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 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
82 typedef int __bitwise fpi_t;
84 /* No special request */
85 #define FPI_NONE ((__force fpi_t)0)
88 * Skip free page reporting notification for the (possibly merged) page.
89 * This does not hinder free page reporting from grabbing the page,
90 * reporting it and marking it "reported" - it only skips notifying
91 * the free page reporting infrastructure about a newly freed page. For
92 * example, used when temporarily pulling a page from a freelist and
93 * putting it back unmodified.
95 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
98 * Place the (possibly merged) page to the tail of the freelist. Will ignore
99 * page shuffling (relevant code - e.g., memory onlining - is expected to
100 * shuffle the whole zone).
102 * Note: No code should rely on this flag for correctness - it's purely
103 * to allow for optimizations when handing back either fresh pages
104 * (memory onlining) or untouched pages (page isolation, free page
107 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
109 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
110 static DEFINE_MUTEX(pcp_batch_high_lock);
111 #define MIN_PERCPU_PAGELIST_FRACTION (8)
113 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
114 DEFINE_PER_CPU(int, numa_node);
115 EXPORT_PER_CPU_SYMBOL(numa_node);
118 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
120 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
122 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
123 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
124 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
125 * defined in <linux/topology.h>.
127 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
128 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
131 /* work_structs for global per-cpu drains */
134 struct work_struct work;
136 static DEFINE_MUTEX(pcpu_drain_mutex);
137 static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
139 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
140 volatile unsigned long latent_entropy __latent_entropy;
141 EXPORT_SYMBOL(latent_entropy);
145 * Array of node states.
147 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
148 [N_POSSIBLE] = NODE_MASK_ALL,
149 [N_ONLINE] = { { [0] = 1UL } },
151 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
152 #ifdef CONFIG_HIGHMEM
153 [N_HIGH_MEMORY] = { { [0] = 1UL } },
155 [N_MEMORY] = { { [0] = 1UL } },
156 [N_CPU] = { { [0] = 1UL } },
159 EXPORT_SYMBOL(node_states);
161 atomic_long_t _totalram_pages __read_mostly;
162 EXPORT_SYMBOL(_totalram_pages);
163 unsigned long totalreserve_pages __read_mostly;
164 unsigned long totalcma_pages __read_mostly;
166 int percpu_pagelist_fraction;
167 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
168 #ifdef CONFIG_INIT_ON_ALLOC_DEFAULT_ON
169 DEFINE_STATIC_KEY_TRUE(init_on_alloc);
171 DEFINE_STATIC_KEY_FALSE(init_on_alloc);
173 EXPORT_SYMBOL(init_on_alloc);
175 #ifdef CONFIG_INIT_ON_FREE_DEFAULT_ON
176 DEFINE_STATIC_KEY_TRUE(init_on_free);
178 DEFINE_STATIC_KEY_FALSE(init_on_free);
180 EXPORT_SYMBOL(init_on_free);
182 static int __init early_init_on_alloc(char *buf)
187 ret = kstrtobool(buf, &bool_result);
190 if (bool_result && page_poisoning_enabled())
191 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_alloc\n");
193 static_branch_enable(&init_on_alloc);
195 static_branch_disable(&init_on_alloc);
198 early_param("init_on_alloc", early_init_on_alloc);
200 static int __init early_init_on_free(char *buf)
205 ret = kstrtobool(buf, &bool_result);
208 if (bool_result && page_poisoning_enabled())
209 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_free\n");
211 static_branch_enable(&init_on_free);
213 static_branch_disable(&init_on_free);
216 early_param("init_on_free", early_init_on_free);
219 * A cached value of the page's pageblock's migratetype, used when the page is
220 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
221 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
222 * Also the migratetype set in the page does not necessarily match the pcplist
223 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
224 * other index - this ensures that it will be put on the correct CMA freelist.
226 static inline int get_pcppage_migratetype(struct page *page)
231 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
233 page->index = migratetype;
236 #ifdef CONFIG_PM_SLEEP
238 * The following functions are used by the suspend/hibernate code to temporarily
239 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
240 * while devices are suspended. To avoid races with the suspend/hibernate code,
241 * they should always be called with system_transition_mutex held
242 * (gfp_allowed_mask also should only be modified with system_transition_mutex
243 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
244 * with that modification).
247 static gfp_t saved_gfp_mask;
249 void pm_restore_gfp_mask(void)
251 WARN_ON(!mutex_is_locked(&system_transition_mutex));
252 if (saved_gfp_mask) {
253 gfp_allowed_mask = saved_gfp_mask;
258 void pm_restrict_gfp_mask(void)
260 WARN_ON(!mutex_is_locked(&system_transition_mutex));
261 WARN_ON(saved_gfp_mask);
262 saved_gfp_mask = gfp_allowed_mask;
263 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
266 bool pm_suspended_storage(void)
268 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
272 #endif /* CONFIG_PM_SLEEP */
274 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
275 unsigned int pageblock_order __read_mostly;
278 static void __free_pages_ok(struct page *page, unsigned int order);
281 * results with 256, 32 in the lowmem_reserve sysctl:
282 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
283 * 1G machine -> (16M dma, 784M normal, 224M high)
284 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
285 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
286 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
288 * TBD: should special case ZONE_DMA32 machines here - in those we normally
289 * don't need any ZONE_NORMAL reservation
291 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
292 #ifdef CONFIG_ZONE_DMA
295 #ifdef CONFIG_ZONE_DMA32
299 #ifdef CONFIG_HIGHMEM
305 static char * const zone_names[MAX_NR_ZONES] = {
306 #ifdef CONFIG_ZONE_DMA
309 #ifdef CONFIG_ZONE_DMA32
313 #ifdef CONFIG_HIGHMEM
317 #ifdef CONFIG_ZONE_DEVICE
322 const char * const migratetype_names[MIGRATE_TYPES] = {
330 #ifdef CONFIG_MEMORY_ISOLATION
335 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
336 [NULL_COMPOUND_DTOR] = NULL,
337 [COMPOUND_PAGE_DTOR] = free_compound_page,
338 #ifdef CONFIG_HUGETLB_PAGE
339 [HUGETLB_PAGE_DTOR] = free_huge_page,
341 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
342 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
346 int min_free_kbytes = 1024;
347 int user_min_free_kbytes = -1;
348 #ifdef CONFIG_DISCONTIGMEM
350 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
351 * are not on separate NUMA nodes. Functionally this works but with
352 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
353 * quite small. By default, do not boost watermarks on discontigmem as in
354 * many cases very high-order allocations like THP are likely to be
355 * unsupported and the premature reclaim offsets the advantage of long-term
356 * fragmentation avoidance.
358 int watermark_boost_factor __read_mostly;
360 int watermark_boost_factor __read_mostly = 15000;
362 int watermark_scale_factor = 10;
364 static unsigned long nr_kernel_pages __initdata;
365 static unsigned long nr_all_pages __initdata;
366 static unsigned long dma_reserve __initdata;
368 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
369 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
370 static unsigned long required_kernelcore __initdata;
371 static unsigned long required_kernelcore_percent __initdata;
372 static unsigned long required_movablecore __initdata;
373 static unsigned long required_movablecore_percent __initdata;
374 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
375 static bool mirrored_kernelcore __meminitdata;
377 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
379 EXPORT_SYMBOL(movable_zone);
382 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
383 unsigned int nr_online_nodes __read_mostly = 1;
384 EXPORT_SYMBOL(nr_node_ids);
385 EXPORT_SYMBOL(nr_online_nodes);
388 int page_group_by_mobility_disabled __read_mostly;
390 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
392 * During boot we initialize deferred pages on-demand, as needed, but once
393 * page_alloc_init_late() has finished, the deferred pages are all initialized,
394 * and we can permanently disable that path.
396 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
399 * Calling kasan_free_pages() only after deferred memory initialization
400 * has completed. Poisoning pages during deferred memory init will greatly
401 * lengthen the process and cause problem in large memory systems as the
402 * deferred pages initialization is done with interrupt disabled.
404 * Assuming that there will be no reference to those newly initialized
405 * pages before they are ever allocated, this should have no effect on
406 * KASAN memory tracking as the poison will be properly inserted at page
407 * allocation time. The only corner case is when pages are allocated by
408 * on-demand allocation and then freed again before the deferred pages
409 * initialization is done, but this is not likely to happen.
411 static inline void kasan_free_nondeferred_pages(struct page *page, int order)
413 if (!static_branch_unlikely(&deferred_pages))
414 kasan_free_pages(page, order);
417 /* Returns true if the struct page for the pfn is uninitialised */
418 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
420 int nid = early_pfn_to_nid(pfn);
422 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
429 * Returns true when the remaining initialisation should be deferred until
430 * later in the boot cycle when it can be parallelised.
432 static bool __meminit
433 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
435 static unsigned long prev_end_pfn, nr_initialised;
438 * prev_end_pfn static that contains the end of previous zone
439 * No need to protect because called very early in boot before smp_init.
441 if (prev_end_pfn != end_pfn) {
442 prev_end_pfn = end_pfn;
446 /* Always populate low zones for address-constrained allocations */
447 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
451 * We start only with one section of pages, more pages are added as
452 * needed until the rest of deferred pages are initialized.
455 if ((nr_initialised > PAGES_PER_SECTION) &&
456 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
457 NODE_DATA(nid)->first_deferred_pfn = pfn;
463 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
465 static inline bool early_page_uninitialised(unsigned long pfn)
470 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
476 /* Return a pointer to the bitmap storing bits affecting a block of pages */
477 static inline unsigned long *get_pageblock_bitmap(struct page *page,
480 #ifdef CONFIG_SPARSEMEM
481 return section_to_usemap(__pfn_to_section(pfn));
483 return page_zone(page)->pageblock_flags;
484 #endif /* CONFIG_SPARSEMEM */
487 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
489 #ifdef CONFIG_SPARSEMEM
490 pfn &= (PAGES_PER_SECTION-1);
492 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
493 #endif /* CONFIG_SPARSEMEM */
494 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
498 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
499 * @page: The page within the block of interest
500 * @pfn: The target page frame number
501 * @mask: mask of bits that the caller is interested in
503 * Return: pageblock_bits flags
505 static __always_inline
506 unsigned long __get_pfnblock_flags_mask(struct page *page,
510 unsigned long *bitmap;
511 unsigned long bitidx, word_bitidx;
514 bitmap = get_pageblock_bitmap(page, pfn);
515 bitidx = pfn_to_bitidx(page, pfn);
516 word_bitidx = bitidx / BITS_PER_LONG;
517 bitidx &= (BITS_PER_LONG-1);
519 word = bitmap[word_bitidx];
520 return (word >> bitidx) & mask;
523 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
526 return __get_pfnblock_flags_mask(page, pfn, mask);
529 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
531 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
535 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
536 * @page: The page within the block of interest
537 * @flags: The flags to set
538 * @pfn: The target page frame number
539 * @mask: mask of bits that the caller is interested in
541 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
545 unsigned long *bitmap;
546 unsigned long bitidx, word_bitidx;
547 unsigned long old_word, word;
549 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
550 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
552 bitmap = get_pageblock_bitmap(page, pfn);
553 bitidx = pfn_to_bitidx(page, pfn);
554 word_bitidx = bitidx / BITS_PER_LONG;
555 bitidx &= (BITS_PER_LONG-1);
557 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
562 word = READ_ONCE(bitmap[word_bitidx]);
564 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
565 if (word == old_word)
571 void set_pageblock_migratetype(struct page *page, int migratetype)
573 if (unlikely(page_group_by_mobility_disabled &&
574 migratetype < MIGRATE_PCPTYPES))
575 migratetype = MIGRATE_UNMOVABLE;
577 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
578 page_to_pfn(page), MIGRATETYPE_MASK);
581 #ifdef CONFIG_DEBUG_VM
582 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
586 unsigned long pfn = page_to_pfn(page);
587 unsigned long sp, start_pfn;
590 seq = zone_span_seqbegin(zone);
591 start_pfn = zone->zone_start_pfn;
592 sp = zone->spanned_pages;
593 if (!zone_spans_pfn(zone, pfn))
595 } while (zone_span_seqretry(zone, seq));
598 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
599 pfn, zone_to_nid(zone), zone->name,
600 start_pfn, start_pfn + sp);
605 static int page_is_consistent(struct zone *zone, struct page *page)
607 if (!pfn_valid_within(page_to_pfn(page)))
609 if (zone != page_zone(page))
615 * Temporary debugging check for pages not lying within a given zone.
617 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
619 if (page_outside_zone_boundaries(zone, page))
621 if (!page_is_consistent(zone, page))
627 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
633 static void bad_page(struct page *page, const char *reason)
635 static unsigned long resume;
636 static unsigned long nr_shown;
637 static unsigned long nr_unshown;
640 * Allow a burst of 60 reports, then keep quiet for that minute;
641 * or allow a steady drip of one report per second.
643 if (nr_shown == 60) {
644 if (time_before(jiffies, resume)) {
650 "BUG: Bad page state: %lu messages suppressed\n",
657 resume = jiffies + 60 * HZ;
659 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
660 current->comm, page_to_pfn(page));
661 __dump_page(page, reason);
662 dump_page_owner(page);
667 /* Leave bad fields for debug, except PageBuddy could make trouble */
668 page_mapcount_reset(page); /* remove PageBuddy */
669 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
673 * Higher-order pages are called "compound pages". They are structured thusly:
675 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
677 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
678 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
680 * The first tail page's ->compound_dtor holds the offset in array of compound
681 * page destructors. See compound_page_dtors.
683 * The first tail page's ->compound_order holds the order of allocation.
684 * This usage means that zero-order pages may not be compound.
687 void free_compound_page(struct page *page)
689 mem_cgroup_uncharge(page);
690 __free_pages_ok(page, compound_order(page));
693 void prep_compound_page(struct page *page, unsigned int order)
696 int nr_pages = 1 << order;
699 for (i = 1; i < nr_pages; i++) {
700 struct page *p = page + i;
701 set_page_count(p, 0);
702 p->mapping = TAIL_MAPPING;
703 set_compound_head(p, page);
706 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
707 set_compound_order(page, order);
708 atomic_set(compound_mapcount_ptr(page), -1);
709 if (hpage_pincount_available(page))
710 atomic_set(compound_pincount_ptr(page), 0);
713 #ifdef CONFIG_DEBUG_PAGEALLOC
714 unsigned int _debug_guardpage_minorder;
716 bool _debug_pagealloc_enabled_early __read_mostly
717 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
718 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
719 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
720 EXPORT_SYMBOL(_debug_pagealloc_enabled);
722 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
724 static int __init early_debug_pagealloc(char *buf)
726 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
728 early_param("debug_pagealloc", early_debug_pagealloc);
730 void init_debug_pagealloc(void)
732 if (!debug_pagealloc_enabled())
735 static_branch_enable(&_debug_pagealloc_enabled);
737 if (!debug_guardpage_minorder())
740 static_branch_enable(&_debug_guardpage_enabled);
743 static int __init debug_guardpage_minorder_setup(char *buf)
747 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
748 pr_err("Bad debug_guardpage_minorder value\n");
751 _debug_guardpage_minorder = res;
752 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
755 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
757 static inline bool set_page_guard(struct zone *zone, struct page *page,
758 unsigned int order, int migratetype)
760 if (!debug_guardpage_enabled())
763 if (order >= debug_guardpage_minorder())
766 __SetPageGuard(page);
767 INIT_LIST_HEAD(&page->lru);
768 set_page_private(page, order);
769 /* Guard pages are not available for any usage */
770 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
775 static inline void clear_page_guard(struct zone *zone, struct page *page,
776 unsigned int order, int migratetype)
778 if (!debug_guardpage_enabled())
781 __ClearPageGuard(page);
783 set_page_private(page, 0);
784 if (!is_migrate_isolate(migratetype))
785 __mod_zone_freepage_state(zone, (1 << order), migratetype);
788 static inline bool set_page_guard(struct zone *zone, struct page *page,
789 unsigned int order, int migratetype) { return false; }
790 static inline void clear_page_guard(struct zone *zone, struct page *page,
791 unsigned int order, int migratetype) {}
794 static inline void set_page_order(struct page *page, unsigned int order)
796 set_page_private(page, order);
797 __SetPageBuddy(page);
801 * This function checks whether a page is free && is the buddy
802 * we can coalesce a page and its buddy if
803 * (a) the buddy is not in a hole (check before calling!) &&
804 * (b) the buddy is in the buddy system &&
805 * (c) a page and its buddy have the same order &&
806 * (d) a page and its buddy are in the same zone.
808 * For recording whether a page is in the buddy system, we set PageBuddy.
809 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
811 * For recording page's order, we use page_private(page).
813 static inline bool page_is_buddy(struct page *page, struct page *buddy,
816 if (!page_is_guard(buddy) && !PageBuddy(buddy))
819 if (page_order(buddy) != order)
823 * zone check is done late to avoid uselessly calculating
824 * zone/node ids for pages that could never merge.
826 if (page_zone_id(page) != page_zone_id(buddy))
829 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
834 #ifdef CONFIG_COMPACTION
835 static inline struct capture_control *task_capc(struct zone *zone)
837 struct capture_control *capc = current->capture_control;
839 return unlikely(capc) &&
840 !(current->flags & PF_KTHREAD) &&
842 capc->cc->zone == zone ? capc : NULL;
846 compaction_capture(struct capture_control *capc, struct page *page,
847 int order, int migratetype)
849 if (!capc || order != capc->cc->order)
852 /* Do not accidentally pollute CMA or isolated regions*/
853 if (is_migrate_cma(migratetype) ||
854 is_migrate_isolate(migratetype))
858 * Do not let lower order allocations polluate a movable pageblock.
859 * This might let an unmovable request use a reclaimable pageblock
860 * and vice-versa but no more than normal fallback logic which can
861 * have trouble finding a high-order free page.
863 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
871 static inline struct capture_control *task_capc(struct zone *zone)
877 compaction_capture(struct capture_control *capc, struct page *page,
878 int order, int migratetype)
882 #endif /* CONFIG_COMPACTION */
884 /* Used for pages not on another list */
885 static inline void add_to_free_list(struct page *page, struct zone *zone,
886 unsigned int order, int migratetype)
888 struct free_area *area = &zone->free_area[order];
890 list_add(&page->lru, &area->free_list[migratetype]);
894 /* Used for pages not on another list */
895 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
896 unsigned int order, int migratetype)
898 struct free_area *area = &zone->free_area[order];
900 list_add_tail(&page->lru, &area->free_list[migratetype]);
905 * Used for pages which are on another list. Move the pages to the tail
906 * of the list - so the moved pages won't immediately be considered for
907 * allocation again (e.g., optimization for memory onlining).
909 static inline void move_to_free_list(struct page *page, struct zone *zone,
910 unsigned int order, int migratetype)
912 struct free_area *area = &zone->free_area[order];
914 list_move_tail(&page->lru, &area->free_list[migratetype]);
917 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
920 /* clear reported state and update reported page count */
921 if (page_reported(page))
922 __ClearPageReported(page);
924 list_del(&page->lru);
925 __ClearPageBuddy(page);
926 set_page_private(page, 0);
927 zone->free_area[order].nr_free--;
931 * If this is not the largest possible page, check if the buddy
932 * of the next-highest order is free. If it is, it's possible
933 * that pages are being freed that will coalesce soon. In case,
934 * that is happening, add the free page to the tail of the list
935 * so it's less likely to be used soon and more likely to be merged
936 * as a higher order page
939 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
940 struct page *page, unsigned int order)
942 struct page *higher_page, *higher_buddy;
943 unsigned long combined_pfn;
945 if (order >= MAX_ORDER - 2)
948 if (!pfn_valid_within(buddy_pfn))
951 combined_pfn = buddy_pfn & pfn;
952 higher_page = page + (combined_pfn - pfn);
953 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
954 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
956 return pfn_valid_within(buddy_pfn) &&
957 page_is_buddy(higher_page, higher_buddy, order + 1);
961 * Freeing function for a buddy system allocator.
963 * The concept of a buddy system is to maintain direct-mapped table
964 * (containing bit values) for memory blocks of various "orders".
965 * The bottom level table contains the map for the smallest allocatable
966 * units of memory (here, pages), and each level above it describes
967 * pairs of units from the levels below, hence, "buddies".
968 * At a high level, all that happens here is marking the table entry
969 * at the bottom level available, and propagating the changes upward
970 * as necessary, plus some accounting needed to play nicely with other
971 * parts of the VM system.
972 * At each level, we keep a list of pages, which are heads of continuous
973 * free pages of length of (1 << order) and marked with PageBuddy.
974 * Page's order is recorded in page_private(page) field.
975 * So when we are allocating or freeing one, we can derive the state of the
976 * other. That is, if we allocate a small block, and both were
977 * free, the remainder of the region must be split into blocks.
978 * If a block is freed, and its buddy is also free, then this
979 * triggers coalescing into a block of larger size.
984 static inline void __free_one_page(struct page *page,
986 struct zone *zone, unsigned int order,
987 int migratetype, fpi_t fpi_flags)
989 struct capture_control *capc = task_capc(zone);
990 unsigned long buddy_pfn;
991 unsigned long combined_pfn;
992 unsigned int max_order;
996 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
998 VM_BUG_ON(!zone_is_initialized(zone));
999 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1001 VM_BUG_ON(migratetype == -1);
1002 if (likely(!is_migrate_isolate(migratetype)))
1003 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1005 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1006 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1009 while (order < max_order - 1) {
1010 if (compaction_capture(capc, page, order, migratetype)) {
1011 __mod_zone_freepage_state(zone, -(1 << order),
1015 buddy_pfn = __find_buddy_pfn(pfn, order);
1016 buddy = page + (buddy_pfn - pfn);
1018 if (!pfn_valid_within(buddy_pfn))
1020 if (!page_is_buddy(page, buddy, order))
1023 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1024 * merge with it and move up one order.
1026 if (page_is_guard(buddy))
1027 clear_page_guard(zone, buddy, order, migratetype);
1029 del_page_from_free_list(buddy, zone, order);
1030 combined_pfn = buddy_pfn & pfn;
1031 page = page + (combined_pfn - pfn);
1035 if (max_order < MAX_ORDER) {
1036 /* If we are here, it means order is >= pageblock_order.
1037 * We want to prevent merge between freepages on isolate
1038 * pageblock and normal pageblock. Without this, pageblock
1039 * isolation could cause incorrect freepage or CMA accounting.
1041 * We don't want to hit this code for the more frequent
1042 * low-order merging.
1044 if (unlikely(has_isolate_pageblock(zone))) {
1047 buddy_pfn = __find_buddy_pfn(pfn, order);
1048 buddy = page + (buddy_pfn - pfn);
1049 buddy_mt = get_pageblock_migratetype(buddy);
1051 if (migratetype != buddy_mt
1052 && (is_migrate_isolate(migratetype) ||
1053 is_migrate_isolate(buddy_mt)))
1057 goto continue_merging;
1061 set_page_order(page, order);
1063 if (fpi_flags & FPI_TO_TAIL)
1065 else if (is_shuffle_order(order))
1066 to_tail = shuffle_pick_tail();
1068 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1071 add_to_free_list_tail(page, zone, order, migratetype);
1073 add_to_free_list(page, zone, order, migratetype);
1075 /* Notify page reporting subsystem of freed page */
1076 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1077 page_reporting_notify_free(order);
1081 * A bad page could be due to a number of fields. Instead of multiple branches,
1082 * try and check multiple fields with one check. The caller must do a detailed
1083 * check if necessary.
1085 static inline bool page_expected_state(struct page *page,
1086 unsigned long check_flags)
1088 if (unlikely(atomic_read(&page->_mapcount) != -1))
1091 if (unlikely((unsigned long)page->mapping |
1092 page_ref_count(page) |
1094 (unsigned long)page->mem_cgroup |
1096 (page->flags & check_flags)))
1102 static const char *page_bad_reason(struct page *page, unsigned long flags)
1104 const char *bad_reason = NULL;
1106 if (unlikely(atomic_read(&page->_mapcount) != -1))
1107 bad_reason = "nonzero mapcount";
1108 if (unlikely(page->mapping != NULL))
1109 bad_reason = "non-NULL mapping";
1110 if (unlikely(page_ref_count(page) != 0))
1111 bad_reason = "nonzero _refcount";
1112 if (unlikely(page->flags & flags)) {
1113 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1114 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1116 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1119 if (unlikely(page->mem_cgroup))
1120 bad_reason = "page still charged to cgroup";
1125 static void check_free_page_bad(struct page *page)
1128 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1131 static inline int check_free_page(struct page *page)
1133 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1136 /* Something has gone sideways, find it */
1137 check_free_page_bad(page);
1141 static int free_tail_pages_check(struct page *head_page, struct page *page)
1146 * We rely page->lru.next never has bit 0 set, unless the page
1147 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1149 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1151 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1155 switch (page - head_page) {
1157 /* the first tail page: ->mapping may be compound_mapcount() */
1158 if (unlikely(compound_mapcount(page))) {
1159 bad_page(page, "nonzero compound_mapcount");
1165 * the second tail page: ->mapping is
1166 * deferred_list.next -- ignore value.
1170 if (page->mapping != TAIL_MAPPING) {
1171 bad_page(page, "corrupted mapping in tail page");
1176 if (unlikely(!PageTail(page))) {
1177 bad_page(page, "PageTail not set");
1180 if (unlikely(compound_head(page) != head_page)) {
1181 bad_page(page, "compound_head not consistent");
1186 page->mapping = NULL;
1187 clear_compound_head(page);
1191 static void kernel_init_free_pages(struct page *page, int numpages)
1195 /* s390's use of memset() could override KASAN redzones. */
1196 kasan_disable_current();
1197 for (i = 0; i < numpages; i++)
1198 clear_highpage(page + i);
1199 kasan_enable_current();
1202 static __always_inline bool free_pages_prepare(struct page *page,
1203 unsigned int order, bool check_free)
1207 VM_BUG_ON_PAGE(PageTail(page), page);
1209 trace_mm_page_free(page, order);
1211 if (unlikely(PageHWPoison(page)) && !order) {
1213 * Do not let hwpoison pages hit pcplists/buddy
1214 * Untie memcg state and reset page's owner
1216 if (memcg_kmem_enabled() && PageKmemcg(page))
1217 __memcg_kmem_uncharge_page(page, order);
1218 reset_page_owner(page, order);
1223 * Check tail pages before head page information is cleared to
1224 * avoid checking PageCompound for order-0 pages.
1226 if (unlikely(order)) {
1227 bool compound = PageCompound(page);
1230 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1233 ClearPageDoubleMap(page);
1234 for (i = 1; i < (1 << order); i++) {
1236 bad += free_tail_pages_check(page, page + i);
1237 if (unlikely(check_free_page(page + i))) {
1241 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1244 if (PageMappingFlags(page))
1245 page->mapping = NULL;
1246 if (memcg_kmem_enabled() && PageKmemcg(page))
1247 __memcg_kmem_uncharge_page(page, order);
1249 bad += check_free_page(page);
1253 page_cpupid_reset_last(page);
1254 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1255 reset_page_owner(page, order);
1257 if (!PageHighMem(page)) {
1258 debug_check_no_locks_freed(page_address(page),
1259 PAGE_SIZE << order);
1260 debug_check_no_obj_freed(page_address(page),
1261 PAGE_SIZE << order);
1263 if (want_init_on_free())
1264 kernel_init_free_pages(page, 1 << order);
1266 kernel_poison_pages(page, 1 << order, 0);
1268 * arch_free_page() can make the page's contents inaccessible. s390
1269 * does this. So nothing which can access the page's contents should
1270 * happen after this.
1272 arch_free_page(page, order);
1274 if (debug_pagealloc_enabled_static())
1275 kernel_map_pages(page, 1 << order, 0);
1277 kasan_free_nondeferred_pages(page, order);
1282 #ifdef CONFIG_DEBUG_VM
1284 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1285 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1286 * moved from pcp lists to free lists.
1288 static bool free_pcp_prepare(struct page *page)
1290 return free_pages_prepare(page, 0, true);
1293 static bool bulkfree_pcp_prepare(struct page *page)
1295 if (debug_pagealloc_enabled_static())
1296 return check_free_page(page);
1302 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1303 * moving from pcp lists to free list in order to reduce overhead. With
1304 * debug_pagealloc enabled, they are checked also immediately when being freed
1307 static bool free_pcp_prepare(struct page *page)
1309 if (debug_pagealloc_enabled_static())
1310 return free_pages_prepare(page, 0, true);
1312 return free_pages_prepare(page, 0, false);
1315 static bool bulkfree_pcp_prepare(struct page *page)
1317 return check_free_page(page);
1319 #endif /* CONFIG_DEBUG_VM */
1321 static inline void prefetch_buddy(struct page *page)
1323 unsigned long pfn = page_to_pfn(page);
1324 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1325 struct page *buddy = page + (buddy_pfn - pfn);
1331 * Frees a number of pages from the PCP lists
1332 * Assumes all pages on list are in same zone, and of same order.
1333 * count is the number of pages to free.
1335 * If the zone was previously in an "all pages pinned" state then look to
1336 * see if this freeing clears that state.
1338 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1339 * pinned" detection logic.
1341 static void free_pcppages_bulk(struct zone *zone, int count,
1342 struct per_cpu_pages *pcp)
1344 int migratetype = 0;
1346 int prefetch_nr = 0;
1347 bool isolated_pageblocks;
1348 struct page *page, *tmp;
1352 * Ensure proper count is passed which otherwise would stuck in the
1353 * below while (list_empty(list)) loop.
1355 count = min(pcp->count, count);
1357 struct list_head *list;
1360 * Remove pages from lists in a round-robin fashion. A
1361 * batch_free count is maintained that is incremented when an
1362 * empty list is encountered. This is so more pages are freed
1363 * off fuller lists instead of spinning excessively around empty
1368 if (++migratetype == MIGRATE_PCPTYPES)
1370 list = &pcp->lists[migratetype];
1371 } while (list_empty(list));
1373 /* This is the only non-empty list. Free them all. */
1374 if (batch_free == MIGRATE_PCPTYPES)
1378 page = list_last_entry(list, struct page, lru);
1379 /* must delete to avoid corrupting pcp list */
1380 list_del(&page->lru);
1383 if (bulkfree_pcp_prepare(page))
1386 list_add_tail(&page->lru, &head);
1389 * We are going to put the page back to the global
1390 * pool, prefetch its buddy to speed up later access
1391 * under zone->lock. It is believed the overhead of
1392 * an additional test and calculating buddy_pfn here
1393 * can be offset by reduced memory latency later. To
1394 * avoid excessive prefetching due to large count, only
1395 * prefetch buddy for the first pcp->batch nr of pages.
1397 if (prefetch_nr++ < pcp->batch)
1398 prefetch_buddy(page);
1399 } while (--count && --batch_free && !list_empty(list));
1402 spin_lock(&zone->lock);
1403 isolated_pageblocks = has_isolate_pageblock(zone);
1406 * Use safe version since after __free_one_page(),
1407 * page->lru.next will not point to original list.
1409 list_for_each_entry_safe(page, tmp, &head, lru) {
1410 int mt = get_pcppage_migratetype(page);
1411 /* MIGRATE_ISOLATE page should not go to pcplists */
1412 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1413 /* Pageblock could have been isolated meanwhile */
1414 if (unlikely(isolated_pageblocks))
1415 mt = get_pageblock_migratetype(page);
1417 __free_one_page(page, page_to_pfn(page), zone, 0, mt, FPI_NONE);
1418 trace_mm_page_pcpu_drain(page, 0, mt);
1420 spin_unlock(&zone->lock);
1423 static void free_one_page(struct zone *zone,
1424 struct page *page, unsigned long pfn,
1428 spin_lock(&zone->lock);
1429 if (unlikely(has_isolate_pageblock(zone) ||
1430 is_migrate_isolate(migratetype))) {
1431 migratetype = get_pfnblock_migratetype(page, pfn);
1433 __free_one_page(page, pfn, zone, order, migratetype, FPI_NONE);
1434 spin_unlock(&zone->lock);
1437 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1438 unsigned long zone, int nid)
1440 mm_zero_struct_page(page);
1441 set_page_links(page, zone, nid, pfn);
1442 init_page_count(page);
1443 page_mapcount_reset(page);
1444 page_cpupid_reset_last(page);
1445 page_kasan_tag_reset(page);
1447 INIT_LIST_HEAD(&page->lru);
1448 #ifdef WANT_PAGE_VIRTUAL
1449 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1450 if (!is_highmem_idx(zone))
1451 set_page_address(page, __va(pfn << PAGE_SHIFT));
1455 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1456 static void __meminit init_reserved_page(unsigned long pfn)
1461 if (!early_page_uninitialised(pfn))
1464 nid = early_pfn_to_nid(pfn);
1465 pgdat = NODE_DATA(nid);
1467 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1468 struct zone *zone = &pgdat->node_zones[zid];
1470 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1473 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1476 static inline void init_reserved_page(unsigned long pfn)
1479 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1482 * Initialised pages do not have PageReserved set. This function is
1483 * called for each range allocated by the bootmem allocator and
1484 * marks the pages PageReserved. The remaining valid pages are later
1485 * sent to the buddy page allocator.
1487 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1489 unsigned long start_pfn = PFN_DOWN(start);
1490 unsigned long end_pfn = PFN_UP(end);
1492 for (; start_pfn < end_pfn; start_pfn++) {
1493 if (pfn_valid(start_pfn)) {
1494 struct page *page = pfn_to_page(start_pfn);
1496 init_reserved_page(start_pfn);
1498 /* Avoid false-positive PageTail() */
1499 INIT_LIST_HEAD(&page->lru);
1502 * no need for atomic set_bit because the struct
1503 * page is not visible yet so nobody should
1506 __SetPageReserved(page);
1511 static void __free_pages_ok(struct page *page, unsigned int order)
1513 unsigned long flags;
1515 unsigned long pfn = page_to_pfn(page);
1517 if (!free_pages_prepare(page, order, true))
1520 migratetype = get_pfnblock_migratetype(page, pfn);
1521 local_irq_save(flags);
1522 __count_vm_events(PGFREE, 1 << order);
1523 free_one_page(page_zone(page), page, pfn, order, migratetype);
1524 local_irq_restore(flags);
1527 void __free_pages_core(struct page *page, unsigned int order)
1529 unsigned int nr_pages = 1 << order;
1530 struct page *p = page;
1534 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1536 __ClearPageReserved(p);
1537 set_page_count(p, 0);
1539 __ClearPageReserved(p);
1540 set_page_count(p, 0);
1542 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1543 set_page_refcounted(page);
1544 __free_pages(page, order);
1547 #ifdef CONFIG_NEED_MULTIPLE_NODES
1549 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1551 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
1554 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1556 int __meminit __early_pfn_to_nid(unsigned long pfn,
1557 struct mminit_pfnnid_cache *state)
1559 unsigned long start_pfn, end_pfn;
1562 if (state->last_start <= pfn && pfn < state->last_end)
1563 return state->last_nid;
1565 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1566 if (nid != NUMA_NO_NODE) {
1567 state->last_start = start_pfn;
1568 state->last_end = end_pfn;
1569 state->last_nid = nid;
1574 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
1576 int __meminit early_pfn_to_nid(unsigned long pfn)
1578 static DEFINE_SPINLOCK(early_pfn_lock);
1581 spin_lock(&early_pfn_lock);
1582 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1584 nid = first_online_node;
1585 spin_unlock(&early_pfn_lock);
1589 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1591 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1594 if (early_page_uninitialised(pfn))
1596 __free_pages_core(page, order);
1600 * Check that the whole (or subset of) a pageblock given by the interval of
1601 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1602 * with the migration of free compaction scanner. The scanners then need to
1603 * use only pfn_valid_within() check for arches that allow holes within
1606 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1608 * It's possible on some configurations to have a setup like node0 node1 node0
1609 * i.e. it's possible that all pages within a zones range of pages do not
1610 * belong to a single zone. We assume that a border between node0 and node1
1611 * can occur within a single pageblock, but not a node0 node1 node0
1612 * interleaving within a single pageblock. It is therefore sufficient to check
1613 * the first and last page of a pageblock and avoid checking each individual
1614 * page in a pageblock.
1616 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1617 unsigned long end_pfn, struct zone *zone)
1619 struct page *start_page;
1620 struct page *end_page;
1622 /* end_pfn is one past the range we are checking */
1625 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1628 start_page = pfn_to_online_page(start_pfn);
1632 if (page_zone(start_page) != zone)
1635 end_page = pfn_to_page(end_pfn);
1637 /* This gives a shorter code than deriving page_zone(end_page) */
1638 if (page_zone_id(start_page) != page_zone_id(end_page))
1644 void set_zone_contiguous(struct zone *zone)
1646 unsigned long block_start_pfn = zone->zone_start_pfn;
1647 unsigned long block_end_pfn;
1649 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1650 for (; block_start_pfn < zone_end_pfn(zone);
1651 block_start_pfn = block_end_pfn,
1652 block_end_pfn += pageblock_nr_pages) {
1654 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1656 if (!__pageblock_pfn_to_page(block_start_pfn,
1657 block_end_pfn, zone))
1662 /* We confirm that there is no hole */
1663 zone->contiguous = true;
1666 void clear_zone_contiguous(struct zone *zone)
1668 zone->contiguous = false;
1671 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1672 static void __init deferred_free_range(unsigned long pfn,
1673 unsigned long nr_pages)
1681 page = pfn_to_page(pfn);
1683 /* Free a large naturally-aligned chunk if possible */
1684 if (nr_pages == pageblock_nr_pages &&
1685 (pfn & (pageblock_nr_pages - 1)) == 0) {
1686 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1687 __free_pages_core(page, pageblock_order);
1691 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1692 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1693 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1694 __free_pages_core(page, 0);
1698 /* Completion tracking for deferred_init_memmap() threads */
1699 static atomic_t pgdat_init_n_undone __initdata;
1700 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1702 static inline void __init pgdat_init_report_one_done(void)
1704 if (atomic_dec_and_test(&pgdat_init_n_undone))
1705 complete(&pgdat_init_all_done_comp);
1709 * Returns true if page needs to be initialized or freed to buddy allocator.
1711 * First we check if pfn is valid on architectures where it is possible to have
1712 * holes within pageblock_nr_pages. On systems where it is not possible, this
1713 * function is optimized out.
1715 * Then, we check if a current large page is valid by only checking the validity
1718 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1720 if (!pfn_valid_within(pfn))
1722 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1728 * Free pages to buddy allocator. Try to free aligned pages in
1729 * pageblock_nr_pages sizes.
1731 static void __init deferred_free_pages(unsigned long pfn,
1732 unsigned long end_pfn)
1734 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1735 unsigned long nr_free = 0;
1737 for (; pfn < end_pfn; pfn++) {
1738 if (!deferred_pfn_valid(pfn)) {
1739 deferred_free_range(pfn - nr_free, nr_free);
1741 } else if (!(pfn & nr_pgmask)) {
1742 deferred_free_range(pfn - nr_free, nr_free);
1748 /* Free the last block of pages to allocator */
1749 deferred_free_range(pfn - nr_free, nr_free);
1753 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1754 * by performing it only once every pageblock_nr_pages.
1755 * Return number of pages initialized.
1757 static unsigned long __init deferred_init_pages(struct zone *zone,
1759 unsigned long end_pfn)
1761 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1762 int nid = zone_to_nid(zone);
1763 unsigned long nr_pages = 0;
1764 int zid = zone_idx(zone);
1765 struct page *page = NULL;
1767 for (; pfn < end_pfn; pfn++) {
1768 if (!deferred_pfn_valid(pfn)) {
1771 } else if (!page || !(pfn & nr_pgmask)) {
1772 page = pfn_to_page(pfn);
1776 __init_single_page(page, pfn, zid, nid);
1783 * This function is meant to pre-load the iterator for the zone init.
1784 * Specifically it walks through the ranges until we are caught up to the
1785 * first_init_pfn value and exits there. If we never encounter the value we
1786 * return false indicating there are no valid ranges left.
1789 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1790 unsigned long *spfn, unsigned long *epfn,
1791 unsigned long first_init_pfn)
1796 * Start out by walking through the ranges in this zone that have
1797 * already been initialized. We don't need to do anything with them
1798 * so we just need to flush them out of the system.
1800 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1801 if (*epfn <= first_init_pfn)
1803 if (*spfn < first_init_pfn)
1804 *spfn = first_init_pfn;
1813 * Initialize and free pages. We do it in two loops: first we initialize
1814 * struct page, then free to buddy allocator, because while we are
1815 * freeing pages we can access pages that are ahead (computing buddy
1816 * page in __free_one_page()).
1818 * In order to try and keep some memory in the cache we have the loop
1819 * broken along max page order boundaries. This way we will not cause
1820 * any issues with the buddy page computation.
1822 static unsigned long __init
1823 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1824 unsigned long *end_pfn)
1826 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1827 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1828 unsigned long nr_pages = 0;
1831 /* First we loop through and initialize the page values */
1832 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1835 if (mo_pfn <= *start_pfn)
1838 t = min(mo_pfn, *end_pfn);
1839 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1841 if (mo_pfn < *end_pfn) {
1842 *start_pfn = mo_pfn;
1847 /* Reset values and now loop through freeing pages as needed */
1850 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1856 t = min(mo_pfn, epfn);
1857 deferred_free_pages(spfn, t);
1867 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1870 unsigned long spfn, epfn;
1871 struct zone *zone = arg;
1874 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1877 * Initialize and free pages in MAX_ORDER sized increments so that we
1878 * can avoid introducing any issues with the buddy allocator.
1880 while (spfn < end_pfn) {
1881 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1886 /* An arch may override for more concurrency. */
1888 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1893 /* Initialise remaining memory on a node */
1894 static int __init deferred_init_memmap(void *data)
1896 pg_data_t *pgdat = data;
1897 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1898 unsigned long spfn = 0, epfn = 0;
1899 unsigned long first_init_pfn, flags;
1900 unsigned long start = jiffies;
1902 int zid, max_threads;
1905 /* Bind memory initialisation thread to a local node if possible */
1906 if (!cpumask_empty(cpumask))
1907 set_cpus_allowed_ptr(current, cpumask);
1909 pgdat_resize_lock(pgdat, &flags);
1910 first_init_pfn = pgdat->first_deferred_pfn;
1911 if (first_init_pfn == ULONG_MAX) {
1912 pgdat_resize_unlock(pgdat, &flags);
1913 pgdat_init_report_one_done();
1917 /* Sanity check boundaries */
1918 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1919 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1920 pgdat->first_deferred_pfn = ULONG_MAX;
1923 * Once we unlock here, the zone cannot be grown anymore, thus if an
1924 * interrupt thread must allocate this early in boot, zone must be
1925 * pre-grown prior to start of deferred page initialization.
1927 pgdat_resize_unlock(pgdat, &flags);
1929 /* Only the highest zone is deferred so find it */
1930 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1931 zone = pgdat->node_zones + zid;
1932 if (first_init_pfn < zone_end_pfn(zone))
1936 /* If the zone is empty somebody else may have cleared out the zone */
1937 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1941 max_threads = deferred_page_init_max_threads(cpumask);
1943 while (spfn < epfn) {
1944 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
1945 struct padata_mt_job job = {
1946 .thread_fn = deferred_init_memmap_chunk,
1949 .size = epfn_align - spfn,
1950 .align = PAGES_PER_SECTION,
1951 .min_chunk = PAGES_PER_SECTION,
1952 .max_threads = max_threads,
1955 padata_do_multithreaded(&job);
1956 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1960 /* Sanity check that the next zone really is unpopulated */
1961 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1963 pr_info("node %d deferred pages initialised in %ums\n",
1964 pgdat->node_id, jiffies_to_msecs(jiffies - start));
1966 pgdat_init_report_one_done();
1971 * If this zone has deferred pages, try to grow it by initializing enough
1972 * deferred pages to satisfy the allocation specified by order, rounded up to
1973 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1974 * of SECTION_SIZE bytes by initializing struct pages in increments of
1975 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1977 * Return true when zone was grown, otherwise return false. We return true even
1978 * when we grow less than requested, to let the caller decide if there are
1979 * enough pages to satisfy the allocation.
1981 * Note: We use noinline because this function is needed only during boot, and
1982 * it is called from a __ref function _deferred_grow_zone. This way we are
1983 * making sure that it is not inlined into permanent text section.
1985 static noinline bool __init
1986 deferred_grow_zone(struct zone *zone, unsigned int order)
1988 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1989 pg_data_t *pgdat = zone->zone_pgdat;
1990 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1991 unsigned long spfn, epfn, flags;
1992 unsigned long nr_pages = 0;
1995 /* Only the last zone may have deferred pages */
1996 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1999 pgdat_resize_lock(pgdat, &flags);
2002 * If someone grew this zone while we were waiting for spinlock, return
2003 * true, as there might be enough pages already.
2005 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2006 pgdat_resize_unlock(pgdat, &flags);
2010 /* If the zone is empty somebody else may have cleared out the zone */
2011 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2012 first_deferred_pfn)) {
2013 pgdat->first_deferred_pfn = ULONG_MAX;
2014 pgdat_resize_unlock(pgdat, &flags);
2015 /* Retry only once. */
2016 return first_deferred_pfn != ULONG_MAX;
2020 * Initialize and free pages in MAX_ORDER sized increments so
2021 * that we can avoid introducing any issues with the buddy
2024 while (spfn < epfn) {
2025 /* update our first deferred PFN for this section */
2026 first_deferred_pfn = spfn;
2028 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2029 touch_nmi_watchdog();
2031 /* We should only stop along section boundaries */
2032 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2035 /* If our quota has been met we can stop here */
2036 if (nr_pages >= nr_pages_needed)
2040 pgdat->first_deferred_pfn = spfn;
2041 pgdat_resize_unlock(pgdat, &flags);
2043 return nr_pages > 0;
2047 * deferred_grow_zone() is __init, but it is called from
2048 * get_page_from_freelist() during early boot until deferred_pages permanently
2049 * disables this call. This is why we have refdata wrapper to avoid warning,
2050 * and to ensure that the function body gets unloaded.
2053 _deferred_grow_zone(struct zone *zone, unsigned int order)
2055 return deferred_grow_zone(zone, order);
2058 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2060 void __init page_alloc_init_late(void)
2065 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2067 /* There will be num_node_state(N_MEMORY) threads */
2068 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2069 for_each_node_state(nid, N_MEMORY) {
2070 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2073 /* Block until all are initialised */
2074 wait_for_completion(&pgdat_init_all_done_comp);
2077 * The number of managed pages has changed due to the initialisation
2078 * so the pcpu batch and high limits needs to be updated or the limits
2079 * will be artificially small.
2081 for_each_populated_zone(zone)
2082 zone_pcp_update(zone);
2085 * We initialized the rest of the deferred pages. Permanently disable
2086 * on-demand struct page initialization.
2088 static_branch_disable(&deferred_pages);
2090 /* Reinit limits that are based on free pages after the kernel is up */
2091 files_maxfiles_init();
2094 /* Discard memblock private memory */
2097 for_each_node_state(nid, N_MEMORY)
2098 shuffle_free_memory(NODE_DATA(nid));
2100 for_each_populated_zone(zone)
2101 set_zone_contiguous(zone);
2105 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2106 void __init init_cma_reserved_pageblock(struct page *page)
2108 unsigned i = pageblock_nr_pages;
2109 struct page *p = page;
2112 __ClearPageReserved(p);
2113 set_page_count(p, 0);
2116 set_pageblock_migratetype(page, MIGRATE_CMA);
2118 if (pageblock_order >= MAX_ORDER) {
2119 i = pageblock_nr_pages;
2122 set_page_refcounted(p);
2123 __free_pages(p, MAX_ORDER - 1);
2124 p += MAX_ORDER_NR_PAGES;
2125 } while (i -= MAX_ORDER_NR_PAGES);
2127 set_page_refcounted(page);
2128 __free_pages(page, pageblock_order);
2131 adjust_managed_page_count(page, pageblock_nr_pages);
2136 * The order of subdivision here is critical for the IO subsystem.
2137 * Please do not alter this order without good reasons and regression
2138 * testing. Specifically, as large blocks of memory are subdivided,
2139 * the order in which smaller blocks are delivered depends on the order
2140 * they're subdivided in this function. This is the primary factor
2141 * influencing the order in which pages are delivered to the IO
2142 * subsystem according to empirical testing, and this is also justified
2143 * by considering the behavior of a buddy system containing a single
2144 * large block of memory acted on by a series of small allocations.
2145 * This behavior is a critical factor in sglist merging's success.
2149 static inline void expand(struct zone *zone, struct page *page,
2150 int low, int high, int migratetype)
2152 unsigned long size = 1 << high;
2154 while (high > low) {
2157 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2160 * Mark as guard pages (or page), that will allow to
2161 * merge back to allocator when buddy will be freed.
2162 * Corresponding page table entries will not be touched,
2163 * pages will stay not present in virtual address space
2165 if (set_page_guard(zone, &page[size], high, migratetype))
2168 add_to_free_list(&page[size], zone, high, migratetype);
2169 set_page_order(&page[size], high);
2173 static void check_new_page_bad(struct page *page)
2175 if (unlikely(page->flags & __PG_HWPOISON)) {
2176 /* Don't complain about hwpoisoned pages */
2177 page_mapcount_reset(page); /* remove PageBuddy */
2182 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2186 * This page is about to be returned from the page allocator
2188 static inline int check_new_page(struct page *page)
2190 if (likely(page_expected_state(page,
2191 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2194 check_new_page_bad(page);
2198 static inline bool free_pages_prezeroed(void)
2200 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2201 page_poisoning_enabled()) || want_init_on_free();
2204 #ifdef CONFIG_DEBUG_VM
2206 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2207 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2208 * also checked when pcp lists are refilled from the free lists.
2210 static inline bool check_pcp_refill(struct page *page)
2212 if (debug_pagealloc_enabled_static())
2213 return check_new_page(page);
2218 static inline bool check_new_pcp(struct page *page)
2220 return check_new_page(page);
2224 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2225 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2226 * enabled, they are also checked when being allocated from the pcp lists.
2228 static inline bool check_pcp_refill(struct page *page)
2230 return check_new_page(page);
2232 static inline bool check_new_pcp(struct page *page)
2234 if (debug_pagealloc_enabled_static())
2235 return check_new_page(page);
2239 #endif /* CONFIG_DEBUG_VM */
2241 static bool check_new_pages(struct page *page, unsigned int order)
2244 for (i = 0; i < (1 << order); i++) {
2245 struct page *p = page + i;
2247 if (unlikely(check_new_page(p)))
2254 inline void post_alloc_hook(struct page *page, unsigned int order,
2257 set_page_private(page, 0);
2258 set_page_refcounted(page);
2260 arch_alloc_page(page, order);
2261 if (debug_pagealloc_enabled_static())
2262 kernel_map_pages(page, 1 << order, 1);
2263 kasan_alloc_pages(page, order);
2264 kernel_poison_pages(page, 1 << order, 1);
2265 set_page_owner(page, order, gfp_flags);
2268 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2269 unsigned int alloc_flags)
2271 post_alloc_hook(page, order, gfp_flags);
2273 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2274 kernel_init_free_pages(page, 1 << order);
2276 if (order && (gfp_flags & __GFP_COMP))
2277 prep_compound_page(page, order);
2280 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2281 * allocate the page. The expectation is that the caller is taking
2282 * steps that will free more memory. The caller should avoid the page
2283 * being used for !PFMEMALLOC purposes.
2285 if (alloc_flags & ALLOC_NO_WATERMARKS)
2286 set_page_pfmemalloc(page);
2288 clear_page_pfmemalloc(page);
2292 * Go through the free lists for the given migratetype and remove
2293 * the smallest available page from the freelists
2295 static __always_inline
2296 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2299 unsigned int current_order;
2300 struct free_area *area;
2303 /* Find a page of the appropriate size in the preferred list */
2304 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2305 area = &(zone->free_area[current_order]);
2306 page = get_page_from_free_area(area, migratetype);
2309 del_page_from_free_list(page, zone, current_order);
2310 expand(zone, page, order, current_order, migratetype);
2311 set_pcppage_migratetype(page, migratetype);
2320 * This array describes the order lists are fallen back to when
2321 * the free lists for the desirable migrate type are depleted
2323 static int fallbacks[MIGRATE_TYPES][3] = {
2324 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2325 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2326 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2328 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2330 #ifdef CONFIG_MEMORY_ISOLATION
2331 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2336 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2339 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2342 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2343 unsigned int order) { return NULL; }
2347 * Move the free pages in a range to the freelist tail of the requested type.
2348 * Note that start_page and end_pages are not aligned on a pageblock
2349 * boundary. If alignment is required, use move_freepages_block()
2351 static int move_freepages(struct zone *zone,
2352 struct page *start_page, struct page *end_page,
2353 int migratetype, int *num_movable)
2357 int pages_moved = 0;
2359 for (page = start_page; page <= end_page;) {
2360 if (!pfn_valid_within(page_to_pfn(page))) {
2365 if (!PageBuddy(page)) {
2367 * We assume that pages that could be isolated for
2368 * migration are movable. But we don't actually try
2369 * isolating, as that would be expensive.
2372 (PageLRU(page) || __PageMovable(page)))
2379 /* Make sure we are not inadvertently changing nodes */
2380 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2381 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2383 order = page_order(page);
2384 move_to_free_list(page, zone, order, migratetype);
2386 pages_moved += 1 << order;
2392 int move_freepages_block(struct zone *zone, struct page *page,
2393 int migratetype, int *num_movable)
2395 unsigned long start_pfn, end_pfn;
2396 struct page *start_page, *end_page;
2401 start_pfn = page_to_pfn(page);
2402 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2403 start_page = pfn_to_page(start_pfn);
2404 end_page = start_page + pageblock_nr_pages - 1;
2405 end_pfn = start_pfn + pageblock_nr_pages - 1;
2407 /* Do not cross zone boundaries */
2408 if (!zone_spans_pfn(zone, start_pfn))
2410 if (!zone_spans_pfn(zone, end_pfn))
2413 return move_freepages(zone, start_page, end_page, migratetype,
2417 static void change_pageblock_range(struct page *pageblock_page,
2418 int start_order, int migratetype)
2420 int nr_pageblocks = 1 << (start_order - pageblock_order);
2422 while (nr_pageblocks--) {
2423 set_pageblock_migratetype(pageblock_page, migratetype);
2424 pageblock_page += pageblock_nr_pages;
2429 * When we are falling back to another migratetype during allocation, try to
2430 * steal extra free pages from the same pageblocks to satisfy further
2431 * allocations, instead of polluting multiple pageblocks.
2433 * If we are stealing a relatively large buddy page, it is likely there will
2434 * be more free pages in the pageblock, so try to steal them all. For
2435 * reclaimable and unmovable allocations, we steal regardless of page size,
2436 * as fragmentation caused by those allocations polluting movable pageblocks
2437 * is worse than movable allocations stealing from unmovable and reclaimable
2440 static bool can_steal_fallback(unsigned int order, int start_mt)
2443 * Leaving this order check is intended, although there is
2444 * relaxed order check in next check. The reason is that
2445 * we can actually steal whole pageblock if this condition met,
2446 * but, below check doesn't guarantee it and that is just heuristic
2447 * so could be changed anytime.
2449 if (order >= pageblock_order)
2452 if (order >= pageblock_order / 2 ||
2453 start_mt == MIGRATE_RECLAIMABLE ||
2454 start_mt == MIGRATE_UNMOVABLE ||
2455 page_group_by_mobility_disabled)
2461 static inline void boost_watermark(struct zone *zone)
2463 unsigned long max_boost;
2465 if (!watermark_boost_factor)
2468 * Don't bother in zones that are unlikely to produce results.
2469 * On small machines, including kdump capture kernels running
2470 * in a small area, boosting the watermark can cause an out of
2471 * memory situation immediately.
2473 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2476 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2477 watermark_boost_factor, 10000);
2480 * high watermark may be uninitialised if fragmentation occurs
2481 * very early in boot so do not boost. We do not fall
2482 * through and boost by pageblock_nr_pages as failing
2483 * allocations that early means that reclaim is not going
2484 * to help and it may even be impossible to reclaim the
2485 * boosted watermark resulting in a hang.
2490 max_boost = max(pageblock_nr_pages, max_boost);
2492 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2497 * This function implements actual steal behaviour. If order is large enough,
2498 * we can steal whole pageblock. If not, we first move freepages in this
2499 * pageblock to our migratetype and determine how many already-allocated pages
2500 * are there in the pageblock with a compatible migratetype. If at least half
2501 * of pages are free or compatible, we can change migratetype of the pageblock
2502 * itself, so pages freed in the future will be put on the correct free list.
2504 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2505 unsigned int alloc_flags, int start_type, bool whole_block)
2507 unsigned int current_order = page_order(page);
2508 int free_pages, movable_pages, alike_pages;
2511 old_block_type = get_pageblock_migratetype(page);
2514 * This can happen due to races and we want to prevent broken
2515 * highatomic accounting.
2517 if (is_migrate_highatomic(old_block_type))
2520 /* Take ownership for orders >= pageblock_order */
2521 if (current_order >= pageblock_order) {
2522 change_pageblock_range(page, current_order, start_type);
2527 * Boost watermarks to increase reclaim pressure to reduce the
2528 * likelihood of future fallbacks. Wake kswapd now as the node
2529 * may be balanced overall and kswapd will not wake naturally.
2531 boost_watermark(zone);
2532 if (alloc_flags & ALLOC_KSWAPD)
2533 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2535 /* We are not allowed to try stealing from the whole block */
2539 free_pages = move_freepages_block(zone, page, start_type,
2542 * Determine how many pages are compatible with our allocation.
2543 * For movable allocation, it's the number of movable pages which
2544 * we just obtained. For other types it's a bit more tricky.
2546 if (start_type == MIGRATE_MOVABLE) {
2547 alike_pages = movable_pages;
2550 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2551 * to MOVABLE pageblock, consider all non-movable pages as
2552 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2553 * vice versa, be conservative since we can't distinguish the
2554 * exact migratetype of non-movable pages.
2556 if (old_block_type == MIGRATE_MOVABLE)
2557 alike_pages = pageblock_nr_pages
2558 - (free_pages + movable_pages);
2563 /* moving whole block can fail due to zone boundary conditions */
2568 * If a sufficient number of pages in the block are either free or of
2569 * comparable migratability as our allocation, claim the whole block.
2571 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2572 page_group_by_mobility_disabled)
2573 set_pageblock_migratetype(page, start_type);
2578 move_to_free_list(page, zone, current_order, start_type);
2582 * Check whether there is a suitable fallback freepage with requested order.
2583 * If only_stealable is true, this function returns fallback_mt only if
2584 * we can steal other freepages all together. This would help to reduce
2585 * fragmentation due to mixed migratetype pages in one pageblock.
2587 int find_suitable_fallback(struct free_area *area, unsigned int order,
2588 int migratetype, bool only_stealable, bool *can_steal)
2593 if (area->nr_free == 0)
2598 fallback_mt = fallbacks[migratetype][i];
2599 if (fallback_mt == MIGRATE_TYPES)
2602 if (free_area_empty(area, fallback_mt))
2605 if (can_steal_fallback(order, migratetype))
2608 if (!only_stealable)
2619 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2620 * there are no empty page blocks that contain a page with a suitable order
2622 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2623 unsigned int alloc_order)
2626 unsigned long max_managed, flags;
2629 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2630 * Check is race-prone but harmless.
2632 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2633 if (zone->nr_reserved_highatomic >= max_managed)
2636 spin_lock_irqsave(&zone->lock, flags);
2638 /* Recheck the nr_reserved_highatomic limit under the lock */
2639 if (zone->nr_reserved_highatomic >= max_managed)
2643 mt = get_pageblock_migratetype(page);
2644 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2645 && !is_migrate_cma(mt)) {
2646 zone->nr_reserved_highatomic += pageblock_nr_pages;
2647 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2648 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2652 spin_unlock_irqrestore(&zone->lock, flags);
2656 * Used when an allocation is about to fail under memory pressure. This
2657 * potentially hurts the reliability of high-order allocations when under
2658 * intense memory pressure but failed atomic allocations should be easier
2659 * to recover from than an OOM.
2661 * If @force is true, try to unreserve a pageblock even though highatomic
2662 * pageblock is exhausted.
2664 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2667 struct zonelist *zonelist = ac->zonelist;
2668 unsigned long flags;
2675 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2678 * Preserve at least one pageblock unless memory pressure
2681 if (!force && zone->nr_reserved_highatomic <=
2685 spin_lock_irqsave(&zone->lock, flags);
2686 for (order = 0; order < MAX_ORDER; order++) {
2687 struct free_area *area = &(zone->free_area[order]);
2689 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2694 * In page freeing path, migratetype change is racy so
2695 * we can counter several free pages in a pageblock
2696 * in this loop althoug we changed the pageblock type
2697 * from highatomic to ac->migratetype. So we should
2698 * adjust the count once.
2700 if (is_migrate_highatomic_page(page)) {
2702 * It should never happen but changes to
2703 * locking could inadvertently allow a per-cpu
2704 * drain to add pages to MIGRATE_HIGHATOMIC
2705 * while unreserving so be safe and watch for
2708 zone->nr_reserved_highatomic -= min(
2710 zone->nr_reserved_highatomic);
2714 * Convert to ac->migratetype and avoid the normal
2715 * pageblock stealing heuristics. Minimally, the caller
2716 * is doing the work and needs the pages. More
2717 * importantly, if the block was always converted to
2718 * MIGRATE_UNMOVABLE or another type then the number
2719 * of pageblocks that cannot be completely freed
2722 set_pageblock_migratetype(page, ac->migratetype);
2723 ret = move_freepages_block(zone, page, ac->migratetype,
2726 spin_unlock_irqrestore(&zone->lock, flags);
2730 spin_unlock_irqrestore(&zone->lock, flags);
2737 * Try finding a free buddy page on the fallback list and put it on the free
2738 * list of requested migratetype, possibly along with other pages from the same
2739 * block, depending on fragmentation avoidance heuristics. Returns true if
2740 * fallback was found so that __rmqueue_smallest() can grab it.
2742 * The use of signed ints for order and current_order is a deliberate
2743 * deviation from the rest of this file, to make the for loop
2744 * condition simpler.
2746 static __always_inline bool
2747 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2748 unsigned int alloc_flags)
2750 struct free_area *area;
2752 int min_order = order;
2758 * Do not steal pages from freelists belonging to other pageblocks
2759 * i.e. orders < pageblock_order. If there are no local zones free,
2760 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2762 if (alloc_flags & ALLOC_NOFRAGMENT)
2763 min_order = pageblock_order;
2766 * Find the largest available free page in the other list. This roughly
2767 * approximates finding the pageblock with the most free pages, which
2768 * would be too costly to do exactly.
2770 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2772 area = &(zone->free_area[current_order]);
2773 fallback_mt = find_suitable_fallback(area, current_order,
2774 start_migratetype, false, &can_steal);
2775 if (fallback_mt == -1)
2779 * We cannot steal all free pages from the pageblock and the
2780 * requested migratetype is movable. In that case it's better to
2781 * steal and split the smallest available page instead of the
2782 * largest available page, because even if the next movable
2783 * allocation falls back into a different pageblock than this
2784 * one, it won't cause permanent fragmentation.
2786 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2787 && current_order > order)
2796 for (current_order = order; current_order < MAX_ORDER;
2798 area = &(zone->free_area[current_order]);
2799 fallback_mt = find_suitable_fallback(area, current_order,
2800 start_migratetype, false, &can_steal);
2801 if (fallback_mt != -1)
2806 * This should not happen - we already found a suitable fallback
2807 * when looking for the largest page.
2809 VM_BUG_ON(current_order == MAX_ORDER);
2812 page = get_page_from_free_area(area, fallback_mt);
2814 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2817 trace_mm_page_alloc_extfrag(page, order, current_order,
2818 start_migratetype, fallback_mt);
2825 * Do the hard work of removing an element from the buddy allocator.
2826 * Call me with the zone->lock already held.
2828 static __always_inline struct page *
2829 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2830 unsigned int alloc_flags)
2836 * Balance movable allocations between regular and CMA areas by
2837 * allocating from CMA when over half of the zone's free memory
2838 * is in the CMA area.
2840 if (alloc_flags & ALLOC_CMA &&
2841 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2842 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2843 page = __rmqueue_cma_fallback(zone, order);
2849 page = __rmqueue_smallest(zone, order, migratetype);
2850 if (unlikely(!page)) {
2851 if (alloc_flags & ALLOC_CMA)
2852 page = __rmqueue_cma_fallback(zone, order);
2854 if (!page && __rmqueue_fallback(zone, order, migratetype,
2859 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2864 * Obtain a specified number of elements from the buddy allocator, all under
2865 * a single hold of the lock, for efficiency. Add them to the supplied list.
2866 * Returns the number of new pages which were placed at *list.
2868 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2869 unsigned long count, struct list_head *list,
2870 int migratetype, unsigned int alloc_flags)
2874 spin_lock(&zone->lock);
2875 for (i = 0; i < count; ++i) {
2876 struct page *page = __rmqueue(zone, order, migratetype,
2878 if (unlikely(page == NULL))
2881 if (unlikely(check_pcp_refill(page)))
2885 * Split buddy pages returned by expand() are received here in
2886 * physical page order. The page is added to the tail of
2887 * caller's list. From the callers perspective, the linked list
2888 * is ordered by page number under some conditions. This is
2889 * useful for IO devices that can forward direction from the
2890 * head, thus also in the physical page order. This is useful
2891 * for IO devices that can merge IO requests if the physical
2892 * pages are ordered properly.
2894 list_add_tail(&page->lru, list);
2896 if (is_migrate_cma(get_pcppage_migratetype(page)))
2897 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2902 * i pages were removed from the buddy list even if some leak due
2903 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2904 * on i. Do not confuse with 'alloced' which is the number of
2905 * pages added to the pcp list.
2907 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2908 spin_unlock(&zone->lock);
2914 * Called from the vmstat counter updater to drain pagesets of this
2915 * currently executing processor on remote nodes after they have
2918 * Note that this function must be called with the thread pinned to
2919 * a single processor.
2921 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2923 unsigned long flags;
2924 int to_drain, batch;
2926 local_irq_save(flags);
2927 batch = READ_ONCE(pcp->batch);
2928 to_drain = min(pcp->count, batch);
2930 free_pcppages_bulk(zone, to_drain, pcp);
2931 local_irq_restore(flags);
2936 * Drain pcplists of the indicated processor and zone.
2938 * The processor must either be the current processor and the
2939 * thread pinned to the current processor or a processor that
2942 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2944 unsigned long flags;
2945 struct per_cpu_pageset *pset;
2946 struct per_cpu_pages *pcp;
2948 local_irq_save(flags);
2949 pset = per_cpu_ptr(zone->pageset, cpu);
2953 free_pcppages_bulk(zone, pcp->count, pcp);
2954 local_irq_restore(flags);
2958 * Drain pcplists of all zones on the indicated processor.
2960 * The processor must either be the current processor and the
2961 * thread pinned to the current processor or a processor that
2964 static void drain_pages(unsigned int cpu)
2968 for_each_populated_zone(zone) {
2969 drain_pages_zone(cpu, zone);
2974 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2976 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2977 * the single zone's pages.
2979 void drain_local_pages(struct zone *zone)
2981 int cpu = smp_processor_id();
2984 drain_pages_zone(cpu, zone);
2989 static void drain_local_pages_wq(struct work_struct *work)
2991 struct pcpu_drain *drain;
2993 drain = container_of(work, struct pcpu_drain, work);
2996 * drain_all_pages doesn't use proper cpu hotplug protection so
2997 * we can race with cpu offline when the WQ can move this from
2998 * a cpu pinned worker to an unbound one. We can operate on a different
2999 * cpu which is allright but we also have to make sure to not move to
3003 drain_local_pages(drain->zone);
3008 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3010 * When zone parameter is non-NULL, spill just the single zone's pages.
3012 * Note that this can be extremely slow as the draining happens in a workqueue.
3014 void drain_all_pages(struct zone *zone)
3019 * Allocate in the BSS so we wont require allocation in
3020 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3022 static cpumask_t cpus_with_pcps;
3025 * Make sure nobody triggers this path before mm_percpu_wq is fully
3028 if (WARN_ON_ONCE(!mm_percpu_wq))
3032 * Do not drain if one is already in progress unless it's specific to
3033 * a zone. Such callers are primarily CMA and memory hotplug and need
3034 * the drain to be complete when the call returns.
3036 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3039 mutex_lock(&pcpu_drain_mutex);
3043 * We don't care about racing with CPU hotplug event
3044 * as offline notification will cause the notified
3045 * cpu to drain that CPU pcps and on_each_cpu_mask
3046 * disables preemption as part of its processing
3048 for_each_online_cpu(cpu) {
3049 struct per_cpu_pageset *pcp;
3051 bool has_pcps = false;
3054 pcp = per_cpu_ptr(zone->pageset, cpu);
3058 for_each_populated_zone(z) {
3059 pcp = per_cpu_ptr(z->pageset, cpu);
3060 if (pcp->pcp.count) {
3068 cpumask_set_cpu(cpu, &cpus_with_pcps);
3070 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3073 for_each_cpu(cpu, &cpus_with_pcps) {
3074 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3077 INIT_WORK(&drain->work, drain_local_pages_wq);
3078 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3080 for_each_cpu(cpu, &cpus_with_pcps)
3081 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3083 mutex_unlock(&pcpu_drain_mutex);
3086 #ifdef CONFIG_HIBERNATION
3089 * Touch the watchdog for every WD_PAGE_COUNT pages.
3091 #define WD_PAGE_COUNT (128*1024)
3093 void mark_free_pages(struct zone *zone)
3095 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3096 unsigned long flags;
3097 unsigned int order, t;
3100 if (zone_is_empty(zone))
3103 spin_lock_irqsave(&zone->lock, flags);
3105 max_zone_pfn = zone_end_pfn(zone);
3106 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3107 if (pfn_valid(pfn)) {
3108 page = pfn_to_page(pfn);
3110 if (!--page_count) {
3111 touch_nmi_watchdog();
3112 page_count = WD_PAGE_COUNT;
3115 if (page_zone(page) != zone)
3118 if (!swsusp_page_is_forbidden(page))
3119 swsusp_unset_page_free(page);
3122 for_each_migratetype_order(order, t) {
3123 list_for_each_entry(page,
3124 &zone->free_area[order].free_list[t], lru) {
3127 pfn = page_to_pfn(page);
3128 for (i = 0; i < (1UL << order); i++) {
3129 if (!--page_count) {
3130 touch_nmi_watchdog();
3131 page_count = WD_PAGE_COUNT;
3133 swsusp_set_page_free(pfn_to_page(pfn + i));
3137 spin_unlock_irqrestore(&zone->lock, flags);
3139 #endif /* CONFIG_PM */
3141 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3145 if (!free_pcp_prepare(page))
3148 migratetype = get_pfnblock_migratetype(page, pfn);
3149 set_pcppage_migratetype(page, migratetype);
3153 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3155 struct zone *zone = page_zone(page);
3156 struct per_cpu_pages *pcp;
3159 migratetype = get_pcppage_migratetype(page);
3160 __count_vm_event(PGFREE);
3163 * We only track unmovable, reclaimable and movable on pcp lists.
3164 * Free ISOLATE pages back to the allocator because they are being
3165 * offlined but treat HIGHATOMIC as movable pages so we can get those
3166 * areas back if necessary. Otherwise, we may have to free
3167 * excessively into the page allocator
3169 if (migratetype >= MIGRATE_PCPTYPES) {
3170 if (unlikely(is_migrate_isolate(migratetype))) {
3171 free_one_page(zone, page, pfn, 0, migratetype);
3174 migratetype = MIGRATE_MOVABLE;
3177 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3178 list_add(&page->lru, &pcp->lists[migratetype]);
3180 if (pcp->count >= pcp->high) {
3181 unsigned long batch = READ_ONCE(pcp->batch);
3182 free_pcppages_bulk(zone, batch, pcp);
3187 * Free a 0-order page
3189 void free_unref_page(struct page *page)
3191 unsigned long flags;
3192 unsigned long pfn = page_to_pfn(page);
3194 if (!free_unref_page_prepare(page, pfn))
3197 local_irq_save(flags);
3198 free_unref_page_commit(page, pfn);
3199 local_irq_restore(flags);
3203 * Free a list of 0-order pages
3205 void free_unref_page_list(struct list_head *list)
3207 struct page *page, *next;
3208 unsigned long flags, pfn;
3209 int batch_count = 0;
3211 /* Prepare pages for freeing */
3212 list_for_each_entry_safe(page, next, list, lru) {
3213 pfn = page_to_pfn(page);
3214 if (!free_unref_page_prepare(page, pfn))
3215 list_del(&page->lru);
3216 set_page_private(page, pfn);
3219 local_irq_save(flags);
3220 list_for_each_entry_safe(page, next, list, lru) {
3221 unsigned long pfn = page_private(page);
3223 set_page_private(page, 0);
3224 trace_mm_page_free_batched(page);
3225 free_unref_page_commit(page, pfn);
3228 * Guard against excessive IRQ disabled times when we get
3229 * a large list of pages to free.
3231 if (++batch_count == SWAP_CLUSTER_MAX) {
3232 local_irq_restore(flags);
3234 local_irq_save(flags);
3237 local_irq_restore(flags);
3241 * split_page takes a non-compound higher-order page, and splits it into
3242 * n (1<<order) sub-pages: page[0..n]
3243 * Each sub-page must be freed individually.
3245 * Note: this is probably too low level an operation for use in drivers.
3246 * Please consult with lkml before using this in your driver.
3248 void split_page(struct page *page, unsigned int order)
3252 VM_BUG_ON_PAGE(PageCompound(page), page);
3253 VM_BUG_ON_PAGE(!page_count(page), page);
3255 for (i = 1; i < (1 << order); i++)
3256 set_page_refcounted(page + i);
3257 split_page_owner(page, 1 << order);
3259 EXPORT_SYMBOL_GPL(split_page);
3261 int __isolate_free_page(struct page *page, unsigned int order)
3263 unsigned long watermark;
3267 BUG_ON(!PageBuddy(page));
3269 zone = page_zone(page);
3270 mt = get_pageblock_migratetype(page);
3272 if (!is_migrate_isolate(mt)) {
3274 * Obey watermarks as if the page was being allocated. We can
3275 * emulate a high-order watermark check with a raised order-0
3276 * watermark, because we already know our high-order page
3279 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3280 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3283 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3286 /* Remove page from free list */
3288 del_page_from_free_list(page, zone, order);
3291 * Set the pageblock if the isolated page is at least half of a
3294 if (order >= pageblock_order - 1) {
3295 struct page *endpage = page + (1 << order) - 1;
3296 for (; page < endpage; page += pageblock_nr_pages) {
3297 int mt = get_pageblock_migratetype(page);
3298 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3299 && !is_migrate_highatomic(mt))
3300 set_pageblock_migratetype(page,
3306 return 1UL << order;
3310 * __putback_isolated_page - Return a now-isolated page back where we got it
3311 * @page: Page that was isolated
3312 * @order: Order of the isolated page
3313 * @mt: The page's pageblock's migratetype
3315 * This function is meant to return a page pulled from the free lists via
3316 * __isolate_free_page back to the free lists they were pulled from.
3318 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3320 struct zone *zone = page_zone(page);
3322 /* zone lock should be held when this function is called */
3323 lockdep_assert_held(&zone->lock);
3325 /* Return isolated page to tail of freelist. */
3326 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3327 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3331 * Update NUMA hit/miss statistics
3333 * Must be called with interrupts disabled.
3335 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3338 enum numa_stat_item local_stat = NUMA_LOCAL;
3340 /* skip numa counters update if numa stats is disabled */
3341 if (!static_branch_likely(&vm_numa_stat_key))
3344 if (zone_to_nid(z) != numa_node_id())
3345 local_stat = NUMA_OTHER;
3347 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3348 __inc_numa_state(z, NUMA_HIT);
3350 __inc_numa_state(z, NUMA_MISS);
3351 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3353 __inc_numa_state(z, local_stat);
3357 /* Remove page from the per-cpu list, caller must protect the list */
3358 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3359 unsigned int alloc_flags,
3360 struct per_cpu_pages *pcp,
3361 struct list_head *list)
3366 if (list_empty(list)) {
3367 pcp->count += rmqueue_bulk(zone, 0,
3369 migratetype, alloc_flags);
3370 if (unlikely(list_empty(list)))
3374 page = list_first_entry(list, struct page, lru);
3375 list_del(&page->lru);
3377 } while (check_new_pcp(page));
3382 /* Lock and remove page from the per-cpu list */
3383 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3384 struct zone *zone, gfp_t gfp_flags,
3385 int migratetype, unsigned int alloc_flags)
3387 struct per_cpu_pages *pcp;
3388 struct list_head *list;
3390 unsigned long flags;
3392 local_irq_save(flags);
3393 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3394 list = &pcp->lists[migratetype];
3395 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3397 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3398 zone_statistics(preferred_zone, zone);
3400 local_irq_restore(flags);
3405 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3408 struct page *rmqueue(struct zone *preferred_zone,
3409 struct zone *zone, unsigned int order,
3410 gfp_t gfp_flags, unsigned int alloc_flags,
3413 unsigned long flags;
3416 if (likely(order == 0)) {
3418 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3419 * we need to skip it when CMA area isn't allowed.
3421 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3422 migratetype != MIGRATE_MOVABLE) {
3423 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3424 migratetype, alloc_flags);
3430 * We most definitely don't want callers attempting to
3431 * allocate greater than order-1 page units with __GFP_NOFAIL.
3433 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3434 spin_lock_irqsave(&zone->lock, flags);
3439 * order-0 request can reach here when the pcplist is skipped
3440 * due to non-CMA allocation context. HIGHATOMIC area is
3441 * reserved for high-order atomic allocation, so order-0
3442 * request should skip it.
3444 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3445 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3447 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3450 page = __rmqueue(zone, order, migratetype, alloc_flags);
3451 } while (page && check_new_pages(page, order));
3452 spin_unlock(&zone->lock);
3455 __mod_zone_freepage_state(zone, -(1 << order),
3456 get_pcppage_migratetype(page));
3458 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3459 zone_statistics(preferred_zone, zone);
3460 local_irq_restore(flags);
3463 /* Separate test+clear to avoid unnecessary atomics */
3464 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3465 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3466 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3469 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3473 local_irq_restore(flags);
3477 #ifdef CONFIG_FAIL_PAGE_ALLOC
3480 struct fault_attr attr;
3482 bool ignore_gfp_highmem;
3483 bool ignore_gfp_reclaim;
3485 } fail_page_alloc = {
3486 .attr = FAULT_ATTR_INITIALIZER,
3487 .ignore_gfp_reclaim = true,
3488 .ignore_gfp_highmem = true,
3492 static int __init setup_fail_page_alloc(char *str)
3494 return setup_fault_attr(&fail_page_alloc.attr, str);
3496 __setup("fail_page_alloc=", setup_fail_page_alloc);
3498 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3500 if (order < fail_page_alloc.min_order)
3502 if (gfp_mask & __GFP_NOFAIL)
3504 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3506 if (fail_page_alloc.ignore_gfp_reclaim &&
3507 (gfp_mask & __GFP_DIRECT_RECLAIM))
3510 return should_fail(&fail_page_alloc.attr, 1 << order);
3513 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3515 static int __init fail_page_alloc_debugfs(void)
3517 umode_t mode = S_IFREG | 0600;
3520 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3521 &fail_page_alloc.attr);
3523 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3524 &fail_page_alloc.ignore_gfp_reclaim);
3525 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3526 &fail_page_alloc.ignore_gfp_highmem);
3527 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3532 late_initcall(fail_page_alloc_debugfs);
3534 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3536 #else /* CONFIG_FAIL_PAGE_ALLOC */
3538 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3543 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3545 static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3547 return __should_fail_alloc_page(gfp_mask, order);
3549 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3551 static inline long __zone_watermark_unusable_free(struct zone *z,
3552 unsigned int order, unsigned int alloc_flags)
3554 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3555 long unusable_free = (1 << order) - 1;
3558 * If the caller does not have rights to ALLOC_HARDER then subtract
3559 * the high-atomic reserves. This will over-estimate the size of the
3560 * atomic reserve but it avoids a search.
3562 if (likely(!alloc_harder))
3563 unusable_free += z->nr_reserved_highatomic;
3566 /* If allocation can't use CMA areas don't use free CMA pages */
3567 if (!(alloc_flags & ALLOC_CMA))
3568 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3571 return unusable_free;
3575 * Return true if free base pages are above 'mark'. For high-order checks it
3576 * will return true of the order-0 watermark is reached and there is at least
3577 * one free page of a suitable size. Checking now avoids taking the zone lock
3578 * to check in the allocation paths if no pages are free.
3580 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3581 int highest_zoneidx, unsigned int alloc_flags,
3586 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3588 /* free_pages may go negative - that's OK */
3589 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3591 if (alloc_flags & ALLOC_HIGH)
3594 if (unlikely(alloc_harder)) {
3596 * OOM victims can try even harder than normal ALLOC_HARDER
3597 * users on the grounds that it's definitely going to be in
3598 * the exit path shortly and free memory. Any allocation it
3599 * makes during the free path will be small and short-lived.
3601 if (alloc_flags & ALLOC_OOM)
3608 * Check watermarks for an order-0 allocation request. If these
3609 * are not met, then a high-order request also cannot go ahead
3610 * even if a suitable page happened to be free.
3612 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3615 /* If this is an order-0 request then the watermark is fine */
3619 /* For a high-order request, check at least one suitable page is free */
3620 for (o = order; o < MAX_ORDER; o++) {
3621 struct free_area *area = &z->free_area[o];
3627 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3628 if (!free_area_empty(area, mt))
3633 if ((alloc_flags & ALLOC_CMA) &&
3634 !free_area_empty(area, MIGRATE_CMA)) {
3638 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3644 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3645 int highest_zoneidx, unsigned int alloc_flags)
3647 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3648 zone_page_state(z, NR_FREE_PAGES));
3651 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3652 unsigned long mark, int highest_zoneidx,
3653 unsigned int alloc_flags, gfp_t gfp_mask)
3657 free_pages = zone_page_state(z, NR_FREE_PAGES);
3660 * Fast check for order-0 only. If this fails then the reserves
3661 * need to be calculated.
3666 fast_free = free_pages;
3667 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3668 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3672 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3676 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3677 * when checking the min watermark. The min watermark is the
3678 * point where boosting is ignored so that kswapd is woken up
3679 * when below the low watermark.
3681 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3682 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3683 mark = z->_watermark[WMARK_MIN];
3684 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3685 alloc_flags, free_pages);
3691 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3692 unsigned long mark, int highest_zoneidx)
3694 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3696 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3697 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3699 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3704 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3706 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3707 node_reclaim_distance;
3709 #else /* CONFIG_NUMA */
3710 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3714 #endif /* CONFIG_NUMA */
3717 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3718 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3719 * premature use of a lower zone may cause lowmem pressure problems that
3720 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3721 * probably too small. It only makes sense to spread allocations to avoid
3722 * fragmentation between the Normal and DMA32 zones.
3724 static inline unsigned int
3725 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3727 unsigned int alloc_flags;
3730 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3733 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3735 #ifdef CONFIG_ZONE_DMA32
3739 if (zone_idx(zone) != ZONE_NORMAL)
3743 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3744 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3745 * on UMA that if Normal is populated then so is DMA32.
3747 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3748 if (nr_online_nodes > 1 && !populated_zone(--zone))
3751 alloc_flags |= ALLOC_NOFRAGMENT;
3752 #endif /* CONFIG_ZONE_DMA32 */
3756 static inline unsigned int current_alloc_flags(gfp_t gfp_mask,
3757 unsigned int alloc_flags)
3760 unsigned int pflags = current->flags;
3762 if (!(pflags & PF_MEMALLOC_NOCMA) &&
3763 gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3764 alloc_flags |= ALLOC_CMA;
3771 * get_page_from_freelist goes through the zonelist trying to allocate
3774 static struct page *
3775 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3776 const struct alloc_context *ac)
3780 struct pglist_data *last_pgdat_dirty_limit = NULL;
3785 * Scan zonelist, looking for a zone with enough free.
3786 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3788 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3789 z = ac->preferred_zoneref;
3790 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3795 if (cpusets_enabled() &&
3796 (alloc_flags & ALLOC_CPUSET) &&
3797 !__cpuset_zone_allowed(zone, gfp_mask))
3800 * When allocating a page cache page for writing, we
3801 * want to get it from a node that is within its dirty
3802 * limit, such that no single node holds more than its
3803 * proportional share of globally allowed dirty pages.
3804 * The dirty limits take into account the node's
3805 * lowmem reserves and high watermark so that kswapd
3806 * should be able to balance it without having to
3807 * write pages from its LRU list.
3809 * XXX: For now, allow allocations to potentially
3810 * exceed the per-node dirty limit in the slowpath
3811 * (spread_dirty_pages unset) before going into reclaim,
3812 * which is important when on a NUMA setup the allowed
3813 * nodes are together not big enough to reach the
3814 * global limit. The proper fix for these situations
3815 * will require awareness of nodes in the
3816 * dirty-throttling and the flusher threads.
3818 if (ac->spread_dirty_pages) {
3819 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3822 if (!node_dirty_ok(zone->zone_pgdat)) {
3823 last_pgdat_dirty_limit = zone->zone_pgdat;
3828 if (no_fallback && nr_online_nodes > 1 &&
3829 zone != ac->preferred_zoneref->zone) {
3833 * If moving to a remote node, retry but allow
3834 * fragmenting fallbacks. Locality is more important
3835 * than fragmentation avoidance.
3837 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3838 if (zone_to_nid(zone) != local_nid) {
3839 alloc_flags &= ~ALLOC_NOFRAGMENT;
3844 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3845 if (!zone_watermark_fast(zone, order, mark,
3846 ac->highest_zoneidx, alloc_flags,
3850 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3852 * Watermark failed for this zone, but see if we can
3853 * grow this zone if it contains deferred pages.
3855 if (static_branch_unlikely(&deferred_pages)) {
3856 if (_deferred_grow_zone(zone, order))
3860 /* Checked here to keep the fast path fast */
3861 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3862 if (alloc_flags & ALLOC_NO_WATERMARKS)
3865 if (node_reclaim_mode == 0 ||
3866 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3869 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3871 case NODE_RECLAIM_NOSCAN:
3874 case NODE_RECLAIM_FULL:
3875 /* scanned but unreclaimable */
3878 /* did we reclaim enough */
3879 if (zone_watermark_ok(zone, order, mark,
3880 ac->highest_zoneidx, alloc_flags))
3888 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3889 gfp_mask, alloc_flags, ac->migratetype);
3891 prep_new_page(page, order, gfp_mask, alloc_flags);
3894 * If this is a high-order atomic allocation then check
3895 * if the pageblock should be reserved for the future
3897 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3898 reserve_highatomic_pageblock(page, zone, order);
3902 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3903 /* Try again if zone has deferred pages */
3904 if (static_branch_unlikely(&deferred_pages)) {
3905 if (_deferred_grow_zone(zone, order))
3913 * It's possible on a UMA machine to get through all zones that are
3914 * fragmented. If avoiding fragmentation, reset and try again.
3917 alloc_flags &= ~ALLOC_NOFRAGMENT;
3924 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3926 unsigned int filter = SHOW_MEM_FILTER_NODES;
3929 * This documents exceptions given to allocations in certain
3930 * contexts that are allowed to allocate outside current's set
3933 if (!(gfp_mask & __GFP_NOMEMALLOC))
3934 if (tsk_is_oom_victim(current) ||
3935 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3936 filter &= ~SHOW_MEM_FILTER_NODES;
3937 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3938 filter &= ~SHOW_MEM_FILTER_NODES;
3940 show_mem(filter, nodemask);
3943 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3945 struct va_format vaf;
3947 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3949 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3952 va_start(args, fmt);
3955 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3956 current->comm, &vaf, gfp_mask, &gfp_mask,
3957 nodemask_pr_args(nodemask));
3960 cpuset_print_current_mems_allowed();
3963 warn_alloc_show_mem(gfp_mask, nodemask);
3966 static inline struct page *
3967 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3968 unsigned int alloc_flags,
3969 const struct alloc_context *ac)
3973 page = get_page_from_freelist(gfp_mask, order,
3974 alloc_flags|ALLOC_CPUSET, ac);
3976 * fallback to ignore cpuset restriction if our nodes
3980 page = get_page_from_freelist(gfp_mask, order,
3986 static inline struct page *
3987 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3988 const struct alloc_context *ac, unsigned long *did_some_progress)
3990 struct oom_control oc = {
3991 .zonelist = ac->zonelist,
3992 .nodemask = ac->nodemask,
3994 .gfp_mask = gfp_mask,
3999 *did_some_progress = 0;
4002 * Acquire the oom lock. If that fails, somebody else is
4003 * making progress for us.
4005 if (!mutex_trylock(&oom_lock)) {
4006 *did_some_progress = 1;
4007 schedule_timeout_uninterruptible(1);
4012 * Go through the zonelist yet one more time, keep very high watermark
4013 * here, this is only to catch a parallel oom killing, we must fail if
4014 * we're still under heavy pressure. But make sure that this reclaim
4015 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4016 * allocation which will never fail due to oom_lock already held.
4018 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4019 ~__GFP_DIRECT_RECLAIM, order,
4020 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4024 /* Coredumps can quickly deplete all memory reserves */
4025 if (current->flags & PF_DUMPCORE)
4027 /* The OOM killer will not help higher order allocs */
4028 if (order > PAGE_ALLOC_COSTLY_ORDER)
4031 * We have already exhausted all our reclaim opportunities without any
4032 * success so it is time to admit defeat. We will skip the OOM killer
4033 * because it is very likely that the caller has a more reasonable
4034 * fallback than shooting a random task.
4036 * The OOM killer may not free memory on a specific node.
4038 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4040 /* The OOM killer does not needlessly kill tasks for lowmem */
4041 if (ac->highest_zoneidx < ZONE_NORMAL)
4043 if (pm_suspended_storage())
4046 * XXX: GFP_NOFS allocations should rather fail than rely on
4047 * other request to make a forward progress.
4048 * We are in an unfortunate situation where out_of_memory cannot
4049 * do much for this context but let's try it to at least get
4050 * access to memory reserved if the current task is killed (see
4051 * out_of_memory). Once filesystems are ready to handle allocation
4052 * failures more gracefully we should just bail out here.
4055 /* Exhausted what can be done so it's blame time */
4056 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4057 *did_some_progress = 1;
4060 * Help non-failing allocations by giving them access to memory
4063 if (gfp_mask & __GFP_NOFAIL)
4064 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4065 ALLOC_NO_WATERMARKS, ac);
4068 mutex_unlock(&oom_lock);
4073 * Maximum number of compaction retries wit a progress before OOM
4074 * killer is consider as the only way to move forward.
4076 #define MAX_COMPACT_RETRIES 16
4078 #ifdef CONFIG_COMPACTION
4079 /* Try memory compaction for high-order allocations before reclaim */
4080 static struct page *
4081 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4082 unsigned int alloc_flags, const struct alloc_context *ac,
4083 enum compact_priority prio, enum compact_result *compact_result)
4085 struct page *page = NULL;
4086 unsigned long pflags;
4087 unsigned int noreclaim_flag;
4092 psi_memstall_enter(&pflags);
4093 noreclaim_flag = memalloc_noreclaim_save();
4095 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4098 memalloc_noreclaim_restore(noreclaim_flag);
4099 psi_memstall_leave(&pflags);
4102 * At least in one zone compaction wasn't deferred or skipped, so let's
4103 * count a compaction stall
4105 count_vm_event(COMPACTSTALL);
4107 /* Prep a captured page if available */
4109 prep_new_page(page, order, gfp_mask, alloc_flags);
4111 /* Try get a page from the freelist if available */
4113 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4116 struct zone *zone = page_zone(page);
4118 zone->compact_blockskip_flush = false;
4119 compaction_defer_reset(zone, order, true);
4120 count_vm_event(COMPACTSUCCESS);
4125 * It's bad if compaction run occurs and fails. The most likely reason
4126 * is that pages exist, but not enough to satisfy watermarks.
4128 count_vm_event(COMPACTFAIL);
4136 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4137 enum compact_result compact_result,
4138 enum compact_priority *compact_priority,
4139 int *compaction_retries)
4141 int max_retries = MAX_COMPACT_RETRIES;
4144 int retries = *compaction_retries;
4145 enum compact_priority priority = *compact_priority;
4150 if (compaction_made_progress(compact_result))
4151 (*compaction_retries)++;
4154 * compaction considers all the zone as desperately out of memory
4155 * so it doesn't really make much sense to retry except when the
4156 * failure could be caused by insufficient priority
4158 if (compaction_failed(compact_result))
4159 goto check_priority;
4162 * compaction was skipped because there are not enough order-0 pages
4163 * to work with, so we retry only if it looks like reclaim can help.
4165 if (compaction_needs_reclaim(compact_result)) {
4166 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4171 * make sure the compaction wasn't deferred or didn't bail out early
4172 * due to locks contention before we declare that we should give up.
4173 * But the next retry should use a higher priority if allowed, so
4174 * we don't just keep bailing out endlessly.
4176 if (compaction_withdrawn(compact_result)) {
4177 goto check_priority;
4181 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4182 * costly ones because they are de facto nofail and invoke OOM
4183 * killer to move on while costly can fail and users are ready
4184 * to cope with that. 1/4 retries is rather arbitrary but we
4185 * would need much more detailed feedback from compaction to
4186 * make a better decision.
4188 if (order > PAGE_ALLOC_COSTLY_ORDER)
4190 if (*compaction_retries <= max_retries) {
4196 * Make sure there are attempts at the highest priority if we exhausted
4197 * all retries or failed at the lower priorities.
4200 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4201 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4203 if (*compact_priority > min_priority) {
4204 (*compact_priority)--;
4205 *compaction_retries = 0;
4209 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4213 static inline struct page *
4214 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4215 unsigned int alloc_flags, const struct alloc_context *ac,
4216 enum compact_priority prio, enum compact_result *compact_result)
4218 *compact_result = COMPACT_SKIPPED;
4223 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4224 enum compact_result compact_result,
4225 enum compact_priority *compact_priority,
4226 int *compaction_retries)
4231 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4235 * There are setups with compaction disabled which would prefer to loop
4236 * inside the allocator rather than hit the oom killer prematurely.
4237 * Let's give them a good hope and keep retrying while the order-0
4238 * watermarks are OK.
4240 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4241 ac->highest_zoneidx, ac->nodemask) {
4242 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4243 ac->highest_zoneidx, alloc_flags))
4248 #endif /* CONFIG_COMPACTION */
4250 #ifdef CONFIG_LOCKDEP
4251 static struct lockdep_map __fs_reclaim_map =
4252 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4254 static bool __need_fs_reclaim(gfp_t gfp_mask)
4256 gfp_mask = current_gfp_context(gfp_mask);
4258 /* no reclaim without waiting on it */
4259 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4262 /* this guy won't enter reclaim */
4263 if (current->flags & PF_MEMALLOC)
4266 /* We're only interested __GFP_FS allocations for now */
4267 if (!(gfp_mask & __GFP_FS))
4270 if (gfp_mask & __GFP_NOLOCKDEP)
4276 void __fs_reclaim_acquire(void)
4278 lock_map_acquire(&__fs_reclaim_map);
4281 void __fs_reclaim_release(void)
4283 lock_map_release(&__fs_reclaim_map);
4286 void fs_reclaim_acquire(gfp_t gfp_mask)
4288 if (__need_fs_reclaim(gfp_mask))
4289 __fs_reclaim_acquire();
4291 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4293 void fs_reclaim_release(gfp_t gfp_mask)
4295 if (__need_fs_reclaim(gfp_mask))
4296 __fs_reclaim_release();
4298 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4301 /* Perform direct synchronous page reclaim */
4302 static unsigned long
4303 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4304 const struct alloc_context *ac)
4306 unsigned int noreclaim_flag;
4307 unsigned long pflags, progress;
4311 /* We now go into synchronous reclaim */
4312 cpuset_memory_pressure_bump();
4313 psi_memstall_enter(&pflags);
4314 fs_reclaim_acquire(gfp_mask);
4315 noreclaim_flag = memalloc_noreclaim_save();
4317 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4320 memalloc_noreclaim_restore(noreclaim_flag);
4321 fs_reclaim_release(gfp_mask);
4322 psi_memstall_leave(&pflags);
4329 /* The really slow allocator path where we enter direct reclaim */
4330 static inline struct page *
4331 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4332 unsigned int alloc_flags, const struct alloc_context *ac,
4333 unsigned long *did_some_progress)
4335 struct page *page = NULL;
4336 bool drained = false;
4338 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4339 if (unlikely(!(*did_some_progress)))
4343 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4346 * If an allocation failed after direct reclaim, it could be because
4347 * pages are pinned on the per-cpu lists or in high alloc reserves.
4348 * Shrink them and try again
4350 if (!page && !drained) {
4351 unreserve_highatomic_pageblock(ac, false);
4352 drain_all_pages(NULL);
4360 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4361 const struct alloc_context *ac)
4365 pg_data_t *last_pgdat = NULL;
4366 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4368 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4370 if (last_pgdat != zone->zone_pgdat)
4371 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4372 last_pgdat = zone->zone_pgdat;
4376 static inline unsigned int
4377 gfp_to_alloc_flags(gfp_t gfp_mask)
4379 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4382 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4383 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4384 * to save two branches.
4386 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4387 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4390 * The caller may dip into page reserves a bit more if the caller
4391 * cannot run direct reclaim, or if the caller has realtime scheduling
4392 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4393 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4395 alloc_flags |= (__force int)
4396 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4398 if (gfp_mask & __GFP_ATOMIC) {
4400 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4401 * if it can't schedule.
4403 if (!(gfp_mask & __GFP_NOMEMALLOC))
4404 alloc_flags |= ALLOC_HARDER;
4406 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4407 * comment for __cpuset_node_allowed().
4409 alloc_flags &= ~ALLOC_CPUSET;
4410 } else if (unlikely(rt_task(current)) && !in_interrupt())
4411 alloc_flags |= ALLOC_HARDER;
4413 alloc_flags = current_alloc_flags(gfp_mask, alloc_flags);
4418 static bool oom_reserves_allowed(struct task_struct *tsk)
4420 if (!tsk_is_oom_victim(tsk))
4424 * !MMU doesn't have oom reaper so give access to memory reserves
4425 * only to the thread with TIF_MEMDIE set
4427 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4434 * Distinguish requests which really need access to full memory
4435 * reserves from oom victims which can live with a portion of it
4437 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4439 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4441 if (gfp_mask & __GFP_MEMALLOC)
4442 return ALLOC_NO_WATERMARKS;
4443 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4444 return ALLOC_NO_WATERMARKS;
4445 if (!in_interrupt()) {
4446 if (current->flags & PF_MEMALLOC)
4447 return ALLOC_NO_WATERMARKS;
4448 else if (oom_reserves_allowed(current))
4455 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4457 return !!__gfp_pfmemalloc_flags(gfp_mask);
4461 * Checks whether it makes sense to retry the reclaim to make a forward progress
4462 * for the given allocation request.
4464 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4465 * without success, or when we couldn't even meet the watermark if we
4466 * reclaimed all remaining pages on the LRU lists.
4468 * Returns true if a retry is viable or false to enter the oom path.
4471 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4472 struct alloc_context *ac, int alloc_flags,
4473 bool did_some_progress, int *no_progress_loops)
4480 * Costly allocations might have made a progress but this doesn't mean
4481 * their order will become available due to high fragmentation so
4482 * always increment the no progress counter for them
4484 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4485 *no_progress_loops = 0;
4487 (*no_progress_loops)++;
4490 * Make sure we converge to OOM if we cannot make any progress
4491 * several times in the row.
4493 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4494 /* Before OOM, exhaust highatomic_reserve */
4495 return unreserve_highatomic_pageblock(ac, true);
4499 * Keep reclaiming pages while there is a chance this will lead
4500 * somewhere. If none of the target zones can satisfy our allocation
4501 * request even if all reclaimable pages are considered then we are
4502 * screwed and have to go OOM.
4504 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4505 ac->highest_zoneidx, ac->nodemask) {
4506 unsigned long available;
4507 unsigned long reclaimable;
4508 unsigned long min_wmark = min_wmark_pages(zone);
4511 available = reclaimable = zone_reclaimable_pages(zone);
4512 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4515 * Would the allocation succeed if we reclaimed all
4516 * reclaimable pages?
4518 wmark = __zone_watermark_ok(zone, order, min_wmark,
4519 ac->highest_zoneidx, alloc_flags, available);
4520 trace_reclaim_retry_zone(z, order, reclaimable,
4521 available, min_wmark, *no_progress_loops, wmark);
4524 * If we didn't make any progress and have a lot of
4525 * dirty + writeback pages then we should wait for
4526 * an IO to complete to slow down the reclaim and
4527 * prevent from pre mature OOM
4529 if (!did_some_progress) {
4530 unsigned long write_pending;
4532 write_pending = zone_page_state_snapshot(zone,
4533 NR_ZONE_WRITE_PENDING);
4535 if (2 * write_pending > reclaimable) {
4536 congestion_wait(BLK_RW_ASYNC, HZ/10);
4548 * Memory allocation/reclaim might be called from a WQ context and the
4549 * current implementation of the WQ concurrency control doesn't
4550 * recognize that a particular WQ is congested if the worker thread is
4551 * looping without ever sleeping. Therefore we have to do a short sleep
4552 * here rather than calling cond_resched().
4554 if (current->flags & PF_WQ_WORKER)
4555 schedule_timeout_uninterruptible(1);
4562 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4565 * It's possible that cpuset's mems_allowed and the nodemask from
4566 * mempolicy don't intersect. This should be normally dealt with by
4567 * policy_nodemask(), but it's possible to race with cpuset update in
4568 * such a way the check therein was true, and then it became false
4569 * before we got our cpuset_mems_cookie here.
4570 * This assumes that for all allocations, ac->nodemask can come only
4571 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4572 * when it does not intersect with the cpuset restrictions) or the
4573 * caller can deal with a violated nodemask.
4575 if (cpusets_enabled() && ac->nodemask &&
4576 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4577 ac->nodemask = NULL;
4582 * When updating a task's mems_allowed or mempolicy nodemask, it is
4583 * possible to race with parallel threads in such a way that our
4584 * allocation can fail while the mask is being updated. If we are about
4585 * to fail, check if the cpuset changed during allocation and if so,
4588 if (read_mems_allowed_retry(cpuset_mems_cookie))
4594 static inline struct page *
4595 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4596 struct alloc_context *ac)
4598 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4599 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4600 struct page *page = NULL;
4601 unsigned int alloc_flags;
4602 unsigned long did_some_progress;
4603 enum compact_priority compact_priority;
4604 enum compact_result compact_result;
4605 int compaction_retries;
4606 int no_progress_loops;
4607 unsigned int cpuset_mems_cookie;
4611 * We also sanity check to catch abuse of atomic reserves being used by
4612 * callers that are not in atomic context.
4614 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4615 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4616 gfp_mask &= ~__GFP_ATOMIC;
4619 compaction_retries = 0;
4620 no_progress_loops = 0;
4621 compact_priority = DEF_COMPACT_PRIORITY;
4622 cpuset_mems_cookie = read_mems_allowed_begin();
4625 * The fast path uses conservative alloc_flags to succeed only until
4626 * kswapd needs to be woken up, and to avoid the cost of setting up
4627 * alloc_flags precisely. So we do that now.
4629 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4632 * We need to recalculate the starting point for the zonelist iterator
4633 * because we might have used different nodemask in the fast path, or
4634 * there was a cpuset modification and we are retrying - otherwise we
4635 * could end up iterating over non-eligible zones endlessly.
4637 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4638 ac->highest_zoneidx, ac->nodemask);
4639 if (!ac->preferred_zoneref->zone)
4642 if (alloc_flags & ALLOC_KSWAPD)
4643 wake_all_kswapds(order, gfp_mask, ac);
4646 * The adjusted alloc_flags might result in immediate success, so try
4649 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4654 * For costly allocations, try direct compaction first, as it's likely
4655 * that we have enough base pages and don't need to reclaim. For non-
4656 * movable high-order allocations, do that as well, as compaction will
4657 * try prevent permanent fragmentation by migrating from blocks of the
4659 * Don't try this for allocations that are allowed to ignore
4660 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4662 if (can_direct_reclaim &&
4664 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4665 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4666 page = __alloc_pages_direct_compact(gfp_mask, order,
4668 INIT_COMPACT_PRIORITY,
4674 * Checks for costly allocations with __GFP_NORETRY, which
4675 * includes some THP page fault allocations
4677 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4679 * If allocating entire pageblock(s) and compaction
4680 * failed because all zones are below low watermarks
4681 * or is prohibited because it recently failed at this
4682 * order, fail immediately unless the allocator has
4683 * requested compaction and reclaim retry.
4686 * - potentially very expensive because zones are far
4687 * below their low watermarks or this is part of very
4688 * bursty high order allocations,
4689 * - not guaranteed to help because isolate_freepages()
4690 * may not iterate over freed pages as part of its
4692 * - unlikely to make entire pageblocks free on its
4695 if (compact_result == COMPACT_SKIPPED ||
4696 compact_result == COMPACT_DEFERRED)
4700 * Looks like reclaim/compaction is worth trying, but
4701 * sync compaction could be very expensive, so keep
4702 * using async compaction.
4704 compact_priority = INIT_COMPACT_PRIORITY;
4709 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4710 if (alloc_flags & ALLOC_KSWAPD)
4711 wake_all_kswapds(order, gfp_mask, ac);
4713 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4715 alloc_flags = current_alloc_flags(gfp_mask, reserve_flags);
4718 * Reset the nodemask and zonelist iterators if memory policies can be
4719 * ignored. These allocations are high priority and system rather than
4722 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4723 ac->nodemask = NULL;
4724 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4725 ac->highest_zoneidx, ac->nodemask);
4728 /* Attempt with potentially adjusted zonelist and alloc_flags */
4729 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4733 /* Caller is not willing to reclaim, we can't balance anything */
4734 if (!can_direct_reclaim)
4737 /* Avoid recursion of direct reclaim */
4738 if (current->flags & PF_MEMALLOC)
4741 /* Try direct reclaim and then allocating */
4742 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4743 &did_some_progress);
4747 /* Try direct compaction and then allocating */
4748 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4749 compact_priority, &compact_result);
4753 /* Do not loop if specifically requested */
4754 if (gfp_mask & __GFP_NORETRY)
4758 * Do not retry costly high order allocations unless they are
4759 * __GFP_RETRY_MAYFAIL
4761 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4764 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4765 did_some_progress > 0, &no_progress_loops))
4769 * It doesn't make any sense to retry for the compaction if the order-0
4770 * reclaim is not able to make any progress because the current
4771 * implementation of the compaction depends on the sufficient amount
4772 * of free memory (see __compaction_suitable)
4774 if (did_some_progress > 0 &&
4775 should_compact_retry(ac, order, alloc_flags,
4776 compact_result, &compact_priority,
4777 &compaction_retries))
4781 /* Deal with possible cpuset update races before we start OOM killing */
4782 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4785 /* Reclaim has failed us, start killing things */
4786 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4790 /* Avoid allocations with no watermarks from looping endlessly */
4791 if (tsk_is_oom_victim(current) &&
4792 (alloc_flags & ALLOC_OOM ||
4793 (gfp_mask & __GFP_NOMEMALLOC)))
4796 /* Retry as long as the OOM killer is making progress */
4797 if (did_some_progress) {
4798 no_progress_loops = 0;
4803 /* Deal with possible cpuset update races before we fail */
4804 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4808 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4811 if (gfp_mask & __GFP_NOFAIL) {
4813 * All existing users of the __GFP_NOFAIL are blockable, so warn
4814 * of any new users that actually require GFP_NOWAIT
4816 if (WARN_ON_ONCE(!can_direct_reclaim))
4820 * PF_MEMALLOC request from this context is rather bizarre
4821 * because we cannot reclaim anything and only can loop waiting
4822 * for somebody to do a work for us
4824 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4827 * non failing costly orders are a hard requirement which we
4828 * are not prepared for much so let's warn about these users
4829 * so that we can identify them and convert them to something
4832 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4835 * Help non-failing allocations by giving them access to memory
4836 * reserves but do not use ALLOC_NO_WATERMARKS because this
4837 * could deplete whole memory reserves which would just make
4838 * the situation worse
4840 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4848 warn_alloc(gfp_mask, ac->nodemask,
4849 "page allocation failure: order:%u", order);
4854 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4855 int preferred_nid, nodemask_t *nodemask,
4856 struct alloc_context *ac, gfp_t *alloc_mask,
4857 unsigned int *alloc_flags)
4859 ac->highest_zoneidx = gfp_zone(gfp_mask);
4860 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4861 ac->nodemask = nodemask;
4862 ac->migratetype = gfp_migratetype(gfp_mask);
4864 if (cpusets_enabled()) {
4865 *alloc_mask |= __GFP_HARDWALL;
4867 * When we are in the interrupt context, it is irrelevant
4868 * to the current task context. It means that any node ok.
4870 if (!in_interrupt() && !ac->nodemask)
4871 ac->nodemask = &cpuset_current_mems_allowed;
4873 *alloc_flags |= ALLOC_CPUSET;
4876 fs_reclaim_acquire(gfp_mask);
4877 fs_reclaim_release(gfp_mask);
4879 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4881 if (should_fail_alloc_page(gfp_mask, order))
4884 *alloc_flags = current_alloc_flags(gfp_mask, *alloc_flags);
4886 /* Dirty zone balancing only done in the fast path */
4887 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4890 * The preferred zone is used for statistics but crucially it is
4891 * also used as the starting point for the zonelist iterator. It
4892 * may get reset for allocations that ignore memory policies.
4894 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4895 ac->highest_zoneidx, ac->nodemask);
4901 * This is the 'heart' of the zoned buddy allocator.
4904 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4905 nodemask_t *nodemask)
4908 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4909 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4910 struct alloc_context ac = { };
4913 * There are several places where we assume that the order value is sane
4914 * so bail out early if the request is out of bound.
4916 if (unlikely(order >= MAX_ORDER)) {
4917 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4921 gfp_mask &= gfp_allowed_mask;
4922 alloc_mask = gfp_mask;
4923 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4927 * Forbid the first pass from falling back to types that fragment
4928 * memory until all local zones are considered.
4930 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4932 /* First allocation attempt */
4933 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4938 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4939 * resp. GFP_NOIO which has to be inherited for all allocation requests
4940 * from a particular context which has been marked by
4941 * memalloc_no{fs,io}_{save,restore}.
4943 alloc_mask = current_gfp_context(gfp_mask);
4944 ac.spread_dirty_pages = false;
4947 * Restore the original nodemask if it was potentially replaced with
4948 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4950 ac.nodemask = nodemask;
4952 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4955 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4956 unlikely(__memcg_kmem_charge_page(page, gfp_mask, order) != 0)) {
4957 __free_pages(page, order);
4961 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4965 EXPORT_SYMBOL(__alloc_pages_nodemask);
4968 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4969 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4970 * you need to access high mem.
4972 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4976 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4979 return (unsigned long) page_address(page);
4981 EXPORT_SYMBOL(__get_free_pages);
4983 unsigned long get_zeroed_page(gfp_t gfp_mask)
4985 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4987 EXPORT_SYMBOL(get_zeroed_page);
4989 static inline void free_the_page(struct page *page, unsigned int order)
4991 if (order == 0) /* Via pcp? */
4992 free_unref_page(page);
4994 __free_pages_ok(page, order);
4997 void __free_pages(struct page *page, unsigned int order)
4999 if (put_page_testzero(page))
5000 free_the_page(page, order);
5001 else if (!PageHead(page))
5003 free_the_page(page + (1 << order), order);
5005 EXPORT_SYMBOL(__free_pages);
5007 void free_pages(unsigned long addr, unsigned int order)
5010 VM_BUG_ON(!virt_addr_valid((void *)addr));
5011 __free_pages(virt_to_page((void *)addr), order);
5015 EXPORT_SYMBOL(free_pages);
5019 * An arbitrary-length arbitrary-offset area of memory which resides
5020 * within a 0 or higher order page. Multiple fragments within that page
5021 * are individually refcounted, in the page's reference counter.
5023 * The page_frag functions below provide a simple allocation framework for
5024 * page fragments. This is used by the network stack and network device
5025 * drivers to provide a backing region of memory for use as either an
5026 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5028 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5031 struct page *page = NULL;
5032 gfp_t gfp = gfp_mask;
5034 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5035 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5037 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5038 PAGE_FRAG_CACHE_MAX_ORDER);
5039 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5041 if (unlikely(!page))
5042 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5044 nc->va = page ? page_address(page) : NULL;
5049 void __page_frag_cache_drain(struct page *page, unsigned int count)
5051 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5053 if (page_ref_sub_and_test(page, count))
5054 free_the_page(page, compound_order(page));
5056 EXPORT_SYMBOL(__page_frag_cache_drain);
5058 void *page_frag_alloc(struct page_frag_cache *nc,
5059 unsigned int fragsz, gfp_t gfp_mask)
5061 unsigned int size = PAGE_SIZE;
5065 if (unlikely(!nc->va)) {
5067 page = __page_frag_cache_refill(nc, gfp_mask);
5071 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5072 /* if size can vary use size else just use PAGE_SIZE */
5075 /* Even if we own the page, we do not use atomic_set().
5076 * This would break get_page_unless_zero() users.
5078 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5080 /* reset page count bias and offset to start of new frag */
5081 nc->pfmemalloc = page_is_pfmemalloc(page);
5082 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5086 offset = nc->offset - fragsz;
5087 if (unlikely(offset < 0)) {
5088 page = virt_to_page(nc->va);
5090 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5093 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5094 /* if size can vary use size else just use PAGE_SIZE */
5097 /* OK, page count is 0, we can safely set it */
5098 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5100 /* reset page count bias and offset to start of new frag */
5101 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5102 offset = size - fragsz;
5106 nc->offset = offset;
5108 return nc->va + offset;
5110 EXPORT_SYMBOL(page_frag_alloc);
5113 * Frees a page fragment allocated out of either a compound or order 0 page.
5115 void page_frag_free(void *addr)
5117 struct page *page = virt_to_head_page(addr);
5119 if (unlikely(put_page_testzero(page)))
5120 free_the_page(page, compound_order(page));
5122 EXPORT_SYMBOL(page_frag_free);
5124 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5128 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5129 unsigned long used = addr + PAGE_ALIGN(size);
5131 split_page(virt_to_page((void *)addr), order);
5132 while (used < alloc_end) {
5137 return (void *)addr;
5141 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5142 * @size: the number of bytes to allocate
5143 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5145 * This function is similar to alloc_pages(), except that it allocates the
5146 * minimum number of pages to satisfy the request. alloc_pages() can only
5147 * allocate memory in power-of-two pages.
5149 * This function is also limited by MAX_ORDER.
5151 * Memory allocated by this function must be released by free_pages_exact().
5153 * Return: pointer to the allocated area or %NULL in case of error.
5155 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5157 unsigned int order = get_order(size);
5160 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5161 gfp_mask &= ~__GFP_COMP;
5163 addr = __get_free_pages(gfp_mask, order);
5164 return make_alloc_exact(addr, order, size);
5166 EXPORT_SYMBOL(alloc_pages_exact);
5169 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5171 * @nid: the preferred node ID where memory should be allocated
5172 * @size: the number of bytes to allocate
5173 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5175 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5178 * Return: pointer to the allocated area or %NULL in case of error.
5180 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5182 unsigned int order = get_order(size);
5185 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5186 gfp_mask &= ~__GFP_COMP;
5188 p = alloc_pages_node(nid, gfp_mask, order);
5191 return make_alloc_exact((unsigned long)page_address(p), order, size);
5195 * free_pages_exact - release memory allocated via alloc_pages_exact()
5196 * @virt: the value returned by alloc_pages_exact.
5197 * @size: size of allocation, same value as passed to alloc_pages_exact().
5199 * Release the memory allocated by a previous call to alloc_pages_exact.
5201 void free_pages_exact(void *virt, size_t size)
5203 unsigned long addr = (unsigned long)virt;
5204 unsigned long end = addr + PAGE_ALIGN(size);
5206 while (addr < end) {
5211 EXPORT_SYMBOL(free_pages_exact);
5214 * nr_free_zone_pages - count number of pages beyond high watermark
5215 * @offset: The zone index of the highest zone
5217 * nr_free_zone_pages() counts the number of pages which are beyond the
5218 * high watermark within all zones at or below a given zone index. For each
5219 * zone, the number of pages is calculated as:
5221 * nr_free_zone_pages = managed_pages - high_pages
5223 * Return: number of pages beyond high watermark.
5225 static unsigned long nr_free_zone_pages(int offset)
5230 /* Just pick one node, since fallback list is circular */
5231 unsigned long sum = 0;
5233 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5235 for_each_zone_zonelist(zone, z, zonelist, offset) {
5236 unsigned long size = zone_managed_pages(zone);
5237 unsigned long high = high_wmark_pages(zone);
5246 * nr_free_buffer_pages - count number of pages beyond high watermark
5248 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5249 * watermark within ZONE_DMA and ZONE_NORMAL.
5251 * Return: number of pages beyond high watermark within ZONE_DMA and
5254 unsigned long nr_free_buffer_pages(void)
5256 return nr_free_zone_pages(gfp_zone(GFP_USER));
5258 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5260 static inline void show_node(struct zone *zone)
5262 if (IS_ENABLED(CONFIG_NUMA))
5263 printk("Node %d ", zone_to_nid(zone));
5266 long si_mem_available(void)
5269 unsigned long pagecache;
5270 unsigned long wmark_low = 0;
5271 unsigned long pages[NR_LRU_LISTS];
5272 unsigned long reclaimable;
5276 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5277 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5280 wmark_low += low_wmark_pages(zone);
5283 * Estimate the amount of memory available for userspace allocations,
5284 * without causing swapping.
5286 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5289 * Not all the page cache can be freed, otherwise the system will
5290 * start swapping. Assume at least half of the page cache, or the
5291 * low watermark worth of cache, needs to stay.
5293 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5294 pagecache -= min(pagecache / 2, wmark_low);
5295 available += pagecache;
5298 * Part of the reclaimable slab and other kernel memory consists of
5299 * items that are in use, and cannot be freed. Cap this estimate at the
5302 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5303 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5304 available += reclaimable - min(reclaimable / 2, wmark_low);
5310 EXPORT_SYMBOL_GPL(si_mem_available);
5312 void si_meminfo(struct sysinfo *val)
5314 val->totalram = totalram_pages();
5315 val->sharedram = global_node_page_state(NR_SHMEM);
5316 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5317 val->bufferram = nr_blockdev_pages();
5318 val->totalhigh = totalhigh_pages();
5319 val->freehigh = nr_free_highpages();
5320 val->mem_unit = PAGE_SIZE;
5323 EXPORT_SYMBOL(si_meminfo);
5326 void si_meminfo_node(struct sysinfo *val, int nid)
5328 int zone_type; /* needs to be signed */
5329 unsigned long managed_pages = 0;
5330 unsigned long managed_highpages = 0;
5331 unsigned long free_highpages = 0;
5332 pg_data_t *pgdat = NODE_DATA(nid);
5334 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5335 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5336 val->totalram = managed_pages;
5337 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5338 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5339 #ifdef CONFIG_HIGHMEM
5340 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5341 struct zone *zone = &pgdat->node_zones[zone_type];
5343 if (is_highmem(zone)) {
5344 managed_highpages += zone_managed_pages(zone);
5345 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5348 val->totalhigh = managed_highpages;
5349 val->freehigh = free_highpages;
5351 val->totalhigh = managed_highpages;
5352 val->freehigh = free_highpages;
5354 val->mem_unit = PAGE_SIZE;
5359 * Determine whether the node should be displayed or not, depending on whether
5360 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5362 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5364 if (!(flags & SHOW_MEM_FILTER_NODES))
5368 * no node mask - aka implicit memory numa policy. Do not bother with
5369 * the synchronization - read_mems_allowed_begin - because we do not
5370 * have to be precise here.
5373 nodemask = &cpuset_current_mems_allowed;
5375 return !node_isset(nid, *nodemask);
5378 #define K(x) ((x) << (PAGE_SHIFT-10))
5380 static void show_migration_types(unsigned char type)
5382 static const char types[MIGRATE_TYPES] = {
5383 [MIGRATE_UNMOVABLE] = 'U',
5384 [MIGRATE_MOVABLE] = 'M',
5385 [MIGRATE_RECLAIMABLE] = 'E',
5386 [MIGRATE_HIGHATOMIC] = 'H',
5388 [MIGRATE_CMA] = 'C',
5390 #ifdef CONFIG_MEMORY_ISOLATION
5391 [MIGRATE_ISOLATE] = 'I',
5394 char tmp[MIGRATE_TYPES + 1];
5398 for (i = 0; i < MIGRATE_TYPES; i++) {
5399 if (type & (1 << i))
5404 printk(KERN_CONT "(%s) ", tmp);
5408 * Show free area list (used inside shift_scroll-lock stuff)
5409 * We also calculate the percentage fragmentation. We do this by counting the
5410 * memory on each free list with the exception of the first item on the list.
5413 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5416 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5418 unsigned long free_pcp = 0;
5423 for_each_populated_zone(zone) {
5424 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5427 for_each_online_cpu(cpu)
5428 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5431 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5432 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5433 " unevictable:%lu dirty:%lu writeback:%lu\n"
5434 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5435 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5436 " free:%lu free_pcp:%lu free_cma:%lu\n",
5437 global_node_page_state(NR_ACTIVE_ANON),
5438 global_node_page_state(NR_INACTIVE_ANON),
5439 global_node_page_state(NR_ISOLATED_ANON),
5440 global_node_page_state(NR_ACTIVE_FILE),
5441 global_node_page_state(NR_INACTIVE_FILE),
5442 global_node_page_state(NR_ISOLATED_FILE),
5443 global_node_page_state(NR_UNEVICTABLE),
5444 global_node_page_state(NR_FILE_DIRTY),
5445 global_node_page_state(NR_WRITEBACK),
5446 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5447 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5448 global_node_page_state(NR_FILE_MAPPED),
5449 global_node_page_state(NR_SHMEM),
5450 global_zone_page_state(NR_PAGETABLE),
5451 global_zone_page_state(NR_BOUNCE),
5452 global_zone_page_state(NR_FREE_PAGES),
5454 global_zone_page_state(NR_FREE_CMA_PAGES));
5456 for_each_online_pgdat(pgdat) {
5457 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5461 " active_anon:%lukB"
5462 " inactive_anon:%lukB"
5463 " active_file:%lukB"
5464 " inactive_file:%lukB"
5465 " unevictable:%lukB"
5466 " isolated(anon):%lukB"
5467 " isolated(file):%lukB"
5472 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5474 " shmem_pmdmapped: %lukB"
5477 " writeback_tmp:%lukB"
5478 " kernel_stack:%lukB"
5479 #ifdef CONFIG_SHADOW_CALL_STACK
5480 " shadow_call_stack:%lukB"
5482 " all_unreclaimable? %s"
5485 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5486 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5487 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5488 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5489 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5490 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5491 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5492 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5493 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5494 K(node_page_state(pgdat, NR_WRITEBACK)),
5495 K(node_page_state(pgdat, NR_SHMEM)),
5496 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5497 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5498 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5500 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5502 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5503 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5504 #ifdef CONFIG_SHADOW_CALL_STACK
5505 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5507 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5511 for_each_populated_zone(zone) {
5514 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5518 for_each_online_cpu(cpu)
5519 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5528 " reserved_highatomic:%luKB"
5529 " active_anon:%lukB"
5530 " inactive_anon:%lukB"
5531 " active_file:%lukB"
5532 " inactive_file:%lukB"
5533 " unevictable:%lukB"
5534 " writepending:%lukB"
5545 K(zone_page_state(zone, NR_FREE_PAGES)),
5546 K(min_wmark_pages(zone)),
5547 K(low_wmark_pages(zone)),
5548 K(high_wmark_pages(zone)),
5549 K(zone->nr_reserved_highatomic),
5550 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5551 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5552 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5553 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5554 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5555 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5556 K(zone->present_pages),
5557 K(zone_managed_pages(zone)),
5558 K(zone_page_state(zone, NR_MLOCK)),
5559 K(zone_page_state(zone, NR_PAGETABLE)),
5560 K(zone_page_state(zone, NR_BOUNCE)),
5562 K(this_cpu_read(zone->pageset->pcp.count)),
5563 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5564 printk("lowmem_reserve[]:");
5565 for (i = 0; i < MAX_NR_ZONES; i++)
5566 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5567 printk(KERN_CONT "\n");
5570 for_each_populated_zone(zone) {
5572 unsigned long nr[MAX_ORDER], flags, total = 0;
5573 unsigned char types[MAX_ORDER];
5575 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5578 printk(KERN_CONT "%s: ", zone->name);
5580 spin_lock_irqsave(&zone->lock, flags);
5581 for (order = 0; order < MAX_ORDER; order++) {
5582 struct free_area *area = &zone->free_area[order];
5585 nr[order] = area->nr_free;
5586 total += nr[order] << order;
5589 for (type = 0; type < MIGRATE_TYPES; type++) {
5590 if (!free_area_empty(area, type))
5591 types[order] |= 1 << type;
5594 spin_unlock_irqrestore(&zone->lock, flags);
5595 for (order = 0; order < MAX_ORDER; order++) {
5596 printk(KERN_CONT "%lu*%lukB ",
5597 nr[order], K(1UL) << order);
5599 show_migration_types(types[order]);
5601 printk(KERN_CONT "= %lukB\n", K(total));
5604 hugetlb_show_meminfo();
5606 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5608 show_swap_cache_info();
5611 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5613 zoneref->zone = zone;
5614 zoneref->zone_idx = zone_idx(zone);
5618 * Builds allocation fallback zone lists.
5620 * Add all populated zones of a node to the zonelist.
5622 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5625 enum zone_type zone_type = MAX_NR_ZONES;
5630 zone = pgdat->node_zones + zone_type;
5631 if (managed_zone(zone)) {
5632 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5633 check_highest_zone(zone_type);
5635 } while (zone_type);
5642 static int __parse_numa_zonelist_order(char *s)
5645 * We used to support different zonlists modes but they turned
5646 * out to be just not useful. Let's keep the warning in place
5647 * if somebody still use the cmd line parameter so that we do
5648 * not fail it silently
5650 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5651 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5657 char numa_zonelist_order[] = "Node";
5660 * sysctl handler for numa_zonelist_order
5662 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5663 void *buffer, size_t *length, loff_t *ppos)
5666 return __parse_numa_zonelist_order(buffer);
5667 return proc_dostring(table, write, buffer, length, ppos);
5671 #define MAX_NODE_LOAD (nr_online_nodes)
5672 static int node_load[MAX_NUMNODES];
5675 * find_next_best_node - find the next node that should appear in a given node's fallback list
5676 * @node: node whose fallback list we're appending
5677 * @used_node_mask: nodemask_t of already used nodes
5679 * We use a number of factors to determine which is the next node that should
5680 * appear on a given node's fallback list. The node should not have appeared
5681 * already in @node's fallback list, and it should be the next closest node
5682 * according to the distance array (which contains arbitrary distance values
5683 * from each node to each node in the system), and should also prefer nodes
5684 * with no CPUs, since presumably they'll have very little allocation pressure
5685 * on them otherwise.
5687 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5689 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5692 int min_val = INT_MAX;
5693 int best_node = NUMA_NO_NODE;
5695 /* Use the local node if we haven't already */
5696 if (!node_isset(node, *used_node_mask)) {
5697 node_set(node, *used_node_mask);
5701 for_each_node_state(n, N_MEMORY) {
5703 /* Don't want a node to appear more than once */
5704 if (node_isset(n, *used_node_mask))
5707 /* Use the distance array to find the distance */
5708 val = node_distance(node, n);
5710 /* Penalize nodes under us ("prefer the next node") */
5713 /* Give preference to headless and unused nodes */
5714 if (!cpumask_empty(cpumask_of_node(n)))
5715 val += PENALTY_FOR_NODE_WITH_CPUS;
5717 /* Slight preference for less loaded node */
5718 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5719 val += node_load[n];
5721 if (val < min_val) {
5728 node_set(best_node, *used_node_mask);
5735 * Build zonelists ordered by node and zones within node.
5736 * This results in maximum locality--normal zone overflows into local
5737 * DMA zone, if any--but risks exhausting DMA zone.
5739 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5742 struct zoneref *zonerefs;
5745 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5747 for (i = 0; i < nr_nodes; i++) {
5750 pg_data_t *node = NODE_DATA(node_order[i]);
5752 nr_zones = build_zonerefs_node(node, zonerefs);
5753 zonerefs += nr_zones;
5755 zonerefs->zone = NULL;
5756 zonerefs->zone_idx = 0;
5760 * Build gfp_thisnode zonelists
5762 static void build_thisnode_zonelists(pg_data_t *pgdat)
5764 struct zoneref *zonerefs;
5767 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5768 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5769 zonerefs += nr_zones;
5770 zonerefs->zone = NULL;
5771 zonerefs->zone_idx = 0;
5775 * Build zonelists ordered by zone and nodes within zones.
5776 * This results in conserving DMA zone[s] until all Normal memory is
5777 * exhausted, but results in overflowing to remote node while memory
5778 * may still exist in local DMA zone.
5781 static void build_zonelists(pg_data_t *pgdat)
5783 static int node_order[MAX_NUMNODES];
5784 int node, load, nr_nodes = 0;
5785 nodemask_t used_mask = NODE_MASK_NONE;
5786 int local_node, prev_node;
5788 /* NUMA-aware ordering of nodes */
5789 local_node = pgdat->node_id;
5790 load = nr_online_nodes;
5791 prev_node = local_node;
5793 memset(node_order, 0, sizeof(node_order));
5794 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5796 * We don't want to pressure a particular node.
5797 * So adding penalty to the first node in same
5798 * distance group to make it round-robin.
5800 if (node_distance(local_node, node) !=
5801 node_distance(local_node, prev_node))
5802 node_load[node] = load;
5804 node_order[nr_nodes++] = node;
5809 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5810 build_thisnode_zonelists(pgdat);
5813 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5815 * Return node id of node used for "local" allocations.
5816 * I.e., first node id of first zone in arg node's generic zonelist.
5817 * Used for initializing percpu 'numa_mem', which is used primarily
5818 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5820 int local_memory_node(int node)
5824 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5825 gfp_zone(GFP_KERNEL),
5827 return zone_to_nid(z->zone);
5831 static void setup_min_unmapped_ratio(void);
5832 static void setup_min_slab_ratio(void);
5833 #else /* CONFIG_NUMA */
5835 static void build_zonelists(pg_data_t *pgdat)
5837 int node, local_node;
5838 struct zoneref *zonerefs;
5841 local_node = pgdat->node_id;
5843 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5844 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5845 zonerefs += nr_zones;
5848 * Now we build the zonelist so that it contains the zones
5849 * of all the other nodes.
5850 * We don't want to pressure a particular node, so when
5851 * building the zones for node N, we make sure that the
5852 * zones coming right after the local ones are those from
5853 * node N+1 (modulo N)
5855 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5856 if (!node_online(node))
5858 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5859 zonerefs += nr_zones;
5861 for (node = 0; node < local_node; node++) {
5862 if (!node_online(node))
5864 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5865 zonerefs += nr_zones;
5868 zonerefs->zone = NULL;
5869 zonerefs->zone_idx = 0;
5872 #endif /* CONFIG_NUMA */
5875 * Boot pageset table. One per cpu which is going to be used for all
5876 * zones and all nodes. The parameters will be set in such a way
5877 * that an item put on a list will immediately be handed over to
5878 * the buddy list. This is safe since pageset manipulation is done
5879 * with interrupts disabled.
5881 * The boot_pagesets must be kept even after bootup is complete for
5882 * unused processors and/or zones. They do play a role for bootstrapping
5883 * hotplugged processors.
5885 * zoneinfo_show() and maybe other functions do
5886 * not check if the processor is online before following the pageset pointer.
5887 * Other parts of the kernel may not check if the zone is available.
5889 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5890 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5891 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5893 static void __build_all_zonelists(void *data)
5896 int __maybe_unused cpu;
5897 pg_data_t *self = data;
5898 static DEFINE_SPINLOCK(lock);
5903 memset(node_load, 0, sizeof(node_load));
5907 * This node is hotadded and no memory is yet present. So just
5908 * building zonelists is fine - no need to touch other nodes.
5910 if (self && !node_online(self->node_id)) {
5911 build_zonelists(self);
5913 for_each_online_node(nid) {
5914 pg_data_t *pgdat = NODE_DATA(nid);
5916 build_zonelists(pgdat);
5919 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5921 * We now know the "local memory node" for each node--
5922 * i.e., the node of the first zone in the generic zonelist.
5923 * Set up numa_mem percpu variable for on-line cpus. During
5924 * boot, only the boot cpu should be on-line; we'll init the
5925 * secondary cpus' numa_mem as they come on-line. During
5926 * node/memory hotplug, we'll fixup all on-line cpus.
5928 for_each_online_cpu(cpu)
5929 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5936 static noinline void __init
5937 build_all_zonelists_init(void)
5941 __build_all_zonelists(NULL);
5944 * Initialize the boot_pagesets that are going to be used
5945 * for bootstrapping processors. The real pagesets for
5946 * each zone will be allocated later when the per cpu
5947 * allocator is available.
5949 * boot_pagesets are used also for bootstrapping offline
5950 * cpus if the system is already booted because the pagesets
5951 * are needed to initialize allocators on a specific cpu too.
5952 * F.e. the percpu allocator needs the page allocator which
5953 * needs the percpu allocator in order to allocate its pagesets
5954 * (a chicken-egg dilemma).
5956 for_each_possible_cpu(cpu)
5957 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5959 mminit_verify_zonelist();
5960 cpuset_init_current_mems_allowed();
5964 * unless system_state == SYSTEM_BOOTING.
5966 * __ref due to call of __init annotated helper build_all_zonelists_init
5967 * [protected by SYSTEM_BOOTING].
5969 void __ref build_all_zonelists(pg_data_t *pgdat)
5971 unsigned long vm_total_pages;
5973 if (system_state == SYSTEM_BOOTING) {
5974 build_all_zonelists_init();
5976 __build_all_zonelists(pgdat);
5977 /* cpuset refresh routine should be here */
5979 /* Get the number of free pages beyond high watermark in all zones. */
5980 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5982 * Disable grouping by mobility if the number of pages in the
5983 * system is too low to allow the mechanism to work. It would be
5984 * more accurate, but expensive to check per-zone. This check is
5985 * made on memory-hotadd so a system can start with mobility
5986 * disabled and enable it later
5988 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5989 page_group_by_mobility_disabled = 1;
5991 page_group_by_mobility_disabled = 0;
5993 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5995 page_group_by_mobility_disabled ? "off" : "on",
5998 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6002 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6003 static bool __meminit
6004 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6006 static struct memblock_region *r;
6008 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6009 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6010 for_each_mem_region(r) {
6011 if (*pfn < memblock_region_memory_end_pfn(r))
6015 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6016 memblock_is_mirror(r)) {
6017 *pfn = memblock_region_memory_end_pfn(r);
6025 * Initially all pages are reserved - free ones are freed
6026 * up by memblock_free_all() once the early boot process is
6027 * done. Non-atomic initialization, single-pass.
6029 * All aligned pageblocks are initialized to the specified migratetype
6030 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6031 * zone stats (e.g., nr_isolate_pageblock) are touched.
6033 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
6034 unsigned long start_pfn,
6035 enum meminit_context context,
6036 struct vmem_altmap *altmap, int migratetype)
6038 unsigned long pfn, end_pfn = start_pfn + size;
6041 if (highest_memmap_pfn < end_pfn - 1)
6042 highest_memmap_pfn = end_pfn - 1;
6044 #ifdef CONFIG_ZONE_DEVICE
6046 * Honor reservation requested by the driver for this ZONE_DEVICE
6047 * memory. We limit the total number of pages to initialize to just
6048 * those that might contain the memory mapping. We will defer the
6049 * ZONE_DEVICE page initialization until after we have released
6052 if (zone == ZONE_DEVICE) {
6056 if (start_pfn == altmap->base_pfn)
6057 start_pfn += altmap->reserve;
6058 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6062 for (pfn = start_pfn; pfn < end_pfn; ) {
6064 * There can be holes in boot-time mem_map[]s handed to this
6065 * function. They do not exist on hotplugged memory.
6067 if (context == MEMINIT_EARLY) {
6068 if (overlap_memmap_init(zone, &pfn))
6070 if (defer_init(nid, pfn, end_pfn))
6074 page = pfn_to_page(pfn);
6075 __init_single_page(page, pfn, zone, nid);
6076 if (context == MEMINIT_HOTPLUG)
6077 __SetPageReserved(page);
6080 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6081 * such that unmovable allocations won't be scattered all
6082 * over the place during system boot.
6084 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6085 set_pageblock_migratetype(page, migratetype);
6092 #ifdef CONFIG_ZONE_DEVICE
6093 void __ref memmap_init_zone_device(struct zone *zone,
6094 unsigned long start_pfn,
6095 unsigned long nr_pages,
6096 struct dev_pagemap *pgmap)
6098 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6099 struct pglist_data *pgdat = zone->zone_pgdat;
6100 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6101 unsigned long zone_idx = zone_idx(zone);
6102 unsigned long start = jiffies;
6103 int nid = pgdat->node_id;
6105 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6109 * The call to memmap_init_zone should have already taken care
6110 * of the pages reserved for the memmap, so we can just jump to
6111 * the end of that region and start processing the device pages.
6114 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6115 nr_pages = end_pfn - start_pfn;
6118 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6119 struct page *page = pfn_to_page(pfn);
6121 __init_single_page(page, pfn, zone_idx, nid);
6124 * Mark page reserved as it will need to wait for onlining
6125 * phase for it to be fully associated with a zone.
6127 * We can use the non-atomic __set_bit operation for setting
6128 * the flag as we are still initializing the pages.
6130 __SetPageReserved(page);
6133 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6134 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6135 * ever freed or placed on a driver-private list.
6137 page->pgmap = pgmap;
6138 page->zone_device_data = NULL;
6141 * Mark the block movable so that blocks are reserved for
6142 * movable at startup. This will force kernel allocations
6143 * to reserve their blocks rather than leaking throughout
6144 * the address space during boot when many long-lived
6145 * kernel allocations are made.
6147 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6148 * because this is done early in section_activate()
6150 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6151 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6156 pr_info("%s initialised %lu pages in %ums\n", __func__,
6157 nr_pages, jiffies_to_msecs(jiffies - start));
6161 static void __meminit zone_init_free_lists(struct zone *zone)
6163 unsigned int order, t;
6164 for_each_migratetype_order(order, t) {
6165 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6166 zone->free_area[order].nr_free = 0;
6170 void __meminit __weak memmap_init(unsigned long size, int nid,
6172 unsigned long range_start_pfn)
6174 unsigned long start_pfn, end_pfn;
6175 unsigned long range_end_pfn = range_start_pfn + size;
6178 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6179 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6180 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6182 if (end_pfn > start_pfn) {
6183 size = end_pfn - start_pfn;
6184 memmap_init_zone(size, nid, zone, start_pfn,
6185 MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6190 static int zone_batchsize(struct zone *zone)
6196 * The per-cpu-pages pools are set to around 1000th of the
6199 batch = zone_managed_pages(zone) / 1024;
6200 /* But no more than a meg. */
6201 if (batch * PAGE_SIZE > 1024 * 1024)
6202 batch = (1024 * 1024) / PAGE_SIZE;
6203 batch /= 4; /* We effectively *= 4 below */
6208 * Clamp the batch to a 2^n - 1 value. Having a power
6209 * of 2 value was found to be more likely to have
6210 * suboptimal cache aliasing properties in some cases.
6212 * For example if 2 tasks are alternately allocating
6213 * batches of pages, one task can end up with a lot
6214 * of pages of one half of the possible page colors
6215 * and the other with pages of the other colors.
6217 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6222 /* The deferral and batching of frees should be suppressed under NOMMU
6225 * The problem is that NOMMU needs to be able to allocate large chunks
6226 * of contiguous memory as there's no hardware page translation to
6227 * assemble apparent contiguous memory from discontiguous pages.
6229 * Queueing large contiguous runs of pages for batching, however,
6230 * causes the pages to actually be freed in smaller chunks. As there
6231 * can be a significant delay between the individual batches being
6232 * recycled, this leads to the once large chunks of space being
6233 * fragmented and becoming unavailable for high-order allocations.
6240 * pcp->high and pcp->batch values are related and dependent on one another:
6241 * ->batch must never be higher then ->high.
6242 * The following function updates them in a safe manner without read side
6245 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6246 * those fields changing asynchronously (acording to the above rule).
6248 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6249 * outside of boot time (or some other assurance that no concurrent updaters
6252 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6253 unsigned long batch)
6255 /* start with a fail safe value for batch */
6259 /* Update high, then batch, in order */
6266 /* a companion to pageset_set_high() */
6267 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6269 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6272 static void pageset_init(struct per_cpu_pageset *p)
6274 struct per_cpu_pages *pcp;
6277 memset(p, 0, sizeof(*p));
6280 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6281 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6284 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6287 pageset_set_batch(p, batch);
6291 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6292 * to the value high for the pageset p.
6294 static void pageset_set_high(struct per_cpu_pageset *p,
6297 unsigned long batch = max(1UL, high / 4);
6298 if ((high / 4) > (PAGE_SHIFT * 8))
6299 batch = PAGE_SHIFT * 8;
6301 pageset_update(&p->pcp, high, batch);
6304 static void pageset_set_high_and_batch(struct zone *zone,
6305 struct per_cpu_pageset *pcp)
6307 if (percpu_pagelist_fraction)
6308 pageset_set_high(pcp,
6309 (zone_managed_pages(zone) /
6310 percpu_pagelist_fraction));
6312 pageset_set_batch(pcp, zone_batchsize(zone));
6315 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6317 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6320 pageset_set_high_and_batch(zone, pcp);
6323 void __meminit setup_zone_pageset(struct zone *zone)
6326 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6327 for_each_possible_cpu(cpu)
6328 zone_pageset_init(zone, cpu);
6332 * Allocate per cpu pagesets and initialize them.
6333 * Before this call only boot pagesets were available.
6335 void __init setup_per_cpu_pageset(void)
6337 struct pglist_data *pgdat;
6339 int __maybe_unused cpu;
6341 for_each_populated_zone(zone)
6342 setup_zone_pageset(zone);
6346 * Unpopulated zones continue using the boot pagesets.
6347 * The numa stats for these pagesets need to be reset.
6348 * Otherwise, they will end up skewing the stats of
6349 * the nodes these zones are associated with.
6351 for_each_possible_cpu(cpu) {
6352 struct per_cpu_pageset *pcp = &per_cpu(boot_pageset, cpu);
6353 memset(pcp->vm_numa_stat_diff, 0,
6354 sizeof(pcp->vm_numa_stat_diff));
6358 for_each_online_pgdat(pgdat)
6359 pgdat->per_cpu_nodestats =
6360 alloc_percpu(struct per_cpu_nodestat);
6363 static __meminit void zone_pcp_init(struct zone *zone)
6366 * per cpu subsystem is not up at this point. The following code
6367 * relies on the ability of the linker to provide the
6368 * offset of a (static) per cpu variable into the per cpu area.
6370 zone->pageset = &boot_pageset;
6372 if (populated_zone(zone))
6373 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6374 zone->name, zone->present_pages,
6375 zone_batchsize(zone));
6378 void __meminit init_currently_empty_zone(struct zone *zone,
6379 unsigned long zone_start_pfn,
6382 struct pglist_data *pgdat = zone->zone_pgdat;
6383 int zone_idx = zone_idx(zone) + 1;
6385 if (zone_idx > pgdat->nr_zones)
6386 pgdat->nr_zones = zone_idx;
6388 zone->zone_start_pfn = zone_start_pfn;
6390 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6391 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6393 (unsigned long)zone_idx(zone),
6394 zone_start_pfn, (zone_start_pfn + size));
6396 zone_init_free_lists(zone);
6397 zone->initialized = 1;
6401 * get_pfn_range_for_nid - Return the start and end page frames for a node
6402 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6403 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6404 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6406 * It returns the start and end page frame of a node based on information
6407 * provided by memblock_set_node(). If called for a node
6408 * with no available memory, a warning is printed and the start and end
6411 void __init get_pfn_range_for_nid(unsigned int nid,
6412 unsigned long *start_pfn, unsigned long *end_pfn)
6414 unsigned long this_start_pfn, this_end_pfn;
6420 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6421 *start_pfn = min(*start_pfn, this_start_pfn);
6422 *end_pfn = max(*end_pfn, this_end_pfn);
6425 if (*start_pfn == -1UL)
6430 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6431 * assumption is made that zones within a node are ordered in monotonic
6432 * increasing memory addresses so that the "highest" populated zone is used
6434 static void __init find_usable_zone_for_movable(void)
6437 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6438 if (zone_index == ZONE_MOVABLE)
6441 if (arch_zone_highest_possible_pfn[zone_index] >
6442 arch_zone_lowest_possible_pfn[zone_index])
6446 VM_BUG_ON(zone_index == -1);
6447 movable_zone = zone_index;
6451 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6452 * because it is sized independent of architecture. Unlike the other zones,
6453 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6454 * in each node depending on the size of each node and how evenly kernelcore
6455 * is distributed. This helper function adjusts the zone ranges
6456 * provided by the architecture for a given node by using the end of the
6457 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6458 * zones within a node are in order of monotonic increases memory addresses
6460 static void __init adjust_zone_range_for_zone_movable(int nid,
6461 unsigned long zone_type,
6462 unsigned long node_start_pfn,
6463 unsigned long node_end_pfn,
6464 unsigned long *zone_start_pfn,
6465 unsigned long *zone_end_pfn)
6467 /* Only adjust if ZONE_MOVABLE is on this node */
6468 if (zone_movable_pfn[nid]) {
6469 /* Size ZONE_MOVABLE */
6470 if (zone_type == ZONE_MOVABLE) {
6471 *zone_start_pfn = zone_movable_pfn[nid];
6472 *zone_end_pfn = min(node_end_pfn,
6473 arch_zone_highest_possible_pfn[movable_zone]);
6475 /* Adjust for ZONE_MOVABLE starting within this range */
6476 } else if (!mirrored_kernelcore &&
6477 *zone_start_pfn < zone_movable_pfn[nid] &&
6478 *zone_end_pfn > zone_movable_pfn[nid]) {
6479 *zone_end_pfn = zone_movable_pfn[nid];
6481 /* Check if this whole range is within ZONE_MOVABLE */
6482 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6483 *zone_start_pfn = *zone_end_pfn;
6488 * Return the number of pages a zone spans in a node, including holes
6489 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6491 static unsigned long __init zone_spanned_pages_in_node(int nid,
6492 unsigned long zone_type,
6493 unsigned long node_start_pfn,
6494 unsigned long node_end_pfn,
6495 unsigned long *zone_start_pfn,
6496 unsigned long *zone_end_pfn)
6498 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6499 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6500 /* When hotadd a new node from cpu_up(), the node should be empty */
6501 if (!node_start_pfn && !node_end_pfn)
6504 /* Get the start and end of the zone */
6505 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6506 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6507 adjust_zone_range_for_zone_movable(nid, zone_type,
6508 node_start_pfn, node_end_pfn,
6509 zone_start_pfn, zone_end_pfn);
6511 /* Check that this node has pages within the zone's required range */
6512 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6515 /* Move the zone boundaries inside the node if necessary */
6516 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6517 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6519 /* Return the spanned pages */
6520 return *zone_end_pfn - *zone_start_pfn;
6524 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6525 * then all holes in the requested range will be accounted for.
6527 unsigned long __init __absent_pages_in_range(int nid,
6528 unsigned long range_start_pfn,
6529 unsigned long range_end_pfn)
6531 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6532 unsigned long start_pfn, end_pfn;
6535 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6536 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6537 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6538 nr_absent -= end_pfn - start_pfn;
6544 * absent_pages_in_range - Return number of page frames in holes within a range
6545 * @start_pfn: The start PFN to start searching for holes
6546 * @end_pfn: The end PFN to stop searching for holes
6548 * Return: the number of pages frames in memory holes within a range.
6550 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6551 unsigned long end_pfn)
6553 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6556 /* Return the number of page frames in holes in a zone on a node */
6557 static unsigned long __init zone_absent_pages_in_node(int nid,
6558 unsigned long zone_type,
6559 unsigned long node_start_pfn,
6560 unsigned long node_end_pfn)
6562 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6563 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6564 unsigned long zone_start_pfn, zone_end_pfn;
6565 unsigned long nr_absent;
6567 /* When hotadd a new node from cpu_up(), the node should be empty */
6568 if (!node_start_pfn && !node_end_pfn)
6571 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6572 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6574 adjust_zone_range_for_zone_movable(nid, zone_type,
6575 node_start_pfn, node_end_pfn,
6576 &zone_start_pfn, &zone_end_pfn);
6577 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6580 * ZONE_MOVABLE handling.
6581 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6584 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6585 unsigned long start_pfn, end_pfn;
6586 struct memblock_region *r;
6588 for_each_mem_region(r) {
6589 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6590 zone_start_pfn, zone_end_pfn);
6591 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6592 zone_start_pfn, zone_end_pfn);
6594 if (zone_type == ZONE_MOVABLE &&
6595 memblock_is_mirror(r))
6596 nr_absent += end_pfn - start_pfn;
6598 if (zone_type == ZONE_NORMAL &&
6599 !memblock_is_mirror(r))
6600 nr_absent += end_pfn - start_pfn;
6607 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6608 unsigned long node_start_pfn,
6609 unsigned long node_end_pfn)
6611 unsigned long realtotalpages = 0, totalpages = 0;
6614 for (i = 0; i < MAX_NR_ZONES; i++) {
6615 struct zone *zone = pgdat->node_zones + i;
6616 unsigned long zone_start_pfn, zone_end_pfn;
6617 unsigned long spanned, absent;
6618 unsigned long size, real_size;
6620 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
6625 absent = zone_absent_pages_in_node(pgdat->node_id, i,
6630 real_size = size - absent;
6633 zone->zone_start_pfn = zone_start_pfn;
6635 zone->zone_start_pfn = 0;
6636 zone->spanned_pages = size;
6637 zone->present_pages = real_size;
6640 realtotalpages += real_size;
6643 pgdat->node_spanned_pages = totalpages;
6644 pgdat->node_present_pages = realtotalpages;
6645 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6649 #ifndef CONFIG_SPARSEMEM
6651 * Calculate the size of the zone->blockflags rounded to an unsigned long
6652 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6653 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6654 * round what is now in bits to nearest long in bits, then return it in
6657 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6659 unsigned long usemapsize;
6661 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6662 usemapsize = roundup(zonesize, pageblock_nr_pages);
6663 usemapsize = usemapsize >> pageblock_order;
6664 usemapsize *= NR_PAGEBLOCK_BITS;
6665 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6667 return usemapsize / 8;
6670 static void __ref setup_usemap(struct pglist_data *pgdat,
6672 unsigned long zone_start_pfn,
6673 unsigned long zonesize)
6675 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6676 zone->pageblock_flags = NULL;
6678 zone->pageblock_flags =
6679 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6681 if (!zone->pageblock_flags)
6682 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6683 usemapsize, zone->name, pgdat->node_id);
6687 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6688 unsigned long zone_start_pfn, unsigned long zonesize) {}
6689 #endif /* CONFIG_SPARSEMEM */
6691 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6693 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6694 void __init set_pageblock_order(void)
6698 /* Check that pageblock_nr_pages has not already been setup */
6699 if (pageblock_order)
6702 if (HPAGE_SHIFT > PAGE_SHIFT)
6703 order = HUGETLB_PAGE_ORDER;
6705 order = MAX_ORDER - 1;
6708 * Assume the largest contiguous order of interest is a huge page.
6709 * This value may be variable depending on boot parameters on IA64 and
6712 pageblock_order = order;
6714 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6717 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6718 * is unused as pageblock_order is set at compile-time. See
6719 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6722 void __init set_pageblock_order(void)
6726 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6728 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6729 unsigned long present_pages)
6731 unsigned long pages = spanned_pages;
6734 * Provide a more accurate estimation if there are holes within
6735 * the zone and SPARSEMEM is in use. If there are holes within the
6736 * zone, each populated memory region may cost us one or two extra
6737 * memmap pages due to alignment because memmap pages for each
6738 * populated regions may not be naturally aligned on page boundary.
6739 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6741 if (spanned_pages > present_pages + (present_pages >> 4) &&
6742 IS_ENABLED(CONFIG_SPARSEMEM))
6743 pages = present_pages;
6745 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6748 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6749 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6751 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6753 spin_lock_init(&ds_queue->split_queue_lock);
6754 INIT_LIST_HEAD(&ds_queue->split_queue);
6755 ds_queue->split_queue_len = 0;
6758 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6761 #ifdef CONFIG_COMPACTION
6762 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6764 init_waitqueue_head(&pgdat->kcompactd_wait);
6767 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6770 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6772 pgdat_resize_init(pgdat);
6774 pgdat_init_split_queue(pgdat);
6775 pgdat_init_kcompactd(pgdat);
6777 init_waitqueue_head(&pgdat->kswapd_wait);
6778 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6780 pgdat_page_ext_init(pgdat);
6781 spin_lock_init(&pgdat->lru_lock);
6782 lruvec_init(&pgdat->__lruvec);
6785 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6786 unsigned long remaining_pages)
6788 atomic_long_set(&zone->managed_pages, remaining_pages);
6789 zone_set_nid(zone, nid);
6790 zone->name = zone_names[idx];
6791 zone->zone_pgdat = NODE_DATA(nid);
6792 spin_lock_init(&zone->lock);
6793 zone_seqlock_init(zone);
6794 zone_pcp_init(zone);
6798 * Set up the zone data structures
6799 * - init pgdat internals
6800 * - init all zones belonging to this node
6802 * NOTE: this function is only called during memory hotplug
6804 #ifdef CONFIG_MEMORY_HOTPLUG
6805 void __ref free_area_init_core_hotplug(int nid)
6808 pg_data_t *pgdat = NODE_DATA(nid);
6810 pgdat_init_internals(pgdat);
6811 for (z = 0; z < MAX_NR_ZONES; z++)
6812 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6817 * Set up the zone data structures:
6818 * - mark all pages reserved
6819 * - mark all memory queues empty
6820 * - clear the memory bitmaps
6822 * NOTE: pgdat should get zeroed by caller.
6823 * NOTE: this function is only called during early init.
6825 static void __init free_area_init_core(struct pglist_data *pgdat)
6828 int nid = pgdat->node_id;
6830 pgdat_init_internals(pgdat);
6831 pgdat->per_cpu_nodestats = &boot_nodestats;
6833 for (j = 0; j < MAX_NR_ZONES; j++) {
6834 struct zone *zone = pgdat->node_zones + j;
6835 unsigned long size, freesize, memmap_pages;
6836 unsigned long zone_start_pfn = zone->zone_start_pfn;
6838 size = zone->spanned_pages;
6839 freesize = zone->present_pages;
6842 * Adjust freesize so that it accounts for how much memory
6843 * is used by this zone for memmap. This affects the watermark
6844 * and per-cpu initialisations
6846 memmap_pages = calc_memmap_size(size, freesize);
6847 if (!is_highmem_idx(j)) {
6848 if (freesize >= memmap_pages) {
6849 freesize -= memmap_pages;
6852 " %s zone: %lu pages used for memmap\n",
6853 zone_names[j], memmap_pages);
6855 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6856 zone_names[j], memmap_pages, freesize);
6859 /* Account for reserved pages */
6860 if (j == 0 && freesize > dma_reserve) {
6861 freesize -= dma_reserve;
6862 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6863 zone_names[0], dma_reserve);
6866 if (!is_highmem_idx(j))
6867 nr_kernel_pages += freesize;
6868 /* Charge for highmem memmap if there are enough kernel pages */
6869 else if (nr_kernel_pages > memmap_pages * 2)
6870 nr_kernel_pages -= memmap_pages;
6871 nr_all_pages += freesize;
6874 * Set an approximate value for lowmem here, it will be adjusted
6875 * when the bootmem allocator frees pages into the buddy system.
6876 * And all highmem pages will be managed by the buddy system.
6878 zone_init_internals(zone, j, nid, freesize);
6883 set_pageblock_order();
6884 setup_usemap(pgdat, zone, zone_start_pfn, size);
6885 init_currently_empty_zone(zone, zone_start_pfn, size);
6886 memmap_init(size, nid, j, zone_start_pfn);
6890 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6891 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6893 unsigned long __maybe_unused start = 0;
6894 unsigned long __maybe_unused offset = 0;
6896 /* Skip empty nodes */
6897 if (!pgdat->node_spanned_pages)
6900 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6901 offset = pgdat->node_start_pfn - start;
6902 /* ia64 gets its own node_mem_map, before this, without bootmem */
6903 if (!pgdat->node_mem_map) {
6904 unsigned long size, end;
6908 * The zone's endpoints aren't required to be MAX_ORDER
6909 * aligned but the node_mem_map endpoints must be in order
6910 * for the buddy allocator to function correctly.
6912 end = pgdat_end_pfn(pgdat);
6913 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6914 size = (end - start) * sizeof(struct page);
6915 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6918 panic("Failed to allocate %ld bytes for node %d memory map\n",
6919 size, pgdat->node_id);
6920 pgdat->node_mem_map = map + offset;
6922 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6923 __func__, pgdat->node_id, (unsigned long)pgdat,
6924 (unsigned long)pgdat->node_mem_map);
6925 #ifndef CONFIG_NEED_MULTIPLE_NODES
6927 * With no DISCONTIG, the global mem_map is just set as node 0's
6929 if (pgdat == NODE_DATA(0)) {
6930 mem_map = NODE_DATA(0)->node_mem_map;
6931 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6937 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6938 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6940 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6941 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6943 pgdat->first_deferred_pfn = ULONG_MAX;
6946 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6949 static void __init free_area_init_node(int nid)
6951 pg_data_t *pgdat = NODE_DATA(nid);
6952 unsigned long start_pfn = 0;
6953 unsigned long end_pfn = 0;
6955 /* pg_data_t should be reset to zero when it's allocated */
6956 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
6958 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6960 pgdat->node_id = nid;
6961 pgdat->node_start_pfn = start_pfn;
6962 pgdat->per_cpu_nodestats = NULL;
6964 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6965 (u64)start_pfn << PAGE_SHIFT,
6966 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6967 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
6969 alloc_node_mem_map(pgdat);
6970 pgdat_set_deferred_range(pgdat);
6972 free_area_init_core(pgdat);
6975 void __init free_area_init_memoryless_node(int nid)
6977 free_area_init_node(nid);
6980 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6982 * Initialize all valid struct pages in the range [spfn, epfn) and mark them
6983 * PageReserved(). Return the number of struct pages that were initialized.
6985 static u64 __init init_unavailable_range(unsigned long spfn, unsigned long epfn)
6990 for (pfn = spfn; pfn < epfn; pfn++) {
6991 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6992 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6993 + pageblock_nr_pages - 1;
6997 * Use a fake node/zone (0) for now. Some of these pages
6998 * (in memblock.reserved but not in memblock.memory) will
6999 * get re-initialized via reserve_bootmem_region() later.
7001 __init_single_page(pfn_to_page(pfn), pfn, 0, 0);
7002 __SetPageReserved(pfn_to_page(pfn));
7010 * Only struct pages that are backed by physical memory are zeroed and
7011 * initialized by going through __init_single_page(). But, there are some
7012 * struct pages which are reserved in memblock allocator and their fields
7013 * may be accessed (for example page_to_pfn() on some configuration accesses
7014 * flags). We must explicitly initialize those struct pages.
7016 * This function also addresses a similar issue where struct pages are left
7017 * uninitialized because the physical address range is not covered by
7018 * memblock.memory or memblock.reserved. That could happen when memblock
7019 * layout is manually configured via memmap=, or when the highest physical
7020 * address (max_pfn) does not end on a section boundary.
7022 static void __init init_unavailable_mem(void)
7024 phys_addr_t start, end;
7026 phys_addr_t next = 0;
7029 * Loop through unavailable ranges not covered by memblock.memory.
7032 for_each_mem_range(i, &start, &end) {
7034 pgcnt += init_unavailable_range(PFN_DOWN(next),
7040 * Early sections always have a fully populated memmap for the whole
7041 * section - see pfn_valid(). If the last section has holes at the
7042 * end and that section is marked "online", the memmap will be
7043 * considered initialized. Make sure that memmap has a well defined
7046 pgcnt += init_unavailable_range(PFN_DOWN(next),
7047 round_up(max_pfn, PAGES_PER_SECTION));
7050 * Struct pages that do not have backing memory. This could be because
7051 * firmware is using some of this memory, or for some other reasons.
7054 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
7057 static inline void __init init_unavailable_mem(void)
7060 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
7062 #if MAX_NUMNODES > 1
7064 * Figure out the number of possible node ids.
7066 void __init setup_nr_node_ids(void)
7068 unsigned int highest;
7070 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7071 nr_node_ids = highest + 1;
7076 * node_map_pfn_alignment - determine the maximum internode alignment
7078 * This function should be called after node map is populated and sorted.
7079 * It calculates the maximum power of two alignment which can distinguish
7082 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7083 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7084 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7085 * shifted, 1GiB is enough and this function will indicate so.
7087 * This is used to test whether pfn -> nid mapping of the chosen memory
7088 * model has fine enough granularity to avoid incorrect mapping for the
7089 * populated node map.
7091 * Return: the determined alignment in pfn's. 0 if there is no alignment
7092 * requirement (single node).
7094 unsigned long __init node_map_pfn_alignment(void)
7096 unsigned long accl_mask = 0, last_end = 0;
7097 unsigned long start, end, mask;
7098 int last_nid = NUMA_NO_NODE;
7101 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7102 if (!start || last_nid < 0 || last_nid == nid) {
7109 * Start with a mask granular enough to pin-point to the
7110 * start pfn and tick off bits one-by-one until it becomes
7111 * too coarse to separate the current node from the last.
7113 mask = ~((1 << __ffs(start)) - 1);
7114 while (mask && last_end <= (start & (mask << 1)))
7117 /* accumulate all internode masks */
7121 /* convert mask to number of pages */
7122 return ~accl_mask + 1;
7126 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7128 * Return: the minimum PFN based on information provided via
7129 * memblock_set_node().
7131 unsigned long __init find_min_pfn_with_active_regions(void)
7133 return PHYS_PFN(memblock_start_of_DRAM());
7137 * early_calculate_totalpages()
7138 * Sum pages in active regions for movable zone.
7139 * Populate N_MEMORY for calculating usable_nodes.
7141 static unsigned long __init early_calculate_totalpages(void)
7143 unsigned long totalpages = 0;
7144 unsigned long start_pfn, end_pfn;
7147 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7148 unsigned long pages = end_pfn - start_pfn;
7150 totalpages += pages;
7152 node_set_state(nid, N_MEMORY);
7158 * Find the PFN the Movable zone begins in each node. Kernel memory
7159 * is spread evenly between nodes as long as the nodes have enough
7160 * memory. When they don't, some nodes will have more kernelcore than
7163 static void __init find_zone_movable_pfns_for_nodes(void)
7166 unsigned long usable_startpfn;
7167 unsigned long kernelcore_node, kernelcore_remaining;
7168 /* save the state before borrow the nodemask */
7169 nodemask_t saved_node_state = node_states[N_MEMORY];
7170 unsigned long totalpages = early_calculate_totalpages();
7171 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7172 struct memblock_region *r;
7174 /* Need to find movable_zone earlier when movable_node is specified. */
7175 find_usable_zone_for_movable();
7178 * If movable_node is specified, ignore kernelcore and movablecore
7181 if (movable_node_is_enabled()) {
7182 for_each_mem_region(r) {
7183 if (!memblock_is_hotpluggable(r))
7186 nid = memblock_get_region_node(r);
7188 usable_startpfn = PFN_DOWN(r->base);
7189 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7190 min(usable_startpfn, zone_movable_pfn[nid]) :
7198 * If kernelcore=mirror is specified, ignore movablecore option
7200 if (mirrored_kernelcore) {
7201 bool mem_below_4gb_not_mirrored = false;
7203 for_each_mem_region(r) {
7204 if (memblock_is_mirror(r))
7207 nid = memblock_get_region_node(r);
7209 usable_startpfn = memblock_region_memory_base_pfn(r);
7211 if (usable_startpfn < 0x100000) {
7212 mem_below_4gb_not_mirrored = true;
7216 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7217 min(usable_startpfn, zone_movable_pfn[nid]) :
7221 if (mem_below_4gb_not_mirrored)
7222 pr_warn("This configuration results in unmirrored kernel memory.\n");
7228 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7229 * amount of necessary memory.
7231 if (required_kernelcore_percent)
7232 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7234 if (required_movablecore_percent)
7235 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7239 * If movablecore= was specified, calculate what size of
7240 * kernelcore that corresponds so that memory usable for
7241 * any allocation type is evenly spread. If both kernelcore
7242 * and movablecore are specified, then the value of kernelcore
7243 * will be used for required_kernelcore if it's greater than
7244 * what movablecore would have allowed.
7246 if (required_movablecore) {
7247 unsigned long corepages;
7250 * Round-up so that ZONE_MOVABLE is at least as large as what
7251 * was requested by the user
7253 required_movablecore =
7254 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7255 required_movablecore = min(totalpages, required_movablecore);
7256 corepages = totalpages - required_movablecore;
7258 required_kernelcore = max(required_kernelcore, corepages);
7262 * If kernelcore was not specified or kernelcore size is larger
7263 * than totalpages, there is no ZONE_MOVABLE.
7265 if (!required_kernelcore || required_kernelcore >= totalpages)
7268 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7269 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7272 /* Spread kernelcore memory as evenly as possible throughout nodes */
7273 kernelcore_node = required_kernelcore / usable_nodes;
7274 for_each_node_state(nid, N_MEMORY) {
7275 unsigned long start_pfn, end_pfn;
7278 * Recalculate kernelcore_node if the division per node
7279 * now exceeds what is necessary to satisfy the requested
7280 * amount of memory for the kernel
7282 if (required_kernelcore < kernelcore_node)
7283 kernelcore_node = required_kernelcore / usable_nodes;
7286 * As the map is walked, we track how much memory is usable
7287 * by the kernel using kernelcore_remaining. When it is
7288 * 0, the rest of the node is usable by ZONE_MOVABLE
7290 kernelcore_remaining = kernelcore_node;
7292 /* Go through each range of PFNs within this node */
7293 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7294 unsigned long size_pages;
7296 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7297 if (start_pfn >= end_pfn)
7300 /* Account for what is only usable for kernelcore */
7301 if (start_pfn < usable_startpfn) {
7302 unsigned long kernel_pages;
7303 kernel_pages = min(end_pfn, usable_startpfn)
7306 kernelcore_remaining -= min(kernel_pages,
7307 kernelcore_remaining);
7308 required_kernelcore -= min(kernel_pages,
7309 required_kernelcore);
7311 /* Continue if range is now fully accounted */
7312 if (end_pfn <= usable_startpfn) {
7315 * Push zone_movable_pfn to the end so
7316 * that if we have to rebalance
7317 * kernelcore across nodes, we will
7318 * not double account here
7320 zone_movable_pfn[nid] = end_pfn;
7323 start_pfn = usable_startpfn;
7327 * The usable PFN range for ZONE_MOVABLE is from
7328 * start_pfn->end_pfn. Calculate size_pages as the
7329 * number of pages used as kernelcore
7331 size_pages = end_pfn - start_pfn;
7332 if (size_pages > kernelcore_remaining)
7333 size_pages = kernelcore_remaining;
7334 zone_movable_pfn[nid] = start_pfn + size_pages;
7337 * Some kernelcore has been met, update counts and
7338 * break if the kernelcore for this node has been
7341 required_kernelcore -= min(required_kernelcore,
7343 kernelcore_remaining -= size_pages;
7344 if (!kernelcore_remaining)
7350 * If there is still required_kernelcore, we do another pass with one
7351 * less node in the count. This will push zone_movable_pfn[nid] further
7352 * along on the nodes that still have memory until kernelcore is
7356 if (usable_nodes && required_kernelcore > usable_nodes)
7360 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7361 for (nid = 0; nid < MAX_NUMNODES; nid++)
7362 zone_movable_pfn[nid] =
7363 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7366 /* restore the node_state */
7367 node_states[N_MEMORY] = saved_node_state;
7370 /* Any regular or high memory on that node ? */
7371 static void check_for_memory(pg_data_t *pgdat, int nid)
7373 enum zone_type zone_type;
7375 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7376 struct zone *zone = &pgdat->node_zones[zone_type];
7377 if (populated_zone(zone)) {
7378 if (IS_ENABLED(CONFIG_HIGHMEM))
7379 node_set_state(nid, N_HIGH_MEMORY);
7380 if (zone_type <= ZONE_NORMAL)
7381 node_set_state(nid, N_NORMAL_MEMORY);
7388 * Some architecturs, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7389 * such cases we allow max_zone_pfn sorted in the descending order
7391 bool __weak arch_has_descending_max_zone_pfns(void)
7397 * free_area_init - Initialise all pg_data_t and zone data
7398 * @max_zone_pfn: an array of max PFNs for each zone
7400 * This will call free_area_init_node() for each active node in the system.
7401 * Using the page ranges provided by memblock_set_node(), the size of each
7402 * zone in each node and their holes is calculated. If the maximum PFN
7403 * between two adjacent zones match, it is assumed that the zone is empty.
7404 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7405 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7406 * starts where the previous one ended. For example, ZONE_DMA32 starts
7407 * at arch_max_dma_pfn.
7409 void __init free_area_init(unsigned long *max_zone_pfn)
7411 unsigned long start_pfn, end_pfn;
7415 /* Record where the zone boundaries are */
7416 memset(arch_zone_lowest_possible_pfn, 0,
7417 sizeof(arch_zone_lowest_possible_pfn));
7418 memset(arch_zone_highest_possible_pfn, 0,
7419 sizeof(arch_zone_highest_possible_pfn));
7421 start_pfn = find_min_pfn_with_active_regions();
7422 descending = arch_has_descending_max_zone_pfns();
7424 for (i = 0; i < MAX_NR_ZONES; i++) {
7426 zone = MAX_NR_ZONES - i - 1;
7430 if (zone == ZONE_MOVABLE)
7433 end_pfn = max(max_zone_pfn[zone], start_pfn);
7434 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7435 arch_zone_highest_possible_pfn[zone] = end_pfn;
7437 start_pfn = end_pfn;
7440 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7441 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7442 find_zone_movable_pfns_for_nodes();
7444 /* Print out the zone ranges */
7445 pr_info("Zone ranges:\n");
7446 for (i = 0; i < MAX_NR_ZONES; i++) {
7447 if (i == ZONE_MOVABLE)
7449 pr_info(" %-8s ", zone_names[i]);
7450 if (arch_zone_lowest_possible_pfn[i] ==
7451 arch_zone_highest_possible_pfn[i])
7454 pr_cont("[mem %#018Lx-%#018Lx]\n",
7455 (u64)arch_zone_lowest_possible_pfn[i]
7457 ((u64)arch_zone_highest_possible_pfn[i]
7458 << PAGE_SHIFT) - 1);
7461 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7462 pr_info("Movable zone start for each node\n");
7463 for (i = 0; i < MAX_NUMNODES; i++) {
7464 if (zone_movable_pfn[i])
7465 pr_info(" Node %d: %#018Lx\n", i,
7466 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7470 * Print out the early node map, and initialize the
7471 * subsection-map relative to active online memory ranges to
7472 * enable future "sub-section" extensions of the memory map.
7474 pr_info("Early memory node ranges\n");
7475 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7476 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7477 (u64)start_pfn << PAGE_SHIFT,
7478 ((u64)end_pfn << PAGE_SHIFT) - 1);
7479 subsection_map_init(start_pfn, end_pfn - start_pfn);
7482 /* Initialise every node */
7483 mminit_verify_pageflags_layout();
7484 setup_nr_node_ids();
7485 init_unavailable_mem();
7486 for_each_online_node(nid) {
7487 pg_data_t *pgdat = NODE_DATA(nid);
7488 free_area_init_node(nid);
7490 /* Any memory on that node */
7491 if (pgdat->node_present_pages)
7492 node_set_state(nid, N_MEMORY);
7493 check_for_memory(pgdat, nid);
7497 static int __init cmdline_parse_core(char *p, unsigned long *core,
7498 unsigned long *percent)
7500 unsigned long long coremem;
7506 /* Value may be a percentage of total memory, otherwise bytes */
7507 coremem = simple_strtoull(p, &endptr, 0);
7508 if (*endptr == '%') {
7509 /* Paranoid check for percent values greater than 100 */
7510 WARN_ON(coremem > 100);
7514 coremem = memparse(p, &p);
7515 /* Paranoid check that UL is enough for the coremem value */
7516 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7518 *core = coremem >> PAGE_SHIFT;
7525 * kernelcore=size sets the amount of memory for use for allocations that
7526 * cannot be reclaimed or migrated.
7528 static int __init cmdline_parse_kernelcore(char *p)
7530 /* parse kernelcore=mirror */
7531 if (parse_option_str(p, "mirror")) {
7532 mirrored_kernelcore = true;
7536 return cmdline_parse_core(p, &required_kernelcore,
7537 &required_kernelcore_percent);
7541 * movablecore=size sets the amount of memory for use for allocations that
7542 * can be reclaimed or migrated.
7544 static int __init cmdline_parse_movablecore(char *p)
7546 return cmdline_parse_core(p, &required_movablecore,
7547 &required_movablecore_percent);
7550 early_param("kernelcore", cmdline_parse_kernelcore);
7551 early_param("movablecore", cmdline_parse_movablecore);
7553 void adjust_managed_page_count(struct page *page, long count)
7555 atomic_long_add(count, &page_zone(page)->managed_pages);
7556 totalram_pages_add(count);
7557 #ifdef CONFIG_HIGHMEM
7558 if (PageHighMem(page))
7559 totalhigh_pages_add(count);
7562 EXPORT_SYMBOL(adjust_managed_page_count);
7564 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7567 unsigned long pages = 0;
7569 start = (void *)PAGE_ALIGN((unsigned long)start);
7570 end = (void *)((unsigned long)end & PAGE_MASK);
7571 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7572 struct page *page = virt_to_page(pos);
7573 void *direct_map_addr;
7576 * 'direct_map_addr' might be different from 'pos'
7577 * because some architectures' virt_to_page()
7578 * work with aliases. Getting the direct map
7579 * address ensures that we get a _writeable_
7580 * alias for the memset().
7582 direct_map_addr = page_address(page);
7583 if ((unsigned int)poison <= 0xFF)
7584 memset(direct_map_addr, poison, PAGE_SIZE);
7586 free_reserved_page(page);
7590 pr_info("Freeing %s memory: %ldK\n",
7591 s, pages << (PAGE_SHIFT - 10));
7596 #ifdef CONFIG_HIGHMEM
7597 void free_highmem_page(struct page *page)
7599 __free_reserved_page(page);
7600 totalram_pages_inc();
7601 atomic_long_inc(&page_zone(page)->managed_pages);
7602 totalhigh_pages_inc();
7607 void __init mem_init_print_info(const char *str)
7609 unsigned long physpages, codesize, datasize, rosize, bss_size;
7610 unsigned long init_code_size, init_data_size;
7612 physpages = get_num_physpages();
7613 codesize = _etext - _stext;
7614 datasize = _edata - _sdata;
7615 rosize = __end_rodata - __start_rodata;
7616 bss_size = __bss_stop - __bss_start;
7617 init_data_size = __init_end - __init_begin;
7618 init_code_size = _einittext - _sinittext;
7621 * Detect special cases and adjust section sizes accordingly:
7622 * 1) .init.* may be embedded into .data sections
7623 * 2) .init.text.* may be out of [__init_begin, __init_end],
7624 * please refer to arch/tile/kernel/vmlinux.lds.S.
7625 * 3) .rodata.* may be embedded into .text or .data sections.
7627 #define adj_init_size(start, end, size, pos, adj) \
7629 if (start <= pos && pos < end && size > adj) \
7633 adj_init_size(__init_begin, __init_end, init_data_size,
7634 _sinittext, init_code_size);
7635 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7636 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7637 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7638 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7640 #undef adj_init_size
7642 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7643 #ifdef CONFIG_HIGHMEM
7647 nr_free_pages() << (PAGE_SHIFT - 10),
7648 physpages << (PAGE_SHIFT - 10),
7649 codesize >> 10, datasize >> 10, rosize >> 10,
7650 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7651 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7652 totalcma_pages << (PAGE_SHIFT - 10),
7653 #ifdef CONFIG_HIGHMEM
7654 totalhigh_pages() << (PAGE_SHIFT - 10),
7656 str ? ", " : "", str ? str : "");
7660 * set_dma_reserve - set the specified number of pages reserved in the first zone
7661 * @new_dma_reserve: The number of pages to mark reserved
7663 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7664 * In the DMA zone, a significant percentage may be consumed by kernel image
7665 * and other unfreeable allocations which can skew the watermarks badly. This
7666 * function may optionally be used to account for unfreeable pages in the
7667 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7668 * smaller per-cpu batchsize.
7670 void __init set_dma_reserve(unsigned long new_dma_reserve)
7672 dma_reserve = new_dma_reserve;
7675 static int page_alloc_cpu_dead(unsigned int cpu)
7678 lru_add_drain_cpu(cpu);
7682 * Spill the event counters of the dead processor
7683 * into the current processors event counters.
7684 * This artificially elevates the count of the current
7687 vm_events_fold_cpu(cpu);
7690 * Zero the differential counters of the dead processor
7691 * so that the vm statistics are consistent.
7693 * This is only okay since the processor is dead and cannot
7694 * race with what we are doing.
7696 cpu_vm_stats_fold(cpu);
7701 int hashdist = HASHDIST_DEFAULT;
7703 static int __init set_hashdist(char *str)
7707 hashdist = simple_strtoul(str, &str, 0);
7710 __setup("hashdist=", set_hashdist);
7713 void __init page_alloc_init(void)
7718 if (num_node_state(N_MEMORY) == 1)
7722 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7723 "mm/page_alloc:dead", NULL,
7724 page_alloc_cpu_dead);
7729 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7730 * or min_free_kbytes changes.
7732 static void calculate_totalreserve_pages(void)
7734 struct pglist_data *pgdat;
7735 unsigned long reserve_pages = 0;
7736 enum zone_type i, j;
7738 for_each_online_pgdat(pgdat) {
7740 pgdat->totalreserve_pages = 0;
7742 for (i = 0; i < MAX_NR_ZONES; i++) {
7743 struct zone *zone = pgdat->node_zones + i;
7745 unsigned long managed_pages = zone_managed_pages(zone);
7747 /* Find valid and maximum lowmem_reserve in the zone */
7748 for (j = i; j < MAX_NR_ZONES; j++) {
7749 if (zone->lowmem_reserve[j] > max)
7750 max = zone->lowmem_reserve[j];
7753 /* we treat the high watermark as reserved pages. */
7754 max += high_wmark_pages(zone);
7756 if (max > managed_pages)
7757 max = managed_pages;
7759 pgdat->totalreserve_pages += max;
7761 reserve_pages += max;
7764 totalreserve_pages = reserve_pages;
7768 * setup_per_zone_lowmem_reserve - called whenever
7769 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7770 * has a correct pages reserved value, so an adequate number of
7771 * pages are left in the zone after a successful __alloc_pages().
7773 static void setup_per_zone_lowmem_reserve(void)
7775 struct pglist_data *pgdat;
7776 enum zone_type j, idx;
7778 for_each_online_pgdat(pgdat) {
7779 for (j = 0; j < MAX_NR_ZONES; j++) {
7780 struct zone *zone = pgdat->node_zones + j;
7781 unsigned long managed_pages = zone_managed_pages(zone);
7783 zone->lowmem_reserve[j] = 0;
7787 struct zone *lower_zone;
7790 lower_zone = pgdat->node_zones + idx;
7792 if (!sysctl_lowmem_reserve_ratio[idx] ||
7793 !zone_managed_pages(lower_zone)) {
7794 lower_zone->lowmem_reserve[j] = 0;
7797 lower_zone->lowmem_reserve[j] =
7798 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7800 managed_pages += zone_managed_pages(lower_zone);
7805 /* update totalreserve_pages */
7806 calculate_totalreserve_pages();
7809 static void __setup_per_zone_wmarks(void)
7811 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7812 unsigned long lowmem_pages = 0;
7814 unsigned long flags;
7816 /* Calculate total number of !ZONE_HIGHMEM pages */
7817 for_each_zone(zone) {
7818 if (!is_highmem(zone))
7819 lowmem_pages += zone_managed_pages(zone);
7822 for_each_zone(zone) {
7825 spin_lock_irqsave(&zone->lock, flags);
7826 tmp = (u64)pages_min * zone_managed_pages(zone);
7827 do_div(tmp, lowmem_pages);
7828 if (is_highmem(zone)) {
7830 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7831 * need highmem pages, so cap pages_min to a small
7834 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7835 * deltas control async page reclaim, and so should
7836 * not be capped for highmem.
7838 unsigned long min_pages;
7840 min_pages = zone_managed_pages(zone) / 1024;
7841 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7842 zone->_watermark[WMARK_MIN] = min_pages;
7845 * If it's a lowmem zone, reserve a number of pages
7846 * proportionate to the zone's size.
7848 zone->_watermark[WMARK_MIN] = tmp;
7852 * Set the kswapd watermarks distance according to the
7853 * scale factor in proportion to available memory, but
7854 * ensure a minimum size on small systems.
7856 tmp = max_t(u64, tmp >> 2,
7857 mult_frac(zone_managed_pages(zone),
7858 watermark_scale_factor, 10000));
7860 zone->watermark_boost = 0;
7861 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7862 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7864 spin_unlock_irqrestore(&zone->lock, flags);
7867 /* update totalreserve_pages */
7868 calculate_totalreserve_pages();
7872 * setup_per_zone_wmarks - called when min_free_kbytes changes
7873 * or when memory is hot-{added|removed}
7875 * Ensures that the watermark[min,low,high] values for each zone are set
7876 * correctly with respect to min_free_kbytes.
7878 void setup_per_zone_wmarks(void)
7880 static DEFINE_SPINLOCK(lock);
7883 __setup_per_zone_wmarks();
7888 * Initialise min_free_kbytes.
7890 * For small machines we want it small (128k min). For large machines
7891 * we want it large (256MB max). But it is not linear, because network
7892 * bandwidth does not increase linearly with machine size. We use
7894 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7895 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7911 int __meminit init_per_zone_wmark_min(void)
7913 unsigned long lowmem_kbytes;
7914 int new_min_free_kbytes;
7916 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7917 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7919 if (new_min_free_kbytes > user_min_free_kbytes) {
7920 min_free_kbytes = new_min_free_kbytes;
7921 if (min_free_kbytes < 128)
7922 min_free_kbytes = 128;
7923 if (min_free_kbytes > 262144)
7924 min_free_kbytes = 262144;
7926 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7927 new_min_free_kbytes, user_min_free_kbytes);
7929 setup_per_zone_wmarks();
7930 refresh_zone_stat_thresholds();
7931 setup_per_zone_lowmem_reserve();
7934 setup_min_unmapped_ratio();
7935 setup_min_slab_ratio();
7938 khugepaged_min_free_kbytes_update();
7942 postcore_initcall(init_per_zone_wmark_min)
7945 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7946 * that we can call two helper functions whenever min_free_kbytes
7949 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7950 void *buffer, size_t *length, loff_t *ppos)
7954 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7959 user_min_free_kbytes = min_free_kbytes;
7960 setup_per_zone_wmarks();
7965 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7966 void *buffer, size_t *length, loff_t *ppos)
7970 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7975 setup_per_zone_wmarks();
7981 static void setup_min_unmapped_ratio(void)
7986 for_each_online_pgdat(pgdat)
7987 pgdat->min_unmapped_pages = 0;
7990 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7991 sysctl_min_unmapped_ratio) / 100;
7995 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7996 void *buffer, size_t *length, loff_t *ppos)
8000 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8004 setup_min_unmapped_ratio();
8009 static void setup_min_slab_ratio(void)
8014 for_each_online_pgdat(pgdat)
8015 pgdat->min_slab_pages = 0;
8018 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8019 sysctl_min_slab_ratio) / 100;
8022 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8023 void *buffer, size_t *length, loff_t *ppos)
8027 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8031 setup_min_slab_ratio();
8038 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8039 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8040 * whenever sysctl_lowmem_reserve_ratio changes.
8042 * The reserve ratio obviously has absolutely no relation with the
8043 * minimum watermarks. The lowmem reserve ratio can only make sense
8044 * if in function of the boot time zone sizes.
8046 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8047 void *buffer, size_t *length, loff_t *ppos)
8051 proc_dointvec_minmax(table, write, buffer, length, ppos);
8053 for (i = 0; i < MAX_NR_ZONES; i++) {
8054 if (sysctl_lowmem_reserve_ratio[i] < 1)
8055 sysctl_lowmem_reserve_ratio[i] = 0;
8058 setup_per_zone_lowmem_reserve();
8062 static void __zone_pcp_update(struct zone *zone)
8066 for_each_possible_cpu(cpu)
8067 pageset_set_high_and_batch(zone,
8068 per_cpu_ptr(zone->pageset, cpu));
8072 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8073 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8074 * pagelist can have before it gets flushed back to buddy allocator.
8076 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8077 void *buffer, size_t *length, loff_t *ppos)
8080 int old_percpu_pagelist_fraction;
8083 mutex_lock(&pcp_batch_high_lock);
8084 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8086 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8087 if (!write || ret < 0)
8090 /* Sanity checking to avoid pcp imbalance */
8091 if (percpu_pagelist_fraction &&
8092 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8093 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8099 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8102 for_each_populated_zone(zone)
8103 __zone_pcp_update(zone);
8105 mutex_unlock(&pcp_batch_high_lock);
8109 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8111 * Returns the number of pages that arch has reserved but
8112 * is not known to alloc_large_system_hash().
8114 static unsigned long __init arch_reserved_kernel_pages(void)
8121 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8122 * machines. As memory size is increased the scale is also increased but at
8123 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8124 * quadruples the scale is increased by one, which means the size of hash table
8125 * only doubles, instead of quadrupling as well.
8126 * Because 32-bit systems cannot have large physical memory, where this scaling
8127 * makes sense, it is disabled on such platforms.
8129 #if __BITS_PER_LONG > 32
8130 #define ADAPT_SCALE_BASE (64ul << 30)
8131 #define ADAPT_SCALE_SHIFT 2
8132 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8136 * allocate a large system hash table from bootmem
8137 * - it is assumed that the hash table must contain an exact power-of-2
8138 * quantity of entries
8139 * - limit is the number of hash buckets, not the total allocation size
8141 void *__init alloc_large_system_hash(const char *tablename,
8142 unsigned long bucketsize,
8143 unsigned long numentries,
8146 unsigned int *_hash_shift,
8147 unsigned int *_hash_mask,
8148 unsigned long low_limit,
8149 unsigned long high_limit)
8151 unsigned long long max = high_limit;
8152 unsigned long log2qty, size;
8157 /* allow the kernel cmdline to have a say */
8159 /* round applicable memory size up to nearest megabyte */
8160 numentries = nr_kernel_pages;
8161 numentries -= arch_reserved_kernel_pages();
8163 /* It isn't necessary when PAGE_SIZE >= 1MB */
8164 if (PAGE_SHIFT < 20)
8165 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8167 #if __BITS_PER_LONG > 32
8169 unsigned long adapt;
8171 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8172 adapt <<= ADAPT_SCALE_SHIFT)
8177 /* limit to 1 bucket per 2^scale bytes of low memory */
8178 if (scale > PAGE_SHIFT)
8179 numentries >>= (scale - PAGE_SHIFT);
8181 numentries <<= (PAGE_SHIFT - scale);
8183 /* Make sure we've got at least a 0-order allocation.. */
8184 if (unlikely(flags & HASH_SMALL)) {
8185 /* Makes no sense without HASH_EARLY */
8186 WARN_ON(!(flags & HASH_EARLY));
8187 if (!(numentries >> *_hash_shift)) {
8188 numentries = 1UL << *_hash_shift;
8189 BUG_ON(!numentries);
8191 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8192 numentries = PAGE_SIZE / bucketsize;
8194 numentries = roundup_pow_of_two(numentries);
8196 /* limit allocation size to 1/16 total memory by default */
8198 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8199 do_div(max, bucketsize);
8201 max = min(max, 0x80000000ULL);
8203 if (numentries < low_limit)
8204 numentries = low_limit;
8205 if (numentries > max)
8208 log2qty = ilog2(numentries);
8210 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8213 size = bucketsize << log2qty;
8214 if (flags & HASH_EARLY) {
8215 if (flags & HASH_ZERO)
8216 table = memblock_alloc(size, SMP_CACHE_BYTES);
8218 table = memblock_alloc_raw(size,
8220 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8221 table = __vmalloc(size, gfp_flags);
8225 * If bucketsize is not a power-of-two, we may free
8226 * some pages at the end of hash table which
8227 * alloc_pages_exact() automatically does
8229 table = alloc_pages_exact(size, gfp_flags);
8230 kmemleak_alloc(table, size, 1, gfp_flags);
8232 } while (!table && size > PAGE_SIZE && --log2qty);
8235 panic("Failed to allocate %s hash table\n", tablename);
8237 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8238 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8239 virt ? "vmalloc" : "linear");
8242 *_hash_shift = log2qty;
8244 *_hash_mask = (1 << log2qty) - 1;
8250 * This function checks whether pageblock includes unmovable pages or not.
8252 * PageLRU check without isolation or lru_lock could race so that
8253 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8254 * check without lock_page also may miss some movable non-lru pages at
8255 * race condition. So you can't expect this function should be exact.
8257 * Returns a page without holding a reference. If the caller wants to
8258 * dereference that page (e.g., dumping), it has to make sure that it
8259 * cannot get removed (e.g., via memory unplug) concurrently.
8262 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8263 int migratetype, int flags)
8265 unsigned long iter = 0;
8266 unsigned long pfn = page_to_pfn(page);
8267 unsigned long offset = pfn % pageblock_nr_pages;
8269 if (is_migrate_cma_page(page)) {
8271 * CMA allocations (alloc_contig_range) really need to mark
8272 * isolate CMA pageblocks even when they are not movable in fact
8273 * so consider them movable here.
8275 if (is_migrate_cma(migratetype))
8281 for (; iter < pageblock_nr_pages - offset; iter++) {
8282 if (!pfn_valid_within(pfn + iter))
8285 page = pfn_to_page(pfn + iter);
8288 * Both, bootmem allocations and memory holes are marked
8289 * PG_reserved and are unmovable. We can even have unmovable
8290 * allocations inside ZONE_MOVABLE, for example when
8291 * specifying "movablecore".
8293 if (PageReserved(page))
8297 * If the zone is movable and we have ruled out all reserved
8298 * pages then it should be reasonably safe to assume the rest
8301 if (zone_idx(zone) == ZONE_MOVABLE)
8305 * Hugepages are not in LRU lists, but they're movable.
8306 * THPs are on the LRU, but need to be counted as #small pages.
8307 * We need not scan over tail pages because we don't
8308 * handle each tail page individually in migration.
8310 if (PageHuge(page) || PageTransCompound(page)) {
8311 struct page *head = compound_head(page);
8312 unsigned int skip_pages;
8314 if (PageHuge(page)) {
8315 if (!hugepage_migration_supported(page_hstate(head)))
8317 } else if (!PageLRU(head) && !__PageMovable(head)) {
8321 skip_pages = compound_nr(head) - (page - head);
8322 iter += skip_pages - 1;
8327 * We can't use page_count without pin a page
8328 * because another CPU can free compound page.
8329 * This check already skips compound tails of THP
8330 * because their page->_refcount is zero at all time.
8332 if (!page_ref_count(page)) {
8333 if (PageBuddy(page))
8334 iter += (1 << page_order(page)) - 1;
8339 * The HWPoisoned page may be not in buddy system, and
8340 * page_count() is not 0.
8342 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8346 * We treat all PageOffline() pages as movable when offlining
8347 * to give drivers a chance to decrement their reference count
8348 * in MEM_GOING_OFFLINE in order to indicate that these pages
8349 * can be offlined as there are no direct references anymore.
8350 * For actually unmovable PageOffline() where the driver does
8351 * not support this, we will fail later when trying to actually
8352 * move these pages that still have a reference count > 0.
8353 * (false negatives in this function only)
8355 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8358 if (__PageMovable(page) || PageLRU(page))
8362 * If there are RECLAIMABLE pages, we need to check
8363 * it. But now, memory offline itself doesn't call
8364 * shrink_node_slabs() and it still to be fixed.
8371 #ifdef CONFIG_CONTIG_ALLOC
8372 static unsigned long pfn_max_align_down(unsigned long pfn)
8374 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8375 pageblock_nr_pages) - 1);
8378 static unsigned long pfn_max_align_up(unsigned long pfn)
8380 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8381 pageblock_nr_pages));
8384 /* [start, end) must belong to a single zone. */
8385 static int __alloc_contig_migrate_range(struct compact_control *cc,
8386 unsigned long start, unsigned long end)
8388 /* This function is based on compact_zone() from compaction.c. */
8389 unsigned int nr_reclaimed;
8390 unsigned long pfn = start;
8391 unsigned int tries = 0;
8393 struct migration_target_control mtc = {
8394 .nid = zone_to_nid(cc->zone),
8395 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8400 while (pfn < end || !list_empty(&cc->migratepages)) {
8401 if (fatal_signal_pending(current)) {
8406 if (list_empty(&cc->migratepages)) {
8407 cc->nr_migratepages = 0;
8408 pfn = isolate_migratepages_range(cc, pfn, end);
8414 } else if (++tries == 5) {
8415 ret = ret < 0 ? ret : -EBUSY;
8419 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8421 cc->nr_migratepages -= nr_reclaimed;
8423 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
8424 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE);
8427 putback_movable_pages(&cc->migratepages);
8434 * alloc_contig_range() -- tries to allocate given range of pages
8435 * @start: start PFN to allocate
8436 * @end: one-past-the-last PFN to allocate
8437 * @migratetype: migratetype of the underlaying pageblocks (either
8438 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8439 * in range must have the same migratetype and it must
8440 * be either of the two.
8441 * @gfp_mask: GFP mask to use during compaction
8443 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8444 * aligned. The PFN range must belong to a single zone.
8446 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8447 * pageblocks in the range. Once isolated, the pageblocks should not
8448 * be modified by others.
8450 * Return: zero on success or negative error code. On success all
8451 * pages which PFN is in [start, end) are allocated for the caller and
8452 * need to be freed with free_contig_range().
8454 int alloc_contig_range(unsigned long start, unsigned long end,
8455 unsigned migratetype, gfp_t gfp_mask)
8457 unsigned long outer_start, outer_end;
8461 struct compact_control cc = {
8462 .nr_migratepages = 0,
8464 .zone = page_zone(pfn_to_page(start)),
8465 .mode = MIGRATE_SYNC,
8466 .ignore_skip_hint = true,
8467 .no_set_skip_hint = true,
8468 .gfp_mask = current_gfp_context(gfp_mask),
8469 .alloc_contig = true,
8471 INIT_LIST_HEAD(&cc.migratepages);
8474 * What we do here is we mark all pageblocks in range as
8475 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8476 * have different sizes, and due to the way page allocator
8477 * work, we align the range to biggest of the two pages so
8478 * that page allocator won't try to merge buddies from
8479 * different pageblocks and change MIGRATE_ISOLATE to some
8480 * other migration type.
8482 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8483 * migrate the pages from an unaligned range (ie. pages that
8484 * we are interested in). This will put all the pages in
8485 * range back to page allocator as MIGRATE_ISOLATE.
8487 * When this is done, we take the pages in range from page
8488 * allocator removing them from the buddy system. This way
8489 * page allocator will never consider using them.
8491 * This lets us mark the pageblocks back as
8492 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8493 * aligned range but not in the unaligned, original range are
8494 * put back to page allocator so that buddy can use them.
8497 ret = start_isolate_page_range(pfn_max_align_down(start),
8498 pfn_max_align_up(end), migratetype, 0);
8503 * In case of -EBUSY, we'd like to know which page causes problem.
8504 * So, just fall through. test_pages_isolated() has a tracepoint
8505 * which will report the busy page.
8507 * It is possible that busy pages could become available before
8508 * the call to test_pages_isolated, and the range will actually be
8509 * allocated. So, if we fall through be sure to clear ret so that
8510 * -EBUSY is not accidentally used or returned to caller.
8512 ret = __alloc_contig_migrate_range(&cc, start, end);
8513 if (ret && ret != -EBUSY)
8518 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8519 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8520 * more, all pages in [start, end) are free in page allocator.
8521 * What we are going to do is to allocate all pages from
8522 * [start, end) (that is remove them from page allocator).
8524 * The only problem is that pages at the beginning and at the
8525 * end of interesting range may be not aligned with pages that
8526 * page allocator holds, ie. they can be part of higher order
8527 * pages. Because of this, we reserve the bigger range and
8528 * once this is done free the pages we are not interested in.
8530 * We don't have to hold zone->lock here because the pages are
8531 * isolated thus they won't get removed from buddy.
8534 lru_add_drain_all();
8537 outer_start = start;
8538 while (!PageBuddy(pfn_to_page(outer_start))) {
8539 if (++order >= MAX_ORDER) {
8540 outer_start = start;
8543 outer_start &= ~0UL << order;
8546 if (outer_start != start) {
8547 order = page_order(pfn_to_page(outer_start));
8550 * outer_start page could be small order buddy page and
8551 * it doesn't include start page. Adjust outer_start
8552 * in this case to report failed page properly
8553 * on tracepoint in test_pages_isolated()
8555 if (outer_start + (1UL << order) <= start)
8556 outer_start = start;
8559 /* Make sure the range is really isolated. */
8560 if (test_pages_isolated(outer_start, end, 0)) {
8561 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8562 __func__, outer_start, end);
8567 /* Grab isolated pages from freelists. */
8568 outer_end = isolate_freepages_range(&cc, outer_start, end);
8574 /* Free head and tail (if any) */
8575 if (start != outer_start)
8576 free_contig_range(outer_start, start - outer_start);
8577 if (end != outer_end)
8578 free_contig_range(end, outer_end - end);
8581 undo_isolate_page_range(pfn_max_align_down(start),
8582 pfn_max_align_up(end), migratetype);
8585 EXPORT_SYMBOL(alloc_contig_range);
8587 static int __alloc_contig_pages(unsigned long start_pfn,
8588 unsigned long nr_pages, gfp_t gfp_mask)
8590 unsigned long end_pfn = start_pfn + nr_pages;
8592 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8596 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8597 unsigned long nr_pages)
8599 unsigned long i, end_pfn = start_pfn + nr_pages;
8602 for (i = start_pfn; i < end_pfn; i++) {
8603 page = pfn_to_online_page(i);
8607 if (page_zone(page) != z)
8610 if (PageReserved(page))
8613 if (page_count(page) > 0)
8622 static bool zone_spans_last_pfn(const struct zone *zone,
8623 unsigned long start_pfn, unsigned long nr_pages)
8625 unsigned long last_pfn = start_pfn + nr_pages - 1;
8627 return zone_spans_pfn(zone, last_pfn);
8631 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8632 * @nr_pages: Number of contiguous pages to allocate
8633 * @gfp_mask: GFP mask to limit search and used during compaction
8635 * @nodemask: Mask for other possible nodes
8637 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8638 * on an applicable zonelist to find a contiguous pfn range which can then be
8639 * tried for allocation with alloc_contig_range(). This routine is intended
8640 * for allocation requests which can not be fulfilled with the buddy allocator.
8642 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8643 * power of two then the alignment is guaranteed to be to the given nr_pages
8644 * (e.g. 1GB request would be aligned to 1GB).
8646 * Allocated pages can be freed with free_contig_range() or by manually calling
8647 * __free_page() on each allocated page.
8649 * Return: pointer to contiguous pages on success, or NULL if not successful.
8651 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8652 int nid, nodemask_t *nodemask)
8654 unsigned long ret, pfn, flags;
8655 struct zonelist *zonelist;
8659 zonelist = node_zonelist(nid, gfp_mask);
8660 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8661 gfp_zone(gfp_mask), nodemask) {
8662 spin_lock_irqsave(&zone->lock, flags);
8664 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8665 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8666 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8668 * We release the zone lock here because
8669 * alloc_contig_range() will also lock the zone
8670 * at some point. If there's an allocation
8671 * spinning on this lock, it may win the race
8672 * and cause alloc_contig_range() to fail...
8674 spin_unlock_irqrestore(&zone->lock, flags);
8675 ret = __alloc_contig_pages(pfn, nr_pages,
8678 return pfn_to_page(pfn);
8679 spin_lock_irqsave(&zone->lock, flags);
8683 spin_unlock_irqrestore(&zone->lock, flags);
8687 #endif /* CONFIG_CONTIG_ALLOC */
8689 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8691 unsigned int count = 0;
8693 for (; nr_pages--; pfn++) {
8694 struct page *page = pfn_to_page(pfn);
8696 count += page_count(page) != 1;
8699 WARN(count != 0, "%d pages are still in use!\n", count);
8701 EXPORT_SYMBOL(free_contig_range);
8704 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8705 * page high values need to be recalulated.
8707 void __meminit zone_pcp_update(struct zone *zone)
8709 mutex_lock(&pcp_batch_high_lock);
8710 __zone_pcp_update(zone);
8711 mutex_unlock(&pcp_batch_high_lock);
8714 void zone_pcp_reset(struct zone *zone)
8716 unsigned long flags;
8718 struct per_cpu_pageset *pset;
8720 /* avoid races with drain_pages() */
8721 local_irq_save(flags);
8722 if (zone->pageset != &boot_pageset) {
8723 for_each_online_cpu(cpu) {
8724 pset = per_cpu_ptr(zone->pageset, cpu);
8725 drain_zonestat(zone, pset);
8727 free_percpu(zone->pageset);
8728 zone->pageset = &boot_pageset;
8730 local_irq_restore(flags);
8733 #ifdef CONFIG_MEMORY_HOTREMOVE
8735 * All pages in the range must be in a single zone, must not contain holes,
8736 * must span full sections, and must be isolated before calling this function.
8738 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8740 unsigned long pfn = start_pfn;
8744 unsigned long flags;
8746 offline_mem_sections(pfn, end_pfn);
8747 zone = page_zone(pfn_to_page(pfn));
8748 spin_lock_irqsave(&zone->lock, flags);
8749 while (pfn < end_pfn) {
8750 page = pfn_to_page(pfn);
8752 * The HWPoisoned page may be not in buddy system, and
8753 * page_count() is not 0.
8755 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8760 * At this point all remaining PageOffline() pages have a
8761 * reference count of 0 and can simply be skipped.
8763 if (PageOffline(page)) {
8764 BUG_ON(page_count(page));
8765 BUG_ON(PageBuddy(page));
8770 BUG_ON(page_count(page));
8771 BUG_ON(!PageBuddy(page));
8772 order = page_order(page);
8773 del_page_from_free_list(page, zone, order);
8774 pfn += (1 << order);
8776 spin_unlock_irqrestore(&zone->lock, flags);
8780 bool is_free_buddy_page(struct page *page)
8782 struct zone *zone = page_zone(page);
8783 unsigned long pfn = page_to_pfn(page);
8784 unsigned long flags;
8787 spin_lock_irqsave(&zone->lock, flags);
8788 for (order = 0; order < MAX_ORDER; order++) {
8789 struct page *page_head = page - (pfn & ((1 << order) - 1));
8791 if (PageBuddy(page_head) && page_order(page_head) >= order)
8794 spin_unlock_irqrestore(&zone->lock, flags);
8796 return order < MAX_ORDER;
8799 #ifdef CONFIG_MEMORY_FAILURE
8801 * Break down a higher-order page in sub-pages, and keep our target out of
8804 static void break_down_buddy_pages(struct zone *zone, struct page *page,
8805 struct page *target, int low, int high,
8808 unsigned long size = 1 << high;
8809 struct page *current_buddy, *next_page;
8811 while (high > low) {
8815 if (target >= &page[size]) {
8816 next_page = page + size;
8817 current_buddy = page;
8820 current_buddy = page + size;
8823 if (set_page_guard(zone, current_buddy, high, migratetype))
8826 if (current_buddy != target) {
8827 add_to_free_list(current_buddy, zone, high, migratetype);
8828 set_page_order(current_buddy, high);
8835 * Take a page that will be marked as poisoned off the buddy allocator.
8837 bool take_page_off_buddy(struct page *page)
8839 struct zone *zone = page_zone(page);
8840 unsigned long pfn = page_to_pfn(page);
8841 unsigned long flags;
8845 spin_lock_irqsave(&zone->lock, flags);
8846 for (order = 0; order < MAX_ORDER; order++) {
8847 struct page *page_head = page - (pfn & ((1 << order) - 1));
8848 int buddy_order = page_order(page_head);
8850 if (PageBuddy(page_head) && buddy_order >= order) {
8851 unsigned long pfn_head = page_to_pfn(page_head);
8852 int migratetype = get_pfnblock_migratetype(page_head,
8855 del_page_from_free_list(page_head, zone, buddy_order);
8856 break_down_buddy_pages(zone, page_head, page, 0,
8857 buddy_order, migratetype);
8861 if (page_count(page_head) > 0)
8864 spin_unlock_irqrestore(&zone->lock, flags);