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>
71 #include <linux/padata.h>
72 #include <linux/khugepaged.h>
74 #include <asm/sections.h>
75 #include <asm/tlbflush.h>
76 #include <asm/div64.h>
79 #include "page_reporting.h"
81 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
82 static DEFINE_MUTEX(pcp_batch_high_lock);
83 #define MIN_PERCPU_PAGELIST_FRACTION (8)
85 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
86 DEFINE_PER_CPU(int, numa_node);
87 EXPORT_PER_CPU_SYMBOL(numa_node);
90 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
92 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
94 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
95 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
96 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
97 * defined in <linux/topology.h>.
99 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
100 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
103 /* work_structs for global per-cpu drains */
106 struct work_struct work;
108 static DEFINE_MUTEX(pcpu_drain_mutex);
109 static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
111 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
112 volatile unsigned long latent_entropy __latent_entropy;
113 EXPORT_SYMBOL(latent_entropy);
117 * Array of node states.
119 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
120 [N_POSSIBLE] = NODE_MASK_ALL,
121 [N_ONLINE] = { { [0] = 1UL } },
123 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
124 #ifdef CONFIG_HIGHMEM
125 [N_HIGH_MEMORY] = { { [0] = 1UL } },
127 [N_MEMORY] = { { [0] = 1UL } },
128 [N_CPU] = { { [0] = 1UL } },
131 EXPORT_SYMBOL(node_states);
133 atomic_long_t _totalram_pages __read_mostly;
134 EXPORT_SYMBOL(_totalram_pages);
135 unsigned long totalreserve_pages __read_mostly;
136 unsigned long totalcma_pages __read_mostly;
138 int percpu_pagelist_fraction;
139 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
140 #ifdef CONFIG_INIT_ON_ALLOC_DEFAULT_ON
141 DEFINE_STATIC_KEY_TRUE(init_on_alloc);
143 DEFINE_STATIC_KEY_FALSE(init_on_alloc);
145 EXPORT_SYMBOL(init_on_alloc);
147 #ifdef CONFIG_INIT_ON_FREE_DEFAULT_ON
148 DEFINE_STATIC_KEY_TRUE(init_on_free);
150 DEFINE_STATIC_KEY_FALSE(init_on_free);
152 EXPORT_SYMBOL(init_on_free);
154 static int __init early_init_on_alloc(char *buf)
159 ret = kstrtobool(buf, &bool_result);
162 if (bool_result && page_poisoning_enabled())
163 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_alloc\n");
165 static_branch_enable(&init_on_alloc);
167 static_branch_disable(&init_on_alloc);
170 early_param("init_on_alloc", early_init_on_alloc);
172 static int __init early_init_on_free(char *buf)
177 ret = kstrtobool(buf, &bool_result);
180 if (bool_result && page_poisoning_enabled())
181 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_free\n");
183 static_branch_enable(&init_on_free);
185 static_branch_disable(&init_on_free);
188 early_param("init_on_free", early_init_on_free);
191 * A cached value of the page's pageblock's migratetype, used when the page is
192 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
193 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
194 * Also the migratetype set in the page does not necessarily match the pcplist
195 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
196 * other index - this ensures that it will be put on the correct CMA freelist.
198 static inline int get_pcppage_migratetype(struct page *page)
203 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
205 page->index = migratetype;
208 #ifdef CONFIG_PM_SLEEP
210 * The following functions are used by the suspend/hibernate code to temporarily
211 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
212 * while devices are suspended. To avoid races with the suspend/hibernate code,
213 * they should always be called with system_transition_mutex held
214 * (gfp_allowed_mask also should only be modified with system_transition_mutex
215 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
216 * with that modification).
219 static gfp_t saved_gfp_mask;
221 void pm_restore_gfp_mask(void)
223 WARN_ON(!mutex_is_locked(&system_transition_mutex));
224 if (saved_gfp_mask) {
225 gfp_allowed_mask = saved_gfp_mask;
230 void pm_restrict_gfp_mask(void)
232 WARN_ON(!mutex_is_locked(&system_transition_mutex));
233 WARN_ON(saved_gfp_mask);
234 saved_gfp_mask = gfp_allowed_mask;
235 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
238 bool pm_suspended_storage(void)
240 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
244 #endif /* CONFIG_PM_SLEEP */
246 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
247 unsigned int pageblock_order __read_mostly;
250 static void __free_pages_ok(struct page *page, unsigned int order);
253 * results with 256, 32 in the lowmem_reserve sysctl:
254 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
255 * 1G machine -> (16M dma, 784M normal, 224M high)
256 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
257 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
258 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
260 * TBD: should special case ZONE_DMA32 machines here - in those we normally
261 * don't need any ZONE_NORMAL reservation
263 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
264 #ifdef CONFIG_ZONE_DMA
267 #ifdef CONFIG_ZONE_DMA32
271 #ifdef CONFIG_HIGHMEM
277 static char * const zone_names[MAX_NR_ZONES] = {
278 #ifdef CONFIG_ZONE_DMA
281 #ifdef CONFIG_ZONE_DMA32
285 #ifdef CONFIG_HIGHMEM
289 #ifdef CONFIG_ZONE_DEVICE
294 const char * const migratetype_names[MIGRATE_TYPES] = {
302 #ifdef CONFIG_MEMORY_ISOLATION
307 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
308 [NULL_COMPOUND_DTOR] = NULL,
309 [COMPOUND_PAGE_DTOR] = free_compound_page,
310 #ifdef CONFIG_HUGETLB_PAGE
311 [HUGETLB_PAGE_DTOR] = free_huge_page,
313 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
314 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
318 int min_free_kbytes = 1024;
319 int user_min_free_kbytes = -1;
320 #ifdef CONFIG_DISCONTIGMEM
322 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
323 * are not on separate NUMA nodes. Functionally this works but with
324 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
325 * quite small. By default, do not boost watermarks on discontigmem as in
326 * many cases very high-order allocations like THP are likely to be
327 * unsupported and the premature reclaim offsets the advantage of long-term
328 * fragmentation avoidance.
330 int watermark_boost_factor __read_mostly;
332 int watermark_boost_factor __read_mostly = 15000;
334 int watermark_scale_factor = 10;
336 static unsigned long nr_kernel_pages __initdata;
337 static unsigned long nr_all_pages __initdata;
338 static unsigned long dma_reserve __initdata;
340 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
341 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
342 static unsigned long required_kernelcore __initdata;
343 static unsigned long required_kernelcore_percent __initdata;
344 static unsigned long required_movablecore __initdata;
345 static unsigned long required_movablecore_percent __initdata;
346 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
347 static bool mirrored_kernelcore __meminitdata;
349 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
351 EXPORT_SYMBOL(movable_zone);
354 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
355 unsigned int nr_online_nodes __read_mostly = 1;
356 EXPORT_SYMBOL(nr_node_ids);
357 EXPORT_SYMBOL(nr_online_nodes);
360 int page_group_by_mobility_disabled __read_mostly;
362 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
364 * During boot we initialize deferred pages on-demand, as needed, but once
365 * page_alloc_init_late() has finished, the deferred pages are all initialized,
366 * and we can permanently disable that path.
368 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
371 * Calling kasan_free_pages() only after deferred memory initialization
372 * has completed. Poisoning pages during deferred memory init will greatly
373 * lengthen the process and cause problem in large memory systems as the
374 * deferred pages initialization is done with interrupt disabled.
376 * Assuming that there will be no reference to those newly initialized
377 * pages before they are ever allocated, this should have no effect on
378 * KASAN memory tracking as the poison will be properly inserted at page
379 * allocation time. The only corner case is when pages are allocated by
380 * on-demand allocation and then freed again before the deferred pages
381 * initialization is done, but this is not likely to happen.
383 static inline void kasan_free_nondeferred_pages(struct page *page, int order)
385 if (!static_branch_unlikely(&deferred_pages))
386 kasan_free_pages(page, order);
389 /* Returns true if the struct page for the pfn is uninitialised */
390 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
392 int nid = early_pfn_to_nid(pfn);
394 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
401 * Returns true when the remaining initialisation should be deferred until
402 * later in the boot cycle when it can be parallelised.
404 static bool __meminit
405 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
407 static unsigned long prev_end_pfn, nr_initialised;
410 * prev_end_pfn static that contains the end of previous zone
411 * No need to protect because called very early in boot before smp_init.
413 if (prev_end_pfn != end_pfn) {
414 prev_end_pfn = end_pfn;
418 /* Always populate low zones for address-constrained allocations */
419 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
423 * We start only with one section of pages, more pages are added as
424 * needed until the rest of deferred pages are initialized.
427 if ((nr_initialised > PAGES_PER_SECTION) &&
428 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
429 NODE_DATA(nid)->first_deferred_pfn = pfn;
435 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
437 static inline bool early_page_uninitialised(unsigned long pfn)
442 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
448 /* Return a pointer to the bitmap storing bits affecting a block of pages */
449 static inline unsigned long *get_pageblock_bitmap(struct page *page,
452 #ifdef CONFIG_SPARSEMEM
453 return section_to_usemap(__pfn_to_section(pfn));
455 return page_zone(page)->pageblock_flags;
456 #endif /* CONFIG_SPARSEMEM */
459 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
461 #ifdef CONFIG_SPARSEMEM
462 pfn &= (PAGES_PER_SECTION-1);
464 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
465 #endif /* CONFIG_SPARSEMEM */
466 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
470 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
471 * @page: The page within the block of interest
472 * @pfn: The target page frame number
473 * @mask: mask of bits that the caller is interested in
475 * Return: pageblock_bits flags
477 static __always_inline
478 unsigned long __get_pfnblock_flags_mask(struct page *page,
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 return (word >> bitidx) & mask;
495 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
498 return __get_pfnblock_flags_mask(page, pfn, mask);
501 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
503 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
507 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
508 * @page: The page within the block of interest
509 * @flags: The flags to set
510 * @pfn: The target page frame number
511 * @mask: mask of bits that the caller is interested in
513 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
517 unsigned long *bitmap;
518 unsigned long bitidx, word_bitidx;
519 unsigned long old_word, word;
521 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
522 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
524 bitmap = get_pageblock_bitmap(page, pfn);
525 bitidx = pfn_to_bitidx(page, pfn);
526 word_bitidx = bitidx / BITS_PER_LONG;
527 bitidx &= (BITS_PER_LONG-1);
529 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
534 word = READ_ONCE(bitmap[word_bitidx]);
536 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
537 if (word == old_word)
543 void set_pageblock_migratetype(struct page *page, int migratetype)
545 if (unlikely(page_group_by_mobility_disabled &&
546 migratetype < MIGRATE_PCPTYPES))
547 migratetype = MIGRATE_UNMOVABLE;
549 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
550 page_to_pfn(page), MIGRATETYPE_MASK);
553 #ifdef CONFIG_DEBUG_VM
554 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
558 unsigned long pfn = page_to_pfn(page);
559 unsigned long sp, start_pfn;
562 seq = zone_span_seqbegin(zone);
563 start_pfn = zone->zone_start_pfn;
564 sp = zone->spanned_pages;
565 if (!zone_spans_pfn(zone, pfn))
567 } while (zone_span_seqretry(zone, seq));
570 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
571 pfn, zone_to_nid(zone), zone->name,
572 start_pfn, start_pfn + sp);
577 static int page_is_consistent(struct zone *zone, struct page *page)
579 if (!pfn_valid_within(page_to_pfn(page)))
581 if (zone != page_zone(page))
587 * Temporary debugging check for pages not lying within a given zone.
589 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
591 if (page_outside_zone_boundaries(zone, page))
593 if (!page_is_consistent(zone, page))
599 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
605 static void bad_page(struct page *page, const char *reason)
607 static unsigned long resume;
608 static unsigned long nr_shown;
609 static unsigned long nr_unshown;
612 * Allow a burst of 60 reports, then keep quiet for that minute;
613 * or allow a steady drip of one report per second.
615 if (nr_shown == 60) {
616 if (time_before(jiffies, resume)) {
622 "BUG: Bad page state: %lu messages suppressed\n",
629 resume = jiffies + 60 * HZ;
631 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
632 current->comm, page_to_pfn(page));
633 __dump_page(page, reason);
634 dump_page_owner(page);
639 /* Leave bad fields for debug, except PageBuddy could make trouble */
640 page_mapcount_reset(page); /* remove PageBuddy */
641 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
645 * Higher-order pages are called "compound pages". They are structured thusly:
647 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
649 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
650 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
652 * The first tail page's ->compound_dtor holds the offset in array of compound
653 * page destructors. See compound_page_dtors.
655 * The first tail page's ->compound_order holds the order of allocation.
656 * This usage means that zero-order pages may not be compound.
659 void free_compound_page(struct page *page)
661 mem_cgroup_uncharge(page);
662 __free_pages_ok(page, compound_order(page));
665 void prep_compound_page(struct page *page, unsigned int order)
668 int nr_pages = 1 << order;
671 for (i = 1; i < nr_pages; i++) {
672 struct page *p = page + i;
673 set_page_count(p, 0);
674 p->mapping = TAIL_MAPPING;
675 set_compound_head(p, page);
678 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
679 set_compound_order(page, order);
680 atomic_set(compound_mapcount_ptr(page), -1);
681 if (hpage_pincount_available(page))
682 atomic_set(compound_pincount_ptr(page), 0);
685 #ifdef CONFIG_DEBUG_PAGEALLOC
686 unsigned int _debug_guardpage_minorder;
688 bool _debug_pagealloc_enabled_early __read_mostly
689 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
690 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
691 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
692 EXPORT_SYMBOL(_debug_pagealloc_enabled);
694 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
696 static int __init early_debug_pagealloc(char *buf)
698 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
700 early_param("debug_pagealloc", early_debug_pagealloc);
702 void init_debug_pagealloc(void)
704 if (!debug_pagealloc_enabled())
707 static_branch_enable(&_debug_pagealloc_enabled);
709 if (!debug_guardpage_minorder())
712 static_branch_enable(&_debug_guardpage_enabled);
715 static int __init debug_guardpage_minorder_setup(char *buf)
719 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
720 pr_err("Bad debug_guardpage_minorder value\n");
723 _debug_guardpage_minorder = res;
724 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
727 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
729 static inline bool set_page_guard(struct zone *zone, struct page *page,
730 unsigned int order, int migratetype)
732 if (!debug_guardpage_enabled())
735 if (order >= debug_guardpage_minorder())
738 __SetPageGuard(page);
739 INIT_LIST_HEAD(&page->lru);
740 set_page_private(page, order);
741 /* Guard pages are not available for any usage */
742 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
747 static inline void clear_page_guard(struct zone *zone, struct page *page,
748 unsigned int order, int migratetype)
750 if (!debug_guardpage_enabled())
753 __ClearPageGuard(page);
755 set_page_private(page, 0);
756 if (!is_migrate_isolate(migratetype))
757 __mod_zone_freepage_state(zone, (1 << order), migratetype);
760 static inline bool set_page_guard(struct zone *zone, struct page *page,
761 unsigned int order, int migratetype) { return false; }
762 static inline void clear_page_guard(struct zone *zone, struct page *page,
763 unsigned int order, int migratetype) {}
766 static inline void set_page_order(struct page *page, unsigned int order)
768 set_page_private(page, order);
769 __SetPageBuddy(page);
773 * This function checks whether a page is free && is the buddy
774 * we can coalesce a page and its buddy if
775 * (a) the buddy is not in a hole (check before calling!) &&
776 * (b) the buddy is in the buddy system &&
777 * (c) a page and its buddy have the same order &&
778 * (d) a page and its buddy are in the same zone.
780 * For recording whether a page is in the buddy system, we set PageBuddy.
781 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
783 * For recording page's order, we use page_private(page).
785 static inline bool page_is_buddy(struct page *page, struct page *buddy,
788 if (!page_is_guard(buddy) && !PageBuddy(buddy))
791 if (page_order(buddy) != order)
795 * zone check is done late to avoid uselessly calculating
796 * zone/node ids for pages that could never merge.
798 if (page_zone_id(page) != page_zone_id(buddy))
801 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
806 #ifdef CONFIG_COMPACTION
807 static inline struct capture_control *task_capc(struct zone *zone)
809 struct capture_control *capc = current->capture_control;
811 return unlikely(capc) &&
812 !(current->flags & PF_KTHREAD) &&
814 capc->cc->zone == zone ? capc : NULL;
818 compaction_capture(struct capture_control *capc, struct page *page,
819 int order, int migratetype)
821 if (!capc || order != capc->cc->order)
824 /* Do not accidentally pollute CMA or isolated regions*/
825 if (is_migrate_cma(migratetype) ||
826 is_migrate_isolate(migratetype))
830 * Do not let lower order allocations polluate a movable pageblock.
831 * This might let an unmovable request use a reclaimable pageblock
832 * and vice-versa but no more than normal fallback logic which can
833 * have trouble finding a high-order free page.
835 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
843 static inline struct capture_control *task_capc(struct zone *zone)
849 compaction_capture(struct capture_control *capc, struct page *page,
850 int order, int migratetype)
854 #endif /* CONFIG_COMPACTION */
856 /* Used for pages not on another list */
857 static inline void add_to_free_list(struct page *page, struct zone *zone,
858 unsigned int order, int migratetype)
860 struct free_area *area = &zone->free_area[order];
862 list_add(&page->lru, &area->free_list[migratetype]);
866 /* Used for pages not on another list */
867 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
868 unsigned int order, int migratetype)
870 struct free_area *area = &zone->free_area[order];
872 list_add_tail(&page->lru, &area->free_list[migratetype]);
876 /* Used for pages which are on another list */
877 static inline void move_to_free_list(struct page *page, struct zone *zone,
878 unsigned int order, int migratetype)
880 struct free_area *area = &zone->free_area[order];
882 list_move(&page->lru, &area->free_list[migratetype]);
885 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
888 /* clear reported state and update reported page count */
889 if (page_reported(page))
890 __ClearPageReported(page);
892 list_del(&page->lru);
893 __ClearPageBuddy(page);
894 set_page_private(page, 0);
895 zone->free_area[order].nr_free--;
899 * If this is not the largest possible page, check if the buddy
900 * of the next-highest order is free. If it is, it's possible
901 * that pages are being freed that will coalesce soon. In case,
902 * that is happening, add the free page to the tail of the list
903 * so it's less likely to be used soon and more likely to be merged
904 * as a higher order page
907 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
908 struct page *page, unsigned int order)
910 struct page *higher_page, *higher_buddy;
911 unsigned long combined_pfn;
913 if (order >= MAX_ORDER - 2)
916 if (!pfn_valid_within(buddy_pfn))
919 combined_pfn = buddy_pfn & pfn;
920 higher_page = page + (combined_pfn - pfn);
921 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
922 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
924 return pfn_valid_within(buddy_pfn) &&
925 page_is_buddy(higher_page, higher_buddy, order + 1);
929 * Freeing function for a buddy system allocator.
931 * The concept of a buddy system is to maintain direct-mapped table
932 * (containing bit values) for memory blocks of various "orders".
933 * The bottom level table contains the map for the smallest allocatable
934 * units of memory (here, pages), and each level above it describes
935 * pairs of units from the levels below, hence, "buddies".
936 * At a high level, all that happens here is marking the table entry
937 * at the bottom level available, and propagating the changes upward
938 * as necessary, plus some accounting needed to play nicely with other
939 * parts of the VM system.
940 * At each level, we keep a list of pages, which are heads of continuous
941 * free pages of length of (1 << order) and marked with PageBuddy.
942 * Page's order is recorded in page_private(page) field.
943 * So when we are allocating or freeing one, we can derive the state of the
944 * other. That is, if we allocate a small block, and both were
945 * free, the remainder of the region must be split into blocks.
946 * If a block is freed, and its buddy is also free, then this
947 * triggers coalescing into a block of larger size.
952 static inline void __free_one_page(struct page *page,
954 struct zone *zone, unsigned int order,
955 int migratetype, bool report)
957 struct capture_control *capc = task_capc(zone);
958 unsigned long buddy_pfn;
959 unsigned long combined_pfn;
960 unsigned int max_order;
964 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
966 VM_BUG_ON(!zone_is_initialized(zone));
967 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
969 VM_BUG_ON(migratetype == -1);
970 if (likely(!is_migrate_isolate(migratetype)))
971 __mod_zone_freepage_state(zone, 1 << order, migratetype);
973 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
974 VM_BUG_ON_PAGE(bad_range(zone, page), page);
977 while (order < max_order - 1) {
978 if (compaction_capture(capc, page, order, migratetype)) {
979 __mod_zone_freepage_state(zone, -(1 << order),
983 buddy_pfn = __find_buddy_pfn(pfn, order);
984 buddy = page + (buddy_pfn - pfn);
986 if (!pfn_valid_within(buddy_pfn))
988 if (!page_is_buddy(page, buddy, order))
991 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
992 * merge with it and move up one order.
994 if (page_is_guard(buddy))
995 clear_page_guard(zone, buddy, order, migratetype);
997 del_page_from_free_list(buddy, zone, order);
998 combined_pfn = buddy_pfn & pfn;
999 page = page + (combined_pfn - pfn);
1003 if (max_order < MAX_ORDER) {
1004 /* If we are here, it means order is >= pageblock_order.
1005 * We want to prevent merge between freepages on isolate
1006 * pageblock and normal pageblock. Without this, pageblock
1007 * isolation could cause incorrect freepage or CMA accounting.
1009 * We don't want to hit this code for the more frequent
1010 * low-order merging.
1012 if (unlikely(has_isolate_pageblock(zone))) {
1015 buddy_pfn = __find_buddy_pfn(pfn, order);
1016 buddy = page + (buddy_pfn - pfn);
1017 buddy_mt = get_pageblock_migratetype(buddy);
1019 if (migratetype != buddy_mt
1020 && (is_migrate_isolate(migratetype) ||
1021 is_migrate_isolate(buddy_mt)))
1025 goto continue_merging;
1029 set_page_order(page, order);
1031 if (is_shuffle_order(order))
1032 to_tail = shuffle_pick_tail();
1034 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1037 add_to_free_list_tail(page, zone, order, migratetype);
1039 add_to_free_list(page, zone, order, migratetype);
1041 /* Notify page reporting subsystem of freed page */
1043 page_reporting_notify_free(order);
1047 * A bad page could be due to a number of fields. Instead of multiple branches,
1048 * try and check multiple fields with one check. The caller must do a detailed
1049 * check if necessary.
1051 static inline bool page_expected_state(struct page *page,
1052 unsigned long check_flags)
1054 if (unlikely(atomic_read(&page->_mapcount) != -1))
1057 if (unlikely((unsigned long)page->mapping |
1058 page_ref_count(page) |
1060 (unsigned long)page->mem_cgroup |
1062 (page->flags & check_flags)))
1068 static const char *page_bad_reason(struct page *page, unsigned long flags)
1070 const char *bad_reason = NULL;
1072 if (unlikely(atomic_read(&page->_mapcount) != -1))
1073 bad_reason = "nonzero mapcount";
1074 if (unlikely(page->mapping != NULL))
1075 bad_reason = "non-NULL mapping";
1076 if (unlikely(page_ref_count(page) != 0))
1077 bad_reason = "nonzero _refcount";
1078 if (unlikely(page->flags & flags)) {
1079 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1080 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1082 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1085 if (unlikely(page->mem_cgroup))
1086 bad_reason = "page still charged to cgroup";
1091 static void check_free_page_bad(struct page *page)
1094 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1097 static inline int check_free_page(struct page *page)
1099 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1102 /* Something has gone sideways, find it */
1103 check_free_page_bad(page);
1107 static int free_tail_pages_check(struct page *head_page, struct page *page)
1112 * We rely page->lru.next never has bit 0 set, unless the page
1113 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1115 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1117 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1121 switch (page - head_page) {
1123 /* the first tail page: ->mapping may be compound_mapcount() */
1124 if (unlikely(compound_mapcount(page))) {
1125 bad_page(page, "nonzero compound_mapcount");
1131 * the second tail page: ->mapping is
1132 * deferred_list.next -- ignore value.
1136 if (page->mapping != TAIL_MAPPING) {
1137 bad_page(page, "corrupted mapping in tail page");
1142 if (unlikely(!PageTail(page))) {
1143 bad_page(page, "PageTail not set");
1146 if (unlikely(compound_head(page) != head_page)) {
1147 bad_page(page, "compound_head not consistent");
1152 page->mapping = NULL;
1153 clear_compound_head(page);
1157 static void kernel_init_free_pages(struct page *page, int numpages)
1161 /* s390's use of memset() could override KASAN redzones. */
1162 kasan_disable_current();
1163 for (i = 0; i < numpages; i++)
1164 clear_highpage(page + i);
1165 kasan_enable_current();
1168 static __always_inline bool free_pages_prepare(struct page *page,
1169 unsigned int order, bool check_free)
1173 VM_BUG_ON_PAGE(PageTail(page), page);
1175 trace_mm_page_free(page, order);
1177 if (unlikely(PageHWPoison(page)) && !order) {
1179 * Do not let hwpoison pages hit pcplists/buddy
1180 * Untie memcg state and reset page's owner
1182 if (memcg_kmem_enabled() && PageKmemcg(page))
1183 __memcg_kmem_uncharge_page(page, order);
1184 reset_page_owner(page, order);
1189 * Check tail pages before head page information is cleared to
1190 * avoid checking PageCompound for order-0 pages.
1192 if (unlikely(order)) {
1193 bool compound = PageCompound(page);
1196 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1199 ClearPageDoubleMap(page);
1200 for (i = 1; i < (1 << order); i++) {
1202 bad += free_tail_pages_check(page, page + i);
1203 if (unlikely(check_free_page(page + i))) {
1207 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1210 if (PageMappingFlags(page))
1211 page->mapping = NULL;
1212 if (memcg_kmem_enabled() && PageKmemcg(page))
1213 __memcg_kmem_uncharge_page(page, order);
1215 bad += check_free_page(page);
1219 page_cpupid_reset_last(page);
1220 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1221 reset_page_owner(page, order);
1223 if (!PageHighMem(page)) {
1224 debug_check_no_locks_freed(page_address(page),
1225 PAGE_SIZE << order);
1226 debug_check_no_obj_freed(page_address(page),
1227 PAGE_SIZE << order);
1229 if (want_init_on_free())
1230 kernel_init_free_pages(page, 1 << order);
1232 kernel_poison_pages(page, 1 << order, 0);
1234 * arch_free_page() can make the page's contents inaccessible. s390
1235 * does this. So nothing which can access the page's contents should
1236 * happen after this.
1238 arch_free_page(page, order);
1240 if (debug_pagealloc_enabled_static())
1241 kernel_map_pages(page, 1 << order, 0);
1243 kasan_free_nondeferred_pages(page, order);
1248 #ifdef CONFIG_DEBUG_VM
1250 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1251 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1252 * moved from pcp lists to free lists.
1254 static bool free_pcp_prepare(struct page *page)
1256 return free_pages_prepare(page, 0, true);
1259 static bool bulkfree_pcp_prepare(struct page *page)
1261 if (debug_pagealloc_enabled_static())
1262 return check_free_page(page);
1268 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1269 * moving from pcp lists to free list in order to reduce overhead. With
1270 * debug_pagealloc enabled, they are checked also immediately when being freed
1273 static bool free_pcp_prepare(struct page *page)
1275 if (debug_pagealloc_enabled_static())
1276 return free_pages_prepare(page, 0, true);
1278 return free_pages_prepare(page, 0, false);
1281 static bool bulkfree_pcp_prepare(struct page *page)
1283 return check_free_page(page);
1285 #endif /* CONFIG_DEBUG_VM */
1287 static inline void prefetch_buddy(struct page *page)
1289 unsigned long pfn = page_to_pfn(page);
1290 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1291 struct page *buddy = page + (buddy_pfn - pfn);
1297 * Frees a number of pages from the PCP lists
1298 * Assumes all pages on list are in same zone, and of same order.
1299 * count is the number of pages to free.
1301 * If the zone was previously in an "all pages pinned" state then look to
1302 * see if this freeing clears that state.
1304 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1305 * pinned" detection logic.
1307 static void free_pcppages_bulk(struct zone *zone, int count,
1308 struct per_cpu_pages *pcp)
1310 int migratetype = 0;
1312 int prefetch_nr = 0;
1313 bool isolated_pageblocks;
1314 struct page *page, *tmp;
1318 * Ensure proper count is passed which otherwise would stuck in the
1319 * below while (list_empty(list)) loop.
1321 count = min(pcp->count, count);
1323 struct list_head *list;
1326 * Remove pages from lists in a round-robin fashion. A
1327 * batch_free count is maintained that is incremented when an
1328 * empty list is encountered. This is so more pages are freed
1329 * off fuller lists instead of spinning excessively around empty
1334 if (++migratetype == MIGRATE_PCPTYPES)
1336 list = &pcp->lists[migratetype];
1337 } while (list_empty(list));
1339 /* This is the only non-empty list. Free them all. */
1340 if (batch_free == MIGRATE_PCPTYPES)
1344 page = list_last_entry(list, struct page, lru);
1345 /* must delete to avoid corrupting pcp list */
1346 list_del(&page->lru);
1349 if (bulkfree_pcp_prepare(page))
1352 list_add_tail(&page->lru, &head);
1355 * We are going to put the page back to the global
1356 * pool, prefetch its buddy to speed up later access
1357 * under zone->lock. It is believed the overhead of
1358 * an additional test and calculating buddy_pfn here
1359 * can be offset by reduced memory latency later. To
1360 * avoid excessive prefetching due to large count, only
1361 * prefetch buddy for the first pcp->batch nr of pages.
1363 if (prefetch_nr++ < pcp->batch)
1364 prefetch_buddy(page);
1365 } while (--count && --batch_free && !list_empty(list));
1368 spin_lock(&zone->lock);
1369 isolated_pageblocks = has_isolate_pageblock(zone);
1372 * Use safe version since after __free_one_page(),
1373 * page->lru.next will not point to original list.
1375 list_for_each_entry_safe(page, tmp, &head, lru) {
1376 int mt = get_pcppage_migratetype(page);
1377 /* MIGRATE_ISOLATE page should not go to pcplists */
1378 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1379 /* Pageblock could have been isolated meanwhile */
1380 if (unlikely(isolated_pageblocks))
1381 mt = get_pageblock_migratetype(page);
1383 __free_one_page(page, page_to_pfn(page), zone, 0, mt, true);
1384 trace_mm_page_pcpu_drain(page, 0, mt);
1386 spin_unlock(&zone->lock);
1389 static void free_one_page(struct zone *zone,
1390 struct page *page, unsigned long pfn,
1394 spin_lock(&zone->lock);
1395 if (unlikely(has_isolate_pageblock(zone) ||
1396 is_migrate_isolate(migratetype))) {
1397 migratetype = get_pfnblock_migratetype(page, pfn);
1399 __free_one_page(page, pfn, zone, order, migratetype, true);
1400 spin_unlock(&zone->lock);
1403 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1404 unsigned long zone, int nid)
1406 mm_zero_struct_page(page);
1407 set_page_links(page, zone, nid, pfn);
1408 init_page_count(page);
1409 page_mapcount_reset(page);
1410 page_cpupid_reset_last(page);
1411 page_kasan_tag_reset(page);
1413 INIT_LIST_HEAD(&page->lru);
1414 #ifdef WANT_PAGE_VIRTUAL
1415 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1416 if (!is_highmem_idx(zone))
1417 set_page_address(page, __va(pfn << PAGE_SHIFT));
1421 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1422 static void __meminit init_reserved_page(unsigned long pfn)
1427 if (!early_page_uninitialised(pfn))
1430 nid = early_pfn_to_nid(pfn);
1431 pgdat = NODE_DATA(nid);
1433 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1434 struct zone *zone = &pgdat->node_zones[zid];
1436 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1439 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1442 static inline void init_reserved_page(unsigned long pfn)
1445 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1448 * Initialised pages do not have PageReserved set. This function is
1449 * called for each range allocated by the bootmem allocator and
1450 * marks the pages PageReserved. The remaining valid pages are later
1451 * sent to the buddy page allocator.
1453 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1455 unsigned long start_pfn = PFN_DOWN(start);
1456 unsigned long end_pfn = PFN_UP(end);
1458 for (; start_pfn < end_pfn; start_pfn++) {
1459 if (pfn_valid(start_pfn)) {
1460 struct page *page = pfn_to_page(start_pfn);
1462 init_reserved_page(start_pfn);
1464 /* Avoid false-positive PageTail() */
1465 INIT_LIST_HEAD(&page->lru);
1468 * no need for atomic set_bit because the struct
1469 * page is not visible yet so nobody should
1472 __SetPageReserved(page);
1477 static void __free_pages_ok(struct page *page, unsigned int order)
1479 unsigned long flags;
1481 unsigned long pfn = page_to_pfn(page);
1483 if (!free_pages_prepare(page, order, true))
1486 migratetype = get_pfnblock_migratetype(page, pfn);
1487 local_irq_save(flags);
1488 __count_vm_events(PGFREE, 1 << order);
1489 free_one_page(page_zone(page), page, pfn, order, migratetype);
1490 local_irq_restore(flags);
1493 void __free_pages_core(struct page *page, unsigned int order)
1495 unsigned int nr_pages = 1 << order;
1496 struct page *p = page;
1500 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1502 __ClearPageReserved(p);
1503 set_page_count(p, 0);
1505 __ClearPageReserved(p);
1506 set_page_count(p, 0);
1508 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1509 set_page_refcounted(page);
1510 __free_pages(page, order);
1513 #ifdef CONFIG_NEED_MULTIPLE_NODES
1515 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1517 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
1520 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1522 int __meminit __early_pfn_to_nid(unsigned long pfn,
1523 struct mminit_pfnnid_cache *state)
1525 unsigned long start_pfn, end_pfn;
1528 if (state->last_start <= pfn && pfn < state->last_end)
1529 return state->last_nid;
1531 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1532 if (nid != NUMA_NO_NODE) {
1533 state->last_start = start_pfn;
1534 state->last_end = end_pfn;
1535 state->last_nid = nid;
1540 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
1542 int __meminit early_pfn_to_nid(unsigned long pfn)
1544 static DEFINE_SPINLOCK(early_pfn_lock);
1547 spin_lock(&early_pfn_lock);
1548 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1550 nid = first_online_node;
1551 spin_unlock(&early_pfn_lock);
1555 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1557 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1560 if (early_page_uninitialised(pfn))
1562 __free_pages_core(page, order);
1566 * Check that the whole (or subset of) a pageblock given by the interval of
1567 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1568 * with the migration of free compaction scanner. The scanners then need to
1569 * use only pfn_valid_within() check for arches that allow holes within
1572 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1574 * It's possible on some configurations to have a setup like node0 node1 node0
1575 * i.e. it's possible that all pages within a zones range of pages do not
1576 * belong to a single zone. We assume that a border between node0 and node1
1577 * can occur within a single pageblock, but not a node0 node1 node0
1578 * interleaving within a single pageblock. It is therefore sufficient to check
1579 * the first and last page of a pageblock and avoid checking each individual
1580 * page in a pageblock.
1582 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1583 unsigned long end_pfn, struct zone *zone)
1585 struct page *start_page;
1586 struct page *end_page;
1588 /* end_pfn is one past the range we are checking */
1591 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1594 start_page = pfn_to_online_page(start_pfn);
1598 if (page_zone(start_page) != zone)
1601 end_page = pfn_to_page(end_pfn);
1603 /* This gives a shorter code than deriving page_zone(end_page) */
1604 if (page_zone_id(start_page) != page_zone_id(end_page))
1610 void set_zone_contiguous(struct zone *zone)
1612 unsigned long block_start_pfn = zone->zone_start_pfn;
1613 unsigned long block_end_pfn;
1615 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1616 for (; block_start_pfn < zone_end_pfn(zone);
1617 block_start_pfn = block_end_pfn,
1618 block_end_pfn += pageblock_nr_pages) {
1620 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1622 if (!__pageblock_pfn_to_page(block_start_pfn,
1623 block_end_pfn, zone))
1628 /* We confirm that there is no hole */
1629 zone->contiguous = true;
1632 void clear_zone_contiguous(struct zone *zone)
1634 zone->contiguous = false;
1637 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1638 static void __init deferred_free_range(unsigned long pfn,
1639 unsigned long nr_pages)
1647 page = pfn_to_page(pfn);
1649 /* Free a large naturally-aligned chunk if possible */
1650 if (nr_pages == pageblock_nr_pages &&
1651 (pfn & (pageblock_nr_pages - 1)) == 0) {
1652 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1653 __free_pages_core(page, pageblock_order);
1657 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1658 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1659 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1660 __free_pages_core(page, 0);
1664 /* Completion tracking for deferred_init_memmap() threads */
1665 static atomic_t pgdat_init_n_undone __initdata;
1666 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1668 static inline void __init pgdat_init_report_one_done(void)
1670 if (atomic_dec_and_test(&pgdat_init_n_undone))
1671 complete(&pgdat_init_all_done_comp);
1675 * Returns true if page needs to be initialized or freed to buddy allocator.
1677 * First we check if pfn is valid on architectures where it is possible to have
1678 * holes within pageblock_nr_pages. On systems where it is not possible, this
1679 * function is optimized out.
1681 * Then, we check if a current large page is valid by only checking the validity
1684 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1686 if (!pfn_valid_within(pfn))
1688 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1694 * Free pages to buddy allocator. Try to free aligned pages in
1695 * pageblock_nr_pages sizes.
1697 static void __init deferred_free_pages(unsigned long pfn,
1698 unsigned long end_pfn)
1700 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1701 unsigned long nr_free = 0;
1703 for (; pfn < end_pfn; pfn++) {
1704 if (!deferred_pfn_valid(pfn)) {
1705 deferred_free_range(pfn - nr_free, nr_free);
1707 } else if (!(pfn & nr_pgmask)) {
1708 deferred_free_range(pfn - nr_free, nr_free);
1714 /* Free the last block of pages to allocator */
1715 deferred_free_range(pfn - nr_free, nr_free);
1719 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1720 * by performing it only once every pageblock_nr_pages.
1721 * Return number of pages initialized.
1723 static unsigned long __init deferred_init_pages(struct zone *zone,
1725 unsigned long end_pfn)
1727 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1728 int nid = zone_to_nid(zone);
1729 unsigned long nr_pages = 0;
1730 int zid = zone_idx(zone);
1731 struct page *page = NULL;
1733 for (; pfn < end_pfn; pfn++) {
1734 if (!deferred_pfn_valid(pfn)) {
1737 } else if (!page || !(pfn & nr_pgmask)) {
1738 page = pfn_to_page(pfn);
1742 __init_single_page(page, pfn, zid, nid);
1749 * This function is meant to pre-load the iterator for the zone init.
1750 * Specifically it walks through the ranges until we are caught up to the
1751 * first_init_pfn value and exits there. If we never encounter the value we
1752 * return false indicating there are no valid ranges left.
1755 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1756 unsigned long *spfn, unsigned long *epfn,
1757 unsigned long first_init_pfn)
1762 * Start out by walking through the ranges in this zone that have
1763 * already been initialized. We don't need to do anything with them
1764 * so we just need to flush them out of the system.
1766 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1767 if (*epfn <= first_init_pfn)
1769 if (*spfn < first_init_pfn)
1770 *spfn = first_init_pfn;
1779 * Initialize and free pages. We do it in two loops: first we initialize
1780 * struct page, then free to buddy allocator, because while we are
1781 * freeing pages we can access pages that are ahead (computing buddy
1782 * page in __free_one_page()).
1784 * In order to try and keep some memory in the cache we have the loop
1785 * broken along max page order boundaries. This way we will not cause
1786 * any issues with the buddy page computation.
1788 static unsigned long __init
1789 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1790 unsigned long *end_pfn)
1792 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1793 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1794 unsigned long nr_pages = 0;
1797 /* First we loop through and initialize the page values */
1798 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1801 if (mo_pfn <= *start_pfn)
1804 t = min(mo_pfn, *end_pfn);
1805 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1807 if (mo_pfn < *end_pfn) {
1808 *start_pfn = mo_pfn;
1813 /* Reset values and now loop through freeing pages as needed */
1816 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1822 t = min(mo_pfn, epfn);
1823 deferred_free_pages(spfn, t);
1833 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1836 unsigned long spfn, epfn;
1837 struct zone *zone = arg;
1840 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1843 * Initialize and free pages in MAX_ORDER sized increments so that we
1844 * can avoid introducing any issues with the buddy allocator.
1846 while (spfn < end_pfn) {
1847 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1852 /* An arch may override for more concurrency. */
1854 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1859 /* Initialise remaining memory on a node */
1860 static int __init deferred_init_memmap(void *data)
1862 pg_data_t *pgdat = data;
1863 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1864 unsigned long spfn = 0, epfn = 0;
1865 unsigned long first_init_pfn, flags;
1866 unsigned long start = jiffies;
1868 int zid, max_threads;
1871 /* Bind memory initialisation thread to a local node if possible */
1872 if (!cpumask_empty(cpumask))
1873 set_cpus_allowed_ptr(current, cpumask);
1875 pgdat_resize_lock(pgdat, &flags);
1876 first_init_pfn = pgdat->first_deferred_pfn;
1877 if (first_init_pfn == ULONG_MAX) {
1878 pgdat_resize_unlock(pgdat, &flags);
1879 pgdat_init_report_one_done();
1883 /* Sanity check boundaries */
1884 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1885 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1886 pgdat->first_deferred_pfn = ULONG_MAX;
1889 * Once we unlock here, the zone cannot be grown anymore, thus if an
1890 * interrupt thread must allocate this early in boot, zone must be
1891 * pre-grown prior to start of deferred page initialization.
1893 pgdat_resize_unlock(pgdat, &flags);
1895 /* Only the highest zone is deferred so find it */
1896 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1897 zone = pgdat->node_zones + zid;
1898 if (first_init_pfn < zone_end_pfn(zone))
1902 /* If the zone is empty somebody else may have cleared out the zone */
1903 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1907 max_threads = deferred_page_init_max_threads(cpumask);
1909 while (spfn < epfn) {
1910 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
1911 struct padata_mt_job job = {
1912 .thread_fn = deferred_init_memmap_chunk,
1915 .size = epfn_align - spfn,
1916 .align = PAGES_PER_SECTION,
1917 .min_chunk = PAGES_PER_SECTION,
1918 .max_threads = max_threads,
1921 padata_do_multithreaded(&job);
1922 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1926 /* Sanity check that the next zone really is unpopulated */
1927 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1929 pr_info("node %d deferred pages initialised in %ums\n",
1930 pgdat->node_id, jiffies_to_msecs(jiffies - start));
1932 pgdat_init_report_one_done();
1937 * If this zone has deferred pages, try to grow it by initializing enough
1938 * deferred pages to satisfy the allocation specified by order, rounded up to
1939 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1940 * of SECTION_SIZE bytes by initializing struct pages in increments of
1941 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1943 * Return true when zone was grown, otherwise return false. We return true even
1944 * when we grow less than requested, to let the caller decide if there are
1945 * enough pages to satisfy the allocation.
1947 * Note: We use noinline because this function is needed only during boot, and
1948 * it is called from a __ref function _deferred_grow_zone. This way we are
1949 * making sure that it is not inlined into permanent text section.
1951 static noinline bool __init
1952 deferred_grow_zone(struct zone *zone, unsigned int order)
1954 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1955 pg_data_t *pgdat = zone->zone_pgdat;
1956 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1957 unsigned long spfn, epfn, flags;
1958 unsigned long nr_pages = 0;
1961 /* Only the last zone may have deferred pages */
1962 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1965 pgdat_resize_lock(pgdat, &flags);
1968 * If someone grew this zone while we were waiting for spinlock, return
1969 * true, as there might be enough pages already.
1971 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1972 pgdat_resize_unlock(pgdat, &flags);
1976 /* If the zone is empty somebody else may have cleared out the zone */
1977 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1978 first_deferred_pfn)) {
1979 pgdat->first_deferred_pfn = ULONG_MAX;
1980 pgdat_resize_unlock(pgdat, &flags);
1981 /* Retry only once. */
1982 return first_deferred_pfn != ULONG_MAX;
1986 * Initialize and free pages in MAX_ORDER sized increments so
1987 * that we can avoid introducing any issues with the buddy
1990 while (spfn < epfn) {
1991 /* update our first deferred PFN for this section */
1992 first_deferred_pfn = spfn;
1994 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1995 touch_nmi_watchdog();
1997 /* We should only stop along section boundaries */
1998 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2001 /* If our quota has been met we can stop here */
2002 if (nr_pages >= nr_pages_needed)
2006 pgdat->first_deferred_pfn = spfn;
2007 pgdat_resize_unlock(pgdat, &flags);
2009 return nr_pages > 0;
2013 * deferred_grow_zone() is __init, but it is called from
2014 * get_page_from_freelist() during early boot until deferred_pages permanently
2015 * disables this call. This is why we have refdata wrapper to avoid warning,
2016 * and to ensure that the function body gets unloaded.
2019 _deferred_grow_zone(struct zone *zone, unsigned int order)
2021 return deferred_grow_zone(zone, order);
2024 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2026 void __init page_alloc_init_late(void)
2031 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2033 /* There will be num_node_state(N_MEMORY) threads */
2034 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2035 for_each_node_state(nid, N_MEMORY) {
2036 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2039 /* Block until all are initialised */
2040 wait_for_completion(&pgdat_init_all_done_comp);
2043 * The number of managed pages has changed due to the initialisation
2044 * so the pcpu batch and high limits needs to be updated or the limits
2045 * will be artificially small.
2047 for_each_populated_zone(zone)
2048 zone_pcp_update(zone);
2051 * We initialized the rest of the deferred pages. Permanently disable
2052 * on-demand struct page initialization.
2054 static_branch_disable(&deferred_pages);
2056 /* Reinit limits that are based on free pages after the kernel is up */
2057 files_maxfiles_init();
2060 /* Discard memblock private memory */
2063 for_each_node_state(nid, N_MEMORY)
2064 shuffle_free_memory(NODE_DATA(nid));
2066 for_each_populated_zone(zone)
2067 set_zone_contiguous(zone);
2071 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2072 void __init init_cma_reserved_pageblock(struct page *page)
2074 unsigned i = pageblock_nr_pages;
2075 struct page *p = page;
2078 __ClearPageReserved(p);
2079 set_page_count(p, 0);
2082 set_pageblock_migratetype(page, MIGRATE_CMA);
2084 if (pageblock_order >= MAX_ORDER) {
2085 i = pageblock_nr_pages;
2088 set_page_refcounted(p);
2089 __free_pages(p, MAX_ORDER - 1);
2090 p += MAX_ORDER_NR_PAGES;
2091 } while (i -= MAX_ORDER_NR_PAGES);
2093 set_page_refcounted(page);
2094 __free_pages(page, pageblock_order);
2097 adjust_managed_page_count(page, pageblock_nr_pages);
2102 * The order of subdivision here is critical for the IO subsystem.
2103 * Please do not alter this order without good reasons and regression
2104 * testing. Specifically, as large blocks of memory are subdivided,
2105 * the order in which smaller blocks are delivered depends on the order
2106 * they're subdivided in this function. This is the primary factor
2107 * influencing the order in which pages are delivered to the IO
2108 * subsystem according to empirical testing, and this is also justified
2109 * by considering the behavior of a buddy system containing a single
2110 * large block of memory acted on by a series of small allocations.
2111 * This behavior is a critical factor in sglist merging's success.
2115 static inline void expand(struct zone *zone, struct page *page,
2116 int low, int high, int migratetype)
2118 unsigned long size = 1 << high;
2120 while (high > low) {
2123 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2126 * Mark as guard pages (or page), that will allow to
2127 * merge back to allocator when buddy will be freed.
2128 * Corresponding page table entries will not be touched,
2129 * pages will stay not present in virtual address space
2131 if (set_page_guard(zone, &page[size], high, migratetype))
2134 add_to_free_list(&page[size], zone, high, migratetype);
2135 set_page_order(&page[size], high);
2139 static void check_new_page_bad(struct page *page)
2141 if (unlikely(page->flags & __PG_HWPOISON)) {
2142 /* Don't complain about hwpoisoned pages */
2143 page_mapcount_reset(page); /* remove PageBuddy */
2148 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2152 * This page is about to be returned from the page allocator
2154 static inline int check_new_page(struct page *page)
2156 if (likely(page_expected_state(page,
2157 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2160 check_new_page_bad(page);
2164 static inline bool free_pages_prezeroed(void)
2166 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2167 page_poisoning_enabled()) || want_init_on_free();
2170 #ifdef CONFIG_DEBUG_VM
2172 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2173 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2174 * also checked when pcp lists are refilled from the free lists.
2176 static inline bool check_pcp_refill(struct page *page)
2178 if (debug_pagealloc_enabled_static())
2179 return check_new_page(page);
2184 static inline bool check_new_pcp(struct page *page)
2186 return check_new_page(page);
2190 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2191 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2192 * enabled, they are also checked when being allocated from the pcp lists.
2194 static inline bool check_pcp_refill(struct page *page)
2196 return check_new_page(page);
2198 static inline bool check_new_pcp(struct page *page)
2200 if (debug_pagealloc_enabled_static())
2201 return check_new_page(page);
2205 #endif /* CONFIG_DEBUG_VM */
2207 static bool check_new_pages(struct page *page, unsigned int order)
2210 for (i = 0; i < (1 << order); i++) {
2211 struct page *p = page + i;
2213 if (unlikely(check_new_page(p)))
2220 inline void post_alloc_hook(struct page *page, unsigned int order,
2223 set_page_private(page, 0);
2224 set_page_refcounted(page);
2226 arch_alloc_page(page, order);
2227 if (debug_pagealloc_enabled_static())
2228 kernel_map_pages(page, 1 << order, 1);
2229 kasan_alloc_pages(page, order);
2230 kernel_poison_pages(page, 1 << order, 1);
2231 set_page_owner(page, order, gfp_flags);
2234 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2235 unsigned int alloc_flags)
2237 post_alloc_hook(page, order, gfp_flags);
2239 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2240 kernel_init_free_pages(page, 1 << order);
2242 if (order && (gfp_flags & __GFP_COMP))
2243 prep_compound_page(page, order);
2246 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2247 * allocate the page. The expectation is that the caller is taking
2248 * steps that will free more memory. The caller should avoid the page
2249 * being used for !PFMEMALLOC purposes.
2251 if (alloc_flags & ALLOC_NO_WATERMARKS)
2252 set_page_pfmemalloc(page);
2254 clear_page_pfmemalloc(page);
2258 * Go through the free lists for the given migratetype and remove
2259 * the smallest available page from the freelists
2261 static __always_inline
2262 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2265 unsigned int current_order;
2266 struct free_area *area;
2269 /* Find a page of the appropriate size in the preferred list */
2270 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2271 area = &(zone->free_area[current_order]);
2272 page = get_page_from_free_area(area, migratetype);
2275 del_page_from_free_list(page, zone, current_order);
2276 expand(zone, page, order, current_order, migratetype);
2277 set_pcppage_migratetype(page, migratetype);
2286 * This array describes the order lists are fallen back to when
2287 * the free lists for the desirable migrate type are depleted
2289 static int fallbacks[MIGRATE_TYPES][3] = {
2290 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2291 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2292 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2294 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2296 #ifdef CONFIG_MEMORY_ISOLATION
2297 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2302 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2305 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2308 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2309 unsigned int order) { return NULL; }
2313 * Move the free pages in a range to the free lists of the requested type.
2314 * Note that start_page and end_pages are not aligned on a pageblock
2315 * boundary. If alignment is required, use move_freepages_block()
2317 static int move_freepages(struct zone *zone,
2318 struct page *start_page, struct page *end_page,
2319 int migratetype, int *num_movable)
2323 int pages_moved = 0;
2325 for (page = start_page; page <= end_page;) {
2326 if (!pfn_valid_within(page_to_pfn(page))) {
2331 if (!PageBuddy(page)) {
2333 * We assume that pages that could be isolated for
2334 * migration are movable. But we don't actually try
2335 * isolating, as that would be expensive.
2338 (PageLRU(page) || __PageMovable(page)))
2345 /* Make sure we are not inadvertently changing nodes */
2346 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2347 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2349 order = page_order(page);
2350 move_to_free_list(page, zone, order, migratetype);
2352 pages_moved += 1 << order;
2358 int move_freepages_block(struct zone *zone, struct page *page,
2359 int migratetype, int *num_movable)
2361 unsigned long start_pfn, end_pfn;
2362 struct page *start_page, *end_page;
2367 start_pfn = page_to_pfn(page);
2368 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2369 start_page = pfn_to_page(start_pfn);
2370 end_page = start_page + pageblock_nr_pages - 1;
2371 end_pfn = start_pfn + pageblock_nr_pages - 1;
2373 /* Do not cross zone boundaries */
2374 if (!zone_spans_pfn(zone, start_pfn))
2376 if (!zone_spans_pfn(zone, end_pfn))
2379 return move_freepages(zone, start_page, end_page, migratetype,
2383 static void change_pageblock_range(struct page *pageblock_page,
2384 int start_order, int migratetype)
2386 int nr_pageblocks = 1 << (start_order - pageblock_order);
2388 while (nr_pageblocks--) {
2389 set_pageblock_migratetype(pageblock_page, migratetype);
2390 pageblock_page += pageblock_nr_pages;
2395 * When we are falling back to another migratetype during allocation, try to
2396 * steal extra free pages from the same pageblocks to satisfy further
2397 * allocations, instead of polluting multiple pageblocks.
2399 * If we are stealing a relatively large buddy page, it is likely there will
2400 * be more free pages in the pageblock, so try to steal them all. For
2401 * reclaimable and unmovable allocations, we steal regardless of page size,
2402 * as fragmentation caused by those allocations polluting movable pageblocks
2403 * is worse than movable allocations stealing from unmovable and reclaimable
2406 static bool can_steal_fallback(unsigned int order, int start_mt)
2409 * Leaving this order check is intended, although there is
2410 * relaxed order check in next check. The reason is that
2411 * we can actually steal whole pageblock if this condition met,
2412 * but, below check doesn't guarantee it and that is just heuristic
2413 * so could be changed anytime.
2415 if (order >= pageblock_order)
2418 if (order >= pageblock_order / 2 ||
2419 start_mt == MIGRATE_RECLAIMABLE ||
2420 start_mt == MIGRATE_UNMOVABLE ||
2421 page_group_by_mobility_disabled)
2427 static inline void boost_watermark(struct zone *zone)
2429 unsigned long max_boost;
2431 if (!watermark_boost_factor)
2434 * Don't bother in zones that are unlikely to produce results.
2435 * On small machines, including kdump capture kernels running
2436 * in a small area, boosting the watermark can cause an out of
2437 * memory situation immediately.
2439 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2442 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2443 watermark_boost_factor, 10000);
2446 * high watermark may be uninitialised if fragmentation occurs
2447 * very early in boot so do not boost. We do not fall
2448 * through and boost by pageblock_nr_pages as failing
2449 * allocations that early means that reclaim is not going
2450 * to help and it may even be impossible to reclaim the
2451 * boosted watermark resulting in a hang.
2456 max_boost = max(pageblock_nr_pages, max_boost);
2458 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2463 * This function implements actual steal behaviour. If order is large enough,
2464 * we can steal whole pageblock. If not, we first move freepages in this
2465 * pageblock to our migratetype and determine how many already-allocated pages
2466 * are there in the pageblock with a compatible migratetype. If at least half
2467 * of pages are free or compatible, we can change migratetype of the pageblock
2468 * itself, so pages freed in the future will be put on the correct free list.
2470 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2471 unsigned int alloc_flags, int start_type, bool whole_block)
2473 unsigned int current_order = page_order(page);
2474 int free_pages, movable_pages, alike_pages;
2477 old_block_type = get_pageblock_migratetype(page);
2480 * This can happen due to races and we want to prevent broken
2481 * highatomic accounting.
2483 if (is_migrate_highatomic(old_block_type))
2486 /* Take ownership for orders >= pageblock_order */
2487 if (current_order >= pageblock_order) {
2488 change_pageblock_range(page, current_order, start_type);
2493 * Boost watermarks to increase reclaim pressure to reduce the
2494 * likelihood of future fallbacks. Wake kswapd now as the node
2495 * may be balanced overall and kswapd will not wake naturally.
2497 boost_watermark(zone);
2498 if (alloc_flags & ALLOC_KSWAPD)
2499 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2501 /* We are not allowed to try stealing from the whole block */
2505 free_pages = move_freepages_block(zone, page, start_type,
2508 * Determine how many pages are compatible with our allocation.
2509 * For movable allocation, it's the number of movable pages which
2510 * we just obtained. For other types it's a bit more tricky.
2512 if (start_type == MIGRATE_MOVABLE) {
2513 alike_pages = movable_pages;
2516 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2517 * to MOVABLE pageblock, consider all non-movable pages as
2518 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2519 * vice versa, be conservative since we can't distinguish the
2520 * exact migratetype of non-movable pages.
2522 if (old_block_type == MIGRATE_MOVABLE)
2523 alike_pages = pageblock_nr_pages
2524 - (free_pages + movable_pages);
2529 /* moving whole block can fail due to zone boundary conditions */
2534 * If a sufficient number of pages in the block are either free or of
2535 * comparable migratability as our allocation, claim the whole block.
2537 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2538 page_group_by_mobility_disabled)
2539 set_pageblock_migratetype(page, start_type);
2544 move_to_free_list(page, zone, current_order, start_type);
2548 * Check whether there is a suitable fallback freepage with requested order.
2549 * If only_stealable is true, this function returns fallback_mt only if
2550 * we can steal other freepages all together. This would help to reduce
2551 * fragmentation due to mixed migratetype pages in one pageblock.
2553 int find_suitable_fallback(struct free_area *area, unsigned int order,
2554 int migratetype, bool only_stealable, bool *can_steal)
2559 if (area->nr_free == 0)
2564 fallback_mt = fallbacks[migratetype][i];
2565 if (fallback_mt == MIGRATE_TYPES)
2568 if (free_area_empty(area, fallback_mt))
2571 if (can_steal_fallback(order, migratetype))
2574 if (!only_stealable)
2585 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2586 * there are no empty page blocks that contain a page with a suitable order
2588 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2589 unsigned int alloc_order)
2592 unsigned long max_managed, flags;
2595 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2596 * Check is race-prone but harmless.
2598 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2599 if (zone->nr_reserved_highatomic >= max_managed)
2602 spin_lock_irqsave(&zone->lock, flags);
2604 /* Recheck the nr_reserved_highatomic limit under the lock */
2605 if (zone->nr_reserved_highatomic >= max_managed)
2609 mt = get_pageblock_migratetype(page);
2610 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2611 && !is_migrate_cma(mt)) {
2612 zone->nr_reserved_highatomic += pageblock_nr_pages;
2613 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2614 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2618 spin_unlock_irqrestore(&zone->lock, flags);
2622 * Used when an allocation is about to fail under memory pressure. This
2623 * potentially hurts the reliability of high-order allocations when under
2624 * intense memory pressure but failed atomic allocations should be easier
2625 * to recover from than an OOM.
2627 * If @force is true, try to unreserve a pageblock even though highatomic
2628 * pageblock is exhausted.
2630 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2633 struct zonelist *zonelist = ac->zonelist;
2634 unsigned long flags;
2641 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2644 * Preserve at least one pageblock unless memory pressure
2647 if (!force && zone->nr_reserved_highatomic <=
2651 spin_lock_irqsave(&zone->lock, flags);
2652 for (order = 0; order < MAX_ORDER; order++) {
2653 struct free_area *area = &(zone->free_area[order]);
2655 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2660 * In page freeing path, migratetype change is racy so
2661 * we can counter several free pages in a pageblock
2662 * in this loop althoug we changed the pageblock type
2663 * from highatomic to ac->migratetype. So we should
2664 * adjust the count once.
2666 if (is_migrate_highatomic_page(page)) {
2668 * It should never happen but changes to
2669 * locking could inadvertently allow a per-cpu
2670 * drain to add pages to MIGRATE_HIGHATOMIC
2671 * while unreserving so be safe and watch for
2674 zone->nr_reserved_highatomic -= min(
2676 zone->nr_reserved_highatomic);
2680 * Convert to ac->migratetype and avoid the normal
2681 * pageblock stealing heuristics. Minimally, the caller
2682 * is doing the work and needs the pages. More
2683 * importantly, if the block was always converted to
2684 * MIGRATE_UNMOVABLE or another type then the number
2685 * of pageblocks that cannot be completely freed
2688 set_pageblock_migratetype(page, ac->migratetype);
2689 ret = move_freepages_block(zone, page, ac->migratetype,
2692 spin_unlock_irqrestore(&zone->lock, flags);
2696 spin_unlock_irqrestore(&zone->lock, flags);
2703 * Try finding a free buddy page on the fallback list and put it on the free
2704 * list of requested migratetype, possibly along with other pages from the same
2705 * block, depending on fragmentation avoidance heuristics. Returns true if
2706 * fallback was found so that __rmqueue_smallest() can grab it.
2708 * The use of signed ints for order and current_order is a deliberate
2709 * deviation from the rest of this file, to make the for loop
2710 * condition simpler.
2712 static __always_inline bool
2713 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2714 unsigned int alloc_flags)
2716 struct free_area *area;
2718 int min_order = order;
2724 * Do not steal pages from freelists belonging to other pageblocks
2725 * i.e. orders < pageblock_order. If there are no local zones free,
2726 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2728 if (alloc_flags & ALLOC_NOFRAGMENT)
2729 min_order = pageblock_order;
2732 * Find the largest available free page in the other list. This roughly
2733 * approximates finding the pageblock with the most free pages, which
2734 * would be too costly to do exactly.
2736 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2738 area = &(zone->free_area[current_order]);
2739 fallback_mt = find_suitable_fallback(area, current_order,
2740 start_migratetype, false, &can_steal);
2741 if (fallback_mt == -1)
2745 * We cannot steal all free pages from the pageblock and the
2746 * requested migratetype is movable. In that case it's better to
2747 * steal and split the smallest available page instead of the
2748 * largest available page, because even if the next movable
2749 * allocation falls back into a different pageblock than this
2750 * one, it won't cause permanent fragmentation.
2752 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2753 && current_order > order)
2762 for (current_order = order; current_order < MAX_ORDER;
2764 area = &(zone->free_area[current_order]);
2765 fallback_mt = find_suitable_fallback(area, current_order,
2766 start_migratetype, false, &can_steal);
2767 if (fallback_mt != -1)
2772 * This should not happen - we already found a suitable fallback
2773 * when looking for the largest page.
2775 VM_BUG_ON(current_order == MAX_ORDER);
2778 page = get_page_from_free_area(area, fallback_mt);
2780 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2783 trace_mm_page_alloc_extfrag(page, order, current_order,
2784 start_migratetype, fallback_mt);
2791 * Do the hard work of removing an element from the buddy allocator.
2792 * Call me with the zone->lock already held.
2794 static __always_inline struct page *
2795 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2796 unsigned int alloc_flags)
2802 * Balance movable allocations between regular and CMA areas by
2803 * allocating from CMA when over half of the zone's free memory
2804 * is in the CMA area.
2806 if (alloc_flags & ALLOC_CMA &&
2807 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2808 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2809 page = __rmqueue_cma_fallback(zone, order);
2815 page = __rmqueue_smallest(zone, order, migratetype);
2816 if (unlikely(!page)) {
2817 if (alloc_flags & ALLOC_CMA)
2818 page = __rmqueue_cma_fallback(zone, order);
2820 if (!page && __rmqueue_fallback(zone, order, migratetype,
2825 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2830 * Obtain a specified number of elements from the buddy allocator, all under
2831 * a single hold of the lock, for efficiency. Add them to the supplied list.
2832 * Returns the number of new pages which were placed at *list.
2834 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2835 unsigned long count, struct list_head *list,
2836 int migratetype, unsigned int alloc_flags)
2840 spin_lock(&zone->lock);
2841 for (i = 0; i < count; ++i) {
2842 struct page *page = __rmqueue(zone, order, migratetype,
2844 if (unlikely(page == NULL))
2847 if (unlikely(check_pcp_refill(page)))
2851 * Split buddy pages returned by expand() are received here in
2852 * physical page order. The page is added to the tail of
2853 * caller's list. From the callers perspective, the linked list
2854 * is ordered by page number under some conditions. This is
2855 * useful for IO devices that can forward direction from the
2856 * head, thus also in the physical page order. This is useful
2857 * for IO devices that can merge IO requests if the physical
2858 * pages are ordered properly.
2860 list_add_tail(&page->lru, list);
2862 if (is_migrate_cma(get_pcppage_migratetype(page)))
2863 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2868 * i pages were removed from the buddy list even if some leak due
2869 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2870 * on i. Do not confuse with 'alloced' which is the number of
2871 * pages added to the pcp list.
2873 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2874 spin_unlock(&zone->lock);
2880 * Called from the vmstat counter updater to drain pagesets of this
2881 * currently executing processor on remote nodes after they have
2884 * Note that this function must be called with the thread pinned to
2885 * a single processor.
2887 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2889 unsigned long flags;
2890 int to_drain, batch;
2892 local_irq_save(flags);
2893 batch = READ_ONCE(pcp->batch);
2894 to_drain = min(pcp->count, batch);
2896 free_pcppages_bulk(zone, to_drain, pcp);
2897 local_irq_restore(flags);
2902 * Drain pcplists of the indicated processor and zone.
2904 * The processor must either be the current processor and the
2905 * thread pinned to the current processor or a processor that
2908 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2910 unsigned long flags;
2911 struct per_cpu_pageset *pset;
2912 struct per_cpu_pages *pcp;
2914 local_irq_save(flags);
2915 pset = per_cpu_ptr(zone->pageset, cpu);
2919 free_pcppages_bulk(zone, pcp->count, pcp);
2920 local_irq_restore(flags);
2924 * Drain pcplists of all zones on the indicated processor.
2926 * The processor must either be the current processor and the
2927 * thread pinned to the current processor or a processor that
2930 static void drain_pages(unsigned int cpu)
2934 for_each_populated_zone(zone) {
2935 drain_pages_zone(cpu, zone);
2940 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2942 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2943 * the single zone's pages.
2945 void drain_local_pages(struct zone *zone)
2947 int cpu = smp_processor_id();
2950 drain_pages_zone(cpu, zone);
2955 static void drain_local_pages_wq(struct work_struct *work)
2957 struct pcpu_drain *drain;
2959 drain = container_of(work, struct pcpu_drain, work);
2962 * drain_all_pages doesn't use proper cpu hotplug protection so
2963 * we can race with cpu offline when the WQ can move this from
2964 * a cpu pinned worker to an unbound one. We can operate on a different
2965 * cpu which is allright but we also have to make sure to not move to
2969 drain_local_pages(drain->zone);
2974 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2976 * When zone parameter is non-NULL, spill just the single zone's pages.
2978 * Note that this can be extremely slow as the draining happens in a workqueue.
2980 void drain_all_pages(struct zone *zone)
2985 * Allocate in the BSS so we wont require allocation in
2986 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2988 static cpumask_t cpus_with_pcps;
2991 * Make sure nobody triggers this path before mm_percpu_wq is fully
2994 if (WARN_ON_ONCE(!mm_percpu_wq))
2998 * Do not drain if one is already in progress unless it's specific to
2999 * a zone. Such callers are primarily CMA and memory hotplug and need
3000 * the drain to be complete when the call returns.
3002 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3005 mutex_lock(&pcpu_drain_mutex);
3009 * We don't care about racing with CPU hotplug event
3010 * as offline notification will cause the notified
3011 * cpu to drain that CPU pcps and on_each_cpu_mask
3012 * disables preemption as part of its processing
3014 for_each_online_cpu(cpu) {
3015 struct per_cpu_pageset *pcp;
3017 bool has_pcps = false;
3020 pcp = per_cpu_ptr(zone->pageset, cpu);
3024 for_each_populated_zone(z) {
3025 pcp = per_cpu_ptr(z->pageset, cpu);
3026 if (pcp->pcp.count) {
3034 cpumask_set_cpu(cpu, &cpus_with_pcps);
3036 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3039 for_each_cpu(cpu, &cpus_with_pcps) {
3040 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3043 INIT_WORK(&drain->work, drain_local_pages_wq);
3044 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3046 for_each_cpu(cpu, &cpus_with_pcps)
3047 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3049 mutex_unlock(&pcpu_drain_mutex);
3052 #ifdef CONFIG_HIBERNATION
3055 * Touch the watchdog for every WD_PAGE_COUNT pages.
3057 #define WD_PAGE_COUNT (128*1024)
3059 void mark_free_pages(struct zone *zone)
3061 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3062 unsigned long flags;
3063 unsigned int order, t;
3066 if (zone_is_empty(zone))
3069 spin_lock_irqsave(&zone->lock, flags);
3071 max_zone_pfn = zone_end_pfn(zone);
3072 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3073 if (pfn_valid(pfn)) {
3074 page = pfn_to_page(pfn);
3076 if (!--page_count) {
3077 touch_nmi_watchdog();
3078 page_count = WD_PAGE_COUNT;
3081 if (page_zone(page) != zone)
3084 if (!swsusp_page_is_forbidden(page))
3085 swsusp_unset_page_free(page);
3088 for_each_migratetype_order(order, t) {
3089 list_for_each_entry(page,
3090 &zone->free_area[order].free_list[t], lru) {
3093 pfn = page_to_pfn(page);
3094 for (i = 0; i < (1UL << order); i++) {
3095 if (!--page_count) {
3096 touch_nmi_watchdog();
3097 page_count = WD_PAGE_COUNT;
3099 swsusp_set_page_free(pfn_to_page(pfn + i));
3103 spin_unlock_irqrestore(&zone->lock, flags);
3105 #endif /* CONFIG_PM */
3107 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3111 if (!free_pcp_prepare(page))
3114 migratetype = get_pfnblock_migratetype(page, pfn);
3115 set_pcppage_migratetype(page, migratetype);
3119 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3121 struct zone *zone = page_zone(page);
3122 struct per_cpu_pages *pcp;
3125 migratetype = get_pcppage_migratetype(page);
3126 __count_vm_event(PGFREE);
3129 * We only track unmovable, reclaimable and movable on pcp lists.
3130 * Free ISOLATE pages back to the allocator because they are being
3131 * offlined but treat HIGHATOMIC as movable pages so we can get those
3132 * areas back if necessary. Otherwise, we may have to free
3133 * excessively into the page allocator
3135 if (migratetype >= MIGRATE_PCPTYPES) {
3136 if (unlikely(is_migrate_isolate(migratetype))) {
3137 free_one_page(zone, page, pfn, 0, migratetype);
3140 migratetype = MIGRATE_MOVABLE;
3143 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3144 list_add(&page->lru, &pcp->lists[migratetype]);
3146 if (pcp->count >= pcp->high) {
3147 unsigned long batch = READ_ONCE(pcp->batch);
3148 free_pcppages_bulk(zone, batch, pcp);
3153 * Free a 0-order page
3155 void free_unref_page(struct page *page)
3157 unsigned long flags;
3158 unsigned long pfn = page_to_pfn(page);
3160 if (!free_unref_page_prepare(page, pfn))
3163 local_irq_save(flags);
3164 free_unref_page_commit(page, pfn);
3165 local_irq_restore(flags);
3169 * Free a list of 0-order pages
3171 void free_unref_page_list(struct list_head *list)
3173 struct page *page, *next;
3174 unsigned long flags, pfn;
3175 int batch_count = 0;
3177 /* Prepare pages for freeing */
3178 list_for_each_entry_safe(page, next, list, lru) {
3179 pfn = page_to_pfn(page);
3180 if (!free_unref_page_prepare(page, pfn))
3181 list_del(&page->lru);
3182 set_page_private(page, pfn);
3185 local_irq_save(flags);
3186 list_for_each_entry_safe(page, next, list, lru) {
3187 unsigned long pfn = page_private(page);
3189 set_page_private(page, 0);
3190 trace_mm_page_free_batched(page);
3191 free_unref_page_commit(page, pfn);
3194 * Guard against excessive IRQ disabled times when we get
3195 * a large list of pages to free.
3197 if (++batch_count == SWAP_CLUSTER_MAX) {
3198 local_irq_restore(flags);
3200 local_irq_save(flags);
3203 local_irq_restore(flags);
3207 * split_page takes a non-compound higher-order page, and splits it into
3208 * n (1<<order) sub-pages: page[0..n]
3209 * Each sub-page must be freed individually.
3211 * Note: this is probably too low level an operation for use in drivers.
3212 * Please consult with lkml before using this in your driver.
3214 void split_page(struct page *page, unsigned int order)
3218 VM_BUG_ON_PAGE(PageCompound(page), page);
3219 VM_BUG_ON_PAGE(!page_count(page), page);
3221 for (i = 1; i < (1 << order); i++)
3222 set_page_refcounted(page + i);
3223 split_page_owner(page, 1 << order);
3225 EXPORT_SYMBOL_GPL(split_page);
3227 int __isolate_free_page(struct page *page, unsigned int order)
3229 unsigned long watermark;
3233 BUG_ON(!PageBuddy(page));
3235 zone = page_zone(page);
3236 mt = get_pageblock_migratetype(page);
3238 if (!is_migrate_isolate(mt)) {
3240 * Obey watermarks as if the page was being allocated. We can
3241 * emulate a high-order watermark check with a raised order-0
3242 * watermark, because we already know our high-order page
3245 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3246 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3249 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3252 /* Remove page from free list */
3254 del_page_from_free_list(page, zone, order);
3257 * Set the pageblock if the isolated page is at least half of a
3260 if (order >= pageblock_order - 1) {
3261 struct page *endpage = page + (1 << order) - 1;
3262 for (; page < endpage; page += pageblock_nr_pages) {
3263 int mt = get_pageblock_migratetype(page);
3264 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3265 && !is_migrate_highatomic(mt))
3266 set_pageblock_migratetype(page,
3272 return 1UL << order;
3276 * __putback_isolated_page - Return a now-isolated page back where we got it
3277 * @page: Page that was isolated
3278 * @order: Order of the isolated page
3279 * @mt: The page's pageblock's migratetype
3281 * This function is meant to return a page pulled from the free lists via
3282 * __isolate_free_page back to the free lists they were pulled from.
3284 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3286 struct zone *zone = page_zone(page);
3288 /* zone lock should be held when this function is called */
3289 lockdep_assert_held(&zone->lock);
3291 /* Return isolated page to tail of freelist. */
3292 __free_one_page(page, page_to_pfn(page), zone, order, mt, false);
3296 * Update NUMA hit/miss statistics
3298 * Must be called with interrupts disabled.
3300 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3303 enum numa_stat_item local_stat = NUMA_LOCAL;
3305 /* skip numa counters update if numa stats is disabled */
3306 if (!static_branch_likely(&vm_numa_stat_key))
3309 if (zone_to_nid(z) != numa_node_id())
3310 local_stat = NUMA_OTHER;
3312 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3313 __inc_numa_state(z, NUMA_HIT);
3315 __inc_numa_state(z, NUMA_MISS);
3316 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3318 __inc_numa_state(z, local_stat);
3322 /* Remove page from the per-cpu list, caller must protect the list */
3323 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3324 unsigned int alloc_flags,
3325 struct per_cpu_pages *pcp,
3326 struct list_head *list)
3331 if (list_empty(list)) {
3332 pcp->count += rmqueue_bulk(zone, 0,
3334 migratetype, alloc_flags);
3335 if (unlikely(list_empty(list)))
3339 page = list_first_entry(list, struct page, lru);
3340 list_del(&page->lru);
3342 } while (check_new_pcp(page));
3347 /* Lock and remove page from the per-cpu list */
3348 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3349 struct zone *zone, gfp_t gfp_flags,
3350 int migratetype, unsigned int alloc_flags)
3352 struct per_cpu_pages *pcp;
3353 struct list_head *list;
3355 unsigned long flags;
3357 local_irq_save(flags);
3358 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3359 list = &pcp->lists[migratetype];
3360 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3362 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3363 zone_statistics(preferred_zone, zone);
3365 local_irq_restore(flags);
3370 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3373 struct page *rmqueue(struct zone *preferred_zone,
3374 struct zone *zone, unsigned int order,
3375 gfp_t gfp_flags, unsigned int alloc_flags,
3378 unsigned long flags;
3381 if (likely(order == 0)) {
3383 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3384 * we need to skip it when CMA area isn't allowed.
3386 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3387 migratetype != MIGRATE_MOVABLE) {
3388 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3389 migratetype, alloc_flags);
3395 * We most definitely don't want callers attempting to
3396 * allocate greater than order-1 page units with __GFP_NOFAIL.
3398 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3399 spin_lock_irqsave(&zone->lock, flags);
3404 * order-0 request can reach here when the pcplist is skipped
3405 * due to non-CMA allocation context. HIGHATOMIC area is
3406 * reserved for high-order atomic allocation, so order-0
3407 * request should skip it.
3409 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3410 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3412 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3415 page = __rmqueue(zone, order, migratetype, alloc_flags);
3416 } while (page && check_new_pages(page, order));
3417 spin_unlock(&zone->lock);
3420 __mod_zone_freepage_state(zone, -(1 << order),
3421 get_pcppage_migratetype(page));
3423 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3424 zone_statistics(preferred_zone, zone);
3425 local_irq_restore(flags);
3428 /* Separate test+clear to avoid unnecessary atomics */
3429 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3430 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3431 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3434 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3438 local_irq_restore(flags);
3442 #ifdef CONFIG_FAIL_PAGE_ALLOC
3445 struct fault_attr attr;
3447 bool ignore_gfp_highmem;
3448 bool ignore_gfp_reclaim;
3450 } fail_page_alloc = {
3451 .attr = FAULT_ATTR_INITIALIZER,
3452 .ignore_gfp_reclaim = true,
3453 .ignore_gfp_highmem = true,
3457 static int __init setup_fail_page_alloc(char *str)
3459 return setup_fault_attr(&fail_page_alloc.attr, str);
3461 __setup("fail_page_alloc=", setup_fail_page_alloc);
3463 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3465 if (order < fail_page_alloc.min_order)
3467 if (gfp_mask & __GFP_NOFAIL)
3469 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3471 if (fail_page_alloc.ignore_gfp_reclaim &&
3472 (gfp_mask & __GFP_DIRECT_RECLAIM))
3475 return should_fail(&fail_page_alloc.attr, 1 << order);
3478 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3480 static int __init fail_page_alloc_debugfs(void)
3482 umode_t mode = S_IFREG | 0600;
3485 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3486 &fail_page_alloc.attr);
3488 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3489 &fail_page_alloc.ignore_gfp_reclaim);
3490 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3491 &fail_page_alloc.ignore_gfp_highmem);
3492 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3497 late_initcall(fail_page_alloc_debugfs);
3499 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3501 #else /* CONFIG_FAIL_PAGE_ALLOC */
3503 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3508 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3510 static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3512 return __should_fail_alloc_page(gfp_mask, order);
3514 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3516 static inline long __zone_watermark_unusable_free(struct zone *z,
3517 unsigned int order, unsigned int alloc_flags)
3519 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3520 long unusable_free = (1 << order) - 1;
3523 * If the caller does not have rights to ALLOC_HARDER then subtract
3524 * the high-atomic reserves. This will over-estimate the size of the
3525 * atomic reserve but it avoids a search.
3527 if (likely(!alloc_harder))
3528 unusable_free += z->nr_reserved_highatomic;
3531 /* If allocation can't use CMA areas don't use free CMA pages */
3532 if (!(alloc_flags & ALLOC_CMA))
3533 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3536 return unusable_free;
3540 * Return true if free base pages are above 'mark'. For high-order checks it
3541 * will return true of the order-0 watermark is reached and there is at least
3542 * one free page of a suitable size. Checking now avoids taking the zone lock
3543 * to check in the allocation paths if no pages are free.
3545 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3546 int highest_zoneidx, unsigned int alloc_flags,
3551 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3553 /* free_pages may go negative - that's OK */
3554 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3556 if (alloc_flags & ALLOC_HIGH)
3559 if (unlikely(alloc_harder)) {
3561 * OOM victims can try even harder than normal ALLOC_HARDER
3562 * users on the grounds that it's definitely going to be in
3563 * the exit path shortly and free memory. Any allocation it
3564 * makes during the free path will be small and short-lived.
3566 if (alloc_flags & ALLOC_OOM)
3573 * Check watermarks for an order-0 allocation request. If these
3574 * are not met, then a high-order request also cannot go ahead
3575 * even if a suitable page happened to be free.
3577 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3580 /* If this is an order-0 request then the watermark is fine */
3584 /* For a high-order request, check at least one suitable page is free */
3585 for (o = order; o < MAX_ORDER; o++) {
3586 struct free_area *area = &z->free_area[o];
3592 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3593 if (!free_area_empty(area, mt))
3598 if ((alloc_flags & ALLOC_CMA) &&
3599 !free_area_empty(area, MIGRATE_CMA)) {
3603 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3609 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3610 int highest_zoneidx, unsigned int alloc_flags)
3612 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3613 zone_page_state(z, NR_FREE_PAGES));
3616 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3617 unsigned long mark, int highest_zoneidx,
3618 unsigned int alloc_flags, gfp_t gfp_mask)
3622 free_pages = zone_page_state(z, NR_FREE_PAGES);
3625 * Fast check for order-0 only. If this fails then the reserves
3626 * need to be calculated.
3631 fast_free = free_pages;
3632 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3633 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3637 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3641 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3642 * when checking the min watermark. The min watermark is the
3643 * point where boosting is ignored so that kswapd is woken up
3644 * when below the low watermark.
3646 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3647 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3648 mark = z->_watermark[WMARK_MIN];
3649 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3650 alloc_flags, free_pages);
3656 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3657 unsigned long mark, int highest_zoneidx)
3659 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3661 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3662 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3664 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3669 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3671 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3672 node_reclaim_distance;
3674 #else /* CONFIG_NUMA */
3675 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3679 #endif /* CONFIG_NUMA */
3682 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3683 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3684 * premature use of a lower zone may cause lowmem pressure problems that
3685 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3686 * probably too small. It only makes sense to spread allocations to avoid
3687 * fragmentation between the Normal and DMA32 zones.
3689 static inline unsigned int
3690 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3692 unsigned int alloc_flags;
3695 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3698 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3700 #ifdef CONFIG_ZONE_DMA32
3704 if (zone_idx(zone) != ZONE_NORMAL)
3708 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3709 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3710 * on UMA that if Normal is populated then so is DMA32.
3712 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3713 if (nr_online_nodes > 1 && !populated_zone(--zone))
3716 alloc_flags |= ALLOC_NOFRAGMENT;
3717 #endif /* CONFIG_ZONE_DMA32 */
3721 static inline unsigned int current_alloc_flags(gfp_t gfp_mask,
3722 unsigned int alloc_flags)
3725 unsigned int pflags = current->flags;
3727 if (!(pflags & PF_MEMALLOC_NOCMA) &&
3728 gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3729 alloc_flags |= ALLOC_CMA;
3736 * get_page_from_freelist goes through the zonelist trying to allocate
3739 static struct page *
3740 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3741 const struct alloc_context *ac)
3745 struct pglist_data *last_pgdat_dirty_limit = NULL;
3750 * Scan zonelist, looking for a zone with enough free.
3751 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3753 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3754 z = ac->preferred_zoneref;
3755 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3760 if (cpusets_enabled() &&
3761 (alloc_flags & ALLOC_CPUSET) &&
3762 !__cpuset_zone_allowed(zone, gfp_mask))
3765 * When allocating a page cache page for writing, we
3766 * want to get it from a node that is within its dirty
3767 * limit, such that no single node holds more than its
3768 * proportional share of globally allowed dirty pages.
3769 * The dirty limits take into account the node's
3770 * lowmem reserves and high watermark so that kswapd
3771 * should be able to balance it without having to
3772 * write pages from its LRU list.
3774 * XXX: For now, allow allocations to potentially
3775 * exceed the per-node dirty limit in the slowpath
3776 * (spread_dirty_pages unset) before going into reclaim,
3777 * which is important when on a NUMA setup the allowed
3778 * nodes are together not big enough to reach the
3779 * global limit. The proper fix for these situations
3780 * will require awareness of nodes in the
3781 * dirty-throttling and the flusher threads.
3783 if (ac->spread_dirty_pages) {
3784 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3787 if (!node_dirty_ok(zone->zone_pgdat)) {
3788 last_pgdat_dirty_limit = zone->zone_pgdat;
3793 if (no_fallback && nr_online_nodes > 1 &&
3794 zone != ac->preferred_zoneref->zone) {
3798 * If moving to a remote node, retry but allow
3799 * fragmenting fallbacks. Locality is more important
3800 * than fragmentation avoidance.
3802 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3803 if (zone_to_nid(zone) != local_nid) {
3804 alloc_flags &= ~ALLOC_NOFRAGMENT;
3809 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3810 if (!zone_watermark_fast(zone, order, mark,
3811 ac->highest_zoneidx, alloc_flags,
3815 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3817 * Watermark failed for this zone, but see if we can
3818 * grow this zone if it contains deferred pages.
3820 if (static_branch_unlikely(&deferred_pages)) {
3821 if (_deferred_grow_zone(zone, order))
3825 /* Checked here to keep the fast path fast */
3826 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3827 if (alloc_flags & ALLOC_NO_WATERMARKS)
3830 if (node_reclaim_mode == 0 ||
3831 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3834 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3836 case NODE_RECLAIM_NOSCAN:
3839 case NODE_RECLAIM_FULL:
3840 /* scanned but unreclaimable */
3843 /* did we reclaim enough */
3844 if (zone_watermark_ok(zone, order, mark,
3845 ac->highest_zoneidx, alloc_flags))
3853 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3854 gfp_mask, alloc_flags, ac->migratetype);
3856 prep_new_page(page, order, gfp_mask, alloc_flags);
3859 * If this is a high-order atomic allocation then check
3860 * if the pageblock should be reserved for the future
3862 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3863 reserve_highatomic_pageblock(page, zone, order);
3867 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3868 /* Try again if zone has deferred pages */
3869 if (static_branch_unlikely(&deferred_pages)) {
3870 if (_deferred_grow_zone(zone, order))
3878 * It's possible on a UMA machine to get through all zones that are
3879 * fragmented. If avoiding fragmentation, reset and try again.
3882 alloc_flags &= ~ALLOC_NOFRAGMENT;
3889 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3891 unsigned int filter = SHOW_MEM_FILTER_NODES;
3894 * This documents exceptions given to allocations in certain
3895 * contexts that are allowed to allocate outside current's set
3898 if (!(gfp_mask & __GFP_NOMEMALLOC))
3899 if (tsk_is_oom_victim(current) ||
3900 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3901 filter &= ~SHOW_MEM_FILTER_NODES;
3902 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3903 filter &= ~SHOW_MEM_FILTER_NODES;
3905 show_mem(filter, nodemask);
3908 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3910 struct va_format vaf;
3912 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3914 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3917 va_start(args, fmt);
3920 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3921 current->comm, &vaf, gfp_mask, &gfp_mask,
3922 nodemask_pr_args(nodemask));
3925 cpuset_print_current_mems_allowed();
3928 warn_alloc_show_mem(gfp_mask, nodemask);
3931 static inline struct page *
3932 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3933 unsigned int alloc_flags,
3934 const struct alloc_context *ac)
3938 page = get_page_from_freelist(gfp_mask, order,
3939 alloc_flags|ALLOC_CPUSET, ac);
3941 * fallback to ignore cpuset restriction if our nodes
3945 page = get_page_from_freelist(gfp_mask, order,
3951 static inline struct page *
3952 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3953 const struct alloc_context *ac, unsigned long *did_some_progress)
3955 struct oom_control oc = {
3956 .zonelist = ac->zonelist,
3957 .nodemask = ac->nodemask,
3959 .gfp_mask = gfp_mask,
3964 *did_some_progress = 0;
3967 * Acquire the oom lock. If that fails, somebody else is
3968 * making progress for us.
3970 if (!mutex_trylock(&oom_lock)) {
3971 *did_some_progress = 1;
3972 schedule_timeout_uninterruptible(1);
3977 * Go through the zonelist yet one more time, keep very high watermark
3978 * here, this is only to catch a parallel oom killing, we must fail if
3979 * we're still under heavy pressure. But make sure that this reclaim
3980 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3981 * allocation which will never fail due to oom_lock already held.
3983 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3984 ~__GFP_DIRECT_RECLAIM, order,
3985 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3989 /* Coredumps can quickly deplete all memory reserves */
3990 if (current->flags & PF_DUMPCORE)
3992 /* The OOM killer will not help higher order allocs */
3993 if (order > PAGE_ALLOC_COSTLY_ORDER)
3996 * We have already exhausted all our reclaim opportunities without any
3997 * success so it is time to admit defeat. We will skip the OOM killer
3998 * because it is very likely that the caller has a more reasonable
3999 * fallback than shooting a random task.
4001 * The OOM killer may not free memory on a specific node.
4003 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4005 /* The OOM killer does not needlessly kill tasks for lowmem */
4006 if (ac->highest_zoneidx < ZONE_NORMAL)
4008 if (pm_suspended_storage())
4011 * XXX: GFP_NOFS allocations should rather fail than rely on
4012 * other request to make a forward progress.
4013 * We are in an unfortunate situation where out_of_memory cannot
4014 * do much for this context but let's try it to at least get
4015 * access to memory reserved if the current task is killed (see
4016 * out_of_memory). Once filesystems are ready to handle allocation
4017 * failures more gracefully we should just bail out here.
4020 /* Exhausted what can be done so it's blame time */
4021 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4022 *did_some_progress = 1;
4025 * Help non-failing allocations by giving them access to memory
4028 if (gfp_mask & __GFP_NOFAIL)
4029 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4030 ALLOC_NO_WATERMARKS, ac);
4033 mutex_unlock(&oom_lock);
4038 * Maximum number of compaction retries wit a progress before OOM
4039 * killer is consider as the only way to move forward.
4041 #define MAX_COMPACT_RETRIES 16
4043 #ifdef CONFIG_COMPACTION
4044 /* Try memory compaction for high-order allocations before reclaim */
4045 static struct page *
4046 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4047 unsigned int alloc_flags, const struct alloc_context *ac,
4048 enum compact_priority prio, enum compact_result *compact_result)
4050 struct page *page = NULL;
4051 unsigned long pflags;
4052 unsigned int noreclaim_flag;
4057 psi_memstall_enter(&pflags);
4058 noreclaim_flag = memalloc_noreclaim_save();
4060 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4063 memalloc_noreclaim_restore(noreclaim_flag);
4064 psi_memstall_leave(&pflags);
4067 * At least in one zone compaction wasn't deferred or skipped, so let's
4068 * count a compaction stall
4070 count_vm_event(COMPACTSTALL);
4072 /* Prep a captured page if available */
4074 prep_new_page(page, order, gfp_mask, alloc_flags);
4076 /* Try get a page from the freelist if available */
4078 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4081 struct zone *zone = page_zone(page);
4083 zone->compact_blockskip_flush = false;
4084 compaction_defer_reset(zone, order, true);
4085 count_vm_event(COMPACTSUCCESS);
4090 * It's bad if compaction run occurs and fails. The most likely reason
4091 * is that pages exist, but not enough to satisfy watermarks.
4093 count_vm_event(COMPACTFAIL);
4101 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4102 enum compact_result compact_result,
4103 enum compact_priority *compact_priority,
4104 int *compaction_retries)
4106 int max_retries = MAX_COMPACT_RETRIES;
4109 int retries = *compaction_retries;
4110 enum compact_priority priority = *compact_priority;
4115 if (compaction_made_progress(compact_result))
4116 (*compaction_retries)++;
4119 * compaction considers all the zone as desperately out of memory
4120 * so it doesn't really make much sense to retry except when the
4121 * failure could be caused by insufficient priority
4123 if (compaction_failed(compact_result))
4124 goto check_priority;
4127 * compaction was skipped because there are not enough order-0 pages
4128 * to work with, so we retry only if it looks like reclaim can help.
4130 if (compaction_needs_reclaim(compact_result)) {
4131 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4136 * make sure the compaction wasn't deferred or didn't bail out early
4137 * due to locks contention before we declare that we should give up.
4138 * But the next retry should use a higher priority if allowed, so
4139 * we don't just keep bailing out endlessly.
4141 if (compaction_withdrawn(compact_result)) {
4142 goto check_priority;
4146 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4147 * costly ones because they are de facto nofail and invoke OOM
4148 * killer to move on while costly can fail and users are ready
4149 * to cope with that. 1/4 retries is rather arbitrary but we
4150 * would need much more detailed feedback from compaction to
4151 * make a better decision.
4153 if (order > PAGE_ALLOC_COSTLY_ORDER)
4155 if (*compaction_retries <= max_retries) {
4161 * Make sure there are attempts at the highest priority if we exhausted
4162 * all retries or failed at the lower priorities.
4165 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4166 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4168 if (*compact_priority > min_priority) {
4169 (*compact_priority)--;
4170 *compaction_retries = 0;
4174 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4178 static inline struct page *
4179 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4180 unsigned int alloc_flags, const struct alloc_context *ac,
4181 enum compact_priority prio, enum compact_result *compact_result)
4183 *compact_result = COMPACT_SKIPPED;
4188 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4189 enum compact_result compact_result,
4190 enum compact_priority *compact_priority,
4191 int *compaction_retries)
4196 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4200 * There are setups with compaction disabled which would prefer to loop
4201 * inside the allocator rather than hit the oom killer prematurely.
4202 * Let's give them a good hope and keep retrying while the order-0
4203 * watermarks are OK.
4205 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4206 ac->highest_zoneidx, ac->nodemask) {
4207 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4208 ac->highest_zoneidx, alloc_flags))
4213 #endif /* CONFIG_COMPACTION */
4215 #ifdef CONFIG_LOCKDEP
4216 static struct lockdep_map __fs_reclaim_map =
4217 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4219 static bool __need_fs_reclaim(gfp_t gfp_mask)
4221 gfp_mask = current_gfp_context(gfp_mask);
4223 /* no reclaim without waiting on it */
4224 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4227 /* this guy won't enter reclaim */
4228 if (current->flags & PF_MEMALLOC)
4231 /* We're only interested __GFP_FS allocations for now */
4232 if (!(gfp_mask & __GFP_FS))
4235 if (gfp_mask & __GFP_NOLOCKDEP)
4241 void __fs_reclaim_acquire(void)
4243 lock_map_acquire(&__fs_reclaim_map);
4246 void __fs_reclaim_release(void)
4248 lock_map_release(&__fs_reclaim_map);
4251 void fs_reclaim_acquire(gfp_t gfp_mask)
4253 if (__need_fs_reclaim(gfp_mask))
4254 __fs_reclaim_acquire();
4256 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4258 void fs_reclaim_release(gfp_t gfp_mask)
4260 if (__need_fs_reclaim(gfp_mask))
4261 __fs_reclaim_release();
4263 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4266 /* Perform direct synchronous page reclaim */
4267 static unsigned long
4268 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4269 const struct alloc_context *ac)
4271 unsigned int noreclaim_flag;
4272 unsigned long pflags, progress;
4276 /* We now go into synchronous reclaim */
4277 cpuset_memory_pressure_bump();
4278 psi_memstall_enter(&pflags);
4279 fs_reclaim_acquire(gfp_mask);
4280 noreclaim_flag = memalloc_noreclaim_save();
4282 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4285 memalloc_noreclaim_restore(noreclaim_flag);
4286 fs_reclaim_release(gfp_mask);
4287 psi_memstall_leave(&pflags);
4294 /* The really slow allocator path where we enter direct reclaim */
4295 static inline struct page *
4296 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4297 unsigned int alloc_flags, const struct alloc_context *ac,
4298 unsigned long *did_some_progress)
4300 struct page *page = NULL;
4301 bool drained = false;
4303 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4304 if (unlikely(!(*did_some_progress)))
4308 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4311 * If an allocation failed after direct reclaim, it could be because
4312 * pages are pinned on the per-cpu lists or in high alloc reserves.
4313 * Shrink them and try again
4315 if (!page && !drained) {
4316 unreserve_highatomic_pageblock(ac, false);
4317 drain_all_pages(NULL);
4325 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4326 const struct alloc_context *ac)
4330 pg_data_t *last_pgdat = NULL;
4331 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4333 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4335 if (last_pgdat != zone->zone_pgdat)
4336 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4337 last_pgdat = zone->zone_pgdat;
4341 static inline unsigned int
4342 gfp_to_alloc_flags(gfp_t gfp_mask)
4344 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4347 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4348 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4349 * to save two branches.
4351 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4352 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4355 * The caller may dip into page reserves a bit more if the caller
4356 * cannot run direct reclaim, or if the caller has realtime scheduling
4357 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4358 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4360 alloc_flags |= (__force int)
4361 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4363 if (gfp_mask & __GFP_ATOMIC) {
4365 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4366 * if it can't schedule.
4368 if (!(gfp_mask & __GFP_NOMEMALLOC))
4369 alloc_flags |= ALLOC_HARDER;
4371 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4372 * comment for __cpuset_node_allowed().
4374 alloc_flags &= ~ALLOC_CPUSET;
4375 } else if (unlikely(rt_task(current)) && !in_interrupt())
4376 alloc_flags |= ALLOC_HARDER;
4378 alloc_flags = current_alloc_flags(gfp_mask, alloc_flags);
4383 static bool oom_reserves_allowed(struct task_struct *tsk)
4385 if (!tsk_is_oom_victim(tsk))
4389 * !MMU doesn't have oom reaper so give access to memory reserves
4390 * only to the thread with TIF_MEMDIE set
4392 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4399 * Distinguish requests which really need access to full memory
4400 * reserves from oom victims which can live with a portion of it
4402 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4404 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4406 if (gfp_mask & __GFP_MEMALLOC)
4407 return ALLOC_NO_WATERMARKS;
4408 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4409 return ALLOC_NO_WATERMARKS;
4410 if (!in_interrupt()) {
4411 if (current->flags & PF_MEMALLOC)
4412 return ALLOC_NO_WATERMARKS;
4413 else if (oom_reserves_allowed(current))
4420 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4422 return !!__gfp_pfmemalloc_flags(gfp_mask);
4426 * Checks whether it makes sense to retry the reclaim to make a forward progress
4427 * for the given allocation request.
4429 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4430 * without success, or when we couldn't even meet the watermark if we
4431 * reclaimed all remaining pages on the LRU lists.
4433 * Returns true if a retry is viable or false to enter the oom path.
4436 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4437 struct alloc_context *ac, int alloc_flags,
4438 bool did_some_progress, int *no_progress_loops)
4445 * Costly allocations might have made a progress but this doesn't mean
4446 * their order will become available due to high fragmentation so
4447 * always increment the no progress counter for them
4449 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4450 *no_progress_loops = 0;
4452 (*no_progress_loops)++;
4455 * Make sure we converge to OOM if we cannot make any progress
4456 * several times in the row.
4458 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4459 /* Before OOM, exhaust highatomic_reserve */
4460 return unreserve_highatomic_pageblock(ac, true);
4464 * Keep reclaiming pages while there is a chance this will lead
4465 * somewhere. If none of the target zones can satisfy our allocation
4466 * request even if all reclaimable pages are considered then we are
4467 * screwed and have to go OOM.
4469 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4470 ac->highest_zoneidx, ac->nodemask) {
4471 unsigned long available;
4472 unsigned long reclaimable;
4473 unsigned long min_wmark = min_wmark_pages(zone);
4476 available = reclaimable = zone_reclaimable_pages(zone);
4477 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4480 * Would the allocation succeed if we reclaimed all
4481 * reclaimable pages?
4483 wmark = __zone_watermark_ok(zone, order, min_wmark,
4484 ac->highest_zoneidx, alloc_flags, available);
4485 trace_reclaim_retry_zone(z, order, reclaimable,
4486 available, min_wmark, *no_progress_loops, wmark);
4489 * If we didn't make any progress and have a lot of
4490 * dirty + writeback pages then we should wait for
4491 * an IO to complete to slow down the reclaim and
4492 * prevent from pre mature OOM
4494 if (!did_some_progress) {
4495 unsigned long write_pending;
4497 write_pending = zone_page_state_snapshot(zone,
4498 NR_ZONE_WRITE_PENDING);
4500 if (2 * write_pending > reclaimable) {
4501 congestion_wait(BLK_RW_ASYNC, HZ/10);
4513 * Memory allocation/reclaim might be called from a WQ context and the
4514 * current implementation of the WQ concurrency control doesn't
4515 * recognize that a particular WQ is congested if the worker thread is
4516 * looping without ever sleeping. Therefore we have to do a short sleep
4517 * here rather than calling cond_resched().
4519 if (current->flags & PF_WQ_WORKER)
4520 schedule_timeout_uninterruptible(1);
4527 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4530 * It's possible that cpuset's mems_allowed and the nodemask from
4531 * mempolicy don't intersect. This should be normally dealt with by
4532 * policy_nodemask(), but it's possible to race with cpuset update in
4533 * such a way the check therein was true, and then it became false
4534 * before we got our cpuset_mems_cookie here.
4535 * This assumes that for all allocations, ac->nodemask can come only
4536 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4537 * when it does not intersect with the cpuset restrictions) or the
4538 * caller can deal with a violated nodemask.
4540 if (cpusets_enabled() && ac->nodemask &&
4541 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4542 ac->nodemask = NULL;
4547 * When updating a task's mems_allowed or mempolicy nodemask, it is
4548 * possible to race with parallel threads in such a way that our
4549 * allocation can fail while the mask is being updated. If we are about
4550 * to fail, check if the cpuset changed during allocation and if so,
4553 if (read_mems_allowed_retry(cpuset_mems_cookie))
4559 static inline struct page *
4560 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4561 struct alloc_context *ac)
4563 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4564 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4565 struct page *page = NULL;
4566 unsigned int alloc_flags;
4567 unsigned long did_some_progress;
4568 enum compact_priority compact_priority;
4569 enum compact_result compact_result;
4570 int compaction_retries;
4571 int no_progress_loops;
4572 unsigned int cpuset_mems_cookie;
4576 * We also sanity check to catch abuse of atomic reserves being used by
4577 * callers that are not in atomic context.
4579 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4580 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4581 gfp_mask &= ~__GFP_ATOMIC;
4584 compaction_retries = 0;
4585 no_progress_loops = 0;
4586 compact_priority = DEF_COMPACT_PRIORITY;
4587 cpuset_mems_cookie = read_mems_allowed_begin();
4590 * The fast path uses conservative alloc_flags to succeed only until
4591 * kswapd needs to be woken up, and to avoid the cost of setting up
4592 * alloc_flags precisely. So we do that now.
4594 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4597 * We need to recalculate the starting point for the zonelist iterator
4598 * because we might have used different nodemask in the fast path, or
4599 * there was a cpuset modification and we are retrying - otherwise we
4600 * could end up iterating over non-eligible zones endlessly.
4602 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4603 ac->highest_zoneidx, ac->nodemask);
4604 if (!ac->preferred_zoneref->zone)
4607 if (alloc_flags & ALLOC_KSWAPD)
4608 wake_all_kswapds(order, gfp_mask, ac);
4611 * The adjusted alloc_flags might result in immediate success, so try
4614 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4619 * For costly allocations, try direct compaction first, as it's likely
4620 * that we have enough base pages and don't need to reclaim. For non-
4621 * movable high-order allocations, do that as well, as compaction will
4622 * try prevent permanent fragmentation by migrating from blocks of the
4624 * Don't try this for allocations that are allowed to ignore
4625 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4627 if (can_direct_reclaim &&
4629 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4630 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4631 page = __alloc_pages_direct_compact(gfp_mask, order,
4633 INIT_COMPACT_PRIORITY,
4639 * Checks for costly allocations with __GFP_NORETRY, which
4640 * includes some THP page fault allocations
4642 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4644 * If allocating entire pageblock(s) and compaction
4645 * failed because all zones are below low watermarks
4646 * or is prohibited because it recently failed at this
4647 * order, fail immediately unless the allocator has
4648 * requested compaction and reclaim retry.
4651 * - potentially very expensive because zones are far
4652 * below their low watermarks or this is part of very
4653 * bursty high order allocations,
4654 * - not guaranteed to help because isolate_freepages()
4655 * may not iterate over freed pages as part of its
4657 * - unlikely to make entire pageblocks free on its
4660 if (compact_result == COMPACT_SKIPPED ||
4661 compact_result == COMPACT_DEFERRED)
4665 * Looks like reclaim/compaction is worth trying, but
4666 * sync compaction could be very expensive, so keep
4667 * using async compaction.
4669 compact_priority = INIT_COMPACT_PRIORITY;
4674 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4675 if (alloc_flags & ALLOC_KSWAPD)
4676 wake_all_kswapds(order, gfp_mask, ac);
4678 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4680 alloc_flags = current_alloc_flags(gfp_mask, reserve_flags);
4683 * Reset the nodemask and zonelist iterators if memory policies can be
4684 * ignored. These allocations are high priority and system rather than
4687 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4688 ac->nodemask = NULL;
4689 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4690 ac->highest_zoneidx, ac->nodemask);
4693 /* Attempt with potentially adjusted zonelist and alloc_flags */
4694 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4698 /* Caller is not willing to reclaim, we can't balance anything */
4699 if (!can_direct_reclaim)
4702 /* Avoid recursion of direct reclaim */
4703 if (current->flags & PF_MEMALLOC)
4706 /* Try direct reclaim and then allocating */
4707 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4708 &did_some_progress);
4712 /* Try direct compaction and then allocating */
4713 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4714 compact_priority, &compact_result);
4718 /* Do not loop if specifically requested */
4719 if (gfp_mask & __GFP_NORETRY)
4723 * Do not retry costly high order allocations unless they are
4724 * __GFP_RETRY_MAYFAIL
4726 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4729 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4730 did_some_progress > 0, &no_progress_loops))
4734 * It doesn't make any sense to retry for the compaction if the order-0
4735 * reclaim is not able to make any progress because the current
4736 * implementation of the compaction depends on the sufficient amount
4737 * of free memory (see __compaction_suitable)
4739 if (did_some_progress > 0 &&
4740 should_compact_retry(ac, order, alloc_flags,
4741 compact_result, &compact_priority,
4742 &compaction_retries))
4746 /* Deal with possible cpuset update races before we start OOM killing */
4747 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4750 /* Reclaim has failed us, start killing things */
4751 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4755 /* Avoid allocations with no watermarks from looping endlessly */
4756 if (tsk_is_oom_victim(current) &&
4757 (alloc_flags & ALLOC_OOM ||
4758 (gfp_mask & __GFP_NOMEMALLOC)))
4761 /* Retry as long as the OOM killer is making progress */
4762 if (did_some_progress) {
4763 no_progress_loops = 0;
4768 /* Deal with possible cpuset update races before we fail */
4769 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4773 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4776 if (gfp_mask & __GFP_NOFAIL) {
4778 * All existing users of the __GFP_NOFAIL are blockable, so warn
4779 * of any new users that actually require GFP_NOWAIT
4781 if (WARN_ON_ONCE(!can_direct_reclaim))
4785 * PF_MEMALLOC request from this context is rather bizarre
4786 * because we cannot reclaim anything and only can loop waiting
4787 * for somebody to do a work for us
4789 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4792 * non failing costly orders are a hard requirement which we
4793 * are not prepared for much so let's warn about these users
4794 * so that we can identify them and convert them to something
4797 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4800 * Help non-failing allocations by giving them access to memory
4801 * reserves but do not use ALLOC_NO_WATERMARKS because this
4802 * could deplete whole memory reserves which would just make
4803 * the situation worse
4805 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4813 warn_alloc(gfp_mask, ac->nodemask,
4814 "page allocation failure: order:%u", order);
4819 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4820 int preferred_nid, nodemask_t *nodemask,
4821 struct alloc_context *ac, gfp_t *alloc_mask,
4822 unsigned int *alloc_flags)
4824 ac->highest_zoneidx = gfp_zone(gfp_mask);
4825 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4826 ac->nodemask = nodemask;
4827 ac->migratetype = gfp_migratetype(gfp_mask);
4829 if (cpusets_enabled()) {
4830 *alloc_mask |= __GFP_HARDWALL;
4832 * When we are in the interrupt context, it is irrelevant
4833 * to the current task context. It means that any node ok.
4835 if (!in_interrupt() && !ac->nodemask)
4836 ac->nodemask = &cpuset_current_mems_allowed;
4838 *alloc_flags |= ALLOC_CPUSET;
4841 fs_reclaim_acquire(gfp_mask);
4842 fs_reclaim_release(gfp_mask);
4844 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4846 if (should_fail_alloc_page(gfp_mask, order))
4849 *alloc_flags = current_alloc_flags(gfp_mask, *alloc_flags);
4851 /* Dirty zone balancing only done in the fast path */
4852 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4855 * The preferred zone is used for statistics but crucially it is
4856 * also used as the starting point for the zonelist iterator. It
4857 * may get reset for allocations that ignore memory policies.
4859 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4860 ac->highest_zoneidx, ac->nodemask);
4866 * This is the 'heart' of the zoned buddy allocator.
4869 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4870 nodemask_t *nodemask)
4873 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4874 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4875 struct alloc_context ac = { };
4878 * There are several places where we assume that the order value is sane
4879 * so bail out early if the request is out of bound.
4881 if (unlikely(order >= MAX_ORDER)) {
4882 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4886 gfp_mask &= gfp_allowed_mask;
4887 alloc_mask = gfp_mask;
4888 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4892 * Forbid the first pass from falling back to types that fragment
4893 * memory until all local zones are considered.
4895 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4897 /* First allocation attempt */
4898 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4903 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4904 * resp. GFP_NOIO which has to be inherited for all allocation requests
4905 * from a particular context which has been marked by
4906 * memalloc_no{fs,io}_{save,restore}.
4908 alloc_mask = current_gfp_context(gfp_mask);
4909 ac.spread_dirty_pages = false;
4912 * Restore the original nodemask if it was potentially replaced with
4913 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4915 ac.nodemask = nodemask;
4917 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4920 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4921 unlikely(__memcg_kmem_charge_page(page, gfp_mask, order) != 0)) {
4922 __free_pages(page, order);
4926 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4930 EXPORT_SYMBOL(__alloc_pages_nodemask);
4933 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4934 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4935 * you need to access high mem.
4937 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4941 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4944 return (unsigned long) page_address(page);
4946 EXPORT_SYMBOL(__get_free_pages);
4948 unsigned long get_zeroed_page(gfp_t gfp_mask)
4950 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4952 EXPORT_SYMBOL(get_zeroed_page);
4954 static inline void free_the_page(struct page *page, unsigned int order)
4956 if (order == 0) /* Via pcp? */
4957 free_unref_page(page);
4959 __free_pages_ok(page, order);
4962 void __free_pages(struct page *page, unsigned int order)
4964 if (put_page_testzero(page))
4965 free_the_page(page, order);
4966 else if (!PageHead(page))
4968 free_the_page(page + (1 << order), order);
4970 EXPORT_SYMBOL(__free_pages);
4972 void free_pages(unsigned long addr, unsigned int order)
4975 VM_BUG_ON(!virt_addr_valid((void *)addr));
4976 __free_pages(virt_to_page((void *)addr), order);
4980 EXPORT_SYMBOL(free_pages);
4984 * An arbitrary-length arbitrary-offset area of memory which resides
4985 * within a 0 or higher order page. Multiple fragments within that page
4986 * are individually refcounted, in the page's reference counter.
4988 * The page_frag functions below provide a simple allocation framework for
4989 * page fragments. This is used by the network stack and network device
4990 * drivers to provide a backing region of memory for use as either an
4991 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4993 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4996 struct page *page = NULL;
4997 gfp_t gfp = gfp_mask;
4999 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5000 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5002 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5003 PAGE_FRAG_CACHE_MAX_ORDER);
5004 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5006 if (unlikely(!page))
5007 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5009 nc->va = page ? page_address(page) : NULL;
5014 void __page_frag_cache_drain(struct page *page, unsigned int count)
5016 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5018 if (page_ref_sub_and_test(page, count))
5019 free_the_page(page, compound_order(page));
5021 EXPORT_SYMBOL(__page_frag_cache_drain);
5023 void *page_frag_alloc(struct page_frag_cache *nc,
5024 unsigned int fragsz, gfp_t gfp_mask)
5026 unsigned int size = PAGE_SIZE;
5030 if (unlikely(!nc->va)) {
5032 page = __page_frag_cache_refill(nc, gfp_mask);
5036 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5037 /* if size can vary use size else just use PAGE_SIZE */
5040 /* Even if we own the page, we do not use atomic_set().
5041 * This would break get_page_unless_zero() users.
5043 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5045 /* reset page count bias and offset to start of new frag */
5046 nc->pfmemalloc = page_is_pfmemalloc(page);
5047 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5051 offset = nc->offset - fragsz;
5052 if (unlikely(offset < 0)) {
5053 page = virt_to_page(nc->va);
5055 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5058 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5059 /* if size can vary use size else just use PAGE_SIZE */
5062 /* OK, page count is 0, we can safely set it */
5063 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5065 /* reset page count bias and offset to start of new frag */
5066 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5067 offset = size - fragsz;
5071 nc->offset = offset;
5073 return nc->va + offset;
5075 EXPORT_SYMBOL(page_frag_alloc);
5078 * Frees a page fragment allocated out of either a compound or order 0 page.
5080 void page_frag_free(void *addr)
5082 struct page *page = virt_to_head_page(addr);
5084 if (unlikely(put_page_testzero(page)))
5085 free_the_page(page, compound_order(page));
5087 EXPORT_SYMBOL(page_frag_free);
5089 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5093 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5094 unsigned long used = addr + PAGE_ALIGN(size);
5096 split_page(virt_to_page((void *)addr), order);
5097 while (used < alloc_end) {
5102 return (void *)addr;
5106 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5107 * @size: the number of bytes to allocate
5108 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5110 * This function is similar to alloc_pages(), except that it allocates the
5111 * minimum number of pages to satisfy the request. alloc_pages() can only
5112 * allocate memory in power-of-two pages.
5114 * This function is also limited by MAX_ORDER.
5116 * Memory allocated by this function must be released by free_pages_exact().
5118 * Return: pointer to the allocated area or %NULL in case of error.
5120 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5122 unsigned int order = get_order(size);
5125 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5126 gfp_mask &= ~__GFP_COMP;
5128 addr = __get_free_pages(gfp_mask, order);
5129 return make_alloc_exact(addr, order, size);
5131 EXPORT_SYMBOL(alloc_pages_exact);
5134 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5136 * @nid: the preferred node ID where memory should be allocated
5137 * @size: the number of bytes to allocate
5138 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5140 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5143 * Return: pointer to the allocated area or %NULL in case of error.
5145 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5147 unsigned int order = get_order(size);
5150 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5151 gfp_mask &= ~__GFP_COMP;
5153 p = alloc_pages_node(nid, gfp_mask, order);
5156 return make_alloc_exact((unsigned long)page_address(p), order, size);
5160 * free_pages_exact - release memory allocated via alloc_pages_exact()
5161 * @virt: the value returned by alloc_pages_exact.
5162 * @size: size of allocation, same value as passed to alloc_pages_exact().
5164 * Release the memory allocated by a previous call to alloc_pages_exact.
5166 void free_pages_exact(void *virt, size_t size)
5168 unsigned long addr = (unsigned long)virt;
5169 unsigned long end = addr + PAGE_ALIGN(size);
5171 while (addr < end) {
5176 EXPORT_SYMBOL(free_pages_exact);
5179 * nr_free_zone_pages - count number of pages beyond high watermark
5180 * @offset: The zone index of the highest zone
5182 * nr_free_zone_pages() counts the number of pages which are beyond the
5183 * high watermark within all zones at or below a given zone index. For each
5184 * zone, the number of pages is calculated as:
5186 * nr_free_zone_pages = managed_pages - high_pages
5188 * Return: number of pages beyond high watermark.
5190 static unsigned long nr_free_zone_pages(int offset)
5195 /* Just pick one node, since fallback list is circular */
5196 unsigned long sum = 0;
5198 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5200 for_each_zone_zonelist(zone, z, zonelist, offset) {
5201 unsigned long size = zone_managed_pages(zone);
5202 unsigned long high = high_wmark_pages(zone);
5211 * nr_free_buffer_pages - count number of pages beyond high watermark
5213 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5214 * watermark within ZONE_DMA and ZONE_NORMAL.
5216 * Return: number of pages beyond high watermark within ZONE_DMA and
5219 unsigned long nr_free_buffer_pages(void)
5221 return nr_free_zone_pages(gfp_zone(GFP_USER));
5223 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5225 static inline void show_node(struct zone *zone)
5227 if (IS_ENABLED(CONFIG_NUMA))
5228 printk("Node %d ", zone_to_nid(zone));
5231 long si_mem_available(void)
5234 unsigned long pagecache;
5235 unsigned long wmark_low = 0;
5236 unsigned long pages[NR_LRU_LISTS];
5237 unsigned long reclaimable;
5241 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5242 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5245 wmark_low += low_wmark_pages(zone);
5248 * Estimate the amount of memory available for userspace allocations,
5249 * without causing swapping.
5251 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5254 * Not all the page cache can be freed, otherwise the system will
5255 * start swapping. Assume at least half of the page cache, or the
5256 * low watermark worth of cache, needs to stay.
5258 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5259 pagecache -= min(pagecache / 2, wmark_low);
5260 available += pagecache;
5263 * Part of the reclaimable slab and other kernel memory consists of
5264 * items that are in use, and cannot be freed. Cap this estimate at the
5267 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5268 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5269 available += reclaimable - min(reclaimable / 2, wmark_low);
5275 EXPORT_SYMBOL_GPL(si_mem_available);
5277 void si_meminfo(struct sysinfo *val)
5279 val->totalram = totalram_pages();
5280 val->sharedram = global_node_page_state(NR_SHMEM);
5281 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5282 val->bufferram = nr_blockdev_pages();
5283 val->totalhigh = totalhigh_pages();
5284 val->freehigh = nr_free_highpages();
5285 val->mem_unit = PAGE_SIZE;
5288 EXPORT_SYMBOL(si_meminfo);
5291 void si_meminfo_node(struct sysinfo *val, int nid)
5293 int zone_type; /* needs to be signed */
5294 unsigned long managed_pages = 0;
5295 unsigned long managed_highpages = 0;
5296 unsigned long free_highpages = 0;
5297 pg_data_t *pgdat = NODE_DATA(nid);
5299 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5300 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5301 val->totalram = managed_pages;
5302 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5303 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5304 #ifdef CONFIG_HIGHMEM
5305 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5306 struct zone *zone = &pgdat->node_zones[zone_type];
5308 if (is_highmem(zone)) {
5309 managed_highpages += zone_managed_pages(zone);
5310 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5313 val->totalhigh = managed_highpages;
5314 val->freehigh = free_highpages;
5316 val->totalhigh = managed_highpages;
5317 val->freehigh = free_highpages;
5319 val->mem_unit = PAGE_SIZE;
5324 * Determine whether the node should be displayed or not, depending on whether
5325 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5327 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5329 if (!(flags & SHOW_MEM_FILTER_NODES))
5333 * no node mask - aka implicit memory numa policy. Do not bother with
5334 * the synchronization - read_mems_allowed_begin - because we do not
5335 * have to be precise here.
5338 nodemask = &cpuset_current_mems_allowed;
5340 return !node_isset(nid, *nodemask);
5343 #define K(x) ((x) << (PAGE_SHIFT-10))
5345 static void show_migration_types(unsigned char type)
5347 static const char types[MIGRATE_TYPES] = {
5348 [MIGRATE_UNMOVABLE] = 'U',
5349 [MIGRATE_MOVABLE] = 'M',
5350 [MIGRATE_RECLAIMABLE] = 'E',
5351 [MIGRATE_HIGHATOMIC] = 'H',
5353 [MIGRATE_CMA] = 'C',
5355 #ifdef CONFIG_MEMORY_ISOLATION
5356 [MIGRATE_ISOLATE] = 'I',
5359 char tmp[MIGRATE_TYPES + 1];
5363 for (i = 0; i < MIGRATE_TYPES; i++) {
5364 if (type & (1 << i))
5369 printk(KERN_CONT "(%s) ", tmp);
5373 * Show free area list (used inside shift_scroll-lock stuff)
5374 * We also calculate the percentage fragmentation. We do this by counting the
5375 * memory on each free list with the exception of the first item on the list.
5378 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5381 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5383 unsigned long free_pcp = 0;
5388 for_each_populated_zone(zone) {
5389 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5392 for_each_online_cpu(cpu)
5393 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5396 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5397 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5398 " unevictable:%lu dirty:%lu writeback:%lu\n"
5399 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5400 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5401 " free:%lu free_pcp:%lu free_cma:%lu\n",
5402 global_node_page_state(NR_ACTIVE_ANON),
5403 global_node_page_state(NR_INACTIVE_ANON),
5404 global_node_page_state(NR_ISOLATED_ANON),
5405 global_node_page_state(NR_ACTIVE_FILE),
5406 global_node_page_state(NR_INACTIVE_FILE),
5407 global_node_page_state(NR_ISOLATED_FILE),
5408 global_node_page_state(NR_UNEVICTABLE),
5409 global_node_page_state(NR_FILE_DIRTY),
5410 global_node_page_state(NR_WRITEBACK),
5411 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5412 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5413 global_node_page_state(NR_FILE_MAPPED),
5414 global_node_page_state(NR_SHMEM),
5415 global_zone_page_state(NR_PAGETABLE),
5416 global_zone_page_state(NR_BOUNCE),
5417 global_zone_page_state(NR_FREE_PAGES),
5419 global_zone_page_state(NR_FREE_CMA_PAGES));
5421 for_each_online_pgdat(pgdat) {
5422 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5426 " active_anon:%lukB"
5427 " inactive_anon:%lukB"
5428 " active_file:%lukB"
5429 " inactive_file:%lukB"
5430 " unevictable:%lukB"
5431 " isolated(anon):%lukB"
5432 " isolated(file):%lukB"
5437 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5439 " shmem_pmdmapped: %lukB"
5442 " writeback_tmp:%lukB"
5443 " kernel_stack:%lukB"
5444 #ifdef CONFIG_SHADOW_CALL_STACK
5445 " shadow_call_stack:%lukB"
5447 " all_unreclaimable? %s"
5450 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5451 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5452 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5453 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5454 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5455 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5456 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5457 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5458 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5459 K(node_page_state(pgdat, NR_WRITEBACK)),
5460 K(node_page_state(pgdat, NR_SHMEM)),
5461 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5462 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5463 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5465 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5467 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5468 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5469 #ifdef CONFIG_SHADOW_CALL_STACK
5470 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5472 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5476 for_each_populated_zone(zone) {
5479 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5483 for_each_online_cpu(cpu)
5484 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5493 " reserved_highatomic:%luKB"
5494 " active_anon:%lukB"
5495 " inactive_anon:%lukB"
5496 " active_file:%lukB"
5497 " inactive_file:%lukB"
5498 " unevictable:%lukB"
5499 " writepending:%lukB"
5510 K(zone_page_state(zone, NR_FREE_PAGES)),
5511 K(min_wmark_pages(zone)),
5512 K(low_wmark_pages(zone)),
5513 K(high_wmark_pages(zone)),
5514 K(zone->nr_reserved_highatomic),
5515 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5516 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5517 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5518 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5519 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5520 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5521 K(zone->present_pages),
5522 K(zone_managed_pages(zone)),
5523 K(zone_page_state(zone, NR_MLOCK)),
5524 K(zone_page_state(zone, NR_PAGETABLE)),
5525 K(zone_page_state(zone, NR_BOUNCE)),
5527 K(this_cpu_read(zone->pageset->pcp.count)),
5528 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5529 printk("lowmem_reserve[]:");
5530 for (i = 0; i < MAX_NR_ZONES; i++)
5531 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5532 printk(KERN_CONT "\n");
5535 for_each_populated_zone(zone) {
5537 unsigned long nr[MAX_ORDER], flags, total = 0;
5538 unsigned char types[MAX_ORDER];
5540 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5543 printk(KERN_CONT "%s: ", zone->name);
5545 spin_lock_irqsave(&zone->lock, flags);
5546 for (order = 0; order < MAX_ORDER; order++) {
5547 struct free_area *area = &zone->free_area[order];
5550 nr[order] = area->nr_free;
5551 total += nr[order] << order;
5554 for (type = 0; type < MIGRATE_TYPES; type++) {
5555 if (!free_area_empty(area, type))
5556 types[order] |= 1 << type;
5559 spin_unlock_irqrestore(&zone->lock, flags);
5560 for (order = 0; order < MAX_ORDER; order++) {
5561 printk(KERN_CONT "%lu*%lukB ",
5562 nr[order], K(1UL) << order);
5564 show_migration_types(types[order]);
5566 printk(KERN_CONT "= %lukB\n", K(total));
5569 hugetlb_show_meminfo();
5571 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5573 show_swap_cache_info();
5576 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5578 zoneref->zone = zone;
5579 zoneref->zone_idx = zone_idx(zone);
5583 * Builds allocation fallback zone lists.
5585 * Add all populated zones of a node to the zonelist.
5587 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5590 enum zone_type zone_type = MAX_NR_ZONES;
5595 zone = pgdat->node_zones + zone_type;
5596 if (managed_zone(zone)) {
5597 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5598 check_highest_zone(zone_type);
5600 } while (zone_type);
5607 static int __parse_numa_zonelist_order(char *s)
5610 * We used to support different zonlists modes but they turned
5611 * out to be just not useful. Let's keep the warning in place
5612 * if somebody still use the cmd line parameter so that we do
5613 * not fail it silently
5615 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5616 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5622 char numa_zonelist_order[] = "Node";
5625 * sysctl handler for numa_zonelist_order
5627 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5628 void *buffer, size_t *length, loff_t *ppos)
5631 return __parse_numa_zonelist_order(buffer);
5632 return proc_dostring(table, write, buffer, length, ppos);
5636 #define MAX_NODE_LOAD (nr_online_nodes)
5637 static int node_load[MAX_NUMNODES];
5640 * find_next_best_node - find the next node that should appear in a given node's fallback list
5641 * @node: node whose fallback list we're appending
5642 * @used_node_mask: nodemask_t of already used nodes
5644 * We use a number of factors to determine which is the next node that should
5645 * appear on a given node's fallback list. The node should not have appeared
5646 * already in @node's fallback list, and it should be the next closest node
5647 * according to the distance array (which contains arbitrary distance values
5648 * from each node to each node in the system), and should also prefer nodes
5649 * with no CPUs, since presumably they'll have very little allocation pressure
5650 * on them otherwise.
5652 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5654 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5657 int min_val = INT_MAX;
5658 int best_node = NUMA_NO_NODE;
5660 /* Use the local node if we haven't already */
5661 if (!node_isset(node, *used_node_mask)) {
5662 node_set(node, *used_node_mask);
5666 for_each_node_state(n, N_MEMORY) {
5668 /* Don't want a node to appear more than once */
5669 if (node_isset(n, *used_node_mask))
5672 /* Use the distance array to find the distance */
5673 val = node_distance(node, n);
5675 /* Penalize nodes under us ("prefer the next node") */
5678 /* Give preference to headless and unused nodes */
5679 if (!cpumask_empty(cpumask_of_node(n)))
5680 val += PENALTY_FOR_NODE_WITH_CPUS;
5682 /* Slight preference for less loaded node */
5683 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5684 val += node_load[n];
5686 if (val < min_val) {
5693 node_set(best_node, *used_node_mask);
5700 * Build zonelists ordered by node and zones within node.
5701 * This results in maximum locality--normal zone overflows into local
5702 * DMA zone, if any--but risks exhausting DMA zone.
5704 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5707 struct zoneref *zonerefs;
5710 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5712 for (i = 0; i < nr_nodes; i++) {
5715 pg_data_t *node = NODE_DATA(node_order[i]);
5717 nr_zones = build_zonerefs_node(node, zonerefs);
5718 zonerefs += nr_zones;
5720 zonerefs->zone = NULL;
5721 zonerefs->zone_idx = 0;
5725 * Build gfp_thisnode zonelists
5727 static void build_thisnode_zonelists(pg_data_t *pgdat)
5729 struct zoneref *zonerefs;
5732 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5733 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5734 zonerefs += nr_zones;
5735 zonerefs->zone = NULL;
5736 zonerefs->zone_idx = 0;
5740 * Build zonelists ordered by zone and nodes within zones.
5741 * This results in conserving DMA zone[s] until all Normal memory is
5742 * exhausted, but results in overflowing to remote node while memory
5743 * may still exist in local DMA zone.
5746 static void build_zonelists(pg_data_t *pgdat)
5748 static int node_order[MAX_NUMNODES];
5749 int node, load, nr_nodes = 0;
5750 nodemask_t used_mask = NODE_MASK_NONE;
5751 int local_node, prev_node;
5753 /* NUMA-aware ordering of nodes */
5754 local_node = pgdat->node_id;
5755 load = nr_online_nodes;
5756 prev_node = local_node;
5758 memset(node_order, 0, sizeof(node_order));
5759 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5761 * We don't want to pressure a particular node.
5762 * So adding penalty to the first node in same
5763 * distance group to make it round-robin.
5765 if (node_distance(local_node, node) !=
5766 node_distance(local_node, prev_node))
5767 node_load[node] = load;
5769 node_order[nr_nodes++] = node;
5774 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5775 build_thisnode_zonelists(pgdat);
5778 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5780 * Return node id of node used for "local" allocations.
5781 * I.e., first node id of first zone in arg node's generic zonelist.
5782 * Used for initializing percpu 'numa_mem', which is used primarily
5783 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5785 int local_memory_node(int node)
5789 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5790 gfp_zone(GFP_KERNEL),
5792 return zone_to_nid(z->zone);
5796 static void setup_min_unmapped_ratio(void);
5797 static void setup_min_slab_ratio(void);
5798 #else /* CONFIG_NUMA */
5800 static void build_zonelists(pg_data_t *pgdat)
5802 int node, local_node;
5803 struct zoneref *zonerefs;
5806 local_node = pgdat->node_id;
5808 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5809 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5810 zonerefs += nr_zones;
5813 * Now we build the zonelist so that it contains the zones
5814 * of all the other nodes.
5815 * We don't want to pressure a particular node, so when
5816 * building the zones for node N, we make sure that the
5817 * zones coming right after the local ones are those from
5818 * node N+1 (modulo N)
5820 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5821 if (!node_online(node))
5823 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5824 zonerefs += nr_zones;
5826 for (node = 0; node < local_node; node++) {
5827 if (!node_online(node))
5829 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5830 zonerefs += nr_zones;
5833 zonerefs->zone = NULL;
5834 zonerefs->zone_idx = 0;
5837 #endif /* CONFIG_NUMA */
5840 * Boot pageset table. One per cpu which is going to be used for all
5841 * zones and all nodes. The parameters will be set in such a way
5842 * that an item put on a list will immediately be handed over to
5843 * the buddy list. This is safe since pageset manipulation is done
5844 * with interrupts disabled.
5846 * The boot_pagesets must be kept even after bootup is complete for
5847 * unused processors and/or zones. They do play a role for bootstrapping
5848 * hotplugged processors.
5850 * zoneinfo_show() and maybe other functions do
5851 * not check if the processor is online before following the pageset pointer.
5852 * Other parts of the kernel may not check if the zone is available.
5854 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5855 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5856 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5858 static void __build_all_zonelists(void *data)
5861 int __maybe_unused cpu;
5862 pg_data_t *self = data;
5863 static DEFINE_SPINLOCK(lock);
5868 memset(node_load, 0, sizeof(node_load));
5872 * This node is hotadded and no memory is yet present. So just
5873 * building zonelists is fine - no need to touch other nodes.
5875 if (self && !node_online(self->node_id)) {
5876 build_zonelists(self);
5878 for_each_online_node(nid) {
5879 pg_data_t *pgdat = NODE_DATA(nid);
5881 build_zonelists(pgdat);
5884 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5886 * We now know the "local memory node" for each node--
5887 * i.e., the node of the first zone in the generic zonelist.
5888 * Set up numa_mem percpu variable for on-line cpus. During
5889 * boot, only the boot cpu should be on-line; we'll init the
5890 * secondary cpus' numa_mem as they come on-line. During
5891 * node/memory hotplug, we'll fixup all on-line cpus.
5893 for_each_online_cpu(cpu)
5894 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5901 static noinline void __init
5902 build_all_zonelists_init(void)
5906 __build_all_zonelists(NULL);
5909 * Initialize the boot_pagesets that are going to be used
5910 * for bootstrapping processors. The real pagesets for
5911 * each zone will be allocated later when the per cpu
5912 * allocator is available.
5914 * boot_pagesets are used also for bootstrapping offline
5915 * cpus if the system is already booted because the pagesets
5916 * are needed to initialize allocators on a specific cpu too.
5917 * F.e. the percpu allocator needs the page allocator which
5918 * needs the percpu allocator in order to allocate its pagesets
5919 * (a chicken-egg dilemma).
5921 for_each_possible_cpu(cpu)
5922 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5924 mminit_verify_zonelist();
5925 cpuset_init_current_mems_allowed();
5929 * unless system_state == SYSTEM_BOOTING.
5931 * __ref due to call of __init annotated helper build_all_zonelists_init
5932 * [protected by SYSTEM_BOOTING].
5934 void __ref build_all_zonelists(pg_data_t *pgdat)
5936 unsigned long vm_total_pages;
5938 if (system_state == SYSTEM_BOOTING) {
5939 build_all_zonelists_init();
5941 __build_all_zonelists(pgdat);
5942 /* cpuset refresh routine should be here */
5944 /* Get the number of free pages beyond high watermark in all zones. */
5945 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5947 * Disable grouping by mobility if the number of pages in the
5948 * system is too low to allow the mechanism to work. It would be
5949 * more accurate, but expensive to check per-zone. This check is
5950 * made on memory-hotadd so a system can start with mobility
5951 * disabled and enable it later
5953 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5954 page_group_by_mobility_disabled = 1;
5956 page_group_by_mobility_disabled = 0;
5958 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5960 page_group_by_mobility_disabled ? "off" : "on",
5963 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5967 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5968 static bool __meminit
5969 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5971 static struct memblock_region *r;
5973 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5974 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5975 for_each_mem_region(r) {
5976 if (*pfn < memblock_region_memory_end_pfn(r))
5980 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5981 memblock_is_mirror(r)) {
5982 *pfn = memblock_region_memory_end_pfn(r);
5990 * Initially all pages are reserved - free ones are freed
5991 * up by memblock_free_all() once the early boot process is
5992 * done. Non-atomic initialization, single-pass.
5994 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5995 unsigned long start_pfn, enum meminit_context context,
5996 struct vmem_altmap *altmap)
5998 unsigned long pfn, end_pfn = start_pfn + size;
6001 if (highest_memmap_pfn < end_pfn - 1)
6002 highest_memmap_pfn = end_pfn - 1;
6004 #ifdef CONFIG_ZONE_DEVICE
6006 * Honor reservation requested by the driver for this ZONE_DEVICE
6007 * memory. We limit the total number of pages to initialize to just
6008 * those that might contain the memory mapping. We will defer the
6009 * ZONE_DEVICE page initialization until after we have released
6012 if (zone == ZONE_DEVICE) {
6016 if (start_pfn == altmap->base_pfn)
6017 start_pfn += altmap->reserve;
6018 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6022 for (pfn = start_pfn; pfn < end_pfn; ) {
6024 * There can be holes in boot-time mem_map[]s handed to this
6025 * function. They do not exist on hotplugged memory.
6027 if (context == MEMINIT_EARLY) {
6028 if (overlap_memmap_init(zone, &pfn))
6030 if (defer_init(nid, pfn, end_pfn))
6034 page = pfn_to_page(pfn);
6035 __init_single_page(page, pfn, zone, nid);
6036 if (context == MEMINIT_HOTPLUG)
6037 __SetPageReserved(page);
6040 * Mark the block movable so that blocks are reserved for
6041 * movable at startup. This will force kernel allocations
6042 * to reserve their blocks rather than leaking throughout
6043 * the address space during boot when many long-lived
6044 * kernel allocations are made.
6046 * bitmap is created for zone's valid pfn range. but memmap
6047 * can be created for invalid pages (for alignment)
6048 * check here not to call set_pageblock_migratetype() against
6051 if (!(pfn & (pageblock_nr_pages - 1))) {
6052 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6059 #ifdef CONFIG_ZONE_DEVICE
6060 void __ref memmap_init_zone_device(struct zone *zone,
6061 unsigned long start_pfn,
6062 unsigned long nr_pages,
6063 struct dev_pagemap *pgmap)
6065 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6066 struct pglist_data *pgdat = zone->zone_pgdat;
6067 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6068 unsigned long zone_idx = zone_idx(zone);
6069 unsigned long start = jiffies;
6070 int nid = pgdat->node_id;
6072 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6076 * The call to memmap_init_zone should have already taken care
6077 * of the pages reserved for the memmap, so we can just jump to
6078 * the end of that region and start processing the device pages.
6081 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6082 nr_pages = end_pfn - start_pfn;
6085 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6086 struct page *page = pfn_to_page(pfn);
6088 __init_single_page(page, pfn, zone_idx, nid);
6091 * Mark page reserved as it will need to wait for onlining
6092 * phase for it to be fully associated with a zone.
6094 * We can use the non-atomic __set_bit operation for setting
6095 * the flag as we are still initializing the pages.
6097 __SetPageReserved(page);
6100 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6101 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6102 * ever freed or placed on a driver-private list.
6104 page->pgmap = pgmap;
6105 page->zone_device_data = NULL;
6108 * Mark the block movable so that blocks are reserved for
6109 * movable at startup. This will force kernel allocations
6110 * to reserve their blocks rather than leaking throughout
6111 * the address space during boot when many long-lived
6112 * kernel allocations are made.
6114 * bitmap is created for zone's valid pfn range. but memmap
6115 * can be created for invalid pages (for alignment)
6116 * check here not to call set_pageblock_migratetype() against
6119 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6120 * because this is done early in section_activate()
6122 if (!(pfn & (pageblock_nr_pages - 1))) {
6123 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6128 pr_info("%s initialised %lu pages in %ums\n", __func__,
6129 nr_pages, jiffies_to_msecs(jiffies - start));
6133 static void __meminit zone_init_free_lists(struct zone *zone)
6135 unsigned int order, t;
6136 for_each_migratetype_order(order, t) {
6137 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6138 zone->free_area[order].nr_free = 0;
6142 void __meminit __weak memmap_init(unsigned long size, int nid,
6144 unsigned long range_start_pfn)
6146 unsigned long start_pfn, end_pfn;
6147 unsigned long range_end_pfn = range_start_pfn + size;
6150 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6151 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6152 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6154 if (end_pfn > start_pfn) {
6155 size = end_pfn - start_pfn;
6156 memmap_init_zone(size, nid, zone, start_pfn,
6157 MEMINIT_EARLY, NULL);
6162 static int zone_batchsize(struct zone *zone)
6168 * The per-cpu-pages pools are set to around 1000th of the
6171 batch = zone_managed_pages(zone) / 1024;
6172 /* But no more than a meg. */
6173 if (batch * PAGE_SIZE > 1024 * 1024)
6174 batch = (1024 * 1024) / PAGE_SIZE;
6175 batch /= 4; /* We effectively *= 4 below */
6180 * Clamp the batch to a 2^n - 1 value. Having a power
6181 * of 2 value was found to be more likely to have
6182 * suboptimal cache aliasing properties in some cases.
6184 * For example if 2 tasks are alternately allocating
6185 * batches of pages, one task can end up with a lot
6186 * of pages of one half of the possible page colors
6187 * and the other with pages of the other colors.
6189 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6194 /* The deferral and batching of frees should be suppressed under NOMMU
6197 * The problem is that NOMMU needs to be able to allocate large chunks
6198 * of contiguous memory as there's no hardware page translation to
6199 * assemble apparent contiguous memory from discontiguous pages.
6201 * Queueing large contiguous runs of pages for batching, however,
6202 * causes the pages to actually be freed in smaller chunks. As there
6203 * can be a significant delay between the individual batches being
6204 * recycled, this leads to the once large chunks of space being
6205 * fragmented and becoming unavailable for high-order allocations.
6212 * pcp->high and pcp->batch values are related and dependent on one another:
6213 * ->batch must never be higher then ->high.
6214 * The following function updates them in a safe manner without read side
6217 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6218 * those fields changing asynchronously (acording to the above rule).
6220 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6221 * outside of boot time (or some other assurance that no concurrent updaters
6224 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6225 unsigned long batch)
6227 /* start with a fail safe value for batch */
6231 /* Update high, then batch, in order */
6238 /* a companion to pageset_set_high() */
6239 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6241 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6244 static void pageset_init(struct per_cpu_pageset *p)
6246 struct per_cpu_pages *pcp;
6249 memset(p, 0, sizeof(*p));
6252 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6253 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6256 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6259 pageset_set_batch(p, batch);
6263 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6264 * to the value high for the pageset p.
6266 static void pageset_set_high(struct per_cpu_pageset *p,
6269 unsigned long batch = max(1UL, high / 4);
6270 if ((high / 4) > (PAGE_SHIFT * 8))
6271 batch = PAGE_SHIFT * 8;
6273 pageset_update(&p->pcp, high, batch);
6276 static void pageset_set_high_and_batch(struct zone *zone,
6277 struct per_cpu_pageset *pcp)
6279 if (percpu_pagelist_fraction)
6280 pageset_set_high(pcp,
6281 (zone_managed_pages(zone) /
6282 percpu_pagelist_fraction));
6284 pageset_set_batch(pcp, zone_batchsize(zone));
6287 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6289 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6292 pageset_set_high_and_batch(zone, pcp);
6295 void __meminit setup_zone_pageset(struct zone *zone)
6298 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6299 for_each_possible_cpu(cpu)
6300 zone_pageset_init(zone, cpu);
6304 * Allocate per cpu pagesets and initialize them.
6305 * Before this call only boot pagesets were available.
6307 void __init setup_per_cpu_pageset(void)
6309 struct pglist_data *pgdat;
6311 int __maybe_unused cpu;
6313 for_each_populated_zone(zone)
6314 setup_zone_pageset(zone);
6318 * Unpopulated zones continue using the boot pagesets.
6319 * The numa stats for these pagesets need to be reset.
6320 * Otherwise, they will end up skewing the stats of
6321 * the nodes these zones are associated with.
6323 for_each_possible_cpu(cpu) {
6324 struct per_cpu_pageset *pcp = &per_cpu(boot_pageset, cpu);
6325 memset(pcp->vm_numa_stat_diff, 0,
6326 sizeof(pcp->vm_numa_stat_diff));
6330 for_each_online_pgdat(pgdat)
6331 pgdat->per_cpu_nodestats =
6332 alloc_percpu(struct per_cpu_nodestat);
6335 static __meminit void zone_pcp_init(struct zone *zone)
6338 * per cpu subsystem is not up at this point. The following code
6339 * relies on the ability of the linker to provide the
6340 * offset of a (static) per cpu variable into the per cpu area.
6342 zone->pageset = &boot_pageset;
6344 if (populated_zone(zone))
6345 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6346 zone->name, zone->present_pages,
6347 zone_batchsize(zone));
6350 void __meminit init_currently_empty_zone(struct zone *zone,
6351 unsigned long zone_start_pfn,
6354 struct pglist_data *pgdat = zone->zone_pgdat;
6355 int zone_idx = zone_idx(zone) + 1;
6357 if (zone_idx > pgdat->nr_zones)
6358 pgdat->nr_zones = zone_idx;
6360 zone->zone_start_pfn = zone_start_pfn;
6362 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6363 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6365 (unsigned long)zone_idx(zone),
6366 zone_start_pfn, (zone_start_pfn + size));
6368 zone_init_free_lists(zone);
6369 zone->initialized = 1;
6373 * get_pfn_range_for_nid - Return the start and end page frames for a node
6374 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6375 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6376 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6378 * It returns the start and end page frame of a node based on information
6379 * provided by memblock_set_node(). If called for a node
6380 * with no available memory, a warning is printed and the start and end
6383 void __init get_pfn_range_for_nid(unsigned int nid,
6384 unsigned long *start_pfn, unsigned long *end_pfn)
6386 unsigned long this_start_pfn, this_end_pfn;
6392 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6393 *start_pfn = min(*start_pfn, this_start_pfn);
6394 *end_pfn = max(*end_pfn, this_end_pfn);
6397 if (*start_pfn == -1UL)
6402 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6403 * assumption is made that zones within a node are ordered in monotonic
6404 * increasing memory addresses so that the "highest" populated zone is used
6406 static void __init find_usable_zone_for_movable(void)
6409 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6410 if (zone_index == ZONE_MOVABLE)
6413 if (arch_zone_highest_possible_pfn[zone_index] >
6414 arch_zone_lowest_possible_pfn[zone_index])
6418 VM_BUG_ON(zone_index == -1);
6419 movable_zone = zone_index;
6423 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6424 * because it is sized independent of architecture. Unlike the other zones,
6425 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6426 * in each node depending on the size of each node and how evenly kernelcore
6427 * is distributed. This helper function adjusts the zone ranges
6428 * provided by the architecture for a given node by using the end of the
6429 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6430 * zones within a node are in order of monotonic increases memory addresses
6432 static void __init adjust_zone_range_for_zone_movable(int nid,
6433 unsigned long zone_type,
6434 unsigned long node_start_pfn,
6435 unsigned long node_end_pfn,
6436 unsigned long *zone_start_pfn,
6437 unsigned long *zone_end_pfn)
6439 /* Only adjust if ZONE_MOVABLE is on this node */
6440 if (zone_movable_pfn[nid]) {
6441 /* Size ZONE_MOVABLE */
6442 if (zone_type == ZONE_MOVABLE) {
6443 *zone_start_pfn = zone_movable_pfn[nid];
6444 *zone_end_pfn = min(node_end_pfn,
6445 arch_zone_highest_possible_pfn[movable_zone]);
6447 /* Adjust for ZONE_MOVABLE starting within this range */
6448 } else if (!mirrored_kernelcore &&
6449 *zone_start_pfn < zone_movable_pfn[nid] &&
6450 *zone_end_pfn > zone_movable_pfn[nid]) {
6451 *zone_end_pfn = zone_movable_pfn[nid];
6453 /* Check if this whole range is within ZONE_MOVABLE */
6454 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6455 *zone_start_pfn = *zone_end_pfn;
6460 * Return the number of pages a zone spans in a node, including holes
6461 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6463 static unsigned long __init zone_spanned_pages_in_node(int nid,
6464 unsigned long zone_type,
6465 unsigned long node_start_pfn,
6466 unsigned long node_end_pfn,
6467 unsigned long *zone_start_pfn,
6468 unsigned long *zone_end_pfn)
6470 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6471 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6472 /* When hotadd a new node from cpu_up(), the node should be empty */
6473 if (!node_start_pfn && !node_end_pfn)
6476 /* Get the start and end of the zone */
6477 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6478 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6479 adjust_zone_range_for_zone_movable(nid, zone_type,
6480 node_start_pfn, node_end_pfn,
6481 zone_start_pfn, zone_end_pfn);
6483 /* Check that this node has pages within the zone's required range */
6484 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6487 /* Move the zone boundaries inside the node if necessary */
6488 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6489 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6491 /* Return the spanned pages */
6492 return *zone_end_pfn - *zone_start_pfn;
6496 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6497 * then all holes in the requested range will be accounted for.
6499 unsigned long __init __absent_pages_in_range(int nid,
6500 unsigned long range_start_pfn,
6501 unsigned long range_end_pfn)
6503 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6504 unsigned long start_pfn, end_pfn;
6507 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6508 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6509 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6510 nr_absent -= end_pfn - start_pfn;
6516 * absent_pages_in_range - Return number of page frames in holes within a range
6517 * @start_pfn: The start PFN to start searching for holes
6518 * @end_pfn: The end PFN to stop searching for holes
6520 * Return: the number of pages frames in memory holes within a range.
6522 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6523 unsigned long end_pfn)
6525 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6528 /* Return the number of page frames in holes in a zone on a node */
6529 static unsigned long __init zone_absent_pages_in_node(int nid,
6530 unsigned long zone_type,
6531 unsigned long node_start_pfn,
6532 unsigned long node_end_pfn)
6534 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6535 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6536 unsigned long zone_start_pfn, zone_end_pfn;
6537 unsigned long nr_absent;
6539 /* When hotadd a new node from cpu_up(), the node should be empty */
6540 if (!node_start_pfn && !node_end_pfn)
6543 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6544 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6546 adjust_zone_range_for_zone_movable(nid, zone_type,
6547 node_start_pfn, node_end_pfn,
6548 &zone_start_pfn, &zone_end_pfn);
6549 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6552 * ZONE_MOVABLE handling.
6553 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6556 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6557 unsigned long start_pfn, end_pfn;
6558 struct memblock_region *r;
6560 for_each_mem_region(r) {
6561 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6562 zone_start_pfn, zone_end_pfn);
6563 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6564 zone_start_pfn, zone_end_pfn);
6566 if (zone_type == ZONE_MOVABLE &&
6567 memblock_is_mirror(r))
6568 nr_absent += end_pfn - start_pfn;
6570 if (zone_type == ZONE_NORMAL &&
6571 !memblock_is_mirror(r))
6572 nr_absent += end_pfn - start_pfn;
6579 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6580 unsigned long node_start_pfn,
6581 unsigned long node_end_pfn)
6583 unsigned long realtotalpages = 0, totalpages = 0;
6586 for (i = 0; i < MAX_NR_ZONES; i++) {
6587 struct zone *zone = pgdat->node_zones + i;
6588 unsigned long zone_start_pfn, zone_end_pfn;
6589 unsigned long spanned, absent;
6590 unsigned long size, real_size;
6592 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
6597 absent = zone_absent_pages_in_node(pgdat->node_id, i,
6602 real_size = size - absent;
6605 zone->zone_start_pfn = zone_start_pfn;
6607 zone->zone_start_pfn = 0;
6608 zone->spanned_pages = size;
6609 zone->present_pages = real_size;
6612 realtotalpages += real_size;
6615 pgdat->node_spanned_pages = totalpages;
6616 pgdat->node_present_pages = realtotalpages;
6617 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6621 #ifndef CONFIG_SPARSEMEM
6623 * Calculate the size of the zone->blockflags rounded to an unsigned long
6624 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6625 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6626 * round what is now in bits to nearest long in bits, then return it in
6629 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6631 unsigned long usemapsize;
6633 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6634 usemapsize = roundup(zonesize, pageblock_nr_pages);
6635 usemapsize = usemapsize >> pageblock_order;
6636 usemapsize *= NR_PAGEBLOCK_BITS;
6637 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6639 return usemapsize / 8;
6642 static void __ref setup_usemap(struct pglist_data *pgdat,
6644 unsigned long zone_start_pfn,
6645 unsigned long zonesize)
6647 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6648 zone->pageblock_flags = NULL;
6650 zone->pageblock_flags =
6651 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6653 if (!zone->pageblock_flags)
6654 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6655 usemapsize, zone->name, pgdat->node_id);
6659 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6660 unsigned long zone_start_pfn, unsigned long zonesize) {}
6661 #endif /* CONFIG_SPARSEMEM */
6663 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6665 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6666 void __init set_pageblock_order(void)
6670 /* Check that pageblock_nr_pages has not already been setup */
6671 if (pageblock_order)
6674 if (HPAGE_SHIFT > PAGE_SHIFT)
6675 order = HUGETLB_PAGE_ORDER;
6677 order = MAX_ORDER - 1;
6680 * Assume the largest contiguous order of interest is a huge page.
6681 * This value may be variable depending on boot parameters on IA64 and
6684 pageblock_order = order;
6686 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6689 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6690 * is unused as pageblock_order is set at compile-time. See
6691 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6694 void __init set_pageblock_order(void)
6698 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6700 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6701 unsigned long present_pages)
6703 unsigned long pages = spanned_pages;
6706 * Provide a more accurate estimation if there are holes within
6707 * the zone and SPARSEMEM is in use. If there are holes within the
6708 * zone, each populated memory region may cost us one or two extra
6709 * memmap pages due to alignment because memmap pages for each
6710 * populated regions may not be naturally aligned on page boundary.
6711 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6713 if (spanned_pages > present_pages + (present_pages >> 4) &&
6714 IS_ENABLED(CONFIG_SPARSEMEM))
6715 pages = present_pages;
6717 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6720 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6721 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6723 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6725 spin_lock_init(&ds_queue->split_queue_lock);
6726 INIT_LIST_HEAD(&ds_queue->split_queue);
6727 ds_queue->split_queue_len = 0;
6730 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6733 #ifdef CONFIG_COMPACTION
6734 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6736 init_waitqueue_head(&pgdat->kcompactd_wait);
6739 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6742 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6744 pgdat_resize_init(pgdat);
6746 pgdat_init_split_queue(pgdat);
6747 pgdat_init_kcompactd(pgdat);
6749 init_waitqueue_head(&pgdat->kswapd_wait);
6750 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6752 pgdat_page_ext_init(pgdat);
6753 spin_lock_init(&pgdat->lru_lock);
6754 lruvec_init(&pgdat->__lruvec);
6757 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6758 unsigned long remaining_pages)
6760 atomic_long_set(&zone->managed_pages, remaining_pages);
6761 zone_set_nid(zone, nid);
6762 zone->name = zone_names[idx];
6763 zone->zone_pgdat = NODE_DATA(nid);
6764 spin_lock_init(&zone->lock);
6765 zone_seqlock_init(zone);
6766 zone_pcp_init(zone);
6770 * Set up the zone data structures
6771 * - init pgdat internals
6772 * - init all zones belonging to this node
6774 * NOTE: this function is only called during memory hotplug
6776 #ifdef CONFIG_MEMORY_HOTPLUG
6777 void __ref free_area_init_core_hotplug(int nid)
6780 pg_data_t *pgdat = NODE_DATA(nid);
6782 pgdat_init_internals(pgdat);
6783 for (z = 0; z < MAX_NR_ZONES; z++)
6784 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6789 * Set up the zone data structures:
6790 * - mark all pages reserved
6791 * - mark all memory queues empty
6792 * - clear the memory bitmaps
6794 * NOTE: pgdat should get zeroed by caller.
6795 * NOTE: this function is only called during early init.
6797 static void __init free_area_init_core(struct pglist_data *pgdat)
6800 int nid = pgdat->node_id;
6802 pgdat_init_internals(pgdat);
6803 pgdat->per_cpu_nodestats = &boot_nodestats;
6805 for (j = 0; j < MAX_NR_ZONES; j++) {
6806 struct zone *zone = pgdat->node_zones + j;
6807 unsigned long size, freesize, memmap_pages;
6808 unsigned long zone_start_pfn = zone->zone_start_pfn;
6810 size = zone->spanned_pages;
6811 freesize = zone->present_pages;
6814 * Adjust freesize so that it accounts for how much memory
6815 * is used by this zone for memmap. This affects the watermark
6816 * and per-cpu initialisations
6818 memmap_pages = calc_memmap_size(size, freesize);
6819 if (!is_highmem_idx(j)) {
6820 if (freesize >= memmap_pages) {
6821 freesize -= memmap_pages;
6824 " %s zone: %lu pages used for memmap\n",
6825 zone_names[j], memmap_pages);
6827 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6828 zone_names[j], memmap_pages, freesize);
6831 /* Account for reserved pages */
6832 if (j == 0 && freesize > dma_reserve) {
6833 freesize -= dma_reserve;
6834 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6835 zone_names[0], dma_reserve);
6838 if (!is_highmem_idx(j))
6839 nr_kernel_pages += freesize;
6840 /* Charge for highmem memmap if there are enough kernel pages */
6841 else if (nr_kernel_pages > memmap_pages * 2)
6842 nr_kernel_pages -= memmap_pages;
6843 nr_all_pages += freesize;
6846 * Set an approximate value for lowmem here, it will be adjusted
6847 * when the bootmem allocator frees pages into the buddy system.
6848 * And all highmem pages will be managed by the buddy system.
6850 zone_init_internals(zone, j, nid, freesize);
6855 set_pageblock_order();
6856 setup_usemap(pgdat, zone, zone_start_pfn, size);
6857 init_currently_empty_zone(zone, zone_start_pfn, size);
6858 memmap_init(size, nid, j, zone_start_pfn);
6862 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6863 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6865 unsigned long __maybe_unused start = 0;
6866 unsigned long __maybe_unused offset = 0;
6868 /* Skip empty nodes */
6869 if (!pgdat->node_spanned_pages)
6872 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6873 offset = pgdat->node_start_pfn - start;
6874 /* ia64 gets its own node_mem_map, before this, without bootmem */
6875 if (!pgdat->node_mem_map) {
6876 unsigned long size, end;
6880 * The zone's endpoints aren't required to be MAX_ORDER
6881 * aligned but the node_mem_map endpoints must be in order
6882 * for the buddy allocator to function correctly.
6884 end = pgdat_end_pfn(pgdat);
6885 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6886 size = (end - start) * sizeof(struct page);
6887 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6890 panic("Failed to allocate %ld bytes for node %d memory map\n",
6891 size, pgdat->node_id);
6892 pgdat->node_mem_map = map + offset;
6894 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6895 __func__, pgdat->node_id, (unsigned long)pgdat,
6896 (unsigned long)pgdat->node_mem_map);
6897 #ifndef CONFIG_NEED_MULTIPLE_NODES
6899 * With no DISCONTIG, the global mem_map is just set as node 0's
6901 if (pgdat == NODE_DATA(0)) {
6902 mem_map = NODE_DATA(0)->node_mem_map;
6903 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6909 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6910 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6912 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6913 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6915 pgdat->first_deferred_pfn = ULONG_MAX;
6918 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6921 static void __init free_area_init_node(int nid)
6923 pg_data_t *pgdat = NODE_DATA(nid);
6924 unsigned long start_pfn = 0;
6925 unsigned long end_pfn = 0;
6927 /* pg_data_t should be reset to zero when it's allocated */
6928 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
6930 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6932 pgdat->node_id = nid;
6933 pgdat->node_start_pfn = start_pfn;
6934 pgdat->per_cpu_nodestats = NULL;
6936 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6937 (u64)start_pfn << PAGE_SHIFT,
6938 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6939 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
6941 alloc_node_mem_map(pgdat);
6942 pgdat_set_deferred_range(pgdat);
6944 free_area_init_core(pgdat);
6947 void __init free_area_init_memoryless_node(int nid)
6949 free_area_init_node(nid);
6952 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6954 * Initialize all valid struct pages in the range [spfn, epfn) and mark them
6955 * PageReserved(). Return the number of struct pages that were initialized.
6957 static u64 __init init_unavailable_range(unsigned long spfn, unsigned long epfn)
6962 for (pfn = spfn; pfn < epfn; pfn++) {
6963 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6964 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6965 + pageblock_nr_pages - 1;
6969 * Use a fake node/zone (0) for now. Some of these pages
6970 * (in memblock.reserved but not in memblock.memory) will
6971 * get re-initialized via reserve_bootmem_region() later.
6973 __init_single_page(pfn_to_page(pfn), pfn, 0, 0);
6974 __SetPageReserved(pfn_to_page(pfn));
6982 * Only struct pages that are backed by physical memory are zeroed and
6983 * initialized by going through __init_single_page(). But, there are some
6984 * struct pages which are reserved in memblock allocator and their fields
6985 * may be accessed (for example page_to_pfn() on some configuration accesses
6986 * flags). We must explicitly initialize those struct pages.
6988 * This function also addresses a similar issue where struct pages are left
6989 * uninitialized because the physical address range is not covered by
6990 * memblock.memory or memblock.reserved. That could happen when memblock
6991 * layout is manually configured via memmap=, or when the highest physical
6992 * address (max_pfn) does not end on a section boundary.
6994 static void __init init_unavailable_mem(void)
6996 phys_addr_t start, end;
6998 phys_addr_t next = 0;
7001 * Loop through unavailable ranges not covered by memblock.memory.
7004 for_each_mem_range(i, &start, &end) {
7006 pgcnt += init_unavailable_range(PFN_DOWN(next),
7012 * Early sections always have a fully populated memmap for the whole
7013 * section - see pfn_valid(). If the last section has holes at the
7014 * end and that section is marked "online", the memmap will be
7015 * considered initialized. Make sure that memmap has a well defined
7018 pgcnt += init_unavailable_range(PFN_DOWN(next),
7019 round_up(max_pfn, PAGES_PER_SECTION));
7022 * Struct pages that do not have backing memory. This could be because
7023 * firmware is using some of this memory, or for some other reasons.
7026 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
7029 static inline void __init init_unavailable_mem(void)
7032 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
7034 #if MAX_NUMNODES > 1
7036 * Figure out the number of possible node ids.
7038 void __init setup_nr_node_ids(void)
7040 unsigned int highest;
7042 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7043 nr_node_ids = highest + 1;
7048 * node_map_pfn_alignment - determine the maximum internode alignment
7050 * This function should be called after node map is populated and sorted.
7051 * It calculates the maximum power of two alignment which can distinguish
7054 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7055 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7056 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7057 * shifted, 1GiB is enough and this function will indicate so.
7059 * This is used to test whether pfn -> nid mapping of the chosen memory
7060 * model has fine enough granularity to avoid incorrect mapping for the
7061 * populated node map.
7063 * Return: the determined alignment in pfn's. 0 if there is no alignment
7064 * requirement (single node).
7066 unsigned long __init node_map_pfn_alignment(void)
7068 unsigned long accl_mask = 0, last_end = 0;
7069 unsigned long start, end, mask;
7070 int last_nid = NUMA_NO_NODE;
7073 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7074 if (!start || last_nid < 0 || last_nid == nid) {
7081 * Start with a mask granular enough to pin-point to the
7082 * start pfn and tick off bits one-by-one until it becomes
7083 * too coarse to separate the current node from the last.
7085 mask = ~((1 << __ffs(start)) - 1);
7086 while (mask && last_end <= (start & (mask << 1)))
7089 /* accumulate all internode masks */
7093 /* convert mask to number of pages */
7094 return ~accl_mask + 1;
7098 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7100 * Return: the minimum PFN based on information provided via
7101 * memblock_set_node().
7103 unsigned long __init find_min_pfn_with_active_regions(void)
7105 return PHYS_PFN(memblock_start_of_DRAM());
7109 * early_calculate_totalpages()
7110 * Sum pages in active regions for movable zone.
7111 * Populate N_MEMORY for calculating usable_nodes.
7113 static unsigned long __init early_calculate_totalpages(void)
7115 unsigned long totalpages = 0;
7116 unsigned long start_pfn, end_pfn;
7119 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7120 unsigned long pages = end_pfn - start_pfn;
7122 totalpages += pages;
7124 node_set_state(nid, N_MEMORY);
7130 * Find the PFN the Movable zone begins in each node. Kernel memory
7131 * is spread evenly between nodes as long as the nodes have enough
7132 * memory. When they don't, some nodes will have more kernelcore than
7135 static void __init find_zone_movable_pfns_for_nodes(void)
7138 unsigned long usable_startpfn;
7139 unsigned long kernelcore_node, kernelcore_remaining;
7140 /* save the state before borrow the nodemask */
7141 nodemask_t saved_node_state = node_states[N_MEMORY];
7142 unsigned long totalpages = early_calculate_totalpages();
7143 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7144 struct memblock_region *r;
7146 /* Need to find movable_zone earlier when movable_node is specified. */
7147 find_usable_zone_for_movable();
7150 * If movable_node is specified, ignore kernelcore and movablecore
7153 if (movable_node_is_enabled()) {
7154 for_each_mem_region(r) {
7155 if (!memblock_is_hotpluggable(r))
7158 nid = memblock_get_region_node(r);
7160 usable_startpfn = PFN_DOWN(r->base);
7161 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7162 min(usable_startpfn, zone_movable_pfn[nid]) :
7170 * If kernelcore=mirror is specified, ignore movablecore option
7172 if (mirrored_kernelcore) {
7173 bool mem_below_4gb_not_mirrored = false;
7175 for_each_mem_region(r) {
7176 if (memblock_is_mirror(r))
7179 nid = memblock_get_region_node(r);
7181 usable_startpfn = memblock_region_memory_base_pfn(r);
7183 if (usable_startpfn < 0x100000) {
7184 mem_below_4gb_not_mirrored = true;
7188 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7189 min(usable_startpfn, zone_movable_pfn[nid]) :
7193 if (mem_below_4gb_not_mirrored)
7194 pr_warn("This configuration results in unmirrored kernel memory.\n");
7200 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7201 * amount of necessary memory.
7203 if (required_kernelcore_percent)
7204 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7206 if (required_movablecore_percent)
7207 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7211 * If movablecore= was specified, calculate what size of
7212 * kernelcore that corresponds so that memory usable for
7213 * any allocation type is evenly spread. If both kernelcore
7214 * and movablecore are specified, then the value of kernelcore
7215 * will be used for required_kernelcore if it's greater than
7216 * what movablecore would have allowed.
7218 if (required_movablecore) {
7219 unsigned long corepages;
7222 * Round-up so that ZONE_MOVABLE is at least as large as what
7223 * was requested by the user
7225 required_movablecore =
7226 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7227 required_movablecore = min(totalpages, required_movablecore);
7228 corepages = totalpages - required_movablecore;
7230 required_kernelcore = max(required_kernelcore, corepages);
7234 * If kernelcore was not specified or kernelcore size is larger
7235 * than totalpages, there is no ZONE_MOVABLE.
7237 if (!required_kernelcore || required_kernelcore >= totalpages)
7240 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7241 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7244 /* Spread kernelcore memory as evenly as possible throughout nodes */
7245 kernelcore_node = required_kernelcore / usable_nodes;
7246 for_each_node_state(nid, N_MEMORY) {
7247 unsigned long start_pfn, end_pfn;
7250 * Recalculate kernelcore_node if the division per node
7251 * now exceeds what is necessary to satisfy the requested
7252 * amount of memory for the kernel
7254 if (required_kernelcore < kernelcore_node)
7255 kernelcore_node = required_kernelcore / usable_nodes;
7258 * As the map is walked, we track how much memory is usable
7259 * by the kernel using kernelcore_remaining. When it is
7260 * 0, the rest of the node is usable by ZONE_MOVABLE
7262 kernelcore_remaining = kernelcore_node;
7264 /* Go through each range of PFNs within this node */
7265 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7266 unsigned long size_pages;
7268 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7269 if (start_pfn >= end_pfn)
7272 /* Account for what is only usable for kernelcore */
7273 if (start_pfn < usable_startpfn) {
7274 unsigned long kernel_pages;
7275 kernel_pages = min(end_pfn, usable_startpfn)
7278 kernelcore_remaining -= min(kernel_pages,
7279 kernelcore_remaining);
7280 required_kernelcore -= min(kernel_pages,
7281 required_kernelcore);
7283 /* Continue if range is now fully accounted */
7284 if (end_pfn <= usable_startpfn) {
7287 * Push zone_movable_pfn to the end so
7288 * that if we have to rebalance
7289 * kernelcore across nodes, we will
7290 * not double account here
7292 zone_movable_pfn[nid] = end_pfn;
7295 start_pfn = usable_startpfn;
7299 * The usable PFN range for ZONE_MOVABLE is from
7300 * start_pfn->end_pfn. Calculate size_pages as the
7301 * number of pages used as kernelcore
7303 size_pages = end_pfn - start_pfn;
7304 if (size_pages > kernelcore_remaining)
7305 size_pages = kernelcore_remaining;
7306 zone_movable_pfn[nid] = start_pfn + size_pages;
7309 * Some kernelcore has been met, update counts and
7310 * break if the kernelcore for this node has been
7313 required_kernelcore -= min(required_kernelcore,
7315 kernelcore_remaining -= size_pages;
7316 if (!kernelcore_remaining)
7322 * If there is still required_kernelcore, we do another pass with one
7323 * less node in the count. This will push zone_movable_pfn[nid] further
7324 * along on the nodes that still have memory until kernelcore is
7328 if (usable_nodes && required_kernelcore > usable_nodes)
7332 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7333 for (nid = 0; nid < MAX_NUMNODES; nid++)
7334 zone_movable_pfn[nid] =
7335 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7338 /* restore the node_state */
7339 node_states[N_MEMORY] = saved_node_state;
7342 /* Any regular or high memory on that node ? */
7343 static void check_for_memory(pg_data_t *pgdat, int nid)
7345 enum zone_type zone_type;
7347 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7348 struct zone *zone = &pgdat->node_zones[zone_type];
7349 if (populated_zone(zone)) {
7350 if (IS_ENABLED(CONFIG_HIGHMEM))
7351 node_set_state(nid, N_HIGH_MEMORY);
7352 if (zone_type <= ZONE_NORMAL)
7353 node_set_state(nid, N_NORMAL_MEMORY);
7360 * Some architecturs, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7361 * such cases we allow max_zone_pfn sorted in the descending order
7363 bool __weak arch_has_descending_max_zone_pfns(void)
7369 * free_area_init - Initialise all pg_data_t and zone data
7370 * @max_zone_pfn: an array of max PFNs for each zone
7372 * This will call free_area_init_node() for each active node in the system.
7373 * Using the page ranges provided by memblock_set_node(), the size of each
7374 * zone in each node and their holes is calculated. If the maximum PFN
7375 * between two adjacent zones match, it is assumed that the zone is empty.
7376 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7377 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7378 * starts where the previous one ended. For example, ZONE_DMA32 starts
7379 * at arch_max_dma_pfn.
7381 void __init free_area_init(unsigned long *max_zone_pfn)
7383 unsigned long start_pfn, end_pfn;
7387 /* Record where the zone boundaries are */
7388 memset(arch_zone_lowest_possible_pfn, 0,
7389 sizeof(arch_zone_lowest_possible_pfn));
7390 memset(arch_zone_highest_possible_pfn, 0,
7391 sizeof(arch_zone_highest_possible_pfn));
7393 start_pfn = find_min_pfn_with_active_regions();
7394 descending = arch_has_descending_max_zone_pfns();
7396 for (i = 0; i < MAX_NR_ZONES; i++) {
7398 zone = MAX_NR_ZONES - i - 1;
7402 if (zone == ZONE_MOVABLE)
7405 end_pfn = max(max_zone_pfn[zone], start_pfn);
7406 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7407 arch_zone_highest_possible_pfn[zone] = end_pfn;
7409 start_pfn = end_pfn;
7412 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7413 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7414 find_zone_movable_pfns_for_nodes();
7416 /* Print out the zone ranges */
7417 pr_info("Zone ranges:\n");
7418 for (i = 0; i < MAX_NR_ZONES; i++) {
7419 if (i == ZONE_MOVABLE)
7421 pr_info(" %-8s ", zone_names[i]);
7422 if (arch_zone_lowest_possible_pfn[i] ==
7423 arch_zone_highest_possible_pfn[i])
7426 pr_cont("[mem %#018Lx-%#018Lx]\n",
7427 (u64)arch_zone_lowest_possible_pfn[i]
7429 ((u64)arch_zone_highest_possible_pfn[i]
7430 << PAGE_SHIFT) - 1);
7433 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7434 pr_info("Movable zone start for each node\n");
7435 for (i = 0; i < MAX_NUMNODES; i++) {
7436 if (zone_movable_pfn[i])
7437 pr_info(" Node %d: %#018Lx\n", i,
7438 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7442 * Print out the early node map, and initialize the
7443 * subsection-map relative to active online memory ranges to
7444 * enable future "sub-section" extensions of the memory map.
7446 pr_info("Early memory node ranges\n");
7447 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7448 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7449 (u64)start_pfn << PAGE_SHIFT,
7450 ((u64)end_pfn << PAGE_SHIFT) - 1);
7451 subsection_map_init(start_pfn, end_pfn - start_pfn);
7454 /* Initialise every node */
7455 mminit_verify_pageflags_layout();
7456 setup_nr_node_ids();
7457 init_unavailable_mem();
7458 for_each_online_node(nid) {
7459 pg_data_t *pgdat = NODE_DATA(nid);
7460 free_area_init_node(nid);
7462 /* Any memory on that node */
7463 if (pgdat->node_present_pages)
7464 node_set_state(nid, N_MEMORY);
7465 check_for_memory(pgdat, nid);
7469 static int __init cmdline_parse_core(char *p, unsigned long *core,
7470 unsigned long *percent)
7472 unsigned long long coremem;
7478 /* Value may be a percentage of total memory, otherwise bytes */
7479 coremem = simple_strtoull(p, &endptr, 0);
7480 if (*endptr == '%') {
7481 /* Paranoid check for percent values greater than 100 */
7482 WARN_ON(coremem > 100);
7486 coremem = memparse(p, &p);
7487 /* Paranoid check that UL is enough for the coremem value */
7488 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7490 *core = coremem >> PAGE_SHIFT;
7497 * kernelcore=size sets the amount of memory for use for allocations that
7498 * cannot be reclaimed or migrated.
7500 static int __init cmdline_parse_kernelcore(char *p)
7502 /* parse kernelcore=mirror */
7503 if (parse_option_str(p, "mirror")) {
7504 mirrored_kernelcore = true;
7508 return cmdline_parse_core(p, &required_kernelcore,
7509 &required_kernelcore_percent);
7513 * movablecore=size sets the amount of memory for use for allocations that
7514 * can be reclaimed or migrated.
7516 static int __init cmdline_parse_movablecore(char *p)
7518 return cmdline_parse_core(p, &required_movablecore,
7519 &required_movablecore_percent);
7522 early_param("kernelcore", cmdline_parse_kernelcore);
7523 early_param("movablecore", cmdline_parse_movablecore);
7525 void adjust_managed_page_count(struct page *page, long count)
7527 atomic_long_add(count, &page_zone(page)->managed_pages);
7528 totalram_pages_add(count);
7529 #ifdef CONFIG_HIGHMEM
7530 if (PageHighMem(page))
7531 totalhigh_pages_add(count);
7534 EXPORT_SYMBOL(adjust_managed_page_count);
7536 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7539 unsigned long pages = 0;
7541 start = (void *)PAGE_ALIGN((unsigned long)start);
7542 end = (void *)((unsigned long)end & PAGE_MASK);
7543 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7544 struct page *page = virt_to_page(pos);
7545 void *direct_map_addr;
7548 * 'direct_map_addr' might be different from 'pos'
7549 * because some architectures' virt_to_page()
7550 * work with aliases. Getting the direct map
7551 * address ensures that we get a _writeable_
7552 * alias for the memset().
7554 direct_map_addr = page_address(page);
7555 if ((unsigned int)poison <= 0xFF)
7556 memset(direct_map_addr, poison, PAGE_SIZE);
7558 free_reserved_page(page);
7562 pr_info("Freeing %s memory: %ldK\n",
7563 s, pages << (PAGE_SHIFT - 10));
7568 #ifdef CONFIG_HIGHMEM
7569 void free_highmem_page(struct page *page)
7571 __free_reserved_page(page);
7572 totalram_pages_inc();
7573 atomic_long_inc(&page_zone(page)->managed_pages);
7574 totalhigh_pages_inc();
7579 void __init mem_init_print_info(const char *str)
7581 unsigned long physpages, codesize, datasize, rosize, bss_size;
7582 unsigned long init_code_size, init_data_size;
7584 physpages = get_num_physpages();
7585 codesize = _etext - _stext;
7586 datasize = _edata - _sdata;
7587 rosize = __end_rodata - __start_rodata;
7588 bss_size = __bss_stop - __bss_start;
7589 init_data_size = __init_end - __init_begin;
7590 init_code_size = _einittext - _sinittext;
7593 * Detect special cases and adjust section sizes accordingly:
7594 * 1) .init.* may be embedded into .data sections
7595 * 2) .init.text.* may be out of [__init_begin, __init_end],
7596 * please refer to arch/tile/kernel/vmlinux.lds.S.
7597 * 3) .rodata.* may be embedded into .text or .data sections.
7599 #define adj_init_size(start, end, size, pos, adj) \
7601 if (start <= pos && pos < end && size > adj) \
7605 adj_init_size(__init_begin, __init_end, init_data_size,
7606 _sinittext, init_code_size);
7607 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7608 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7609 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7610 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7612 #undef adj_init_size
7614 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7615 #ifdef CONFIG_HIGHMEM
7619 nr_free_pages() << (PAGE_SHIFT - 10),
7620 physpages << (PAGE_SHIFT - 10),
7621 codesize >> 10, datasize >> 10, rosize >> 10,
7622 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7623 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7624 totalcma_pages << (PAGE_SHIFT - 10),
7625 #ifdef CONFIG_HIGHMEM
7626 totalhigh_pages() << (PAGE_SHIFT - 10),
7628 str ? ", " : "", str ? str : "");
7632 * set_dma_reserve - set the specified number of pages reserved in the first zone
7633 * @new_dma_reserve: The number of pages to mark reserved
7635 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7636 * In the DMA zone, a significant percentage may be consumed by kernel image
7637 * and other unfreeable allocations which can skew the watermarks badly. This
7638 * function may optionally be used to account for unfreeable pages in the
7639 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7640 * smaller per-cpu batchsize.
7642 void __init set_dma_reserve(unsigned long new_dma_reserve)
7644 dma_reserve = new_dma_reserve;
7647 static int page_alloc_cpu_dead(unsigned int cpu)
7650 lru_add_drain_cpu(cpu);
7654 * Spill the event counters of the dead processor
7655 * into the current processors event counters.
7656 * This artificially elevates the count of the current
7659 vm_events_fold_cpu(cpu);
7662 * Zero the differential counters of the dead processor
7663 * so that the vm statistics are consistent.
7665 * This is only okay since the processor is dead and cannot
7666 * race with what we are doing.
7668 cpu_vm_stats_fold(cpu);
7673 int hashdist = HASHDIST_DEFAULT;
7675 static int __init set_hashdist(char *str)
7679 hashdist = simple_strtoul(str, &str, 0);
7682 __setup("hashdist=", set_hashdist);
7685 void __init page_alloc_init(void)
7690 if (num_node_state(N_MEMORY) == 1)
7694 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7695 "mm/page_alloc:dead", NULL,
7696 page_alloc_cpu_dead);
7701 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7702 * or min_free_kbytes changes.
7704 static void calculate_totalreserve_pages(void)
7706 struct pglist_data *pgdat;
7707 unsigned long reserve_pages = 0;
7708 enum zone_type i, j;
7710 for_each_online_pgdat(pgdat) {
7712 pgdat->totalreserve_pages = 0;
7714 for (i = 0; i < MAX_NR_ZONES; i++) {
7715 struct zone *zone = pgdat->node_zones + i;
7717 unsigned long managed_pages = zone_managed_pages(zone);
7719 /* Find valid and maximum lowmem_reserve in the zone */
7720 for (j = i; j < MAX_NR_ZONES; j++) {
7721 if (zone->lowmem_reserve[j] > max)
7722 max = zone->lowmem_reserve[j];
7725 /* we treat the high watermark as reserved pages. */
7726 max += high_wmark_pages(zone);
7728 if (max > managed_pages)
7729 max = managed_pages;
7731 pgdat->totalreserve_pages += max;
7733 reserve_pages += max;
7736 totalreserve_pages = reserve_pages;
7740 * setup_per_zone_lowmem_reserve - called whenever
7741 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7742 * has a correct pages reserved value, so an adequate number of
7743 * pages are left in the zone after a successful __alloc_pages().
7745 static void setup_per_zone_lowmem_reserve(void)
7747 struct pglist_data *pgdat;
7748 enum zone_type j, idx;
7750 for_each_online_pgdat(pgdat) {
7751 for (j = 0; j < MAX_NR_ZONES; j++) {
7752 struct zone *zone = pgdat->node_zones + j;
7753 unsigned long managed_pages = zone_managed_pages(zone);
7755 zone->lowmem_reserve[j] = 0;
7759 struct zone *lower_zone;
7762 lower_zone = pgdat->node_zones + idx;
7764 if (!sysctl_lowmem_reserve_ratio[idx] ||
7765 !zone_managed_pages(lower_zone)) {
7766 lower_zone->lowmem_reserve[j] = 0;
7769 lower_zone->lowmem_reserve[j] =
7770 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7772 managed_pages += zone_managed_pages(lower_zone);
7777 /* update totalreserve_pages */
7778 calculate_totalreserve_pages();
7781 static void __setup_per_zone_wmarks(void)
7783 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7784 unsigned long lowmem_pages = 0;
7786 unsigned long flags;
7788 /* Calculate total number of !ZONE_HIGHMEM pages */
7789 for_each_zone(zone) {
7790 if (!is_highmem(zone))
7791 lowmem_pages += zone_managed_pages(zone);
7794 for_each_zone(zone) {
7797 spin_lock_irqsave(&zone->lock, flags);
7798 tmp = (u64)pages_min * zone_managed_pages(zone);
7799 do_div(tmp, lowmem_pages);
7800 if (is_highmem(zone)) {
7802 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7803 * need highmem pages, so cap pages_min to a small
7806 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7807 * deltas control async page reclaim, and so should
7808 * not be capped for highmem.
7810 unsigned long min_pages;
7812 min_pages = zone_managed_pages(zone) / 1024;
7813 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7814 zone->_watermark[WMARK_MIN] = min_pages;
7817 * If it's a lowmem zone, reserve a number of pages
7818 * proportionate to the zone's size.
7820 zone->_watermark[WMARK_MIN] = tmp;
7824 * Set the kswapd watermarks distance according to the
7825 * scale factor in proportion to available memory, but
7826 * ensure a minimum size on small systems.
7828 tmp = max_t(u64, tmp >> 2,
7829 mult_frac(zone_managed_pages(zone),
7830 watermark_scale_factor, 10000));
7832 zone->watermark_boost = 0;
7833 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7834 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7836 spin_unlock_irqrestore(&zone->lock, flags);
7839 /* update totalreserve_pages */
7840 calculate_totalreserve_pages();
7844 * setup_per_zone_wmarks - called when min_free_kbytes changes
7845 * or when memory is hot-{added|removed}
7847 * Ensures that the watermark[min,low,high] values for each zone are set
7848 * correctly with respect to min_free_kbytes.
7850 void setup_per_zone_wmarks(void)
7852 static DEFINE_SPINLOCK(lock);
7855 __setup_per_zone_wmarks();
7860 * Initialise min_free_kbytes.
7862 * For small machines we want it small (128k min). For large machines
7863 * we want it large (256MB max). But it is not linear, because network
7864 * bandwidth does not increase linearly with machine size. We use
7866 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7867 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7883 int __meminit init_per_zone_wmark_min(void)
7885 unsigned long lowmem_kbytes;
7886 int new_min_free_kbytes;
7888 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7889 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7891 if (new_min_free_kbytes > user_min_free_kbytes) {
7892 min_free_kbytes = new_min_free_kbytes;
7893 if (min_free_kbytes < 128)
7894 min_free_kbytes = 128;
7895 if (min_free_kbytes > 262144)
7896 min_free_kbytes = 262144;
7898 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7899 new_min_free_kbytes, user_min_free_kbytes);
7901 setup_per_zone_wmarks();
7902 refresh_zone_stat_thresholds();
7903 setup_per_zone_lowmem_reserve();
7906 setup_min_unmapped_ratio();
7907 setup_min_slab_ratio();
7910 khugepaged_min_free_kbytes_update();
7914 postcore_initcall(init_per_zone_wmark_min)
7917 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7918 * that we can call two helper functions whenever min_free_kbytes
7921 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7922 void *buffer, size_t *length, loff_t *ppos)
7926 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7931 user_min_free_kbytes = min_free_kbytes;
7932 setup_per_zone_wmarks();
7937 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7938 void *buffer, size_t *length, loff_t *ppos)
7942 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7947 setup_per_zone_wmarks();
7953 static void setup_min_unmapped_ratio(void)
7958 for_each_online_pgdat(pgdat)
7959 pgdat->min_unmapped_pages = 0;
7962 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7963 sysctl_min_unmapped_ratio) / 100;
7967 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7968 void *buffer, size_t *length, loff_t *ppos)
7972 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7976 setup_min_unmapped_ratio();
7981 static void setup_min_slab_ratio(void)
7986 for_each_online_pgdat(pgdat)
7987 pgdat->min_slab_pages = 0;
7990 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7991 sysctl_min_slab_ratio) / 100;
7994 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7995 void *buffer, size_t *length, loff_t *ppos)
7999 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8003 setup_min_slab_ratio();
8010 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8011 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8012 * whenever sysctl_lowmem_reserve_ratio changes.
8014 * The reserve ratio obviously has absolutely no relation with the
8015 * minimum watermarks. The lowmem reserve ratio can only make sense
8016 * if in function of the boot time zone sizes.
8018 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8019 void *buffer, size_t *length, loff_t *ppos)
8023 proc_dointvec_minmax(table, write, buffer, length, ppos);
8025 for (i = 0; i < MAX_NR_ZONES; i++) {
8026 if (sysctl_lowmem_reserve_ratio[i] < 1)
8027 sysctl_lowmem_reserve_ratio[i] = 0;
8030 setup_per_zone_lowmem_reserve();
8034 static void __zone_pcp_update(struct zone *zone)
8038 for_each_possible_cpu(cpu)
8039 pageset_set_high_and_batch(zone,
8040 per_cpu_ptr(zone->pageset, cpu));
8044 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8045 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8046 * pagelist can have before it gets flushed back to buddy allocator.
8048 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8049 void *buffer, size_t *length, loff_t *ppos)
8052 int old_percpu_pagelist_fraction;
8055 mutex_lock(&pcp_batch_high_lock);
8056 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8058 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8059 if (!write || ret < 0)
8062 /* Sanity checking to avoid pcp imbalance */
8063 if (percpu_pagelist_fraction &&
8064 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8065 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8071 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8074 for_each_populated_zone(zone)
8075 __zone_pcp_update(zone);
8077 mutex_unlock(&pcp_batch_high_lock);
8081 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8083 * Returns the number of pages that arch has reserved but
8084 * is not known to alloc_large_system_hash().
8086 static unsigned long __init arch_reserved_kernel_pages(void)
8093 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8094 * machines. As memory size is increased the scale is also increased but at
8095 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8096 * quadruples the scale is increased by one, which means the size of hash table
8097 * only doubles, instead of quadrupling as well.
8098 * Because 32-bit systems cannot have large physical memory, where this scaling
8099 * makes sense, it is disabled on such platforms.
8101 #if __BITS_PER_LONG > 32
8102 #define ADAPT_SCALE_BASE (64ul << 30)
8103 #define ADAPT_SCALE_SHIFT 2
8104 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8108 * allocate a large system hash table from bootmem
8109 * - it is assumed that the hash table must contain an exact power-of-2
8110 * quantity of entries
8111 * - limit is the number of hash buckets, not the total allocation size
8113 void *__init alloc_large_system_hash(const char *tablename,
8114 unsigned long bucketsize,
8115 unsigned long numentries,
8118 unsigned int *_hash_shift,
8119 unsigned int *_hash_mask,
8120 unsigned long low_limit,
8121 unsigned long high_limit)
8123 unsigned long long max = high_limit;
8124 unsigned long log2qty, size;
8129 /* allow the kernel cmdline to have a say */
8131 /* round applicable memory size up to nearest megabyte */
8132 numentries = nr_kernel_pages;
8133 numentries -= arch_reserved_kernel_pages();
8135 /* It isn't necessary when PAGE_SIZE >= 1MB */
8136 if (PAGE_SHIFT < 20)
8137 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8139 #if __BITS_PER_LONG > 32
8141 unsigned long adapt;
8143 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8144 adapt <<= ADAPT_SCALE_SHIFT)
8149 /* limit to 1 bucket per 2^scale bytes of low memory */
8150 if (scale > PAGE_SHIFT)
8151 numentries >>= (scale - PAGE_SHIFT);
8153 numentries <<= (PAGE_SHIFT - scale);
8155 /* Make sure we've got at least a 0-order allocation.. */
8156 if (unlikely(flags & HASH_SMALL)) {
8157 /* Makes no sense without HASH_EARLY */
8158 WARN_ON(!(flags & HASH_EARLY));
8159 if (!(numentries >> *_hash_shift)) {
8160 numentries = 1UL << *_hash_shift;
8161 BUG_ON(!numentries);
8163 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8164 numentries = PAGE_SIZE / bucketsize;
8166 numentries = roundup_pow_of_two(numentries);
8168 /* limit allocation size to 1/16 total memory by default */
8170 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8171 do_div(max, bucketsize);
8173 max = min(max, 0x80000000ULL);
8175 if (numentries < low_limit)
8176 numentries = low_limit;
8177 if (numentries > max)
8180 log2qty = ilog2(numentries);
8182 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8185 size = bucketsize << log2qty;
8186 if (flags & HASH_EARLY) {
8187 if (flags & HASH_ZERO)
8188 table = memblock_alloc(size, SMP_CACHE_BYTES);
8190 table = memblock_alloc_raw(size,
8192 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8193 table = __vmalloc(size, gfp_flags);
8197 * If bucketsize is not a power-of-two, we may free
8198 * some pages at the end of hash table which
8199 * alloc_pages_exact() automatically does
8201 table = alloc_pages_exact(size, gfp_flags);
8202 kmemleak_alloc(table, size, 1, gfp_flags);
8204 } while (!table && size > PAGE_SIZE && --log2qty);
8207 panic("Failed to allocate %s hash table\n", tablename);
8209 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8210 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8211 virt ? "vmalloc" : "linear");
8214 *_hash_shift = log2qty;
8216 *_hash_mask = (1 << log2qty) - 1;
8222 * This function checks whether pageblock includes unmovable pages or not.
8224 * PageLRU check without isolation or lru_lock could race so that
8225 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8226 * check without lock_page also may miss some movable non-lru pages at
8227 * race condition. So you can't expect this function should be exact.
8229 * Returns a page without holding a reference. If the caller wants to
8230 * dereference that page (e.g., dumping), it has to make sure that it
8231 * cannot get removed (e.g., via memory unplug) concurrently.
8234 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8235 int migratetype, int flags)
8237 unsigned long iter = 0;
8238 unsigned long pfn = page_to_pfn(page);
8239 unsigned long offset = pfn % pageblock_nr_pages;
8241 if (is_migrate_cma_page(page)) {
8243 * CMA allocations (alloc_contig_range) really need to mark
8244 * isolate CMA pageblocks even when they are not movable in fact
8245 * so consider them movable here.
8247 if (is_migrate_cma(migratetype))
8253 for (; iter < pageblock_nr_pages - offset; iter++) {
8254 if (!pfn_valid_within(pfn + iter))
8257 page = pfn_to_page(pfn + iter);
8260 * Both, bootmem allocations and memory holes are marked
8261 * PG_reserved and are unmovable. We can even have unmovable
8262 * allocations inside ZONE_MOVABLE, for example when
8263 * specifying "movablecore".
8265 if (PageReserved(page))
8269 * If the zone is movable and we have ruled out all reserved
8270 * pages then it should be reasonably safe to assume the rest
8273 if (zone_idx(zone) == ZONE_MOVABLE)
8277 * Hugepages are not in LRU lists, but they're movable.
8278 * THPs are on the LRU, but need to be counted as #small pages.
8279 * We need not scan over tail pages because we don't
8280 * handle each tail page individually in migration.
8282 if (PageHuge(page) || PageTransCompound(page)) {
8283 struct page *head = compound_head(page);
8284 unsigned int skip_pages;
8286 if (PageHuge(page)) {
8287 if (!hugepage_migration_supported(page_hstate(head)))
8289 } else if (!PageLRU(head) && !__PageMovable(head)) {
8293 skip_pages = compound_nr(head) - (page - head);
8294 iter += skip_pages - 1;
8299 * We can't use page_count without pin a page
8300 * because another CPU can free compound page.
8301 * This check already skips compound tails of THP
8302 * because their page->_refcount is zero at all time.
8304 if (!page_ref_count(page)) {
8305 if (PageBuddy(page))
8306 iter += (1 << page_order(page)) - 1;
8311 * The HWPoisoned page may be not in buddy system, and
8312 * page_count() is not 0.
8314 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8318 * We treat all PageOffline() pages as movable when offlining
8319 * to give drivers a chance to decrement their reference count
8320 * in MEM_GOING_OFFLINE in order to indicate that these pages
8321 * can be offlined as there are no direct references anymore.
8322 * For actually unmovable PageOffline() where the driver does
8323 * not support this, we will fail later when trying to actually
8324 * move these pages that still have a reference count > 0.
8325 * (false negatives in this function only)
8327 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8330 if (__PageMovable(page) || PageLRU(page))
8334 * If there are RECLAIMABLE pages, we need to check
8335 * it. But now, memory offline itself doesn't call
8336 * shrink_node_slabs() and it still to be fixed.
8343 #ifdef CONFIG_CONTIG_ALLOC
8344 static unsigned long pfn_max_align_down(unsigned long pfn)
8346 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8347 pageblock_nr_pages) - 1);
8350 static unsigned long pfn_max_align_up(unsigned long pfn)
8352 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8353 pageblock_nr_pages));
8356 /* [start, end) must belong to a single zone. */
8357 static int __alloc_contig_migrate_range(struct compact_control *cc,
8358 unsigned long start, unsigned long end)
8360 /* This function is based on compact_zone() from compaction.c. */
8361 unsigned int nr_reclaimed;
8362 unsigned long pfn = start;
8363 unsigned int tries = 0;
8365 struct migration_target_control mtc = {
8366 .nid = zone_to_nid(cc->zone),
8367 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8372 while (pfn < end || !list_empty(&cc->migratepages)) {
8373 if (fatal_signal_pending(current)) {
8378 if (list_empty(&cc->migratepages)) {
8379 cc->nr_migratepages = 0;
8380 pfn = isolate_migratepages_range(cc, pfn, end);
8386 } else if (++tries == 5) {
8387 ret = ret < 0 ? ret : -EBUSY;
8391 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8393 cc->nr_migratepages -= nr_reclaimed;
8395 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
8396 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE);
8399 putback_movable_pages(&cc->migratepages);
8406 * alloc_contig_range() -- tries to allocate given range of pages
8407 * @start: start PFN to allocate
8408 * @end: one-past-the-last PFN to allocate
8409 * @migratetype: migratetype of the underlaying pageblocks (either
8410 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8411 * in range must have the same migratetype and it must
8412 * be either of the two.
8413 * @gfp_mask: GFP mask to use during compaction
8415 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8416 * aligned. The PFN range must belong to a single zone.
8418 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8419 * pageblocks in the range. Once isolated, the pageblocks should not
8420 * be modified by others.
8422 * Return: zero on success or negative error code. On success all
8423 * pages which PFN is in [start, end) are allocated for the caller and
8424 * need to be freed with free_contig_range().
8426 int alloc_contig_range(unsigned long start, unsigned long end,
8427 unsigned migratetype, gfp_t gfp_mask)
8429 unsigned long outer_start, outer_end;
8433 struct compact_control cc = {
8434 .nr_migratepages = 0,
8436 .zone = page_zone(pfn_to_page(start)),
8437 .mode = MIGRATE_SYNC,
8438 .ignore_skip_hint = true,
8439 .no_set_skip_hint = true,
8440 .gfp_mask = current_gfp_context(gfp_mask),
8441 .alloc_contig = true,
8443 INIT_LIST_HEAD(&cc.migratepages);
8446 * What we do here is we mark all pageblocks in range as
8447 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8448 * have different sizes, and due to the way page allocator
8449 * work, we align the range to biggest of the two pages so
8450 * that page allocator won't try to merge buddies from
8451 * different pageblocks and change MIGRATE_ISOLATE to some
8452 * other migration type.
8454 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8455 * migrate the pages from an unaligned range (ie. pages that
8456 * we are interested in). This will put all the pages in
8457 * range back to page allocator as MIGRATE_ISOLATE.
8459 * When this is done, we take the pages in range from page
8460 * allocator removing them from the buddy system. This way
8461 * page allocator will never consider using them.
8463 * This lets us mark the pageblocks back as
8464 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8465 * aligned range but not in the unaligned, original range are
8466 * put back to page allocator so that buddy can use them.
8469 ret = start_isolate_page_range(pfn_max_align_down(start),
8470 pfn_max_align_up(end), migratetype, 0);
8475 * In case of -EBUSY, we'd like to know which page causes problem.
8476 * So, just fall through. test_pages_isolated() has a tracepoint
8477 * which will report the busy page.
8479 * It is possible that busy pages could become available before
8480 * the call to test_pages_isolated, and the range will actually be
8481 * allocated. So, if we fall through be sure to clear ret so that
8482 * -EBUSY is not accidentally used or returned to caller.
8484 ret = __alloc_contig_migrate_range(&cc, start, end);
8485 if (ret && ret != -EBUSY)
8490 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8491 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8492 * more, all pages in [start, end) are free in page allocator.
8493 * What we are going to do is to allocate all pages from
8494 * [start, end) (that is remove them from page allocator).
8496 * The only problem is that pages at the beginning and at the
8497 * end of interesting range may be not aligned with pages that
8498 * page allocator holds, ie. they can be part of higher order
8499 * pages. Because of this, we reserve the bigger range and
8500 * once this is done free the pages we are not interested in.
8502 * We don't have to hold zone->lock here because the pages are
8503 * isolated thus they won't get removed from buddy.
8506 lru_add_drain_all();
8509 outer_start = start;
8510 while (!PageBuddy(pfn_to_page(outer_start))) {
8511 if (++order >= MAX_ORDER) {
8512 outer_start = start;
8515 outer_start &= ~0UL << order;
8518 if (outer_start != start) {
8519 order = page_order(pfn_to_page(outer_start));
8522 * outer_start page could be small order buddy page and
8523 * it doesn't include start page. Adjust outer_start
8524 * in this case to report failed page properly
8525 * on tracepoint in test_pages_isolated()
8527 if (outer_start + (1UL << order) <= start)
8528 outer_start = start;
8531 /* Make sure the range is really isolated. */
8532 if (test_pages_isolated(outer_start, end, 0)) {
8533 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8534 __func__, outer_start, end);
8539 /* Grab isolated pages from freelists. */
8540 outer_end = isolate_freepages_range(&cc, outer_start, end);
8546 /* Free head and tail (if any) */
8547 if (start != outer_start)
8548 free_contig_range(outer_start, start - outer_start);
8549 if (end != outer_end)
8550 free_contig_range(end, outer_end - end);
8553 undo_isolate_page_range(pfn_max_align_down(start),
8554 pfn_max_align_up(end), migratetype);
8557 EXPORT_SYMBOL(alloc_contig_range);
8559 static int __alloc_contig_pages(unsigned long start_pfn,
8560 unsigned long nr_pages, gfp_t gfp_mask)
8562 unsigned long end_pfn = start_pfn + nr_pages;
8564 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8568 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8569 unsigned long nr_pages)
8571 unsigned long i, end_pfn = start_pfn + nr_pages;
8574 for (i = start_pfn; i < end_pfn; i++) {
8575 page = pfn_to_online_page(i);
8579 if (page_zone(page) != z)
8582 if (PageReserved(page))
8585 if (page_count(page) > 0)
8594 static bool zone_spans_last_pfn(const struct zone *zone,
8595 unsigned long start_pfn, unsigned long nr_pages)
8597 unsigned long last_pfn = start_pfn + nr_pages - 1;
8599 return zone_spans_pfn(zone, last_pfn);
8603 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8604 * @nr_pages: Number of contiguous pages to allocate
8605 * @gfp_mask: GFP mask to limit search and used during compaction
8607 * @nodemask: Mask for other possible nodes
8609 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8610 * on an applicable zonelist to find a contiguous pfn range which can then be
8611 * tried for allocation with alloc_contig_range(). This routine is intended
8612 * for allocation requests which can not be fulfilled with the buddy allocator.
8614 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8615 * power of two then the alignment is guaranteed to be to the given nr_pages
8616 * (e.g. 1GB request would be aligned to 1GB).
8618 * Allocated pages can be freed with free_contig_range() or by manually calling
8619 * __free_page() on each allocated page.
8621 * Return: pointer to contiguous pages on success, or NULL if not successful.
8623 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8624 int nid, nodemask_t *nodemask)
8626 unsigned long ret, pfn, flags;
8627 struct zonelist *zonelist;
8631 zonelist = node_zonelist(nid, gfp_mask);
8632 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8633 gfp_zone(gfp_mask), nodemask) {
8634 spin_lock_irqsave(&zone->lock, flags);
8636 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8637 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8638 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8640 * We release the zone lock here because
8641 * alloc_contig_range() will also lock the zone
8642 * at some point. If there's an allocation
8643 * spinning on this lock, it may win the race
8644 * and cause alloc_contig_range() to fail...
8646 spin_unlock_irqrestore(&zone->lock, flags);
8647 ret = __alloc_contig_pages(pfn, nr_pages,
8650 return pfn_to_page(pfn);
8651 spin_lock_irqsave(&zone->lock, flags);
8655 spin_unlock_irqrestore(&zone->lock, flags);
8659 #endif /* CONFIG_CONTIG_ALLOC */
8661 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8663 unsigned int count = 0;
8665 for (; nr_pages--; pfn++) {
8666 struct page *page = pfn_to_page(pfn);
8668 count += page_count(page) != 1;
8671 WARN(count != 0, "%d pages are still in use!\n", count);
8673 EXPORT_SYMBOL(free_contig_range);
8676 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8677 * page high values need to be recalulated.
8679 void __meminit zone_pcp_update(struct zone *zone)
8681 mutex_lock(&pcp_batch_high_lock);
8682 __zone_pcp_update(zone);
8683 mutex_unlock(&pcp_batch_high_lock);
8686 void zone_pcp_reset(struct zone *zone)
8688 unsigned long flags;
8690 struct per_cpu_pageset *pset;
8692 /* avoid races with drain_pages() */
8693 local_irq_save(flags);
8694 if (zone->pageset != &boot_pageset) {
8695 for_each_online_cpu(cpu) {
8696 pset = per_cpu_ptr(zone->pageset, cpu);
8697 drain_zonestat(zone, pset);
8699 free_percpu(zone->pageset);
8700 zone->pageset = &boot_pageset;
8702 local_irq_restore(flags);
8705 #ifdef CONFIG_MEMORY_HOTREMOVE
8707 * All pages in the range must be in a single zone and isolated
8708 * before calling this.
8711 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8717 unsigned long flags;
8718 unsigned long offlined_pages = 0;
8720 /* find the first valid pfn */
8721 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8725 return offlined_pages;
8727 offline_mem_sections(pfn, end_pfn);
8728 zone = page_zone(pfn_to_page(pfn));
8729 spin_lock_irqsave(&zone->lock, flags);
8731 while (pfn < end_pfn) {
8732 if (!pfn_valid(pfn)) {
8736 page = pfn_to_page(pfn);
8738 * The HWPoisoned page may be not in buddy system, and
8739 * page_count() is not 0.
8741 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8747 * At this point all remaining PageOffline() pages have a
8748 * reference count of 0 and can simply be skipped.
8750 if (PageOffline(page)) {
8751 BUG_ON(page_count(page));
8752 BUG_ON(PageBuddy(page));
8758 BUG_ON(page_count(page));
8759 BUG_ON(!PageBuddy(page));
8760 order = page_order(page);
8761 offlined_pages += 1 << order;
8762 del_page_from_free_list(page, zone, order);
8763 pfn += (1 << order);
8765 spin_unlock_irqrestore(&zone->lock, flags);
8767 return offlined_pages;
8771 bool is_free_buddy_page(struct page *page)
8773 struct zone *zone = page_zone(page);
8774 unsigned long pfn = page_to_pfn(page);
8775 unsigned long flags;
8778 spin_lock_irqsave(&zone->lock, flags);
8779 for (order = 0; order < MAX_ORDER; order++) {
8780 struct page *page_head = page - (pfn & ((1 << order) - 1));
8782 if (PageBuddy(page_head) && page_order(page_head) >= order)
8785 spin_unlock_irqrestore(&zone->lock, flags);
8787 return order < MAX_ORDER;
8790 #ifdef CONFIG_MEMORY_FAILURE
8792 * Break down a higher-order page in sub-pages, and keep our target out of
8795 static void break_down_buddy_pages(struct zone *zone, struct page *page,
8796 struct page *target, int low, int high,
8799 unsigned long size = 1 << high;
8800 struct page *current_buddy, *next_page;
8802 while (high > low) {
8806 if (target >= &page[size]) {
8807 next_page = page + size;
8808 current_buddy = page;
8811 current_buddy = page + size;
8814 if (set_page_guard(zone, current_buddy, high, migratetype))
8817 if (current_buddy != target) {
8818 add_to_free_list(current_buddy, zone, high, migratetype);
8819 set_page_order(current_buddy, high);
8826 * Take a page that will be marked as poisoned off the buddy allocator.
8828 bool take_page_off_buddy(struct page *page)
8830 struct zone *zone = page_zone(page);
8831 unsigned long pfn = page_to_pfn(page);
8832 unsigned long flags;
8836 spin_lock_irqsave(&zone->lock, flags);
8837 for (order = 0; order < MAX_ORDER; order++) {
8838 struct page *page_head = page - (pfn & ((1 << order) - 1));
8839 int buddy_order = page_order(page_head);
8841 if (PageBuddy(page_head) && buddy_order >= order) {
8842 unsigned long pfn_head = page_to_pfn(page_head);
8843 int migratetype = get_pfnblock_migratetype(page_head,
8846 del_page_from_free_list(page_head, zone, buddy_order);
8847 break_down_buddy_pages(zone, page_head, page, 0,
8848 buddy_order, migratetype);
8852 if (page_count(page_head) > 0)
8855 spin_unlock_irqrestore(&zone->lock, flags);