to the allocator because they are being
3035 * offlined but treat HIGHATOMIC as movable pages so we can get those
3036 * areas back if necessary. Otherwise, we may have to free
3037 * excessively into the page allocator
3039 if (migratetype >= MIGRATE_PCPTYPES) {
3040 if (unlikely(is_migrate_isolate(migratetype))) {
3041 free_one_page(zone, page, pfn, 0, migratetype);
3044 migratetype = MIGRATE_MOVABLE;
3047 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3048 list_add(&page->lru, &pcp->lists[migratetype]);
3050 if (pcp->count >= pcp->high) {
3051 unsigned long batch = READ_ONCE(pcp->batch);
3052 free_pcppages_bulk(zone, batch, pcp);
3057 * Free a 0-order page
3059 void free_unref_page(struct page *page)
3061 unsigned long flags;
3062 unsigned long pfn = page_to_pfn(page);
3064 if (!free_unref_page_prepare(page, pfn))
3067 local_irq_save(flags);
3068 free_unref_page_commit(page, pfn);
3069 local_irq_restore(flags);
3073 * Free a list of 0-order pages
3075 void free_unref_page_list(struct list_head *list)
3077 struct page *page, *next;
3078 unsigned long flags, pfn;
3079 int batch_count = 0;
3081 /* Prepare pages for freeing */
3082 list_for_each_entry_safe(page, next, list, lru) {
3083 pfn = page_to_pfn(page);
3084 if (!free_unref_page_prepare(page, pfn))
3085 list_del(&page->lru);
3086 set_page_private(page, pfn);
3089 local_irq_save(flags);
3090 list_for_each_entry_safe(page, next, list, lru) {
3091 unsigned long pfn = page_private(page);
3093 set_page_private(page, 0);
3094 trace_mm_page_free_batched(page);
3095 free_unref_page_commit(page, pfn);
3098 * Guard against excessive IRQ disabled times when we get
3099 * a large list of pages to free.
3101 if (++batch_count == SWAP_CLUSTER_MAX) {
3102 local_irq_restore(flags);
3104 local_irq_save(flags);
3107 local_irq_restore(flags);
3111 * split_page takes a non-compound higher-order page, and splits it into
3112 * n (1<<order) sub-pages: page[0..n]
3113 * Each sub-page must be freed individually.
3115 * Note: this is probably too low level an operation for use in drivers.
3116 * Please consult with lkml before using this in your driver.
3118 void split_page(struct page *page, unsigned int order)
3122 VM_BUG_ON_PAGE(PageCompound(page), page);
3123 VM_BUG_ON_PAGE(!page_count(page), page);
3125 for (i = 1; i < (1 << order); i++)
3126 set_page_refcounted(page + i);
3127 split_page_owner(page, order);
3129 EXPORT_SYMBOL_GPL(split_page);
3131 int __isolate_free_page(struct page *page, unsigned int order)
3133 struct free_area *area = &page_zone(page)->free_area[order];
3134 unsigned long watermark;
3138 BUG_ON(!PageBuddy(page));
3140 zone = page_zone(page);
3141 mt = get_pageblock_migratetype(page);
3143 if (!is_migrate_isolate(mt)) {
3145 * Obey watermarks as if the page was being allocated. We can
3146 * emulate a high-order watermark check with a raised order-0
3147 * watermark, because we already know our high-order page
3150 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3151 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3154 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3157 /* Remove page from free list */
3159 del_page_from_free_area(page, area);
3162 * Set the pageblock if the isolated page is at least half of a
3165 if (order >= pageblock_order - 1) {
3166 struct page *endpage = page + (1 << order) - 1;
3167 for (; page < endpage; page += pageblock_nr_pages) {
3168 int mt = get_pageblock_migratetype(page);
3169 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3170 && !is_migrate_highatomic(mt))
3171 set_pageblock_migratetype(page,
3177 return 1UL << order;
3181 * Update NUMA hit/miss statistics
3183 * Must be called with interrupts disabled.
3185 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3188 enum numa_stat_item local_stat = NUMA_LOCAL;
3190 /* skip numa counters update if numa stats is disabled */
3191 if (!static_branch_likely(&vm_numa_stat_key))
3194 if (zone_to_nid(z) != numa_node_id())
3195 local_stat = NUMA_OTHER;
3197 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3198 __inc_numa_state(z, NUMA_HIT);
3200 __inc_numa_state(z, NUMA_MISS);
3201 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3203 __inc_numa_state(z, local_stat);
3207 /* Remove page from the per-cpu list, caller must protect the list */
3208 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3209 unsigned int alloc_flags,
3210 struct per_cpu_pages *pcp,
3211 struct list_head *list)
3216 if (list_empty(list)) {
3217 pcp->count += rmqueue_bulk(zone, 0,
3219 migratetype, alloc_flags);
3220 if (unlikely(list_empty(list)))
3224 page = list_first_entry(list, struct page, lru);
3225 list_del(&page->lru);
3227 } while (check_new_pcp(page));
3232 /* Lock and remove page from the per-cpu list */
3233 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3234 struct zone *zone, gfp_t gfp_flags,
3235 int migratetype, unsigned int alloc_flags)
3237 struct per_cpu_pages *pcp;
3238 struct list_head *list;
3240 unsigned long flags;
3242 local_irq_save(flags);
3243 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3244 list = &pcp->lists[migratetype];
3245 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3247 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3248 zone_statistics(preferred_zone, zone);
3250 local_irq_restore(flags);
3255 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3258 struct page *rmqueue(struct zone *preferred_zone,
3259 struct zone *zone, unsigned int order,
3260 gfp_t gfp_flags, unsigned int alloc_flags,
3263 unsigned long flags;
3266 if (likely(order == 0)) {
3267 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3268 migratetype, alloc_flags);
3273 * We most definitely don't want callers attempting to
3274 * allocate greater than order-1 page units with __GFP_NOFAIL.
3276 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3277 spin_lock_irqsave(&zone->lock, flags);
3281 if (alloc_flags & ALLOC_HARDER) {
3282 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3284 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3287 page = __rmqueue(zone, order, migratetype, alloc_flags);
3288 } while (page && check_new_pages(page, order));
3289 spin_unlock(&zone->lock);
3292 __mod_zone_freepage_state(zone, -(1 << order),
3293 get_pcppage_migratetype(page));
3295 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3296 zone_statistics(preferred_zone, zone);
3297 local_irq_restore(flags);
3300 /* Separate test+clear to avoid unnecessary atomics */
3301 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3302 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3303 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3306 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3310 local_irq_restore(flags);
3314 #ifdef CONFIG_FAIL_PAGE_ALLOC
3317 struct fault_attr attr;
3319 bool ignore_gfp_highmem;
3320 bool ignore_gfp_reclaim;
3322 } fail_page_alloc = {
3323 .attr = FAULT_ATTR_INITIALIZER,
3324 .ignore_gfp_reclaim = true,
3325 .ignore_gfp_highmem = true,
3329 static int __init setup_fail_page_alloc(char *str)
3331 return setup_fault_attr(&fail_page_alloc.attr, str);
3333 __setup("fail_page_alloc=", setup_fail_page_alloc);
3335 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3337 if (order < fail_page_alloc.min_order)
3339 if (gfp_mask & __GFP_NOFAIL)
3341 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3343 if (fail_page_alloc.ignore_gfp_reclaim &&
3344 (gfp_mask & __GFP_DIRECT_RECLAIM))
3347 return should_fail(&fail_page_alloc.attr, 1 << order);
3350 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3352 static int __init fail_page_alloc_debugfs(void)
3354 umode_t mode = S_IFREG | 0600;
3357 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3358 &fail_page_alloc.attr);
3360 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3361 &fail_page_alloc.ignore_gfp_reclaim);
3362 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3363 &fail_page_alloc.ignore_gfp_highmem);
3364 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3369 late_initcall(fail_page_alloc_debugfs);
3371 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3373 #else /* CONFIG_FAIL_PAGE_ALLOC */
3375 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3380 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3382 static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3384 return __should_fail_alloc_page(gfp_mask, order);
3386 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3389 * Return true if free base pages are above 'mark'. For high-order checks it
3390 * will return true of the order-0 watermark is reached and there is at least
3391 * one free page of a suitable size. Checking now avoids taking the zone lock
3392 * to check in the allocation paths if no pages are free.
3394 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3395 int classzone_idx, unsigned int alloc_flags,
3400 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3402 /* free_pages may go negative - that's OK */
3403 free_pages -= (1 << order) - 1;
3405 if (alloc_flags & ALLOC_HIGH)
3409 * If the caller does not have rights to ALLOC_HARDER then subtract
3410 * the high-atomic reserves. This will over-estimate the size of the
3411 * atomic reserve but it avoids a search.
3413 if (likely(!alloc_harder)) {
3414 free_pages -= z->nr_reserved_highatomic;
3417 * OOM victims can try even harder than normal ALLOC_HARDER
3418 * users on the grounds that it's definitely going to be in
3419 * the exit path shortly and free memory. Any allocation it
3420 * makes during the free path will be small and short-lived.
3422 if (alloc_flags & ALLOC_OOM)
3430 /* If allocation can't use CMA areas don't use free CMA pages */
3431 if (!(alloc_flags & ALLOC_CMA))
3432 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3436 * Check watermarks for an order-0 allocation request. If these
3437 * are not met, then a high-order request also cannot go ahead
3438 * even if a suitable page happened to be free.
3440 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3443 /* If this is an order-0 request then the watermark is fine */
3447 /* For a high-order request, check at least one suitable page is free */
3448 for (o = order; o < MAX_ORDER; o++) {
3449 struct free_area *area = &z->free_area[o];
3455 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3456 if (!free_area_empty(area, mt))
3461 if ((alloc_flags & ALLOC_CMA) &&
3462 !free_area_empty(area, MIGRATE_CMA)) {
3467 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3473 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3474 int classzone_idx, unsigned int alloc_flags)
3476 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3477 zone_page_state(z, NR_FREE_PAGES));
3480 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3481 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3483 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3487 /* If allocation can't use CMA areas don't use free CMA pages */
3488 if (!(alloc_flags & ALLOC_CMA))
3489 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3493 * Fast check for order-0 only. If this fails then the reserves
3494 * need to be calculated. There is a corner case where the check
3495 * passes but only the high-order atomic reserve are free. If
3496 * the caller is !atomic then it'll uselessly search the free
3497 * list. That corner case is then slower but it is harmless.
3499 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3502 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3506 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3507 unsigned long mark, int classzone_idx)
3509 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3511 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3512 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3514 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3519 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3521 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3522 node_reclaim_distance;
3524 #else /* CONFIG_NUMA */
3525 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3529 #endif /* CONFIG_NUMA */
3532 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3533 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3534 * premature use of a lower zone may cause lowmem pressure problems that
3535 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3536 * probably too small. It only makes sense to spread allocations to avoid
3537 * fragmentation between the Normal and DMA32 zones.
3539 static inline unsigned int
3540 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3542 unsigned int alloc_flags = 0;
3544 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3545 alloc_flags |= ALLOC_KSWAPD;
3547 #ifdef CONFIG_ZONE_DMA32
3551 if (zone_idx(zone) != ZONE_NORMAL)
3555 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3556 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3557 * on UMA that if Normal is populated then so is DMA32.
3559 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3560 if (nr_online_nodes > 1 && !populated_zone(--zone))
3563 alloc_flags |= ALLOC_NOFRAGMENT;
3564 #endif /* CONFIG_ZONE_DMA32 */
3569 * get_page_from_freelist goes through the zonelist trying to allocate
3572 static struct page *
3573 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3574 const struct alloc_context *ac)
3578 struct pglist_data *last_pgdat_dirty_limit = NULL;
3583 * Scan zonelist, looking for a zone with enough free.
3584 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3586 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3587 z = ac->preferred_zoneref;
3588 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3593 if (cpusets_enabled() &&
3594 (alloc_flags & ALLOC_CPUSET) &&
3595 !__cpuset_zone_allowed(zone, gfp_mask))
3598 * When allocating a page cache page for writing, we
3599 * want to get it from a node that is within its dirty
3600 * limit, such that no single node holds more than its
3601 * proportional share of globally allowed dirty pages.
3602 * The dirty limits take into account the node's
3603 * lowmem reserves and high watermark so that kswapd
3604 * should be able to balance it without having to
3605 * write pages from its LRU list.
3607 * XXX: For now, allow allocations to potentially
3608 * exceed the per-node dirty limit in the slowpath
3609 * (spread_dirty_pages unset) before going into reclaim,
3610 * which is important when on a NUMA setup the allowed
3611 * nodes are together not big enough to reach the
3612 * global limit. The proper fix for these situations
3613 * will require awareness of nodes in the
3614 * dirty-throttling and the flusher threads.
3616 if (ac->spread_dirty_pages) {
3617 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3620 if (!node_dirty_ok(zone->zone_pgdat)) {
3621 last_pgdat_dirty_limit = zone->zone_pgdat;
3626 if (no_fallback && nr_online_nodes > 1 &&
3627 zone != ac->preferred_zoneref->zone) {
3631 * If moving to a remote node, retry but allow
3632 * fragmenting fallbacks. Locality is more important
3633 * than fragmentation avoidance.
3635 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3636 if (zone_to_nid(zone) != local_nid) {
3637 alloc_flags &= ~ALLOC_NOFRAGMENT;
3642 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3643 if (!zone_watermark_fast(zone, order, mark,
3644 ac_classzone_idx(ac), alloc_flags)) {
3647 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3649 * Watermark failed for this zone, but see if we can
3650 * grow this zone if it contains deferred pages.
3652 if (static_branch_unlikely(&deferred_pages)) {
3653 if (_deferred_grow_zone(zone, order))
3657 /* Checked here to keep the fast path fast */
3658 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3659 if (alloc_flags & ALLOC_NO_WATERMARKS)
3662 if (node_reclaim_mode == 0 ||
3663 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3666 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3668 case NODE_RECLAIM_NOSCAN:
3671 case NODE_RECLAIM_FULL:
3672 /* scanned but unreclaimable */
3675 /* did we reclaim enough */
3676 if (zone_watermark_ok(zone, order, mark,
3677 ac_classzone_idx(ac), alloc_flags))
3685 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3686 gfp_mask, alloc_flags, ac->migratetype);
3688 prep_new_page(page, order, gfp_mask, alloc_flags);
3691 * If this is a high-order atomic allocation then check
3692 * if the pageblock should be reserved for the future
3694 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3695 reserve_highatomic_pageblock(page, zone, order);
3699 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3700 /* Try again if zone has deferred pages */
3701 if (static_branch_unlikely(&deferred_pages)) {
3702 if (_deferred_grow_zone(zone, order))
3710 * It's possible on a UMA machine to get through all zones that are
3711 * fragmented. If avoiding fragmentation, reset and try again.
3714 alloc_flags &= ~ALLOC_NOFRAGMENT;
3721 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3723 unsigned int filter = SHOW_MEM_FILTER_NODES;
3726 * This documents exceptions given to allocations in certain
3727 * contexts that are allowed to allocate outside current's set
3730 if (!(gfp_mask & __GFP_NOMEMALLOC))
3731 if (tsk_is_oom_victim(current) ||
3732 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3733 filter &= ~SHOW_MEM_FILTER_NODES;
3734 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3735 filter &= ~SHOW_MEM_FILTER_NODES;
3737 show_mem(filter, nodemask);
3740 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3742 struct va_format vaf;
3744 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3746 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3749 va_start(args, fmt);
3752 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3753 current->comm, &vaf, gfp_mask, &gfp_mask,
3754 nodemask_pr_args(nodemask));
3757 cpuset_print_current_mems_allowed();
3760 warn_alloc_show_mem(gfp_mask, nodemask);
3763 static inline struct page *
3764 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3765 unsigned int alloc_flags,
3766 const struct alloc_context *ac)
3770 page = get_page_from_freelist(gfp_mask, order,
3771 alloc_flags|ALLOC_CPUSET, ac);
3773 * fallback to ignore cpuset restriction if our nodes
3777 page = get_page_from_freelist(gfp_mask, order,
3783 static inline struct page *
3784 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3785 const struct alloc_context *ac, unsigned long *did_some_progress)
3787 struct oom_control oc = {
3788 .zonelist = ac->zonelist,
3789 .nodemask = ac->nodemask,
3791 .gfp_mask = gfp_mask,
3796 *did_some_progress = 0;
3799 * Acquire the oom lock. If that fails, somebody else is
3800 * making progress for us.
3802 if (!mutex_trylock(&oom_lock)) {
3803 *did_some_progress = 1;
3804 schedule_timeout_uninterruptible(1);
3809 * Go through the zonelist yet one more time, keep very high watermark
3810 * here, this is only to catch a parallel oom killing, we must fail if
3811 * we're still under heavy pressure. But make sure that this reclaim
3812 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3813 * allocation which will never fail due to oom_lock already held.
3815 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3816 ~__GFP_DIRECT_RECLAIM, order,
3817 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3821 /* Coredumps can quickly deplete all memory reserves */
3822 if (current->flags & PF_DUMPCORE)
3824 /* The OOM killer will not help higher order allocs */
3825 if (order > PAGE_ALLOC_COSTLY_ORDER)
3828 * We have already exhausted all our reclaim opportunities without any
3829 * success so it is time to admit defeat. We will skip the OOM killer
3830 * because it is very likely that the caller has a more reasonable
3831 * fallback than shooting a random task.
3833 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3835 /* The OOM killer does not needlessly kill tasks for lowmem */
3836 if (ac->high_zoneidx < ZONE_NORMAL)
3838 if (pm_suspended_storage())
3841 * XXX: GFP_NOFS allocations should rather fail than rely on
3842 * other request to make a forward progress.
3843 * We are in an unfortunate situation where out_of_memory cannot
3844 * do much for this context but let's try it to at least get
3845 * access to memory reserved if the current task is killed (see
3846 * out_of_memory). Once filesystems are ready to handle allocation
3847 * failures more gracefully we should just bail out here.
3850 /* The OOM killer may not free memory on a specific node */
3851 if (gfp_mask & __GFP_THISNODE)
3854 /* Exhausted what can be done so it's blame time */
3855 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3856 *did_some_progress = 1;
3859 * Help non-failing allocations by giving them access to memory
3862 if (gfp_mask & __GFP_NOFAIL)
3863 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3864 ALLOC_NO_WATERMARKS, ac);
3867 mutex_unlock(&oom_lock);
3872 * Maximum number of compaction retries wit a progress before OOM
3873 * killer is consider as the only way to move forward.
3875 #define MAX_COMPACT_RETRIES 16
3877 #ifdef CONFIG_COMPACTION
3878 /* Try memory compaction for high-order allocations before reclaim */
3879 static struct page *
3880 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3881 unsigned int alloc_flags, const struct alloc_context *ac,
3882 enum compact_priority prio, enum compact_result *compact_result)
3884 struct page *page = NULL;
3885 unsigned long pflags;
3886 unsigned int noreclaim_flag;
3891 psi_memstall_enter(&pflags);
3892 noreclaim_flag = memalloc_noreclaim_save();
3894 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3897 memalloc_noreclaim_restore(noreclaim_flag);
3898 psi_memstall_leave(&pflags);
3901 * At least in one zone compaction wasn't deferred or skipped, so let's
3902 * count a compaction stall
3904 count_vm_event(COMPACTSTALL);
3906 /* Prep a captured page if available */
3908 prep_new_page(page, order, gfp_mask, alloc_flags);
3910 /* Try get a page from the freelist if available */
3912 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3915 struct zone *zone = page_zone(page);
3917 zone->compact_blockskip_flush = false;
3918 compaction_defer_reset(zone, order, true);
3919 count_vm_event(COMPACTSUCCESS);
3924 * It's bad if compaction run occurs and fails. The most likely reason
3925 * is that pages exist, but not enough to satisfy watermarks.
3927 count_vm_event(COMPACTFAIL);
3935 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3936 enum compact_result compact_result,
3937 enum compact_priority *compact_priority,
3938 int *compaction_retries)
3940 int max_retries = MAX_COMPACT_RETRIES;
3943 int retries = *compaction_retries;
3944 enum compact_priority priority = *compact_priority;
3949 if (compaction_made_progress(compact_result))
3950 (*compaction_retries)++;
3953 * compaction considers all the zone as desperately out of memory
3954 * so it doesn't really make much sense to retry except when the
3955 * failure could be caused by insufficient priority
3957 if (compaction_failed(compact_result))
3958 goto check_priority;
3961 * compaction was skipped because there are not enough order-0 pages
3962 * to work with, so we retry only if it looks like reclaim can help.
3964 if (compaction_needs_reclaim(compact_result)) {
3965 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3970 * make sure the compaction wasn't deferred or didn't bail out early
3971 * due to locks contention before we declare that we should give up.
3972 * But the next retry should use a higher priority if allowed, so
3973 * we don't just keep bailing out endlessly.
3975 if (compaction_withdrawn(compact_result)) {
3976 goto check_priority;
3980 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3981 * costly ones because they are de facto nofail and invoke OOM
3982 * killer to move on while costly can fail and users are ready
3983 * to cope with that. 1/4 retries is rather arbitrary but we
3984 * would need much more detailed feedback from compaction to
3985 * make a better decision.
3987 if (order > PAGE_ALLOC_COSTLY_ORDER)
3989 if (*compaction_retries <= max_retries) {
3995 * Make sure there are attempts at the highest priority if we exhausted
3996 * all retries or failed at the lower priorities.
3999 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4000 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4002 if (*compact_priority > min_priority) {
4003 (*compact_priority)--;
4004 *compaction_retries = 0;
4008 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4012 static inline struct page *
4013 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4014 unsigned int alloc_flags, const struct alloc_context *ac,
4015 enum compact_priority prio, enum compact_result *compact_result)
4017 *compact_result = COMPACT_SKIPPED;
4022 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4023 enum compact_result compact_result,
4024 enum compact_priority *compact_priority,
4025 int *compaction_retries)
4030 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4034 * There are setups with compaction disabled which would prefer to loop
4035 * inside the allocator rather than hit the oom killer prematurely.
4036 * Let's give them a good hope and keep retrying while the order-0
4037 * watermarks are OK.
4039 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4041 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4042 ac_classzone_idx(ac), alloc_flags))
4047 #endif /* CONFIG_COMPACTION */
4049 #ifdef CONFIG_LOCKDEP
4050 static struct lockdep_map __fs_reclaim_map =
4051 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4053 static bool __need_fs_reclaim(gfp_t gfp_mask)
4055 gfp_mask = current_gfp_context(gfp_mask);
4057 /* no reclaim without waiting on it */
4058 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4061 /* this guy won't enter reclaim */
4062 if (current->flags & PF_MEMALLOC)
4065 /* We're only interested __GFP_FS allocations for now */
4066 if (!(gfp_mask & __GFP_FS))
4069 if (gfp_mask & __GFP_NOLOCKDEP)
4075 void __fs_reclaim_acquire(void)
4077 lock_map_acquire(&__fs_reclaim_map);
4080 void __fs_reclaim_release(void)
4082 lock_map_release(&__fs_reclaim_map);
4085 void fs_reclaim_acquire(gfp_t gfp_mask)
4087 if (__need_fs_reclaim(gfp_mask))
4088 __fs_reclaim_acquire();
4090 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4092 void fs_reclaim_release(gfp_t gfp_mask)
4094 if (__need_fs_reclaim(gfp_mask))
4095 __fs_reclaim_release();
4097 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4100 /* Perform direct synchronous page reclaim */
4102 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4103 const struct alloc_context *ac)
4106 unsigned int noreclaim_flag;
4107 unsigned long pflags;
4111 /* We now go into synchronous reclaim */
4112 cpuset_memory_pressure_bump();
4113 psi_memstall_enter(&pflags);
4114 fs_reclaim_acquire(gfp_mask);
4115 noreclaim_flag = memalloc_noreclaim_save();
4117 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4120 memalloc_noreclaim_restore(noreclaim_flag);
4121 fs_reclaim_release(gfp_mask);
4122 psi_memstall_leave(&pflags);
4129 /* The really slow allocator path where we enter direct reclaim */
4130 static inline struct page *
4131 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4132 unsigned int alloc_flags, const struct alloc_context *ac,
4133 unsigned long *did_some_progress)
4135 struct page *page = NULL;
4136 bool drained = false;
4138 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4139 if (unlikely(!(*did_some_progress)))
4143 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4146 * If an allocation failed after direct reclaim, it could be because
4147 * pages are pinned on the per-cpu lists or in high alloc reserves.
4148 * Shrink them them and try again
4150 if (!page && !drained) {
4151 unreserve_highatomic_pageblock(ac, false);
4152 drain_all_pages(NULL);
4160 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4161 const struct alloc_context *ac)
4165 pg_data_t *last_pgdat = NULL;
4166 enum zone_type high_zoneidx = ac->high_zoneidx;
4168 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
4170 if (last_pgdat != zone->zone_pgdat)
4171 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
4172 last_pgdat = zone->zone_pgdat;
4176 static inline unsigned int
4177 gfp_to_alloc_flags(gfp_t gfp_mask)
4179 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4181 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
4182 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4185 * The caller may dip into page reserves a bit more if the caller
4186 * cannot run direct reclaim, or if the caller has realtime scheduling
4187 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4188 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4190 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
4192 if (gfp_mask & __GFP_ATOMIC) {
4194 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4195 * if it can't schedule.
4197 if (!(gfp_mask & __GFP_NOMEMALLOC))
4198 alloc_flags |= ALLOC_HARDER;
4200 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4201 * comment for __cpuset_node_allowed().
4203 alloc_flags &= ~ALLOC_CPUSET;
4204 } else if (unlikely(rt_task(current)) && !in_interrupt())
4205 alloc_flags |= ALLOC_HARDER;
4207 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4208 alloc_flags |= ALLOC_KSWAPD;
4211 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4212 alloc_flags |= ALLOC_CMA;
4217 static bool oom_reserves_allowed(struct task_struct *tsk)
4219 if (!tsk_is_oom_victim(tsk))
4223 * !MMU doesn't have oom reaper so give access to memory reserves
4224 * only to the thread with TIF_MEMDIE set
4226 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4233 * Distinguish requests which really need access to full memory
4234 * reserves from oom victims which can live with a portion of it
4236 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4238 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4240 if (gfp_mask & __GFP_MEMALLOC)
4241 return ALLOC_NO_WATERMARKS;
4242 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4243 return ALLOC_NO_WATERMARKS;
4244 if (!in_interrupt()) {
4245 if (current->flags & PF_MEMALLOC)
4246 return ALLOC_NO_WATERMARKS;
4247 else if (oom_reserves_allowed(current))
4254 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4256 return !!__gfp_pfmemalloc_flags(gfp_mask);
4260 * Checks whether it makes sense to retry the reclaim to make a forward progress
4261 * for the given allocation request.
4263 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4264 * without success, or when we couldn't even meet the watermark if we
4265 * reclaimed all remaining pages on the LRU lists.
4267 * Returns true if a retry is viable or false to enter the oom path.
4270 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4271 struct alloc_context *ac, int alloc_flags,
4272 bool did_some_progress, int *no_progress_loops)
4279 * Costly allocations might have made a progress but this doesn't mean
4280 * their order will become available due to high fragmentation so
4281 * always increment the no progress counter for them
4283 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4284 *no_progress_loops = 0;
4286 (*no_progress_loops)++;
4289 * Make sure we converge to OOM if we cannot make any progress
4290 * several times in the row.
4292 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4293 /* Before OOM, exhaust highatomic_reserve */
4294 return unreserve_highatomic_pageblock(ac, true);
4298 * Keep reclaiming pages while there is a chance this will lead
4299 * somewhere. If none of the target zones can satisfy our allocation
4300 * request even if all reclaimable pages are considered then we are
4301 * screwed and have to go OOM.
4303 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4305 unsigned long available;
4306 unsigned long reclaimable;
4307 unsigned long min_wmark = min_wmark_pages(zone);
4310 available = reclaimable = zone_reclaimable_pages(zone);
4311 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4314 * Would the allocation succeed if we reclaimed all
4315 * reclaimable pages?
4317 wmark = __zone_watermark_ok(zone, order, min_wmark,
4318 ac_classzone_idx(ac), alloc_flags, available);
4319 trace_reclaim_retry_zone(z, order, reclaimable,
4320 available, min_wmark, *no_progress_loops, wmark);
4323 * If we didn't make any progress and have a lot of
4324 * dirty + writeback pages then we should wait for
4325 * an IO to complete to slow down the reclaim and
4326 * prevent from pre mature OOM
4328 if (!did_some_progress) {
4329 unsigned long write_pending;
4331 write_pending = zone_page_state_snapshot(zone,
4332 NR_ZONE_WRITE_PENDING);
4334 if (2 * write_pending > reclaimable) {
4335 congestion_wait(BLK_RW_ASYNC, HZ/10);
4347 * Memory allocation/reclaim might be called from a WQ context and the
4348 * current implementation of the WQ concurrency control doesn't
4349 * recognize that a particular WQ is congested if the worker thread is
4350 * looping without ever sleeping. Therefore we have to do a short sleep
4351 * here rather than calling cond_resched().
4353 if (current->flags & PF_WQ_WORKER)
4354 schedule_timeout_uninterruptible(1);
4361 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4364 * It's possible that cpuset's mems_allowed and the nodemask from
4365 * mempolicy don't intersect. This should be normally dealt with by
4366 * policy_nodemask(), but it's possible to race with cpuset update in
4367 * such a way the check therein was true, and then it became false
4368 * before we got our cpuset_mems_cookie here.
4369 * This assumes that for all allocations, ac->nodemask can come only
4370 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4371 * when it does not intersect with the cpuset restrictions) or the
4372 * caller can deal with a violated nodemask.
4374 if (cpusets_enabled() && ac->nodemask &&
4375 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4376 ac->nodemask = NULL;
4381 * When updating a task's mems_allowed or mempolicy nodemask, it is
4382 * possible to race with parallel threads in such a way that our
4383 * allocation can fail while the mask is being updated. If we are about
4384 * to fail, check if the cpuset changed during allocation and if so,
4387 if (read_mems_allowed_retry(cpuset_mems_cookie))
4393 static inline struct page *
4394 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4395 struct alloc_context *ac)
4397 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4398 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4399 struct page *page = NULL;
4400 unsigned int alloc_flags;
4401 unsigned long did_some_progress;
4402 enum compact_priority compact_priority;
4403 enum compact_result compact_result;
4404 int compaction_retries;
4405 int no_progress_loops;
4406 unsigned int cpuset_mems_cookie;
4410 * We also sanity check to catch abuse of atomic reserves being used by
4411 * callers that are not in atomic context.
4413 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4414 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4415 gfp_mask &= ~__GFP_ATOMIC;
4418 compaction_retries = 0;
4419 no_progress_loops = 0;
4420 compact_priority = DEF_COMPACT_PRIORITY;
4421 cpuset_mems_cookie = read_mems_allowed_begin();
4424 * The fast path uses conservative alloc_flags to succeed only until
4425 * kswapd needs to be woken up, and to avoid the cost of setting up
4426 * alloc_flags precisely. So we do that now.
4428 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4431 * We need to recalculate the starting point for the zonelist iterator
4432 * because we might have used different nodemask in the fast path, or
4433 * there was a cpuset modification and we are retrying - otherwise we
4434 * could end up iterating over non-eligible zones endlessly.
4436 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4437 ac->high_zoneidx, ac->nodemask);
4438 if (!ac->preferred_zoneref->zone)
4441 if (alloc_flags & ALLOC_KSWAPD)
4442 wake_all_kswapds(order, gfp_mask, ac);
4445 * The adjusted alloc_flags might result in immediate success, so try
4448 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4453 * For costly allocations, try direct compaction first, as it's likely
4454 * that we have enough base pages and don't need to reclaim. For non-
4455 * movable high-order allocations, do that as well, as compaction will
4456 * try prevent permanent fragmentation by migrating from blocks of the
4458 * Don't try this for allocations that are allowed to ignore
4459 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4461 if (can_direct_reclaim &&
4463 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4464 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4465 page = __alloc_pages_direct_compact(gfp_mask, order,
4467 INIT_COMPACT_PRIORITY,
4472 if (order >= pageblock_order && (gfp_mask & __GFP_IO) &&
4473 !(gfp_mask & __GFP_RETRY_MAYFAIL)) {
4475 * If allocating entire pageblock(s) and compaction
4476 * failed because all zones are below low watermarks
4477 * or is prohibited because it recently failed at this
4478 * order, fail immediately unless the allocator has
4479 * requested compaction and reclaim retry.
4482 * - potentially very expensive because zones are far
4483 * below their low watermarks or this is part of very
4484 * bursty high order allocations,
4485 * - not guaranteed to help because isolate_freepages()
4486 * may not iterate over freed pages as part of its
4488 * - unlikely to make entire pageblocks free on its
4491 if (compact_result == COMPACT_SKIPPED ||
4492 compact_result == COMPACT_DEFERRED)
4497 * Checks for costly allocations with __GFP_NORETRY, which
4498 * includes THP page fault allocations
4500 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4502 * If compaction is deferred for high-order allocations,
4503 * it is because sync compaction recently failed. If
4504 * this is the case and the caller requested a THP
4505 * allocation, we do not want to heavily disrupt the
4506 * system, so we fail the allocation instead of entering
4509 if (compact_result == COMPACT_DEFERRED)
4513 * Looks like reclaim/compaction is worth trying, but
4514 * sync compaction could be very expensive, so keep
4515 * using async compaction.
4517 compact_priority = INIT_COMPACT_PRIORITY;
4522 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4523 if (alloc_flags & ALLOC_KSWAPD)
4524 wake_all_kswapds(order, gfp_mask, ac);
4526 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4528 alloc_flags = reserve_flags;
4531 * Reset the nodemask and zonelist iterators if memory policies can be
4532 * ignored. These allocations are high priority and system rather than
4535 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4536 ac->nodemask = NULL;
4537 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4538 ac->high_zoneidx, ac->nodemask);
4541 /* Attempt with potentially adjusted zonelist and alloc_flags */
4542 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4546 /* Caller is not willing to reclaim, we can't balance anything */
4547 if (!can_direct_reclaim)
4550 /* Avoid recursion of direct reclaim */
4551 if (current->flags & PF_MEMALLOC)
4554 /* Try direct reclaim and then allocating */
4555 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4556 &did_some_progress);
4560 /* Try direct compaction and then allocating */
4561 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4562 compact_priority, &compact_result);
4566 /* Do not loop if specifically requested */
4567 if (gfp_mask & __GFP_NORETRY)
4571 * Do not retry costly high order allocations unless they are
4572 * __GFP_RETRY_MAYFAIL
4574 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4577 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4578 did_some_progress > 0, &no_progress_loops))
4582 * It doesn't make any sense to retry for the compaction if the order-0
4583 * reclaim is not able to make any progress because the current
4584 * implementation of the compaction depends on the sufficient amount
4585 * of free memory (see __compaction_suitable)
4587 if (did_some_progress > 0 &&
4588 should_compact_retry(ac, order, alloc_flags,
4589 compact_result, &compact_priority,
4590 &compaction_retries))
4594 /* Deal with possible cpuset update races before we start OOM killing */
4595 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4598 /* Reclaim has failed us, start killing things */
4599 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4603 /* Avoid allocations with no watermarks from looping endlessly */
4604 if (tsk_is_oom_victim(current) &&
4605 (alloc_flags == ALLOC_OOM ||
4606 (gfp_mask & __GFP_NOMEMALLOC)))
4609 /* Retry as long as the OOM killer is making progress */
4610 if (did_some_progress) {
4611 no_progress_loops = 0;
4616 /* Deal with possible cpuset update races before we fail */
4617 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4621 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4624 if (gfp_mask & __GFP_NOFAIL) {
4626 * All existing users of the __GFP_NOFAIL are blockable, so warn
4627 * of any new users that actually require GFP_NOWAIT
4629 if (WARN_ON_ONCE(!can_direct_reclaim))
4633 * PF_MEMALLOC request from this context is rather bizarre
4634 * because we cannot reclaim anything and only can loop waiting
4635 * for somebody to do a work for us
4637 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4640 * non failing costly orders are a hard requirement which we
4641 * are not prepared for much so let's warn about these users
4642 * so that we can identify them and convert them to something
4645 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4648 * Help non-failing allocations by giving them access to memory
4649 * reserves but do not use ALLOC_NO_WATERMARKS because this
4650 * could deplete whole memory reserves which would just make
4651 * the situation worse
4653 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4661 warn_alloc(gfp_mask, ac->nodemask,
4662 "page allocation failure: order:%u", order);
4667 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4668 int preferred_nid, nodemask_t *nodemask,
4669 struct alloc_context *ac, gfp_t *alloc_mask,
4670 unsigned int *alloc_flags)
4672 ac->high_zoneidx = gfp_zone(gfp_mask);
4673 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4674 ac->nodemask = nodemask;
4675 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4677 if (cpusets_enabled()) {
4678 *alloc_mask |= __GFP_HARDWALL;
4680 ac->nodemask = &cpuset_current_mems_allowed;
4682 *alloc_flags |= ALLOC_CPUSET;
4685 fs_reclaim_acquire(gfp_mask);
4686 fs_reclaim_release(gfp_mask);
4688 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4690 if (should_fail_alloc_page(gfp_mask, order))
4693 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4694 *alloc_flags |= ALLOC_CMA;
4699 /* Determine whether to spread dirty pages and what the first usable zone */
4700 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4702 /* Dirty zone balancing only done in the fast path */
4703 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4706 * The preferred zone is used for statistics but crucially it is
4707 * also used as the starting point for the zonelist iterator. It
4708 * may get reset for allocations that ignore memory policies.
4710 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4711 ac->high_zoneidx, ac->nodemask);
4715 * This is the 'heart' of the zoned buddy allocator.
4718 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4719 nodemask_t *nodemask)
4722 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4723 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4724 struct alloc_context ac = { };
4727 * There are several places where we assume that the order value is sane
4728 * so bail out early if the request is out of bound.
4730 if (unlikely(order >= MAX_ORDER)) {
4731 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4735 gfp_mask &= gfp_allowed_mask;
4736 alloc_mask = gfp_mask;
4737 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4740 finalise_ac(gfp_mask, &ac);
4743 * Forbid the first pass from falling back to types that fragment
4744 * memory until all local zones are considered.
4746 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4748 /* First allocation attempt */
4749 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4754 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4755 * resp. GFP_NOIO which has to be inherited for all allocation requests
4756 * from a particular context which has been marked by
4757 * memalloc_no{fs,io}_{save,restore}.
4759 alloc_mask = current_gfp_context(gfp_mask);
4760 ac.spread_dirty_pages = false;
4763 * Restore the original nodemask if it was potentially replaced with
4764 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4766 if (unlikely(ac.nodemask != nodemask))
4767 ac.nodemask = nodemask;
4769 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4772 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4773 unlikely(__memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4774 __free_pages(page, order);
4778 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4782 EXPORT_SYMBOL(__alloc_pages_nodemask);
4785 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4786 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4787 * you need to access high mem.
4789 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4793 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4796 return (unsigned long) page_address(page);
4798 EXPORT_SYMBOL(__get_free_pages);
4800 unsigned long get_zeroed_page(gfp_t gfp_mask)
4802 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4804 EXPORT_SYMBOL(get_zeroed_page);
4806 static inline void free_the_page(struct page *page, unsigned int order)
4808 if (order == 0) /* Via pcp? */
4809 free_unref_page(page);
4811 __free_pages_ok(page, order);
4814 void __free_pages(struct page *page, unsigned int order)
4816 if (put_page_testzero(page))
4817 free_the_page(page, order);
4819 EXPORT_SYMBOL(__free_pages);
4821 void free_pages(unsigned long addr, unsigned int order)
4824 VM_BUG_ON(!virt_addr_valid((void *)addr));
4825 __free_pages(virt_to_page((void *)addr), order);
4829 EXPORT_SYMBOL(free_pages);
4833 * An arbitrary-length arbitrary-offset area of memory which resides
4834 * within a 0 or higher order page. Multiple fragments within that page
4835 * are individually refcounted, in the page's reference counter.
4837 * The page_frag functions below provide a simple allocation framework for
4838 * page fragments. This is used by the network stack and network device
4839 * drivers to provide a backing region of memory for use as either an
4840 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4842 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4845 struct page *page = NULL;
4846 gfp_t gfp = gfp_mask;
4848 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4849 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4851 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4852 PAGE_FRAG_CACHE_MAX_ORDER);
4853 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4855 if (unlikely(!page))
4856 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4858 nc->va = page ? page_address(page) : NULL;
4863 void __page_frag_cache_drain(struct page *page, unsigned int count)
4865 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4867 if (page_ref_sub_and_test(page, count))
4868 free_the_page(page, compound_order(page));
4870 EXPORT_SYMBOL(__page_frag_cache_drain);
4872 void *page_frag_alloc(struct page_frag_cache *nc,
4873 unsigned int fragsz, gfp_t gfp_mask)
4875 unsigned int size = PAGE_SIZE;
4879 if (unlikely(!nc->va)) {
4881 page = __page_frag_cache_refill(nc, gfp_mask);
4885 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4886 /* if size can vary use size else just use PAGE_SIZE */
4889 /* Even if we own the page, we do not use atomic_set().
4890 * This would break get_page_unless_zero() users.
4892 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4894 /* reset page count bias and offset to start of new frag */
4895 nc->pfmemalloc = page_is_pfmemalloc(page);
4896 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4900 offset = nc->offset - fragsz;
4901 if (unlikely(offset < 0)) {
4902 page = virt_to_page(nc->va);
4904 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4907 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4908 /* if size can vary use size else just use PAGE_SIZE */
4911 /* OK, page count is 0, we can safely set it */
4912 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4914 /* reset page count bias and offset to start of new frag */
4915 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4916 offset = size - fragsz;
4920 nc->offset = offset;
4922 return nc->va + offset;
4924 EXPORT_SYMBOL(page_frag_alloc);
4927 * Frees a page fragment allocated out of either a compound or order 0 page.
4929 void page_frag_free(void *addr)
4931 struct page *page = virt_to_head_page(addr);
4933 if (unlikely(put_page_testzero(page)))
4934 free_the_page(page, compound_order(page));
4936 EXPORT_SYMBOL(page_frag_free);
4938 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4942 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4943 unsigned long used = addr + PAGE_ALIGN(size);
4945 split_page(virt_to_page((void *)addr), order);
4946 while (used < alloc_end) {
4951 return (void *)addr;
4955 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4956 * @size: the number of bytes to allocate
4957 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4959 * This function is similar to alloc_pages(), except that it allocates the
4960 * minimum number of pages to satisfy the request. alloc_pages() can only
4961 * allocate memory in power-of-two pages.
4963 * This function is also limited by MAX_ORDER.
4965 * Memory allocated by this function must be released by free_pages_exact().
4967 * Return: pointer to the allocated area or %NULL in case of error.
4969 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4971 unsigned int order = get_order(size);
4974 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
4975 gfp_mask &= ~__GFP_COMP;
4977 addr = __get_free_pages(gfp_mask, order);
4978 return make_alloc_exact(addr, order, size);
4980 EXPORT_SYMBOL(alloc_pages_exact);
4983 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4985 * @nid: the preferred node ID where memory should be allocated
4986 * @size: the number of bytes to allocate
4987 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4989 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4992 * Return: pointer to the allocated area or %NULL in case of error.
4994 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4996 unsigned int order = get_order(size);
4999 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5000 gfp_mask &= ~__GFP_COMP;
5002 p = alloc_pages_node(nid, gfp_mask, order);
5005 return make_alloc_exact((unsigned long)page_address(p), order, size);
5009 * free_pages_exact - release memory allocated via alloc_pages_exact()
5010 * @virt: the value returned by alloc_pages_exact.
5011 * @size: size of allocation, same value as passed to alloc_pages_exact().
5013 * Release the memory allocated by a previous call to alloc_pages_exact.
5015 void free_pages_exact(void *virt, size_t size)
5017 unsigned long addr = (unsigned long)virt;
5018 unsigned long end = addr + PAGE_ALIGN(size);
5020 while (addr < end) {
5025 EXPORT_SYMBOL(free_pages_exact);
5028 * nr_free_zone_pages - count number of pages beyond high watermark
5029 * @offset: The zone index of the highest zone
5031 * nr_free_zone_pages() counts the number of pages which are beyond the
5032 * high watermark within all zones at or below a given zone index. For each
5033 * zone, the number of pages is calculated as:
5035 * nr_free_zone_pages = managed_pages - high_pages
5037 * Return: number of pages beyond high watermark.
5039 static unsigned long nr_free_zone_pages(int offset)
5044 /* Just pick one node, since fallback list is circular */
5045 unsigned long sum = 0;
5047 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5049 for_each_zone_zonelist(zone, z, zonelist, offset) {
5050 unsigned long size = zone_managed_pages(zone);
5051 unsigned long high = high_wmark_pages(zone);
5060 * nr_free_buffer_pages - count number of pages beyond high watermark
5062 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5063 * watermark within ZONE_DMA and ZONE_NORMAL.
5065 * Return: number of pages beyond high watermark within ZONE_DMA and
5068 unsigned long nr_free_buffer_pages(void)
5070 return nr_free_zone_pages(gfp_zone(GFP_USER));
5072 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5075 * nr_free_pagecache_pages - count number of pages beyond high watermark
5077 * nr_free_pagecache_pages() counts the number of pages which are beyond the
5078 * high watermark within all zones.
5080 * Return: number of pages beyond high watermark within all zones.
5082 unsigned long nr_free_pagecache_pages(void)
5084 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5087 static inline void show_node(struct zone *zone)
5089 if (IS_ENABLED(CONFIG_NUMA))
5090 printk("Node %d ", zone_to_nid(zone));
5093 long si_mem_available(void)
5096 unsigned long pagecache;
5097 unsigned long wmark_low = 0;
5098 unsigned long pages[NR_LRU_LISTS];
5099 unsigned long reclaimable;
5103 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5104 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5107 wmark_low += low_wmark_pages(zone);
5110 * Estimate the amount of memory available for userspace allocations,
5111 * without causing swapping.
5113 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5116 * Not all the page cache can be freed, otherwise the system will
5117 * start swapping. Assume at least half of the page cache, or the
5118 * low watermark worth of cache, needs to stay.
5120 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5121 pagecache -= min(pagecache / 2, wmark_low);
5122 available += pagecache;
5125 * Part of the reclaimable slab and other kernel memory consists of
5126 * items that are in use, and cannot be freed. Cap this estimate at the
5129 reclaimable = global_node_page_state(NR_SLAB_RECLAIMABLE) +
5130 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5131 available += reclaimable - min(reclaimable / 2, wmark_low);
5137 EXPORT_SYMBOL_GPL(si_mem_available);
5139 void si_meminfo(struct sysinfo *val)
5141 val->totalram = totalram_pages();
5142 val->sharedram = global_node_page_state(NR_SHMEM);
5143 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5144 val->bufferram = nr_blockdev_pages();
5145 val->totalhigh = totalhigh_pages();
5146 val->freehigh = nr_free_highpages();
5147 val->mem_unit = PAGE_SIZE;
5150 EXPORT_SYMBOL(si_meminfo);
5153 void si_meminfo_node(struct sysinfo *val, int nid)
5155 int zone_type; /* needs to be signed */
5156 unsigned long managed_pages = 0;
5157 unsigned long managed_highpages = 0;
5158 unsigned long free_highpages = 0;
5159 pg_data_t *pgdat = NODE_DATA(nid);
5161 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5162 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5163 val->totalram = managed_pages;
5164 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5165 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5166 #ifdef CONFIG_HIGHMEM
5167 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5168 struct zone *zone = &pgdat->node_zones[zone_type];
5170 if (is_highmem(zone)) {
5171 managed_highpages += zone_managed_pages(zone);
5172 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5175 val->totalhigh = managed_highpages;
5176 val->freehigh = free_highpages;
5178 val->totalhigh = managed_highpages;
5179 val->freehigh = free_highpages;
5181 val->mem_unit = PAGE_SIZE;
5186 * Determine whether the node should be displayed or not, depending on whether
5187 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5189 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5191 if (!(flags & SHOW_MEM_FILTER_NODES))
5195 * no node mask - aka implicit memory numa policy. Do not bother with
5196 * the synchronization - read_mems_allowed_begin - because we do not
5197 * have to be precise here.
5200 nodemask = &cpuset_current_mems_allowed;
5202 return !node_isset(nid, *nodemask);
5205 #define K(x) ((x) << (PAGE_SHIFT-10))
5207 static void show_migration_types(unsigned char type)
5209 static const char types[MIGRATE_TYPES] = {
5210 [MIGRATE_UNMOVABLE] = 'U',
5211 [MIGRATE_MOVABLE] = 'M',
5212 [MIGRATE_RECLAIMABLE] = 'E',
5213 [MIGRATE_HIGHATOMIC] = 'H',
5215 [MIGRATE_CMA] = 'C',
5217 #ifdef CONFIG_MEMORY_ISOLATION
5218 [MIGRATE_ISOLATE] = 'I',
5221 char tmp[MIGRATE_TYPES + 1];
5225 for (i = 0; i < MIGRATE_TYPES; i++) {
5226 if (type & (1 << i))
5231 printk(KERN_CONT "(%s) ", tmp);
5235 * Show free area list (used inside shift_scroll-lock stuff)
5236 * We also calculate the percentage fragmentation. We do this by counting the
5237 * memory on each free list with the exception of the first item on the list.
5240 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5243 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5245 unsigned long free_pcp = 0;
5250 for_each_populated_zone(zone) {
5251 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5254 for_each_online_cpu(cpu)
5255 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5258 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5259 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5260 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
5261 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5262 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5263 " free:%lu free_pcp:%lu free_cma:%lu\n",
5264 global_node_page_state(NR_ACTIVE_ANON),
5265 global_node_page_state(NR_INACTIVE_ANON),
5266 global_node_page_state(NR_ISOLATED_ANON),
5267 global_node_page_state(NR_ACTIVE_FILE),
5268 global_node_page_state(NR_INACTIVE_FILE),
5269 global_node_page_state(NR_ISOLATED_FILE),
5270 global_node_page_state(NR_UNEVICTABLE),
5271 global_node_page_state(NR_FILE_DIRTY),
5272 global_node_page_state(NR_WRITEBACK),
5273 global_node_page_state(NR_UNSTABLE_NFS),
5274 global_node_page_state(NR_SLAB_RECLAIMABLE),
5275 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
5276 global_node_page_state(NR_FILE_MAPPED),
5277 global_node_page_state(NR_SHMEM),
5278 global_zone_page_state(NR_PAGETABLE),
5279 global_zone_page_state(NR_BOUNCE),
5280 global_zone_page_state(NR_FREE_PAGES),
5282 global_zone_page_state(NR_FREE_CMA_PAGES));
5284 for_each_online_pgdat(pgdat) {
5285 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5289 " active_anon:%lukB"
5290 " inactive_anon:%lukB"
5291 " active_file:%lukB"
5292 " inactive_file:%lukB"
5293 " unevictable:%lukB"
5294 " isolated(anon):%lukB"
5295 " isolated(file):%lukB"
5300 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5302 " shmem_pmdmapped: %lukB"
5305 " writeback_tmp:%lukB"
5307 " all_unreclaimable? %s"
5310 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5311 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5312 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5313 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5314 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5315 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5316 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5317 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5318 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5319 K(node_page_state(pgdat, NR_WRITEBACK)),
5320 K(node_page_state(pgdat, NR_SHMEM)),
5321 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5322 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5323 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5325 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5327 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5328 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
5329 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5333 for_each_populated_zone(zone) {
5336 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5340 for_each_online_cpu(cpu)
5341 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5350 " active_anon:%lukB"
5351 " inactive_anon:%lukB"
5352 " active_file:%lukB"
5353 " inactive_file:%lukB"
5354 " unevictable:%lukB"
5355 " writepending:%lukB"
5359 " kernel_stack:%lukB"
5367 K(zone_page_state(zone, NR_FREE_PAGES)),
5368 K(min_wmark_pages(zone)),
5369 K(low_wmark_pages(zone)),
5370 K(high_wmark_pages(zone)),
5371 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5372 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5373 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5374 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5375 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5376 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5377 K(zone->present_pages),
5378 K(zone_managed_pages(zone)),
5379 K(zone_page_state(zone, NR_MLOCK)),
5380 zone_page_state(zone, NR_KERNEL_STACK_KB),
5381 K(zone_page_state(zone, NR_PAGETABLE)),
5382 K(zone_page_state(zone, NR_BOUNCE)),
5384 K(this_cpu_read(zone->pageset->pcp.count)),
5385 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5386 printk("lowmem_reserve[]:");
5387 for (i = 0; i < MAX_NR_ZONES; i++)
5388 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5389 printk(KERN_CONT "\n");
5392 for_each_populated_zone(zone) {
5394 unsigned long nr[MAX_ORDER], flags, total = 0;
5395 unsigned char types[MAX_ORDER];
5397 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5400 printk(KERN_CONT "%s: ", zone->name);
5402 spin_lock_irqsave(&zone->lock, flags);
5403 for (order = 0; order < MAX_ORDER; order++) {
5404 struct free_area *area = &zone->free_area[order];
5407 nr[order] = area->nr_free;
5408 total += nr[order] << order;
5411 for (type = 0; type < MIGRATE_TYPES; type++) {
5412 if (!free_area_empty(area, type))
5413 types[order] |= 1 << type;
5416 spin_unlock_irqrestore(&zone->lock, flags);
5417 for (order = 0; order < MAX_ORDER; order++) {
5418 printk(KERN_CONT "%lu*%lukB ",
5419 nr[order], K(1UL) << order);
5421 show_migration_types(types[order]);
5423 printk(KERN_CONT "= %lukB\n", K(total));
5426 hugetlb_show_meminfo();
5428 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5430 show_swap_cache_info();
5433 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5435 zoneref->zone = zone;
5436 zoneref->zone_idx = zone_idx(zone);
5440 * Builds allocation fallback zone lists.
5442 * Add all populated zones of a node to the zonelist.
5444 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5447 enum zone_type zone_type = MAX_NR_ZONES;
5452 zone = pgdat->node_zones + zone_type;
5453 if (managed_zone(zone)) {
5454 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5455 check_highest_zone(zone_type);
5457 } while (zone_type);
5464 static int __parse_numa_zonelist_order(char *s)
5467 * We used to support different zonlists modes but they turned
5468 * out to be just not useful. Let's keep the warning in place
5469 * if somebody still use the cmd line parameter so that we do
5470 * not fail it silently
5472 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5473 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5479 static __init int setup_numa_zonelist_order(char *s)
5484 return __parse_numa_zonelist_order(s);
5486 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5488 char numa_zonelist_order[] = "Node";
5491 * sysctl handler for numa_zonelist_order
5493 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5494 void __user *buffer, size_t *length,
5501 return proc_dostring(table, write, buffer, length, ppos);
5502 str = memdup_user_nul(buffer, 16);
5504 return PTR_ERR(str);
5506 ret = __parse_numa_zonelist_order(str);
5512 #define MAX_NODE_LOAD (nr_online_nodes)
5513 static int node_load[MAX_NUMNODES];
5516 * find_next_best_node - find the next node that should appear in a given node's fallback list
5517 * @node: node whose fallback list we're appending
5518 * @used_node_mask: nodemask_t of already used nodes
5520 * We use a number of factors to determine which is the next node that should
5521 * appear on a given node's fallback list. The node should not have appeared
5522 * already in @node's fallback list, and it should be the next closest node
5523 * according to the distance array (which contains arbitrary distance values
5524 * from each node to each node in the system), and should also prefer nodes
5525 * with no CPUs, since presumably they'll have very little allocation pressure
5526 * on them otherwise.
5528 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5530 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5533 int min_val = INT_MAX;
5534 int best_node = NUMA_NO_NODE;
5535 const struct cpumask *tmp = cpumask_of_node(0);
5537 /* Use the local node if we haven't already */
5538 if (!node_isset(node, *used_node_mask)) {
5539 node_set(node, *used_node_mask);
5543 for_each_node_state(n, N_MEMORY) {
5545 /* Don't want a node to appear more than once */
5546 if (node_isset(n, *used_node_mask))
5549 /* Use the distance array to find the distance */
5550 val = node_distance(node, n);
5552 /* Penalize nodes under us ("prefer the next node") */
5555 /* Give preference to headless and unused nodes */
5556 tmp = cpumask_of_node(n);
5557 if (!cpumask_empty(tmp))
5558 val += PENALTY_FOR_NODE_WITH_CPUS;
5560 /* Slight preference for less loaded node */
5561 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5562 val += node_load[n];
5564 if (val < min_val) {
5571 node_set(best_node, *used_node_mask);
5578 * Build zonelists ordered by node and zones within node.
5579 * This results in maximum locality--normal zone overflows into local
5580 * DMA zone, if any--but risks exhausting DMA zone.
5582 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5585 struct zoneref *zonerefs;
5588 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5590 for (i = 0; i < nr_nodes; i++) {
5593 pg_data_t *node = NODE_DATA(node_order[i]);
5595 nr_zones = build_zonerefs_node(node, zonerefs);
5596 zonerefs += nr_zones;
5598 zonerefs->zone = NULL;
5599 zonerefs->zone_idx = 0;
5603 * Build gfp_thisnode zonelists
5605 static void build_thisnode_zonelists(pg_data_t *pgdat)
5607 struct zoneref *zonerefs;
5610 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5611 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5612 zonerefs += nr_zones;
5613 zonerefs->zone = NULL;
5614 zonerefs->zone_idx = 0;
5618 * Build zonelists ordered by zone and nodes within zones.
5619 * This results in conserving DMA zone[s] until all Normal memory is
5620 * exhausted, but results in overflowing to remote node while memory
5621 * may still exist in local DMA zone.
5624 static void build_zonelists(pg_data_t *pgdat)
5626 static int node_order[MAX_NUMNODES];
5627 int node, load, nr_nodes = 0;
5628 nodemask_t used_mask;
5629 int local_node, prev_node;
5631 /* NUMA-aware ordering of nodes */
5632 local_node = pgdat->node_id;
5633 load = nr_online_nodes;
5634 prev_node = local_node;
5635 nodes_clear(used_mask);
5637 memset(node_order, 0, sizeof(node_order));
5638 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5640 * We don't want to pressure a particular node.
5641 * So adding penalty to the first node in same
5642 * distance group to make it round-robin.
5644 if (node_distance(local_node, node) !=
5645 node_distance(local_node, prev_node))
5646 node_load[node] = load;
5648 node_order[nr_nodes++] = node;
5653 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5654 build_thisnode_zonelists(pgdat);
5657 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5659 * Return node id of node used for "local" allocations.
5660 * I.e., first node id of first zone in arg node's generic zonelist.
5661 * Used for initializing percpu 'numa_mem', which is used primarily
5662 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5664 int local_memory_node(int node)
5668 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5669 gfp_zone(GFP_KERNEL),
5671 return zone_to_nid(z->zone);
5675 static void setup_min_unmapped_ratio(void);
5676 static void setup_min_slab_ratio(void);
5677 #else /* CONFIG_NUMA */
5679 static void build_zonelists(pg_data_t *pgdat)
5681 int node, local_node;
5682 struct zoneref *zonerefs;
5685 local_node = pgdat->node_id;
5687 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5688 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5689 zonerefs += nr_zones;
5692 * Now we build the zonelist so that it contains the zones
5693 * of all the other nodes.
5694 * We don't want to pressure a particular node, so when
5695 * building the zones for node N, we make sure that the
5696 * zones coming right after the local ones are those from
5697 * node N+1 (modulo N)
5699 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5700 if (!node_online(node))
5702 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5703 zonerefs += nr_zones;
5705 for (node = 0; node < local_node; node++) {
5706 if (!node_online(node))
5708 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5709 zonerefs += nr_zones;
5712 zonerefs->zone = NULL;
5713 zonerefs->zone_idx = 0;
5716 #endif /* CONFIG_NUMA */
5719 * Boot pageset table. One per cpu which is going to be used for all
5720 * zones and all nodes. The parameters will be set in such a way
5721 * that an item put on a list will immediately be handed over to
5722 * the buddy list. This is safe since pageset manipulation is done
5723 * with interrupts disabled.
5725 * The boot_pagesets must be kept even after bootup is complete for
5726 * unused processors and/or zones. They do play a role for bootstrapping
5727 * hotplugged processors.
5729 * zoneinfo_show() and maybe other functions do
5730 * not check if the processor is online before following the pageset pointer.
5731 * Other parts of the kernel may not check if the zone is available.
5733 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5734 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5735 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5737 static void __build_all_zonelists(void *data)
5740 int __maybe_unused cpu;
5741 pg_data_t *self = data;
5742 static DEFINE_SPINLOCK(lock);
5747 memset(node_load, 0, sizeof(node_load));
5751 * This node is hotadded and no memory is yet present. So just
5752 * building zonelists is fine - no need to touch other nodes.
5754 if (self && !node_online(self->node_id)) {
5755 build_zonelists(self);
5757 for_each_online_node(nid) {
5758 pg_data_t *pgdat = NODE_DATA(nid);
5760 build_zonelists(pgdat);
5763 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5765 * We now know the "local memory node" for each node--
5766 * i.e., the node of the first zone in the generic zonelist.
5767 * Set up numa_mem percpu variable for on-line cpus. During
5768 * boot, only the boot cpu should be on-line; we'll init the
5769 * secondary cpus' numa_mem as they come on-line. During
5770 * node/memory hotplug, we'll fixup all on-line cpus.
5772 for_each_online_cpu(cpu)
5773 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5780 static noinline void __init
5781 build_all_zonelists_init(void)
5785 __build_all_zonelists(NULL);
5788 * Initialize the boot_pagesets that are going to be used
5789 * for bootstrapping processors. The real pagesets for
5790 * each zone will be allocated later when the per cpu
5791 * allocator is available.
5793 * boot_pagesets are used also for bootstrapping offline
5794 * cpus if the system is already booted because the pagesets
5795 * are needed to initialize allocators on a specific cpu too.
5796 * F.e. the percpu allocator needs the page allocator which
5797 * needs the percpu allocator in order to allocate its pagesets
5798 * (a chicken-egg dilemma).
5800 for_each_possible_cpu(cpu)
5801 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5803 mminit_verify_zonelist();
5804 cpuset_init_current_mems_allowed();
5808 * unless system_state == SYSTEM_BOOTING.
5810 * __ref due to call of __init annotated helper build_all_zonelists_init
5811 * [protected by SYSTEM_BOOTING].
5813 void __ref build_all_zonelists(pg_data_t *pgdat)
5815 if (system_state == SYSTEM_BOOTING) {
5816 build_all_zonelists_init();
5818 __build_all_zonelists(pgdat);
5819 /* cpuset refresh routine should be here */
5821 vm_total_pages = nr_free_pagecache_pages();
5823 * Disable grouping by mobility if the number of pages in the
5824 * system is too low to allow the mechanism to work. It would be
5825 * more accurate, but expensive to check per-zone. This check is
5826 * made on memory-hotadd so a system can start with mobility
5827 * disabled and enable it later
5829 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5830 page_group_by_mobility_disabled = 1;
5832 page_group_by_mobility_disabled = 0;
5834 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5836 page_group_by_mobility_disabled ? "off" : "on",
5839 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5843 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5844 static bool __meminit
5845 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5847 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5848 static struct memblock_region *r;
5850 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5851 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5852 for_each_memblock(memory, r) {
5853 if (*pfn < memblock_region_memory_end_pfn(r))
5857 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5858 memblock_is_mirror(r)) {
5859 *pfn = memblock_region_memory_end_pfn(r);
5868 * Initially all pages are reserved - free ones are freed
5869 * up by memblock_free_all() once the early boot process is
5870 * done. Non-atomic initialization, single-pass.
5872 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5873 unsigned long start_pfn, enum memmap_context context,
5874 struct vmem_altmap *altmap)
5876 unsigned long pfn, end_pfn = start_pfn + size;
5879 if (highest_memmap_pfn < end_pfn - 1)
5880 highest_memmap_pfn = end_pfn - 1;
5882 #ifdef CONFIG_ZONE_DEVICE
5884 * Honor reservation requested by the driver for this ZONE_DEVICE
5885 * memory. We limit the total number of pages to initialize to just
5886 * those that might contain the memory mapping. We will defer the
5887 * ZONE_DEVICE page initialization until after we have released
5890 if (zone == ZONE_DEVICE) {
5894 if (start_pfn == altmap->base_pfn)
5895 start_pfn += altmap->reserve;
5896 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5900 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5902 * There can be holes in boot-time mem_map[]s handed to this
5903 * function. They do not exist on hotplugged memory.
5905 if (context == MEMMAP_EARLY) {
5906 if (!early_pfn_valid(pfn))
5908 if (!early_pfn_in_nid(pfn, nid))
5910 if (overlap_memmap_init(zone, &pfn))
5912 if (defer_init(nid, pfn, end_pfn))
5916 page = pfn_to_page(pfn);
5917 __init_single_page(page, pfn, zone, nid);
5918 if (context == MEMMAP_HOTPLUG)
5919 __SetPageReserved(page);
5922 * Mark the block movable so that blocks are reserved for
5923 * movable at startup. This will force kernel allocations
5924 * to reserve their blocks rather than leaking throughout
5925 * the address space during boot when many long-lived
5926 * kernel allocations are made.
5928 * bitmap is created for zone's valid pfn range. but memmap
5929 * can be created for invalid pages (for alignment)
5930 * check here not to call set_pageblock_migratetype() against
5933 if (!(pfn & (pageblock_nr_pages - 1))) {
5934 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5940 #ifdef CONFIG_ZONE_DEVICE
5941 void __ref memmap_init_zone_device(struct zone *zone,
5942 unsigned long start_pfn,
5944 struct dev_pagemap *pgmap)
5946 unsigned long pfn, end_pfn = start_pfn + size;
5947 struct pglist_data *pgdat = zone->zone_pgdat;
5948 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
5949 unsigned long zone_idx = zone_idx(zone);
5950 unsigned long start = jiffies;
5951 int nid = pgdat->node_id;
5953 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
5957 * The call to memmap_init_zone should have already taken care
5958 * of the pages reserved for the memmap, so we can just jump to
5959 * the end of that region and start processing the device pages.
5962 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5963 size = end_pfn - start_pfn;
5966 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5967 struct page *page = pfn_to_page(pfn);
5969 __init_single_page(page, pfn, zone_idx, nid);
5972 * Mark page reserved as it will need to wait for onlining
5973 * phase for it to be fully associated with a zone.
5975 * We can use the non-atomic __set_bit operation for setting
5976 * the flag as we are still initializing the pages.
5978 __SetPageReserved(page);
5981 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
5982 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
5983 * ever freed or placed on a driver-private list.
5985 page->pgmap = pgmap;
5986 page->zone_device_data = NULL;
5989 * Mark the block movable so that blocks are reserved for
5990 * movable at startup. This will force kernel allocations
5991 * to reserve their blocks rather than leaking throughout
5992 * the address space during boot when many long-lived
5993 * kernel allocations are made.
5995 * bitmap is created for zone's valid pfn range. but memmap
5996 * can be created for invalid pages (for alignment)
5997 * check here not to call set_pageblock_migratetype() against
6000 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
6001 * because this is done early in section_activate()
6003 if (!(pfn & (pageblock_nr_pages - 1))) {
6004 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6009 pr_info("%s initialised %lu pages in %ums\n", __func__,
6010 size, jiffies_to_msecs(jiffies - start));
6014 static void __meminit zone_init_free_lists(struct zone *zone)
6016 unsigned int order, t;
6017 for_each_migratetype_order(order, t) {
6018 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6019 zone->free_area[order].nr_free = 0;
6023 void __meminit __weak memmap_init(unsigned long size, int nid,
6024 unsigned long zone, unsigned long start_pfn)
6026 memmap_init_zone(size, nid, zone, start_pfn, MEMMAP_EARLY, NULL);
6029 static int zone_batchsize(struct zone *zone)
6035 * The per-cpu-pages pools are set to around 1000th of the
6038 batch = zone_managed_pages(zone) / 1024;
6039 /* But no more than a meg. */
6040 if (batch * PAGE_SIZE > 1024 * 1024)
6041 batch = (1024 * 1024) / PAGE_SIZE;
6042 batch /= 4; /* We effectively *= 4 below */
6047 * Clamp the batch to a 2^n - 1 value. Having a power
6048 * of 2 value was found to be more likely to have
6049 * suboptimal cache aliasing properties in some cases.
6051 * For example if 2 tasks are alternately allocating
6052 * batches of pages, one task can end up with a lot
6053 * of pages of one half of the possible page colors
6054 * and the other with pages of the other colors.
6056 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6061 /* The deferral and batching of frees should be suppressed under NOMMU
6064 * The problem is that NOMMU needs to be able to allocate large chunks
6065 * of contiguous memory as there's no hardware page translation to
6066 * assemble apparent contiguous memory from discontiguous pages.
6068 * Queueing large contiguous runs of pages for batching, however,
6069 * causes the pages to actually be freed in smaller chunks. As there
6070 * can be a significant delay between the individual batches being
6071 * recycled, this leads to the once large chunks of space being
6072 * fragmented and becoming unavailable for high-order allocations.
6079 * pcp->high and pcp->batch values are related and dependent on one another:
6080 * ->batch must never be higher then ->high.
6081 * The following function updates them in a safe manner without read side
6084 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6085 * those fields changing asynchronously (acording the the above rule).
6087 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6088 * outside of boot time (or some other assurance that no concurrent updaters
6091 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6092 unsigned long batch)
6094 /* start with a fail safe value for batch */
6098 /* Update high, then batch, in order */
6105 /* a companion to pageset_set_high() */
6106 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6108 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6111 static void pageset_init(struct per_cpu_pageset *p)
6113 struct per_cpu_pages *pcp;
6116 memset(p, 0, sizeof(*p));
6119 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6120 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6123 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6126 pageset_set_batch(p, batch);
6130 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6131 * to the value high for the pageset p.
6133 static void pageset_set_high(struct per_cpu_pageset *p,
6136 unsigned long batch = max(1UL, high / 4);
6137 if ((high / 4) > (PAGE_SHIFT * 8))
6138 batch = PAGE_SHIFT * 8;
6140 pageset_update(&p->pcp, high, batch);
6143 static void pageset_set_high_and_batch(struct zone *zone,
6144 struct per_cpu_pageset *pcp)
6146 if (percpu_pagelist_fraction)
6147 pageset_set_high(pcp,
6148 (zone_managed_pages(zone) /
6149 percpu_pagelist_fraction));
6151 pageset_set_batch(pcp, zone_batchsize(zone));
6154 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6156 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6159 pageset_set_high_and_batch(zone, pcp);
6162 void __meminit setup_zone_pageset(struct zone *zone)
6165 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6166 for_each_possible_cpu(cpu)
6167 zone_pageset_init(zone, cpu);
6171 * Allocate per cpu pagesets and initialize them.
6172 * Before this call only boot pagesets were available.
6174 void __init setup_per_cpu_pageset(void)
6176 struct pglist_data *pgdat;
6179 for_each_populated_zone(zone)
6180 setup_zone_pageset(zone);
6182 for_each_online_pgdat(pgdat)
6183 pgdat->per_cpu_nodestats =
6184 alloc_percpu(struct per_cpu_nodestat);
6187 static __meminit void zone_pcp_init(struct zone *zone)
6190 * per cpu subsystem is not up at this point. The following code
6191 * relies on the ability of the linker to provide the
6192 * offset of a (static) per cpu variable into the per cpu area.
6194 zone->pageset = &boot_pageset;
6196 if (populated_zone(zone))
6197 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6198 zone->name, zone->present_pages,
6199 zone_batchsize(zone));
6202 void __meminit init_currently_empty_zone(struct zone *zone,
6203 unsigned long zone_start_pfn,
6206 struct pglist_data *pgdat = zone->zone_pgdat;
6207 int zone_idx = zone_idx(zone) + 1;
6209 if (zone_idx > pgdat->nr_zones)
6210 pgdat->nr_zones = zone_idx;
6212 zone->zone_start_pfn = zone_start_pfn;
6214 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6215 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6217 (unsigned long)zone_idx(zone),
6218 zone_start_pfn, (zone_start_pfn + size));
6220 zone_init_free_lists(zone);
6221 zone->initialized = 1;
6224 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6225 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
6228 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
6230 int __meminit __early_pfn_to_nid(unsigned long pfn,
6231 struct mminit_pfnnid_cache *state)
6233 unsigned long start_pfn, end_pfn;
6236 if (state->last_start <= pfn && pfn < state->last_end)
6237 return state->last_nid;
6239 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
6240 if (nid != NUMA_NO_NODE) {
6241 state->last_start = start_pfn;
6242 state->last_end = end_pfn;
6243 state->last_nid = nid;
6248 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
6251 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
6252 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
6253 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
6255 * If an architecture guarantees that all ranges registered contain no holes
6256 * and may be freed, this this function may be used instead of calling
6257 * memblock_free_early_nid() manually.
6259 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
6261 unsigned long start_pfn, end_pfn;
6264 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
6265 start_pfn = min(start_pfn, max_low_pfn);
6266 end_pfn = min(end_pfn, max_low_pfn);
6268 if (start_pfn < end_pfn)
6269 memblock_free_early_nid(PFN_PHYS(start_pfn),
6270 (end_pfn - start_pfn) << PAGE_SHIFT,
6276 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6277 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6279 * If an architecture guarantees that all ranges registered contain no holes and may
6280 * be freed, this function may be used instead of calling memory_present() manually.
6282 void __init sparse_memory_present_with_active_regions(int nid)
6284 unsigned long start_pfn, end_pfn;
6287 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
6288 memory_present(this_nid, start_pfn, end_pfn);
6292 * get_pfn_range_for_nid - Return the start and end page frames for a node
6293 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6294 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6295 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6297 * It returns the start and end page frame of a node based on information
6298 * provided by memblock_set_node(). If called for a node
6299 * with no available memory, a warning is printed and the start and end
6302 void __init get_pfn_range_for_nid(unsigned int nid,
6303 unsigned long *start_pfn, unsigned long *end_pfn)
6305 unsigned long this_start_pfn, this_end_pfn;
6311 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6312 *start_pfn = min(*start_pfn, this_start_pfn);
6313 *end_pfn = max(*end_pfn, this_end_pfn);
6316 if (*start_pfn == -1UL)
6321 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6322 * assumption is made that zones within a node are ordered in monotonic
6323 * increasing memory addresses so that the "highest" populated zone is used
6325 static void __init find_usable_zone_for_movable(void)
6328 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6329 if (zone_index == ZONE_MOVABLE)
6332 if (arch_zone_highest_possible_pfn[zone_index] >
6333 arch_zone_lowest_possible_pfn[zone_index])
6337 VM_BUG_ON(zone_index == -1);
6338 movable_zone = zone_index;
6342 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6343 * because it is sized independent of architecture. Unlike the other zones,
6344 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6345 * in each node depending on the size of each node and how evenly kernelcore
6346 * is distributed. This helper function adjusts the zone ranges
6347 * provided by the architecture for a given node by using the end of the
6348 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6349 * zones within a node are in order of monotonic increases memory addresses
6351 static void __init adjust_zone_range_for_zone_movable(int nid,
6352 unsigned long zone_type,
6353 unsigned long node_start_pfn,
6354 unsigned long node_end_pfn,
6355 unsigned long *zone_start_pfn,
6356 unsigned long *zone_end_pfn)
6358 /* Only adjust if ZONE_MOVABLE is on this node */
6359 if (zone_movable_pfn[nid]) {
6360 /* Size ZONE_MOVABLE */
6361 if (zone_type == ZONE_MOVABLE) {
6362 *zone_start_pfn = zone_movable_pfn[nid];
6363 *zone_end_pfn = min(node_end_pfn,
6364 arch_zone_highest_possible_pfn[movable_zone]);
6366 /* Adjust for ZONE_MOVABLE starting within this range */
6367 } else if (!mirrored_kernelcore &&
6368 *zone_start_pfn < zone_movable_pfn[nid] &&
6369 *zone_end_pfn > zone_movable_pfn[nid]) {
6370 *zone_end_pfn = zone_movable_pfn[nid];
6372 /* Check if this whole range is within ZONE_MOVABLE */
6373 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6374 *zone_start_pfn = *zone_end_pfn;
6379 * Return the number of pages a zone spans in a node, including holes
6380 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6382 static unsigned long __init zone_spanned_pages_in_node(int nid,
6383 unsigned long zone_type,
6384 unsigned long node_start_pfn,
6385 unsigned long node_end_pfn,
6386 unsigned long *zone_start_pfn,
6387 unsigned long *zone_end_pfn,
6388 unsigned long *ignored)
6390 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6391 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6392 /* When hotadd a new node from cpu_up(), the node should be empty */
6393 if (!node_start_pfn && !node_end_pfn)
6396 /* Get the start and end of the zone */
6397 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6398 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6399 adjust_zone_range_for_zone_movable(nid, zone_type,
6400 node_start_pfn, node_end_pfn,
6401 zone_start_pfn, zone_end_pfn);
6403 /* Check that this node has pages within the zone's required range */
6404 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6407 /* Move the zone boundaries inside the node if necessary */
6408 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6409 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6411 /* Return the spanned pages */
6412 return *zone_end_pfn - *zone_start_pfn;
6416 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6417 * then all holes in the requested range will be accounted for.
6419 unsigned long __init __absent_pages_in_range(int nid,
6420 unsigned long range_start_pfn,
6421 unsigned long range_end_pfn)
6423 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6424 unsigned long start_pfn, end_pfn;
6427 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6428 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6429 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6430 nr_absent -= end_pfn - start_pfn;
6436 * absent_pages_in_range - Return number of page frames in holes within a range
6437 * @start_pfn: The start PFN to start searching for holes
6438 * @end_pfn: The end PFN to stop searching for holes
6440 * Return: the number of pages frames in memory holes within a range.
6442 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6443 unsigned long end_pfn)
6445 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6448 /* Return the number of page frames in holes in a zone on a node */
6449 static unsigned long __init zone_absent_pages_in_node(int nid,
6450 unsigned long zone_type,
6451 unsigned long node_start_pfn,
6452 unsigned long node_end_pfn,
6453 unsigned long *ignored)
6455 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6456 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6457 unsigned long zone_start_pfn, zone_end_pfn;
6458 unsigned long nr_absent;
6460 /* When hotadd a new node from cpu_up(), the node should be empty */
6461 if (!node_start_pfn && !node_end_pfn)
6464 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6465 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6467 adjust_zone_range_for_zone_movable(nid, zone_type,
6468 node_start_pfn, node_end_pfn,
6469 &zone_start_pfn, &zone_end_pfn);
6470 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6473 * ZONE_MOVABLE handling.
6474 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6477 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6478 unsigned long start_pfn, end_pfn;
6479 struct memblock_region *r;
6481 for_each_memblock(memory, r) {
6482 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6483 zone_start_pfn, zone_end_pfn);
6484 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6485 zone_start_pfn, zone_end_pfn);
6487 if (zone_type == ZONE_MOVABLE &&
6488 memblock_is_mirror(r))
6489 nr_absent += end_pfn - start_pfn;
6491 if (zone_type == ZONE_NORMAL &&
6492 !memblock_is_mirror(r))
6493 nr_absent += end_pfn - start_pfn;
6500 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6501 static inline unsigned long __init zone_spanned_pages_in_node(int nid,
6502 unsigned long zone_type,
6503 unsigned long node_start_pfn,
6504 unsigned long node_end_pfn,
6505 unsigned long *zone_start_pfn,
6506 unsigned long *zone_end_pfn,
6507 unsigned long *zones_size)
6511 *zone_start_pfn = node_start_pfn;
6512 for (zone = 0; zone < zone_type; zone++)
6513 *zone_start_pfn += zones_size[zone];
6515 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6517 return zones_size[zone_type];
6520 static inline unsigned long __init zone_absent_pages_in_node(int nid,
6521 unsigned long zone_type,
6522 unsigned long node_start_pfn,
6523 unsigned long node_end_pfn,
6524 unsigned long *zholes_size)
6529 return zholes_size[zone_type];
6532 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6534 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6535 unsigned long node_start_pfn,
6536 unsigned long node_end_pfn,
6537 unsigned long *zones_size,
6538 unsigned long *zholes_size)
6540 unsigned long realtotalpages = 0, totalpages = 0;
6543 for (i = 0; i < MAX_NR_ZONES; i++) {
6544 struct zone *zone = pgdat->node_zones + i;
6545 unsigned long zone_start_pfn, zone_end_pfn;
6546 unsigned long size, real_size;
6548 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6554 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6555 node_start_pfn, node_end_pfn,
6558 zone->zone_start_pfn = zone_start_pfn;
6560 zone->zone_start_pfn = 0;
6561 zone->spanned_pages = size;
6562 zone->present_pages = real_size;
6565 realtotalpages += real_size;
6568 pgdat->node_spanned_pages = totalpages;
6569 pgdat->node_present_pages = realtotalpages;
6570 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6574 #ifndef CONFIG_SPARSEMEM
6576 * Calculate the size of the zone->blockflags rounded to an unsigned long
6577 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6578 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6579 * round what is now in bits to nearest long in bits, then return it in
6582 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6584 unsigned long usemapsize;
6586 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6587 usemapsize = roundup(zonesize, pageblock_nr_pages);
6588 usemapsize = usemapsize >> pageblock_order;
6589 usemapsize *= NR_PAGEBLOCK_BITS;
6590 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6592 return usemapsize / 8;
6595 static void __ref setup_usemap(struct pglist_data *pgdat,
6597 unsigned long zone_start_pfn,
6598 unsigned long zonesize)
6600 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6601 zone->pageblock_flags = NULL;
6603 zone->pageblock_flags =
6604 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6606 if (!zone->pageblock_flags)
6607 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6608 usemapsize, zone->name, pgdat->node_id);
6612 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6613 unsigned long zone_start_pfn, unsigned long zonesize) {}
6614 #endif /* CONFIG_SPARSEMEM */
6616 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6618 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6619 void __init set_pageblock_order(void)
6623 /* Check that pageblock_nr_pages has not already been setup */
6624 if (pageblock_order)
6627 if (HPAGE_SHIFT > PAGE_SHIFT)
6628 order = HUGETLB_PAGE_ORDER;
6630 order = MAX_ORDER - 1;
6633 * Assume the largest contiguous order of interest is a huge page.
6634 * This value may be variable depending on boot parameters on IA64 and
6637 pageblock_order = order;
6639 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6642 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6643 * is unused as pageblock_order is set at compile-time. See
6644 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6647 void __init set_pageblock_order(void)
6651 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6653 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6654 unsigned long present_pages)
6656 unsigned long pages = spanned_pages;
6659 * Provide a more accurate estimation if there are holes within
6660 * the zone and SPARSEMEM is in use. If there are holes within the
6661 * zone, each populated memory region may cost us one or two extra
6662 * memmap pages due to alignment because memmap pages for each
6663 * populated regions may not be naturally aligned on page boundary.
6664 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6666 if (spanned_pages > present_pages + (present_pages >> 4) &&
6667 IS_ENABLED(CONFIG_SPARSEMEM))
6668 pages = present_pages;
6670 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6673 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6674 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6676 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6678 spin_lock_init(&ds_queue->split_queue_lock);
6679 INIT_LIST_HEAD(&ds_queue->split_queue);
6680 ds_queue->split_queue_len = 0;
6683 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6686 #ifdef CONFIG_COMPACTION
6687 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6689 init_waitqueue_head(&pgdat->kcompactd_wait);
6692 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6695 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6697 pgdat_resize_init(pgdat);
6699 pgdat_init_split_queue(pgdat);
6700 pgdat_init_kcompactd(pgdat);
6702 init_waitqueue_head(&pgdat->kswapd_wait);
6703 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6705 pgdat_page_ext_init(pgdat);
6706 spin_lock_init(&pgdat->lru_lock);
6707 lruvec_init(node_lruvec(pgdat));
6710 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6711 unsigned long remaining_pages)
6713 atomic_long_set(&zone->managed_pages, remaining_pages);
6714 zone_set_nid(zone, nid);
6715 zone->name = zone_names[idx];
6716 zone->zone_pgdat = NODE_DATA(nid);
6717 spin_lock_init(&zone->lock);
6718 zone_seqlock_init(zone);
6719 zone_pcp_init(zone);
6723 * Set up the zone data structures
6724 * - init pgdat internals
6725 * - init all zones belonging to this node
6727 * NOTE: this function is only called during memory hotplug
6729 #ifdef CONFIG_MEMORY_HOTPLUG
6730 void __ref free_area_init_core_hotplug(int nid)
6733 pg_data_t *pgdat = NODE_DATA(nid);
6735 pgdat_init_internals(pgdat);
6736 for (z = 0; z < MAX_NR_ZONES; z++)
6737 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6742 * Set up the zone data structures:
6743 * - mark all pages reserved
6744 * - mark all memory queues empty
6745 * - clear the memory bitmaps
6747 * NOTE: pgdat should get zeroed by caller.
6748 * NOTE: this function is only called during early init.
6750 static void __init free_area_init_core(struct pglist_data *pgdat)
6753 int nid = pgdat->node_id;
6755 pgdat_init_internals(pgdat);
6756 pgdat->per_cpu_nodestats = &boot_nodestats;
6758 for (j = 0; j < MAX_NR_ZONES; j++) {
6759 struct zone *zone = pgdat->node_zones + j;
6760 unsigned long size, freesize, memmap_pages;
6761 unsigned long zone_start_pfn = zone->zone_start_pfn;
6763 size = zone->spanned_pages;
6764 freesize = zone->present_pages;
6767 * Adjust freesize so that it accounts for how much memory
6768 * is used by this zone for memmap. This affects the watermark
6769 * and per-cpu initialisations
6771 memmap_pages = calc_memmap_size(size, freesize);
6772 if (!is_highmem_idx(j)) {
6773 if (freesize >= memmap_pages) {
6774 freesize -= memmap_pages;
6777 " %s zone: %lu pages used for memmap\n",
6778 zone_names[j], memmap_pages);
6780 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6781 zone_names[j], memmap_pages, freesize);
6784 /* Account for reserved pages */
6785 if (j == 0 && freesize > dma_reserve) {
6786 freesize -= dma_reserve;
6787 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6788 zone_names[0], dma_reserve);
6791 if (!is_highmem_idx(j))
6792 nr_kernel_pages += freesize;
6793 /* Charge for highmem memmap if there are enough kernel pages */
6794 else if (nr_kernel_pages > memmap_pages * 2)
6795 nr_kernel_pages -= memmap_pages;
6796 nr_all_pages += freesize;
6799 * Set an approximate value for lowmem here, it will be adjusted
6800 * when the bootmem allocator frees pages into the buddy system.
6801 * And all highmem pages will be managed by the buddy system.
6803 zone_init_internals(zone, j, nid, freesize);
6808 set_pageblock_order();
6809 setup_usemap(pgdat, zone, zone_start_pfn, size);
6810 init_currently_empty_zone(zone, zone_start_pfn, size);
6811 memmap_init(size, nid, j, zone_start_pfn);
6815 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6816 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6818 unsigned long __maybe_unused start = 0;
6819 unsigned long __maybe_unused offset = 0;
6821 /* Skip empty nodes */
6822 if (!pgdat->node_spanned_pages)
6825 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6826 offset = pgdat->node_start_pfn - start;
6827 /* ia64 gets its own node_mem_map, before this, without bootmem */
6828 if (!pgdat->node_mem_map) {
6829 unsigned long size, end;
6833 * The zone's endpoints aren't required to be MAX_ORDER
6834 * aligned but the node_mem_map endpoints must be in order
6835 * for the buddy allocator to function correctly.
6837 end = pgdat_end_pfn(pgdat);
6838 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6839 size = (end - start) * sizeof(struct page);
6840 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6843 panic("Failed to allocate %ld bytes for node %d memory map\n",
6844 size, pgdat->node_id);
6845 pgdat->node_mem_map = map + offset;
6847 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6848 __func__, pgdat->node_id, (unsigned long)pgdat,
6849 (unsigned long)pgdat->node_mem_map);
6850 #ifndef CONFIG_NEED_MULTIPLE_NODES
6852 * With no DISCONTIG, the global mem_map is just set as node 0's
6854 if (pgdat == NODE_DATA(0)) {
6855 mem_map = NODE_DATA(0)->node_mem_map;
6856 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6857 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6859 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6864 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6865 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6867 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6868 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6870 pgdat->first_deferred_pfn = ULONG_MAX;
6873 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6876 void __init free_area_init_node(int nid, unsigned long *zones_size,
6877 unsigned long node_start_pfn,
6878 unsigned long *zholes_size)
6880 pg_data_t *pgdat = NODE_DATA(nid);
6881 unsigned long start_pfn = 0;
6882 unsigned long end_pfn = 0;
6884 /* pg_data_t should be reset to zero when it's allocated */
6885 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6887 pgdat->node_id = nid;
6888 pgdat->node_start_pfn = node_start_pfn;
6889 pgdat->per_cpu_nodestats = NULL;
6890 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6891 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6892 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6893 (u64)start_pfn << PAGE_SHIFT,
6894 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6896 start_pfn = node_start_pfn;
6898 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6899 zones_size, zholes_size);
6901 alloc_node_mem_map(pgdat);
6902 pgdat_set_deferred_range(pgdat);
6904 free_area_init_core(pgdat);
6907 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6909 * Zero all valid struct pages in range [spfn, epfn), return number of struct
6912 static u64 zero_pfn_range(unsigned long spfn, unsigned long epfn)
6917 for (pfn = spfn; pfn < epfn; pfn++) {
6918 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6919 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6920 + pageblock_nr_pages - 1;
6923 mm_zero_struct_page(pfn_to_page(pfn));
6931 * Only struct pages that are backed by physical memory are zeroed and
6932 * initialized by going through __init_single_page(). But, there are some
6933 * struct pages which are reserved in memblock allocator and their fields
6934 * may be accessed (for example page_to_pfn() on some configuration accesses
6935 * flags). We must explicitly zero those struct pages.
6937 * This function also addresses a similar issue where struct pages are left
6938 * uninitialized because the physical address range is not covered by
6939 * memblock.memory or memblock.reserved. That could happen when memblock
6940 * layout is manually configured via memmap=, or when the highest physical
6941 * address (max_pfn) does not end on a section boundary.
6943 void __init zero_resv_unavail(void)
6945 phys_addr_t start, end;
6947 phys_addr_t next = 0;
6950 * Loop through unavailable ranges not covered by memblock.memory.
6953 for_each_mem_range(i, &memblock.memory, NULL,
6954 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
6956 pgcnt += zero_pfn_range(PFN_DOWN(next), PFN_UP(start));
6961 * Early sections always have a fully populated memmap for the whole
6962 * section - see pfn_valid(). If the last section has holes at the
6963 * end and that section is marked "online", the memmap will be
6964 * considered initialized. Make sure that memmap has a well defined
6967 pgcnt += zero_pfn_range(PFN_DOWN(next),
6968 round_up(max_pfn, PAGES_PER_SECTION));
6971 * Struct pages that do not have backing memory. This could be because
6972 * firmware is using some of this memory, or for some other reasons.
6975 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
6977 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
6979 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6981 #if MAX_NUMNODES > 1
6983 * Figure out the number of possible node ids.
6985 void __init setup_nr_node_ids(void)
6987 unsigned int highest;
6989 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6990 nr_node_ids = highest + 1;
6995 * node_map_pfn_alignment - determine the maximum internode alignment
6997 * This function should be called after node map is populated and sorted.
6998 * It calculates the maximum power of two alignment which can distinguish
7001 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7002 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7003 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7004 * shifted, 1GiB is enough and this function will indicate so.
7006 * This is used to test whether pfn -> nid mapping of the chosen memory
7007 * model has fine enough granularity to avoid incorrect mapping for the
7008 * populated node map.
7010 * Return: the determined alignment in pfn's. 0 if there is no alignment
7011 * requirement (single node).
7013 unsigned long __init node_map_pfn_alignment(void)
7015 unsigned long accl_mask = 0, last_end = 0;
7016 unsigned long start, end, mask;
7017 int last_nid = NUMA_NO_NODE;
7020 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7021 if (!start || last_nid < 0 || last_nid == nid) {
7028 * Start with a mask granular enough to pin-point to the
7029 * start pfn and tick off bits one-by-one until it becomes
7030 * too coarse to separate the current node from the last.
7032 mask = ~((1 << __ffs(start)) - 1);
7033 while (mask && last_end <= (start & (mask << 1)))
7036 /* accumulate all internode masks */
7040 /* convert mask to number of pages */
7041 return ~accl_mask + 1;
7044 /* Find the lowest pfn for a node */
7045 static unsigned long __init find_min_pfn_for_node(int nid)
7047 unsigned long min_pfn = ULONG_MAX;
7048 unsigned long start_pfn;
7051 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
7052 min_pfn = min(min_pfn, start_pfn);
7054 if (min_pfn == ULONG_MAX) {
7055 pr_warn("Could not find start_pfn for node %d\n", nid);
7063 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7065 * Return: the minimum PFN based on information provided via
7066 * memblock_set_node().
7068 unsigned long __init find_min_pfn_with_active_regions(void)
7070 return find_min_pfn_for_node(MAX_NUMNODES);
7074 * early_calculate_totalpages()
7075 * Sum pages in active regions for movable zone.
7076 * Populate N_MEMORY for calculating usable_nodes.
7078 static unsigned long __init early_calculate_totalpages(void)
7080 unsigned long totalpages = 0;
7081 unsigned long start_pfn, end_pfn;
7084 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7085 unsigned long pages = end_pfn - start_pfn;
7087 totalpages += pages;
7089 node_set_state(nid, N_MEMORY);
7095 * Find the PFN the Movable zone begins in each node. Kernel memory
7096 * is spread evenly between nodes as long as the nodes have enough
7097 * memory. When they don't, some nodes will have more kernelcore than
7100 static void __init find_zone_movable_pfns_for_nodes(void)
7103 unsigned long usable_startpfn;
7104 unsigned long kernelcore_node, kernelcore_remaining;
7105 /* save the state before borrow the nodemask */
7106 nodemask_t saved_node_state = node_states[N_MEMORY];
7107 unsigned long totalpages = early_calculate_totalpages();
7108 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7109 struct memblock_region *r;
7111 /* Need to find movable_zone earlier when movable_node is specified. */
7112 find_usable_zone_for_movable();
7115 * If movable_node is specified, ignore kernelcore and movablecore
7118 if (movable_node_is_enabled()) {
7119 for_each_memblock(memory, r) {
7120 if (!memblock_is_hotpluggable(r))
7125 usable_startpfn = PFN_DOWN(r->base);
7126 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7127 min(usable_startpfn, zone_movable_pfn[nid]) :
7135 * If kernelcore=mirror is specified, ignore movablecore option
7137 if (mirrored_kernelcore) {
7138 bool mem_below_4gb_not_mirrored = false;
7140 for_each_memblock(memory, r) {
7141 if (memblock_is_mirror(r))
7146 usable_startpfn = memblock_region_memory_base_pfn(r);
7148 if (usable_startpfn < 0x100000) {
7149 mem_below_4gb_not_mirrored = true;
7153 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7154 min(usable_startpfn, zone_movable_pfn[nid]) :
7158 if (mem_below_4gb_not_mirrored)
7159 pr_warn("This configuration results in unmirrored kernel memory.");
7165 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7166 * amount of necessary memory.
7168 if (required_kernelcore_percent)
7169 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7171 if (required_movablecore_percent)
7172 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7176 * If movablecore= was specified, calculate what size of
7177 * kernelcore that corresponds so that memory usable for
7178 * any allocation type is evenly spread. If both kernelcore
7179 * and movablecore are specified, then the value of kernelcore
7180 * will be used for required_kernelcore if it's greater than
7181 * what movablecore would have allowed.
7183 if (required_movablecore) {
7184 unsigned long corepages;
7187 * Round-up so that ZONE_MOVABLE is at least as large as what
7188 * was requested by the user
7190 required_movablecore =
7191 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7192 required_movablecore = min(totalpages, required_movablecore);
7193 corepages = totalpages - required_movablecore;
7195 required_kernelcore = max(required_kernelcore, corepages);
7199 * If kernelcore was not specified or kernelcore size is larger
7200 * than totalpages, there is no ZONE_MOVABLE.
7202 if (!required_kernelcore || required_kernelcore >= totalpages)
7205 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7206 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7209 /* Spread kernelcore memory as evenly as possible throughout nodes */
7210 kernelcore_node = required_kernelcore / usable_nodes;
7211 for_each_node_state(nid, N_MEMORY) {
7212 unsigned long start_pfn, end_pfn;
7215 * Recalculate kernelcore_node if the division per node
7216 * now exceeds what is necessary to satisfy the requested
7217 * amount of memory for the kernel
7219 if (required_kernelcore < kernelcore_node)
7220 kernelcore_node = required_kernelcore / usable_nodes;
7223 * As the map is walked, we track how much memory is usable
7224 * by the kernel using kernelcore_remaining. When it is
7225 * 0, the rest of the node is usable by ZONE_MOVABLE
7227 kernelcore_remaining = kernelcore_node;
7229 /* Go through each range of PFNs within this node */
7230 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7231 unsigned long size_pages;
7233 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7234 if (start_pfn >= end_pfn)
7237 /* Account for what is only usable for kernelcore */
7238 if (start_pfn < usable_startpfn) {
7239 unsigned long kernel_pages;
7240 kernel_pages = min(end_pfn, usable_startpfn)
7243 kernelcore_remaining -= min(kernel_pages,
7244 kernelcore_remaining);
7245 required_kernelcore -= min(kernel_pages,
7246 required_kernelcore);
7248 /* Continue if range is now fully accounted */
7249 if (end_pfn <= usable_startpfn) {
7252 * Push zone_movable_pfn to the end so
7253 * that if we have to rebalance
7254 * kernelcore across nodes, we will
7255 * not double account here
7257 zone_movable_pfn[nid] = end_pfn;
7260 start_pfn = usable_startpfn;
7264 * The usable PFN range for ZONE_MOVABLE is from
7265 * start_pfn->end_pfn. Calculate size_pages as the
7266 * number of pages used as kernelcore
7268 size_pages = end_pfn - start_pfn;
7269 if (size_pages > kernelcore_remaining)
7270 size_pages = kernelcore_remaining;
7271 zone_movable_pfn[nid] = start_pfn + size_pages;
7274 * Some kernelcore has been met, update counts and
7275 * break if the kernelcore for this node has been
7278 required_kernelcore -= min(required_kernelcore,
7280 kernelcore_remaining -= size_pages;
7281 if (!kernelcore_remaining)
7287 * If there is still required_kernelcore, we do another pass with one
7288 * less node in the count. This will push zone_movable_pfn[nid] further
7289 * along on the nodes that still have memory until kernelcore is
7293 if (usable_nodes && required_kernelcore > usable_nodes)
7297 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7298 for (nid = 0; nid < MAX_NUMNODES; nid++)
7299 zone_movable_pfn[nid] =
7300 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7303 /* restore the node_state */
7304 node_states[N_MEMORY] = saved_node_state;
7307 /* Any regular or high memory on that node ? */
7308 static void check_for_memory(pg_data_t *pgdat, int nid)
7310 enum zone_type zone_type;
7312 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7313 struct zone *zone = &pgdat->node_zones[zone_type];
7314 if (populated_zone(zone)) {
7315 if (IS_ENABLED(CONFIG_HIGHMEM))
7316 node_set_state(nid, N_HIGH_MEMORY);
7317 if (zone_type <= ZONE_NORMAL)
7318 node_set_state(nid, N_NORMAL_MEMORY);
7325 * free_area_init_nodes - Initialise all pg_data_t and zone data
7326 * @max_zone_pfn: an array of max PFNs for each zone
7328 * This will call free_area_init_node() for each active node in the system.
7329 * Using the page ranges provided by memblock_set_node(), the size of each
7330 * zone in each node and their holes is calculated. If the maximum PFN
7331 * between two adjacent zones match, it is assumed that the zone is empty.
7332 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7333 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7334 * starts where the previous one ended. For example, ZONE_DMA32 starts
7335 * at arch_max_dma_pfn.
7337 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
7339 unsigned long start_pfn, end_pfn;
7342 /* Record where the zone boundaries are */
7343 memset(arch_zone_lowest_possible_pfn, 0,
7344 sizeof(arch_zone_lowest_possible_pfn));
7345 memset(arch_zone_highest_possible_pfn, 0,
7346 sizeof(arch_zone_highest_possible_pfn));
7348 start_pfn = find_min_pfn_with_active_regions();
7350 for (i = 0; i < MAX_NR_ZONES; i++) {
7351 if (i == ZONE_MOVABLE)
7354 end_pfn = max(max_zone_pfn[i], start_pfn);
7355 arch_zone_lowest_possible_pfn[i] = start_pfn;
7356 arch_zone_highest_possible_pfn[i] = end_pfn;
7358 start_pfn = end_pfn;
7361 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7362 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7363 find_zone_movable_pfns_for_nodes();
7365 /* Print out the zone ranges */
7366 pr_info("Zone ranges:\n");
7367 for (i = 0; i < MAX_NR_ZONES; i++) {
7368 if (i == ZONE_MOVABLE)
7370 pr_info(" %-8s ", zone_names[i]);
7371 if (arch_zone_lowest_possible_pfn[i] ==
7372 arch_zone_highest_possible_pfn[i])
7375 pr_cont("[mem %#018Lx-%#018Lx]\n",
7376 (u64)arch_zone_lowest_possible_pfn[i]
7378 ((u64)arch_zone_highest_possible_pfn[i]
7379 << PAGE_SHIFT) - 1);
7382 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7383 pr_info("Movable zone start for each node\n");
7384 for (i = 0; i < MAX_NUMNODES; i++) {
7385 if (zone_movable_pfn[i])
7386 pr_info(" Node %d: %#018Lx\n", i,
7387 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7391 * Print out the early node map, and initialize the
7392 * subsection-map relative to active online memory ranges to
7393 * enable future "sub-section" extensions of the memory map.
7395 pr_info("Early memory node ranges\n");
7396 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7397 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7398 (u64)start_pfn << PAGE_SHIFT,
7399 ((u64)end_pfn << PAGE_SHIFT) - 1);
7400 subsection_map_init(start_pfn, end_pfn - start_pfn);
7403 /* Initialise every node */
7404 mminit_verify_pageflags_layout();
7405 setup_nr_node_ids();
7406 zero_resv_unavail();
7407 for_each_online_node(nid) {
7408 pg_data_t *pgdat = NODE_DATA(nid);
7409 free_area_init_node(nid, NULL,
7410 find_min_pfn_for_node(nid), NULL);
7412 /* Any memory on that node */
7413 if (pgdat->node_present_pages)
7414 node_set_state(nid, N_MEMORY);
7415 check_for_memory(pgdat, nid);
7419 static int __init cmdline_parse_core(char *p, unsigned long *core,
7420 unsigned long *percent)
7422 unsigned long long coremem;
7428 /* Value may be a percentage of total memory, otherwise bytes */
7429 coremem = simple_strtoull(p, &endptr, 0);
7430 if (*endptr == '%') {
7431 /* Paranoid check for percent values greater than 100 */
7432 WARN_ON(coremem > 100);
7436 coremem = memparse(p, &p);
7437 /* Paranoid check that UL is enough for the coremem value */
7438 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7440 *core = coremem >> PAGE_SHIFT;
7447 * kernelcore=size sets the amount of memory for use for allocations that
7448 * cannot be reclaimed or migrated.
7450 static int __init cmdline_parse_kernelcore(char *p)
7452 /* parse kernelcore=mirror */
7453 if (parse_option_str(p, "mirror")) {
7454 mirrored_kernelcore = true;
7458 return cmdline_parse_core(p, &required_kernelcore,
7459 &required_kernelcore_percent);
7463 * movablecore=size sets the amount of memory for use for allocations that
7464 * can be reclaimed or migrated.
7466 static int __init cmdline_parse_movablecore(char *p)
7468 return cmdline_parse_core(p, &required_movablecore,
7469 &required_movablecore_percent);
7472 early_param("kernelcore", cmdline_parse_kernelcore);
7473 early_param("movablecore", cmdline_parse_movablecore);
7475 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
7477 void adjust_managed_page_count(struct page *page, long count)
7479 atomic_long_add(count, &page_zone(page)->managed_pages);
7480 totalram_pages_add(count);
7481 #ifdef CONFIG_HIGHMEM
7482 if (PageHighMem(page))
7483 totalhigh_pages_add(count);
7486 EXPORT_SYMBOL(adjust_managed_page_count);
7488 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7491 unsigned long pages = 0;
7493 start = (void *)PAGE_ALIGN((unsigned long)start);
7494 end = (void *)((unsigned long)end & PAGE_MASK);
7495 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7496 struct page *page = virt_to_page(pos);
7497 void *direct_map_addr;
7500 * 'direct_map_addr' might be different from 'pos'
7501 * because some architectures' virt_to_page()
7502 * work with aliases. Getting the direct map
7503 * address ensures that we get a _writeable_
7504 * alias for the memset().
7506 direct_map_addr = page_address(page);
7507 if ((unsigned int)poison <= 0xFF)
7508 memset(direct_map_addr, poison, PAGE_SIZE);
7510 free_reserved_page(page);
7514 pr_info("Freeing %s memory: %ldK\n",
7515 s, pages << (PAGE_SHIFT - 10));
7520 #ifdef CONFIG_HIGHMEM
7521 void free_highmem_page(struct page *page)
7523 __free_reserved_page(page);
7524 totalram_pages_inc();
7525 atomic_long_inc(&page_zone(page)->managed_pages);
7526 totalhigh_pages_inc();
7531 void __init mem_init_print_info(const char *str)
7533 unsigned long physpages, codesize, datasize, rosize, bss_size;
7534 unsigned long init_code_size, init_data_size;
7536 physpages = get_num_physpages();
7537 codesize = _etext - _stext;
7538 datasize = _edata - _sdata;
7539 rosize = __end_rodata - __start_rodata;
7540 bss_size = __bss_stop - __bss_start;
7541 init_data_size = __init_end - __init_begin;
7542 init_code_size = _einittext - _sinittext;
7545 * Detect special cases and adjust section sizes accordingly:
7546 * 1) .init.* may be embedded into .data sections
7547 * 2) .init.text.* may be out of [__init_begin, __init_end],
7548 * please refer to arch/tile/kernel/vmlinux.lds.S.
7549 * 3) .rodata.* may be embedded into .text or .data sections.
7551 #define adj_init_size(start, end, size, pos, adj) \
7553 if (start <= pos && pos < end && size > adj) \
7557 adj_init_size(__init_begin, __init_end, init_data_size,
7558 _sinittext, init_code_size);
7559 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7560 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7561 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7562 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7564 #undef adj_init_size
7566 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7567 #ifdef CONFIG_HIGHMEM
7571 nr_free_pages() << (PAGE_SHIFT - 10),
7572 physpages << (PAGE_SHIFT - 10),
7573 codesize >> 10, datasize >> 10, rosize >> 10,
7574 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7575 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7576 totalcma_pages << (PAGE_SHIFT - 10),
7577 #ifdef CONFIG_HIGHMEM
7578 totalhigh_pages() << (PAGE_SHIFT - 10),
7580 str ? ", " : "", str ? str : "");
7584 * set_dma_reserve - set the specified number of pages reserved in the first zone
7585 * @new_dma_reserve: The number of pages to mark reserved
7587 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7588 * In the DMA zone, a significant percentage may be consumed by kernel image
7589 * and other unfreeable allocations which can skew the watermarks badly. This
7590 * function may optionally be used to account for unfreeable pages in the
7591 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7592 * smaller per-cpu batchsize.
7594 void __init set_dma_reserve(unsigned long new_dma_reserve)
7596 dma_reserve = new_dma_reserve;
7599 void __init free_area_init(unsigned long *zones_size)
7601 zero_resv_unavail();
7602 free_area_init_node(0, zones_size,
7603 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
7606 static int page_alloc_cpu_dead(unsigned int cpu)
7609 lru_add_drain_cpu(cpu);
7613 * Spill the event counters of the dead processor
7614 * into the current processors event counters.
7615 * This artificially elevates the count of the current
7618 vm_events_fold_cpu(cpu);
7621 * Zero the differential counters of the dead processor
7622 * so that the vm statistics are consistent.
7624 * This is only okay since the processor is dead and cannot
7625 * race with what we are doing.
7627 cpu_vm_stats_fold(cpu);
7632 int hashdist = HASHDIST_DEFAULT;
7634 static int __init set_hashdist(char *str)
7638 hashdist = simple_strtoul(str, &str, 0);
7641 __setup("hashdist=", set_hashdist);
7644 void __init page_alloc_init(void)
7649 if (num_node_state(N_MEMORY) == 1)
7653 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7654 "mm/page_alloc:dead", NULL,
7655 page_alloc_cpu_dead);
7660 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7661 * or min_free_kbytes changes.
7663 static void calculate_totalreserve_pages(void)
7665 struct pglist_data *pgdat;
7666 unsigned long reserve_pages = 0;
7667 enum zone_type i, j;
7669 for_each_online_pgdat(pgdat) {
7671 pgdat->totalreserve_pages = 0;
7673 for (i = 0; i < MAX_NR_ZONES; i++) {
7674 struct zone *zone = pgdat->node_zones + i;
7676 unsigned long managed_pages = zone_managed_pages(zone);
7678 /* Find valid and maximum lowmem_reserve in the zone */
7679 for (j = i; j < MAX_NR_ZONES; j++) {
7680 if (zone->lowmem_reserve[j] > max)
7681 max = zone->lowmem_reserve[j];
7684 /* we treat the high watermark as reserved pages. */
7685 max += high_wmark_pages(zone);
7687 if (max > managed_pages)
7688 max = managed_pages;
7690 pgdat->totalreserve_pages += max;
7692 reserve_pages += max;
7695 totalreserve_pages = reserve_pages;
7699 * setup_per_zone_lowmem_reserve - called whenever
7700 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7701 * has a correct pages reserved value, so an adequate number of
7702 * pages are left in the zone after a successful __alloc_pages().
7704 static void setup_per_zone_lowmem_reserve(void)
7706 struct pglist_data *pgdat;
7707 enum zone_type j, idx;
7709 for_each_online_pgdat(pgdat) {
7710 for (j = 0; j < MAX_NR_ZONES; j++) {
7711 struct zone *zone = pgdat->node_zones + j;
7712 unsigned long managed_pages = zone_managed_pages(zone);
7714 zone->lowmem_reserve[j] = 0;
7718 struct zone *lower_zone;
7721 lower_zone = pgdat->node_zones + idx;
7723 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7724 sysctl_lowmem_reserve_ratio[idx] = 0;
7725 lower_zone->lowmem_reserve[j] = 0;
7727 lower_zone->lowmem_reserve[j] =
7728 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7730 managed_pages += zone_managed_pages(lower_zone);
7735 /* update totalreserve_pages */
7736 calculate_totalreserve_pages();
7739 static void __setup_per_zone_wmarks(void)
7741 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7742 unsigned long lowmem_pages = 0;
7744 unsigned long flags;
7746 /* Calculate total number of !ZONE_HIGHMEM pages */
7747 for_each_zone(zone) {
7748 if (!is_highmem(zone))
7749 lowmem_pages += zone_managed_pages(zone);
7752 for_each_zone(zone) {
7755 spin_lock_irqsave(&zone->lock, flags);
7756 tmp = (u64)pages_min * zone_managed_pages(zone);
7757 do_div(tmp, lowmem_pages);
7758 if (is_highmem(zone)) {
7760 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7761 * need highmem pages, so cap pages_min to a small
7764 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7765 * deltas control async page reclaim, and so should
7766 * not be capped for highmem.
7768 unsigned long min_pages;
7770 min_pages = zone_managed_pages(zone) / 1024;
7771 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7772 zone->_watermark[WMARK_MIN] = min_pages;
7775 * If it's a lowmem zone, reserve a number of pages
7776 * proportionate to the zone's size.
7778 zone->_watermark[WMARK_MIN] = tmp;
7782 * Set the kswapd watermarks distance according to the
7783 * scale factor in proportion to available memory, but
7784 * ensure a minimum size on small systems.
7786 tmp = max_t(u64, tmp >> 2,
7787 mult_frac(zone_managed_pages(zone),
7788 watermark_scale_factor, 10000));
7790 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7791 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7792 zone->watermark_boost = 0;
7794 spin_unlock_irqrestore(&zone->lock, flags);
7797 /* update totalreserve_pages */
7798 calculate_totalreserve_pages();
7802 * setup_per_zone_wmarks - called when min_free_kbytes changes
7803 * or when memory is hot-{added|removed}
7805 * Ensures that the watermark[min,low,high] values for each zone are set
7806 * correctly with respect to min_free_kbytes.
7808 void setup_per_zone_wmarks(void)
7810 static DEFINE_SPINLOCK(lock);
7813 __setup_per_zone_wmarks();
7818 * Initialise min_free_kbytes.
7820 * For small machines we want it small (128k min). For large machines
7821 * we want it large (64MB max). But it is not linear, because network
7822 * bandwidth does not increase linearly with machine size. We use
7824 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7825 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7841 int __meminit init_per_zone_wmark_min(void)
7843 unsigned long lowmem_kbytes;
7844 int new_min_free_kbytes;
7846 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7847 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7849 if (new_min_free_kbytes > user_min_free_kbytes) {
7850 min_free_kbytes = new_min_free_kbytes;
7851 if (min_free_kbytes < 128)
7852 min_free_kbytes = 128;
7853 if (min_free_kbytes > 65536)
7854 min_free_kbytes = 65536;
7856 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7857 new_min_free_kbytes, user_min_free_kbytes);
7859 setup_per_zone_wmarks();
7860 refresh_zone_stat_thresholds();
7861 setup_per_zone_lowmem_reserve();
7864 setup_min_unmapped_ratio();
7865 setup_min_slab_ratio();
7870 core_initcall(init_per_zone_wmark_min)
7873 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7874 * that we can call two helper functions whenever min_free_kbytes
7877 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7878 void __user *buffer, size_t *length, loff_t *ppos)
7882 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7887 user_min_free_kbytes = min_free_kbytes;
7888 setup_per_zone_wmarks();
7893 int watermark_boost_factor_sysctl_handler(struct ctl_table *table, int write,
7894 void __user *buffer, size_t *length, loff_t *ppos)
7898 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7905 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7906 void __user *buffer, size_t *length, loff_t *ppos)
7910 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7915 setup_per_zone_wmarks();
7921 static void setup_min_unmapped_ratio(void)
7926 for_each_online_pgdat(pgdat)
7927 pgdat->min_unmapped_pages = 0;
7930 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7931 sysctl_min_unmapped_ratio) / 100;
7935 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7936 void __user *buffer, size_t *length, loff_t *ppos)
7940 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7944 setup_min_unmapped_ratio();
7949 static void setup_min_slab_ratio(void)
7954 for_each_online_pgdat(pgdat)
7955 pgdat->min_slab_pages = 0;
7958 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7959 sysctl_min_slab_ratio) / 100;
7962 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7963 void __user *buffer, size_t *length, loff_t *ppos)
7967 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7971 setup_min_slab_ratio();
7978 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7979 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7980 * whenever sysctl_lowmem_reserve_ratio changes.
7982 * The reserve ratio obviously has absolutely no relation with the
7983 * minimum watermarks. The lowmem reserve ratio can only make sense
7984 * if in function of the boot time zone sizes.
7986 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7987 void __user *buffer, size_t *length, loff_t *ppos)
7989 proc_dointvec_minmax(table, write, buffer, length, ppos);
7990 setup_per_zone_lowmem_reserve();
7995 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7996 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7997 * pagelist can have before it gets flushed back to buddy allocator.
7999 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8000 void __user *buffer, size_t *length, loff_t *ppos)
8003 int old_percpu_pagelist_fraction;
8006 mutex_lock(&pcp_batch_high_lock);
8007 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8009 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8010 if (!write || ret < 0)
8013 /* Sanity checking to avoid pcp imbalance */
8014 if (percpu_pagelist_fraction &&
8015 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8016 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8022 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8025 for_each_populated_zone(zone) {
8028 for_each_possible_cpu(cpu)
8029 pageset_set_high_and_batch(zone,
8030 per_cpu_ptr(zone->pageset, cpu));
8033 mutex_unlock(&pcp_batch_high_lock);
8037 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8039 * Returns the number of pages that arch has reserved but
8040 * is not known to alloc_large_system_hash().
8042 static unsigned long __init arch_reserved_kernel_pages(void)
8049 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8050 * machines. As memory size is increased the scale is also increased but at
8051 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8052 * quadruples the scale is increased by one, which means the size of hash table
8053 * only doubles, instead of quadrupling as well.
8054 * Because 32-bit systems cannot have large physical memory, where this scaling
8055 * makes sense, it is disabled on such platforms.
8057 #if __BITS_PER_LONG > 32
8058 #define ADAPT_SCALE_BASE (64ul << 30)
8059 #define ADAPT_SCALE_SHIFT 2
8060 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8064 * allocate a large system hash table from bootmem
8065 * - it is assumed that the hash table must contain an exact power-of-2
8066 * quantity of entries
8067 * - limit is the number of hash buckets, not the total allocation size
8069 void *__init alloc_large_system_hash(const char *tablename,
8070 unsigned long bucketsize,
8071 unsigned long numentries,
8074 unsigned int *_hash_shift,
8075 unsigned int *_hash_mask,
8076 unsigned long low_limit,
8077 unsigned long high_limit)
8079 unsigned long long max = high_limit;
8080 unsigned long log2qty, size;
8085 /* allow the kernel cmdline to have a say */
8087 /* round applicable memory size up to nearest megabyte */
8088 numentries = nr_kernel_pages;
8089 numentries -= arch_reserved_kernel_pages();
8091 /* It isn't necessary when PAGE_SIZE >= 1MB */
8092 if (PAGE_SHIFT < 20)
8093 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8095 #if __BITS_PER_LONG > 32
8097 unsigned long adapt;
8099 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8100 adapt <<= ADAPT_SCALE_SHIFT)
8105 /* limit to 1 bucket per 2^scale bytes of low memory */
8106 if (scale > PAGE_SHIFT)
8107 numentries >>= (scale - PAGE_SHIFT);
8109 numentries <<= (PAGE_SHIFT - scale);
8111 /* Make sure we've got at least a 0-order allocation.. */
8112 if (unlikely(flags & HASH_SMALL)) {
8113 /* Makes no sense without HASH_EARLY */
8114 WARN_ON(!(flags & HASH_EARLY));
8115 if (!(numentries >> *_hash_shift)) {
8116 numentries = 1UL << *_hash_shift;
8117 BUG_ON(!numentries);
8119 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8120 numentries = PAGE_SIZE / bucketsize;
8122 numentries = roundup_pow_of_two(numentries);
8124 /* limit allocation size to 1/16 total memory by default */
8126 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8127 do_div(max, bucketsize);
8129 max = min(max, 0x80000000ULL);
8131 if (numentries < low_limit)
8132 numentries = low_limit;
8133 if (numentries > max)
8136 log2qty = ilog2(numentries);
8138 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8141 size = bucketsize << log2qty;
8142 if (flags & HASH_EARLY) {
8143 if (flags & HASH_ZERO)
8144 table = memblock_alloc(size, SMP_CACHE_BYTES);
8146 table = memblock_alloc_raw(size,
8148 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8149 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
8153 * If bucketsize is not a power-of-two, we may free
8154 * some pages at the end of hash table which
8155 * alloc_pages_exact() automatically does
8157 table = alloc_pages_exact(size, gfp_flags);
8158 kmemleak_alloc(table, size, 1, gfp_flags);
8160 } while (!table && size > PAGE_SIZE && --log2qty);
8163 panic("Failed to allocate %s hash table\n", tablename);
8165 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8166 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8167 virt ? "vmalloc" : "linear");
8170 *_hash_shift = log2qty;
8172 *_hash_mask = (1 << log2qty) - 1;
8178 * This function checks whether pageblock includes unmovable pages or not.
8179 * If @count is not zero, it is okay to include less @count unmovable pages
8181 * PageLRU check without isolation or lru_lock could race so that
8182 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8183 * check without lock_page also may miss some movable non-lru pages at
8184 * race condition. So you can't expect this function should be exact.
8186 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
8187 int migratetype, int flags)
8189 unsigned long found;
8190 unsigned long iter = 0;
8191 unsigned long pfn = page_to_pfn(page);
8192 const char *reason = "unmovable page";
8195 * TODO we could make this much more efficient by not checking every
8196 * page in the range if we know all of them are in MOVABLE_ZONE and
8197 * that the movable zone guarantees that pages are migratable but
8198 * the later is not the case right now unfortunatelly. E.g. movablecore
8199 * can still lead to having bootmem allocations in zone_movable.
8202 if (is_migrate_cma_page(page)) {
8204 * CMA allocations (alloc_contig_range) really need to mark
8205 * isolate CMA pageblocks even when they are not movable in fact
8206 * so consider them movable here.
8208 if (is_migrate_cma(migratetype))
8211 reason = "CMA page";
8215 for (found = 0; iter < pageblock_nr_pages; iter++) {
8216 unsigned long check = pfn + iter;
8218 if (!pfn_valid_within(check))
8221 page = pfn_to_page(check);
8223 if (PageReserved(page))
8227 * If the zone is movable and we have ruled out all reserved
8228 * pages then it should be reasonably safe to assume the rest
8231 if (zone_idx(zone) == ZONE_MOVABLE)
8235 * Hugepages are not in LRU lists, but they're movable.
8236 * We need not scan over tail pages because we don't
8237 * handle each tail page individually in migration.
8239 if (PageHuge(page)) {
8240 struct page *head = compound_head(page);
8241 unsigned int skip_pages;
8243 if (!hugepage_migration_supported(page_hstate(head)))
8246 skip_pages = compound_nr(head) - (page - head);
8247 iter += skip_pages - 1;
8252 * We can't use page_count without pin a page
8253 * because another CPU can free compound page.
8254 * This check already skips compound tails of THP
8255 * because their page->_refcount is zero at all time.
8257 if (!page_ref_count(page)) {
8258 if (PageBuddy(page))
8259 iter += (1 << page_order(page)) - 1;
8264 * The HWPoisoned page may be not in buddy system, and
8265 * page_count() is not 0.
8267 if ((flags & SKIP_HWPOISON) && PageHWPoison(page))
8270 if (__PageMovable(page))
8276 * If there are RECLAIMABLE pages, we need to check
8277 * it. But now, memory offline itself doesn't call
8278 * shrink_node_slabs() and it still to be fixed.
8281 * If the page is not RAM, page_count()should be 0.
8282 * we don't need more check. This is an _used_ not-movable page.
8284 * The problematic thing here is PG_reserved pages. PG_reserved
8285 * is set to both of a memory hole page and a _used_ kernel
8293 WARN_ON_ONCE(zone_idx(zone) == ZONE_MOVABLE);
8294 if (flags & REPORT_FAILURE)
8295 dump_page(pfn_to_page(pfn + iter), reason);
8299 #ifdef CONFIG_CONTIG_ALLOC
8300 static unsigned long pfn_max_align_down(unsigned long pfn)
8302 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8303 pageblock_nr_pages) - 1);
8306 static unsigned long pfn_max_align_up(unsigned long pfn)
8308 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8309 pageblock_nr_pages));
8312 /* [start, end) must belong to a single zone. */
8313 static int __alloc_contig_migrate_range(struct compact_control *cc,
8314 unsigned long start, unsigned long end)
8316 /* This function is based on compact_zone() from compaction.c. */
8317 unsigned long nr_reclaimed;
8318 unsigned long pfn = start;
8319 unsigned int tries = 0;
8324 while (pfn < end || !list_empty(&cc->migratepages)) {
8325 if (fatal_signal_pending(current)) {
8330 if (list_empty(&cc->migratepages)) {
8331 cc->nr_migratepages = 0;
8332 pfn = isolate_migratepages_range(cc, pfn, end);
8338 } else if (++tries == 5) {
8339 ret = ret < 0 ? ret : -EBUSY;
8343 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8345 cc->nr_migratepages -= nr_reclaimed;
8347 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
8348 NULL, 0, cc->mode, MR_CONTIG_RANGE);
8351 putback_movable_pages(&cc->migratepages);
8358 * alloc_contig_range() -- tries to allocate given range of pages
8359 * @start: start PFN to allocate
8360 * @end: one-past-the-last PFN to allocate
8361 * @migratetype: migratetype of the underlaying pageblocks (either
8362 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8363 * in range must have the same migratetype and it must
8364 * be either of the two.
8365 * @gfp_mask: GFP mask to use during compaction
8367 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8368 * aligned. The PFN range must belong to a single zone.
8370 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8371 * pageblocks in the range. Once isolated, the pageblocks should not
8372 * be modified by others.
8374 * Return: zero on success or negative error code. On success all
8375 * pages which PFN is in [start, end) are allocated for the caller and
8376 * need to be freed with free_contig_range().
8378 int alloc_contig_range(unsigned long start, unsigned long end,
8379 unsigned migratetype, gfp_t gfp_mask)
8381 unsigned long outer_start, outer_end;
8385 struct compact_control cc = {
8386 .nr_migratepages = 0,
8388 .zone = page_zone(pfn_to_page(start)),
8389 .mode = MIGRATE_SYNC,
8390 .ignore_skip_hint = true,
8391 .no_set_skip_hint = true,
8392 .gfp_mask = current_gfp_context(gfp_mask),
8394 INIT_LIST_HEAD(&cc.migratepages);
8397 * What we do here is we mark all pageblocks in range as
8398 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8399 * have different sizes, and due to the way page allocator
8400 * work, we align the range to biggest of the two pages so
8401 * that page allocator won't try to merge buddies from
8402 * different pageblocks and change MIGRATE_ISOLATE to some
8403 * other migration type.
8405 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8406 * migrate the pages from an unaligned range (ie. pages that
8407 * we are interested in). This will put all the pages in
8408 * range back to page allocator as MIGRATE_ISOLATE.
8410 * When this is done, we take the pages in range from page
8411 * allocator removing them from the buddy system. This way
8412 * page allocator will never consider using them.
8414 * This lets us mark the pageblocks back as
8415 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8416 * aligned range but not in the unaligned, original range are
8417 * put back to page allocator so that buddy can use them.
8420 ret = start_isolate_page_range(pfn_max_align_down(start),
8421 pfn_max_align_up(end), migratetype, 0);
8426 * In case of -EBUSY, we'd like to know which page causes problem.
8427 * So, just fall through. test_pages_isolated() has a tracepoint
8428 * which will report the busy page.
8430 * It is possible that busy pages could become available before
8431 * the call to test_pages_isolated, and the range will actually be
8432 * allocated. So, if we fall through be sure to clear ret so that
8433 * -EBUSY is not accidentally used or returned to caller.
8435 ret = __alloc_contig_migrate_range(&cc, start, end);
8436 if (ret && ret != -EBUSY)
8441 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8442 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8443 * more, all pages in [start, end) are free in page allocator.
8444 * What we are going to do is to allocate all pages from
8445 * [start, end) (that is remove them from page allocator).
8447 * The only problem is that pages at the beginning and at the
8448 * end of interesting range may be not aligned with pages that
8449 * page allocator holds, ie. they can be part of higher order
8450 * pages. Because of this, we reserve the bigger range and
8451 * once this is done free the pages we are not interested in.
8453 * We don't have to hold zone->lock here because the pages are
8454 * isolated thus they won't get removed from buddy.
8457 lru_add_drain_all();
8460 outer_start = start;
8461 while (!PageBuddy(pfn_to_page(outer_start))) {
8462 if (++order >= MAX_ORDER) {
8463 outer_start = start;
8466 outer_start &= ~0UL << order;
8469 if (outer_start != start) {
8470 order = page_order(pfn_to_page(outer_start));
8473 * outer_start page could be small order buddy page and
8474 * it doesn't include start page. Adjust outer_start
8475 * in this case to report failed page properly
8476 * on tracepoint in test_pages_isolated()
8478 if (outer_start + (1UL << order) <= start)
8479 outer_start = start;
8482 /* Make sure the range is really isolated. */
8483 if (test_pages_isolated(outer_start, end, false)) {
8484 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8485 __func__, outer_start, end);
8490 /* Grab isolated pages from freelists. */
8491 outer_end = isolate_freepages_range(&cc, outer_start, end);
8497 /* Free head and tail (if any) */
8498 if (start != outer_start)
8499 free_contig_range(outer_start, start - outer_start);
8500 if (end != outer_end)
8501 free_contig_range(end, outer_end - end);
8504 undo_isolate_page_range(pfn_max_align_down(start),
8505 pfn_max_align_up(end), migratetype);
8508 #endif /* CONFIG_CONTIG_ALLOC */
8510 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8512 unsigned int count = 0;
8514 for (; nr_pages--; pfn++) {
8515 struct page *page = pfn_to_page(pfn);
8517 count += page_count(page) != 1;
8520 WARN(count != 0, "%d pages are still in use!\n", count);
8524 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8525 * page high values need to be recalulated.
8527 void __meminit zone_pcp_update(struct zone *zone)
8530 mutex_lock(&pcp_batch_high_lock);
8531 for_each_possible_cpu(cpu)
8532 pageset_set_high_and_batch(zone,
8533 per_cpu_ptr(zone->pageset, cpu));
8534 mutex_unlock(&pcp_batch_high_lock);
8537 void zone_pcp_reset(struct zone *zone)
8539 unsigned long flags;
8541 struct per_cpu_pageset *pset;
8543 /* avoid races with drain_pages() */
8544 local_irq_save(flags);
8545 if (zone->pageset != &boot_pageset) {
8546 for_each_online_cpu(cpu) {
8547 pset = per_cpu_ptr(zone->pageset, cpu);
8548 drain_zonestat(zone, pset);
8550 free_percpu(zone->pageset);
8551 zone->pageset = &boot_pageset;
8553 local_irq_restore(flags);
8556 #ifdef CONFIG_MEMORY_HOTREMOVE
8558 * All pages in the range must be in a single zone and isolated
8559 * before calling this.
8562 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8566 unsigned int order, i;
8568 unsigned long flags;
8569 unsigned long offlined_pages = 0;
8571 /* find the first valid pfn */
8572 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8576 return offlined_pages;
8578 offline_mem_sections(pfn, end_pfn);
8579 zone = page_zone(pfn_to_page(pfn));
8580 spin_lock_irqsave(&zone->lock, flags);
8582 while (pfn < end_pfn) {
8583 if (!pfn_valid(pfn)) {
8587 page = pfn_to_page(pfn);
8589 * The HWPoisoned page may be not in buddy system, and
8590 * page_count() is not 0.
8592 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8594 SetPageReserved(page);
8599 BUG_ON(page_count(page));
8600 BUG_ON(!PageBuddy(page));
8601 order = page_order(page);
8602 offlined_pages += 1 << order;
8603 #ifdef CONFIG_DEBUG_VM
8604 pr_info("remove from free list %lx %d %lx\n",
8605 pfn, 1 << order, end_pfn);
8607 del_page_from_free_area(page, &zone->free_area[order]);
8608 for (i = 0; i < (1 << order); i++)
8609 SetPageReserved((page+i));
8610 pfn += (1 << order);
8612 spin_unlock_irqrestore(&zone->lock, flags);
8614 return offlined_pages;
8618 bool is_free_buddy_page(struct page *page)
8620 struct zone *zone = page_zone(page);
8621 unsigned long pfn = page_to_pfn(page);
8622 unsigned long flags;
8625 spin_lock_irqsave(&zone->lock, flags);
8626 for (order = 0; order < MAX_ORDER; order++) {
8627 struct page *page_head = page - (pfn & ((1 << order) - 1));
8629 if (PageBuddy(page_head) && page_order(page_head) >= order)
8632 spin_unlock_irqrestore(&zone->lock, flags);
8634 return order < MAX_ORDER;
8637 #ifdef CONFIG_MEMORY_FAILURE
8639 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8640 * test is performed under the zone lock to prevent a race against page
8643 bool set_hwpoison_free_buddy_page(struct page *page)
8645 struct zone *zone = page_zone(page);
8646 unsigned long pfn = page_to_pfn(page);
8647 unsigned long flags;
8649 bool hwpoisoned = false;
8651 spin_lock_irqsave(&zone->lock, flags);
8652 for (order = 0; order < MAX_ORDER; order++) {
8653 struct page *page_head = page - (pfn & ((1 << order) - 1));
8655 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8656 if (!TestSetPageHWPoison(page))
8661 spin_unlock_irqrestore(&zone->lock, flags);