1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/page_alloc.c
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
18 #include <linux/stddef.h>
20 #include <linux/highmem.h>
21 #include <linux/swap.h>
22 #include <linux/interrupt.h>
23 #include <linux/pagemap.h>
24 #include <linux/jiffies.h>
25 #include <linux/memblock.h>
26 #include <linux/compiler.h>
27 #include <linux/kernel.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/memremap.h>
46 #include <linux/stop_machine.h>
47 #include <linux/random.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <trace/events/oom.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/sched/mm.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
66 #include <linux/memcontrol.h>
67 #include <linux/ftrace.h>
68 #include <linux/lockdep.h>
69 #include <linux/nmi.h>
70 #include <linux/psi.h>
72 #include <asm/sections.h>
73 #include <asm/tlbflush.h>
74 #include <asm/div64.h>
77 #include "page_reporting.h"
79 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
80 static DEFINE_MUTEX(pcp_batch_high_lock);
81 #define MIN_PERCPU_PAGELIST_FRACTION (8)
83 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
84 DEFINE_PER_CPU(int, numa_node);
85 EXPORT_PER_CPU_SYMBOL(numa_node);
88 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
90 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
92 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
93 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
94 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
95 * defined in <linux/topology.h>.
97 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
98 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
101 /* work_structs for global per-cpu drains */
104 struct work_struct work;
106 static DEFINE_MUTEX(pcpu_drain_mutex);
107 static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
109 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
110 volatile unsigned long latent_entropy __latent_entropy;
111 EXPORT_SYMBOL(latent_entropy);
115 * Array of node states.
117 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
118 [N_POSSIBLE] = NODE_MASK_ALL,
119 [N_ONLINE] = { { [0] = 1UL } },
121 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
122 #ifdef CONFIG_HIGHMEM
123 [N_HIGH_MEMORY] = { { [0] = 1UL } },
125 [N_MEMORY] = { { [0] = 1UL } },
126 [N_CPU] = { { [0] = 1UL } },
129 EXPORT_SYMBOL(node_states);
131 atomic_long_t _totalram_pages __read_mostly;
132 EXPORT_SYMBOL(_totalram_pages);
133 unsigned long totalreserve_pages __read_mostly;
134 unsigned long totalcma_pages __read_mostly;
136 int percpu_pagelist_fraction;
137 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
138 #ifdef CONFIG_INIT_ON_ALLOC_DEFAULT_ON
139 DEFINE_STATIC_KEY_TRUE(init_on_alloc);
141 DEFINE_STATIC_KEY_FALSE(init_on_alloc);
143 EXPORT_SYMBOL(init_on_alloc);
145 #ifdef CONFIG_INIT_ON_FREE_DEFAULT_ON
146 DEFINE_STATIC_KEY_TRUE(init_on_free);
148 DEFINE_STATIC_KEY_FALSE(init_on_free);
150 EXPORT_SYMBOL(init_on_free);
152 static int __init early_init_on_alloc(char *buf)
159 ret = kstrtobool(buf, &bool_result);
160 if (bool_result && page_poisoning_enabled())
161 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_alloc\n");
163 static_branch_enable(&init_on_alloc);
165 static_branch_disable(&init_on_alloc);
168 early_param("init_on_alloc", early_init_on_alloc);
170 static int __init early_init_on_free(char *buf)
177 ret = kstrtobool(buf, &bool_result);
178 if (bool_result && page_poisoning_enabled())
179 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_free\n");
181 static_branch_enable(&init_on_free);
183 static_branch_disable(&init_on_free);
186 early_param("init_on_free", early_init_on_free);
189 * A cached value of the page's pageblock's migratetype, used when the page is
190 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
191 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
192 * Also the migratetype set in the page does not necessarily match the pcplist
193 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
194 * other index - this ensures that it will be put on the correct CMA freelist.
196 static inline int get_pcppage_migratetype(struct page *page)
201 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
203 page->index = migratetype;
206 #ifdef CONFIG_PM_SLEEP
208 * The following functions are used by the suspend/hibernate code to temporarily
209 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
210 * while devices are suspended. To avoid races with the suspend/hibernate code,
211 * they should always be called with system_transition_mutex held
212 * (gfp_allowed_mask also should only be modified with system_transition_mutex
213 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
214 * with that modification).
217 static gfp_t saved_gfp_mask;
219 void pm_restore_gfp_mask(void)
221 WARN_ON(!mutex_is_locked(&system_transition_mutex));
222 if (saved_gfp_mask) {
223 gfp_allowed_mask = saved_gfp_mask;
228 void pm_restrict_gfp_mask(void)
230 WARN_ON(!mutex_is_locked(&system_transition_mutex));
231 WARN_ON(saved_gfp_mask);
232 saved_gfp_mask = gfp_allowed_mask;
233 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
236 bool pm_suspended_storage(void)
238 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
242 #endif /* CONFIG_PM_SLEEP */
244 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
245 unsigned int pageblock_order __read_mostly;
248 static void __free_pages_ok(struct page *page, unsigned int order);
251 * results with 256, 32 in the lowmem_reserve sysctl:
252 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
253 * 1G machine -> (16M dma, 784M normal, 224M high)
254 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
255 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
256 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
258 * TBD: should special case ZONE_DMA32 machines here - in those we normally
259 * don't need any ZONE_NORMAL reservation
261 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
262 #ifdef CONFIG_ZONE_DMA
265 #ifdef CONFIG_ZONE_DMA32
269 #ifdef CONFIG_HIGHMEM
275 static char * const zone_names[MAX_NR_ZONES] = {
276 #ifdef CONFIG_ZONE_DMA
279 #ifdef CONFIG_ZONE_DMA32
283 #ifdef CONFIG_HIGHMEM
287 #ifdef CONFIG_ZONE_DEVICE
292 const char * const migratetype_names[MIGRATE_TYPES] = {
300 #ifdef CONFIG_MEMORY_ISOLATION
305 compound_page_dtor * const compound_page_dtors[] = {
308 #ifdef CONFIG_HUGETLB_PAGE
311 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
316 int min_free_kbytes = 1024;
317 int user_min_free_kbytes = -1;
318 #ifdef CONFIG_DISCONTIGMEM
320 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
321 * are not on separate NUMA nodes. Functionally this works but with
322 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
323 * quite small. By default, do not boost watermarks on discontigmem as in
324 * many cases very high-order allocations like THP are likely to be
325 * unsupported and the premature reclaim offsets the advantage of long-term
326 * fragmentation avoidance.
328 int watermark_boost_factor __read_mostly;
330 int watermark_boost_factor __read_mostly = 15000;
332 int watermark_scale_factor = 10;
334 static unsigned long nr_kernel_pages __initdata;
335 static unsigned long nr_all_pages __initdata;
336 static unsigned long dma_reserve __initdata;
338 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
339 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
340 static unsigned long required_kernelcore __initdata;
341 static unsigned long required_kernelcore_percent __initdata;
342 static unsigned long required_movablecore __initdata;
343 static unsigned long required_movablecore_percent __initdata;
344 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
345 static bool mirrored_kernelcore __meminitdata;
347 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
349 EXPORT_SYMBOL(movable_zone);
352 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
353 unsigned int nr_online_nodes __read_mostly = 1;
354 EXPORT_SYMBOL(nr_node_ids);
355 EXPORT_SYMBOL(nr_online_nodes);
358 int page_group_by_mobility_disabled __read_mostly;
360 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
362 * During boot we initialize deferred pages on-demand, as needed, but once
363 * page_alloc_init_late() has finished, the deferred pages are all initialized,
364 * and we can permanently disable that path.
366 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
369 * Calling kasan_free_pages() only after deferred memory initialization
370 * has completed. Poisoning pages during deferred memory init will greatly
371 * lengthen the process and cause problem in large memory systems as the
372 * deferred pages initialization is done with interrupt disabled.
374 * Assuming that there will be no reference to those newly initialized
375 * pages before they are ever allocated, this should have no effect on
376 * KASAN memory tracking as the poison will be properly inserted at page
377 * allocation time. The only corner case is when pages are allocated by
378 * on-demand allocation and then freed again before the deferred pages
379 * initialization is done, but this is not likely to happen.
381 static inline void kasan_free_nondeferred_pages(struct page *page, int order)
383 if (!static_branch_unlikely(&deferred_pages))
384 kasan_free_pages(page, order);
387 /* Returns true if the struct page for the pfn is uninitialised */
388 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
390 int nid = early_pfn_to_nid(pfn);
392 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
399 * Returns true when the remaining initialisation should be deferred until
400 * later in the boot cycle when it can be parallelised.
402 static bool __meminit
403 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
405 static unsigned long prev_end_pfn, nr_initialised;
408 * prev_end_pfn static that contains the end of previous zone
409 * No need to protect because called very early in boot before smp_init.
411 if (prev_end_pfn != end_pfn) {
412 prev_end_pfn = end_pfn;
416 /* Always populate low zones for address-constrained allocations */
417 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
421 * We start only with one section of pages, more pages are added as
422 * needed until the rest of deferred pages are initialized.
425 if ((nr_initialised > PAGES_PER_SECTION) &&
426 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
427 NODE_DATA(nid)->first_deferred_pfn = pfn;
433 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
435 static inline bool early_page_uninitialised(unsigned long pfn)
440 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
446 /* Return a pointer to the bitmap storing bits affecting a block of pages */
447 static inline unsigned long *get_pageblock_bitmap(struct page *page,
450 #ifdef CONFIG_SPARSEMEM
451 return section_to_usemap(__pfn_to_section(pfn));
453 return page_zone(page)->pageblock_flags;
454 #endif /* CONFIG_SPARSEMEM */
457 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
459 #ifdef CONFIG_SPARSEMEM
460 pfn &= (PAGES_PER_SECTION-1);
461 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
463 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
464 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
465 #endif /* CONFIG_SPARSEMEM */
469 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
470 * @page: The page within the block of interest
471 * @pfn: The target page frame number
472 * @end_bitidx: The last bit of interest to retrieve
473 * @mask: mask of bits that the caller is interested in
475 * Return: pageblock_bits flags
477 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
479 unsigned long end_bitidx,
482 unsigned long *bitmap;
483 unsigned long bitidx, word_bitidx;
486 bitmap = get_pageblock_bitmap(page, pfn);
487 bitidx = pfn_to_bitidx(page, pfn);
488 word_bitidx = bitidx / BITS_PER_LONG;
489 bitidx &= (BITS_PER_LONG-1);
491 word = bitmap[word_bitidx];
492 bitidx += end_bitidx;
493 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
496 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
497 unsigned long end_bitidx,
500 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
503 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
505 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
509 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
510 * @page: The page within the block of interest
511 * @flags: The flags to set
512 * @pfn: The target page frame number
513 * @end_bitidx: The last bit of interest
514 * @mask: mask of bits that the caller is interested in
516 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
518 unsigned long end_bitidx,
521 unsigned long *bitmap;
522 unsigned long bitidx, word_bitidx;
523 unsigned long old_word, word;
525 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
526 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
528 bitmap = get_pageblock_bitmap(page, pfn);
529 bitidx = pfn_to_bitidx(page, pfn);
530 word_bitidx = bitidx / BITS_PER_LONG;
531 bitidx &= (BITS_PER_LONG-1);
533 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
535 bitidx += end_bitidx;
536 mask <<= (BITS_PER_LONG - bitidx - 1);
537 flags <<= (BITS_PER_LONG - bitidx - 1);
539 word = READ_ONCE(bitmap[word_bitidx]);
541 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
542 if (word == old_word)
548 void set_pageblock_migratetype(struct page *page, int migratetype)
550 if (unlikely(page_group_by_mobility_disabled &&
551 migratetype < MIGRATE_PCPTYPES))
552 migratetype = MIGRATE_UNMOVABLE;
554 set_pageblock_flags_group(page, (unsigned long)migratetype,
555 PB_migrate, PB_migrate_end);
558 #ifdef CONFIG_DEBUG_VM
559 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
563 unsigned long pfn = page_to_pfn(page);
564 unsigned long sp, start_pfn;
567 seq = zone_span_seqbegin(zone);
568 start_pfn = zone->zone_start_pfn;
569 sp = zone->spanned_pages;
570 if (!zone_spans_pfn(zone, pfn))
572 } while (zone_span_seqretry(zone, seq));
575 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
576 pfn, zone_to_nid(zone), zone->name,
577 start_pfn, start_pfn + sp);
582 static int page_is_consistent(struct zone *zone, struct page *page)
584 if (!pfn_valid_within(page_to_pfn(page)))
586 if (zone != page_zone(page))
592 * Temporary debugging check for pages not lying within a given zone.
594 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
596 if (page_outside_zone_boundaries(zone, page))
598 if (!page_is_consistent(zone, page))
604 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
610 static void bad_page(struct page *page, const char *reason,
611 unsigned long bad_flags)
613 static unsigned long resume;
614 static unsigned long nr_shown;
615 static unsigned long nr_unshown;
618 * Allow a burst of 60 reports, then keep quiet for that minute;
619 * or allow a steady drip of one report per second.
621 if (nr_shown == 60) {
622 if (time_before(jiffies, resume)) {
628 "BUG: Bad page state: %lu messages suppressed\n",
635 resume = jiffies + 60 * HZ;
637 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
638 current->comm, page_to_pfn(page));
639 __dump_page(page, reason);
640 bad_flags &= page->flags;
642 pr_alert("bad because of flags: %#lx(%pGp)\n",
643 bad_flags, &bad_flags);
644 dump_page_owner(page);
649 /* Leave bad fields for debug, except PageBuddy could make trouble */
650 page_mapcount_reset(page); /* remove PageBuddy */
651 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
655 * Higher-order pages are called "compound pages". They are structured thusly:
657 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
659 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
660 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
662 * The first tail page's ->compound_dtor holds the offset in array of compound
663 * page destructors. See compound_page_dtors.
665 * The first tail page's ->compound_order holds the order of allocation.
666 * This usage means that zero-order pages may not be compound.
669 void free_compound_page(struct page *page)
671 mem_cgroup_uncharge(page);
672 __free_pages_ok(page, compound_order(page));
675 void prep_compound_page(struct page *page, unsigned int order)
678 int nr_pages = 1 << order;
680 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
681 set_compound_order(page, order);
683 for (i = 1; i < nr_pages; i++) {
684 struct page *p = page + i;
685 set_page_count(p, 0);
686 p->mapping = TAIL_MAPPING;
687 set_compound_head(p, page);
689 atomic_set(compound_mapcount_ptr(page), -1);
690 if (hpage_pincount_available(page))
691 atomic_set(compound_pincount_ptr(page), 0);
694 #ifdef CONFIG_DEBUG_PAGEALLOC
695 unsigned int _debug_guardpage_minorder;
697 bool _debug_pagealloc_enabled_early __read_mostly
698 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
699 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
700 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
701 EXPORT_SYMBOL(_debug_pagealloc_enabled);
703 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
705 static int __init early_debug_pagealloc(char *buf)
707 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
709 early_param("debug_pagealloc", early_debug_pagealloc);
711 void init_debug_pagealloc(void)
713 if (!debug_pagealloc_enabled())
716 static_branch_enable(&_debug_pagealloc_enabled);
718 if (!debug_guardpage_minorder())
721 static_branch_enable(&_debug_guardpage_enabled);
724 static int __init debug_guardpage_minorder_setup(char *buf)
728 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
729 pr_err("Bad debug_guardpage_minorder value\n");
732 _debug_guardpage_minorder = res;
733 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
736 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
738 static inline bool set_page_guard(struct zone *zone, struct page *page,
739 unsigned int order, int migratetype)
741 if (!debug_guardpage_enabled())
744 if (order >= debug_guardpage_minorder())
747 __SetPageGuard(page);
748 INIT_LIST_HEAD(&page->lru);
749 set_page_private(page, order);
750 /* Guard pages are not available for any usage */
751 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
756 static inline void clear_page_guard(struct zone *zone, struct page *page,
757 unsigned int order, int migratetype)
759 if (!debug_guardpage_enabled())
762 __ClearPageGuard(page);
764 set_page_private(page, 0);
765 if (!is_migrate_isolate(migratetype))
766 __mod_zone_freepage_state(zone, (1 << order), migratetype);
769 static inline bool set_page_guard(struct zone *zone, struct page *page,
770 unsigned int order, int migratetype) { return false; }
771 static inline void clear_page_guard(struct zone *zone, struct page *page,
772 unsigned int order, int migratetype) {}
775 static inline void set_page_order(struct page *page, unsigned int order)
777 set_page_private(page, order);
778 __SetPageBuddy(page);
782 * This function checks whether a page is free && is the buddy
783 * we can coalesce a page and its buddy if
784 * (a) the buddy is not in a hole (check before calling!) &&
785 * (b) the buddy is in the buddy system &&
786 * (c) a page and its buddy have the same order &&
787 * (d) a page and its buddy are in the same zone.
789 * For recording whether a page is in the buddy system, we set PageBuddy.
790 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
792 * For recording page's order, we use page_private(page).
794 static inline bool page_is_buddy(struct page *page, struct page *buddy,
797 if (!page_is_guard(buddy) && !PageBuddy(buddy))
800 if (page_order(buddy) != order)
804 * zone check is done late to avoid uselessly calculating
805 * zone/node ids for pages that could never merge.
807 if (page_zone_id(page) != page_zone_id(buddy))
810 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
815 #ifdef CONFIG_COMPACTION
816 static inline struct capture_control *task_capc(struct zone *zone)
818 struct capture_control *capc = current->capture_control;
821 !(current->flags & PF_KTHREAD) &&
823 capc->cc->zone == zone &&
824 capc->cc->direct_compaction ? capc : NULL;
828 compaction_capture(struct capture_control *capc, struct page *page,
829 int order, int migratetype)
831 if (!capc || order != capc->cc->order)
834 /* Do not accidentally pollute CMA or isolated regions*/
835 if (is_migrate_cma(migratetype) ||
836 is_migrate_isolate(migratetype))
840 * Do not let lower order allocations polluate a movable pageblock.
841 * This might let an unmovable request use a reclaimable pageblock
842 * and vice-versa but no more than normal fallback logic which can
843 * have trouble finding a high-order free page.
845 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
853 static inline struct capture_control *task_capc(struct zone *zone)
859 compaction_capture(struct capture_control *capc, struct page *page,
860 int order, int migratetype)
864 #endif /* CONFIG_COMPACTION */
866 /* Used for pages not on another list */
867 static inline void add_to_free_list(struct page *page, struct zone *zone,
868 unsigned int order, int migratetype)
870 struct free_area *area = &zone->free_area[order];
872 list_add(&page->lru, &area->free_list[migratetype]);
876 /* Used for pages not on another list */
877 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
878 unsigned int order, int migratetype)
880 struct free_area *area = &zone->free_area[order];
882 list_add_tail(&page->lru, &area->free_list[migratetype]);
886 /* Used for pages which are on another list */
887 static inline void move_to_free_list(struct page *page, struct zone *zone,
888 unsigned int order, int migratetype)
890 struct free_area *area = &zone->free_area[order];
892 list_move(&page->lru, &area->free_list[migratetype]);
895 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
898 /* clear reported state and update reported page count */
899 if (page_reported(page))
900 __ClearPageReported(page);
902 list_del(&page->lru);
903 __ClearPageBuddy(page);
904 set_page_private(page, 0);
905 zone->free_area[order].nr_free--;
909 * If this is not the largest possible page, check if the buddy
910 * of the next-highest order is free. If it is, it's possible
911 * that pages are being freed that will coalesce soon. In case,
912 * that is happening, add the free page to the tail of the list
913 * so it's less likely to be used soon and more likely to be merged
914 * as a higher order page
917 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
918 struct page *page, unsigned int order)
920 struct page *higher_page, *higher_buddy;
921 unsigned long combined_pfn;
923 if (order >= MAX_ORDER - 2)
926 if (!pfn_valid_within(buddy_pfn))
929 combined_pfn = buddy_pfn & pfn;
930 higher_page = page + (combined_pfn - pfn);
931 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
932 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
934 return pfn_valid_within(buddy_pfn) &&
935 page_is_buddy(higher_page, higher_buddy, order + 1);
939 * Freeing function for a buddy system allocator.
941 * The concept of a buddy system is to maintain direct-mapped table
942 * (containing bit values) for memory blocks of various "orders".
943 * The bottom level table contains the map for the smallest allocatable
944 * units of memory (here, pages), and each level above it describes
945 * pairs of units from the levels below, hence, "buddies".
946 * At a high level, all that happens here is marking the table entry
947 * at the bottom level available, and propagating the changes upward
948 * as necessary, plus some accounting needed to play nicely with other
949 * parts of the VM system.
950 * At each level, we keep a list of pages, which are heads of continuous
951 * free pages of length of (1 << order) and marked with PageBuddy.
952 * Page's order is recorded in page_private(page) field.
953 * So when we are allocating or freeing one, we can derive the state of the
954 * other. That is, if we allocate a small block, and both were
955 * free, the remainder of the region must be split into blocks.
956 * If a block is freed, and its buddy is also free, then this
957 * triggers coalescing into a block of larger size.
962 static inline void __free_one_page(struct page *page,
964 struct zone *zone, unsigned int order,
965 int migratetype, bool report)
967 struct capture_control *capc = task_capc(zone);
968 unsigned long uninitialized_var(buddy_pfn);
969 unsigned long combined_pfn;
970 unsigned int max_order;
974 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
976 VM_BUG_ON(!zone_is_initialized(zone));
977 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
979 VM_BUG_ON(migratetype == -1);
980 if (likely(!is_migrate_isolate(migratetype)))
981 __mod_zone_freepage_state(zone, 1 << order, migratetype);
983 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
984 VM_BUG_ON_PAGE(bad_range(zone, page), page);
987 while (order < max_order - 1) {
988 if (compaction_capture(capc, page, order, migratetype)) {
989 __mod_zone_freepage_state(zone, -(1 << order),
993 buddy_pfn = __find_buddy_pfn(pfn, order);
994 buddy = page + (buddy_pfn - pfn);
996 if (!pfn_valid_within(buddy_pfn))
998 if (!page_is_buddy(page, buddy, order))
1001 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1002 * merge with it and move up one order.
1004 if (page_is_guard(buddy))
1005 clear_page_guard(zone, buddy, order, migratetype);
1007 del_page_from_free_list(buddy, zone, order);
1008 combined_pfn = buddy_pfn & pfn;
1009 page = page + (combined_pfn - pfn);
1013 if (max_order < MAX_ORDER) {
1014 /* If we are here, it means order is >= pageblock_order.
1015 * We want to prevent merge between freepages on isolate
1016 * pageblock and normal pageblock. Without this, pageblock
1017 * isolation could cause incorrect freepage or CMA accounting.
1019 * We don't want to hit this code for the more frequent
1020 * low-order merging.
1022 if (unlikely(has_isolate_pageblock(zone))) {
1025 buddy_pfn = __find_buddy_pfn(pfn, order);
1026 buddy = page + (buddy_pfn - pfn);
1027 buddy_mt = get_pageblock_migratetype(buddy);
1029 if (migratetype != buddy_mt
1030 && (is_migrate_isolate(migratetype) ||
1031 is_migrate_isolate(buddy_mt)))
1035 goto continue_merging;
1039 set_page_order(page, order);
1041 if (is_shuffle_order(order))
1042 to_tail = shuffle_pick_tail();
1044 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1047 add_to_free_list_tail(page, zone, order, migratetype);
1049 add_to_free_list(page, zone, order, migratetype);
1051 /* Notify page reporting subsystem of freed page */
1053 page_reporting_notify_free(order);
1057 * A bad page could be due to a number of fields. Instead of multiple branches,
1058 * try and check multiple fields with one check. The caller must do a detailed
1059 * check if necessary.
1061 static inline bool page_expected_state(struct page *page,
1062 unsigned long check_flags)
1064 if (unlikely(atomic_read(&page->_mapcount) != -1))
1067 if (unlikely((unsigned long)page->mapping |
1068 page_ref_count(page) |
1070 (unsigned long)page->mem_cgroup |
1072 (page->flags & check_flags)))
1078 static void free_pages_check_bad(struct page *page)
1080 const char *bad_reason;
1081 unsigned long bad_flags;
1086 if (unlikely(atomic_read(&page->_mapcount) != -1))
1087 bad_reason = "nonzero mapcount";
1088 if (unlikely(page->mapping != NULL))
1089 bad_reason = "non-NULL mapping";
1090 if (unlikely(page_ref_count(page) != 0))
1091 bad_reason = "nonzero _refcount";
1092 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
1093 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1094 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
1097 if (unlikely(page->mem_cgroup))
1098 bad_reason = "page still charged to cgroup";
1100 bad_page(page, bad_reason, bad_flags);
1103 static inline int free_pages_check(struct page *page)
1105 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1108 /* Something has gone sideways, find it */
1109 free_pages_check_bad(page);
1113 static int free_tail_pages_check(struct page *head_page, struct page *page)
1118 * We rely page->lru.next never has bit 0 set, unless the page
1119 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1121 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1123 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1127 switch (page - head_page) {
1129 /* the first tail page: ->mapping may be compound_mapcount() */
1130 if (unlikely(compound_mapcount(page))) {
1131 bad_page(page, "nonzero compound_mapcount", 0);
1137 * the second tail page: ->mapping is
1138 * deferred_list.next -- ignore value.
1142 if (page->mapping != TAIL_MAPPING) {
1143 bad_page(page, "corrupted mapping in tail page", 0);
1148 if (unlikely(!PageTail(page))) {
1149 bad_page(page, "PageTail not set", 0);
1152 if (unlikely(compound_head(page) != head_page)) {
1153 bad_page(page, "compound_head not consistent", 0);
1158 page->mapping = NULL;
1159 clear_compound_head(page);
1163 static void kernel_init_free_pages(struct page *page, int numpages)
1167 for (i = 0; i < numpages; i++)
1168 clear_highpage(page + i);
1171 static __always_inline bool free_pages_prepare(struct page *page,
1172 unsigned int order, bool check_free)
1176 VM_BUG_ON_PAGE(PageTail(page), page);
1178 trace_mm_page_free(page, order);
1181 * Check tail pages before head page information is cleared to
1182 * avoid checking PageCompound for order-0 pages.
1184 if (unlikely(order)) {
1185 bool compound = PageCompound(page);
1188 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1191 ClearPageDoubleMap(page);
1192 for (i = 1; i < (1 << order); i++) {
1194 bad += free_tail_pages_check(page, page + i);
1195 if (unlikely(free_pages_check(page + i))) {
1199 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1202 if (PageMappingFlags(page))
1203 page->mapping = NULL;
1204 if (memcg_kmem_enabled() && PageKmemcg(page))
1205 __memcg_kmem_uncharge_page(page, order);
1207 bad += free_pages_check(page);
1211 page_cpupid_reset_last(page);
1212 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1213 reset_page_owner(page, order);
1215 if (!PageHighMem(page)) {
1216 debug_check_no_locks_freed(page_address(page),
1217 PAGE_SIZE << order);
1218 debug_check_no_obj_freed(page_address(page),
1219 PAGE_SIZE << order);
1221 if (want_init_on_free())
1222 kernel_init_free_pages(page, 1 << order);
1224 kernel_poison_pages(page, 1 << order, 0);
1226 * arch_free_page() can make the page's contents inaccessible. s390
1227 * does this. So nothing which can access the page's contents should
1228 * happen after this.
1230 arch_free_page(page, order);
1232 if (debug_pagealloc_enabled_static())
1233 kernel_map_pages(page, 1 << order, 0);
1235 kasan_free_nondeferred_pages(page, order);
1240 #ifdef CONFIG_DEBUG_VM
1242 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1243 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1244 * moved from pcp lists to free lists.
1246 static bool free_pcp_prepare(struct page *page)
1248 return free_pages_prepare(page, 0, true);
1251 static bool bulkfree_pcp_prepare(struct page *page)
1253 if (debug_pagealloc_enabled_static())
1254 return free_pages_check(page);
1260 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1261 * moving from pcp lists to free list in order to reduce overhead. With
1262 * debug_pagealloc enabled, they are checked also immediately when being freed
1265 static bool free_pcp_prepare(struct page *page)
1267 if (debug_pagealloc_enabled_static())
1268 return free_pages_prepare(page, 0, true);
1270 return free_pages_prepare(page, 0, false);
1273 static bool bulkfree_pcp_prepare(struct page *page)
1275 return free_pages_check(page);
1277 #endif /* CONFIG_DEBUG_VM */
1279 static inline void prefetch_buddy(struct page *page)
1281 unsigned long pfn = page_to_pfn(page);
1282 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1283 struct page *buddy = page + (buddy_pfn - pfn);
1289 * Frees a number of pages from the PCP lists
1290 * Assumes all pages on list are in same zone, and of same order.
1291 * count is the number of pages to free.
1293 * If the zone was previously in an "all pages pinned" state then look to
1294 * see if this freeing clears that state.
1296 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1297 * pinned" detection logic.
1299 static void free_pcppages_bulk(struct zone *zone, int count,
1300 struct per_cpu_pages *pcp)
1302 int migratetype = 0;
1304 int prefetch_nr = 0;
1305 bool isolated_pageblocks;
1306 struct page *page, *tmp;
1310 struct list_head *list;
1313 * Remove pages from lists in a round-robin fashion. A
1314 * batch_free count is maintained that is incremented when an
1315 * empty list is encountered. This is so more pages are freed
1316 * off fuller lists instead of spinning excessively around empty
1321 if (++migratetype == MIGRATE_PCPTYPES)
1323 list = &pcp->lists[migratetype];
1324 } while (list_empty(list));
1326 /* This is the only non-empty list. Free them all. */
1327 if (batch_free == MIGRATE_PCPTYPES)
1331 page = list_last_entry(list, struct page, lru);
1332 /* must delete to avoid corrupting pcp list */
1333 list_del(&page->lru);
1336 if (bulkfree_pcp_prepare(page))
1339 list_add_tail(&page->lru, &head);
1342 * We are going to put the page back to the global
1343 * pool, prefetch its buddy to speed up later access
1344 * under zone->lock. It is believed the overhead of
1345 * an additional test and calculating buddy_pfn here
1346 * can be offset by reduced memory latency later. To
1347 * avoid excessive prefetching due to large count, only
1348 * prefetch buddy for the first pcp->batch nr of pages.
1350 if (prefetch_nr++ < pcp->batch)
1351 prefetch_buddy(page);
1352 } while (--count && --batch_free && !list_empty(list));
1355 spin_lock(&zone->lock);
1356 isolated_pageblocks = has_isolate_pageblock(zone);
1359 * Use safe version since after __free_one_page(),
1360 * page->lru.next will not point to original list.
1362 list_for_each_entry_safe(page, tmp, &head, lru) {
1363 int mt = get_pcppage_migratetype(page);
1364 /* MIGRATE_ISOLATE page should not go to pcplists */
1365 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1366 /* Pageblock could have been isolated meanwhile */
1367 if (unlikely(isolated_pageblocks))
1368 mt = get_pageblock_migratetype(page);
1370 __free_one_page(page, page_to_pfn(page), zone, 0, mt, true);
1371 trace_mm_page_pcpu_drain(page, 0, mt);
1373 spin_unlock(&zone->lock);
1376 static void free_one_page(struct zone *zone,
1377 struct page *page, unsigned long pfn,
1381 spin_lock(&zone->lock);
1382 if (unlikely(has_isolate_pageblock(zone) ||
1383 is_migrate_isolate(migratetype))) {
1384 migratetype = get_pfnblock_migratetype(page, pfn);
1386 __free_one_page(page, pfn, zone, order, migratetype, true);
1387 spin_unlock(&zone->lock);
1390 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1391 unsigned long zone, int nid)
1393 mm_zero_struct_page(page);
1394 set_page_links(page, zone, nid, pfn);
1395 init_page_count(page);
1396 page_mapcount_reset(page);
1397 page_cpupid_reset_last(page);
1398 page_kasan_tag_reset(page);
1400 INIT_LIST_HEAD(&page->lru);
1401 #ifdef WANT_PAGE_VIRTUAL
1402 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1403 if (!is_highmem_idx(zone))
1404 set_page_address(page, __va(pfn << PAGE_SHIFT));
1408 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1409 static void __meminit init_reserved_page(unsigned long pfn)
1414 if (!early_page_uninitialised(pfn))
1417 nid = early_pfn_to_nid(pfn);
1418 pgdat = NODE_DATA(nid);
1420 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1421 struct zone *zone = &pgdat->node_zones[zid];
1423 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1426 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1429 static inline void init_reserved_page(unsigned long pfn)
1432 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1435 * Initialised pages do not have PageReserved set. This function is
1436 * called for each range allocated by the bootmem allocator and
1437 * marks the pages PageReserved. The remaining valid pages are later
1438 * sent to the buddy page allocator.
1440 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1442 unsigned long start_pfn = PFN_DOWN(start);
1443 unsigned long end_pfn = PFN_UP(end);
1445 for (; start_pfn < end_pfn; start_pfn++) {
1446 if (pfn_valid(start_pfn)) {
1447 struct page *page = pfn_to_page(start_pfn);
1449 init_reserved_page(start_pfn);
1451 /* Avoid false-positive PageTail() */
1452 INIT_LIST_HEAD(&page->lru);
1455 * no need for atomic set_bit because the struct
1456 * page is not visible yet so nobody should
1459 __SetPageReserved(page);
1464 static void __free_pages_ok(struct page *page, unsigned int order)
1466 unsigned long flags;
1468 unsigned long pfn = page_to_pfn(page);
1470 if (!free_pages_prepare(page, order, true))
1473 migratetype = get_pfnblock_migratetype(page, pfn);
1474 local_irq_save(flags);
1475 __count_vm_events(PGFREE, 1 << order);
1476 free_one_page(page_zone(page), page, pfn, order, migratetype);
1477 local_irq_restore(flags);
1480 void __free_pages_core(struct page *page, unsigned int order)
1482 unsigned int nr_pages = 1 << order;
1483 struct page *p = page;
1487 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1489 __ClearPageReserved(p);
1490 set_page_count(p, 0);
1492 __ClearPageReserved(p);
1493 set_page_count(p, 0);
1495 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1496 set_page_refcounted(page);
1497 __free_pages(page, order);
1500 #ifdef CONFIG_NEED_MULTIPLE_NODES
1502 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1504 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
1507 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1509 int __meminit __early_pfn_to_nid(unsigned long pfn,
1510 struct mminit_pfnnid_cache *state)
1512 unsigned long start_pfn, end_pfn;
1515 if (state->last_start <= pfn && pfn < state->last_end)
1516 return state->last_nid;
1518 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1519 if (nid != NUMA_NO_NODE) {
1520 state->last_start = start_pfn;
1521 state->last_end = end_pfn;
1522 state->last_nid = nid;
1527 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
1529 int __meminit early_pfn_to_nid(unsigned long pfn)
1531 static DEFINE_SPINLOCK(early_pfn_lock);
1534 spin_lock(&early_pfn_lock);
1535 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1537 nid = first_online_node;
1538 spin_unlock(&early_pfn_lock);
1542 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1544 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1547 if (early_page_uninitialised(pfn))
1549 __free_pages_core(page, order);
1553 * Check that the whole (or subset of) a pageblock given by the interval of
1554 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1555 * with the migration of free compaction scanner. The scanners then need to
1556 * use only pfn_valid_within() check for arches that allow holes within
1559 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1561 * It's possible on some configurations to have a setup like node0 node1 node0
1562 * i.e. it's possible that all pages within a zones range of pages do not
1563 * belong to a single zone. We assume that a border between node0 and node1
1564 * can occur within a single pageblock, but not a node0 node1 node0
1565 * interleaving within a single pageblock. It is therefore sufficient to check
1566 * the first and last page of a pageblock and avoid checking each individual
1567 * page in a pageblock.
1569 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1570 unsigned long end_pfn, struct zone *zone)
1572 struct page *start_page;
1573 struct page *end_page;
1575 /* end_pfn is one past the range we are checking */
1578 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1581 start_page = pfn_to_online_page(start_pfn);
1585 if (page_zone(start_page) != zone)
1588 end_page = pfn_to_page(end_pfn);
1590 /* This gives a shorter code than deriving page_zone(end_page) */
1591 if (page_zone_id(start_page) != page_zone_id(end_page))
1597 void set_zone_contiguous(struct zone *zone)
1599 unsigned long block_start_pfn = zone->zone_start_pfn;
1600 unsigned long block_end_pfn;
1602 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1603 for (; block_start_pfn < zone_end_pfn(zone);
1604 block_start_pfn = block_end_pfn,
1605 block_end_pfn += pageblock_nr_pages) {
1607 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1609 if (!__pageblock_pfn_to_page(block_start_pfn,
1610 block_end_pfn, zone))
1615 /* We confirm that there is no hole */
1616 zone->contiguous = true;
1619 void clear_zone_contiguous(struct zone *zone)
1621 zone->contiguous = false;
1624 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1625 static void __init deferred_free_range(unsigned long pfn,
1626 unsigned long nr_pages)
1634 page = pfn_to_page(pfn);
1636 /* Free a large naturally-aligned chunk if possible */
1637 if (nr_pages == pageblock_nr_pages &&
1638 (pfn & (pageblock_nr_pages - 1)) == 0) {
1639 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1640 __free_pages_core(page, pageblock_order);
1644 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1645 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1646 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1647 __free_pages_core(page, 0);
1651 /* Completion tracking for deferred_init_memmap() threads */
1652 static atomic_t pgdat_init_n_undone __initdata;
1653 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1655 static inline void __init pgdat_init_report_one_done(void)
1657 if (atomic_dec_and_test(&pgdat_init_n_undone))
1658 complete(&pgdat_init_all_done_comp);
1662 * Returns true if page needs to be initialized or freed to buddy allocator.
1664 * First we check if pfn is valid on architectures where it is possible to have
1665 * holes within pageblock_nr_pages. On systems where it is not possible, this
1666 * function is optimized out.
1668 * Then, we check if a current large page is valid by only checking the validity
1671 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1673 if (!pfn_valid_within(pfn))
1675 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1681 * Free pages to buddy allocator. Try to free aligned pages in
1682 * pageblock_nr_pages sizes.
1684 static void __init deferred_free_pages(unsigned long pfn,
1685 unsigned long end_pfn)
1687 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1688 unsigned long nr_free = 0;
1690 for (; pfn < end_pfn; pfn++) {
1691 if (!deferred_pfn_valid(pfn)) {
1692 deferred_free_range(pfn - nr_free, nr_free);
1694 } else if (!(pfn & nr_pgmask)) {
1695 deferred_free_range(pfn - nr_free, nr_free);
1697 touch_nmi_watchdog();
1702 /* Free the last block of pages to allocator */
1703 deferred_free_range(pfn - nr_free, nr_free);
1707 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1708 * by performing it only once every pageblock_nr_pages.
1709 * Return number of pages initialized.
1711 static unsigned long __init deferred_init_pages(struct zone *zone,
1713 unsigned long end_pfn)
1715 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1716 int nid = zone_to_nid(zone);
1717 unsigned long nr_pages = 0;
1718 int zid = zone_idx(zone);
1719 struct page *page = NULL;
1721 for (; pfn < end_pfn; pfn++) {
1722 if (!deferred_pfn_valid(pfn)) {
1725 } else if (!page || !(pfn & nr_pgmask)) {
1726 page = pfn_to_page(pfn);
1727 touch_nmi_watchdog();
1731 __init_single_page(page, pfn, zid, nid);
1738 * This function is meant to pre-load the iterator for the zone init.
1739 * Specifically it walks through the ranges until we are caught up to the
1740 * first_init_pfn value and exits there. If we never encounter the value we
1741 * return false indicating there are no valid ranges left.
1744 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1745 unsigned long *spfn, unsigned long *epfn,
1746 unsigned long first_init_pfn)
1751 * Start out by walking through the ranges in this zone that have
1752 * already been initialized. We don't need to do anything with them
1753 * so we just need to flush them out of the system.
1755 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1756 if (*epfn <= first_init_pfn)
1758 if (*spfn < first_init_pfn)
1759 *spfn = first_init_pfn;
1768 * Initialize and free pages. We do it in two loops: first we initialize
1769 * struct page, then free to buddy allocator, because while we are
1770 * freeing pages we can access pages that are ahead (computing buddy
1771 * page in __free_one_page()).
1773 * In order to try and keep some memory in the cache we have the loop
1774 * broken along max page order boundaries. This way we will not cause
1775 * any issues with the buddy page computation.
1777 static unsigned long __init
1778 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1779 unsigned long *end_pfn)
1781 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1782 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1783 unsigned long nr_pages = 0;
1786 /* First we loop through and initialize the page values */
1787 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1790 if (mo_pfn <= *start_pfn)
1793 t = min(mo_pfn, *end_pfn);
1794 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1796 if (mo_pfn < *end_pfn) {
1797 *start_pfn = mo_pfn;
1802 /* Reset values and now loop through freeing pages as needed */
1805 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1811 t = min(mo_pfn, epfn);
1812 deferred_free_pages(spfn, t);
1821 /* Initialise remaining memory on a node */
1822 static int __init deferred_init_memmap(void *data)
1824 pg_data_t *pgdat = data;
1825 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1826 unsigned long spfn = 0, epfn = 0, nr_pages = 0;
1827 unsigned long first_init_pfn, flags;
1828 unsigned long start = jiffies;
1833 /* Bind memory initialisation thread to a local node if possible */
1834 if (!cpumask_empty(cpumask))
1835 set_cpus_allowed_ptr(current, cpumask);
1837 pgdat_resize_lock(pgdat, &flags);
1838 first_init_pfn = pgdat->first_deferred_pfn;
1839 if (first_init_pfn == ULONG_MAX) {
1840 pgdat_resize_unlock(pgdat, &flags);
1841 pgdat_init_report_one_done();
1845 /* Sanity check boundaries */
1846 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1847 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1848 pgdat->first_deferred_pfn = ULONG_MAX;
1850 /* Only the highest zone is deferred so find it */
1851 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1852 zone = pgdat->node_zones + zid;
1853 if (first_init_pfn < zone_end_pfn(zone))
1857 /* If the zone is empty somebody else may have cleared out the zone */
1858 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1863 * Initialize and free pages in MAX_ORDER sized increments so
1864 * that we can avoid introducing any issues with the buddy
1868 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1870 pgdat_resize_unlock(pgdat, &flags);
1872 /* Sanity check that the next zone really is unpopulated */
1873 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1875 pr_info("node %d initialised, %lu pages in %ums\n",
1876 pgdat->node_id, nr_pages, jiffies_to_msecs(jiffies - start));
1878 pgdat_init_report_one_done();
1883 * If this zone has deferred pages, try to grow it by initializing enough
1884 * deferred pages to satisfy the allocation specified by order, rounded up to
1885 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1886 * of SECTION_SIZE bytes by initializing struct pages in increments of
1887 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1889 * Return true when zone was grown, otherwise return false. We return true even
1890 * when we grow less than requested, to let the caller decide if there are
1891 * enough pages to satisfy the allocation.
1893 * Note: We use noinline because this function is needed only during boot, and
1894 * it is called from a __ref function _deferred_grow_zone. This way we are
1895 * making sure that it is not inlined into permanent text section.
1897 static noinline bool __init
1898 deferred_grow_zone(struct zone *zone, unsigned int order)
1900 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1901 pg_data_t *pgdat = zone->zone_pgdat;
1902 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1903 unsigned long spfn, epfn, flags;
1904 unsigned long nr_pages = 0;
1907 /* Only the last zone may have deferred pages */
1908 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1911 pgdat_resize_lock(pgdat, &flags);
1914 * If deferred pages have been initialized while we were waiting for
1915 * the lock, return true, as the zone was grown. The caller will retry
1916 * this zone. We won't return to this function since the caller also
1917 * has this static branch.
1919 if (!static_branch_unlikely(&deferred_pages)) {
1920 pgdat_resize_unlock(pgdat, &flags);
1925 * If someone grew this zone while we were waiting for spinlock, return
1926 * true, as there might be enough pages already.
1928 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1929 pgdat_resize_unlock(pgdat, &flags);
1933 /* If the zone is empty somebody else may have cleared out the zone */
1934 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1935 first_deferred_pfn)) {
1936 pgdat->first_deferred_pfn = ULONG_MAX;
1937 pgdat_resize_unlock(pgdat, &flags);
1938 /* Retry only once. */
1939 return first_deferred_pfn != ULONG_MAX;
1943 * Initialize and free pages in MAX_ORDER sized increments so
1944 * that we can avoid introducing any issues with the buddy
1947 while (spfn < epfn) {
1948 /* update our first deferred PFN for this section */
1949 first_deferred_pfn = spfn;
1951 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1953 /* We should only stop along section boundaries */
1954 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
1957 /* If our quota has been met we can stop here */
1958 if (nr_pages >= nr_pages_needed)
1962 pgdat->first_deferred_pfn = spfn;
1963 pgdat_resize_unlock(pgdat, &flags);
1965 return nr_pages > 0;
1969 * deferred_grow_zone() is __init, but it is called from
1970 * get_page_from_freelist() during early boot until deferred_pages permanently
1971 * disables this call. This is why we have refdata wrapper to avoid warning,
1972 * and to ensure that the function body gets unloaded.
1975 _deferred_grow_zone(struct zone *zone, unsigned int order)
1977 return deferred_grow_zone(zone, order);
1980 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1982 void __init page_alloc_init_late(void)
1987 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1989 /* There will be num_node_state(N_MEMORY) threads */
1990 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1991 for_each_node_state(nid, N_MEMORY) {
1992 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1995 /* Block until all are initialised */
1996 wait_for_completion(&pgdat_init_all_done_comp);
1999 * The number of managed pages has changed due to the initialisation
2000 * so the pcpu batch and high limits needs to be updated or the limits
2001 * will be artificially small.
2003 for_each_populated_zone(zone)
2004 zone_pcp_update(zone);
2007 * We initialized the rest of the deferred pages. Permanently disable
2008 * on-demand struct page initialization.
2010 static_branch_disable(&deferred_pages);
2012 /* Reinit limits that are based on free pages after the kernel is up */
2013 files_maxfiles_init();
2016 /* Discard memblock private memory */
2019 for_each_node_state(nid, N_MEMORY)
2020 shuffle_free_memory(NODE_DATA(nid));
2022 for_each_populated_zone(zone)
2023 set_zone_contiguous(zone);
2027 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2028 void __init init_cma_reserved_pageblock(struct page *page)
2030 unsigned i = pageblock_nr_pages;
2031 struct page *p = page;
2034 __ClearPageReserved(p);
2035 set_page_count(p, 0);
2038 set_pageblock_migratetype(page, MIGRATE_CMA);
2040 if (pageblock_order >= MAX_ORDER) {
2041 i = pageblock_nr_pages;
2044 set_page_refcounted(p);
2045 __free_pages(p, MAX_ORDER - 1);
2046 p += MAX_ORDER_NR_PAGES;
2047 } while (i -= MAX_ORDER_NR_PAGES);
2049 set_page_refcounted(page);
2050 __free_pages(page, pageblock_order);
2053 adjust_managed_page_count(page, pageblock_nr_pages);
2058 * The order of subdivision here is critical for the IO subsystem.
2059 * Please do not alter this order without good reasons and regression
2060 * testing. Specifically, as large blocks of memory are subdivided,
2061 * the order in which smaller blocks are delivered depends on the order
2062 * they're subdivided in this function. This is the primary factor
2063 * influencing the order in which pages are delivered to the IO
2064 * subsystem according to empirical testing, and this is also justified
2065 * by considering the behavior of a buddy system containing a single
2066 * large block of memory acted on by a series of small allocations.
2067 * This behavior is a critical factor in sglist merging's success.
2071 static inline void expand(struct zone *zone, struct page *page,
2072 int low, int high, int migratetype)
2074 unsigned long size = 1 << high;
2076 while (high > low) {
2079 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2082 * Mark as guard pages (or page), that will allow to
2083 * merge back to allocator when buddy will be freed.
2084 * Corresponding page table entries will not be touched,
2085 * pages will stay not present in virtual address space
2087 if (set_page_guard(zone, &page[size], high, migratetype))
2090 add_to_free_list(&page[size], zone, high, migratetype);
2091 set_page_order(&page[size], high);
2095 static void check_new_page_bad(struct page *page)
2097 const char *bad_reason = NULL;
2098 unsigned long bad_flags = 0;
2100 if (unlikely(atomic_read(&page->_mapcount) != -1))
2101 bad_reason = "nonzero mapcount";
2102 if (unlikely(page->mapping != NULL))
2103 bad_reason = "non-NULL mapping";
2104 if (unlikely(page_ref_count(page) != 0))
2105 bad_reason = "nonzero _refcount";
2106 if (unlikely(page->flags & __PG_HWPOISON)) {
2107 bad_reason = "HWPoisoned (hardware-corrupted)";
2108 bad_flags = __PG_HWPOISON;
2109 /* Don't complain about hwpoisoned pages */
2110 page_mapcount_reset(page); /* remove PageBuddy */
2113 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
2114 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
2115 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
2118 if (unlikely(page->mem_cgroup))
2119 bad_reason = "page still charged to cgroup";
2121 bad_page(page, bad_reason, bad_flags);
2125 * This page is about to be returned from the page allocator
2127 static inline int check_new_page(struct page *page)
2129 if (likely(page_expected_state(page,
2130 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2133 check_new_page_bad(page);
2137 static inline bool free_pages_prezeroed(void)
2139 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2140 page_poisoning_enabled()) || want_init_on_free();
2143 #ifdef CONFIG_DEBUG_VM
2145 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2146 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2147 * also checked when pcp lists are refilled from the free lists.
2149 static inline bool check_pcp_refill(struct page *page)
2151 if (debug_pagealloc_enabled_static())
2152 return check_new_page(page);
2157 static inline bool check_new_pcp(struct page *page)
2159 return check_new_page(page);
2163 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2164 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2165 * enabled, they are also checked when being allocated from the pcp lists.
2167 static inline bool check_pcp_refill(struct page *page)
2169 return check_new_page(page);
2171 static inline bool check_new_pcp(struct page *page)
2173 if (debug_pagealloc_enabled_static())
2174 return check_new_page(page);
2178 #endif /* CONFIG_DEBUG_VM */
2180 static bool check_new_pages(struct page *page, unsigned int order)
2183 for (i = 0; i < (1 << order); i++) {
2184 struct page *p = page + i;
2186 if (unlikely(check_new_page(p)))
2193 inline void post_alloc_hook(struct page *page, unsigned int order,
2196 set_page_private(page, 0);
2197 set_page_refcounted(page);
2199 arch_alloc_page(page, order);
2200 if (debug_pagealloc_enabled_static())
2201 kernel_map_pages(page, 1 << order, 1);
2202 kasan_alloc_pages(page, order);
2203 kernel_poison_pages(page, 1 << order, 1);
2204 set_page_owner(page, order, gfp_flags);
2207 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2208 unsigned int alloc_flags)
2210 post_alloc_hook(page, order, gfp_flags);
2212 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2213 kernel_init_free_pages(page, 1 << order);
2215 if (order && (gfp_flags & __GFP_COMP))
2216 prep_compound_page(page, order);
2219 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2220 * allocate the page. The expectation is that the caller is taking
2221 * steps that will free more memory. The caller should avoid the page
2222 * being used for !PFMEMALLOC purposes.
2224 if (alloc_flags & ALLOC_NO_WATERMARKS)
2225 set_page_pfmemalloc(page);
2227 clear_page_pfmemalloc(page);
2231 * Go through the free lists for the given migratetype and remove
2232 * the smallest available page from the freelists
2234 static __always_inline
2235 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2238 unsigned int current_order;
2239 struct free_area *area;
2242 /* Find a page of the appropriate size in the preferred list */
2243 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2244 area = &(zone->free_area[current_order]);
2245 page = get_page_from_free_area(area, migratetype);
2248 del_page_from_free_list(page, zone, current_order);
2249 expand(zone, page, order, current_order, migratetype);
2250 set_pcppage_migratetype(page, migratetype);
2259 * This array describes the order lists are fallen back to when
2260 * the free lists for the desirable migrate type are depleted
2262 static int fallbacks[MIGRATE_TYPES][4] = {
2263 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2264 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2265 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2267 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2269 #ifdef CONFIG_MEMORY_ISOLATION
2270 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2275 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2278 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2281 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2282 unsigned int order) { return NULL; }
2286 * Move the free pages in a range to the free lists of the requested type.
2287 * Note that start_page and end_pages are not aligned on a pageblock
2288 * boundary. If alignment is required, use move_freepages_block()
2290 static int move_freepages(struct zone *zone,
2291 struct page *start_page, struct page *end_page,
2292 int migratetype, int *num_movable)
2296 int pages_moved = 0;
2298 for (page = start_page; page <= end_page;) {
2299 if (!pfn_valid_within(page_to_pfn(page))) {
2304 if (!PageBuddy(page)) {
2306 * We assume that pages that could be isolated for
2307 * migration are movable. But we don't actually try
2308 * isolating, as that would be expensive.
2311 (PageLRU(page) || __PageMovable(page)))
2318 /* Make sure we are not inadvertently changing nodes */
2319 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2320 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2322 order = page_order(page);
2323 move_to_free_list(page, zone, order, migratetype);
2325 pages_moved += 1 << order;
2331 int move_freepages_block(struct zone *zone, struct page *page,
2332 int migratetype, int *num_movable)
2334 unsigned long start_pfn, end_pfn;
2335 struct page *start_page, *end_page;
2340 start_pfn = page_to_pfn(page);
2341 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2342 start_page = pfn_to_page(start_pfn);
2343 end_page = start_page + pageblock_nr_pages - 1;
2344 end_pfn = start_pfn + pageblock_nr_pages - 1;
2346 /* Do not cross zone boundaries */
2347 if (!zone_spans_pfn(zone, start_pfn))
2349 if (!zone_spans_pfn(zone, end_pfn))
2352 return move_freepages(zone, start_page, end_page, migratetype,
2356 static void change_pageblock_range(struct page *pageblock_page,
2357 int start_order, int migratetype)
2359 int nr_pageblocks = 1 << (start_order - pageblock_order);
2361 while (nr_pageblocks--) {
2362 set_pageblock_migratetype(pageblock_page, migratetype);
2363 pageblock_page += pageblock_nr_pages;
2368 * When we are falling back to another migratetype during allocation, try to
2369 * steal extra free pages from the same pageblocks to satisfy further
2370 * allocations, instead of polluting multiple pageblocks.
2372 * If we are stealing a relatively large buddy page, it is likely there will
2373 * be more free pages in the pageblock, so try to steal them all. For
2374 * reclaimable and unmovable allocations, we steal regardless of page size,
2375 * as fragmentation caused by those allocations polluting movable pageblocks
2376 * is worse than movable allocations stealing from unmovable and reclaimable
2379 static bool can_steal_fallback(unsigned int order, int start_mt)
2382 * Leaving this order check is intended, although there is
2383 * relaxed order check in next check. The reason is that
2384 * we can actually steal whole pageblock if this condition met,
2385 * but, below check doesn't guarantee it and that is just heuristic
2386 * so could be changed anytime.
2388 if (order >= pageblock_order)
2391 if (order >= pageblock_order / 2 ||
2392 start_mt == MIGRATE_RECLAIMABLE ||
2393 start_mt == MIGRATE_UNMOVABLE ||
2394 page_group_by_mobility_disabled)
2400 static inline void boost_watermark(struct zone *zone)
2402 unsigned long max_boost;
2404 if (!watermark_boost_factor)
2407 * Don't bother in zones that are unlikely to produce results.
2408 * On small machines, including kdump capture kernels running
2409 * in a small area, boosting the watermark can cause an out of
2410 * memory situation immediately.
2412 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2415 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2416 watermark_boost_factor, 10000);
2419 * high watermark may be uninitialised if fragmentation occurs
2420 * very early in boot so do not boost. We do not fall
2421 * through and boost by pageblock_nr_pages as failing
2422 * allocations that early means that reclaim is not going
2423 * to help and it may even be impossible to reclaim the
2424 * boosted watermark resulting in a hang.
2429 max_boost = max(pageblock_nr_pages, max_boost);
2431 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2436 * This function implements actual steal behaviour. If order is large enough,
2437 * we can steal whole pageblock. If not, we first move freepages in this
2438 * pageblock to our migratetype and determine how many already-allocated pages
2439 * are there in the pageblock with a compatible migratetype. If at least half
2440 * of pages are free or compatible, we can change migratetype of the pageblock
2441 * itself, so pages freed in the future will be put on the correct free list.
2443 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2444 unsigned int alloc_flags, int start_type, bool whole_block)
2446 unsigned int current_order = page_order(page);
2447 int free_pages, movable_pages, alike_pages;
2450 old_block_type = get_pageblock_migratetype(page);
2453 * This can happen due to races and we want to prevent broken
2454 * highatomic accounting.
2456 if (is_migrate_highatomic(old_block_type))
2459 /* Take ownership for orders >= pageblock_order */
2460 if (current_order >= pageblock_order) {
2461 change_pageblock_range(page, current_order, start_type);
2466 * Boost watermarks to increase reclaim pressure to reduce the
2467 * likelihood of future fallbacks. Wake kswapd now as the node
2468 * may be balanced overall and kswapd will not wake naturally.
2470 boost_watermark(zone);
2471 if (alloc_flags & ALLOC_KSWAPD)
2472 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2474 /* We are not allowed to try stealing from the whole block */
2478 free_pages = move_freepages_block(zone, page, start_type,
2481 * Determine how many pages are compatible with our allocation.
2482 * For movable allocation, it's the number of movable pages which
2483 * we just obtained. For other types it's a bit more tricky.
2485 if (start_type == MIGRATE_MOVABLE) {
2486 alike_pages = movable_pages;
2489 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2490 * to MOVABLE pageblock, consider all non-movable pages as
2491 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2492 * vice versa, be conservative since we can't distinguish the
2493 * exact migratetype of non-movable pages.
2495 if (old_block_type == MIGRATE_MOVABLE)
2496 alike_pages = pageblock_nr_pages
2497 - (free_pages + movable_pages);
2502 /* moving whole block can fail due to zone boundary conditions */
2507 * If a sufficient number of pages in the block are either free or of
2508 * comparable migratability as our allocation, claim the whole block.
2510 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2511 page_group_by_mobility_disabled)
2512 set_pageblock_migratetype(page, start_type);
2517 move_to_free_list(page, zone, current_order, start_type);
2521 * Check whether there is a suitable fallback freepage with requested order.
2522 * If only_stealable is true, this function returns fallback_mt only if
2523 * we can steal other freepages all together. This would help to reduce
2524 * fragmentation due to mixed migratetype pages in one pageblock.
2526 int find_suitable_fallback(struct free_area *area, unsigned int order,
2527 int migratetype, bool only_stealable, bool *can_steal)
2532 if (area->nr_free == 0)
2537 fallback_mt = fallbacks[migratetype][i];
2538 if (fallback_mt == MIGRATE_TYPES)
2541 if (free_area_empty(area, fallback_mt))
2544 if (can_steal_fallback(order, migratetype))
2547 if (!only_stealable)
2558 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2559 * there are no empty page blocks that contain a page with a suitable order
2561 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2562 unsigned int alloc_order)
2565 unsigned long max_managed, flags;
2568 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2569 * Check is race-prone but harmless.
2571 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2572 if (zone->nr_reserved_highatomic >= max_managed)
2575 spin_lock_irqsave(&zone->lock, flags);
2577 /* Recheck the nr_reserved_highatomic limit under the lock */
2578 if (zone->nr_reserved_highatomic >= max_managed)
2582 mt = get_pageblock_migratetype(page);
2583 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2584 && !is_migrate_cma(mt)) {
2585 zone->nr_reserved_highatomic += pageblock_nr_pages;
2586 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2587 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2591 spin_unlock_irqrestore(&zone->lock, flags);
2595 * Used when an allocation is about to fail under memory pressure. This
2596 * potentially hurts the reliability of high-order allocations when under
2597 * intense memory pressure but failed atomic allocations should be easier
2598 * to recover from than an OOM.
2600 * If @force is true, try to unreserve a pageblock even though highatomic
2601 * pageblock is exhausted.
2603 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2606 struct zonelist *zonelist = ac->zonelist;
2607 unsigned long flags;
2614 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2617 * Preserve at least one pageblock unless memory pressure
2620 if (!force && zone->nr_reserved_highatomic <=
2624 spin_lock_irqsave(&zone->lock, flags);
2625 for (order = 0; order < MAX_ORDER; order++) {
2626 struct free_area *area = &(zone->free_area[order]);
2628 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2633 * In page freeing path, migratetype change is racy so
2634 * we can counter several free pages in a pageblock
2635 * in this loop althoug we changed the pageblock type
2636 * from highatomic to ac->migratetype. So we should
2637 * adjust the count once.
2639 if (is_migrate_highatomic_page(page)) {
2641 * It should never happen but changes to
2642 * locking could inadvertently allow a per-cpu
2643 * drain to add pages to MIGRATE_HIGHATOMIC
2644 * while unreserving so be safe and watch for
2647 zone->nr_reserved_highatomic -= min(
2649 zone->nr_reserved_highatomic);
2653 * Convert to ac->migratetype and avoid the normal
2654 * pageblock stealing heuristics. Minimally, the caller
2655 * is doing the work and needs the pages. More
2656 * importantly, if the block was always converted to
2657 * MIGRATE_UNMOVABLE or another type then the number
2658 * of pageblocks that cannot be completely freed
2661 set_pageblock_migratetype(page, ac->migratetype);
2662 ret = move_freepages_block(zone, page, ac->migratetype,
2665 spin_unlock_irqrestore(&zone->lock, flags);
2669 spin_unlock_irqrestore(&zone->lock, flags);
2676 * Try finding a free buddy page on the fallback list and put it on the free
2677 * list of requested migratetype, possibly along with other pages from the same
2678 * block, depending on fragmentation avoidance heuristics. Returns true if
2679 * fallback was found so that __rmqueue_smallest() can grab it.
2681 * The use of signed ints for order and current_order is a deliberate
2682 * deviation from the rest of this file, to make the for loop
2683 * condition simpler.
2685 static __always_inline bool
2686 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2687 unsigned int alloc_flags)
2689 struct free_area *area;
2691 int min_order = order;
2697 * Do not steal pages from freelists belonging to other pageblocks
2698 * i.e. orders < pageblock_order. If there are no local zones free,
2699 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2701 if (alloc_flags & ALLOC_NOFRAGMENT)
2702 min_order = pageblock_order;
2705 * Find the largest available free page in the other list. This roughly
2706 * approximates finding the pageblock with the most free pages, which
2707 * would be too costly to do exactly.
2709 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2711 area = &(zone->free_area[current_order]);
2712 fallback_mt = find_suitable_fallback(area, current_order,
2713 start_migratetype, false, &can_steal);
2714 if (fallback_mt == -1)
2718 * We cannot steal all free pages from the pageblock and the
2719 * requested migratetype is movable. In that case it's better to
2720 * steal and split the smallest available page instead of the
2721 * largest available page, because even if the next movable
2722 * allocation falls back into a different pageblock than this
2723 * one, it won't cause permanent fragmentation.
2725 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2726 && current_order > order)
2735 for (current_order = order; current_order < MAX_ORDER;
2737 area = &(zone->free_area[current_order]);
2738 fallback_mt = find_suitable_fallback(area, current_order,
2739 start_migratetype, false, &can_steal);
2740 if (fallback_mt != -1)
2745 * This should not happen - we already found a suitable fallback
2746 * when looking for the largest page.
2748 VM_BUG_ON(current_order == MAX_ORDER);
2751 page = get_page_from_free_area(area, fallback_mt);
2753 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2756 trace_mm_page_alloc_extfrag(page, order, current_order,
2757 start_migratetype, fallback_mt);
2764 * Do the hard work of removing an element from the buddy allocator.
2765 * Call me with the zone->lock already held.
2767 static __always_inline struct page *
2768 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2769 unsigned int alloc_flags)
2774 page = __rmqueue_smallest(zone, order, migratetype);
2775 if (unlikely(!page)) {
2776 if (migratetype == MIGRATE_MOVABLE)
2777 page = __rmqueue_cma_fallback(zone, order);
2779 if (!page && __rmqueue_fallback(zone, order, migratetype,
2784 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2789 * Obtain a specified number of elements from the buddy allocator, all under
2790 * a single hold of the lock, for efficiency. Add them to the supplied list.
2791 * Returns the number of new pages which were placed at *list.
2793 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2794 unsigned long count, struct list_head *list,
2795 int migratetype, unsigned int alloc_flags)
2799 spin_lock(&zone->lock);
2800 for (i = 0; i < count; ++i) {
2801 struct page *page = __rmqueue(zone, order, migratetype,
2803 if (unlikely(page == NULL))
2806 if (unlikely(check_pcp_refill(page)))
2810 * Split buddy pages returned by expand() are received here in
2811 * physical page order. The page is added to the tail of
2812 * caller's list. From the callers perspective, the linked list
2813 * is ordered by page number under some conditions. This is
2814 * useful for IO devices that can forward direction from the
2815 * head, thus also in the physical page order. This is useful
2816 * for IO devices that can merge IO requests if the physical
2817 * pages are ordered properly.
2819 list_add_tail(&page->lru, list);
2821 if (is_migrate_cma(get_pcppage_migratetype(page)))
2822 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2827 * i pages were removed from the buddy list even if some leak due
2828 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2829 * on i. Do not confuse with 'alloced' which is the number of
2830 * pages added to the pcp list.
2832 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2833 spin_unlock(&zone->lock);
2839 * Called from the vmstat counter updater to drain pagesets of this
2840 * currently executing processor on remote nodes after they have
2843 * Note that this function must be called with the thread pinned to
2844 * a single processor.
2846 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2848 unsigned long flags;
2849 int to_drain, batch;
2851 local_irq_save(flags);
2852 batch = READ_ONCE(pcp->batch);
2853 to_drain = min(pcp->count, batch);
2855 free_pcppages_bulk(zone, to_drain, pcp);
2856 local_irq_restore(flags);
2861 * Drain pcplists of the indicated processor and zone.
2863 * The processor must either be the current processor and the
2864 * thread pinned to the current processor or a processor that
2867 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2869 unsigned long flags;
2870 struct per_cpu_pageset *pset;
2871 struct per_cpu_pages *pcp;
2873 local_irq_save(flags);
2874 pset = per_cpu_ptr(zone->pageset, cpu);
2878 free_pcppages_bulk(zone, pcp->count, pcp);
2879 local_irq_restore(flags);
2883 * Drain pcplists of all zones on the indicated processor.
2885 * The processor must either be the current processor and the
2886 * thread pinned to the current processor or a processor that
2889 static void drain_pages(unsigned int cpu)
2893 for_each_populated_zone(zone) {
2894 drain_pages_zone(cpu, zone);
2899 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2901 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2902 * the single zone's pages.
2904 void drain_local_pages(struct zone *zone)
2906 int cpu = smp_processor_id();
2909 drain_pages_zone(cpu, zone);
2914 static void drain_local_pages_wq(struct work_struct *work)
2916 struct pcpu_drain *drain;
2918 drain = container_of(work, struct pcpu_drain, work);
2921 * drain_all_pages doesn't use proper cpu hotplug protection so
2922 * we can race with cpu offline when the WQ can move this from
2923 * a cpu pinned worker to an unbound one. We can operate on a different
2924 * cpu which is allright but we also have to make sure to not move to
2928 drain_local_pages(drain->zone);
2933 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2935 * When zone parameter is non-NULL, spill just the single zone's pages.
2937 * Note that this can be extremely slow as the draining happens in a workqueue.
2939 void drain_all_pages(struct zone *zone)
2944 * Allocate in the BSS so we wont require allocation in
2945 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2947 static cpumask_t cpus_with_pcps;
2950 * Make sure nobody triggers this path before mm_percpu_wq is fully
2953 if (WARN_ON_ONCE(!mm_percpu_wq))
2957 * Do not drain if one is already in progress unless it's specific to
2958 * a zone. Such callers are primarily CMA and memory hotplug and need
2959 * the drain to be complete when the call returns.
2961 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2964 mutex_lock(&pcpu_drain_mutex);
2968 * We don't care about racing with CPU hotplug event
2969 * as offline notification will cause the notified
2970 * cpu to drain that CPU pcps and on_each_cpu_mask
2971 * disables preemption as part of its processing
2973 for_each_online_cpu(cpu) {
2974 struct per_cpu_pageset *pcp;
2976 bool has_pcps = false;
2979 pcp = per_cpu_ptr(zone->pageset, cpu);
2983 for_each_populated_zone(z) {
2984 pcp = per_cpu_ptr(z->pageset, cpu);
2985 if (pcp->pcp.count) {
2993 cpumask_set_cpu(cpu, &cpus_with_pcps);
2995 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2998 for_each_cpu(cpu, &cpus_with_pcps) {
2999 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3002 INIT_WORK(&drain->work, drain_local_pages_wq);
3003 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3005 for_each_cpu(cpu, &cpus_with_pcps)
3006 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3008 mutex_unlock(&pcpu_drain_mutex);
3011 #ifdef CONFIG_HIBERNATION
3014 * Touch the watchdog for every WD_PAGE_COUNT pages.
3016 #define WD_PAGE_COUNT (128*1024)
3018 void mark_free_pages(struct zone *zone)
3020 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3021 unsigned long flags;
3022 unsigned int order, t;
3025 if (zone_is_empty(zone))
3028 spin_lock_irqsave(&zone->lock, flags);
3030 max_zone_pfn = zone_end_pfn(zone);
3031 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3032 if (pfn_valid(pfn)) {
3033 page = pfn_to_page(pfn);
3035 if (!--page_count) {
3036 touch_nmi_watchdog();
3037 page_count = WD_PAGE_COUNT;
3040 if (page_zone(page) != zone)
3043 if (!swsusp_page_is_forbidden(page))
3044 swsusp_unset_page_free(page);
3047 for_each_migratetype_order(order, t) {
3048 list_for_each_entry(page,
3049 &zone->free_area[order].free_list[t], lru) {
3052 pfn = page_to_pfn(page);
3053 for (i = 0; i < (1UL << order); i++) {
3054 if (!--page_count) {
3055 touch_nmi_watchdog();
3056 page_count = WD_PAGE_COUNT;
3058 swsusp_set_page_free(pfn_to_page(pfn + i));
3062 spin_unlock_irqrestore(&zone->lock, flags);
3064 #endif /* CONFIG_PM */
3066 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3070 if (!free_pcp_prepare(page))
3073 migratetype = get_pfnblock_migratetype(page, pfn);
3074 set_pcppage_migratetype(page, migratetype);
3078 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3080 struct zone *zone = page_zone(page);
3081 struct per_cpu_pages *pcp;
3084 migratetype = get_pcppage_migratetype(page);
3085 __count_vm_event(PGFREE);
3088 * We only track unmovable, reclaimable and movable on pcp lists.
3089 * Free ISOLATE pages back to the allocator because they are being
3090 * offlined but treat HIGHATOMIC as movable pages so we can get those
3091 * areas back if necessary. Otherwise, we may have to free
3092 * excessively into the page allocator
3094 if (migratetype >= MIGRATE_PCPTYPES) {
3095 if (unlikely(is_migrate_isolate(migratetype))) {
3096 free_one_page(zone, page, pfn, 0, migratetype);
3099 migratetype = MIGRATE_MOVABLE;
3102 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3103 list_add(&page->lru, &pcp->lists[migratetype]);
3105 if (pcp->count >= pcp->high) {
3106 unsigned long batch = READ_ONCE(pcp->batch);
3107 free_pcppages_bulk(zone, batch, pcp);
3112 * Free a 0-order page
3114 void free_unref_page(struct page *page)
3116 unsigned long flags;
3117 unsigned long pfn = page_to_pfn(page);
3119 if (!free_unref_page_prepare(page, pfn))
3122 local_irq_save(flags);
3123 free_unref_page_commit(page, pfn);
3124 local_irq_restore(flags);
3128 * Free a list of 0-order pages
3130 void free_unref_page_list(struct list_head *list)
3132 struct page *page, *next;
3133 unsigned long flags, pfn;
3134 int batch_count = 0;
3136 /* Prepare pages for freeing */
3137 list_for_each_entry_safe(page, next, list, lru) {
3138 pfn = page_to_pfn(page);
3139 if (!free_unref_page_prepare(page, pfn))
3140 list_del(&page->lru);
3141 set_page_private(page, pfn);
3144 local_irq_save(flags);
3145 list_for_each_entry_safe(page, next, list, lru) {
3146 unsigned long pfn = page_private(page);
3148 set_page_private(page, 0);
3149 trace_mm_page_free_batched(page);
3150 free_unref_page_commit(page, pfn);
3153 * Guard against excessive IRQ disabled times when we get
3154 * a large list of pages to free.
3156 if (++batch_count == SWAP_CLUSTER_MAX) {
3157 local_irq_restore(flags);
3159 local_irq_save(flags);
3162 local_irq_restore(flags);
3166 * split_page takes a non-compound higher-order page, and splits it into
3167 * n (1<<order) sub-pages: page[0..n]
3168 * Each sub-page must be freed individually.
3170 * Note: this is probably too low level an operation for use in drivers.
3171 * Please consult with lkml before using this in your driver.
3173 void split_page(struct page *page, unsigned int order)
3177 VM_BUG_ON_PAGE(PageCompound(page), page);
3178 VM_BUG_ON_PAGE(!page_count(page), page);
3180 for (i = 1; i < (1 << order); i++)
3181 set_page_refcounted(page + i);
3182 split_page_owner(page, order);
3184 EXPORT_SYMBOL_GPL(split_page);
3186 int __isolate_free_page(struct page *page, unsigned int order)
3188 unsigned long watermark;
3192 BUG_ON(!PageBuddy(page));
3194 zone = page_zone(page);
3195 mt = get_pageblock_migratetype(page);
3197 if (!is_migrate_isolate(mt)) {
3199 * Obey watermarks as if the page was being allocated. We can
3200 * emulate a high-order watermark check with a raised order-0
3201 * watermark, because we already know our high-order page
3204 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3205 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3208 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3211 /* Remove page from free list */
3213 del_page_from_free_list(page, zone, order);
3216 * Set the pageblock if the isolated page is at least half of a
3219 if (order >= pageblock_order - 1) {
3220 struct page *endpage = page + (1 << order) - 1;
3221 for (; page < endpage; page += pageblock_nr_pages) {
3222 int mt = get_pageblock_migratetype(page);
3223 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3224 && !is_migrate_highatomic(mt))
3225 set_pageblock_migratetype(page,
3231 return 1UL << order;
3235 * __putback_isolated_page - Return a now-isolated page back where we got it
3236 * @page: Page that was isolated
3237 * @order: Order of the isolated page
3238 * @mt: The page's pageblock's migratetype
3240 * This function is meant to return a page pulled from the free lists via
3241 * __isolate_free_page back to the free lists they were pulled from.
3243 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3245 struct zone *zone = page_zone(page);
3247 /* zone lock should be held when this function is called */
3248 lockdep_assert_held(&zone->lock);
3250 /* Return isolated page to tail of freelist. */
3251 __free_one_page(page, page_to_pfn(page), zone, order, mt, false);
3255 * Update NUMA hit/miss statistics
3257 * Must be called with interrupts disabled.
3259 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3262 enum numa_stat_item local_stat = NUMA_LOCAL;
3264 /* skip numa counters update if numa stats is disabled */
3265 if (!static_branch_likely(&vm_numa_stat_key))
3268 if (zone_to_nid(z) != numa_node_id())
3269 local_stat = NUMA_OTHER;
3271 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3272 __inc_numa_state(z, NUMA_HIT);
3274 __inc_numa_state(z, NUMA_MISS);
3275 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3277 __inc_numa_state(z, local_stat);
3281 /* Remove page from the per-cpu list, caller must protect the list */
3282 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3283 unsigned int alloc_flags,
3284 struct per_cpu_pages *pcp,
3285 struct list_head *list)
3290 if (list_empty(list)) {
3291 pcp->count += rmqueue_bulk(zone, 0,
3293 migratetype, alloc_flags);
3294 if (unlikely(list_empty(list)))
3298 page = list_first_entry(list, struct page, lru);
3299 list_del(&page->lru);
3301 } while (check_new_pcp(page));
3306 /* Lock and remove page from the per-cpu list */
3307 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3308 struct zone *zone, gfp_t gfp_flags,
3309 int migratetype, unsigned int alloc_flags)
3311 struct per_cpu_pages *pcp;
3312 struct list_head *list;
3314 unsigned long flags;
3316 local_irq_save(flags);
3317 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3318 list = &pcp->lists[migratetype];
3319 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3321 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3322 zone_statistics(preferred_zone, zone);
3324 local_irq_restore(flags);
3329 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3332 struct page *rmqueue(struct zone *preferred_zone,
3333 struct zone *zone, unsigned int order,
3334 gfp_t gfp_flags, unsigned int alloc_flags,
3337 unsigned long flags;
3340 if (likely(order == 0)) {
3341 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3342 migratetype, alloc_flags);
3347 * We most definitely don't want callers attempting to
3348 * allocate greater than order-1 page units with __GFP_NOFAIL.
3350 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3351 spin_lock_irqsave(&zone->lock, flags);
3355 if (alloc_flags & ALLOC_HARDER) {
3356 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3358 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3361 page = __rmqueue(zone, order, migratetype, alloc_flags);
3362 } while (page && check_new_pages(page, order));
3363 spin_unlock(&zone->lock);
3366 __mod_zone_freepage_state(zone, -(1 << order),
3367 get_pcppage_migratetype(page));
3369 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3370 zone_statistics(preferred_zone, zone);
3371 local_irq_restore(flags);
3374 /* Separate test+clear to avoid unnecessary atomics */
3375 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3376 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3377 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3380 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3384 local_irq_restore(flags);
3388 #ifdef CONFIG_FAIL_PAGE_ALLOC
3391 struct fault_attr attr;
3393 bool ignore_gfp_highmem;
3394 bool ignore_gfp_reclaim;
3396 } fail_page_alloc = {
3397 .attr = FAULT_ATTR_INITIALIZER,
3398 .ignore_gfp_reclaim = true,
3399 .ignore_gfp_highmem = true,
3403 static int __init setup_fail_page_alloc(char *str)
3405 return setup_fault_attr(&fail_page_alloc.attr, str);
3407 __setup("fail_page_alloc=", setup_fail_page_alloc);
3409 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3411 if (order < fail_page_alloc.min_order)
3413 if (gfp_mask & __GFP_NOFAIL)
3415 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3417 if (fail_page_alloc.ignore_gfp_reclaim &&
3418 (gfp_mask & __GFP_DIRECT_RECLAIM))
3421 return should_fail(&fail_page_alloc.attr, 1 << order);
3424 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3426 static int __init fail_page_alloc_debugfs(void)
3428 umode_t mode = S_IFREG | 0600;
3431 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3432 &fail_page_alloc.attr);
3434 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3435 &fail_page_alloc.ignore_gfp_reclaim);
3436 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3437 &fail_page_alloc.ignore_gfp_highmem);
3438 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3443 late_initcall(fail_page_alloc_debugfs);
3445 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3447 #else /* CONFIG_FAIL_PAGE_ALLOC */
3449 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3454 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3456 static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3458 return __should_fail_alloc_page(gfp_mask, order);
3460 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3463 * Return true if free base pages are above 'mark'. For high-order checks it
3464 * will return true of the order-0 watermark is reached and there is at least
3465 * one free page of a suitable size. Checking now avoids taking the zone lock
3466 * to check in the allocation paths if no pages are free.
3468 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3469 int classzone_idx, unsigned int alloc_flags,
3474 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3476 /* free_pages may go negative - that's OK */
3477 free_pages -= (1 << order) - 1;
3479 if (alloc_flags & ALLOC_HIGH)
3483 * If the caller does not have rights to ALLOC_HARDER then subtract
3484 * the high-atomic reserves. This will over-estimate the size of the
3485 * atomic reserve but it avoids a search.
3487 if (likely(!alloc_harder)) {
3488 free_pages -= z->nr_reserved_highatomic;
3491 * OOM victims can try even harder than normal ALLOC_HARDER
3492 * users on the grounds that it's definitely going to be in
3493 * the exit path shortly and free memory. Any allocation it
3494 * makes during the free path will be small and short-lived.
3496 if (alloc_flags & ALLOC_OOM)
3504 /* If allocation can't use CMA areas don't use free CMA pages */
3505 if (!(alloc_flags & ALLOC_CMA))
3506 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3510 * Check watermarks for an order-0 allocation request. If these
3511 * are not met, then a high-order request also cannot go ahead
3512 * even if a suitable page happened to be free.
3514 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3517 /* If this is an order-0 request then the watermark is fine */
3521 /* For a high-order request, check at least one suitable page is free */
3522 for (o = order; o < MAX_ORDER; o++) {
3523 struct free_area *area = &z->free_area[o];
3529 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3530 if (!free_area_empty(area, mt))
3535 if ((alloc_flags & ALLOC_CMA) &&
3536 !free_area_empty(area, MIGRATE_CMA)) {
3540 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3546 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3547 int classzone_idx, unsigned int alloc_flags)
3549 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3550 zone_page_state(z, NR_FREE_PAGES));
3553 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3554 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3556 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3560 /* If allocation can't use CMA areas don't use free CMA pages */
3561 if (!(alloc_flags & ALLOC_CMA))
3562 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3566 * Fast check for order-0 only. If this fails then the reserves
3567 * need to be calculated. There is a corner case where the check
3568 * passes but only the high-order atomic reserve are free. If
3569 * the caller is !atomic then it'll uselessly search the free
3570 * list. That corner case is then slower but it is harmless.
3572 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3575 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3579 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3580 unsigned long mark, int classzone_idx)
3582 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3584 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3585 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3587 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3592 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3594 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3595 node_reclaim_distance;
3597 #else /* CONFIG_NUMA */
3598 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3602 #endif /* CONFIG_NUMA */
3605 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3606 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3607 * premature use of a lower zone may cause lowmem pressure problems that
3608 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3609 * probably too small. It only makes sense to spread allocations to avoid
3610 * fragmentation between the Normal and DMA32 zones.
3612 static inline unsigned int
3613 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3615 unsigned int alloc_flags;
3618 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3621 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3623 #ifdef CONFIG_ZONE_DMA32
3627 if (zone_idx(zone) != ZONE_NORMAL)
3631 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3632 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3633 * on UMA that if Normal is populated then so is DMA32.
3635 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3636 if (nr_online_nodes > 1 && !populated_zone(--zone))
3639 alloc_flags |= ALLOC_NOFRAGMENT;
3640 #endif /* CONFIG_ZONE_DMA32 */
3645 * get_page_from_freelist goes through the zonelist trying to allocate
3648 static struct page *
3649 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3650 const struct alloc_context *ac)
3654 struct pglist_data *last_pgdat_dirty_limit = NULL;
3659 * Scan zonelist, looking for a zone with enough free.
3660 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3662 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3663 z = ac->preferred_zoneref;
3664 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3669 if (cpusets_enabled() &&
3670 (alloc_flags & ALLOC_CPUSET) &&
3671 !__cpuset_zone_allowed(zone, gfp_mask))
3674 * When allocating a page cache page for writing, we
3675 * want to get it from a node that is within its dirty
3676 * limit, such that no single node holds more than its
3677 * proportional share of globally allowed dirty pages.
3678 * The dirty limits take into account the node's
3679 * lowmem reserves and high watermark so that kswapd
3680 * should be able to balance it without having to
3681 * write pages from its LRU list.
3683 * XXX: For now, allow allocations to potentially
3684 * exceed the per-node dirty limit in the slowpath
3685 * (spread_dirty_pages unset) before going into reclaim,
3686 * which is important when on a NUMA setup the allowed
3687 * nodes are together not big enough to reach the
3688 * global limit. The proper fix for these situations
3689 * will require awareness of nodes in the
3690 * dirty-throttling and the flusher threads.
3692 if (ac->spread_dirty_pages) {
3693 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3696 if (!node_dirty_ok(zone->zone_pgdat)) {
3697 last_pgdat_dirty_limit = zone->zone_pgdat;
3702 if (no_fallback && nr_online_nodes > 1 &&
3703 zone != ac->preferred_zoneref->zone) {
3707 * If moving to a remote node, retry but allow
3708 * fragmenting fallbacks. Locality is more important
3709 * than fragmentation avoidance.
3711 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3712 if (zone_to_nid(zone) != local_nid) {
3713 alloc_flags &= ~ALLOC_NOFRAGMENT;
3718 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3719 if (!zone_watermark_fast(zone, order, mark,
3720 ac_classzone_idx(ac), alloc_flags)) {
3723 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3725 * Watermark failed for this zone, but see if we can
3726 * grow this zone if it contains deferred pages.
3728 if (static_branch_unlikely(&deferred_pages)) {
3729 if (_deferred_grow_zone(zone, order))
3733 /* Checked here to keep the fast path fast */
3734 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3735 if (alloc_flags & ALLOC_NO_WATERMARKS)
3738 if (node_reclaim_mode == 0 ||
3739 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3742 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3744 case NODE_RECLAIM_NOSCAN:
3747 case NODE_RECLAIM_FULL:
3748 /* scanned but unreclaimable */
3751 /* did we reclaim enough */
3752 if (zone_watermark_ok(zone, order, mark,
3753 ac_classzone_idx(ac), alloc_flags))
3761 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3762 gfp_mask, alloc_flags, ac->migratetype);
3764 prep_new_page(page, order, gfp_mask, alloc_flags);
3767 * If this is a high-order atomic allocation then check
3768 * if the pageblock should be reserved for the future
3770 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3771 reserve_highatomic_pageblock(page, zone, order);
3775 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3776 /* Try again if zone has deferred pages */
3777 if (static_branch_unlikely(&deferred_pages)) {
3778 if (_deferred_grow_zone(zone, order))
3786 * It's possible on a UMA machine to get through all zones that are
3787 * fragmented. If avoiding fragmentation, reset and try again.
3790 alloc_flags &= ~ALLOC_NOFRAGMENT;
3797 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3799 unsigned int filter = SHOW_MEM_FILTER_NODES;
3802 * This documents exceptions given to allocations in certain
3803 * contexts that are allowed to allocate outside current's set
3806 if (!(gfp_mask & __GFP_NOMEMALLOC))
3807 if (tsk_is_oom_victim(current) ||
3808 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3809 filter &= ~SHOW_MEM_FILTER_NODES;
3810 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3811 filter &= ~SHOW_MEM_FILTER_NODES;
3813 show_mem(filter, nodemask);
3816 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3818 struct va_format vaf;
3820 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3822 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3825 va_start(args, fmt);
3828 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3829 current->comm, &vaf, gfp_mask, &gfp_mask,
3830 nodemask_pr_args(nodemask));
3833 cpuset_print_current_mems_allowed();
3836 warn_alloc_show_mem(gfp_mask, nodemask);
3839 static inline struct page *
3840 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3841 unsigned int alloc_flags,
3842 const struct alloc_context *ac)
3846 page = get_page_from_freelist(gfp_mask, order,
3847 alloc_flags|ALLOC_CPUSET, ac);
3849 * fallback to ignore cpuset restriction if our nodes
3853 page = get_page_from_freelist(gfp_mask, order,
3859 static inline struct page *
3860 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3861 const struct alloc_context *ac, unsigned long *did_some_progress)
3863 struct oom_control oc = {
3864 .zonelist = ac->zonelist,
3865 .nodemask = ac->nodemask,
3867 .gfp_mask = gfp_mask,
3872 *did_some_progress = 0;
3875 * Acquire the oom lock. If that fails, somebody else is
3876 * making progress for us.
3878 if (!mutex_trylock(&oom_lock)) {
3879 *did_some_progress = 1;
3880 schedule_timeout_uninterruptible(1);
3885 * Go through the zonelist yet one more time, keep very high watermark
3886 * here, this is only to catch a parallel oom killing, we must fail if
3887 * we're still under heavy pressure. But make sure that this reclaim
3888 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3889 * allocation which will never fail due to oom_lock already held.
3891 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3892 ~__GFP_DIRECT_RECLAIM, order,
3893 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3897 /* Coredumps can quickly deplete all memory reserves */
3898 if (current->flags & PF_DUMPCORE)
3900 /* The OOM killer will not help higher order allocs */
3901 if (order > PAGE_ALLOC_COSTLY_ORDER)
3904 * We have already exhausted all our reclaim opportunities without any
3905 * success so it is time to admit defeat. We will skip the OOM killer
3906 * because it is very likely that the caller has a more reasonable
3907 * fallback than shooting a random task.
3909 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3911 /* The OOM killer does not needlessly kill tasks for lowmem */
3912 if (ac->high_zoneidx < ZONE_NORMAL)
3914 if (pm_suspended_storage())
3917 * XXX: GFP_NOFS allocations should rather fail than rely on
3918 * other request to make a forward progress.
3919 * We are in an unfortunate situation where out_of_memory cannot
3920 * do much for this context but let's try it to at least get
3921 * access to memory reserved if the current task is killed (see
3922 * out_of_memory). Once filesystems are ready to handle allocation
3923 * failures more gracefully we should just bail out here.
3926 /* The OOM killer may not free memory on a specific node */
3927 if (gfp_mask & __GFP_THISNODE)
3930 /* Exhausted what can be done so it's blame time */
3931 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3932 *did_some_progress = 1;
3935 * Help non-failing allocations by giving them access to memory
3938 if (gfp_mask & __GFP_NOFAIL)
3939 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3940 ALLOC_NO_WATERMARKS, ac);
3943 mutex_unlock(&oom_lock);
3948 * Maximum number of compaction retries wit a progress before OOM
3949 * killer is consider as the only way to move forward.
3951 #define MAX_COMPACT_RETRIES 16
3953 #ifdef CONFIG_COMPACTION
3954 /* Try memory compaction for high-order allocations before reclaim */
3955 static struct page *
3956 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3957 unsigned int alloc_flags, const struct alloc_context *ac,
3958 enum compact_priority prio, enum compact_result *compact_result)
3960 struct page *page = NULL;
3961 unsigned long pflags;
3962 unsigned int noreclaim_flag;
3967 psi_memstall_enter(&pflags);
3968 noreclaim_flag = memalloc_noreclaim_save();
3970 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3973 memalloc_noreclaim_restore(noreclaim_flag);
3974 psi_memstall_leave(&pflags);
3977 * At least in one zone compaction wasn't deferred or skipped, so let's
3978 * count a compaction stall
3980 count_vm_event(COMPACTSTALL);
3982 /* Prep a captured page if available */
3984 prep_new_page(page, order, gfp_mask, alloc_flags);
3986 /* Try get a page from the freelist if available */
3988 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3991 struct zone *zone = page_zone(page);
3993 zone->compact_blockskip_flush = false;
3994 compaction_defer_reset(zone, order, true);
3995 count_vm_event(COMPACTSUCCESS);
4000 * It's bad if compaction run occurs and fails. The most likely reason
4001 * is that pages exist, but not enough to satisfy watermarks.
4003 count_vm_event(COMPACTFAIL);
4011 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4012 enum compact_result compact_result,
4013 enum compact_priority *compact_priority,
4014 int *compaction_retries)
4016 int max_retries = MAX_COMPACT_RETRIES;
4019 int retries = *compaction_retries;
4020 enum compact_priority priority = *compact_priority;
4025 if (compaction_made_progress(compact_result))
4026 (*compaction_retries)++;
4029 * compaction considers all the zone as desperately out of memory
4030 * so it doesn't really make much sense to retry except when the
4031 * failure could be caused by insufficient priority
4033 if (compaction_failed(compact_result))
4034 goto check_priority;
4037 * compaction was skipped because there are not enough order-0 pages
4038 * to work with, so we retry only if it looks like reclaim can help.
4040 if (compaction_needs_reclaim(compact_result)) {
4041 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4046 * make sure the compaction wasn't deferred or didn't bail out early
4047 * due to locks contention before we declare that we should give up.
4048 * But the next retry should use a higher priority if allowed, so
4049 * we don't just keep bailing out endlessly.
4051 if (compaction_withdrawn(compact_result)) {
4052 goto check_priority;
4056 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4057 * costly ones because they are de facto nofail and invoke OOM
4058 * killer to move on while costly can fail and users are ready
4059 * to cope with that. 1/4 retries is rather arbitrary but we
4060 * would need much more detailed feedback from compaction to
4061 * make a better decision.
4063 if (order > PAGE_ALLOC_COSTLY_ORDER)
4065 if (*compaction_retries <= max_retries) {
4071 * Make sure there are attempts at the highest priority if we exhausted
4072 * all retries or failed at the lower priorities.
4075 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4076 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4078 if (*compact_priority > min_priority) {
4079 (*compact_priority)--;
4080 *compaction_retries = 0;
4084 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4088 static inline struct page *
4089 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4090 unsigned int alloc_flags, const struct alloc_context *ac,
4091 enum compact_priority prio, enum compact_result *compact_result)
4093 *compact_result = COMPACT_SKIPPED;
4098 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4099 enum compact_result compact_result,
4100 enum compact_priority *compact_priority,
4101 int *compaction_retries)
4106 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4110 * There are setups with compaction disabled which would prefer to loop
4111 * inside the allocator rather than hit the oom killer prematurely.
4112 * Let's give them a good hope and keep retrying while the order-0
4113 * watermarks are OK.
4115 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4117 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4118 ac_classzone_idx(ac), alloc_flags))
4123 #endif /* CONFIG_COMPACTION */
4125 #ifdef CONFIG_LOCKDEP
4126 static struct lockdep_map __fs_reclaim_map =
4127 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4129 static bool __need_fs_reclaim(gfp_t gfp_mask)
4131 gfp_mask = current_gfp_context(gfp_mask);
4133 /* no reclaim without waiting on it */
4134 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4137 /* this guy won't enter reclaim */
4138 if (current->flags & PF_MEMALLOC)
4141 /* We're only interested __GFP_FS allocations for now */
4142 if (!(gfp_mask & __GFP_FS))
4145 if (gfp_mask & __GFP_NOLOCKDEP)
4151 void __fs_reclaim_acquire(void)
4153 lock_map_acquire(&__fs_reclaim_map);
4156 void __fs_reclaim_release(void)
4158 lock_map_release(&__fs_reclaim_map);
4161 void fs_reclaim_acquire(gfp_t gfp_mask)
4163 if (__need_fs_reclaim(gfp_mask))
4164 __fs_reclaim_acquire();
4166 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4168 void fs_reclaim_release(gfp_t gfp_mask)
4170 if (__need_fs_reclaim(gfp_mask))
4171 __fs_reclaim_release();
4173 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4176 /* Perform direct synchronous page reclaim */
4178 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4179 const struct alloc_context *ac)
4182 unsigned int noreclaim_flag;
4183 unsigned long pflags;
4187 /* We now go into synchronous reclaim */
4188 cpuset_memory_pressure_bump();
4189 psi_memstall_enter(&pflags);
4190 fs_reclaim_acquire(gfp_mask);
4191 noreclaim_flag = memalloc_noreclaim_save();
4193 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4196 memalloc_noreclaim_restore(noreclaim_flag);
4197 fs_reclaim_release(gfp_mask);
4198 psi_memstall_leave(&pflags);
4205 /* The really slow allocator path where we enter direct reclaim */
4206 static inline struct page *
4207 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4208 unsigned int alloc_flags, const struct alloc_context *ac,
4209 unsigned long *did_some_progress)
4211 struct page *page = NULL;
4212 bool drained = false;
4214 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4215 if (unlikely(!(*did_some_progress)))
4219 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4222 * If an allocation failed after direct reclaim, it could be because
4223 * pages are pinned on the per-cpu lists or in high alloc reserves.
4224 * Shrink them them and try again
4226 if (!page && !drained) {
4227 unreserve_highatomic_pageblock(ac, false);
4228 drain_all_pages(NULL);
4236 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4237 const struct alloc_context *ac)
4241 pg_data_t *last_pgdat = NULL;
4242 enum zone_type high_zoneidx = ac->high_zoneidx;
4244 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
4246 if (last_pgdat != zone->zone_pgdat)
4247 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
4248 last_pgdat = zone->zone_pgdat;
4252 static inline unsigned int
4253 gfp_to_alloc_flags(gfp_t gfp_mask)
4255 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4258 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4259 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4260 * to save two branches.
4262 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4263 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4266 * The caller may dip into page reserves a bit more if the caller
4267 * cannot run direct reclaim, or if the caller has realtime scheduling
4268 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4269 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4271 alloc_flags |= (__force int)
4272 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4274 if (gfp_mask & __GFP_ATOMIC) {
4276 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4277 * if it can't schedule.
4279 if (!(gfp_mask & __GFP_NOMEMALLOC))
4280 alloc_flags |= ALLOC_HARDER;
4282 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4283 * comment for __cpuset_node_allowed().
4285 alloc_flags &= ~ALLOC_CPUSET;
4286 } else if (unlikely(rt_task(current)) && !in_interrupt())
4287 alloc_flags |= ALLOC_HARDER;
4290 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4291 alloc_flags |= ALLOC_CMA;
4296 static bool oom_reserves_allowed(struct task_struct *tsk)
4298 if (!tsk_is_oom_victim(tsk))
4302 * !MMU doesn't have oom reaper so give access to memory reserves
4303 * only to the thread with TIF_MEMDIE set
4305 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4312 * Distinguish requests which really need access to full memory
4313 * reserves from oom victims which can live with a portion of it
4315 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4317 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4319 if (gfp_mask & __GFP_MEMALLOC)
4320 return ALLOC_NO_WATERMARKS;
4321 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4322 return ALLOC_NO_WATERMARKS;
4323 if (!in_interrupt()) {
4324 if (current->flags & PF_MEMALLOC)
4325 return ALLOC_NO_WATERMARKS;
4326 else if (oom_reserves_allowed(current))
4333 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4335 return !!__gfp_pfmemalloc_flags(gfp_mask);
4339 * Checks whether it makes sense to retry the reclaim to make a forward progress
4340 * for the given allocation request.
4342 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4343 * without success, or when we couldn't even meet the watermark if we
4344 * reclaimed all remaining pages on the LRU lists.
4346 * Returns true if a retry is viable or false to enter the oom path.
4349 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4350 struct alloc_context *ac, int alloc_flags,
4351 bool did_some_progress, int *no_progress_loops)
4358 * Costly allocations might have made a progress but this doesn't mean
4359 * their order will become available due to high fragmentation so
4360 * always increment the no progress counter for them
4362 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4363 *no_progress_loops = 0;
4365 (*no_progress_loops)++;
4368 * Make sure we converge to OOM if we cannot make any progress
4369 * several times in the row.
4371 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4372 /* Before OOM, exhaust highatomic_reserve */
4373 return unreserve_highatomic_pageblock(ac, true);
4377 * Keep reclaiming pages while there is a chance this will lead
4378 * somewhere. If none of the target zones can satisfy our allocation
4379 * request even if all reclaimable pages are considered then we are
4380 * screwed and have to go OOM.
4382 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4384 unsigned long available;
4385 unsigned long reclaimable;
4386 unsigned long min_wmark = min_wmark_pages(zone);
4389 available = reclaimable = zone_reclaimable_pages(zone);
4390 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4393 * Would the allocation succeed if we reclaimed all
4394 * reclaimable pages?
4396 wmark = __zone_watermark_ok(zone, order, min_wmark,
4397 ac_classzone_idx(ac), alloc_flags, available);
4398 trace_reclaim_retry_zone(z, order, reclaimable,
4399 available, min_wmark, *no_progress_loops, wmark);
4402 * If we didn't make any progress and have a lot of
4403 * dirty + writeback pages then we should wait for
4404 * an IO to complete to slow down the reclaim and
4405 * prevent from pre mature OOM
4407 if (!did_some_progress) {
4408 unsigned long write_pending;
4410 write_pending = zone_page_state_snapshot(zone,
4411 NR_ZONE_WRITE_PENDING);
4413 if (2 * write_pending > reclaimable) {
4414 congestion_wait(BLK_RW_ASYNC, HZ/10);
4426 * Memory allocation/reclaim might be called from a WQ context and the
4427 * current implementation of the WQ concurrency control doesn't
4428 * recognize that a particular WQ is congested if the worker thread is
4429 * looping without ever sleeping. Therefore we have to do a short sleep
4430 * here rather than calling cond_resched().
4432 if (current->flags & PF_WQ_WORKER)
4433 schedule_timeout_uninterruptible(1);
4440 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4443 * It's possible that cpuset's mems_allowed and the nodemask from
4444 * mempolicy don't intersect. This should be normally dealt with by
4445 * policy_nodemask(), but it's possible to race with cpuset update in
4446 * such a way the check therein was true, and then it became false
4447 * before we got our cpuset_mems_cookie here.
4448 * This assumes that for all allocations, ac->nodemask can come only
4449 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4450 * when it does not intersect with the cpuset restrictions) or the
4451 * caller can deal with a violated nodemask.
4453 if (cpusets_enabled() && ac->nodemask &&
4454 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4455 ac->nodemask = NULL;
4460 * When updating a task's mems_allowed or mempolicy nodemask, it is
4461 * possible to race with parallel threads in such a way that our
4462 * allocation can fail while the mask is being updated. If we are about
4463 * to fail, check if the cpuset changed during allocation and if so,
4466 if (read_mems_allowed_retry(cpuset_mems_cookie))
4472 static inline struct page *
4473 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4474 struct alloc_context *ac)
4476 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4477 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4478 struct page *page = NULL;
4479 unsigned int alloc_flags;
4480 unsigned long did_some_progress;
4481 enum compact_priority compact_priority;
4482 enum compact_result compact_result;
4483 int compaction_retries;
4484 int no_progress_loops;
4485 unsigned int cpuset_mems_cookie;
4489 * We also sanity check to catch abuse of atomic reserves being used by
4490 * callers that are not in atomic context.
4492 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4493 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4494 gfp_mask &= ~__GFP_ATOMIC;
4497 compaction_retries = 0;
4498 no_progress_loops = 0;
4499 compact_priority = DEF_COMPACT_PRIORITY;
4500 cpuset_mems_cookie = read_mems_allowed_begin();
4503 * The fast path uses conservative alloc_flags to succeed only until
4504 * kswapd needs to be woken up, and to avoid the cost of setting up
4505 * alloc_flags precisely. So we do that now.
4507 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4510 * We need to recalculate the starting point for the zonelist iterator
4511 * because we might have used different nodemask in the fast path, or
4512 * there was a cpuset modification and we are retrying - otherwise we
4513 * could end up iterating over non-eligible zones endlessly.
4515 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4516 ac->high_zoneidx, ac->nodemask);
4517 if (!ac->preferred_zoneref->zone)
4520 if (alloc_flags & ALLOC_KSWAPD)
4521 wake_all_kswapds(order, gfp_mask, ac);
4524 * The adjusted alloc_flags might result in immediate success, so try
4527 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4532 * For costly allocations, try direct compaction first, as it's likely
4533 * that we have enough base pages and don't need to reclaim. For non-
4534 * movable high-order allocations, do that as well, as compaction will
4535 * try prevent permanent fragmentation by migrating from blocks of the
4537 * Don't try this for allocations that are allowed to ignore
4538 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4540 if (can_direct_reclaim &&
4542 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4543 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4544 page = __alloc_pages_direct_compact(gfp_mask, order,
4546 INIT_COMPACT_PRIORITY,
4552 * Checks for costly allocations with __GFP_NORETRY, which
4553 * includes some THP page fault allocations
4555 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4557 * If allocating entire pageblock(s) and compaction
4558 * failed because all zones are below low watermarks
4559 * or is prohibited because it recently failed at this
4560 * order, fail immediately unless the allocator has
4561 * requested compaction and reclaim retry.
4564 * - potentially very expensive because zones are far
4565 * below their low watermarks or this is part of very
4566 * bursty high order allocations,
4567 * - not guaranteed to help because isolate_freepages()
4568 * may not iterate over freed pages as part of its
4570 * - unlikely to make entire pageblocks free on its
4573 if (compact_result == COMPACT_SKIPPED ||
4574 compact_result == COMPACT_DEFERRED)
4578 * Looks like reclaim/compaction is worth trying, but
4579 * sync compaction could be very expensive, so keep
4580 * using async compaction.
4582 compact_priority = INIT_COMPACT_PRIORITY;
4587 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4588 if (alloc_flags & ALLOC_KSWAPD)
4589 wake_all_kswapds(order, gfp_mask, ac);
4591 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4593 alloc_flags = reserve_flags;
4596 * Reset the nodemask and zonelist iterators if memory policies can be
4597 * ignored. These allocations are high priority and system rather than
4600 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4601 ac->nodemask = NULL;
4602 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4603 ac->high_zoneidx, ac->nodemask);
4606 /* Attempt with potentially adjusted zonelist and alloc_flags */
4607 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4611 /* Caller is not willing to reclaim, we can't balance anything */
4612 if (!can_direct_reclaim)
4615 /* Avoid recursion of direct reclaim */
4616 if (current->flags & PF_MEMALLOC)
4619 /* Try direct reclaim and then allocating */
4620 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4621 &did_some_progress);
4625 /* Try direct compaction and then allocating */
4626 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4627 compact_priority, &compact_result);
4631 /* Do not loop if specifically requested */
4632 if (gfp_mask & __GFP_NORETRY)
4636 * Do not retry costly high order allocations unless they are
4637 * __GFP_RETRY_MAYFAIL
4639 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4642 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4643 did_some_progress > 0, &no_progress_loops))
4647 * It doesn't make any sense to retry for the compaction if the order-0
4648 * reclaim is not able to make any progress because the current
4649 * implementation of the compaction depends on the sufficient amount
4650 * of free memory (see __compaction_suitable)
4652 if (did_some_progress > 0 &&
4653 should_compact_retry(ac, order, alloc_flags,
4654 compact_result, &compact_priority,
4655 &compaction_retries))
4659 /* Deal with possible cpuset update races before we start OOM killing */
4660 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4663 /* Reclaim has failed us, start killing things */
4664 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4668 /* Avoid allocations with no watermarks from looping endlessly */
4669 if (tsk_is_oom_victim(current) &&
4670 (alloc_flags == ALLOC_OOM ||
4671 (gfp_mask & __GFP_NOMEMALLOC)))
4674 /* Retry as long as the OOM killer is making progress */
4675 if (did_some_progress) {
4676 no_progress_loops = 0;
4681 /* Deal with possible cpuset update races before we fail */
4682 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4686 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4689 if (gfp_mask & __GFP_NOFAIL) {
4691 * All existing users of the __GFP_NOFAIL are blockable, so warn
4692 * of any new users that actually require GFP_NOWAIT
4694 if (WARN_ON_ONCE(!can_direct_reclaim))
4698 * PF_MEMALLOC request from this context is rather bizarre
4699 * because we cannot reclaim anything and only can loop waiting
4700 * for somebody to do a work for us
4702 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4705 * non failing costly orders are a hard requirement which we
4706 * are not prepared for much so let's warn about these users
4707 * so that we can identify them and convert them to something
4710 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4713 * Help non-failing allocations by giving them access to memory
4714 * reserves but do not use ALLOC_NO_WATERMARKS because this
4715 * could deplete whole memory reserves which would just make
4716 * the situation worse
4718 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4726 warn_alloc(gfp_mask, ac->nodemask,
4727 "page allocation failure: order:%u", order);
4732 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4733 int preferred_nid, nodemask_t *nodemask,
4734 struct alloc_context *ac, gfp_t *alloc_mask,
4735 unsigned int *alloc_flags)
4737 ac->high_zoneidx = gfp_zone(gfp_mask);
4738 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4739 ac->nodemask = nodemask;
4740 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4742 if (cpusets_enabled()) {
4743 *alloc_mask |= __GFP_HARDWALL;
4745 ac->nodemask = &cpuset_current_mems_allowed;
4747 *alloc_flags |= ALLOC_CPUSET;
4750 fs_reclaim_acquire(gfp_mask);
4751 fs_reclaim_release(gfp_mask);
4753 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4755 if (should_fail_alloc_page(gfp_mask, order))
4758 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4759 *alloc_flags |= ALLOC_CMA;
4764 /* Determine whether to spread dirty pages and what the first usable zone */
4765 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4767 /* Dirty zone balancing only done in the fast path */
4768 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4771 * The preferred zone is used for statistics but crucially it is
4772 * also used as the starting point for the zonelist iterator. It
4773 * may get reset for allocations that ignore memory policies.
4775 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4776 ac->high_zoneidx, ac->nodemask);
4780 * This is the 'heart' of the zoned buddy allocator.
4783 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4784 nodemask_t *nodemask)
4787 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4788 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4789 struct alloc_context ac = { };
4792 * There are several places where we assume that the order value is sane
4793 * so bail out early if the request is out of bound.
4795 if (unlikely(order >= MAX_ORDER)) {
4796 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4800 gfp_mask &= gfp_allowed_mask;
4801 alloc_mask = gfp_mask;
4802 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4805 finalise_ac(gfp_mask, &ac);
4808 * Forbid the first pass from falling back to types that fragment
4809 * memory until all local zones are considered.
4811 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4813 /* First allocation attempt */
4814 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4819 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4820 * resp. GFP_NOIO which has to be inherited for all allocation requests
4821 * from a particular context which has been marked by
4822 * memalloc_no{fs,io}_{save,restore}.
4824 alloc_mask = current_gfp_context(gfp_mask);
4825 ac.spread_dirty_pages = false;
4828 * Restore the original nodemask if it was potentially replaced with
4829 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4831 ac.nodemask = nodemask;
4833 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4836 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4837 unlikely(__memcg_kmem_charge_page(page, gfp_mask, order) != 0)) {
4838 __free_pages(page, order);
4842 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4846 EXPORT_SYMBOL(__alloc_pages_nodemask);
4849 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4850 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4851 * you need to access high mem.
4853 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4857 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4860 return (unsigned long) page_address(page);
4862 EXPORT_SYMBOL(__get_free_pages);
4864 unsigned long get_zeroed_page(gfp_t gfp_mask)
4866 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4868 EXPORT_SYMBOL(get_zeroed_page);
4870 static inline void free_the_page(struct page *page, unsigned int order)
4872 if (order == 0) /* Via pcp? */
4873 free_unref_page(page);
4875 __free_pages_ok(page, order);
4878 void __free_pages(struct page *page, unsigned int order)
4880 if (put_page_testzero(page))
4881 free_the_page(page, order);
4883 EXPORT_SYMBOL(__free_pages);
4885 void free_pages(unsigned long addr, unsigned int order)
4888 VM_BUG_ON(!virt_addr_valid((void *)addr));
4889 __free_pages(virt_to_page((void *)addr), order);
4893 EXPORT_SYMBOL(free_pages);
4897 * An arbitrary-length arbitrary-offset area of memory which resides
4898 * within a 0 or higher order page. Multiple fragments within that page
4899 * are individually refcounted, in the page's reference counter.
4901 * The page_frag functions below provide a simple allocation framework for
4902 * page fragments. This is used by the network stack and network device
4903 * drivers to provide a backing region of memory for use as either an
4904 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4906 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4909 struct page *page = NULL;
4910 gfp_t gfp = gfp_mask;
4912 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4913 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4915 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4916 PAGE_FRAG_CACHE_MAX_ORDER);
4917 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4919 if (unlikely(!page))
4920 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4922 nc->va = page ? page_address(page) : NULL;
4927 void __page_frag_cache_drain(struct page *page, unsigned int count)
4929 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4931 if (page_ref_sub_and_test(page, count))
4932 free_the_page(page, compound_order(page));
4934 EXPORT_SYMBOL(__page_frag_cache_drain);
4936 void *page_frag_alloc(struct page_frag_cache *nc,
4937 unsigned int fragsz, gfp_t gfp_mask)
4939 unsigned int size = PAGE_SIZE;
4943 if (unlikely(!nc->va)) {
4945 page = __page_frag_cache_refill(nc, gfp_mask);
4949 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4950 /* if size can vary use size else just use PAGE_SIZE */
4953 /* Even if we own the page, we do not use atomic_set().
4954 * This would break get_page_unless_zero() users.
4956 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4958 /* reset page count bias and offset to start of new frag */
4959 nc->pfmemalloc = page_is_pfmemalloc(page);
4960 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4964 offset = nc->offset - fragsz;
4965 if (unlikely(offset < 0)) {
4966 page = virt_to_page(nc->va);
4968 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4971 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4972 /* if size can vary use size else just use PAGE_SIZE */
4975 /* OK, page count is 0, we can safely set it */
4976 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4978 /* reset page count bias and offset to start of new frag */
4979 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4980 offset = size - fragsz;
4984 nc->offset = offset;
4986 return nc->va + offset;
4988 EXPORT_SYMBOL(page_frag_alloc);
4991 * Frees a page fragment allocated out of either a compound or order 0 page.
4993 void page_frag_free(void *addr)
4995 struct page *page = virt_to_head_page(addr);
4997 if (unlikely(put_page_testzero(page)))
4998 free_the_page(page, compound_order(page));
5000 EXPORT_SYMBOL(page_frag_free);
5002 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5006 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5007 unsigned long used = addr + PAGE_ALIGN(size);
5009 split_page(virt_to_page((void *)addr), order);
5010 while (used < alloc_end) {
5015 return (void *)addr;
5019 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5020 * @size: the number of bytes to allocate
5021 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5023 * This function is similar to alloc_pages(), except that it allocates the
5024 * minimum number of pages to satisfy the request. alloc_pages() can only
5025 * allocate memory in power-of-two pages.
5027 * This function is also limited by MAX_ORDER.
5029 * Memory allocated by this function must be released by free_pages_exact().
5031 * Return: pointer to the allocated area or %NULL in case of error.
5033 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5035 unsigned int order = get_order(size);
5038 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5039 gfp_mask &= ~__GFP_COMP;
5041 addr = __get_free_pages(gfp_mask, order);
5042 return make_alloc_exact(addr, order, size);
5044 EXPORT_SYMBOL(alloc_pages_exact);
5047 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5049 * @nid: the preferred node ID where memory should be allocated
5050 * @size: the number of bytes to allocate
5051 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5053 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5056 * Return: pointer to the allocated area or %NULL in case of error.
5058 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5060 unsigned int order = get_order(size);
5063 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5064 gfp_mask &= ~__GFP_COMP;
5066 p = alloc_pages_node(nid, gfp_mask, order);
5069 return make_alloc_exact((unsigned long)page_address(p), order, size);
5073 * free_pages_exact - release memory allocated via alloc_pages_exact()
5074 * @virt: the value returned by alloc_pages_exact.
5075 * @size: size of allocation, same value as passed to alloc_pages_exact().
5077 * Release the memory allocated by a previous call to alloc_pages_exact.
5079 void free_pages_exact(void *virt, size_t size)
5081 unsigned long addr = (unsigned long)virt;
5082 unsigned long end = addr + PAGE_ALIGN(size);
5084 while (addr < end) {
5089 EXPORT_SYMBOL(free_pages_exact);
5092 * nr_free_zone_pages - count number of pages beyond high watermark
5093 * @offset: The zone index of the highest zone
5095 * nr_free_zone_pages() counts the number of pages which are beyond the
5096 * high watermark within all zones at or below a given zone index. For each
5097 * zone, the number of pages is calculated as:
5099 * nr_free_zone_pages = managed_pages - high_pages
5101 * Return: number of pages beyond high watermark.
5103 static unsigned long nr_free_zone_pages(int offset)
5108 /* Just pick one node, since fallback list is circular */
5109 unsigned long sum = 0;
5111 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5113 for_each_zone_zonelist(zone, z, zonelist, offset) {
5114 unsigned long size = zone_managed_pages(zone);
5115 unsigned long high = high_wmark_pages(zone);
5124 * nr_free_buffer_pages - count number of pages beyond high watermark
5126 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5127 * watermark within ZONE_DMA and ZONE_NORMAL.
5129 * Return: number of pages beyond high watermark within ZONE_DMA and
5132 unsigned long nr_free_buffer_pages(void)
5134 return nr_free_zone_pages(gfp_zone(GFP_USER));
5136 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5139 * nr_free_pagecache_pages - count number of pages beyond high watermark
5141 * nr_free_pagecache_pages() counts the number of pages which are beyond the
5142 * high watermark within all zones.
5144 * Return: number of pages beyond high watermark within all zones.
5146 unsigned long nr_free_pagecache_pages(void)
5148 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5151 static inline void show_node(struct zone *zone)
5153 if (IS_ENABLED(CONFIG_NUMA))
5154 printk("Node %d ", zone_to_nid(zone));
5157 long si_mem_available(void)
5160 unsigned long pagecache;
5161 unsigned long wmark_low = 0;
5162 unsigned long pages[NR_LRU_LISTS];
5163 unsigned long reclaimable;
5167 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5168 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5171 wmark_low += low_wmark_pages(zone);
5174 * Estimate the amount of memory available for userspace allocations,
5175 * without causing swapping.
5177 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5180 * Not all the page cache can be freed, otherwise the system will
5181 * start swapping. Assume at least half of the page cache, or the
5182 * low watermark worth of cache, needs to stay.
5184 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5185 pagecache -= min(pagecache / 2, wmark_low);
5186 available += pagecache;
5189 * Part of the reclaimable slab and other kernel memory consists of
5190 * items that are in use, and cannot be freed. Cap this estimate at the
5193 reclaimable = global_node_page_state(NR_SLAB_RECLAIMABLE) +
5194 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5195 available += reclaimable - min(reclaimable / 2, wmark_low);
5201 EXPORT_SYMBOL_GPL(si_mem_available);
5203 void si_meminfo(struct sysinfo *val)
5205 val->totalram = totalram_pages();
5206 val->sharedram = global_node_page_state(NR_SHMEM);
5207 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5208 val->bufferram = nr_blockdev_pages();
5209 val->totalhigh = totalhigh_pages();
5210 val->freehigh = nr_free_highpages();
5211 val->mem_unit = PAGE_SIZE;
5214 EXPORT_SYMBOL(si_meminfo);
5217 void si_meminfo_node(struct sysinfo *val, int nid)
5219 int zone_type; /* needs to be signed */
5220 unsigned long managed_pages = 0;
5221 unsigned long managed_highpages = 0;
5222 unsigned long free_highpages = 0;
5223 pg_data_t *pgdat = NODE_DATA(nid);
5225 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5226 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5227 val->totalram = managed_pages;
5228 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5229 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5230 #ifdef CONFIG_HIGHMEM
5231 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5232 struct zone *zone = &pgdat->node_zones[zone_type];
5234 if (is_highmem(zone)) {
5235 managed_highpages += zone_managed_pages(zone);
5236 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5239 val->totalhigh = managed_highpages;
5240 val->freehigh = free_highpages;
5242 val->totalhigh = managed_highpages;
5243 val->freehigh = free_highpages;
5245 val->mem_unit = PAGE_SIZE;
5250 * Determine whether the node should be displayed or not, depending on whether
5251 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5253 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5255 if (!(flags & SHOW_MEM_FILTER_NODES))
5259 * no node mask - aka implicit memory numa policy. Do not bother with
5260 * the synchronization - read_mems_allowed_begin - because we do not
5261 * have to be precise here.
5264 nodemask = &cpuset_current_mems_allowed;
5266 return !node_isset(nid, *nodemask);
5269 #define K(x) ((x) << (PAGE_SHIFT-10))
5271 static void show_migration_types(unsigned char type)
5273 static const char types[MIGRATE_TYPES] = {
5274 [MIGRATE_UNMOVABLE] = 'U',
5275 [MIGRATE_MOVABLE] = 'M',
5276 [MIGRATE_RECLAIMABLE] = 'E',
5277 [MIGRATE_HIGHATOMIC] = 'H',
5279 [MIGRATE_CMA] = 'C',
5281 #ifdef CONFIG_MEMORY_ISOLATION
5282 [MIGRATE_ISOLATE] = 'I',
5285 char tmp[MIGRATE_TYPES + 1];
5289 for (i = 0; i < MIGRATE_TYPES; i++) {
5290 if (type & (1 << i))
5295 printk(KERN_CONT "(%s) ", tmp);
5299 * Show free area list (used inside shift_scroll-lock stuff)
5300 * We also calculate the percentage fragmentation. We do this by counting the
5301 * memory on each free list with the exception of the first item on the list.
5304 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5307 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5309 unsigned long free_pcp = 0;
5314 for_each_populated_zone(zone) {
5315 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5318 for_each_online_cpu(cpu)
5319 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5322 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5323 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5324 " unevictable:%lu dirty:%lu writeback:%lu\n"
5325 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5326 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5327 " free:%lu free_pcp:%lu free_cma:%lu\n",
5328 global_node_page_state(NR_ACTIVE_ANON),
5329 global_node_page_state(NR_INACTIVE_ANON),
5330 global_node_page_state(NR_ISOLATED_ANON),
5331 global_node_page_state(NR_ACTIVE_FILE),
5332 global_node_page_state(NR_INACTIVE_FILE),
5333 global_node_page_state(NR_ISOLATED_FILE),
5334 global_node_page_state(NR_UNEVICTABLE),
5335 global_node_page_state(NR_FILE_DIRTY),
5336 global_node_page_state(NR_WRITEBACK),
5337 global_node_page_state(NR_SLAB_RECLAIMABLE),
5338 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
5339 global_node_page_state(NR_FILE_MAPPED),
5340 global_node_page_state(NR_SHMEM),
5341 global_zone_page_state(NR_PAGETABLE),
5342 global_zone_page_state(NR_BOUNCE),
5343 global_zone_page_state(NR_FREE_PAGES),
5345 global_zone_page_state(NR_FREE_CMA_PAGES));
5347 for_each_online_pgdat(pgdat) {
5348 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5352 " active_anon:%lukB"
5353 " inactive_anon:%lukB"
5354 " active_file:%lukB"
5355 " inactive_file:%lukB"
5356 " unevictable:%lukB"
5357 " isolated(anon):%lukB"
5358 " isolated(file):%lukB"
5363 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5365 " shmem_pmdmapped: %lukB"
5368 " writeback_tmp:%lukB"
5369 " all_unreclaimable? %s"
5372 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5373 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5374 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5375 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5376 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5377 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5378 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5379 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5380 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5381 K(node_page_state(pgdat, NR_WRITEBACK)),
5382 K(node_page_state(pgdat, NR_SHMEM)),
5383 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5384 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5385 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5387 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5389 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5390 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5394 for_each_populated_zone(zone) {
5397 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5401 for_each_online_cpu(cpu)
5402 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5411 " reserved_highatomic:%luKB"
5412 " active_anon:%lukB"
5413 " inactive_anon:%lukB"
5414 " active_file:%lukB"
5415 " inactive_file:%lukB"
5416 " unevictable:%lukB"
5417 " writepending:%lukB"
5421 " kernel_stack:%lukB"
5422 #ifdef CONFIG_SHADOW_CALL_STACK
5423 " shadow_call_stack:%lukB"
5432 K(zone_page_state(zone, NR_FREE_PAGES)),
5433 K(min_wmark_pages(zone)),
5434 K(low_wmark_pages(zone)),
5435 K(high_wmark_pages(zone)),
5436 K(zone->nr_reserved_highatomic),
5437 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5438 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5439 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5440 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5441 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5442 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5443 K(zone->present_pages),
5444 K(zone_managed_pages(zone)),
5445 K(zone_page_state(zone, NR_MLOCK)),
5446 zone_page_state(zone, NR_KERNEL_STACK_KB),
5447 #ifdef CONFIG_SHADOW_CALL_STACK
5448 zone_page_state(zone, NR_KERNEL_SCS_KB),
5450 K(zone_page_state(zone, NR_PAGETABLE)),
5451 K(zone_page_state(zone, NR_BOUNCE)),
5453 K(this_cpu_read(zone->pageset->pcp.count)),
5454 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5455 printk("lowmem_reserve[]:");
5456 for (i = 0; i < MAX_NR_ZONES; i++)
5457 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5458 printk(KERN_CONT "\n");
5461 for_each_populated_zone(zone) {
5463 unsigned long nr[MAX_ORDER], flags, total = 0;
5464 unsigned char types[MAX_ORDER];
5466 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5469 printk(KERN_CONT "%s: ", zone->name);
5471 spin_lock_irqsave(&zone->lock, flags);
5472 for (order = 0; order < MAX_ORDER; order++) {
5473 struct free_area *area = &zone->free_area[order];
5476 nr[order] = area->nr_free;
5477 total += nr[order] << order;
5480 for (type = 0; type < MIGRATE_TYPES; type++) {
5481 if (!free_area_empty(area, type))
5482 types[order] |= 1 << type;
5485 spin_unlock_irqrestore(&zone->lock, flags);
5486 for (order = 0; order < MAX_ORDER; order++) {
5487 printk(KERN_CONT "%lu*%lukB ",
5488 nr[order], K(1UL) << order);
5490 show_migration_types(types[order]);
5492 printk(KERN_CONT "= %lukB\n", K(total));
5495 hugetlb_show_meminfo();
5497 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5499 show_swap_cache_info();
5502 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5504 zoneref->zone = zone;
5505 zoneref->zone_idx = zone_idx(zone);
5509 * Builds allocation fallback zone lists.
5511 * Add all populated zones of a node to the zonelist.
5513 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5516 enum zone_type zone_type = MAX_NR_ZONES;
5521 zone = pgdat->node_zones + zone_type;
5522 if (managed_zone(zone)) {
5523 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5524 check_highest_zone(zone_type);
5526 } while (zone_type);
5533 static int __parse_numa_zonelist_order(char *s)
5536 * We used to support different zonlists modes but they turned
5537 * out to be just not useful. Let's keep the warning in place
5538 * if somebody still use the cmd line parameter so that we do
5539 * not fail it silently
5541 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5542 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5548 static __init int setup_numa_zonelist_order(char *s)
5553 return __parse_numa_zonelist_order(s);
5555 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5557 char numa_zonelist_order[] = "Node";
5560 * sysctl handler for numa_zonelist_order
5562 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5563 void __user *buffer, size_t *length,
5570 return proc_dostring(table, write, buffer, length, ppos);
5571 str = memdup_user_nul(buffer, 16);
5573 return PTR_ERR(str);
5575 ret = __parse_numa_zonelist_order(str);
5581 #define MAX_NODE_LOAD (nr_online_nodes)
5582 static int node_load[MAX_NUMNODES];
5585 * find_next_best_node - find the next node that should appear in a given node's fallback list
5586 * @node: node whose fallback list we're appending
5587 * @used_node_mask: nodemask_t of already used nodes
5589 * We use a number of factors to determine which is the next node that should
5590 * appear on a given node's fallback list. The node should not have appeared
5591 * already in @node's fallback list, and it should be the next closest node
5592 * according to the distance array (which contains arbitrary distance values
5593 * from each node to each node in the system), and should also prefer nodes
5594 * with no CPUs, since presumably they'll have very little allocation pressure
5595 * on them otherwise.
5597 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5599 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5602 int min_val = INT_MAX;
5603 int best_node = NUMA_NO_NODE;
5604 const struct cpumask *tmp = cpumask_of_node(0);
5606 /* Use the local node if we haven't already */
5607 if (!node_isset(node, *used_node_mask)) {
5608 node_set(node, *used_node_mask);
5612 for_each_node_state(n, N_MEMORY) {
5614 /* Don't want a node to appear more than once */
5615 if (node_isset(n, *used_node_mask))
5618 /* Use the distance array to find the distance */
5619 val = node_distance(node, n);
5621 /* Penalize nodes under us ("prefer the next node") */
5624 /* Give preference to headless and unused nodes */
5625 tmp = cpumask_of_node(n);
5626 if (!cpumask_empty(tmp))
5627 val += PENALTY_FOR_NODE_WITH_CPUS;
5629 /* Slight preference for less loaded node */
5630 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5631 val += node_load[n];
5633 if (val < min_val) {
5640 node_set(best_node, *used_node_mask);
5647 * Build zonelists ordered by node and zones within node.
5648 * This results in maximum locality--normal zone overflows into local
5649 * DMA zone, if any--but risks exhausting DMA zone.
5651 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5654 struct zoneref *zonerefs;
5657 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5659 for (i = 0; i < nr_nodes; i++) {
5662 pg_data_t *node = NODE_DATA(node_order[i]);
5664 nr_zones = build_zonerefs_node(node, zonerefs);
5665 zonerefs += nr_zones;
5667 zonerefs->zone = NULL;
5668 zonerefs->zone_idx = 0;
5672 * Build gfp_thisnode zonelists
5674 static void build_thisnode_zonelists(pg_data_t *pgdat)
5676 struct zoneref *zonerefs;
5679 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5680 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5681 zonerefs += nr_zones;
5682 zonerefs->zone = NULL;
5683 zonerefs->zone_idx = 0;
5687 * Build zonelists ordered by zone and nodes within zones.
5688 * This results in conserving DMA zone[s] until all Normal memory is
5689 * exhausted, but results in overflowing to remote node while memory
5690 * may still exist in local DMA zone.
5693 static void build_zonelists(pg_data_t *pgdat)
5695 static int node_order[MAX_NUMNODES];
5696 int node, load, nr_nodes = 0;
5697 nodemask_t used_mask;
5698 int local_node, prev_node;
5700 /* NUMA-aware ordering of nodes */
5701 local_node = pgdat->node_id;
5702 load = nr_online_nodes;
5703 prev_node = local_node;
5704 nodes_clear(used_mask);
5706 memset(node_order, 0, sizeof(node_order));
5707 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5709 * We don't want to pressure a particular node.
5710 * So adding penalty to the first node in same
5711 * distance group to make it round-robin.
5713 if (node_distance(local_node, node) !=
5714 node_distance(local_node, prev_node))
5715 node_load[node] = load;
5717 node_order[nr_nodes++] = node;
5722 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5723 build_thisnode_zonelists(pgdat);
5726 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5728 * Return node id of node used for "local" allocations.
5729 * I.e., first node id of first zone in arg node's generic zonelist.
5730 * Used for initializing percpu 'numa_mem', which is used primarily
5731 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5733 int local_memory_node(int node)
5737 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5738 gfp_zone(GFP_KERNEL),
5740 return zone_to_nid(z->zone);
5744 static void setup_min_unmapped_ratio(void);
5745 static void setup_min_slab_ratio(void);
5746 #else /* CONFIG_NUMA */
5748 static void build_zonelists(pg_data_t *pgdat)
5750 int node, local_node;
5751 struct zoneref *zonerefs;
5754 local_node = pgdat->node_id;
5756 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5757 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5758 zonerefs += nr_zones;
5761 * Now we build the zonelist so that it contains the zones
5762 * of all the other nodes.
5763 * We don't want to pressure a particular node, so when
5764 * building the zones for node N, we make sure that the
5765 * zones coming right after the local ones are those from
5766 * node N+1 (modulo N)
5768 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5769 if (!node_online(node))
5771 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5772 zonerefs += nr_zones;
5774 for (node = 0; node < local_node; node++) {
5775 if (!node_online(node))
5777 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5778 zonerefs += nr_zones;
5781 zonerefs->zone = NULL;
5782 zonerefs->zone_idx = 0;
5785 #endif /* CONFIG_NUMA */
5788 * Boot pageset table. One per cpu which is going to be used for all
5789 * zones and all nodes. The parameters will be set in such a way
5790 * that an item put on a list will immediately be handed over to
5791 * the buddy list. This is safe since pageset manipulation is done
5792 * with interrupts disabled.
5794 * The boot_pagesets must be kept even after bootup is complete for
5795 * unused processors and/or zones. They do play a role for bootstrapping
5796 * hotplugged processors.
5798 * zoneinfo_show() and maybe other functions do
5799 * not check if the processor is online before following the pageset pointer.
5800 * Other parts of the kernel may not check if the zone is available.
5802 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5803 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5804 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5806 static void __build_all_zonelists(void *data)
5809 int __maybe_unused cpu;
5810 pg_data_t *self = data;
5811 static DEFINE_SPINLOCK(lock);
5816 memset(node_load, 0, sizeof(node_load));
5820 * This node is hotadded and no memory is yet present. So just
5821 * building zonelists is fine - no need to touch other nodes.
5823 if (self && !node_online(self->node_id)) {
5824 build_zonelists(self);
5826 for_each_online_node(nid) {
5827 pg_data_t *pgdat = NODE_DATA(nid);
5829 build_zonelists(pgdat);
5832 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5834 * We now know the "local memory node" for each node--
5835 * i.e., the node of the first zone in the generic zonelist.
5836 * Set up numa_mem percpu variable for on-line cpus. During
5837 * boot, only the boot cpu should be on-line; we'll init the
5838 * secondary cpus' numa_mem as they come on-line. During
5839 * node/memory hotplug, we'll fixup all on-line cpus.
5841 for_each_online_cpu(cpu)
5842 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5849 static noinline void __init
5850 build_all_zonelists_init(void)
5854 __build_all_zonelists(NULL);
5857 * Initialize the boot_pagesets that are going to be used
5858 * for bootstrapping processors. The real pagesets for
5859 * each zone will be allocated later when the per cpu
5860 * allocator is available.
5862 * boot_pagesets are used also for bootstrapping offline
5863 * cpus if the system is already booted because the pagesets
5864 * are needed to initialize allocators on a specific cpu too.
5865 * F.e. the percpu allocator needs the page allocator which
5866 * needs the percpu allocator in order to allocate its pagesets
5867 * (a chicken-egg dilemma).
5869 for_each_possible_cpu(cpu)
5870 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5872 mminit_verify_zonelist();
5873 cpuset_init_current_mems_allowed();
5877 * unless system_state == SYSTEM_BOOTING.
5879 * __ref due to call of __init annotated helper build_all_zonelists_init
5880 * [protected by SYSTEM_BOOTING].
5882 void __ref build_all_zonelists(pg_data_t *pgdat)
5884 if (system_state == SYSTEM_BOOTING) {
5885 build_all_zonelists_init();
5887 __build_all_zonelists(pgdat);
5888 /* cpuset refresh routine should be here */
5890 vm_total_pages = nr_free_pagecache_pages();
5892 * Disable grouping by mobility if the number of pages in the
5893 * system is too low to allow the mechanism to work. It would be
5894 * more accurate, but expensive to check per-zone. This check is
5895 * made on memory-hotadd so a system can start with mobility
5896 * disabled and enable it later
5898 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5899 page_group_by_mobility_disabled = 1;
5901 page_group_by_mobility_disabled = 0;
5903 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5905 page_group_by_mobility_disabled ? "off" : "on",
5908 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5912 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5913 static bool __meminit
5914 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5916 static struct memblock_region *r;
5918 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5919 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5920 for_each_memblock(memory, r) {
5921 if (*pfn < memblock_region_memory_end_pfn(r))
5925 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5926 memblock_is_mirror(r)) {
5927 *pfn = memblock_region_memory_end_pfn(r);
5935 * Initially all pages are reserved - free ones are freed
5936 * up by memblock_free_all() once the early boot process is
5937 * done. Non-atomic initialization, single-pass.
5939 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5940 unsigned long start_pfn, enum memmap_context context,
5941 struct vmem_altmap *altmap)
5943 unsigned long pfn, end_pfn = start_pfn + size;
5946 if (highest_memmap_pfn < end_pfn - 1)
5947 highest_memmap_pfn = end_pfn - 1;
5949 #ifdef CONFIG_ZONE_DEVICE
5951 * Honor reservation requested by the driver for this ZONE_DEVICE
5952 * memory. We limit the total number of pages to initialize to just
5953 * those that might contain the memory mapping. We will defer the
5954 * ZONE_DEVICE page initialization until after we have released
5957 if (zone == ZONE_DEVICE) {
5961 if (start_pfn == altmap->base_pfn)
5962 start_pfn += altmap->reserve;
5963 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5967 for (pfn = start_pfn; pfn < end_pfn; ) {
5969 * There can be holes in boot-time mem_map[]s handed to this
5970 * function. They do not exist on hotplugged memory.
5972 if (context == MEMMAP_EARLY) {
5973 if (overlap_memmap_init(zone, &pfn))
5975 if (defer_init(nid, pfn, end_pfn))
5979 page = pfn_to_page(pfn);
5980 __init_single_page(page, pfn, zone, nid);
5981 if (context == MEMMAP_HOTPLUG)
5982 __SetPageReserved(page);
5985 * Mark the block movable so that blocks are reserved for
5986 * movable at startup. This will force kernel allocations
5987 * to reserve their blocks rather than leaking throughout
5988 * the address space during boot when many long-lived
5989 * kernel allocations are made.
5991 * bitmap is created for zone's valid pfn range. but memmap
5992 * can be created for invalid pages (for alignment)
5993 * check here not to call set_pageblock_migratetype() against
5996 if (!(pfn & (pageblock_nr_pages - 1))) {
5997 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6004 #ifdef CONFIG_ZONE_DEVICE
6005 void __ref memmap_init_zone_device(struct zone *zone,
6006 unsigned long start_pfn,
6007 unsigned long nr_pages,
6008 struct dev_pagemap *pgmap)
6010 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6011 struct pglist_data *pgdat = zone->zone_pgdat;
6012 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6013 unsigned long zone_idx = zone_idx(zone);
6014 unsigned long start = jiffies;
6015 int nid = pgdat->node_id;
6017 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6021 * The call to memmap_init_zone should have already taken care
6022 * of the pages reserved for the memmap, so we can just jump to
6023 * the end of that region and start processing the device pages.
6026 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6027 nr_pages = end_pfn - start_pfn;
6030 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6031 struct page *page = pfn_to_page(pfn);
6033 __init_single_page(page, pfn, zone_idx, nid);
6036 * Mark page reserved as it will need to wait for onlining
6037 * phase for it to be fully associated with a zone.
6039 * We can use the non-atomic __set_bit operation for setting
6040 * the flag as we are still initializing the pages.
6042 __SetPageReserved(page);
6045 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6046 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6047 * ever freed or placed on a driver-private list.
6049 page->pgmap = pgmap;
6050 page->zone_device_data = NULL;
6053 * Mark the block movable so that blocks are reserved for
6054 * movable at startup. This will force kernel allocations
6055 * to reserve their blocks rather than leaking throughout
6056 * the address space during boot when many long-lived
6057 * kernel allocations are made.
6059 * bitmap is created for zone's valid pfn range. but memmap
6060 * can be created for invalid pages (for alignment)
6061 * check here not to call set_pageblock_migratetype() against
6064 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
6065 * because this is done early in section_activate()
6067 if (!(pfn & (pageblock_nr_pages - 1))) {
6068 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6073 pr_info("%s initialised %lu pages in %ums\n", __func__,
6074 nr_pages, jiffies_to_msecs(jiffies - start));
6078 static void __meminit zone_init_free_lists(struct zone *zone)
6080 unsigned int order, t;
6081 for_each_migratetype_order(order, t) {
6082 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6083 zone->free_area[order].nr_free = 0;
6087 void __meminit __weak memmap_init(unsigned long size, int nid,
6089 unsigned long range_start_pfn)
6091 unsigned long start_pfn, end_pfn;
6092 unsigned long range_end_pfn = range_start_pfn + size;
6095 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6096 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6097 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6099 if (end_pfn > start_pfn) {
6100 size = end_pfn - start_pfn;
6101 memmap_init_zone(size, nid, zone, start_pfn,
6102 MEMMAP_EARLY, NULL);
6107 static int zone_batchsize(struct zone *zone)
6113 * The per-cpu-pages pools are set to around 1000th of the
6116 batch = zone_managed_pages(zone) / 1024;
6117 /* But no more than a meg. */
6118 if (batch * PAGE_SIZE > 1024 * 1024)
6119 batch = (1024 * 1024) / PAGE_SIZE;
6120 batch /= 4; /* We effectively *= 4 below */
6125 * Clamp the batch to a 2^n - 1 value. Having a power
6126 * of 2 value was found to be more likely to have
6127 * suboptimal cache aliasing properties in some cases.
6129 * For example if 2 tasks are alternately allocating
6130 * batches of pages, one task can end up with a lot
6131 * of pages of one half of the possible page colors
6132 * and the other with pages of the other colors.
6134 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6139 /* The deferral and batching of frees should be suppressed under NOMMU
6142 * The problem is that NOMMU needs to be able to allocate large chunks
6143 * of contiguous memory as there's no hardware page translation to
6144 * assemble apparent contiguous memory from discontiguous pages.
6146 * Queueing large contiguous runs of pages for batching, however,
6147 * causes the pages to actually be freed in smaller chunks. As there
6148 * can be a significant delay between the individual batches being
6149 * recycled, this leads to the once large chunks of space being
6150 * fragmented and becoming unavailable for high-order allocations.
6157 * pcp->high and pcp->batch values are related and dependent on one another:
6158 * ->batch must never be higher then ->high.
6159 * The following function updates them in a safe manner without read side
6162 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6163 * those fields changing asynchronously (acording the the above rule).
6165 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6166 * outside of boot time (or some other assurance that no concurrent updaters
6169 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6170 unsigned long batch)
6172 /* start with a fail safe value for batch */
6176 /* Update high, then batch, in order */
6183 /* a companion to pageset_set_high() */
6184 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6186 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6189 static void pageset_init(struct per_cpu_pageset *p)
6191 struct per_cpu_pages *pcp;
6194 memset(p, 0, sizeof(*p));
6197 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6198 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6201 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6204 pageset_set_batch(p, batch);
6208 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6209 * to the value high for the pageset p.
6211 static void pageset_set_high(struct per_cpu_pageset *p,
6214 unsigned long batch = max(1UL, high / 4);
6215 if ((high / 4) > (PAGE_SHIFT * 8))
6216 batch = PAGE_SHIFT * 8;
6218 pageset_update(&p->pcp, high, batch);
6221 static void pageset_set_high_and_batch(struct zone *zone,
6222 struct per_cpu_pageset *pcp)
6224 if (percpu_pagelist_fraction)
6225 pageset_set_high(pcp,
6226 (zone_managed_pages(zone) /
6227 percpu_pagelist_fraction));
6229 pageset_set_batch(pcp, zone_batchsize(zone));
6232 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6234 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6237 pageset_set_high_and_batch(zone, pcp);
6240 void __meminit setup_zone_pageset(struct zone *zone)
6243 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6244 for_each_possible_cpu(cpu)
6245 zone_pageset_init(zone, cpu);
6249 * Allocate per cpu pagesets and initialize them.
6250 * Before this call only boot pagesets were available.
6252 void __init setup_per_cpu_pageset(void)
6254 struct pglist_data *pgdat;
6257 for_each_populated_zone(zone)
6258 setup_zone_pageset(zone);
6260 for_each_online_pgdat(pgdat)
6261 pgdat->per_cpu_nodestats =
6262 alloc_percpu(struct per_cpu_nodestat);
6265 static __meminit void zone_pcp_init(struct zone *zone)
6268 * per cpu subsystem is not up at this point. The following code
6269 * relies on the ability of the linker to provide the
6270 * offset of a (static) per cpu variable into the per cpu area.
6272 zone->pageset = &boot_pageset;
6274 if (populated_zone(zone))
6275 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6276 zone->name, zone->present_pages,
6277 zone_batchsize(zone));
6280 void __meminit init_currently_empty_zone(struct zone *zone,
6281 unsigned long zone_start_pfn,
6284 struct pglist_data *pgdat = zone->zone_pgdat;
6285 int zone_idx = zone_idx(zone) + 1;
6287 if (zone_idx > pgdat->nr_zones)
6288 pgdat->nr_zones = zone_idx;
6290 zone->zone_start_pfn = zone_start_pfn;
6292 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6293 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6295 (unsigned long)zone_idx(zone),
6296 zone_start_pfn, (zone_start_pfn + size));
6298 zone_init_free_lists(zone);
6299 zone->initialized = 1;
6303 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
6304 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
6305 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
6307 * If an architecture guarantees that all ranges registered contain no holes
6308 * and may be freed, this this function may be used instead of calling
6309 * memblock_free_early_nid() manually.
6311 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
6313 unsigned long start_pfn, end_pfn;
6316 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
6317 start_pfn = min(start_pfn, max_low_pfn);
6318 end_pfn = min(end_pfn, max_low_pfn);
6320 if (start_pfn < end_pfn)
6321 memblock_free_early_nid(PFN_PHYS(start_pfn),
6322 (end_pfn - start_pfn) << PAGE_SHIFT,
6328 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6329 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6331 * If an architecture guarantees that all ranges registered contain no holes and may
6332 * be freed, this function may be used instead of calling memory_present() manually.
6334 void __init sparse_memory_present_with_active_regions(int nid)
6336 unsigned long start_pfn, end_pfn;
6339 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
6340 memory_present(this_nid, start_pfn, end_pfn);
6344 * get_pfn_range_for_nid - Return the start and end page frames for a node
6345 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6346 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6347 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6349 * It returns the start and end page frame of a node based on information
6350 * provided by memblock_set_node(). If called for a node
6351 * with no available memory, a warning is printed and the start and end
6354 void __init get_pfn_range_for_nid(unsigned int nid,
6355 unsigned long *start_pfn, unsigned long *end_pfn)
6357 unsigned long this_start_pfn, this_end_pfn;
6363 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6364 *start_pfn = min(*start_pfn, this_start_pfn);
6365 *end_pfn = max(*end_pfn, this_end_pfn);
6368 if (*start_pfn == -1UL)
6373 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6374 * assumption is made that zones within a node are ordered in monotonic
6375 * increasing memory addresses so that the "highest" populated zone is used
6377 static void __init find_usable_zone_for_movable(void)
6380 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6381 if (zone_index == ZONE_MOVABLE)
6384 if (arch_zone_highest_possible_pfn[zone_index] >
6385 arch_zone_lowest_possible_pfn[zone_index])
6389 VM_BUG_ON(zone_index == -1);
6390 movable_zone = zone_index;
6394 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6395 * because it is sized independent of architecture. Unlike the other zones,
6396 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6397 * in each node depending on the size of each node and how evenly kernelcore
6398 * is distributed. This helper function adjusts the zone ranges
6399 * provided by the architecture for a given node by using the end of the
6400 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6401 * zones within a node are in order of monotonic increases memory addresses
6403 static void __init adjust_zone_range_for_zone_movable(int nid,
6404 unsigned long zone_type,
6405 unsigned long node_start_pfn,
6406 unsigned long node_end_pfn,
6407 unsigned long *zone_start_pfn,
6408 unsigned long *zone_end_pfn)
6410 /* Only adjust if ZONE_MOVABLE is on this node */
6411 if (zone_movable_pfn[nid]) {
6412 /* Size ZONE_MOVABLE */
6413 if (zone_type == ZONE_MOVABLE) {
6414 *zone_start_pfn = zone_movable_pfn[nid];
6415 *zone_end_pfn = min(node_end_pfn,
6416 arch_zone_highest_possible_pfn[movable_zone]);
6418 /* Adjust for ZONE_MOVABLE starting within this range */
6419 } else if (!mirrored_kernelcore &&
6420 *zone_start_pfn < zone_movable_pfn[nid] &&
6421 *zone_end_pfn > zone_movable_pfn[nid]) {
6422 *zone_end_pfn = zone_movable_pfn[nid];
6424 /* Check if this whole range is within ZONE_MOVABLE */
6425 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6426 *zone_start_pfn = *zone_end_pfn;
6431 * Return the number of pages a zone spans in a node, including holes
6432 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6434 static unsigned long __init zone_spanned_pages_in_node(int nid,
6435 unsigned long zone_type,
6436 unsigned long node_start_pfn,
6437 unsigned long node_end_pfn,
6438 unsigned long *zone_start_pfn,
6439 unsigned long *zone_end_pfn,
6440 unsigned long *ignored)
6442 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6443 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6444 /* When hotadd a new node from cpu_up(), the node should be empty */
6445 if (!node_start_pfn && !node_end_pfn)
6448 /* Get the start and end of the zone */
6449 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6450 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6451 adjust_zone_range_for_zone_movable(nid, zone_type,
6452 node_start_pfn, node_end_pfn,
6453 zone_start_pfn, zone_end_pfn);
6455 /* Check that this node has pages within the zone's required range */
6456 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6459 /* Move the zone boundaries inside the node if necessary */
6460 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6461 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6463 /* Return the spanned pages */
6464 return *zone_end_pfn - *zone_start_pfn;
6468 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6469 * then all holes in the requested range will be accounted for.
6471 unsigned long __init __absent_pages_in_range(int nid,
6472 unsigned long range_start_pfn,
6473 unsigned long range_end_pfn)
6475 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6476 unsigned long start_pfn, end_pfn;
6479 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6480 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6481 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6482 nr_absent -= end_pfn - start_pfn;
6488 * absent_pages_in_range - Return number of page frames in holes within a range
6489 * @start_pfn: The start PFN to start searching for holes
6490 * @end_pfn: The end PFN to stop searching for holes
6492 * Return: the number of pages frames in memory holes within a range.
6494 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6495 unsigned long end_pfn)
6497 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6500 /* Return the number of page frames in holes in a zone on a node */
6501 static unsigned long __init zone_absent_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 *ignored)
6507 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6508 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6509 unsigned long zone_start_pfn, zone_end_pfn;
6510 unsigned long nr_absent;
6512 /* When hotadd a new node from cpu_up(), the node should be empty */
6513 if (!node_start_pfn && !node_end_pfn)
6516 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6517 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6519 adjust_zone_range_for_zone_movable(nid, zone_type,
6520 node_start_pfn, node_end_pfn,
6521 &zone_start_pfn, &zone_end_pfn);
6522 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6525 * ZONE_MOVABLE handling.
6526 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6529 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6530 unsigned long start_pfn, end_pfn;
6531 struct memblock_region *r;
6533 for_each_memblock(memory, r) {
6534 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6535 zone_start_pfn, zone_end_pfn);
6536 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6537 zone_start_pfn, zone_end_pfn);
6539 if (zone_type == ZONE_MOVABLE &&
6540 memblock_is_mirror(r))
6541 nr_absent += end_pfn - start_pfn;
6543 if (zone_type == ZONE_NORMAL &&
6544 !memblock_is_mirror(r))
6545 nr_absent += end_pfn - start_pfn;
6552 static inline unsigned long __init compat_zone_spanned_pages_in_node(int nid,
6553 unsigned long zone_type,
6554 unsigned long node_start_pfn,
6555 unsigned long node_end_pfn,
6556 unsigned long *zone_start_pfn,
6557 unsigned long *zone_end_pfn,
6558 unsigned long *zones_size)
6562 *zone_start_pfn = node_start_pfn;
6563 for (zone = 0; zone < zone_type; zone++)
6564 *zone_start_pfn += zones_size[zone];
6566 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6568 return zones_size[zone_type];
6571 static inline unsigned long __init compat_zone_absent_pages_in_node(int nid,
6572 unsigned long zone_type,
6573 unsigned long node_start_pfn,
6574 unsigned long node_end_pfn,
6575 unsigned long *zholes_size)
6580 return zholes_size[zone_type];
6583 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6584 unsigned long node_start_pfn,
6585 unsigned long node_end_pfn,
6586 unsigned long *zones_size,
6587 unsigned long *zholes_size,
6590 unsigned long realtotalpages = 0, totalpages = 0;
6593 for (i = 0; i < MAX_NR_ZONES; i++) {
6594 struct zone *zone = pgdat->node_zones + i;
6595 unsigned long zone_start_pfn, zone_end_pfn;
6596 unsigned long spanned, absent;
6597 unsigned long size, real_size;
6600 spanned = compat_zone_spanned_pages_in_node(
6607 absent = compat_zone_absent_pages_in_node(
6613 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
6619 absent = zone_absent_pages_in_node(pgdat->node_id, i,
6626 real_size = size - absent;
6629 zone->zone_start_pfn = zone_start_pfn;
6631 zone->zone_start_pfn = 0;
6632 zone->spanned_pages = size;
6633 zone->present_pages = real_size;
6636 realtotalpages += real_size;
6639 pgdat->node_spanned_pages = totalpages;
6640 pgdat->node_present_pages = realtotalpages;
6641 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6645 #ifndef CONFIG_SPARSEMEM
6647 * Calculate the size of the zone->blockflags rounded to an unsigned long
6648 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6649 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6650 * round what is now in bits to nearest long in bits, then return it in
6653 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6655 unsigned long usemapsize;
6657 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6658 usemapsize = roundup(zonesize, pageblock_nr_pages);
6659 usemapsize = usemapsize >> pageblock_order;
6660 usemapsize *= NR_PAGEBLOCK_BITS;
6661 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6663 return usemapsize / 8;
6666 static void __ref setup_usemap(struct pglist_data *pgdat,
6668 unsigned long zone_start_pfn,
6669 unsigned long zonesize)
6671 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6672 zone->pageblock_flags = NULL;
6674 zone->pageblock_flags =
6675 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6677 if (!zone->pageblock_flags)
6678 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6679 usemapsize, zone->name, pgdat->node_id);
6683 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6684 unsigned long zone_start_pfn, unsigned long zonesize) {}
6685 #endif /* CONFIG_SPARSEMEM */
6687 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6689 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6690 void __init set_pageblock_order(void)
6694 /* Check that pageblock_nr_pages has not already been setup */
6695 if (pageblock_order)
6698 if (HPAGE_SHIFT > PAGE_SHIFT)
6699 order = HUGETLB_PAGE_ORDER;
6701 order = MAX_ORDER - 1;
6704 * Assume the largest contiguous order of interest is a huge page.
6705 * This value may be variable depending on boot parameters on IA64 and
6708 pageblock_order = order;
6710 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6713 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6714 * is unused as pageblock_order is set at compile-time. See
6715 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6718 void __init set_pageblock_order(void)
6722 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6724 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6725 unsigned long present_pages)
6727 unsigned long pages = spanned_pages;
6730 * Provide a more accurate estimation if there are holes within
6731 * the zone and SPARSEMEM is in use. If there are holes within the
6732 * zone, each populated memory region may cost us one or two extra
6733 * memmap pages due to alignment because memmap pages for each
6734 * populated regions may not be naturally aligned on page boundary.
6735 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6737 if (spanned_pages > present_pages + (present_pages >> 4) &&
6738 IS_ENABLED(CONFIG_SPARSEMEM))
6739 pages = present_pages;
6741 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6744 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6745 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6747 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6749 spin_lock_init(&ds_queue->split_queue_lock);
6750 INIT_LIST_HEAD(&ds_queue->split_queue);
6751 ds_queue->split_queue_len = 0;
6754 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6757 #ifdef CONFIG_COMPACTION
6758 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6760 init_waitqueue_head(&pgdat->kcompactd_wait);
6763 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6766 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6768 pgdat_resize_init(pgdat);
6770 pgdat_init_split_queue(pgdat);
6771 pgdat_init_kcompactd(pgdat);
6773 init_waitqueue_head(&pgdat->kswapd_wait);
6774 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6776 pgdat_page_ext_init(pgdat);
6777 spin_lock_init(&pgdat->lru_lock);
6778 lruvec_init(&pgdat->__lruvec);
6781 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6782 unsigned long remaining_pages)
6784 atomic_long_set(&zone->managed_pages, remaining_pages);
6785 zone_set_nid(zone, nid);
6786 zone->name = zone_names[idx];
6787 zone->zone_pgdat = NODE_DATA(nid);
6788 spin_lock_init(&zone->lock);
6789 zone_seqlock_init(zone);
6790 zone_pcp_init(zone);
6794 * Set up the zone data structures
6795 * - init pgdat internals
6796 * - init all zones belonging to this node
6798 * NOTE: this function is only called during memory hotplug
6800 #ifdef CONFIG_MEMORY_HOTPLUG
6801 void __ref free_area_init_core_hotplug(int nid)
6804 pg_data_t *pgdat = NODE_DATA(nid);
6806 pgdat_init_internals(pgdat);
6807 for (z = 0; z < MAX_NR_ZONES; z++)
6808 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6813 * Set up the zone data structures:
6814 * - mark all pages reserved
6815 * - mark all memory queues empty
6816 * - clear the memory bitmaps
6818 * NOTE: pgdat should get zeroed by caller.
6819 * NOTE: this function is only called during early init.
6821 static void __init free_area_init_core(struct pglist_data *pgdat)
6824 int nid = pgdat->node_id;
6826 pgdat_init_internals(pgdat);
6827 pgdat->per_cpu_nodestats = &boot_nodestats;
6829 for (j = 0; j < MAX_NR_ZONES; j++) {
6830 struct zone *zone = pgdat->node_zones + j;
6831 unsigned long size, freesize, memmap_pages;
6832 unsigned long zone_start_pfn = zone->zone_start_pfn;
6834 size = zone->spanned_pages;
6835 freesize = zone->present_pages;
6838 * Adjust freesize so that it accounts for how much memory
6839 * is used by this zone for memmap. This affects the watermark
6840 * and per-cpu initialisations
6842 memmap_pages = calc_memmap_size(size, freesize);
6843 if (!is_highmem_idx(j)) {
6844 if (freesize >= memmap_pages) {
6845 freesize -= memmap_pages;
6848 " %s zone: %lu pages used for memmap\n",
6849 zone_names[j], memmap_pages);
6851 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6852 zone_names[j], memmap_pages, freesize);
6855 /* Account for reserved pages */
6856 if (j == 0 && freesize > dma_reserve) {
6857 freesize -= dma_reserve;
6858 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6859 zone_names[0], dma_reserve);
6862 if (!is_highmem_idx(j))
6863 nr_kernel_pages += freesize;
6864 /* Charge for highmem memmap if there are enough kernel pages */
6865 else if (nr_kernel_pages > memmap_pages * 2)
6866 nr_kernel_pages -= memmap_pages;
6867 nr_all_pages += freesize;
6870 * Set an approximate value for lowmem here, it will be adjusted
6871 * when the bootmem allocator frees pages into the buddy system.
6872 * And all highmem pages will be managed by the buddy system.
6874 zone_init_internals(zone, j, nid, freesize);
6879 set_pageblock_order();
6880 setup_usemap(pgdat, zone, zone_start_pfn, size);
6881 init_currently_empty_zone(zone, zone_start_pfn, size);
6882 memmap_init(size, nid, j, zone_start_pfn);
6886 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6887 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6889 unsigned long __maybe_unused start = 0;
6890 unsigned long __maybe_unused offset = 0;
6892 /* Skip empty nodes */
6893 if (!pgdat->node_spanned_pages)
6896 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6897 offset = pgdat->node_start_pfn - start;
6898 /* ia64 gets its own node_mem_map, before this, without bootmem */
6899 if (!pgdat->node_mem_map) {
6900 unsigned long size, end;
6904 * The zone's endpoints aren't required to be MAX_ORDER
6905 * aligned but the node_mem_map endpoints must be in order
6906 * for the buddy allocator to function correctly.
6908 end = pgdat_end_pfn(pgdat);
6909 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6910 size = (end - start) * sizeof(struct page);
6911 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6914 panic("Failed to allocate %ld bytes for node %d memory map\n",
6915 size, pgdat->node_id);
6916 pgdat->node_mem_map = map + offset;
6918 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6919 __func__, pgdat->node_id, (unsigned long)pgdat,
6920 (unsigned long)pgdat->node_mem_map);
6921 #ifndef CONFIG_NEED_MULTIPLE_NODES
6923 * With no DISCONTIG, the global mem_map is just set as node 0's
6925 if (pgdat == NODE_DATA(0)) {
6926 mem_map = NODE_DATA(0)->node_mem_map;
6927 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6933 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6934 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6936 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6937 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6939 pgdat->first_deferred_pfn = ULONG_MAX;
6942 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6945 static void __init __free_area_init_node(int nid, unsigned long *zones_size,
6946 unsigned long node_start_pfn,
6947 unsigned long *zholes_size,
6950 pg_data_t *pgdat = NODE_DATA(nid);
6951 unsigned long start_pfn = 0;
6952 unsigned long end_pfn = 0;
6954 /* pg_data_t should be reset to zero when it's allocated */
6955 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6957 pgdat->node_id = nid;
6958 pgdat->node_start_pfn = node_start_pfn;
6959 pgdat->per_cpu_nodestats = NULL;
6961 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6962 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6963 (u64)start_pfn << PAGE_SHIFT,
6964 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6966 start_pfn = node_start_pfn;
6968 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6969 zones_size, zholes_size, compat);
6971 alloc_node_mem_map(pgdat);
6972 pgdat_set_deferred_range(pgdat);
6974 free_area_init_core(pgdat);
6977 void __init free_area_init_node(int nid, unsigned long *zones_size,
6978 unsigned long node_start_pfn,
6979 unsigned long *zholes_size)
6981 __free_area_init_node(nid, zones_size, node_start_pfn, zholes_size,
6985 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6987 * Initialize all valid struct pages in the range [spfn, epfn) and mark them
6988 * PageReserved(). Return the number of struct pages that were initialized.
6990 static u64 __init init_unavailable_range(unsigned long spfn, unsigned long epfn)
6995 for (pfn = spfn; pfn < epfn; pfn++) {
6996 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6997 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6998 + pageblock_nr_pages - 1;
7002 * Use a fake node/zone (0) for now. Some of these pages
7003 * (in memblock.reserved but not in memblock.memory) will
7004 * get re-initialized via reserve_bootmem_region() later.
7006 __init_single_page(pfn_to_page(pfn), pfn, 0, 0);
7007 __SetPageReserved(pfn_to_page(pfn));
7015 * Only struct pages that are backed by physical memory are zeroed and
7016 * initialized by going through __init_single_page(). But, there are some
7017 * struct pages which are reserved in memblock allocator and their fields
7018 * may be accessed (for example page_to_pfn() on some configuration accesses
7019 * flags). We must explicitly initialize those struct pages.
7021 * This function also addresses a similar issue where struct pages are left
7022 * uninitialized because the physical address range is not covered by
7023 * memblock.memory or memblock.reserved. That could happen when memblock
7024 * layout is manually configured via memmap=, or when the highest physical
7025 * address (max_pfn) does not end on a section boundary.
7027 static void __init init_unavailable_mem(void)
7029 phys_addr_t start, end;
7031 phys_addr_t next = 0;
7034 * Loop through unavailable ranges not covered by memblock.memory.
7037 for_each_mem_range(i, &memblock.memory, NULL,
7038 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
7040 pgcnt += init_unavailable_range(PFN_DOWN(next),
7046 * Early sections always have a fully populated memmap for the whole
7047 * section - see pfn_valid(). If the last section has holes at the
7048 * end and that section is marked "online", the memmap will be
7049 * considered initialized. Make sure that memmap has a well defined
7052 pgcnt += init_unavailable_range(PFN_DOWN(next),
7053 round_up(max_pfn, PAGES_PER_SECTION));
7056 * Struct pages that do not have backing memory. This could be because
7057 * firmware is using some of this memory, or for some other reasons.
7060 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
7063 static inline void __init init_unavailable_mem(void)
7066 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
7068 #if MAX_NUMNODES > 1
7070 * Figure out the number of possible node ids.
7072 void __init setup_nr_node_ids(void)
7074 unsigned int highest;
7076 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7077 nr_node_ids = highest + 1;
7082 * node_map_pfn_alignment - determine the maximum internode alignment
7084 * This function should be called after node map is populated and sorted.
7085 * It calculates the maximum power of two alignment which can distinguish
7088 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7089 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7090 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7091 * shifted, 1GiB is enough and this function will indicate so.
7093 * This is used to test whether pfn -> nid mapping of the chosen memory
7094 * model has fine enough granularity to avoid incorrect mapping for the
7095 * populated node map.
7097 * Return: the determined alignment in pfn's. 0 if there is no alignment
7098 * requirement (single node).
7100 unsigned long __init node_map_pfn_alignment(void)
7102 unsigned long accl_mask = 0, last_end = 0;
7103 unsigned long start, end, mask;
7104 int last_nid = NUMA_NO_NODE;
7107 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7108 if (!start || last_nid < 0 || last_nid == nid) {
7115 * Start with a mask granular enough to pin-point to the
7116 * start pfn and tick off bits one-by-one until it becomes
7117 * too coarse to separate the current node from the last.
7119 mask = ~((1 << __ffs(start)) - 1);
7120 while (mask && last_end <= (start & (mask << 1)))
7123 /* accumulate all internode masks */
7127 /* convert mask to number of pages */
7128 return ~accl_mask + 1;
7131 /* Find the lowest pfn for a node */
7132 static unsigned long __init find_min_pfn_for_node(int nid)
7134 unsigned long min_pfn = ULONG_MAX;
7135 unsigned long start_pfn;
7138 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
7139 min_pfn = min(min_pfn, start_pfn);
7141 if (min_pfn == ULONG_MAX) {
7142 pr_warn("Could not find start_pfn for node %d\n", nid);
7150 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7152 * Return: the minimum PFN based on information provided via
7153 * memblock_set_node().
7155 unsigned long __init find_min_pfn_with_active_regions(void)
7157 return find_min_pfn_for_node(MAX_NUMNODES);
7161 * early_calculate_totalpages()
7162 * Sum pages in active regions for movable zone.
7163 * Populate N_MEMORY for calculating usable_nodes.
7165 static unsigned long __init early_calculate_totalpages(void)
7167 unsigned long totalpages = 0;
7168 unsigned long start_pfn, end_pfn;
7171 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7172 unsigned long pages = end_pfn - start_pfn;
7174 totalpages += pages;
7176 node_set_state(nid, N_MEMORY);
7182 * Find the PFN the Movable zone begins in each node. Kernel memory
7183 * is spread evenly between nodes as long as the nodes have enough
7184 * memory. When they don't, some nodes will have more kernelcore than
7187 static void __init find_zone_movable_pfns_for_nodes(void)
7190 unsigned long usable_startpfn;
7191 unsigned long kernelcore_node, kernelcore_remaining;
7192 /* save the state before borrow the nodemask */
7193 nodemask_t saved_node_state = node_states[N_MEMORY];
7194 unsigned long totalpages = early_calculate_totalpages();
7195 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7196 struct memblock_region *r;
7198 /* Need to find movable_zone earlier when movable_node is specified. */
7199 find_usable_zone_for_movable();
7202 * If movable_node is specified, ignore kernelcore and movablecore
7205 if (movable_node_is_enabled()) {
7206 for_each_memblock(memory, r) {
7207 if (!memblock_is_hotpluggable(r))
7210 nid = memblock_get_region_node(r);
7212 usable_startpfn = PFN_DOWN(r->base);
7213 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7214 min(usable_startpfn, zone_movable_pfn[nid]) :
7222 * If kernelcore=mirror is specified, ignore movablecore option
7224 if (mirrored_kernelcore) {
7225 bool mem_below_4gb_not_mirrored = false;
7227 for_each_memblock(memory, r) {
7228 if (memblock_is_mirror(r))
7231 nid = memblock_get_region_node(r);
7233 usable_startpfn = memblock_region_memory_base_pfn(r);
7235 if (usable_startpfn < 0x100000) {
7236 mem_below_4gb_not_mirrored = true;
7240 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7241 min(usable_startpfn, zone_movable_pfn[nid]) :
7245 if (mem_below_4gb_not_mirrored)
7246 pr_warn("This configuration results in unmirrored kernel memory.");
7252 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7253 * amount of necessary memory.
7255 if (required_kernelcore_percent)
7256 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7258 if (required_movablecore_percent)
7259 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7263 * If movablecore= was specified, calculate what size of
7264 * kernelcore that corresponds so that memory usable for
7265 * any allocation type is evenly spread. If both kernelcore
7266 * and movablecore are specified, then the value of kernelcore
7267 * will be used for required_kernelcore if it's greater than
7268 * what movablecore would have allowed.
7270 if (required_movablecore) {
7271 unsigned long corepages;
7274 * Round-up so that ZONE_MOVABLE is at least as large as what
7275 * was requested by the user
7277 required_movablecore =
7278 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7279 required_movablecore = min(totalpages, required_movablecore);
7280 corepages = totalpages - required_movablecore;
7282 required_kernelcore = max(required_kernelcore, corepages);
7286 * If kernelcore was not specified or kernelcore size is larger
7287 * than totalpages, there is no ZONE_MOVABLE.
7289 if (!required_kernelcore || required_kernelcore >= totalpages)
7292 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7293 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7296 /* Spread kernelcore memory as evenly as possible throughout nodes */
7297 kernelcore_node = required_kernelcore / usable_nodes;
7298 for_each_node_state(nid, N_MEMORY) {
7299 unsigned long start_pfn, end_pfn;
7302 * Recalculate kernelcore_node if the division per node
7303 * now exceeds what is necessary to satisfy the requested
7304 * amount of memory for the kernel
7306 if (required_kernelcore < kernelcore_node)
7307 kernelcore_node = required_kernelcore / usable_nodes;
7310 * As the map is walked, we track how much memory is usable
7311 * by the kernel using kernelcore_remaining. When it is
7312 * 0, the rest of the node is usable by ZONE_MOVABLE
7314 kernelcore_remaining = kernelcore_node;
7316 /* Go through each range of PFNs within this node */
7317 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7318 unsigned long size_pages;
7320 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7321 if (start_pfn >= end_pfn)
7324 /* Account for what is only usable for kernelcore */
7325 if (start_pfn < usable_startpfn) {
7326 unsigned long kernel_pages;
7327 kernel_pages = min(end_pfn, usable_startpfn)
7330 kernelcore_remaining -= min(kernel_pages,
7331 kernelcore_remaining);
7332 required_kernelcore -= min(kernel_pages,
7333 required_kernelcore);
7335 /* Continue if range is now fully accounted */
7336 if (end_pfn <= usable_startpfn) {
7339 * Push zone_movable_pfn to the end so
7340 * that if we have to rebalance
7341 * kernelcore across nodes, we will
7342 * not double account here
7344 zone_movable_pfn[nid] = end_pfn;
7347 start_pfn = usable_startpfn;
7351 * The usable PFN range for ZONE_MOVABLE is from
7352 * start_pfn->end_pfn. Calculate size_pages as the
7353 * number of pages used as kernelcore
7355 size_pages = end_pfn - start_pfn;
7356 if (size_pages > kernelcore_remaining)
7357 size_pages = kernelcore_remaining;
7358 zone_movable_pfn[nid] = start_pfn + size_pages;
7361 * Some kernelcore has been met, update counts and
7362 * break if the kernelcore for this node has been
7365 required_kernelcore -= min(required_kernelcore,
7367 kernelcore_remaining -= size_pages;
7368 if (!kernelcore_remaining)
7374 * If there is still required_kernelcore, we do another pass with one
7375 * less node in the count. This will push zone_movable_pfn[nid] further
7376 * along on the nodes that still have memory until kernelcore is
7380 if (usable_nodes && required_kernelcore > usable_nodes)
7384 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7385 for (nid = 0; nid < MAX_NUMNODES; nid++)
7386 zone_movable_pfn[nid] =
7387 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7390 /* restore the node_state */
7391 node_states[N_MEMORY] = saved_node_state;
7394 /* Any regular or high memory on that node ? */
7395 static void check_for_memory(pg_data_t *pgdat, int nid)
7397 enum zone_type zone_type;
7399 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7400 struct zone *zone = &pgdat->node_zones[zone_type];
7401 if (populated_zone(zone)) {
7402 if (IS_ENABLED(CONFIG_HIGHMEM))
7403 node_set_state(nid, N_HIGH_MEMORY);
7404 if (zone_type <= ZONE_NORMAL)
7405 node_set_state(nid, N_NORMAL_MEMORY);
7412 * free_area_init - Initialise all pg_data_t and zone data
7413 * @max_zone_pfn: an array of max PFNs for each zone
7415 * This will call free_area_init_node() for each active node in the system.
7416 * Using the page ranges provided by memblock_set_node(), the size of each
7417 * zone in each node and their holes is calculated. If the maximum PFN
7418 * between two adjacent zones match, it is assumed that the zone is empty.
7419 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7420 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7421 * starts where the previous one ended. For example, ZONE_DMA32 starts
7422 * at arch_max_dma_pfn.
7424 void __init free_area_init(unsigned long *max_zone_pfn)
7426 unsigned long start_pfn, end_pfn;
7429 /* Record where the zone boundaries are */
7430 memset(arch_zone_lowest_possible_pfn, 0,
7431 sizeof(arch_zone_lowest_possible_pfn));
7432 memset(arch_zone_highest_possible_pfn, 0,
7433 sizeof(arch_zone_highest_possible_pfn));
7435 start_pfn = find_min_pfn_with_active_regions();
7437 for (i = 0; i < MAX_NR_ZONES; i++) {
7438 if (i == ZONE_MOVABLE)
7441 end_pfn = max(max_zone_pfn[i], start_pfn);
7442 arch_zone_lowest_possible_pfn[i] = start_pfn;
7443 arch_zone_highest_possible_pfn[i] = end_pfn;
7445 start_pfn = end_pfn;
7448 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7449 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7450 find_zone_movable_pfns_for_nodes();
7452 /* Print out the zone ranges */
7453 pr_info("Zone ranges:\n");
7454 for (i = 0; i < MAX_NR_ZONES; i++) {
7455 if (i == ZONE_MOVABLE)
7457 pr_info(" %-8s ", zone_names[i]);
7458 if (arch_zone_lowest_possible_pfn[i] ==
7459 arch_zone_highest_possible_pfn[i])
7462 pr_cont("[mem %#018Lx-%#018Lx]\n",
7463 (u64)arch_zone_lowest_possible_pfn[i]
7465 ((u64)arch_zone_highest_possible_pfn[i]
7466 << PAGE_SHIFT) - 1);
7469 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7470 pr_info("Movable zone start for each node\n");
7471 for (i = 0; i < MAX_NUMNODES; i++) {
7472 if (zone_movable_pfn[i])
7473 pr_info(" Node %d: %#018Lx\n", i,
7474 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7478 * Print out the early node map, and initialize the
7479 * subsection-map relative to active online memory ranges to
7480 * enable future "sub-section" extensions of the memory map.
7482 pr_info("Early memory node ranges\n");
7483 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7484 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7485 (u64)start_pfn << PAGE_SHIFT,
7486 ((u64)end_pfn << PAGE_SHIFT) - 1);
7487 subsection_map_init(start_pfn, end_pfn - start_pfn);
7490 /* Initialise every node */
7491 mminit_verify_pageflags_layout();
7492 setup_nr_node_ids();
7493 init_unavailable_mem();
7494 for_each_online_node(nid) {
7495 pg_data_t *pgdat = NODE_DATA(nid);
7496 __free_area_init_node(nid, NULL,
7497 find_min_pfn_for_node(nid), NULL, false);
7499 /* Any memory on that node */
7500 if (pgdat->node_present_pages)
7501 node_set_state(nid, N_MEMORY);
7502 check_for_memory(pgdat, nid);
7506 static int __init cmdline_parse_core(char *p, unsigned long *core,
7507 unsigned long *percent)
7509 unsigned long long coremem;
7515 /* Value may be a percentage of total memory, otherwise bytes */
7516 coremem = simple_strtoull(p, &endptr, 0);
7517 if (*endptr == '%') {
7518 /* Paranoid check for percent values greater than 100 */
7519 WARN_ON(coremem > 100);
7523 coremem = memparse(p, &p);
7524 /* Paranoid check that UL is enough for the coremem value */
7525 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7527 *core = coremem >> PAGE_SHIFT;
7534 * kernelcore=size sets the amount of memory for use for allocations that
7535 * cannot be reclaimed or migrated.
7537 static int __init cmdline_parse_kernelcore(char *p)
7539 /* parse kernelcore=mirror */
7540 if (parse_option_str(p, "mirror")) {
7541 mirrored_kernelcore = true;
7545 return cmdline_parse_core(p, &required_kernelcore,
7546 &required_kernelcore_percent);
7550 * movablecore=size sets the amount of memory for use for allocations that
7551 * can be reclaimed or migrated.
7553 static int __init cmdline_parse_movablecore(char *p)
7555 return cmdline_parse_core(p, &required_movablecore,
7556 &required_movablecore_percent);
7559 early_param("kernelcore", cmdline_parse_kernelcore);
7560 early_param("movablecore", cmdline_parse_movablecore);
7562 void adjust_managed_page_count(struct page *page, long count)
7564 atomic_long_add(count, &page_zone(page)->managed_pages);
7565 totalram_pages_add(count);
7566 #ifdef CONFIG_HIGHMEM
7567 if (PageHighMem(page))
7568 totalhigh_pages_add(count);
7571 EXPORT_SYMBOL(adjust_managed_page_count);
7573 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7576 unsigned long pages = 0;
7578 start = (void *)PAGE_ALIGN((unsigned long)start);
7579 end = (void *)((unsigned long)end & PAGE_MASK);
7580 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7581 struct page *page = virt_to_page(pos);
7582 void *direct_map_addr;
7585 * 'direct_map_addr' might be different from 'pos'
7586 * because some architectures' virt_to_page()
7587 * work with aliases. Getting the direct map
7588 * address ensures that we get a _writeable_
7589 * alias for the memset().
7591 direct_map_addr = page_address(page);
7592 if ((unsigned int)poison <= 0xFF)
7593 memset(direct_map_addr, poison, PAGE_SIZE);
7595 free_reserved_page(page);
7599 pr_info("Freeing %s memory: %ldK\n",
7600 s, pages << (PAGE_SHIFT - 10));
7605 #ifdef CONFIG_HIGHMEM
7606 void free_highmem_page(struct page *page)
7608 __free_reserved_page(page);
7609 totalram_pages_inc();
7610 atomic_long_inc(&page_zone(page)->managed_pages);
7611 totalhigh_pages_inc();
7616 void __init mem_init_print_info(const char *str)
7618 unsigned long physpages, codesize, datasize, rosize, bss_size;
7619 unsigned long init_code_size, init_data_size;
7621 physpages = get_num_physpages();
7622 codesize = _etext - _stext;
7623 datasize = _edata - _sdata;
7624 rosize = __end_rodata - __start_rodata;
7625 bss_size = __bss_stop - __bss_start;
7626 init_data_size = __init_end - __init_begin;
7627 init_code_size = _einittext - _sinittext;
7630 * Detect special cases and adjust section sizes accordingly:
7631 * 1) .init.* may be embedded into .data sections
7632 * 2) .init.text.* may be out of [__init_begin, __init_end],
7633 * please refer to arch/tile/kernel/vmlinux.lds.S.
7634 * 3) .rodata.* may be embedded into .text or .data sections.
7636 #define adj_init_size(start, end, size, pos, adj) \
7638 if (start <= pos && pos < end && size > adj) \
7642 adj_init_size(__init_begin, __init_end, init_data_size,
7643 _sinittext, init_code_size);
7644 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7645 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7646 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7647 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7649 #undef adj_init_size
7651 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7652 #ifdef CONFIG_HIGHMEM
7656 nr_free_pages() << (PAGE_SHIFT - 10),
7657 physpages << (PAGE_SHIFT - 10),
7658 codesize >> 10, datasize >> 10, rosize >> 10,
7659 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7660 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7661 totalcma_pages << (PAGE_SHIFT - 10),
7662 #ifdef CONFIG_HIGHMEM
7663 totalhigh_pages() << (PAGE_SHIFT - 10),
7665 str ? ", " : "", str ? str : "");
7669 * set_dma_reserve - set the specified number of pages reserved in the first zone
7670 * @new_dma_reserve: The number of pages to mark reserved
7672 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7673 * In the DMA zone, a significant percentage may be consumed by kernel image
7674 * and other unfreeable allocations which can skew the watermarks badly. This
7675 * function may optionally be used to account for unfreeable pages in the
7676 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7677 * smaller per-cpu batchsize.
7679 void __init set_dma_reserve(unsigned long new_dma_reserve)
7681 dma_reserve = new_dma_reserve;
7684 static int page_alloc_cpu_dead(unsigned int cpu)
7687 lru_add_drain_cpu(cpu);
7691 * Spill the event counters of the dead processor
7692 * into the current processors event counters.
7693 * This artificially elevates the count of the current
7696 vm_events_fold_cpu(cpu);
7699 * Zero the differential counters of the dead processor
7700 * so that the vm statistics are consistent.
7702 * This is only okay since the processor is dead and cannot
7703 * race with what we are doing.
7705 cpu_vm_stats_fold(cpu);
7710 int hashdist = HASHDIST_DEFAULT;
7712 static int __init set_hashdist(char *str)
7716 hashdist = simple_strtoul(str, &str, 0);
7719 __setup("hashdist=", set_hashdist);
7722 void __init page_alloc_init(void)
7727 if (num_node_state(N_MEMORY) == 1)
7731 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7732 "mm/page_alloc:dead", NULL,
7733 page_alloc_cpu_dead);
7738 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7739 * or min_free_kbytes changes.
7741 static void calculate_totalreserve_pages(void)
7743 struct pglist_data *pgdat;
7744 unsigned long reserve_pages = 0;
7745 enum zone_type i, j;
7747 for_each_online_pgdat(pgdat) {
7749 pgdat->totalreserve_pages = 0;
7751 for (i = 0; i < MAX_NR_ZONES; i++) {
7752 struct zone *zone = pgdat->node_zones + i;
7754 unsigned long managed_pages = zone_managed_pages(zone);
7756 /* Find valid and maximum lowmem_reserve in the zone */
7757 for (j = i; j < MAX_NR_ZONES; j++) {
7758 if (zone->lowmem_reserve[j] > max)
7759 max = zone->lowmem_reserve[j];
7762 /* we treat the high watermark as reserved pages. */
7763 max += high_wmark_pages(zone);
7765 if (max > managed_pages)
7766 max = managed_pages;
7768 pgdat->totalreserve_pages += max;
7770 reserve_pages += max;
7773 totalreserve_pages = reserve_pages;
7777 * setup_per_zone_lowmem_reserve - called whenever
7778 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7779 * has a correct pages reserved value, so an adequate number of
7780 * pages are left in the zone after a successful __alloc_pages().
7782 static void setup_per_zone_lowmem_reserve(void)
7784 struct pglist_data *pgdat;
7785 enum zone_type j, idx;
7787 for_each_online_pgdat(pgdat) {
7788 for (j = 0; j < MAX_NR_ZONES; j++) {
7789 struct zone *zone = pgdat->node_zones + j;
7790 unsigned long managed_pages = zone_managed_pages(zone);
7792 zone->lowmem_reserve[j] = 0;
7796 struct zone *lower_zone;
7799 lower_zone = pgdat->node_zones + idx;
7801 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7802 sysctl_lowmem_reserve_ratio[idx] = 0;
7803 lower_zone->lowmem_reserve[j] = 0;
7805 lower_zone->lowmem_reserve[j] =
7806 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7808 managed_pages += zone_managed_pages(lower_zone);
7813 /* update totalreserve_pages */
7814 calculate_totalreserve_pages();
7817 static void __setup_per_zone_wmarks(void)
7819 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7820 unsigned long lowmem_pages = 0;
7822 unsigned long flags;
7824 /* Calculate total number of !ZONE_HIGHMEM pages */
7825 for_each_zone(zone) {
7826 if (!is_highmem(zone))
7827 lowmem_pages += zone_managed_pages(zone);
7830 for_each_zone(zone) {
7833 spin_lock_irqsave(&zone->lock, flags);
7834 tmp = (u64)pages_min * zone_managed_pages(zone);
7835 do_div(tmp, lowmem_pages);
7836 if (is_highmem(zone)) {
7838 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7839 * need highmem pages, so cap pages_min to a small
7842 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7843 * deltas control async page reclaim, and so should
7844 * not be capped for highmem.
7846 unsigned long min_pages;
7848 min_pages = zone_managed_pages(zone) / 1024;
7849 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7850 zone->_watermark[WMARK_MIN] = min_pages;
7853 * If it's a lowmem zone, reserve a number of pages
7854 * proportionate to the zone's size.
7856 zone->_watermark[WMARK_MIN] = tmp;
7860 * Set the kswapd watermarks distance according to the
7861 * scale factor in proportion to available memory, but
7862 * ensure a minimum size on small systems.
7864 tmp = max_t(u64, tmp >> 2,
7865 mult_frac(zone_managed_pages(zone),
7866 watermark_scale_factor, 10000));
7868 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7869 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7870 zone->watermark_boost = 0;
7872 spin_unlock_irqrestore(&zone->lock, flags);
7875 /* update totalreserve_pages */
7876 calculate_totalreserve_pages();
7880 * setup_per_zone_wmarks - called when min_free_kbytes changes
7881 * or when memory is hot-{added|removed}
7883 * Ensures that the watermark[min,low,high] values for each zone are set
7884 * correctly with respect to min_free_kbytes.
7886 void setup_per_zone_wmarks(void)
7888 static DEFINE_SPINLOCK(lock);
7891 __setup_per_zone_wmarks();
7896 * Initialise min_free_kbytes.
7898 * For small machines we want it small (128k min). For large machines
7899 * we want it large (64MB max). But it is not linear, because network
7900 * bandwidth does not increase linearly with machine size. We use
7902 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7903 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7919 int __meminit init_per_zone_wmark_min(void)
7921 unsigned long lowmem_kbytes;
7922 int new_min_free_kbytes;
7924 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7925 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7927 if (new_min_free_kbytes > user_min_free_kbytes) {
7928 min_free_kbytes = new_min_free_kbytes;
7929 if (min_free_kbytes < 128)
7930 min_free_kbytes = 128;
7931 if (min_free_kbytes > 262144)
7932 min_free_kbytes = 262144;
7934 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7935 new_min_free_kbytes, user_min_free_kbytes);
7937 setup_per_zone_wmarks();
7938 refresh_zone_stat_thresholds();
7939 setup_per_zone_lowmem_reserve();
7942 setup_min_unmapped_ratio();
7943 setup_min_slab_ratio();
7948 core_initcall(init_per_zone_wmark_min)
7951 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7952 * that we can call two helper functions whenever min_free_kbytes
7955 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7956 void __user *buffer, size_t *length, loff_t *ppos)
7960 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7965 user_min_free_kbytes = min_free_kbytes;
7966 setup_per_zone_wmarks();
7971 int watermark_boost_factor_sysctl_handler(struct ctl_table *table, int write,
7972 void __user *buffer, size_t *length, loff_t *ppos)
7976 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7983 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7984 void __user *buffer, size_t *length, loff_t *ppos)
7988 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7993 setup_per_zone_wmarks();
7999 static void setup_min_unmapped_ratio(void)
8004 for_each_online_pgdat(pgdat)
8005 pgdat->min_unmapped_pages = 0;
8008 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8009 sysctl_min_unmapped_ratio) / 100;
8013 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8014 void __user *buffer, size_t *length, loff_t *ppos)
8018 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8022 setup_min_unmapped_ratio();
8027 static void setup_min_slab_ratio(void)
8032 for_each_online_pgdat(pgdat)
8033 pgdat->min_slab_pages = 0;
8036 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8037 sysctl_min_slab_ratio) / 100;
8040 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8041 void __user *buffer, size_t *length, loff_t *ppos)
8045 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8049 setup_min_slab_ratio();
8056 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8057 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8058 * whenever sysctl_lowmem_reserve_ratio changes.
8060 * The reserve ratio obviously has absolutely no relation with the
8061 * minimum watermarks. The lowmem reserve ratio can only make sense
8062 * if in function of the boot time zone sizes.
8064 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8065 void __user *buffer, size_t *length, loff_t *ppos)
8067 proc_dointvec_minmax(table, write, buffer, length, ppos);
8068 setup_per_zone_lowmem_reserve();
8072 static void __zone_pcp_update(struct zone *zone)
8076 for_each_possible_cpu(cpu)
8077 pageset_set_high_and_batch(zone,
8078 per_cpu_ptr(zone->pageset, cpu));
8082 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8083 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8084 * pagelist can have before it gets flushed back to buddy allocator.
8086 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8087 void __user *buffer, size_t *length, loff_t *ppos)
8090 int old_percpu_pagelist_fraction;
8093 mutex_lock(&pcp_batch_high_lock);
8094 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8096 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8097 if (!write || ret < 0)
8100 /* Sanity checking to avoid pcp imbalance */
8101 if (percpu_pagelist_fraction &&
8102 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8103 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8109 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8112 for_each_populated_zone(zone)
8113 __zone_pcp_update(zone);
8115 mutex_unlock(&pcp_batch_high_lock);
8119 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8121 * Returns the number of pages that arch has reserved but
8122 * is not known to alloc_large_system_hash().
8124 static unsigned long __init arch_reserved_kernel_pages(void)
8131 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8132 * machines. As memory size is increased the scale is also increased but at
8133 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8134 * quadruples the scale is increased by one, which means the size of hash table
8135 * only doubles, instead of quadrupling as well.
8136 * Because 32-bit systems cannot have large physical memory, where this scaling
8137 * makes sense, it is disabled on such platforms.
8139 #if __BITS_PER_LONG > 32
8140 #define ADAPT_SCALE_BASE (64ul << 30)
8141 #define ADAPT_SCALE_SHIFT 2
8142 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8146 * allocate a large system hash table from bootmem
8147 * - it is assumed that the hash table must contain an exact power-of-2
8148 * quantity of entries
8149 * - limit is the number of hash buckets, not the total allocation size
8151 void *__init alloc_large_system_hash(const char *tablename,
8152 unsigned long bucketsize,
8153 unsigned long numentries,
8156 unsigned int *_hash_shift,
8157 unsigned int *_hash_mask,
8158 unsigned long low_limit,
8159 unsigned long high_limit)
8161 unsigned long long max = high_limit;
8162 unsigned long log2qty, size;
8167 /* allow the kernel cmdline to have a say */
8169 /* round applicable memory size up to nearest megabyte */
8170 numentries = nr_kernel_pages;
8171 numentries -= arch_reserved_kernel_pages();
8173 /* It isn't necessary when PAGE_SIZE >= 1MB */
8174 if (PAGE_SHIFT < 20)
8175 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8177 #if __BITS_PER_LONG > 32
8179 unsigned long adapt;
8181 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8182 adapt <<= ADAPT_SCALE_SHIFT)
8187 /* limit to 1 bucket per 2^scale bytes of low memory */
8188 if (scale > PAGE_SHIFT)
8189 numentries >>= (scale - PAGE_SHIFT);
8191 numentries <<= (PAGE_SHIFT - scale);
8193 /* Make sure we've got at least a 0-order allocation.. */
8194 if (unlikely(flags & HASH_SMALL)) {
8195 /* Makes no sense without HASH_EARLY */
8196 WARN_ON(!(flags & HASH_EARLY));
8197 if (!(numentries >> *_hash_shift)) {
8198 numentries = 1UL << *_hash_shift;
8199 BUG_ON(!numentries);
8201 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8202 numentries = PAGE_SIZE / bucketsize;
8204 numentries = roundup_pow_of_two(numentries);
8206 /* limit allocation size to 1/16 total memory by default */
8208 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8209 do_div(max, bucketsize);
8211 max = min(max, 0x80000000ULL);
8213 if (numentries < low_limit)
8214 numentries = low_limit;
8215 if (numentries > max)
8218 log2qty = ilog2(numentries);
8220 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8223 size = bucketsize << log2qty;
8224 if (flags & HASH_EARLY) {
8225 if (flags & HASH_ZERO)
8226 table = memblock_alloc(size, SMP_CACHE_BYTES);
8228 table = memblock_alloc_raw(size,
8230 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8231 table = __vmalloc(size, gfp_flags);
8235 * If bucketsize is not a power-of-two, we may free
8236 * some pages at the end of hash table which
8237 * alloc_pages_exact() automatically does
8239 table = alloc_pages_exact(size, gfp_flags);
8240 kmemleak_alloc(table, size, 1, gfp_flags);
8242 } while (!table && size > PAGE_SIZE && --log2qty);
8245 panic("Failed to allocate %s hash table\n", tablename);
8247 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8248 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8249 virt ? "vmalloc" : "linear");
8252 *_hash_shift = log2qty;
8254 *_hash_mask = (1 << log2qty) - 1;
8260 * This function checks whether pageblock includes unmovable pages or not.
8262 * PageLRU check without isolation or lru_lock could race so that
8263 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8264 * check without lock_page also may miss some movable non-lru pages at
8265 * race condition. So you can't expect this function should be exact.
8267 * Returns a page without holding a reference. If the caller wants to
8268 * dereference that page (e.g., dumping), it has to make sure that that it
8269 * cannot get removed (e.g., via memory unplug) concurrently.
8272 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8273 int migratetype, int flags)
8275 unsigned long iter = 0;
8276 unsigned long pfn = page_to_pfn(page);
8279 * TODO we could make this much more efficient by not checking every
8280 * page in the range if we know all of them are in MOVABLE_ZONE and
8281 * that the movable zone guarantees that pages are migratable but
8282 * the later is not the case right now unfortunatelly. E.g. movablecore
8283 * can still lead to having bootmem allocations in zone_movable.
8286 if (is_migrate_cma_page(page)) {
8288 * CMA allocations (alloc_contig_range) really need to mark
8289 * isolate CMA pageblocks even when they are not movable in fact
8290 * so consider them movable here.
8292 if (is_migrate_cma(migratetype))
8298 for (; iter < pageblock_nr_pages; iter++) {
8299 if (!pfn_valid_within(pfn + iter))
8302 page = pfn_to_page(pfn + iter);
8304 if (PageReserved(page))
8308 * If the zone is movable and we have ruled out all reserved
8309 * pages then it should be reasonably safe to assume the rest
8312 if (zone_idx(zone) == ZONE_MOVABLE)
8316 * Hugepages are not in LRU lists, but they're movable.
8317 * THPs are on the LRU, but need to be counted as #small pages.
8318 * We need not scan over tail pages because we don't
8319 * handle each tail page individually in migration.
8321 if (PageHuge(page) || PageTransCompound(page)) {
8322 struct page *head = compound_head(page);
8323 unsigned int skip_pages;
8325 if (PageHuge(page)) {
8326 if (!hugepage_migration_supported(page_hstate(head)))
8328 } else if (!PageLRU(head) && !__PageMovable(head)) {
8332 skip_pages = compound_nr(head) - (page - head);
8333 iter += skip_pages - 1;
8338 * We can't use page_count without pin a page
8339 * because another CPU can free compound page.
8340 * This check already skips compound tails of THP
8341 * because their page->_refcount is zero at all time.
8343 if (!page_ref_count(page)) {
8344 if (PageBuddy(page))
8345 iter += (1 << page_order(page)) - 1;
8350 * The HWPoisoned page may be not in buddy system, and
8351 * page_count() is not 0.
8353 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8356 if (__PageMovable(page) || PageLRU(page))
8360 * If there are RECLAIMABLE pages, we need to check
8361 * it. But now, memory offline itself doesn't call
8362 * shrink_node_slabs() and it still to be fixed.
8365 * If the page is not RAM, page_count()should be 0.
8366 * we don't need more check. This is an _used_ not-movable page.
8368 * The problematic thing here is PG_reserved pages. PG_reserved
8369 * is set to both of a memory hole page and a _used_ kernel
8377 #ifdef CONFIG_CONTIG_ALLOC
8378 static unsigned long pfn_max_align_down(unsigned long pfn)
8380 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8381 pageblock_nr_pages) - 1);
8384 static unsigned long pfn_max_align_up(unsigned long pfn)
8386 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8387 pageblock_nr_pages));
8390 /* [start, end) must belong to a single zone. */
8391 static int __alloc_contig_migrate_range(struct compact_control *cc,
8392 unsigned long start, unsigned long end)
8394 /* This function is based on compact_zone() from compaction.c. */
8395 unsigned long nr_reclaimed;
8396 unsigned long pfn = start;
8397 unsigned int tries = 0;
8402 while (pfn < end || !list_empty(&cc->migratepages)) {
8403 if (fatal_signal_pending(current)) {
8408 if (list_empty(&cc->migratepages)) {
8409 cc->nr_migratepages = 0;
8410 pfn = isolate_migratepages_range(cc, pfn, end);
8416 } else if (++tries == 5) {
8417 ret = ret < 0 ? ret : -EBUSY;
8421 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8423 cc->nr_migratepages -= nr_reclaimed;
8425 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
8426 NULL, 0, cc->mode, MR_CONTIG_RANGE);
8429 putback_movable_pages(&cc->migratepages);
8436 * alloc_contig_range() -- tries to allocate given range of pages
8437 * @start: start PFN to allocate
8438 * @end: one-past-the-last PFN to allocate
8439 * @migratetype: migratetype of the underlaying pageblocks (either
8440 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8441 * in range must have the same migratetype and it must
8442 * be either of the two.
8443 * @gfp_mask: GFP mask to use during compaction
8445 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8446 * aligned. The PFN range must belong to a single zone.
8448 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8449 * pageblocks in the range. Once isolated, the pageblocks should not
8450 * be modified by others.
8452 * Return: zero on success or negative error code. On success all
8453 * pages which PFN is in [start, end) are allocated for the caller and
8454 * need to be freed with free_contig_range().
8456 int alloc_contig_range(unsigned long start, unsigned long end,
8457 unsigned migratetype, gfp_t gfp_mask)
8459 unsigned long outer_start, outer_end;
8463 struct compact_control cc = {
8464 .nr_migratepages = 0,
8466 .zone = page_zone(pfn_to_page(start)),
8467 .mode = MIGRATE_SYNC,
8468 .ignore_skip_hint = true,
8469 .no_set_skip_hint = true,
8470 .gfp_mask = current_gfp_context(gfp_mask),
8471 .alloc_contig = true,
8473 INIT_LIST_HEAD(&cc.migratepages);
8476 * What we do here is we mark all pageblocks in range as
8477 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8478 * have different sizes, and due to the way page allocator
8479 * work, we align the range to biggest of the two pages so
8480 * that page allocator won't try to merge buddies from
8481 * different pageblocks and change MIGRATE_ISOLATE to some
8482 * other migration type.
8484 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8485 * migrate the pages from an unaligned range (ie. pages that
8486 * we are interested in). This will put all the pages in
8487 * range back to page allocator as MIGRATE_ISOLATE.
8489 * When this is done, we take the pages in range from page
8490 * allocator removing them from the buddy system. This way
8491 * page allocator will never consider using them.
8493 * This lets us mark the pageblocks back as
8494 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8495 * aligned range but not in the unaligned, original range are
8496 * put back to page allocator so that buddy can use them.
8499 ret = start_isolate_page_range(pfn_max_align_down(start),
8500 pfn_max_align_up(end), migratetype, 0);
8505 * In case of -EBUSY, we'd like to know which page causes problem.
8506 * So, just fall through. test_pages_isolated() has a tracepoint
8507 * which will report the busy page.
8509 * It is possible that busy pages could become available before
8510 * the call to test_pages_isolated, and the range will actually be
8511 * allocated. So, if we fall through be sure to clear ret so that
8512 * -EBUSY is not accidentally used or returned to caller.
8514 ret = __alloc_contig_migrate_range(&cc, start, end);
8515 if (ret && ret != -EBUSY)
8520 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8521 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8522 * more, all pages in [start, end) are free in page allocator.
8523 * What we are going to do is to allocate all pages from
8524 * [start, end) (that is remove them from page allocator).
8526 * The only problem is that pages at the beginning and at the
8527 * end of interesting range may be not aligned with pages that
8528 * page allocator holds, ie. they can be part of higher order
8529 * pages. Because of this, we reserve the bigger range and
8530 * once this is done free the pages we are not interested in.
8532 * We don't have to hold zone->lock here because the pages are
8533 * isolated thus they won't get removed from buddy.
8536 lru_add_drain_all();
8539 outer_start = start;
8540 while (!PageBuddy(pfn_to_page(outer_start))) {
8541 if (++order >= MAX_ORDER) {
8542 outer_start = start;
8545 outer_start &= ~0UL << order;
8548 if (outer_start != start) {
8549 order = page_order(pfn_to_page(outer_start));
8552 * outer_start page could be small order buddy page and
8553 * it doesn't include start page. Adjust outer_start
8554 * in this case to report failed page properly
8555 * on tracepoint in test_pages_isolated()
8557 if (outer_start + (1UL << order) <= start)
8558 outer_start = start;
8561 /* Make sure the range is really isolated. */
8562 if (test_pages_isolated(outer_start, end, 0)) {
8563 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8564 __func__, outer_start, end);
8569 /* Grab isolated pages from freelists. */
8570 outer_end = isolate_freepages_range(&cc, outer_start, end);
8576 /* Free head and tail (if any) */
8577 if (start != outer_start)
8578 free_contig_range(outer_start, start - outer_start);
8579 if (end != outer_end)
8580 free_contig_range(end, outer_end - end);
8583 undo_isolate_page_range(pfn_max_align_down(start),
8584 pfn_max_align_up(end), migratetype);
8588 static int __alloc_contig_pages(unsigned long start_pfn,
8589 unsigned long nr_pages, gfp_t gfp_mask)
8591 unsigned long end_pfn = start_pfn + nr_pages;
8593 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8597 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8598 unsigned long nr_pages)
8600 unsigned long i, end_pfn = start_pfn + nr_pages;
8603 for (i = start_pfn; i < end_pfn; i++) {
8604 page = pfn_to_online_page(i);
8608 if (page_zone(page) != z)
8611 if (PageReserved(page))
8614 if (page_count(page) > 0)
8623 static bool zone_spans_last_pfn(const struct zone *zone,
8624 unsigned long start_pfn, unsigned long nr_pages)
8626 unsigned long last_pfn = start_pfn + nr_pages - 1;
8628 return zone_spans_pfn(zone, last_pfn);
8632 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8633 * @nr_pages: Number of contiguous pages to allocate
8634 * @gfp_mask: GFP mask to limit search and used during compaction
8636 * @nodemask: Mask for other possible nodes
8638 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8639 * on an applicable zonelist to find a contiguous pfn range which can then be
8640 * tried for allocation with alloc_contig_range(). This routine is intended
8641 * for allocation requests which can not be fulfilled with the buddy allocator.
8643 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8644 * power of two then the alignment is guaranteed to be to the given nr_pages
8645 * (e.g. 1GB request would be aligned to 1GB).
8647 * Allocated pages can be freed with free_contig_range() or by manually calling
8648 * __free_page() on each allocated page.
8650 * Return: pointer to contiguous pages on success, or NULL if not successful.
8652 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8653 int nid, nodemask_t *nodemask)
8655 unsigned long ret, pfn, flags;
8656 struct zonelist *zonelist;
8660 zonelist = node_zonelist(nid, gfp_mask);
8661 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8662 gfp_zone(gfp_mask), nodemask) {
8663 spin_lock_irqsave(&zone->lock, flags);
8665 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8666 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8667 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8669 * We release the zone lock here because
8670 * alloc_contig_range() will also lock the zone
8671 * at some point. If there's an allocation
8672 * spinning on this lock, it may win the race
8673 * and cause alloc_contig_range() to fail...
8675 spin_unlock_irqrestore(&zone->lock, flags);
8676 ret = __alloc_contig_pages(pfn, nr_pages,
8679 return pfn_to_page(pfn);
8680 spin_lock_irqsave(&zone->lock, flags);
8684 spin_unlock_irqrestore(&zone->lock, flags);
8688 #endif /* CONFIG_CONTIG_ALLOC */
8690 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8692 unsigned int count = 0;
8694 for (; nr_pages--; pfn++) {
8695 struct page *page = pfn_to_page(pfn);
8697 count += page_count(page) != 1;
8700 WARN(count != 0, "%d pages are still in use!\n", count);
8704 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8705 * page high values need to be recalulated.
8707 void __meminit zone_pcp_update(struct zone *zone)
8709 mutex_lock(&pcp_batch_high_lock);
8710 __zone_pcp_update(zone);
8711 mutex_unlock(&pcp_batch_high_lock);
8714 void zone_pcp_reset(struct zone *zone)
8716 unsigned long flags;
8718 struct per_cpu_pageset *pset;
8720 /* avoid races with drain_pages() */
8721 local_irq_save(flags);
8722 if (zone->pageset != &boot_pageset) {
8723 for_each_online_cpu(cpu) {
8724 pset = per_cpu_ptr(zone->pageset, cpu);
8725 drain_zonestat(zone, pset);
8727 free_percpu(zone->pageset);
8728 zone->pageset = &boot_pageset;
8730 local_irq_restore(flags);
8733 #ifdef CONFIG_MEMORY_HOTREMOVE
8735 * All pages in the range must be in a single zone and isolated
8736 * before calling this.
8739 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8745 unsigned long flags;
8746 unsigned long offlined_pages = 0;
8748 /* find the first valid pfn */
8749 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8753 return offlined_pages;
8755 offline_mem_sections(pfn, end_pfn);
8756 zone = page_zone(pfn_to_page(pfn));
8757 spin_lock_irqsave(&zone->lock, flags);
8759 while (pfn < end_pfn) {
8760 if (!pfn_valid(pfn)) {
8764 page = pfn_to_page(pfn);
8766 * The HWPoisoned page may be not in buddy system, and
8767 * page_count() is not 0.
8769 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8775 BUG_ON(page_count(page));
8776 BUG_ON(!PageBuddy(page));
8777 order = page_order(page);
8778 offlined_pages += 1 << order;
8779 del_page_from_free_list(page, zone, order);
8780 pfn += (1 << order);
8782 spin_unlock_irqrestore(&zone->lock, flags);
8784 return offlined_pages;
8788 bool is_free_buddy_page(struct page *page)
8790 struct zone *zone = page_zone(page);
8791 unsigned long pfn = page_to_pfn(page);
8792 unsigned long flags;
8795 spin_lock_irqsave(&zone->lock, flags);
8796 for (order = 0; order < MAX_ORDER; order++) {
8797 struct page *page_head = page - (pfn & ((1 << order) - 1));
8799 if (PageBuddy(page_head) && page_order(page_head) >= order)
8802 spin_unlock_irqrestore(&zone->lock, flags);
8804 return order < MAX_ORDER;
8807 #ifdef CONFIG_MEMORY_FAILURE
8809 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8810 * test is performed under the zone lock to prevent a race against page
8813 bool set_hwpoison_free_buddy_page(struct page *page)
8815 struct zone *zone = page_zone(page);
8816 unsigned long pfn = page_to_pfn(page);
8817 unsigned long flags;
8819 bool hwpoisoned = false;
8821 spin_lock_irqsave(&zone->lock, flags);
8822 for (order = 0; order < MAX_ORDER; order++) {
8823 struct page *page_head = page - (pfn & ((1 << order) - 1));
8825 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8826 if (!TestSetPageHWPoison(page))
8831 spin_unlock_irqrestore(&zone->lock, flags);