1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/page_alloc.c
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
18 #include <linux/stddef.h>
20 #include <linux/highmem.h>
21 #include <linux/swap.h>
22 #include <linux/interrupt.h>
23 #include <linux/pagemap.h>
24 #include <linux/jiffies.h>
25 #include <linux/memblock.h>
26 #include <linux/compiler.h>
27 #include <linux/kernel.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/memremap.h>
46 #include <linux/stop_machine.h>
47 #include <linux/random.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <trace/events/oom.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/sched/mm.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
66 #include <linux/memcontrol.h>
67 #include <linux/ftrace.h>
68 #include <linux/lockdep.h>
69 #include <linux/nmi.h>
70 #include <linux/psi.h>
72 #include <asm/sections.h>
73 #include <asm/tlbflush.h>
74 #include <asm/div64.h>
77 #include "page_reporting.h"
79 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
80 static DEFINE_MUTEX(pcp_batch_high_lock);
81 #define MIN_PERCPU_PAGELIST_FRACTION (8)
83 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
84 DEFINE_PER_CPU(int, numa_node);
85 EXPORT_PER_CPU_SYMBOL(numa_node);
88 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
90 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
92 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
93 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
94 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
95 * defined in <linux/topology.h>.
97 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
98 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
101 /* work_structs for global per-cpu drains */
104 struct work_struct work;
106 static DEFINE_MUTEX(pcpu_drain_mutex);
107 static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
109 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
110 volatile unsigned long latent_entropy __latent_entropy;
111 EXPORT_SYMBOL(latent_entropy);
115 * Array of node states.
117 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
118 [N_POSSIBLE] = NODE_MASK_ALL,
119 [N_ONLINE] = { { [0] = 1UL } },
121 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
122 #ifdef CONFIG_HIGHMEM
123 [N_HIGH_MEMORY] = { { [0] = 1UL } },
125 [N_MEMORY] = { { [0] = 1UL } },
126 [N_CPU] = { { [0] = 1UL } },
129 EXPORT_SYMBOL(node_states);
131 atomic_long_t _totalram_pages __read_mostly;
132 EXPORT_SYMBOL(_totalram_pages);
133 unsigned long totalreserve_pages __read_mostly;
134 unsigned long totalcma_pages __read_mostly;
136 int percpu_pagelist_fraction;
137 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
138 #ifdef CONFIG_INIT_ON_ALLOC_DEFAULT_ON
139 DEFINE_STATIC_KEY_TRUE(init_on_alloc);
141 DEFINE_STATIC_KEY_FALSE(init_on_alloc);
143 EXPORT_SYMBOL(init_on_alloc);
145 #ifdef CONFIG_INIT_ON_FREE_DEFAULT_ON
146 DEFINE_STATIC_KEY_TRUE(init_on_free);
148 DEFINE_STATIC_KEY_FALSE(init_on_free);
150 EXPORT_SYMBOL(init_on_free);
152 static int __init early_init_on_alloc(char *buf)
159 ret = kstrtobool(buf, &bool_result);
160 if (bool_result && page_poisoning_enabled())
161 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_alloc\n");
163 static_branch_enable(&init_on_alloc);
165 static_branch_disable(&init_on_alloc);
168 early_param("init_on_alloc", early_init_on_alloc);
170 static int __init early_init_on_free(char *buf)
177 ret = kstrtobool(buf, &bool_result);
178 if (bool_result && page_poisoning_enabled())
179 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_free\n");
181 static_branch_enable(&init_on_free);
183 static_branch_disable(&init_on_free);
186 early_param("init_on_free", early_init_on_free);
189 * A cached value of the page's pageblock's migratetype, used when the page is
190 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
191 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
192 * Also the migratetype set in the page does not necessarily match the pcplist
193 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
194 * other index - this ensures that it will be put on the correct CMA freelist.
196 static inline int get_pcppage_migratetype(struct page *page)
201 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
203 page->index = migratetype;
206 #ifdef CONFIG_PM_SLEEP
208 * The following functions are used by the suspend/hibernate code to temporarily
209 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
210 * while devices are suspended. To avoid races with the suspend/hibernate code,
211 * they should always be called with system_transition_mutex held
212 * (gfp_allowed_mask also should only be modified with system_transition_mutex
213 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
214 * with that modification).
217 static gfp_t saved_gfp_mask;
219 void pm_restore_gfp_mask(void)
221 WARN_ON(!mutex_is_locked(&system_transition_mutex));
222 if (saved_gfp_mask) {
223 gfp_allowed_mask = saved_gfp_mask;
228 void pm_restrict_gfp_mask(void)
230 WARN_ON(!mutex_is_locked(&system_transition_mutex));
231 WARN_ON(saved_gfp_mask);
232 saved_gfp_mask = gfp_allowed_mask;
233 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
236 bool pm_suspended_storage(void)
238 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
242 #endif /* CONFIG_PM_SLEEP */
244 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
245 unsigned int pageblock_order __read_mostly;
248 static void __free_pages_ok(struct page *page, unsigned int order);
251 * results with 256, 32 in the lowmem_reserve sysctl:
252 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
253 * 1G machine -> (16M dma, 784M normal, 224M high)
254 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
255 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
256 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
258 * TBD: should special case ZONE_DMA32 machines here - in those we normally
259 * don't need any ZONE_NORMAL reservation
261 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
262 #ifdef CONFIG_ZONE_DMA
265 #ifdef CONFIG_ZONE_DMA32
269 #ifdef CONFIG_HIGHMEM
275 static char * const zone_names[MAX_NR_ZONES] = {
276 #ifdef CONFIG_ZONE_DMA
279 #ifdef CONFIG_ZONE_DMA32
283 #ifdef CONFIG_HIGHMEM
287 #ifdef CONFIG_ZONE_DEVICE
292 const char * const migratetype_names[MIGRATE_TYPES] = {
300 #ifdef CONFIG_MEMORY_ISOLATION
305 compound_page_dtor * const compound_page_dtors[] = {
308 #ifdef CONFIG_HUGETLB_PAGE
311 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
316 int min_free_kbytes = 1024;
317 int user_min_free_kbytes = -1;
318 #ifdef CONFIG_DISCONTIGMEM
320 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
321 * are not on separate NUMA nodes. Functionally this works but with
322 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
323 * quite small. By default, do not boost watermarks on discontigmem as in
324 * many cases very high-order allocations like THP are likely to be
325 * unsupported and the premature reclaim offsets the advantage of long-term
326 * fragmentation avoidance.
328 int watermark_boost_factor __read_mostly;
330 int watermark_boost_factor __read_mostly = 15000;
332 int watermark_scale_factor = 10;
334 static unsigned long nr_kernel_pages __initdata;
335 static unsigned long nr_all_pages __initdata;
336 static unsigned long dma_reserve __initdata;
338 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
339 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
340 static unsigned long required_kernelcore __initdata;
341 static unsigned long required_kernelcore_percent __initdata;
342 static unsigned long required_movablecore __initdata;
343 static unsigned long required_movablecore_percent __initdata;
344 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
345 static bool mirrored_kernelcore __meminitdata;
347 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
349 EXPORT_SYMBOL(movable_zone);
352 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
353 unsigned int nr_online_nodes __read_mostly = 1;
354 EXPORT_SYMBOL(nr_node_ids);
355 EXPORT_SYMBOL(nr_online_nodes);
358 int page_group_by_mobility_disabled __read_mostly;
360 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
362 * During boot we initialize deferred pages on-demand, as needed, but once
363 * page_alloc_init_late() has finished, the deferred pages are all initialized,
364 * and we can permanently disable that path.
366 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
369 * Calling kasan_free_pages() only after deferred memory initialization
370 * has completed. Poisoning pages during deferred memory init will greatly
371 * lengthen the process and cause problem in large memory systems as the
372 * deferred pages initialization is done with interrupt disabled.
374 * Assuming that there will be no reference to those newly initialized
375 * pages before they are ever allocated, this should have no effect on
376 * KASAN memory tracking as the poison will be properly inserted at page
377 * allocation time. The only corner case is when pages are allocated by
378 * on-demand allocation and then freed again before the deferred pages
379 * initialization is done, but this is not likely to happen.
381 static inline void kasan_free_nondeferred_pages(struct page *page, int order)
383 if (!static_branch_unlikely(&deferred_pages))
384 kasan_free_pages(page, order);
387 /* Returns true if the struct page for the pfn is uninitialised */
388 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
390 int nid = early_pfn_to_nid(pfn);
392 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
399 * Returns true when the remaining initialisation should be deferred until
400 * later in the boot cycle when it can be parallelised.
402 static bool __meminit
403 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
405 static unsigned long prev_end_pfn, nr_initialised;
408 * prev_end_pfn static that contains the end of previous zone
409 * No need to protect because called very early in boot before smp_init.
411 if (prev_end_pfn != end_pfn) {
412 prev_end_pfn = end_pfn;
416 /* Always populate low zones for address-constrained allocations */
417 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
421 * We start only with one section of pages, more pages are added as
422 * needed until the rest of deferred pages are initialized.
425 if ((nr_initialised > PAGES_PER_SECTION) &&
426 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
427 NODE_DATA(nid)->first_deferred_pfn = pfn;
433 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
435 static inline bool early_page_uninitialised(unsigned long pfn)
440 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
446 /* Return a pointer to the bitmap storing bits affecting a block of pages */
447 static inline unsigned long *get_pageblock_bitmap(struct page *page,
450 #ifdef CONFIG_SPARSEMEM
451 return section_to_usemap(__pfn_to_section(pfn));
453 return page_zone(page)->pageblock_flags;
454 #endif /* CONFIG_SPARSEMEM */
457 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
459 #ifdef CONFIG_SPARSEMEM
460 pfn &= (PAGES_PER_SECTION-1);
461 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
463 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
464 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
465 #endif /* CONFIG_SPARSEMEM */
469 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
470 * @page: The page within the block of interest
471 * @pfn: The target page frame number
472 * @end_bitidx: The last bit of interest to retrieve
473 * @mask: mask of bits that the caller is interested in
475 * Return: pageblock_bits flags
477 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
479 unsigned long end_bitidx,
482 unsigned long *bitmap;
483 unsigned long bitidx, word_bitidx;
486 bitmap = get_pageblock_bitmap(page, pfn);
487 bitidx = pfn_to_bitidx(page, pfn);
488 word_bitidx = bitidx / BITS_PER_LONG;
489 bitidx &= (BITS_PER_LONG-1);
491 word = bitmap[word_bitidx];
492 bitidx += end_bitidx;
493 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
496 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
497 unsigned long end_bitidx,
500 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
503 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
505 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
509 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
510 * @page: The page within the block of interest
511 * @flags: The flags to set
512 * @pfn: The target page frame number
513 * @end_bitidx: The last bit of interest
514 * @mask: mask of bits that the caller is interested in
516 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
518 unsigned long end_bitidx,
521 unsigned long *bitmap;
522 unsigned long bitidx, word_bitidx;
523 unsigned long old_word, word;
525 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
526 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
528 bitmap = get_pageblock_bitmap(page, pfn);
529 bitidx = pfn_to_bitidx(page, pfn);
530 word_bitidx = bitidx / BITS_PER_LONG;
531 bitidx &= (BITS_PER_LONG-1);
533 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
535 bitidx += end_bitidx;
536 mask <<= (BITS_PER_LONG - bitidx - 1);
537 flags <<= (BITS_PER_LONG - bitidx - 1);
539 word = READ_ONCE(bitmap[word_bitidx]);
541 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
542 if (word == old_word)
548 void set_pageblock_migratetype(struct page *page, int migratetype)
550 if (unlikely(page_group_by_mobility_disabled &&
551 migratetype < MIGRATE_PCPTYPES))
552 migratetype = MIGRATE_UNMOVABLE;
554 set_pageblock_flags_group(page, (unsigned long)migratetype,
555 PB_migrate, PB_migrate_end);
558 #ifdef CONFIG_DEBUG_VM
559 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
563 unsigned long pfn = page_to_pfn(page);
564 unsigned long sp, start_pfn;
567 seq = zone_span_seqbegin(zone);
568 start_pfn = zone->zone_start_pfn;
569 sp = zone->spanned_pages;
570 if (!zone_spans_pfn(zone, pfn))
572 } while (zone_span_seqretry(zone, seq));
575 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
576 pfn, zone_to_nid(zone), zone->name,
577 start_pfn, start_pfn + sp);
582 static int page_is_consistent(struct zone *zone, struct page *page)
584 if (!pfn_valid_within(page_to_pfn(page)))
586 if (zone != page_zone(page))
592 * Temporary debugging check for pages not lying within a given zone.
594 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
596 if (page_outside_zone_boundaries(zone, page))
598 if (!page_is_consistent(zone, page))
604 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
610 static void bad_page(struct page *page, const char *reason)
612 static unsigned long resume;
613 static unsigned long nr_shown;
614 static unsigned long nr_unshown;
617 * Allow a burst of 60 reports, then keep quiet for that minute;
618 * or allow a steady drip of one report per second.
620 if (nr_shown == 60) {
621 if (time_before(jiffies, resume)) {
627 "BUG: Bad page state: %lu messages suppressed\n",
634 resume = jiffies + 60 * HZ;
636 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
637 current->comm, page_to_pfn(page));
638 __dump_page(page, reason);
639 dump_page_owner(page);
644 /* Leave bad fields for debug, except PageBuddy could make trouble */
645 page_mapcount_reset(page); /* remove PageBuddy */
646 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
650 * Higher-order pages are called "compound pages". They are structured thusly:
652 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
654 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
655 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
657 * The first tail page's ->compound_dtor holds the offset in array of compound
658 * page destructors. See compound_page_dtors.
660 * The first tail page's ->compound_order holds the order of allocation.
661 * This usage means that zero-order pages may not be compound.
664 void free_compound_page(struct page *page)
666 mem_cgroup_uncharge(page);
667 __free_pages_ok(page, compound_order(page));
670 void prep_compound_page(struct page *page, unsigned int order)
673 int nr_pages = 1 << order;
675 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
676 set_compound_order(page, order);
678 for (i = 1; i < nr_pages; i++) {
679 struct page *p = page + i;
680 set_page_count(p, 0);
681 p->mapping = TAIL_MAPPING;
682 set_compound_head(p, page);
684 atomic_set(compound_mapcount_ptr(page), -1);
685 if (hpage_pincount_available(page))
686 atomic_set(compound_pincount_ptr(page), 0);
689 #ifdef CONFIG_DEBUG_PAGEALLOC
690 unsigned int _debug_guardpage_minorder;
692 bool _debug_pagealloc_enabled_early __read_mostly
693 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
694 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
695 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
696 EXPORT_SYMBOL(_debug_pagealloc_enabled);
698 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
700 static int __init early_debug_pagealloc(char *buf)
702 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
704 early_param("debug_pagealloc", early_debug_pagealloc);
706 void init_debug_pagealloc(void)
708 if (!debug_pagealloc_enabled())
711 static_branch_enable(&_debug_pagealloc_enabled);
713 if (!debug_guardpage_minorder())
716 static_branch_enable(&_debug_guardpage_enabled);
719 static int __init debug_guardpage_minorder_setup(char *buf)
723 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
724 pr_err("Bad debug_guardpage_minorder value\n");
727 _debug_guardpage_minorder = res;
728 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
731 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
733 static inline bool set_page_guard(struct zone *zone, struct page *page,
734 unsigned int order, int migratetype)
736 if (!debug_guardpage_enabled())
739 if (order >= debug_guardpage_minorder())
742 __SetPageGuard(page);
743 INIT_LIST_HEAD(&page->lru);
744 set_page_private(page, order);
745 /* Guard pages are not available for any usage */
746 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
751 static inline void clear_page_guard(struct zone *zone, struct page *page,
752 unsigned int order, int migratetype)
754 if (!debug_guardpage_enabled())
757 __ClearPageGuard(page);
759 set_page_private(page, 0);
760 if (!is_migrate_isolate(migratetype))
761 __mod_zone_freepage_state(zone, (1 << order), migratetype);
764 static inline bool set_page_guard(struct zone *zone, struct page *page,
765 unsigned int order, int migratetype) { return false; }
766 static inline void clear_page_guard(struct zone *zone, struct page *page,
767 unsigned int order, int migratetype) {}
770 static inline void set_page_order(struct page *page, unsigned int order)
772 set_page_private(page, order);
773 __SetPageBuddy(page);
777 * This function checks whether a page is free && is the buddy
778 * we can coalesce a page and its buddy if
779 * (a) the buddy is not in a hole (check before calling!) &&
780 * (b) the buddy is in the buddy system &&
781 * (c) a page and its buddy have the same order &&
782 * (d) a page and its buddy are in the same zone.
784 * For recording whether a page is in the buddy system, we set PageBuddy.
785 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
787 * For recording page's order, we use page_private(page).
789 static inline bool page_is_buddy(struct page *page, struct page *buddy,
792 if (!page_is_guard(buddy) && !PageBuddy(buddy))
795 if (page_order(buddy) != order)
799 * zone check is done late to avoid uselessly calculating
800 * zone/node ids for pages that could never merge.
802 if (page_zone_id(page) != page_zone_id(buddy))
805 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
810 #ifdef CONFIG_COMPACTION
811 static inline struct capture_control *task_capc(struct zone *zone)
813 struct capture_control *capc = current->capture_control;
816 !(current->flags & PF_KTHREAD) &&
818 capc->cc->zone == zone &&
819 capc->cc->direct_compaction ? capc : NULL;
823 compaction_capture(struct capture_control *capc, struct page *page,
824 int order, int migratetype)
826 if (!capc || order != capc->cc->order)
829 /* Do not accidentally pollute CMA or isolated regions*/
830 if (is_migrate_cma(migratetype) ||
831 is_migrate_isolate(migratetype))
835 * Do not let lower order allocations polluate a movable pageblock.
836 * This might let an unmovable request use a reclaimable pageblock
837 * and vice-versa but no more than normal fallback logic which can
838 * have trouble finding a high-order free page.
840 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
848 static inline struct capture_control *task_capc(struct zone *zone)
854 compaction_capture(struct capture_control *capc, struct page *page,
855 int order, int migratetype)
859 #endif /* CONFIG_COMPACTION */
861 /* Used for pages not on another list */
862 static inline void add_to_free_list(struct page *page, struct zone *zone,
863 unsigned int order, int migratetype)
865 struct free_area *area = &zone->free_area[order];
867 list_add(&page->lru, &area->free_list[migratetype]);
871 /* Used for pages not on another list */
872 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
873 unsigned int order, int migratetype)
875 struct free_area *area = &zone->free_area[order];
877 list_add_tail(&page->lru, &area->free_list[migratetype]);
881 /* Used for pages which are on another list */
882 static inline void move_to_free_list(struct page *page, struct zone *zone,
883 unsigned int order, int migratetype)
885 struct free_area *area = &zone->free_area[order];
887 list_move(&page->lru, &area->free_list[migratetype]);
890 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
893 /* clear reported state and update reported page count */
894 if (page_reported(page))
895 __ClearPageReported(page);
897 list_del(&page->lru);
898 __ClearPageBuddy(page);
899 set_page_private(page, 0);
900 zone->free_area[order].nr_free--;
904 * If this is not the largest possible page, check if the buddy
905 * of the next-highest order is free. If it is, it's possible
906 * that pages are being freed that will coalesce soon. In case,
907 * that is happening, add the free page to the tail of the list
908 * so it's less likely to be used soon and more likely to be merged
909 * as a higher order page
912 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
913 struct page *page, unsigned int order)
915 struct page *higher_page, *higher_buddy;
916 unsigned long combined_pfn;
918 if (order >= MAX_ORDER - 2)
921 if (!pfn_valid_within(buddy_pfn))
924 combined_pfn = buddy_pfn & pfn;
925 higher_page = page + (combined_pfn - pfn);
926 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
927 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
929 return pfn_valid_within(buddy_pfn) &&
930 page_is_buddy(higher_page, higher_buddy, order + 1);
934 * Freeing function for a buddy system allocator.
936 * The concept of a buddy system is to maintain direct-mapped table
937 * (containing bit values) for memory blocks of various "orders".
938 * The bottom level table contains the map for the smallest allocatable
939 * units of memory (here, pages), and each level above it describes
940 * pairs of units from the levels below, hence, "buddies".
941 * At a high level, all that happens here is marking the table entry
942 * at the bottom level available, and propagating the changes upward
943 * as necessary, plus some accounting needed to play nicely with other
944 * parts of the VM system.
945 * At each level, we keep a list of pages, which are heads of continuous
946 * free pages of length of (1 << order) and marked with PageBuddy.
947 * Page's order is recorded in page_private(page) field.
948 * So when we are allocating or freeing one, we can derive the state of the
949 * other. That is, if we allocate a small block, and both were
950 * free, the remainder of the region must be split into blocks.
951 * If a block is freed, and its buddy is also free, then this
952 * triggers coalescing into a block of larger size.
957 static inline void __free_one_page(struct page *page,
959 struct zone *zone, unsigned int order,
960 int migratetype, bool report)
962 struct capture_control *capc = task_capc(zone);
963 unsigned long uninitialized_var(buddy_pfn);
964 unsigned long combined_pfn;
965 unsigned int max_order;
969 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
971 VM_BUG_ON(!zone_is_initialized(zone));
972 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
974 VM_BUG_ON(migratetype == -1);
975 if (likely(!is_migrate_isolate(migratetype)))
976 __mod_zone_freepage_state(zone, 1 << order, migratetype);
978 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
979 VM_BUG_ON_PAGE(bad_range(zone, page), page);
982 while (order < max_order - 1) {
983 if (compaction_capture(capc, page, order, migratetype)) {
984 __mod_zone_freepage_state(zone, -(1 << order),
988 buddy_pfn = __find_buddy_pfn(pfn, order);
989 buddy = page + (buddy_pfn - pfn);
991 if (!pfn_valid_within(buddy_pfn))
993 if (!page_is_buddy(page, buddy, order))
996 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
997 * merge with it and move up one order.
999 if (page_is_guard(buddy))
1000 clear_page_guard(zone, buddy, order, migratetype);
1002 del_page_from_free_list(buddy, zone, order);
1003 combined_pfn = buddy_pfn & pfn;
1004 page = page + (combined_pfn - pfn);
1008 if (max_order < MAX_ORDER) {
1009 /* If we are here, it means order is >= pageblock_order.
1010 * We want to prevent merge between freepages on isolate
1011 * pageblock and normal pageblock. Without this, pageblock
1012 * isolation could cause incorrect freepage or CMA accounting.
1014 * We don't want to hit this code for the more frequent
1015 * low-order merging.
1017 if (unlikely(has_isolate_pageblock(zone))) {
1020 buddy_pfn = __find_buddy_pfn(pfn, order);
1021 buddy = page + (buddy_pfn - pfn);
1022 buddy_mt = get_pageblock_migratetype(buddy);
1024 if (migratetype != buddy_mt
1025 && (is_migrate_isolate(migratetype) ||
1026 is_migrate_isolate(buddy_mt)))
1030 goto continue_merging;
1034 set_page_order(page, order);
1036 if (is_shuffle_order(order))
1037 to_tail = shuffle_pick_tail();
1039 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1042 add_to_free_list_tail(page, zone, order, migratetype);
1044 add_to_free_list(page, zone, order, migratetype);
1046 /* Notify page reporting subsystem of freed page */
1048 page_reporting_notify_free(order);
1052 * A bad page could be due to a number of fields. Instead of multiple branches,
1053 * try and check multiple fields with one check. The caller must do a detailed
1054 * check if necessary.
1056 static inline bool page_expected_state(struct page *page,
1057 unsigned long check_flags)
1059 if (unlikely(atomic_read(&page->_mapcount) != -1))
1062 if (unlikely((unsigned long)page->mapping |
1063 page_ref_count(page) |
1065 (unsigned long)page->mem_cgroup |
1067 (page->flags & check_flags)))
1073 static void check_free_page_bad(struct page *page)
1075 const char *bad_reason = NULL;
1077 if (unlikely(atomic_read(&page->_mapcount) != -1))
1078 bad_reason = "nonzero mapcount";
1079 if (unlikely(page->mapping != NULL))
1080 bad_reason = "non-NULL mapping";
1081 if (unlikely(page_ref_count(page) != 0))
1082 bad_reason = "nonzero _refcount";
1083 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE))
1084 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1086 if (unlikely(page->mem_cgroup))
1087 bad_reason = "page still charged to cgroup";
1089 bad_page(page, bad_reason);
1092 static inline int check_free_page(struct page *page)
1094 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1097 /* Something has gone sideways, find it */
1098 check_free_page_bad(page);
1102 static int free_tail_pages_check(struct page *head_page, struct page *page)
1107 * We rely page->lru.next never has bit 0 set, unless the page
1108 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1110 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1112 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1116 switch (page - head_page) {
1118 /* the first tail page: ->mapping may be compound_mapcount() */
1119 if (unlikely(compound_mapcount(page))) {
1120 bad_page(page, "nonzero compound_mapcount");
1126 * the second tail page: ->mapping is
1127 * deferred_list.next -- ignore value.
1131 if (page->mapping != TAIL_MAPPING) {
1132 bad_page(page, "corrupted mapping in tail page");
1137 if (unlikely(!PageTail(page))) {
1138 bad_page(page, "PageTail not set");
1141 if (unlikely(compound_head(page) != head_page)) {
1142 bad_page(page, "compound_head not consistent");
1147 page->mapping = NULL;
1148 clear_compound_head(page);
1152 static void kernel_init_free_pages(struct page *page, int numpages)
1156 for (i = 0; i < numpages; i++)
1157 clear_highpage(page + i);
1160 static __always_inline bool free_pages_prepare(struct page *page,
1161 unsigned int order, bool check_free)
1165 VM_BUG_ON_PAGE(PageTail(page), page);
1167 trace_mm_page_free(page, order);
1170 * Check tail pages before head page information is cleared to
1171 * avoid checking PageCompound for order-0 pages.
1173 if (unlikely(order)) {
1174 bool compound = PageCompound(page);
1177 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1180 ClearPageDoubleMap(page);
1181 for (i = 1; i < (1 << order); i++) {
1183 bad += free_tail_pages_check(page, page + i);
1184 if (unlikely(check_free_page(page + i))) {
1188 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1191 if (PageMappingFlags(page))
1192 page->mapping = NULL;
1193 if (memcg_kmem_enabled() && PageKmemcg(page))
1194 __memcg_kmem_uncharge_page(page, order);
1196 bad += check_free_page(page);
1200 page_cpupid_reset_last(page);
1201 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1202 reset_page_owner(page, order);
1204 if (!PageHighMem(page)) {
1205 debug_check_no_locks_freed(page_address(page),
1206 PAGE_SIZE << order);
1207 debug_check_no_obj_freed(page_address(page),
1208 PAGE_SIZE << order);
1210 if (want_init_on_free())
1211 kernel_init_free_pages(page, 1 << order);
1213 kernel_poison_pages(page, 1 << order, 0);
1215 * arch_free_page() can make the page's contents inaccessible. s390
1216 * does this. So nothing which can access the page's contents should
1217 * happen after this.
1219 arch_free_page(page, order);
1221 if (debug_pagealloc_enabled_static())
1222 kernel_map_pages(page, 1 << order, 0);
1224 kasan_free_nondeferred_pages(page, order);
1229 #ifdef CONFIG_DEBUG_VM
1231 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1232 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1233 * moved from pcp lists to free lists.
1235 static bool free_pcp_prepare(struct page *page)
1237 return free_pages_prepare(page, 0, true);
1240 static bool bulkfree_pcp_prepare(struct page *page)
1242 if (debug_pagealloc_enabled_static())
1243 return check_free_page(page);
1249 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1250 * moving from pcp lists to free list in order to reduce overhead. With
1251 * debug_pagealloc enabled, they are checked also immediately when being freed
1254 static bool free_pcp_prepare(struct page *page)
1256 if (debug_pagealloc_enabled_static())
1257 return free_pages_prepare(page, 0, true);
1259 return free_pages_prepare(page, 0, false);
1262 static bool bulkfree_pcp_prepare(struct page *page)
1264 return check_free_page(page);
1266 #endif /* CONFIG_DEBUG_VM */
1268 static inline void prefetch_buddy(struct page *page)
1270 unsigned long pfn = page_to_pfn(page);
1271 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1272 struct page *buddy = page + (buddy_pfn - pfn);
1278 * Frees a number of pages from the PCP lists
1279 * Assumes all pages on list are in same zone, and of same order.
1280 * count is the number of pages to free.
1282 * If the zone was previously in an "all pages pinned" state then look to
1283 * see if this freeing clears that state.
1285 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1286 * pinned" detection logic.
1288 static void free_pcppages_bulk(struct zone *zone, int count,
1289 struct per_cpu_pages *pcp)
1291 int migratetype = 0;
1293 int prefetch_nr = 0;
1294 bool isolated_pageblocks;
1295 struct page *page, *tmp;
1299 struct list_head *list;
1302 * Remove pages from lists in a round-robin fashion. A
1303 * batch_free count is maintained that is incremented when an
1304 * empty list is encountered. This is so more pages are freed
1305 * off fuller lists instead of spinning excessively around empty
1310 if (++migratetype == MIGRATE_PCPTYPES)
1312 list = &pcp->lists[migratetype];
1313 } while (list_empty(list));
1315 /* This is the only non-empty list. Free them all. */
1316 if (batch_free == MIGRATE_PCPTYPES)
1320 page = list_last_entry(list, struct page, lru);
1321 /* must delete to avoid corrupting pcp list */
1322 list_del(&page->lru);
1325 if (bulkfree_pcp_prepare(page))
1328 list_add_tail(&page->lru, &head);
1331 * We are going to put the page back to the global
1332 * pool, prefetch its buddy to speed up later access
1333 * under zone->lock. It is believed the overhead of
1334 * an additional test and calculating buddy_pfn here
1335 * can be offset by reduced memory latency later. To
1336 * avoid excessive prefetching due to large count, only
1337 * prefetch buddy for the first pcp->batch nr of pages.
1339 if (prefetch_nr++ < pcp->batch)
1340 prefetch_buddy(page);
1341 } while (--count && --batch_free && !list_empty(list));
1344 spin_lock(&zone->lock);
1345 isolated_pageblocks = has_isolate_pageblock(zone);
1348 * Use safe version since after __free_one_page(),
1349 * page->lru.next will not point to original list.
1351 list_for_each_entry_safe(page, tmp, &head, lru) {
1352 int mt = get_pcppage_migratetype(page);
1353 /* MIGRATE_ISOLATE page should not go to pcplists */
1354 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1355 /* Pageblock could have been isolated meanwhile */
1356 if (unlikely(isolated_pageblocks))
1357 mt = get_pageblock_migratetype(page);
1359 __free_one_page(page, page_to_pfn(page), zone, 0, mt, true);
1360 trace_mm_page_pcpu_drain(page, 0, mt);
1362 spin_unlock(&zone->lock);
1365 static void free_one_page(struct zone *zone,
1366 struct page *page, unsigned long pfn,
1370 spin_lock(&zone->lock);
1371 if (unlikely(has_isolate_pageblock(zone) ||
1372 is_migrate_isolate(migratetype))) {
1373 migratetype = get_pfnblock_migratetype(page, pfn);
1375 __free_one_page(page, pfn, zone, order, migratetype, true);
1376 spin_unlock(&zone->lock);
1379 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1380 unsigned long zone, int nid)
1382 mm_zero_struct_page(page);
1383 set_page_links(page, zone, nid, pfn);
1384 init_page_count(page);
1385 page_mapcount_reset(page);
1386 page_cpupid_reset_last(page);
1387 page_kasan_tag_reset(page);
1389 INIT_LIST_HEAD(&page->lru);
1390 #ifdef WANT_PAGE_VIRTUAL
1391 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1392 if (!is_highmem_idx(zone))
1393 set_page_address(page, __va(pfn << PAGE_SHIFT));
1397 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1398 static void __meminit init_reserved_page(unsigned long pfn)
1403 if (!early_page_uninitialised(pfn))
1406 nid = early_pfn_to_nid(pfn);
1407 pgdat = NODE_DATA(nid);
1409 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1410 struct zone *zone = &pgdat->node_zones[zid];
1412 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1415 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1418 static inline void init_reserved_page(unsigned long pfn)
1421 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1424 * Initialised pages do not have PageReserved set. This function is
1425 * called for each range allocated by the bootmem allocator and
1426 * marks the pages PageReserved. The remaining valid pages are later
1427 * sent to the buddy page allocator.
1429 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1431 unsigned long start_pfn = PFN_DOWN(start);
1432 unsigned long end_pfn = PFN_UP(end);
1434 for (; start_pfn < end_pfn; start_pfn++) {
1435 if (pfn_valid(start_pfn)) {
1436 struct page *page = pfn_to_page(start_pfn);
1438 init_reserved_page(start_pfn);
1440 /* Avoid false-positive PageTail() */
1441 INIT_LIST_HEAD(&page->lru);
1444 * no need for atomic set_bit because the struct
1445 * page is not visible yet so nobody should
1448 __SetPageReserved(page);
1453 static void __free_pages_ok(struct page *page, unsigned int order)
1455 unsigned long flags;
1457 unsigned long pfn = page_to_pfn(page);
1459 if (!free_pages_prepare(page, order, true))
1462 migratetype = get_pfnblock_migratetype(page, pfn);
1463 local_irq_save(flags);
1464 __count_vm_events(PGFREE, 1 << order);
1465 free_one_page(page_zone(page), page, pfn, order, migratetype);
1466 local_irq_restore(flags);
1469 void __free_pages_core(struct page *page, unsigned int order)
1471 unsigned int nr_pages = 1 << order;
1472 struct page *p = page;
1476 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1478 __ClearPageReserved(p);
1479 set_page_count(p, 0);
1481 __ClearPageReserved(p);
1482 set_page_count(p, 0);
1484 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1485 set_page_refcounted(page);
1486 __free_pages(page, order);
1489 #ifdef CONFIG_NEED_MULTIPLE_NODES
1491 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1493 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
1496 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1498 int __meminit __early_pfn_to_nid(unsigned long pfn,
1499 struct mminit_pfnnid_cache *state)
1501 unsigned long start_pfn, end_pfn;
1504 if (state->last_start <= pfn && pfn < state->last_end)
1505 return state->last_nid;
1507 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1508 if (nid != NUMA_NO_NODE) {
1509 state->last_start = start_pfn;
1510 state->last_end = end_pfn;
1511 state->last_nid = nid;
1516 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
1518 int __meminit early_pfn_to_nid(unsigned long pfn)
1520 static DEFINE_SPINLOCK(early_pfn_lock);
1523 spin_lock(&early_pfn_lock);
1524 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1526 nid = first_online_node;
1527 spin_unlock(&early_pfn_lock);
1531 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1533 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1536 if (early_page_uninitialised(pfn))
1538 __free_pages_core(page, order);
1542 * Check that the whole (or subset of) a pageblock given by the interval of
1543 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1544 * with the migration of free compaction scanner. The scanners then need to
1545 * use only pfn_valid_within() check for arches that allow holes within
1548 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1550 * It's possible on some configurations to have a setup like node0 node1 node0
1551 * i.e. it's possible that all pages within a zones range of pages do not
1552 * belong to a single zone. We assume that a border between node0 and node1
1553 * can occur within a single pageblock, but not a node0 node1 node0
1554 * interleaving within a single pageblock. It is therefore sufficient to check
1555 * the first and last page of a pageblock and avoid checking each individual
1556 * page in a pageblock.
1558 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1559 unsigned long end_pfn, struct zone *zone)
1561 struct page *start_page;
1562 struct page *end_page;
1564 /* end_pfn is one past the range we are checking */
1567 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1570 start_page = pfn_to_online_page(start_pfn);
1574 if (page_zone(start_page) != zone)
1577 end_page = pfn_to_page(end_pfn);
1579 /* This gives a shorter code than deriving page_zone(end_page) */
1580 if (page_zone_id(start_page) != page_zone_id(end_page))
1586 void set_zone_contiguous(struct zone *zone)
1588 unsigned long block_start_pfn = zone->zone_start_pfn;
1589 unsigned long block_end_pfn;
1591 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1592 for (; block_start_pfn < zone_end_pfn(zone);
1593 block_start_pfn = block_end_pfn,
1594 block_end_pfn += pageblock_nr_pages) {
1596 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1598 if (!__pageblock_pfn_to_page(block_start_pfn,
1599 block_end_pfn, zone))
1604 /* We confirm that there is no hole */
1605 zone->contiguous = true;
1608 void clear_zone_contiguous(struct zone *zone)
1610 zone->contiguous = false;
1613 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1614 static void __init deferred_free_range(unsigned long pfn,
1615 unsigned long nr_pages)
1623 page = pfn_to_page(pfn);
1625 /* Free a large naturally-aligned chunk if possible */
1626 if (nr_pages == pageblock_nr_pages &&
1627 (pfn & (pageblock_nr_pages - 1)) == 0) {
1628 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1629 __free_pages_core(page, pageblock_order);
1633 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1634 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1635 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1636 __free_pages_core(page, 0);
1640 /* Completion tracking for deferred_init_memmap() threads */
1641 static atomic_t pgdat_init_n_undone __initdata;
1642 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1644 static inline void __init pgdat_init_report_one_done(void)
1646 if (atomic_dec_and_test(&pgdat_init_n_undone))
1647 complete(&pgdat_init_all_done_comp);
1651 * Returns true if page needs to be initialized or freed to buddy allocator.
1653 * First we check if pfn is valid on architectures where it is possible to have
1654 * holes within pageblock_nr_pages. On systems where it is not possible, this
1655 * function is optimized out.
1657 * Then, we check if a current large page is valid by only checking the validity
1660 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1662 if (!pfn_valid_within(pfn))
1664 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1670 * Free pages to buddy allocator. Try to free aligned pages in
1671 * pageblock_nr_pages sizes.
1673 static void __init deferred_free_pages(unsigned long pfn,
1674 unsigned long end_pfn)
1676 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1677 unsigned long nr_free = 0;
1679 for (; pfn < end_pfn; pfn++) {
1680 if (!deferred_pfn_valid(pfn)) {
1681 deferred_free_range(pfn - nr_free, nr_free);
1683 } else if (!(pfn & nr_pgmask)) {
1684 deferred_free_range(pfn - nr_free, nr_free);
1686 touch_nmi_watchdog();
1691 /* Free the last block of pages to allocator */
1692 deferred_free_range(pfn - nr_free, nr_free);
1696 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1697 * by performing it only once every pageblock_nr_pages.
1698 * Return number of pages initialized.
1700 static unsigned long __init deferred_init_pages(struct zone *zone,
1702 unsigned long end_pfn)
1704 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1705 int nid = zone_to_nid(zone);
1706 unsigned long nr_pages = 0;
1707 int zid = zone_idx(zone);
1708 struct page *page = NULL;
1710 for (; pfn < end_pfn; pfn++) {
1711 if (!deferred_pfn_valid(pfn)) {
1714 } else if (!page || !(pfn & nr_pgmask)) {
1715 page = pfn_to_page(pfn);
1716 touch_nmi_watchdog();
1720 __init_single_page(page, pfn, zid, nid);
1727 * This function is meant to pre-load the iterator for the zone init.
1728 * Specifically it walks through the ranges until we are caught up to the
1729 * first_init_pfn value and exits there. If we never encounter the value we
1730 * return false indicating there are no valid ranges left.
1733 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1734 unsigned long *spfn, unsigned long *epfn,
1735 unsigned long first_init_pfn)
1740 * Start out by walking through the ranges in this zone that have
1741 * already been initialized. We don't need to do anything with them
1742 * so we just need to flush them out of the system.
1744 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1745 if (*epfn <= first_init_pfn)
1747 if (*spfn < first_init_pfn)
1748 *spfn = first_init_pfn;
1757 * Initialize and free pages. We do it in two loops: first we initialize
1758 * struct page, then free to buddy allocator, because while we are
1759 * freeing pages we can access pages that are ahead (computing buddy
1760 * page in __free_one_page()).
1762 * In order to try and keep some memory in the cache we have the loop
1763 * broken along max page order boundaries. This way we will not cause
1764 * any issues with the buddy page computation.
1766 static unsigned long __init
1767 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1768 unsigned long *end_pfn)
1770 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1771 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1772 unsigned long nr_pages = 0;
1775 /* First we loop through and initialize the page values */
1776 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1779 if (mo_pfn <= *start_pfn)
1782 t = min(mo_pfn, *end_pfn);
1783 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1785 if (mo_pfn < *end_pfn) {
1786 *start_pfn = mo_pfn;
1791 /* Reset values and now loop through freeing pages as needed */
1794 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1800 t = min(mo_pfn, epfn);
1801 deferred_free_pages(spfn, t);
1810 /* Initialise remaining memory on a node */
1811 static int __init deferred_init_memmap(void *data)
1813 pg_data_t *pgdat = data;
1814 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1815 unsigned long spfn = 0, epfn = 0, nr_pages = 0;
1816 unsigned long first_init_pfn, flags;
1817 unsigned long start = jiffies;
1822 /* Bind memory initialisation thread to a local node if possible */
1823 if (!cpumask_empty(cpumask))
1824 set_cpus_allowed_ptr(current, cpumask);
1826 pgdat_resize_lock(pgdat, &flags);
1827 first_init_pfn = pgdat->first_deferred_pfn;
1828 if (first_init_pfn == ULONG_MAX) {
1829 pgdat_resize_unlock(pgdat, &flags);
1830 pgdat_init_report_one_done();
1834 /* Sanity check boundaries */
1835 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1836 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1837 pgdat->first_deferred_pfn = ULONG_MAX;
1839 /* Only the highest zone is deferred so find it */
1840 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1841 zone = pgdat->node_zones + zid;
1842 if (first_init_pfn < zone_end_pfn(zone))
1846 /* If the zone is empty somebody else may have cleared out the zone */
1847 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1852 * Initialize and free pages in MAX_ORDER sized increments so
1853 * that we can avoid introducing any issues with the buddy
1857 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1859 pgdat_resize_unlock(pgdat, &flags);
1861 /* Sanity check that the next zone really is unpopulated */
1862 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1864 pr_info("node %d initialised, %lu pages in %ums\n",
1865 pgdat->node_id, nr_pages, jiffies_to_msecs(jiffies - start));
1867 pgdat_init_report_one_done();
1872 * If this zone has deferred pages, try to grow it by initializing enough
1873 * deferred pages to satisfy the allocation specified by order, rounded up to
1874 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1875 * of SECTION_SIZE bytes by initializing struct pages in increments of
1876 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1878 * Return true when zone was grown, otherwise return false. We return true even
1879 * when we grow less than requested, to let the caller decide if there are
1880 * enough pages to satisfy the allocation.
1882 * Note: We use noinline because this function is needed only during boot, and
1883 * it is called from a __ref function _deferred_grow_zone. This way we are
1884 * making sure that it is not inlined into permanent text section.
1886 static noinline bool __init
1887 deferred_grow_zone(struct zone *zone, unsigned int order)
1889 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1890 pg_data_t *pgdat = zone->zone_pgdat;
1891 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1892 unsigned long spfn, epfn, flags;
1893 unsigned long nr_pages = 0;
1896 /* Only the last zone may have deferred pages */
1897 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1900 pgdat_resize_lock(pgdat, &flags);
1903 * If deferred pages have been initialized while we were waiting for
1904 * the lock, return true, as the zone was grown. The caller will retry
1905 * this zone. We won't return to this function since the caller also
1906 * has this static branch.
1908 if (!static_branch_unlikely(&deferred_pages)) {
1909 pgdat_resize_unlock(pgdat, &flags);
1914 * If someone grew this zone while we were waiting for spinlock, return
1915 * true, as there might be enough pages already.
1917 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1918 pgdat_resize_unlock(pgdat, &flags);
1922 /* If the zone is empty somebody else may have cleared out the zone */
1923 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1924 first_deferred_pfn)) {
1925 pgdat->first_deferred_pfn = ULONG_MAX;
1926 pgdat_resize_unlock(pgdat, &flags);
1927 /* Retry only once. */
1928 return first_deferred_pfn != ULONG_MAX;
1932 * Initialize and free pages in MAX_ORDER sized increments so
1933 * that we can avoid introducing any issues with the buddy
1936 while (spfn < epfn) {
1937 /* update our first deferred PFN for this section */
1938 first_deferred_pfn = spfn;
1940 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1942 /* We should only stop along section boundaries */
1943 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
1946 /* If our quota has been met we can stop here */
1947 if (nr_pages >= nr_pages_needed)
1951 pgdat->first_deferred_pfn = spfn;
1952 pgdat_resize_unlock(pgdat, &flags);
1954 return nr_pages > 0;
1958 * deferred_grow_zone() is __init, but it is called from
1959 * get_page_from_freelist() during early boot until deferred_pages permanently
1960 * disables this call. This is why we have refdata wrapper to avoid warning,
1961 * and to ensure that the function body gets unloaded.
1964 _deferred_grow_zone(struct zone *zone, unsigned int order)
1966 return deferred_grow_zone(zone, order);
1969 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1971 void __init page_alloc_init_late(void)
1976 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1978 /* There will be num_node_state(N_MEMORY) threads */
1979 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1980 for_each_node_state(nid, N_MEMORY) {
1981 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1984 /* Block until all are initialised */
1985 wait_for_completion(&pgdat_init_all_done_comp);
1988 * The number of managed pages has changed due to the initialisation
1989 * so the pcpu batch and high limits needs to be updated or the limits
1990 * will be artificially small.
1992 for_each_populated_zone(zone)
1993 zone_pcp_update(zone);
1996 * We initialized the rest of the deferred pages. Permanently disable
1997 * on-demand struct page initialization.
1999 static_branch_disable(&deferred_pages);
2001 /* Reinit limits that are based on free pages after the kernel is up */
2002 files_maxfiles_init();
2005 /* Discard memblock private memory */
2008 for_each_node_state(nid, N_MEMORY)
2009 shuffle_free_memory(NODE_DATA(nid));
2011 for_each_populated_zone(zone)
2012 set_zone_contiguous(zone);
2016 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2017 void __init init_cma_reserved_pageblock(struct page *page)
2019 unsigned i = pageblock_nr_pages;
2020 struct page *p = page;
2023 __ClearPageReserved(p);
2024 set_page_count(p, 0);
2027 set_pageblock_migratetype(page, MIGRATE_CMA);
2029 if (pageblock_order >= MAX_ORDER) {
2030 i = pageblock_nr_pages;
2033 set_page_refcounted(p);
2034 __free_pages(p, MAX_ORDER - 1);
2035 p += MAX_ORDER_NR_PAGES;
2036 } while (i -= MAX_ORDER_NR_PAGES);
2038 set_page_refcounted(page);
2039 __free_pages(page, pageblock_order);
2042 adjust_managed_page_count(page, pageblock_nr_pages);
2047 * The order of subdivision here is critical for the IO subsystem.
2048 * Please do not alter this order without good reasons and regression
2049 * testing. Specifically, as large blocks of memory are subdivided,
2050 * the order in which smaller blocks are delivered depends on the order
2051 * they're subdivided in this function. This is the primary factor
2052 * influencing the order in which pages are delivered to the IO
2053 * subsystem according to empirical testing, and this is also justified
2054 * by considering the behavior of a buddy system containing a single
2055 * large block of memory acted on by a series of small allocations.
2056 * This behavior is a critical factor in sglist merging's success.
2060 static inline void expand(struct zone *zone, struct page *page,
2061 int low, int high, int migratetype)
2063 unsigned long size = 1 << high;
2065 while (high > low) {
2068 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2071 * Mark as guard pages (or page), that will allow to
2072 * merge back to allocator when buddy will be freed.
2073 * Corresponding page table entries will not be touched,
2074 * pages will stay not present in virtual address space
2076 if (set_page_guard(zone, &page[size], high, migratetype))
2079 add_to_free_list(&page[size], zone, high, migratetype);
2080 set_page_order(&page[size], high);
2084 static void check_new_page_bad(struct page *page)
2086 const char *bad_reason = NULL;
2088 if (unlikely(page->flags & __PG_HWPOISON)) {
2089 /* Don't complain about hwpoisoned pages */
2090 page_mapcount_reset(page); /* remove PageBuddy */
2093 if (unlikely(atomic_read(&page->_mapcount) != -1))
2094 bad_reason = "nonzero mapcount";
2095 if (unlikely(page->mapping != NULL))
2096 bad_reason = "non-NULL mapping";
2097 if (unlikely(page_ref_count(page) != 0))
2098 bad_reason = "nonzero _refcount";
2099 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP))
2100 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
2102 if (unlikely(page->mem_cgroup))
2103 bad_reason = "page still charged to cgroup";
2105 bad_page(page, bad_reason);
2109 * This page is about to be returned from the page allocator
2111 static inline int check_new_page(struct page *page)
2113 if (likely(page_expected_state(page,
2114 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2117 check_new_page_bad(page);
2121 static inline bool free_pages_prezeroed(void)
2123 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2124 page_poisoning_enabled()) || want_init_on_free();
2127 #ifdef CONFIG_DEBUG_VM
2129 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2130 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2131 * also checked when pcp lists are refilled from the free lists.
2133 static inline bool check_pcp_refill(struct page *page)
2135 if (debug_pagealloc_enabled_static())
2136 return check_new_page(page);
2141 static inline bool check_new_pcp(struct page *page)
2143 return check_new_page(page);
2147 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2148 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2149 * enabled, they are also checked when being allocated from the pcp lists.
2151 static inline bool check_pcp_refill(struct page *page)
2153 return check_new_page(page);
2155 static inline bool check_new_pcp(struct page *page)
2157 if (debug_pagealloc_enabled_static())
2158 return check_new_page(page);
2162 #endif /* CONFIG_DEBUG_VM */
2164 static bool check_new_pages(struct page *page, unsigned int order)
2167 for (i = 0; i < (1 << order); i++) {
2168 struct page *p = page + i;
2170 if (unlikely(check_new_page(p)))
2177 inline void post_alloc_hook(struct page *page, unsigned int order,
2180 set_page_private(page, 0);
2181 set_page_refcounted(page);
2183 arch_alloc_page(page, order);
2184 if (debug_pagealloc_enabled_static())
2185 kernel_map_pages(page, 1 << order, 1);
2186 kasan_alloc_pages(page, order);
2187 kernel_poison_pages(page, 1 << order, 1);
2188 set_page_owner(page, order, gfp_flags);
2191 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2192 unsigned int alloc_flags)
2194 post_alloc_hook(page, order, gfp_flags);
2196 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2197 kernel_init_free_pages(page, 1 << order);
2199 if (order && (gfp_flags & __GFP_COMP))
2200 prep_compound_page(page, order);
2203 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2204 * allocate the page. The expectation is that the caller is taking
2205 * steps that will free more memory. The caller should avoid the page
2206 * being used for !PFMEMALLOC purposes.
2208 if (alloc_flags & ALLOC_NO_WATERMARKS)
2209 set_page_pfmemalloc(page);
2211 clear_page_pfmemalloc(page);
2215 * Go through the free lists for the given migratetype and remove
2216 * the smallest available page from the freelists
2218 static __always_inline
2219 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2222 unsigned int current_order;
2223 struct free_area *area;
2226 /* Find a page of the appropriate size in the preferred list */
2227 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2228 area = &(zone->free_area[current_order]);
2229 page = get_page_from_free_area(area, migratetype);
2232 del_page_from_free_list(page, zone, current_order);
2233 expand(zone, page, order, current_order, migratetype);
2234 set_pcppage_migratetype(page, migratetype);
2243 * This array describes the order lists are fallen back to when
2244 * the free lists for the desirable migrate type are depleted
2246 static int fallbacks[MIGRATE_TYPES][4] = {
2247 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2248 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2249 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2251 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2253 #ifdef CONFIG_MEMORY_ISOLATION
2254 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2259 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2262 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2265 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2266 unsigned int order) { return NULL; }
2270 * Move the free pages in a range to the free lists of the requested type.
2271 * Note that start_page and end_pages are not aligned on a pageblock
2272 * boundary. If alignment is required, use move_freepages_block()
2274 static int move_freepages(struct zone *zone,
2275 struct page *start_page, struct page *end_page,
2276 int migratetype, int *num_movable)
2280 int pages_moved = 0;
2282 for (page = start_page; page <= end_page;) {
2283 if (!pfn_valid_within(page_to_pfn(page))) {
2288 if (!PageBuddy(page)) {
2290 * We assume that pages that could be isolated for
2291 * migration are movable. But we don't actually try
2292 * isolating, as that would be expensive.
2295 (PageLRU(page) || __PageMovable(page)))
2302 /* Make sure we are not inadvertently changing nodes */
2303 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2304 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2306 order = page_order(page);
2307 move_to_free_list(page, zone, order, migratetype);
2309 pages_moved += 1 << order;
2315 int move_freepages_block(struct zone *zone, struct page *page,
2316 int migratetype, int *num_movable)
2318 unsigned long start_pfn, end_pfn;
2319 struct page *start_page, *end_page;
2324 start_pfn = page_to_pfn(page);
2325 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2326 start_page = pfn_to_page(start_pfn);
2327 end_page = start_page + pageblock_nr_pages - 1;
2328 end_pfn = start_pfn + pageblock_nr_pages - 1;
2330 /* Do not cross zone boundaries */
2331 if (!zone_spans_pfn(zone, start_pfn))
2333 if (!zone_spans_pfn(zone, end_pfn))
2336 return move_freepages(zone, start_page, end_page, migratetype,
2340 static void change_pageblock_range(struct page *pageblock_page,
2341 int start_order, int migratetype)
2343 int nr_pageblocks = 1 << (start_order - pageblock_order);
2345 while (nr_pageblocks--) {
2346 set_pageblock_migratetype(pageblock_page, migratetype);
2347 pageblock_page += pageblock_nr_pages;
2352 * When we are falling back to another migratetype during allocation, try to
2353 * steal extra free pages from the same pageblocks to satisfy further
2354 * allocations, instead of polluting multiple pageblocks.
2356 * If we are stealing a relatively large buddy page, it is likely there will
2357 * be more free pages in the pageblock, so try to steal them all. For
2358 * reclaimable and unmovable allocations, we steal regardless of page size,
2359 * as fragmentation caused by those allocations polluting movable pageblocks
2360 * is worse than movable allocations stealing from unmovable and reclaimable
2363 static bool can_steal_fallback(unsigned int order, int start_mt)
2366 * Leaving this order check is intended, although there is
2367 * relaxed order check in next check. The reason is that
2368 * we can actually steal whole pageblock if this condition met,
2369 * but, below check doesn't guarantee it and that is just heuristic
2370 * so could be changed anytime.
2372 if (order >= pageblock_order)
2375 if (order >= pageblock_order / 2 ||
2376 start_mt == MIGRATE_RECLAIMABLE ||
2377 start_mt == MIGRATE_UNMOVABLE ||
2378 page_group_by_mobility_disabled)
2384 static inline void boost_watermark(struct zone *zone)
2386 unsigned long max_boost;
2388 if (!watermark_boost_factor)
2391 * Don't bother in zones that are unlikely to produce results.
2392 * On small machines, including kdump capture kernels running
2393 * in a small area, boosting the watermark can cause an out of
2394 * memory situation immediately.
2396 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2399 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2400 watermark_boost_factor, 10000);
2403 * high watermark may be uninitialised if fragmentation occurs
2404 * very early in boot so do not boost. We do not fall
2405 * through and boost by pageblock_nr_pages as failing
2406 * allocations that early means that reclaim is not going
2407 * to help and it may even be impossible to reclaim the
2408 * boosted watermark resulting in a hang.
2413 max_boost = max(pageblock_nr_pages, max_boost);
2415 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2420 * This function implements actual steal behaviour. If order is large enough,
2421 * we can steal whole pageblock. If not, we first move freepages in this
2422 * pageblock to our migratetype and determine how many already-allocated pages
2423 * are there in the pageblock with a compatible migratetype. If at least half
2424 * of pages are free or compatible, we can change migratetype of the pageblock
2425 * itself, so pages freed in the future will be put on the correct free list.
2427 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2428 unsigned int alloc_flags, int start_type, bool whole_block)
2430 unsigned int current_order = page_order(page);
2431 int free_pages, movable_pages, alike_pages;
2434 old_block_type = get_pageblock_migratetype(page);
2437 * This can happen due to races and we want to prevent broken
2438 * highatomic accounting.
2440 if (is_migrate_highatomic(old_block_type))
2443 /* Take ownership for orders >= pageblock_order */
2444 if (current_order >= pageblock_order) {
2445 change_pageblock_range(page, current_order, start_type);
2450 * Boost watermarks to increase reclaim pressure to reduce the
2451 * likelihood of future fallbacks. Wake kswapd now as the node
2452 * may be balanced overall and kswapd will not wake naturally.
2454 boost_watermark(zone);
2455 if (alloc_flags & ALLOC_KSWAPD)
2456 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2458 /* We are not allowed to try stealing from the whole block */
2462 free_pages = move_freepages_block(zone, page, start_type,
2465 * Determine how many pages are compatible with our allocation.
2466 * For movable allocation, it's the number of movable pages which
2467 * we just obtained. For other types it's a bit more tricky.
2469 if (start_type == MIGRATE_MOVABLE) {
2470 alike_pages = movable_pages;
2473 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2474 * to MOVABLE pageblock, consider all non-movable pages as
2475 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2476 * vice versa, be conservative since we can't distinguish the
2477 * exact migratetype of non-movable pages.
2479 if (old_block_type == MIGRATE_MOVABLE)
2480 alike_pages = pageblock_nr_pages
2481 - (free_pages + movable_pages);
2486 /* moving whole block can fail due to zone boundary conditions */
2491 * If a sufficient number of pages in the block are either free or of
2492 * comparable migratability as our allocation, claim the whole block.
2494 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2495 page_group_by_mobility_disabled)
2496 set_pageblock_migratetype(page, start_type);
2501 move_to_free_list(page, zone, current_order, start_type);
2505 * Check whether there is a suitable fallback freepage with requested order.
2506 * If only_stealable is true, this function returns fallback_mt only if
2507 * we can steal other freepages all together. This would help to reduce
2508 * fragmentation due to mixed migratetype pages in one pageblock.
2510 int find_suitable_fallback(struct free_area *area, unsigned int order,
2511 int migratetype, bool only_stealable, bool *can_steal)
2516 if (area->nr_free == 0)
2521 fallback_mt = fallbacks[migratetype][i];
2522 if (fallback_mt == MIGRATE_TYPES)
2525 if (free_area_empty(area, fallback_mt))
2528 if (can_steal_fallback(order, migratetype))
2531 if (!only_stealable)
2542 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2543 * there are no empty page blocks that contain a page with a suitable order
2545 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2546 unsigned int alloc_order)
2549 unsigned long max_managed, flags;
2552 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2553 * Check is race-prone but harmless.
2555 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2556 if (zone->nr_reserved_highatomic >= max_managed)
2559 spin_lock_irqsave(&zone->lock, flags);
2561 /* Recheck the nr_reserved_highatomic limit under the lock */
2562 if (zone->nr_reserved_highatomic >= max_managed)
2566 mt = get_pageblock_migratetype(page);
2567 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2568 && !is_migrate_cma(mt)) {
2569 zone->nr_reserved_highatomic += pageblock_nr_pages;
2570 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2571 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2575 spin_unlock_irqrestore(&zone->lock, flags);
2579 * Used when an allocation is about to fail under memory pressure. This
2580 * potentially hurts the reliability of high-order allocations when under
2581 * intense memory pressure but failed atomic allocations should be easier
2582 * to recover from than an OOM.
2584 * If @force is true, try to unreserve a pageblock even though highatomic
2585 * pageblock is exhausted.
2587 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2590 struct zonelist *zonelist = ac->zonelist;
2591 unsigned long flags;
2598 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2601 * Preserve at least one pageblock unless memory pressure
2604 if (!force && zone->nr_reserved_highatomic <=
2608 spin_lock_irqsave(&zone->lock, flags);
2609 for (order = 0; order < MAX_ORDER; order++) {
2610 struct free_area *area = &(zone->free_area[order]);
2612 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2617 * In page freeing path, migratetype change is racy so
2618 * we can counter several free pages in a pageblock
2619 * in this loop althoug we changed the pageblock type
2620 * from highatomic to ac->migratetype. So we should
2621 * adjust the count once.
2623 if (is_migrate_highatomic_page(page)) {
2625 * It should never happen but changes to
2626 * locking could inadvertently allow a per-cpu
2627 * drain to add pages to MIGRATE_HIGHATOMIC
2628 * while unreserving so be safe and watch for
2631 zone->nr_reserved_highatomic -= min(
2633 zone->nr_reserved_highatomic);
2637 * Convert to ac->migratetype and avoid the normal
2638 * pageblock stealing heuristics. Minimally, the caller
2639 * is doing the work and needs the pages. More
2640 * importantly, if the block was always converted to
2641 * MIGRATE_UNMOVABLE or another type then the number
2642 * of pageblocks that cannot be completely freed
2645 set_pageblock_migratetype(page, ac->migratetype);
2646 ret = move_freepages_block(zone, page, ac->migratetype,
2649 spin_unlock_irqrestore(&zone->lock, flags);
2653 spin_unlock_irqrestore(&zone->lock, flags);
2660 * Try finding a free buddy page on the fallback list and put it on the free
2661 * list of requested migratetype, possibly along with other pages from the same
2662 * block, depending on fragmentation avoidance heuristics. Returns true if
2663 * fallback was found so that __rmqueue_smallest() can grab it.
2665 * The use of signed ints for order and current_order is a deliberate
2666 * deviation from the rest of this file, to make the for loop
2667 * condition simpler.
2669 static __always_inline bool
2670 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2671 unsigned int alloc_flags)
2673 struct free_area *area;
2675 int min_order = order;
2681 * Do not steal pages from freelists belonging to other pageblocks
2682 * i.e. orders < pageblock_order. If there are no local zones free,
2683 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2685 if (alloc_flags & ALLOC_NOFRAGMENT)
2686 min_order = pageblock_order;
2689 * Find the largest available free page in the other list. This roughly
2690 * approximates finding the pageblock with the most free pages, which
2691 * would be too costly to do exactly.
2693 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2695 area = &(zone->free_area[current_order]);
2696 fallback_mt = find_suitable_fallback(area, current_order,
2697 start_migratetype, false, &can_steal);
2698 if (fallback_mt == -1)
2702 * We cannot steal all free pages from the pageblock and the
2703 * requested migratetype is movable. In that case it's better to
2704 * steal and split the smallest available page instead of the
2705 * largest available page, because even if the next movable
2706 * allocation falls back into a different pageblock than this
2707 * one, it won't cause permanent fragmentation.
2709 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2710 && current_order > order)
2719 for (current_order = order; current_order < MAX_ORDER;
2721 area = &(zone->free_area[current_order]);
2722 fallback_mt = find_suitable_fallback(area, current_order,
2723 start_migratetype, false, &can_steal);
2724 if (fallback_mt != -1)
2729 * This should not happen - we already found a suitable fallback
2730 * when looking for the largest page.
2732 VM_BUG_ON(current_order == MAX_ORDER);
2735 page = get_page_from_free_area(area, fallback_mt);
2737 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2740 trace_mm_page_alloc_extfrag(page, order, current_order,
2741 start_migratetype, fallback_mt);
2748 * Do the hard work of removing an element from the buddy allocator.
2749 * Call me with the zone->lock already held.
2751 static __always_inline struct page *
2752 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2753 unsigned int alloc_flags)
2758 page = __rmqueue_smallest(zone, order, migratetype);
2759 if (unlikely(!page)) {
2760 if (migratetype == MIGRATE_MOVABLE)
2761 page = __rmqueue_cma_fallback(zone, order);
2763 if (!page && __rmqueue_fallback(zone, order, migratetype,
2768 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2773 * Obtain a specified number of elements from the buddy allocator, all under
2774 * a single hold of the lock, for efficiency. Add them to the supplied list.
2775 * Returns the number of new pages which were placed at *list.
2777 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2778 unsigned long count, struct list_head *list,
2779 int migratetype, unsigned int alloc_flags)
2783 spin_lock(&zone->lock);
2784 for (i = 0; i < count; ++i) {
2785 struct page *page = __rmqueue(zone, order, migratetype,
2787 if (unlikely(page == NULL))
2790 if (unlikely(check_pcp_refill(page)))
2794 * Split buddy pages returned by expand() are received here in
2795 * physical page order. The page is added to the tail of
2796 * caller's list. From the callers perspective, the linked list
2797 * is ordered by page number under some conditions. This is
2798 * useful for IO devices that can forward direction from the
2799 * head, thus also in the physical page order. This is useful
2800 * for IO devices that can merge IO requests if the physical
2801 * pages are ordered properly.
2803 list_add_tail(&page->lru, list);
2805 if (is_migrate_cma(get_pcppage_migratetype(page)))
2806 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2811 * i pages were removed from the buddy list even if some leak due
2812 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2813 * on i. Do not confuse with 'alloced' which is the number of
2814 * pages added to the pcp list.
2816 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2817 spin_unlock(&zone->lock);
2823 * Called from the vmstat counter updater to drain pagesets of this
2824 * currently executing processor on remote nodes after they have
2827 * Note that this function must be called with the thread pinned to
2828 * a single processor.
2830 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2832 unsigned long flags;
2833 int to_drain, batch;
2835 local_irq_save(flags);
2836 batch = READ_ONCE(pcp->batch);
2837 to_drain = min(pcp->count, batch);
2839 free_pcppages_bulk(zone, to_drain, pcp);
2840 local_irq_restore(flags);
2845 * Drain pcplists of the indicated processor and zone.
2847 * The processor must either be the current processor and the
2848 * thread pinned to the current processor or a processor that
2851 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2853 unsigned long flags;
2854 struct per_cpu_pageset *pset;
2855 struct per_cpu_pages *pcp;
2857 local_irq_save(flags);
2858 pset = per_cpu_ptr(zone->pageset, cpu);
2862 free_pcppages_bulk(zone, pcp->count, pcp);
2863 local_irq_restore(flags);
2867 * Drain pcplists of all zones on the indicated processor.
2869 * The processor must either be the current processor and the
2870 * thread pinned to the current processor or a processor that
2873 static void drain_pages(unsigned int cpu)
2877 for_each_populated_zone(zone) {
2878 drain_pages_zone(cpu, zone);
2883 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2885 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2886 * the single zone's pages.
2888 void drain_local_pages(struct zone *zone)
2890 int cpu = smp_processor_id();
2893 drain_pages_zone(cpu, zone);
2898 static void drain_local_pages_wq(struct work_struct *work)
2900 struct pcpu_drain *drain;
2902 drain = container_of(work, struct pcpu_drain, work);
2905 * drain_all_pages doesn't use proper cpu hotplug protection so
2906 * we can race with cpu offline when the WQ can move this from
2907 * a cpu pinned worker to an unbound one. We can operate on a different
2908 * cpu which is allright but we also have to make sure to not move to
2912 drain_local_pages(drain->zone);
2917 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2919 * When zone parameter is non-NULL, spill just the single zone's pages.
2921 * Note that this can be extremely slow as the draining happens in a workqueue.
2923 void drain_all_pages(struct zone *zone)
2928 * Allocate in the BSS so we wont require allocation in
2929 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2931 static cpumask_t cpus_with_pcps;
2934 * Make sure nobody triggers this path before mm_percpu_wq is fully
2937 if (WARN_ON_ONCE(!mm_percpu_wq))
2941 * Do not drain if one is already in progress unless it's specific to
2942 * a zone. Such callers are primarily CMA and memory hotplug and need
2943 * the drain to be complete when the call returns.
2945 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2948 mutex_lock(&pcpu_drain_mutex);
2952 * We don't care about racing with CPU hotplug event
2953 * as offline notification will cause the notified
2954 * cpu to drain that CPU pcps and on_each_cpu_mask
2955 * disables preemption as part of its processing
2957 for_each_online_cpu(cpu) {
2958 struct per_cpu_pageset *pcp;
2960 bool has_pcps = false;
2963 pcp = per_cpu_ptr(zone->pageset, cpu);
2967 for_each_populated_zone(z) {
2968 pcp = per_cpu_ptr(z->pageset, cpu);
2969 if (pcp->pcp.count) {
2977 cpumask_set_cpu(cpu, &cpus_with_pcps);
2979 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2982 for_each_cpu(cpu, &cpus_with_pcps) {
2983 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
2986 INIT_WORK(&drain->work, drain_local_pages_wq);
2987 queue_work_on(cpu, mm_percpu_wq, &drain->work);
2989 for_each_cpu(cpu, &cpus_with_pcps)
2990 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
2992 mutex_unlock(&pcpu_drain_mutex);
2995 #ifdef CONFIG_HIBERNATION
2998 * Touch the watchdog for every WD_PAGE_COUNT pages.
3000 #define WD_PAGE_COUNT (128*1024)
3002 void mark_free_pages(struct zone *zone)
3004 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3005 unsigned long flags;
3006 unsigned int order, t;
3009 if (zone_is_empty(zone))
3012 spin_lock_irqsave(&zone->lock, flags);
3014 max_zone_pfn = zone_end_pfn(zone);
3015 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3016 if (pfn_valid(pfn)) {
3017 page = pfn_to_page(pfn);
3019 if (!--page_count) {
3020 touch_nmi_watchdog();
3021 page_count = WD_PAGE_COUNT;
3024 if (page_zone(page) != zone)
3027 if (!swsusp_page_is_forbidden(page))
3028 swsusp_unset_page_free(page);
3031 for_each_migratetype_order(order, t) {
3032 list_for_each_entry(page,
3033 &zone->free_area[order].free_list[t], lru) {
3036 pfn = page_to_pfn(page);
3037 for (i = 0; i < (1UL << order); i++) {
3038 if (!--page_count) {
3039 touch_nmi_watchdog();
3040 page_count = WD_PAGE_COUNT;
3042 swsusp_set_page_free(pfn_to_page(pfn + i));
3046 spin_unlock_irqrestore(&zone->lock, flags);
3048 #endif /* CONFIG_PM */
3050 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3054 if (!free_pcp_prepare(page))
3057 migratetype = get_pfnblock_migratetype(page, pfn);
3058 set_pcppage_migratetype(page, migratetype);
3062 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3064 struct zone *zone = page_zone(page);
3065 struct per_cpu_pages *pcp;
3068 migratetype = get_pcppage_migratetype(page);
3069 __count_vm_event(PGFREE);
3072 * We only track unmovable, reclaimable and movable on pcp lists.
3073 * Free ISOLATE pages back to the allocator because they are being
3074 * offlined but treat HIGHATOMIC as movable pages so we can get those
3075 * areas back if necessary. Otherwise, we may have to free
3076 * excessively into the page allocator
3078 if (migratetype >= MIGRATE_PCPTYPES) {
3079 if (unlikely(is_migrate_isolate(migratetype))) {
3080 free_one_page(zone, page, pfn, 0, migratetype);
3083 migratetype = MIGRATE_MOVABLE;
3086 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3087 list_add(&page->lru, &pcp->lists[migratetype]);
3089 if (pcp->count >= pcp->high) {
3090 unsigned long batch = READ_ONCE(pcp->batch);
3091 free_pcppages_bulk(zone, batch, pcp);
3096 * Free a 0-order page
3098 void free_unref_page(struct page *page)
3100 unsigned long flags;
3101 unsigned long pfn = page_to_pfn(page);
3103 if (!free_unref_page_prepare(page, pfn))
3106 local_irq_save(flags);
3107 free_unref_page_commit(page, pfn);
3108 local_irq_restore(flags);
3112 * Free a list of 0-order pages
3114 void free_unref_page_list(struct list_head *list)
3116 struct page *page, *next;
3117 unsigned long flags, pfn;
3118 int batch_count = 0;
3120 /* Prepare pages for freeing */
3121 list_for_each_entry_safe(page, next, list, lru) {
3122 pfn = page_to_pfn(page);
3123 if (!free_unref_page_prepare(page, pfn))
3124 list_del(&page->lru);
3125 set_page_private(page, pfn);
3128 local_irq_save(flags);
3129 list_for_each_entry_safe(page, next, list, lru) {
3130 unsigned long pfn = page_private(page);
3132 set_page_private(page, 0);
3133 trace_mm_page_free_batched(page);
3134 free_unref_page_commit(page, pfn);
3137 * Guard against excessive IRQ disabled times when we get
3138 * a large list of pages to free.
3140 if (++batch_count == SWAP_CLUSTER_MAX) {
3141 local_irq_restore(flags);
3143 local_irq_save(flags);
3146 local_irq_restore(flags);
3150 * split_page takes a non-compound higher-order page, and splits it into
3151 * n (1<<order) sub-pages: page[0..n]
3152 * Each sub-page must be freed individually.
3154 * Note: this is probably too low level an operation for use in drivers.
3155 * Please consult with lkml before using this in your driver.
3157 void split_page(struct page *page, unsigned int order)
3161 VM_BUG_ON_PAGE(PageCompound(page), page);
3162 VM_BUG_ON_PAGE(!page_count(page), page);
3164 for (i = 1; i < (1 << order); i++)
3165 set_page_refcounted(page + i);
3166 split_page_owner(page, order);
3168 EXPORT_SYMBOL_GPL(split_page);
3170 int __isolate_free_page(struct page *page, unsigned int order)
3172 unsigned long watermark;
3176 BUG_ON(!PageBuddy(page));
3178 zone = page_zone(page);
3179 mt = get_pageblock_migratetype(page);
3181 if (!is_migrate_isolate(mt)) {
3183 * Obey watermarks as if the page was being allocated. We can
3184 * emulate a high-order watermark check with a raised order-0
3185 * watermark, because we already know our high-order page
3188 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3189 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3192 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3195 /* Remove page from free list */
3197 del_page_from_free_list(page, zone, order);
3200 * Set the pageblock if the isolated page is at least half of a
3203 if (order >= pageblock_order - 1) {
3204 struct page *endpage = page + (1 << order) - 1;
3205 for (; page < endpage; page += pageblock_nr_pages) {
3206 int mt = get_pageblock_migratetype(page);
3207 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3208 && !is_migrate_highatomic(mt))
3209 set_pageblock_migratetype(page,
3215 return 1UL << order;
3219 * __putback_isolated_page - Return a now-isolated page back where we got it
3220 * @page: Page that was isolated
3221 * @order: Order of the isolated page
3222 * @mt: The page's pageblock's migratetype
3224 * This function is meant to return a page pulled from the free lists via
3225 * __isolate_free_page back to the free lists they were pulled from.
3227 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3229 struct zone *zone = page_zone(page);
3231 /* zone lock should be held when this function is called */
3232 lockdep_assert_held(&zone->lock);
3234 /* Return isolated page to tail of freelist. */
3235 __free_one_page(page, page_to_pfn(page), zone, order, mt, false);
3239 * Update NUMA hit/miss statistics
3241 * Must be called with interrupts disabled.
3243 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3246 enum numa_stat_item local_stat = NUMA_LOCAL;
3248 /* skip numa counters update if numa stats is disabled */
3249 if (!static_branch_likely(&vm_numa_stat_key))
3252 if (zone_to_nid(z) != numa_node_id())
3253 local_stat = NUMA_OTHER;
3255 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3256 __inc_numa_state(z, NUMA_HIT);
3258 __inc_numa_state(z, NUMA_MISS);
3259 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3261 __inc_numa_state(z, local_stat);
3265 /* Remove page from the per-cpu list, caller must protect the list */
3266 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3267 unsigned int alloc_flags,
3268 struct per_cpu_pages *pcp,
3269 struct list_head *list)
3274 if (list_empty(list)) {
3275 pcp->count += rmqueue_bulk(zone, 0,
3277 migratetype, alloc_flags);
3278 if (unlikely(list_empty(list)))
3282 page = list_first_entry(list, struct page, lru);
3283 list_del(&page->lru);
3285 } while (check_new_pcp(page));
3290 /* Lock and remove page from the per-cpu list */
3291 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3292 struct zone *zone, gfp_t gfp_flags,
3293 int migratetype, unsigned int alloc_flags)
3295 struct per_cpu_pages *pcp;
3296 struct list_head *list;
3298 unsigned long flags;
3300 local_irq_save(flags);
3301 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3302 list = &pcp->lists[migratetype];
3303 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3305 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3306 zone_statistics(preferred_zone, zone);
3308 local_irq_restore(flags);
3313 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3316 struct page *rmqueue(struct zone *preferred_zone,
3317 struct zone *zone, unsigned int order,
3318 gfp_t gfp_flags, unsigned int alloc_flags,
3321 unsigned long flags;
3324 if (likely(order == 0)) {
3325 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3326 migratetype, alloc_flags);
3331 * We most definitely don't want callers attempting to
3332 * allocate greater than order-1 page units with __GFP_NOFAIL.
3334 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3335 spin_lock_irqsave(&zone->lock, flags);
3339 if (alloc_flags & ALLOC_HARDER) {
3340 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3342 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3345 page = __rmqueue(zone, order, migratetype, alloc_flags);
3346 } while (page && check_new_pages(page, order));
3347 spin_unlock(&zone->lock);
3350 __mod_zone_freepage_state(zone, -(1 << order),
3351 get_pcppage_migratetype(page));
3353 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3354 zone_statistics(preferred_zone, zone);
3355 local_irq_restore(flags);
3358 /* Separate test+clear to avoid unnecessary atomics */
3359 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3360 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3361 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3364 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3368 local_irq_restore(flags);
3372 #ifdef CONFIG_FAIL_PAGE_ALLOC
3375 struct fault_attr attr;
3377 bool ignore_gfp_highmem;
3378 bool ignore_gfp_reclaim;
3380 } fail_page_alloc = {
3381 .attr = FAULT_ATTR_INITIALIZER,
3382 .ignore_gfp_reclaim = true,
3383 .ignore_gfp_highmem = true,
3387 static int __init setup_fail_page_alloc(char *str)
3389 return setup_fault_attr(&fail_page_alloc.attr, str);
3391 __setup("fail_page_alloc=", setup_fail_page_alloc);
3393 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3395 if (order < fail_page_alloc.min_order)
3397 if (gfp_mask & __GFP_NOFAIL)
3399 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3401 if (fail_page_alloc.ignore_gfp_reclaim &&
3402 (gfp_mask & __GFP_DIRECT_RECLAIM))
3405 return should_fail(&fail_page_alloc.attr, 1 << order);
3408 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3410 static int __init fail_page_alloc_debugfs(void)
3412 umode_t mode = S_IFREG | 0600;
3415 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3416 &fail_page_alloc.attr);
3418 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3419 &fail_page_alloc.ignore_gfp_reclaim);
3420 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3421 &fail_page_alloc.ignore_gfp_highmem);
3422 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3427 late_initcall(fail_page_alloc_debugfs);
3429 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3431 #else /* CONFIG_FAIL_PAGE_ALLOC */
3433 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3438 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3440 static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3442 return __should_fail_alloc_page(gfp_mask, order);
3444 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3447 * Return true if free base pages are above 'mark'. For high-order checks it
3448 * will return true of the order-0 watermark is reached and there is at least
3449 * one free page of a suitable size. Checking now avoids taking the zone lock
3450 * to check in the allocation paths if no pages are free.
3452 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3453 int classzone_idx, unsigned int alloc_flags,
3458 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3460 /* free_pages may go negative - that's OK */
3461 free_pages -= (1 << order) - 1;
3463 if (alloc_flags & ALLOC_HIGH)
3467 * If the caller does not have rights to ALLOC_HARDER then subtract
3468 * the high-atomic reserves. This will over-estimate the size of the
3469 * atomic reserve but it avoids a search.
3471 if (likely(!alloc_harder)) {
3472 free_pages -= z->nr_reserved_highatomic;
3475 * OOM victims can try even harder than normal ALLOC_HARDER
3476 * users on the grounds that it's definitely going to be in
3477 * the exit path shortly and free memory. Any allocation it
3478 * makes during the free path will be small and short-lived.
3480 if (alloc_flags & ALLOC_OOM)
3488 /* If allocation can't use CMA areas don't use free CMA pages */
3489 if (!(alloc_flags & ALLOC_CMA))
3490 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3494 * Check watermarks for an order-0 allocation request. If these
3495 * are not met, then a high-order request also cannot go ahead
3496 * even if a suitable page happened to be free.
3498 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3501 /* If this is an order-0 request then the watermark is fine */
3505 /* For a high-order request, check at least one suitable page is free */
3506 for (o = order; o < MAX_ORDER; o++) {
3507 struct free_area *area = &z->free_area[o];
3513 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3514 if (!free_area_empty(area, mt))
3519 if ((alloc_flags & ALLOC_CMA) &&
3520 !free_area_empty(area, MIGRATE_CMA)) {
3524 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3530 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3531 int classzone_idx, unsigned int alloc_flags)
3533 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3534 zone_page_state(z, NR_FREE_PAGES));
3537 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3538 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3540 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3544 /* If allocation can't use CMA areas don't use free CMA pages */
3545 if (!(alloc_flags & ALLOC_CMA))
3546 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3550 * Fast check for order-0 only. If this fails then the reserves
3551 * need to be calculated. There is a corner case where the check
3552 * passes but only the high-order atomic reserve are free. If
3553 * the caller is !atomic then it'll uselessly search the free
3554 * list. That corner case is then slower but it is harmless.
3556 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3559 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3563 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3564 unsigned long mark, int classzone_idx)
3566 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3568 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3569 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3571 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3576 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3578 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3579 node_reclaim_distance;
3581 #else /* CONFIG_NUMA */
3582 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3586 #endif /* CONFIG_NUMA */
3589 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3590 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3591 * premature use of a lower zone may cause lowmem pressure problems that
3592 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3593 * probably too small. It only makes sense to spread allocations to avoid
3594 * fragmentation between the Normal and DMA32 zones.
3596 static inline unsigned int
3597 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3599 unsigned int alloc_flags;
3602 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3605 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3607 #ifdef CONFIG_ZONE_DMA32
3611 if (zone_idx(zone) != ZONE_NORMAL)
3615 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3616 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3617 * on UMA that if Normal is populated then so is DMA32.
3619 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3620 if (nr_online_nodes > 1 && !populated_zone(--zone))
3623 alloc_flags |= ALLOC_NOFRAGMENT;
3624 #endif /* CONFIG_ZONE_DMA32 */
3629 * get_page_from_freelist goes through the zonelist trying to allocate
3632 static struct page *
3633 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3634 const struct alloc_context *ac)
3638 struct pglist_data *last_pgdat_dirty_limit = NULL;
3643 * Scan zonelist, looking for a zone with enough free.
3644 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3646 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3647 z = ac->preferred_zoneref;
3648 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3653 if (cpusets_enabled() &&
3654 (alloc_flags & ALLOC_CPUSET) &&
3655 !__cpuset_zone_allowed(zone, gfp_mask))
3658 * When allocating a page cache page for writing, we
3659 * want to get it from a node that is within its dirty
3660 * limit, such that no single node holds more than its
3661 * proportional share of globally allowed dirty pages.
3662 * The dirty limits take into account the node's
3663 * lowmem reserves and high watermark so that kswapd
3664 * should be able to balance it without having to
3665 * write pages from its LRU list.
3667 * XXX: For now, allow allocations to potentially
3668 * exceed the per-node dirty limit in the slowpath
3669 * (spread_dirty_pages unset) before going into reclaim,
3670 * which is important when on a NUMA setup the allowed
3671 * nodes are together not big enough to reach the
3672 * global limit. The proper fix for these situations
3673 * will require awareness of nodes in the
3674 * dirty-throttling and the flusher threads.
3676 if (ac->spread_dirty_pages) {
3677 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3680 if (!node_dirty_ok(zone->zone_pgdat)) {
3681 last_pgdat_dirty_limit = zone->zone_pgdat;
3686 if (no_fallback && nr_online_nodes > 1 &&
3687 zone != ac->preferred_zoneref->zone) {
3691 * If moving to a remote node, retry but allow
3692 * fragmenting fallbacks. Locality is more important
3693 * than fragmentation avoidance.
3695 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3696 if (zone_to_nid(zone) != local_nid) {
3697 alloc_flags &= ~ALLOC_NOFRAGMENT;
3702 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3703 if (!zone_watermark_fast(zone, order, mark,
3704 ac_classzone_idx(ac), alloc_flags)) {
3707 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3709 * Watermark failed for this zone, but see if we can
3710 * grow this zone if it contains deferred pages.
3712 if (static_branch_unlikely(&deferred_pages)) {
3713 if (_deferred_grow_zone(zone, order))
3717 /* Checked here to keep the fast path fast */
3718 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3719 if (alloc_flags & ALLOC_NO_WATERMARKS)
3722 if (node_reclaim_mode == 0 ||
3723 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3726 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3728 case NODE_RECLAIM_NOSCAN:
3731 case NODE_RECLAIM_FULL:
3732 /* scanned but unreclaimable */
3735 /* did we reclaim enough */
3736 if (zone_watermark_ok(zone, order, mark,
3737 ac_classzone_idx(ac), alloc_flags))
3745 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3746 gfp_mask, alloc_flags, ac->migratetype);
3748 prep_new_page(page, order, gfp_mask, alloc_flags);
3751 * If this is a high-order atomic allocation then check
3752 * if the pageblock should be reserved for the future
3754 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3755 reserve_highatomic_pageblock(page, zone, order);
3759 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3760 /* Try again if zone has deferred pages */
3761 if (static_branch_unlikely(&deferred_pages)) {
3762 if (_deferred_grow_zone(zone, order))
3770 * It's possible on a UMA machine to get through all zones that are
3771 * fragmented. If avoiding fragmentation, reset and try again.
3774 alloc_flags &= ~ALLOC_NOFRAGMENT;
3781 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3783 unsigned int filter = SHOW_MEM_FILTER_NODES;
3786 * This documents exceptions given to allocations in certain
3787 * contexts that are allowed to allocate outside current's set
3790 if (!(gfp_mask & __GFP_NOMEMALLOC))
3791 if (tsk_is_oom_victim(current) ||
3792 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3793 filter &= ~SHOW_MEM_FILTER_NODES;
3794 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3795 filter &= ~SHOW_MEM_FILTER_NODES;
3797 show_mem(filter, nodemask);
3800 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3802 struct va_format vaf;
3804 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3806 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3809 va_start(args, fmt);
3812 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3813 current->comm, &vaf, gfp_mask, &gfp_mask,
3814 nodemask_pr_args(nodemask));
3817 cpuset_print_current_mems_allowed();
3820 warn_alloc_show_mem(gfp_mask, nodemask);
3823 static inline struct page *
3824 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3825 unsigned int alloc_flags,
3826 const struct alloc_context *ac)
3830 page = get_page_from_freelist(gfp_mask, order,
3831 alloc_flags|ALLOC_CPUSET, ac);
3833 * fallback to ignore cpuset restriction if our nodes
3837 page = get_page_from_freelist(gfp_mask, order,
3843 static inline struct page *
3844 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3845 const struct alloc_context *ac, unsigned long *did_some_progress)
3847 struct oom_control oc = {
3848 .zonelist = ac->zonelist,
3849 .nodemask = ac->nodemask,
3851 .gfp_mask = gfp_mask,
3856 *did_some_progress = 0;
3859 * Acquire the oom lock. If that fails, somebody else is
3860 * making progress for us.
3862 if (!mutex_trylock(&oom_lock)) {
3863 *did_some_progress = 1;
3864 schedule_timeout_uninterruptible(1);
3869 * Go through the zonelist yet one more time, keep very high watermark
3870 * here, this is only to catch a parallel oom killing, we must fail if
3871 * we're still under heavy pressure. But make sure that this reclaim
3872 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3873 * allocation which will never fail due to oom_lock already held.
3875 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3876 ~__GFP_DIRECT_RECLAIM, order,
3877 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3881 /* Coredumps can quickly deplete all memory reserves */
3882 if (current->flags & PF_DUMPCORE)
3884 /* The OOM killer will not help higher order allocs */
3885 if (order > PAGE_ALLOC_COSTLY_ORDER)
3888 * We have already exhausted all our reclaim opportunities without any
3889 * success so it is time to admit defeat. We will skip the OOM killer
3890 * because it is very likely that the caller has a more reasonable
3891 * fallback than shooting a random task.
3893 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3895 /* The OOM killer does not needlessly kill tasks for lowmem */
3896 if (ac->high_zoneidx < ZONE_NORMAL)
3898 if (pm_suspended_storage())
3901 * XXX: GFP_NOFS allocations should rather fail than rely on
3902 * other request to make a forward progress.
3903 * We are in an unfortunate situation where out_of_memory cannot
3904 * do much for this context but let's try it to at least get
3905 * access to memory reserved if the current task is killed (see
3906 * out_of_memory). Once filesystems are ready to handle allocation
3907 * failures more gracefully we should just bail out here.
3910 /* The OOM killer may not free memory on a specific node */
3911 if (gfp_mask & __GFP_THISNODE)
3914 /* Exhausted what can be done so it's blame time */
3915 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3916 *did_some_progress = 1;
3919 * Help non-failing allocations by giving them access to memory
3922 if (gfp_mask & __GFP_NOFAIL)
3923 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3924 ALLOC_NO_WATERMARKS, ac);
3927 mutex_unlock(&oom_lock);
3932 * Maximum number of compaction retries wit a progress before OOM
3933 * killer is consider as the only way to move forward.
3935 #define MAX_COMPACT_RETRIES 16
3937 #ifdef CONFIG_COMPACTION
3938 /* Try memory compaction for high-order allocations before reclaim */
3939 static struct page *
3940 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3941 unsigned int alloc_flags, const struct alloc_context *ac,
3942 enum compact_priority prio, enum compact_result *compact_result)
3944 struct page *page = NULL;
3945 unsigned long pflags;
3946 unsigned int noreclaim_flag;
3951 psi_memstall_enter(&pflags);
3952 noreclaim_flag = memalloc_noreclaim_save();
3954 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3957 memalloc_noreclaim_restore(noreclaim_flag);
3958 psi_memstall_leave(&pflags);
3961 * At least in one zone compaction wasn't deferred or skipped, so let's
3962 * count a compaction stall
3964 count_vm_event(COMPACTSTALL);
3966 /* Prep a captured page if available */
3968 prep_new_page(page, order, gfp_mask, alloc_flags);
3970 /* Try get a page from the freelist if available */
3972 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3975 struct zone *zone = page_zone(page);
3977 zone->compact_blockskip_flush = false;
3978 compaction_defer_reset(zone, order, true);
3979 count_vm_event(COMPACTSUCCESS);
3984 * It's bad if compaction run occurs and fails. The most likely reason
3985 * is that pages exist, but not enough to satisfy watermarks.
3987 count_vm_event(COMPACTFAIL);
3995 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3996 enum compact_result compact_result,
3997 enum compact_priority *compact_priority,
3998 int *compaction_retries)
4000 int max_retries = MAX_COMPACT_RETRIES;
4003 int retries = *compaction_retries;
4004 enum compact_priority priority = *compact_priority;
4009 if (compaction_made_progress(compact_result))
4010 (*compaction_retries)++;
4013 * compaction considers all the zone as desperately out of memory
4014 * so it doesn't really make much sense to retry except when the
4015 * failure could be caused by insufficient priority
4017 if (compaction_failed(compact_result))
4018 goto check_priority;
4021 * compaction was skipped because there are not enough order-0 pages
4022 * to work with, so we retry only if it looks like reclaim can help.
4024 if (compaction_needs_reclaim(compact_result)) {
4025 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4030 * make sure the compaction wasn't deferred or didn't bail out early
4031 * due to locks contention before we declare that we should give up.
4032 * But the next retry should use a higher priority if allowed, so
4033 * we don't just keep bailing out endlessly.
4035 if (compaction_withdrawn(compact_result)) {
4036 goto check_priority;
4040 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4041 * costly ones because they are de facto nofail and invoke OOM
4042 * killer to move on while costly can fail and users are ready
4043 * to cope with that. 1/4 retries is rather arbitrary but we
4044 * would need much more detailed feedback from compaction to
4045 * make a better decision.
4047 if (order > PAGE_ALLOC_COSTLY_ORDER)
4049 if (*compaction_retries <= max_retries) {
4055 * Make sure there are attempts at the highest priority if we exhausted
4056 * all retries or failed at the lower priorities.
4059 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4060 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4062 if (*compact_priority > min_priority) {
4063 (*compact_priority)--;
4064 *compaction_retries = 0;
4068 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4072 static inline struct page *
4073 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4074 unsigned int alloc_flags, const struct alloc_context *ac,
4075 enum compact_priority prio, enum compact_result *compact_result)
4077 *compact_result = COMPACT_SKIPPED;
4082 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4083 enum compact_result compact_result,
4084 enum compact_priority *compact_priority,
4085 int *compaction_retries)
4090 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4094 * There are setups with compaction disabled which would prefer to loop
4095 * inside the allocator rather than hit the oom killer prematurely.
4096 * Let's give them a good hope and keep retrying while the order-0
4097 * watermarks are OK.
4099 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4101 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4102 ac_classzone_idx(ac), alloc_flags))
4107 #endif /* CONFIG_COMPACTION */
4109 #ifdef CONFIG_LOCKDEP
4110 static struct lockdep_map __fs_reclaim_map =
4111 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4113 static bool __need_fs_reclaim(gfp_t gfp_mask)
4115 gfp_mask = current_gfp_context(gfp_mask);
4117 /* no reclaim without waiting on it */
4118 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4121 /* this guy won't enter reclaim */
4122 if (current->flags & PF_MEMALLOC)
4125 /* We're only interested __GFP_FS allocations for now */
4126 if (!(gfp_mask & __GFP_FS))
4129 if (gfp_mask & __GFP_NOLOCKDEP)
4135 void __fs_reclaim_acquire(void)
4137 lock_map_acquire(&__fs_reclaim_map);
4140 void __fs_reclaim_release(void)
4142 lock_map_release(&__fs_reclaim_map);
4145 void fs_reclaim_acquire(gfp_t gfp_mask)
4147 if (__need_fs_reclaim(gfp_mask))
4148 __fs_reclaim_acquire();
4150 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4152 void fs_reclaim_release(gfp_t gfp_mask)
4154 if (__need_fs_reclaim(gfp_mask))
4155 __fs_reclaim_release();
4157 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4160 /* Perform direct synchronous page reclaim */
4162 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4163 const struct alloc_context *ac)
4166 unsigned int noreclaim_flag;
4167 unsigned long pflags;
4171 /* We now go into synchronous reclaim */
4172 cpuset_memory_pressure_bump();
4173 psi_memstall_enter(&pflags);
4174 fs_reclaim_acquire(gfp_mask);
4175 noreclaim_flag = memalloc_noreclaim_save();
4177 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4180 memalloc_noreclaim_restore(noreclaim_flag);
4181 fs_reclaim_release(gfp_mask);
4182 psi_memstall_leave(&pflags);
4189 /* The really slow allocator path where we enter direct reclaim */
4190 static inline struct page *
4191 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4192 unsigned int alloc_flags, const struct alloc_context *ac,
4193 unsigned long *did_some_progress)
4195 struct page *page = NULL;
4196 bool drained = false;
4198 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4199 if (unlikely(!(*did_some_progress)))
4203 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4206 * If an allocation failed after direct reclaim, it could be because
4207 * pages are pinned on the per-cpu lists or in high alloc reserves.
4208 * Shrink them them and try again
4210 if (!page && !drained) {
4211 unreserve_highatomic_pageblock(ac, false);
4212 drain_all_pages(NULL);
4220 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4221 const struct alloc_context *ac)
4225 pg_data_t *last_pgdat = NULL;
4226 enum zone_type high_zoneidx = ac->high_zoneidx;
4228 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
4230 if (last_pgdat != zone->zone_pgdat)
4231 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
4232 last_pgdat = zone->zone_pgdat;
4236 static inline unsigned int
4237 gfp_to_alloc_flags(gfp_t gfp_mask)
4239 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4242 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4243 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4244 * to save two branches.
4246 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4247 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4250 * The caller may dip into page reserves a bit more if the caller
4251 * cannot run direct reclaim, or if the caller has realtime scheduling
4252 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4253 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4255 alloc_flags |= (__force int)
4256 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4258 if (gfp_mask & __GFP_ATOMIC) {
4260 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4261 * if it can't schedule.
4263 if (!(gfp_mask & __GFP_NOMEMALLOC))
4264 alloc_flags |= ALLOC_HARDER;
4266 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4267 * comment for __cpuset_node_allowed().
4269 alloc_flags &= ~ALLOC_CPUSET;
4270 } else if (unlikely(rt_task(current)) && !in_interrupt())
4271 alloc_flags |= ALLOC_HARDER;
4274 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4275 alloc_flags |= ALLOC_CMA;
4280 static bool oom_reserves_allowed(struct task_struct *tsk)
4282 if (!tsk_is_oom_victim(tsk))
4286 * !MMU doesn't have oom reaper so give access to memory reserves
4287 * only to the thread with TIF_MEMDIE set
4289 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4296 * Distinguish requests which really need access to full memory
4297 * reserves from oom victims which can live with a portion of it
4299 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4301 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4303 if (gfp_mask & __GFP_MEMALLOC)
4304 return ALLOC_NO_WATERMARKS;
4305 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4306 return ALLOC_NO_WATERMARKS;
4307 if (!in_interrupt()) {
4308 if (current->flags & PF_MEMALLOC)
4309 return ALLOC_NO_WATERMARKS;
4310 else if (oom_reserves_allowed(current))
4317 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4319 return !!__gfp_pfmemalloc_flags(gfp_mask);
4323 * Checks whether it makes sense to retry the reclaim to make a forward progress
4324 * for the given allocation request.
4326 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4327 * without success, or when we couldn't even meet the watermark if we
4328 * reclaimed all remaining pages on the LRU lists.
4330 * Returns true if a retry is viable or false to enter the oom path.
4333 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4334 struct alloc_context *ac, int alloc_flags,
4335 bool did_some_progress, int *no_progress_loops)
4342 * Costly allocations might have made a progress but this doesn't mean
4343 * their order will become available due to high fragmentation so
4344 * always increment the no progress counter for them
4346 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4347 *no_progress_loops = 0;
4349 (*no_progress_loops)++;
4352 * Make sure we converge to OOM if we cannot make any progress
4353 * several times in the row.
4355 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4356 /* Before OOM, exhaust highatomic_reserve */
4357 return unreserve_highatomic_pageblock(ac, true);
4361 * Keep reclaiming pages while there is a chance this will lead
4362 * somewhere. If none of the target zones can satisfy our allocation
4363 * request even if all reclaimable pages are considered then we are
4364 * screwed and have to go OOM.
4366 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4368 unsigned long available;
4369 unsigned long reclaimable;
4370 unsigned long min_wmark = min_wmark_pages(zone);
4373 available = reclaimable = zone_reclaimable_pages(zone);
4374 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4377 * Would the allocation succeed if we reclaimed all
4378 * reclaimable pages?
4380 wmark = __zone_watermark_ok(zone, order, min_wmark,
4381 ac_classzone_idx(ac), alloc_flags, available);
4382 trace_reclaim_retry_zone(z, order, reclaimable,
4383 available, min_wmark, *no_progress_loops, wmark);
4386 * If we didn't make any progress and have a lot of
4387 * dirty + writeback pages then we should wait for
4388 * an IO to complete to slow down the reclaim and
4389 * prevent from pre mature OOM
4391 if (!did_some_progress) {
4392 unsigned long write_pending;
4394 write_pending = zone_page_state_snapshot(zone,
4395 NR_ZONE_WRITE_PENDING);
4397 if (2 * write_pending > reclaimable) {
4398 congestion_wait(BLK_RW_ASYNC, HZ/10);
4410 * Memory allocation/reclaim might be called from a WQ context and the
4411 * current implementation of the WQ concurrency control doesn't
4412 * recognize that a particular WQ is congested if the worker thread is
4413 * looping without ever sleeping. Therefore we have to do a short sleep
4414 * here rather than calling cond_resched().
4416 if (current->flags & PF_WQ_WORKER)
4417 schedule_timeout_uninterruptible(1);
4424 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4427 * It's possible that cpuset's mems_allowed and the nodemask from
4428 * mempolicy don't intersect. This should be normally dealt with by
4429 * policy_nodemask(), but it's possible to race with cpuset update in
4430 * such a way the check therein was true, and then it became false
4431 * before we got our cpuset_mems_cookie here.
4432 * This assumes that for all allocations, ac->nodemask can come only
4433 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4434 * when it does not intersect with the cpuset restrictions) or the
4435 * caller can deal with a violated nodemask.
4437 if (cpusets_enabled() && ac->nodemask &&
4438 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4439 ac->nodemask = NULL;
4444 * When updating a task's mems_allowed or mempolicy nodemask, it is
4445 * possible to race with parallel threads in such a way that our
4446 * allocation can fail while the mask is being updated. If we are about
4447 * to fail, check if the cpuset changed during allocation and if so,
4450 if (read_mems_allowed_retry(cpuset_mems_cookie))
4456 static inline struct page *
4457 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4458 struct alloc_context *ac)
4460 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4461 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4462 struct page *page = NULL;
4463 unsigned int alloc_flags;
4464 unsigned long did_some_progress;
4465 enum compact_priority compact_priority;
4466 enum compact_result compact_result;
4467 int compaction_retries;
4468 int no_progress_loops;
4469 unsigned int cpuset_mems_cookie;
4473 * We also sanity check to catch abuse of atomic reserves being used by
4474 * callers that are not in atomic context.
4476 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4477 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4478 gfp_mask &= ~__GFP_ATOMIC;
4481 compaction_retries = 0;
4482 no_progress_loops = 0;
4483 compact_priority = DEF_COMPACT_PRIORITY;
4484 cpuset_mems_cookie = read_mems_allowed_begin();
4487 * The fast path uses conservative alloc_flags to succeed only until
4488 * kswapd needs to be woken up, and to avoid the cost of setting up
4489 * alloc_flags precisely. So we do that now.
4491 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4494 * We need to recalculate the starting point for the zonelist iterator
4495 * because we might have used different nodemask in the fast path, or
4496 * there was a cpuset modification and we are retrying - otherwise we
4497 * could end up iterating over non-eligible zones endlessly.
4499 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4500 ac->high_zoneidx, ac->nodemask);
4501 if (!ac->preferred_zoneref->zone)
4504 if (alloc_flags & ALLOC_KSWAPD)
4505 wake_all_kswapds(order, gfp_mask, ac);
4508 * The adjusted alloc_flags might result in immediate success, so try
4511 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4516 * For costly allocations, try direct compaction first, as it's likely
4517 * that we have enough base pages and don't need to reclaim. For non-
4518 * movable high-order allocations, do that as well, as compaction will
4519 * try prevent permanent fragmentation by migrating from blocks of the
4521 * Don't try this for allocations that are allowed to ignore
4522 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4524 if (can_direct_reclaim &&
4526 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4527 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4528 page = __alloc_pages_direct_compact(gfp_mask, order,
4530 INIT_COMPACT_PRIORITY,
4536 * Checks for costly allocations with __GFP_NORETRY, which
4537 * includes some THP page fault allocations
4539 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4541 * If allocating entire pageblock(s) and compaction
4542 * failed because all zones are below low watermarks
4543 * or is prohibited because it recently failed at this
4544 * order, fail immediately unless the allocator has
4545 * requested compaction and reclaim retry.
4548 * - potentially very expensive because zones are far
4549 * below their low watermarks or this is part of very
4550 * bursty high order allocations,
4551 * - not guaranteed to help because isolate_freepages()
4552 * may not iterate over freed pages as part of its
4554 * - unlikely to make entire pageblocks free on its
4557 if (compact_result == COMPACT_SKIPPED ||
4558 compact_result == COMPACT_DEFERRED)
4562 * Looks like reclaim/compaction is worth trying, but
4563 * sync compaction could be very expensive, so keep
4564 * using async compaction.
4566 compact_priority = INIT_COMPACT_PRIORITY;
4571 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4572 if (alloc_flags & ALLOC_KSWAPD)
4573 wake_all_kswapds(order, gfp_mask, ac);
4575 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4577 alloc_flags = reserve_flags;
4580 * Reset the nodemask and zonelist iterators if memory policies can be
4581 * ignored. These allocations are high priority and system rather than
4584 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4585 ac->nodemask = NULL;
4586 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4587 ac->high_zoneidx, ac->nodemask);
4590 /* Attempt with potentially adjusted zonelist and alloc_flags */
4591 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4595 /* Caller is not willing to reclaim, we can't balance anything */
4596 if (!can_direct_reclaim)
4599 /* Avoid recursion of direct reclaim */
4600 if (current->flags & PF_MEMALLOC)
4603 /* Try direct reclaim and then allocating */
4604 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4605 &did_some_progress);
4609 /* Try direct compaction and then allocating */
4610 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4611 compact_priority, &compact_result);
4615 /* Do not loop if specifically requested */
4616 if (gfp_mask & __GFP_NORETRY)
4620 * Do not retry costly high order allocations unless they are
4621 * __GFP_RETRY_MAYFAIL
4623 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4626 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4627 did_some_progress > 0, &no_progress_loops))
4631 * It doesn't make any sense to retry for the compaction if the order-0
4632 * reclaim is not able to make any progress because the current
4633 * implementation of the compaction depends on the sufficient amount
4634 * of free memory (see __compaction_suitable)
4636 if (did_some_progress > 0 &&
4637 should_compact_retry(ac, order, alloc_flags,
4638 compact_result, &compact_priority,
4639 &compaction_retries))
4643 /* Deal with possible cpuset update races before we start OOM killing */
4644 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4647 /* Reclaim has failed us, start killing things */
4648 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4652 /* Avoid allocations with no watermarks from looping endlessly */
4653 if (tsk_is_oom_victim(current) &&
4654 (alloc_flags == ALLOC_OOM ||
4655 (gfp_mask & __GFP_NOMEMALLOC)))
4658 /* Retry as long as the OOM killer is making progress */
4659 if (did_some_progress) {
4660 no_progress_loops = 0;
4665 /* Deal with possible cpuset update races before we fail */
4666 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4670 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4673 if (gfp_mask & __GFP_NOFAIL) {
4675 * All existing users of the __GFP_NOFAIL are blockable, so warn
4676 * of any new users that actually require GFP_NOWAIT
4678 if (WARN_ON_ONCE(!can_direct_reclaim))
4682 * PF_MEMALLOC request from this context is rather bizarre
4683 * because we cannot reclaim anything and only can loop waiting
4684 * for somebody to do a work for us
4686 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4689 * non failing costly orders are a hard requirement which we
4690 * are not prepared for much so let's warn about these users
4691 * so that we can identify them and convert them to something
4694 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4697 * Help non-failing allocations by giving them access to memory
4698 * reserves but do not use ALLOC_NO_WATERMARKS because this
4699 * could deplete whole memory reserves which would just make
4700 * the situation worse
4702 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4710 warn_alloc(gfp_mask, ac->nodemask,
4711 "page allocation failure: order:%u", order);
4716 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4717 int preferred_nid, nodemask_t *nodemask,
4718 struct alloc_context *ac, gfp_t *alloc_mask,
4719 unsigned int *alloc_flags)
4721 ac->high_zoneidx = gfp_zone(gfp_mask);
4722 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4723 ac->nodemask = nodemask;
4724 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4726 if (cpusets_enabled()) {
4727 *alloc_mask |= __GFP_HARDWALL;
4729 ac->nodemask = &cpuset_current_mems_allowed;
4731 *alloc_flags |= ALLOC_CPUSET;
4734 fs_reclaim_acquire(gfp_mask);
4735 fs_reclaim_release(gfp_mask);
4737 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4739 if (should_fail_alloc_page(gfp_mask, order))
4742 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4743 *alloc_flags |= ALLOC_CMA;
4748 /* Determine whether to spread dirty pages and what the first usable zone */
4749 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4751 /* Dirty zone balancing only done in the fast path */
4752 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4755 * The preferred zone is used for statistics but crucially it is
4756 * also used as the starting point for the zonelist iterator. It
4757 * may get reset for allocations that ignore memory policies.
4759 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4760 ac->high_zoneidx, ac->nodemask);
4764 * This is the 'heart' of the zoned buddy allocator.
4767 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4768 nodemask_t *nodemask)
4771 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4772 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4773 struct alloc_context ac = { };
4776 * There are several places where we assume that the order value is sane
4777 * so bail out early if the request is out of bound.
4779 if (unlikely(order >= MAX_ORDER)) {
4780 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4784 gfp_mask &= gfp_allowed_mask;
4785 alloc_mask = gfp_mask;
4786 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4789 finalise_ac(gfp_mask, &ac);
4792 * Forbid the first pass from falling back to types that fragment
4793 * memory until all local zones are considered.
4795 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4797 /* First allocation attempt */
4798 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4803 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4804 * resp. GFP_NOIO which has to be inherited for all allocation requests
4805 * from a particular context which has been marked by
4806 * memalloc_no{fs,io}_{save,restore}.
4808 alloc_mask = current_gfp_context(gfp_mask);
4809 ac.spread_dirty_pages = false;
4812 * Restore the original nodemask if it was potentially replaced with
4813 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4815 ac.nodemask = nodemask;
4817 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4820 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4821 unlikely(__memcg_kmem_charge_page(page, gfp_mask, order) != 0)) {
4822 __free_pages(page, order);
4826 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4830 EXPORT_SYMBOL(__alloc_pages_nodemask);
4833 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4834 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4835 * you need to access high mem.
4837 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4841 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4844 return (unsigned long) page_address(page);
4846 EXPORT_SYMBOL(__get_free_pages);
4848 unsigned long get_zeroed_page(gfp_t gfp_mask)
4850 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4852 EXPORT_SYMBOL(get_zeroed_page);
4854 static inline void free_the_page(struct page *page, unsigned int order)
4856 if (order == 0) /* Via pcp? */
4857 free_unref_page(page);
4859 __free_pages_ok(page, order);
4862 void __free_pages(struct page *page, unsigned int order)
4864 if (put_page_testzero(page))
4865 free_the_page(page, order);
4867 EXPORT_SYMBOL(__free_pages);
4869 void free_pages(unsigned long addr, unsigned int order)
4872 VM_BUG_ON(!virt_addr_valid((void *)addr));
4873 __free_pages(virt_to_page((void *)addr), order);
4877 EXPORT_SYMBOL(free_pages);
4881 * An arbitrary-length arbitrary-offset area of memory which resides
4882 * within a 0 or higher order page. Multiple fragments within that page
4883 * are individually refcounted, in the page's reference counter.
4885 * The page_frag functions below provide a simple allocation framework for
4886 * page fragments. This is used by the network stack and network device
4887 * drivers to provide a backing region of memory for use as either an
4888 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4890 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4893 struct page *page = NULL;
4894 gfp_t gfp = gfp_mask;
4896 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4897 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4899 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4900 PAGE_FRAG_CACHE_MAX_ORDER);
4901 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4903 if (unlikely(!page))
4904 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4906 nc->va = page ? page_address(page) : NULL;
4911 void __page_frag_cache_drain(struct page *page, unsigned int count)
4913 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4915 if (page_ref_sub_and_test(page, count))
4916 free_the_page(page, compound_order(page));
4918 EXPORT_SYMBOL(__page_frag_cache_drain);
4920 void *page_frag_alloc(struct page_frag_cache *nc,
4921 unsigned int fragsz, gfp_t gfp_mask)
4923 unsigned int size = PAGE_SIZE;
4927 if (unlikely(!nc->va)) {
4929 page = __page_frag_cache_refill(nc, gfp_mask);
4933 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4934 /* if size can vary use size else just use PAGE_SIZE */
4937 /* Even if we own the page, we do not use atomic_set().
4938 * This would break get_page_unless_zero() users.
4940 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4942 /* reset page count bias and offset to start of new frag */
4943 nc->pfmemalloc = page_is_pfmemalloc(page);
4944 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4948 offset = nc->offset - fragsz;
4949 if (unlikely(offset < 0)) {
4950 page = virt_to_page(nc->va);
4952 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4955 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4956 /* if size can vary use size else just use PAGE_SIZE */
4959 /* OK, page count is 0, we can safely set it */
4960 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4962 /* reset page count bias and offset to start of new frag */
4963 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4964 offset = size - fragsz;
4968 nc->offset = offset;
4970 return nc->va + offset;
4972 EXPORT_SYMBOL(page_frag_alloc);
4975 * Frees a page fragment allocated out of either a compound or order 0 page.
4977 void page_frag_free(void *addr)
4979 struct page *page = virt_to_head_page(addr);
4981 if (unlikely(put_page_testzero(page)))
4982 free_the_page(page, compound_order(page));
4984 EXPORT_SYMBOL(page_frag_free);
4986 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4990 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4991 unsigned long used = addr + PAGE_ALIGN(size);
4993 split_page(virt_to_page((void *)addr), order);
4994 while (used < alloc_end) {
4999 return (void *)addr;
5003 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5004 * @size: the number of bytes to allocate
5005 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5007 * This function is similar to alloc_pages(), except that it allocates the
5008 * minimum number of pages to satisfy the request. alloc_pages() can only
5009 * allocate memory in power-of-two pages.
5011 * This function is also limited by MAX_ORDER.
5013 * Memory allocated by this function must be released by free_pages_exact().
5015 * Return: pointer to the allocated area or %NULL in case of error.
5017 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5019 unsigned int order = get_order(size);
5022 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5023 gfp_mask &= ~__GFP_COMP;
5025 addr = __get_free_pages(gfp_mask, order);
5026 return make_alloc_exact(addr, order, size);
5028 EXPORT_SYMBOL(alloc_pages_exact);
5031 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5033 * @nid: the preferred node ID where memory should be allocated
5034 * @size: the number of bytes to allocate
5035 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5037 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5040 * Return: pointer to the allocated area or %NULL in case of error.
5042 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5044 unsigned int order = get_order(size);
5047 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5048 gfp_mask &= ~__GFP_COMP;
5050 p = alloc_pages_node(nid, gfp_mask, order);
5053 return make_alloc_exact((unsigned long)page_address(p), order, size);
5057 * free_pages_exact - release memory allocated via alloc_pages_exact()
5058 * @virt: the value returned by alloc_pages_exact.
5059 * @size: size of allocation, same value as passed to alloc_pages_exact().
5061 * Release the memory allocated by a previous call to alloc_pages_exact.
5063 void free_pages_exact(void *virt, size_t size)
5065 unsigned long addr = (unsigned long)virt;
5066 unsigned long end = addr + PAGE_ALIGN(size);
5068 while (addr < end) {
5073 EXPORT_SYMBOL(free_pages_exact);
5076 * nr_free_zone_pages - count number of pages beyond high watermark
5077 * @offset: The zone index of the highest zone
5079 * nr_free_zone_pages() counts the number of pages which are beyond the
5080 * high watermark within all zones at or below a given zone index. For each
5081 * zone, the number of pages is calculated as:
5083 * nr_free_zone_pages = managed_pages - high_pages
5085 * Return: number of pages beyond high watermark.
5087 static unsigned long nr_free_zone_pages(int offset)
5092 /* Just pick one node, since fallback list is circular */
5093 unsigned long sum = 0;
5095 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5097 for_each_zone_zonelist(zone, z, zonelist, offset) {
5098 unsigned long size = zone_managed_pages(zone);
5099 unsigned long high = high_wmark_pages(zone);
5108 * nr_free_buffer_pages - count number of pages beyond high watermark
5110 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5111 * watermark within ZONE_DMA and ZONE_NORMAL.
5113 * Return: number of pages beyond high watermark within ZONE_DMA and
5116 unsigned long nr_free_buffer_pages(void)
5118 return nr_free_zone_pages(gfp_zone(GFP_USER));
5120 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5123 * nr_free_pagecache_pages - count number of pages beyond high watermark
5125 * nr_free_pagecache_pages() counts the number of pages which are beyond the
5126 * high watermark within all zones.
5128 * Return: number of pages beyond high watermark within all zones.
5130 unsigned long nr_free_pagecache_pages(void)
5132 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5135 static inline void show_node(struct zone *zone)
5137 if (IS_ENABLED(CONFIG_NUMA))
5138 printk("Node %d ", zone_to_nid(zone));
5141 long si_mem_available(void)
5144 unsigned long pagecache;
5145 unsigned long wmark_low = 0;
5146 unsigned long pages[NR_LRU_LISTS];
5147 unsigned long reclaimable;
5151 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5152 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5155 wmark_low += low_wmark_pages(zone);
5158 * Estimate the amount of memory available for userspace allocations,
5159 * without causing swapping.
5161 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5164 * Not all the page cache can be freed, otherwise the system will
5165 * start swapping. Assume at least half of the page cache, or the
5166 * low watermark worth of cache, needs to stay.
5168 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5169 pagecache -= min(pagecache / 2, wmark_low);
5170 available += pagecache;
5173 * Part of the reclaimable slab and other kernel memory consists of
5174 * items that are in use, and cannot be freed. Cap this estimate at the
5177 reclaimable = global_node_page_state(NR_SLAB_RECLAIMABLE) +
5178 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5179 available += reclaimable - min(reclaimable / 2, wmark_low);
5185 EXPORT_SYMBOL_GPL(si_mem_available);
5187 void si_meminfo(struct sysinfo *val)
5189 val->totalram = totalram_pages();
5190 val->sharedram = global_node_page_state(NR_SHMEM);
5191 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5192 val->bufferram = nr_blockdev_pages();
5193 val->totalhigh = totalhigh_pages();
5194 val->freehigh = nr_free_highpages();
5195 val->mem_unit = PAGE_SIZE;
5198 EXPORT_SYMBOL(si_meminfo);
5201 void si_meminfo_node(struct sysinfo *val, int nid)
5203 int zone_type; /* needs to be signed */
5204 unsigned long managed_pages = 0;
5205 unsigned long managed_highpages = 0;
5206 unsigned long free_highpages = 0;
5207 pg_data_t *pgdat = NODE_DATA(nid);
5209 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5210 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5211 val->totalram = managed_pages;
5212 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5213 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5214 #ifdef CONFIG_HIGHMEM
5215 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5216 struct zone *zone = &pgdat->node_zones[zone_type];
5218 if (is_highmem(zone)) {
5219 managed_highpages += zone_managed_pages(zone);
5220 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5223 val->totalhigh = managed_highpages;
5224 val->freehigh = free_highpages;
5226 val->totalhigh = managed_highpages;
5227 val->freehigh = free_highpages;
5229 val->mem_unit = PAGE_SIZE;
5234 * Determine whether the node should be displayed or not, depending on whether
5235 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5237 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5239 if (!(flags & SHOW_MEM_FILTER_NODES))
5243 * no node mask - aka implicit memory numa policy. Do not bother with
5244 * the synchronization - read_mems_allowed_begin - because we do not
5245 * have to be precise here.
5248 nodemask = &cpuset_current_mems_allowed;
5250 return !node_isset(nid, *nodemask);
5253 #define K(x) ((x) << (PAGE_SHIFT-10))
5255 static void show_migration_types(unsigned char type)
5257 static const char types[MIGRATE_TYPES] = {
5258 [MIGRATE_UNMOVABLE] = 'U',
5259 [MIGRATE_MOVABLE] = 'M',
5260 [MIGRATE_RECLAIMABLE] = 'E',
5261 [MIGRATE_HIGHATOMIC] = 'H',
5263 [MIGRATE_CMA] = 'C',
5265 #ifdef CONFIG_MEMORY_ISOLATION
5266 [MIGRATE_ISOLATE] = 'I',
5269 char tmp[MIGRATE_TYPES + 1];
5273 for (i = 0; i < MIGRATE_TYPES; i++) {
5274 if (type & (1 << i))
5279 printk(KERN_CONT "(%s) ", tmp);
5283 * Show free area list (used inside shift_scroll-lock stuff)
5284 * We also calculate the percentage fragmentation. We do this by counting the
5285 * memory on each free list with the exception of the first item on the list.
5288 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5291 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5293 unsigned long free_pcp = 0;
5298 for_each_populated_zone(zone) {
5299 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5302 for_each_online_cpu(cpu)
5303 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5306 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5307 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5308 " unevictable:%lu dirty:%lu writeback:%lu\n"
5309 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5310 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5311 " free:%lu free_pcp:%lu free_cma:%lu\n",
5312 global_node_page_state(NR_ACTIVE_ANON),
5313 global_node_page_state(NR_INACTIVE_ANON),
5314 global_node_page_state(NR_ISOLATED_ANON),
5315 global_node_page_state(NR_ACTIVE_FILE),
5316 global_node_page_state(NR_INACTIVE_FILE),
5317 global_node_page_state(NR_ISOLATED_FILE),
5318 global_node_page_state(NR_UNEVICTABLE),
5319 global_node_page_state(NR_FILE_DIRTY),
5320 global_node_page_state(NR_WRITEBACK),
5321 global_node_page_state(NR_SLAB_RECLAIMABLE),
5322 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
5323 global_node_page_state(NR_FILE_MAPPED),
5324 global_node_page_state(NR_SHMEM),
5325 global_zone_page_state(NR_PAGETABLE),
5326 global_zone_page_state(NR_BOUNCE),
5327 global_zone_page_state(NR_FREE_PAGES),
5329 global_zone_page_state(NR_FREE_CMA_PAGES));
5331 for_each_online_pgdat(pgdat) {
5332 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5336 " active_anon:%lukB"
5337 " inactive_anon:%lukB"
5338 " active_file:%lukB"
5339 " inactive_file:%lukB"
5340 " unevictable:%lukB"
5341 " isolated(anon):%lukB"
5342 " isolated(file):%lukB"
5347 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5349 " shmem_pmdmapped: %lukB"
5352 " writeback_tmp:%lukB"
5353 " all_unreclaimable? %s"
5356 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5357 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5358 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5359 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5360 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5361 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5362 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5363 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5364 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5365 K(node_page_state(pgdat, NR_WRITEBACK)),
5366 K(node_page_state(pgdat, NR_SHMEM)),
5367 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5368 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5369 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5371 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5373 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5374 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5378 for_each_populated_zone(zone) {
5381 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5385 for_each_online_cpu(cpu)
5386 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5395 " reserved_highatomic:%luKB"
5396 " active_anon:%lukB"
5397 " inactive_anon:%lukB"
5398 " active_file:%lukB"
5399 " inactive_file:%lukB"
5400 " unevictable:%lukB"
5401 " writepending:%lukB"
5405 " kernel_stack:%lukB"
5406 #ifdef CONFIG_SHADOW_CALL_STACK
5407 " shadow_call_stack:%lukB"
5416 K(zone_page_state(zone, NR_FREE_PAGES)),
5417 K(min_wmark_pages(zone)),
5418 K(low_wmark_pages(zone)),
5419 K(high_wmark_pages(zone)),
5420 K(zone->nr_reserved_highatomic),
5421 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5422 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5423 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5424 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5425 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5426 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5427 K(zone->present_pages),
5428 K(zone_managed_pages(zone)),
5429 K(zone_page_state(zone, NR_MLOCK)),
5430 zone_page_state(zone, NR_KERNEL_STACK_KB),
5431 #ifdef CONFIG_SHADOW_CALL_STACK
5432 zone_page_state(zone, NR_KERNEL_SCS_KB),
5434 K(zone_page_state(zone, NR_PAGETABLE)),
5435 K(zone_page_state(zone, NR_BOUNCE)),
5437 K(this_cpu_read(zone->pageset->pcp.count)),
5438 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5439 printk("lowmem_reserve[]:");
5440 for (i = 0; i < MAX_NR_ZONES; i++)
5441 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5442 printk(KERN_CONT "\n");
5445 for_each_populated_zone(zone) {
5447 unsigned long nr[MAX_ORDER], flags, total = 0;
5448 unsigned char types[MAX_ORDER];
5450 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5453 printk(KERN_CONT "%s: ", zone->name);
5455 spin_lock_irqsave(&zone->lock, flags);
5456 for (order = 0; order < MAX_ORDER; order++) {
5457 struct free_area *area = &zone->free_area[order];
5460 nr[order] = area->nr_free;
5461 total += nr[order] << order;
5464 for (type = 0; type < MIGRATE_TYPES; type++) {
5465 if (!free_area_empty(area, type))
5466 types[order] |= 1 << type;
5469 spin_unlock_irqrestore(&zone->lock, flags);
5470 for (order = 0; order < MAX_ORDER; order++) {
5471 printk(KERN_CONT "%lu*%lukB ",
5472 nr[order], K(1UL) << order);
5474 show_migration_types(types[order]);
5476 printk(KERN_CONT "= %lukB\n", K(total));
5479 hugetlb_show_meminfo();
5481 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5483 show_swap_cache_info();
5486 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5488 zoneref->zone = zone;
5489 zoneref->zone_idx = zone_idx(zone);
5493 * Builds allocation fallback zone lists.
5495 * Add all populated zones of a node to the zonelist.
5497 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5500 enum zone_type zone_type = MAX_NR_ZONES;
5505 zone = pgdat->node_zones + zone_type;
5506 if (managed_zone(zone)) {
5507 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5508 check_highest_zone(zone_type);
5510 } while (zone_type);
5517 static int __parse_numa_zonelist_order(char *s)
5520 * We used to support different zonlists modes but they turned
5521 * out to be just not useful. Let's keep the warning in place
5522 * if somebody still use the cmd line parameter so that we do
5523 * not fail it silently
5525 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5526 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5532 static __init int setup_numa_zonelist_order(char *s)
5537 return __parse_numa_zonelist_order(s);
5539 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5541 char numa_zonelist_order[] = "Node";
5544 * sysctl handler for numa_zonelist_order
5546 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5547 void __user *buffer, size_t *length,
5554 return proc_dostring(table, write, buffer, length, ppos);
5555 str = memdup_user_nul(buffer, 16);
5557 return PTR_ERR(str);
5559 ret = __parse_numa_zonelist_order(str);
5565 #define MAX_NODE_LOAD (nr_online_nodes)
5566 static int node_load[MAX_NUMNODES];
5569 * find_next_best_node - find the next node that should appear in a given node's fallback list
5570 * @node: node whose fallback list we're appending
5571 * @used_node_mask: nodemask_t of already used nodes
5573 * We use a number of factors to determine which is the next node that should
5574 * appear on a given node's fallback list. The node should not have appeared
5575 * already in @node's fallback list, and it should be the next closest node
5576 * according to the distance array (which contains arbitrary distance values
5577 * from each node to each node in the system), and should also prefer nodes
5578 * with no CPUs, since presumably they'll have very little allocation pressure
5579 * on them otherwise.
5581 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5583 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5586 int min_val = INT_MAX;
5587 int best_node = NUMA_NO_NODE;
5588 const struct cpumask *tmp = cpumask_of_node(0);
5590 /* Use the local node if we haven't already */
5591 if (!node_isset(node, *used_node_mask)) {
5592 node_set(node, *used_node_mask);
5596 for_each_node_state(n, N_MEMORY) {
5598 /* Don't want a node to appear more than once */
5599 if (node_isset(n, *used_node_mask))
5602 /* Use the distance array to find the distance */
5603 val = node_distance(node, n);
5605 /* Penalize nodes under us ("prefer the next node") */
5608 /* Give preference to headless and unused nodes */
5609 tmp = cpumask_of_node(n);
5610 if (!cpumask_empty(tmp))
5611 val += PENALTY_FOR_NODE_WITH_CPUS;
5613 /* Slight preference for less loaded node */
5614 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5615 val += node_load[n];
5617 if (val < min_val) {
5624 node_set(best_node, *used_node_mask);
5631 * Build zonelists ordered by node and zones within node.
5632 * This results in maximum locality--normal zone overflows into local
5633 * DMA zone, if any--but risks exhausting DMA zone.
5635 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5638 struct zoneref *zonerefs;
5641 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5643 for (i = 0; i < nr_nodes; i++) {
5646 pg_data_t *node = NODE_DATA(node_order[i]);
5648 nr_zones = build_zonerefs_node(node, zonerefs);
5649 zonerefs += nr_zones;
5651 zonerefs->zone = NULL;
5652 zonerefs->zone_idx = 0;
5656 * Build gfp_thisnode zonelists
5658 static void build_thisnode_zonelists(pg_data_t *pgdat)
5660 struct zoneref *zonerefs;
5663 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5664 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5665 zonerefs += nr_zones;
5666 zonerefs->zone = NULL;
5667 zonerefs->zone_idx = 0;
5671 * Build zonelists ordered by zone and nodes within zones.
5672 * This results in conserving DMA zone[s] until all Normal memory is
5673 * exhausted, but results in overflowing to remote node while memory
5674 * may still exist in local DMA zone.
5677 static void build_zonelists(pg_data_t *pgdat)
5679 static int node_order[MAX_NUMNODES];
5680 int node, load, nr_nodes = 0;
5681 nodemask_t used_mask;
5682 int local_node, prev_node;
5684 /* NUMA-aware ordering of nodes */
5685 local_node = pgdat->node_id;
5686 load = nr_online_nodes;
5687 prev_node = local_node;
5688 nodes_clear(used_mask);
5690 memset(node_order, 0, sizeof(node_order));
5691 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5693 * We don't want to pressure a particular node.
5694 * So adding penalty to the first node in same
5695 * distance group to make it round-robin.
5697 if (node_distance(local_node, node) !=
5698 node_distance(local_node, prev_node))
5699 node_load[node] = load;
5701 node_order[nr_nodes++] = node;
5706 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5707 build_thisnode_zonelists(pgdat);
5710 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5712 * Return node id of node used for "local" allocations.
5713 * I.e., first node id of first zone in arg node's generic zonelist.
5714 * Used for initializing percpu 'numa_mem', which is used primarily
5715 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5717 int local_memory_node(int node)
5721 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5722 gfp_zone(GFP_KERNEL),
5724 return zone_to_nid(z->zone);
5728 static void setup_min_unmapped_ratio(void);
5729 static void setup_min_slab_ratio(void);
5730 #else /* CONFIG_NUMA */
5732 static void build_zonelists(pg_data_t *pgdat)
5734 int node, local_node;
5735 struct zoneref *zonerefs;
5738 local_node = pgdat->node_id;
5740 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5741 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5742 zonerefs += nr_zones;
5745 * Now we build the zonelist so that it contains the zones
5746 * of all the other nodes.
5747 * We don't want to pressure a particular node, so when
5748 * building the zones for node N, we make sure that the
5749 * zones coming right after the local ones are those from
5750 * node N+1 (modulo N)
5752 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5753 if (!node_online(node))
5755 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5756 zonerefs += nr_zones;
5758 for (node = 0; node < local_node; node++) {
5759 if (!node_online(node))
5761 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5762 zonerefs += nr_zones;
5765 zonerefs->zone = NULL;
5766 zonerefs->zone_idx = 0;
5769 #endif /* CONFIG_NUMA */
5772 * Boot pageset table. One per cpu which is going to be used for all
5773 * zones and all nodes. The parameters will be set in such a way
5774 * that an item put on a list will immediately be handed over to
5775 * the buddy list. This is safe since pageset manipulation is done
5776 * with interrupts disabled.
5778 * The boot_pagesets must be kept even after bootup is complete for
5779 * unused processors and/or zones. They do play a role for bootstrapping
5780 * hotplugged processors.
5782 * zoneinfo_show() and maybe other functions do
5783 * not check if the processor is online before following the pageset pointer.
5784 * Other parts of the kernel may not check if the zone is available.
5786 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5787 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5788 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5790 static void __build_all_zonelists(void *data)
5793 int __maybe_unused cpu;
5794 pg_data_t *self = data;
5795 static DEFINE_SPINLOCK(lock);
5800 memset(node_load, 0, sizeof(node_load));
5804 * This node is hotadded and no memory is yet present. So just
5805 * building zonelists is fine - no need to touch other nodes.
5807 if (self && !node_online(self->node_id)) {
5808 build_zonelists(self);
5810 for_each_online_node(nid) {
5811 pg_data_t *pgdat = NODE_DATA(nid);
5813 build_zonelists(pgdat);
5816 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5818 * We now know the "local memory node" for each node--
5819 * i.e., the node of the first zone in the generic zonelist.
5820 * Set up numa_mem percpu variable for on-line cpus. During
5821 * boot, only the boot cpu should be on-line; we'll init the
5822 * secondary cpus' numa_mem as they come on-line. During
5823 * node/memory hotplug, we'll fixup all on-line cpus.
5825 for_each_online_cpu(cpu)
5826 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5833 static noinline void __init
5834 build_all_zonelists_init(void)
5838 __build_all_zonelists(NULL);
5841 * Initialize the boot_pagesets that are going to be used
5842 * for bootstrapping processors. The real pagesets for
5843 * each zone will be allocated later when the per cpu
5844 * allocator is available.
5846 * boot_pagesets are used also for bootstrapping offline
5847 * cpus if the system is already booted because the pagesets
5848 * are needed to initialize allocators on a specific cpu too.
5849 * F.e. the percpu allocator needs the page allocator which
5850 * needs the percpu allocator in order to allocate its pagesets
5851 * (a chicken-egg dilemma).
5853 for_each_possible_cpu(cpu)
5854 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5856 mminit_verify_zonelist();
5857 cpuset_init_current_mems_allowed();
5861 * unless system_state == SYSTEM_BOOTING.
5863 * __ref due to call of __init annotated helper build_all_zonelists_init
5864 * [protected by SYSTEM_BOOTING].
5866 void __ref build_all_zonelists(pg_data_t *pgdat)
5868 if (system_state == SYSTEM_BOOTING) {
5869 build_all_zonelists_init();
5871 __build_all_zonelists(pgdat);
5872 /* cpuset refresh routine should be here */
5874 vm_total_pages = nr_free_pagecache_pages();
5876 * Disable grouping by mobility if the number of pages in the
5877 * system is too low to allow the mechanism to work. It would be
5878 * more accurate, but expensive to check per-zone. This check is
5879 * made on memory-hotadd so a system can start with mobility
5880 * disabled and enable it later
5882 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5883 page_group_by_mobility_disabled = 1;
5885 page_group_by_mobility_disabled = 0;
5887 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5889 page_group_by_mobility_disabled ? "off" : "on",
5892 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5896 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5897 static bool __meminit
5898 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5900 static struct memblock_region *r;
5902 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5903 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5904 for_each_memblock(memory, r) {
5905 if (*pfn < memblock_region_memory_end_pfn(r))
5909 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5910 memblock_is_mirror(r)) {
5911 *pfn = memblock_region_memory_end_pfn(r);
5919 * Initially all pages are reserved - free ones are freed
5920 * up by memblock_free_all() once the early boot process is
5921 * done. Non-atomic initialization, single-pass.
5923 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5924 unsigned long start_pfn, enum memmap_context context,
5925 struct vmem_altmap *altmap)
5927 unsigned long pfn, end_pfn = start_pfn + size;
5930 if (highest_memmap_pfn < end_pfn - 1)
5931 highest_memmap_pfn = end_pfn - 1;
5933 #ifdef CONFIG_ZONE_DEVICE
5935 * Honor reservation requested by the driver for this ZONE_DEVICE
5936 * memory. We limit the total number of pages to initialize to just
5937 * those that might contain the memory mapping. We will defer the
5938 * ZONE_DEVICE page initialization until after we have released
5941 if (zone == ZONE_DEVICE) {
5945 if (start_pfn == altmap->base_pfn)
5946 start_pfn += altmap->reserve;
5947 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5951 for (pfn = start_pfn; pfn < end_pfn; ) {
5953 * There can be holes in boot-time mem_map[]s handed to this
5954 * function. They do not exist on hotplugged memory.
5956 if (context == MEMMAP_EARLY) {
5957 if (overlap_memmap_init(zone, &pfn))
5959 if (defer_init(nid, pfn, end_pfn))
5963 page = pfn_to_page(pfn);
5964 __init_single_page(page, pfn, zone, nid);
5965 if (context == MEMMAP_HOTPLUG)
5966 __SetPageReserved(page);
5969 * Mark the block movable so that blocks are reserved for
5970 * movable at startup. This will force kernel allocations
5971 * to reserve their blocks rather than leaking throughout
5972 * the address space during boot when many long-lived
5973 * kernel allocations are made.
5975 * bitmap is created for zone's valid pfn range. but memmap
5976 * can be created for invalid pages (for alignment)
5977 * check here not to call set_pageblock_migratetype() against
5980 if (!(pfn & (pageblock_nr_pages - 1))) {
5981 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5988 #ifdef CONFIG_ZONE_DEVICE
5989 void __ref memmap_init_zone_device(struct zone *zone,
5990 unsigned long start_pfn,
5991 unsigned long nr_pages,
5992 struct dev_pagemap *pgmap)
5994 unsigned long pfn, end_pfn = start_pfn + nr_pages;
5995 struct pglist_data *pgdat = zone->zone_pgdat;
5996 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
5997 unsigned long zone_idx = zone_idx(zone);
5998 unsigned long start = jiffies;
5999 int nid = pgdat->node_id;
6001 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6005 * The call to memmap_init_zone should have already taken care
6006 * of the pages reserved for the memmap, so we can just jump to
6007 * the end of that region and start processing the device pages.
6010 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6011 nr_pages = end_pfn - start_pfn;
6014 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6015 struct page *page = pfn_to_page(pfn);
6017 __init_single_page(page, pfn, zone_idx, nid);
6020 * Mark page reserved as it will need to wait for onlining
6021 * phase for it to be fully associated with a zone.
6023 * We can use the non-atomic __set_bit operation for setting
6024 * the flag as we are still initializing the pages.
6026 __SetPageReserved(page);
6029 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6030 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6031 * ever freed or placed on a driver-private list.
6033 page->pgmap = pgmap;
6034 page->zone_device_data = NULL;
6037 * Mark the block movable so that blocks are reserved for
6038 * movable at startup. This will force kernel allocations
6039 * to reserve their blocks rather than leaking throughout
6040 * the address space during boot when many long-lived
6041 * kernel allocations are made.
6043 * bitmap is created for zone's valid pfn range. but memmap
6044 * can be created for invalid pages (for alignment)
6045 * check here not to call set_pageblock_migratetype() against
6048 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
6049 * because this is done early in section_activate()
6051 if (!(pfn & (pageblock_nr_pages - 1))) {
6052 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6057 pr_info("%s initialised %lu pages in %ums\n", __func__,
6058 nr_pages, jiffies_to_msecs(jiffies - start));
6062 static void __meminit zone_init_free_lists(struct zone *zone)
6064 unsigned int order, t;
6065 for_each_migratetype_order(order, t) {
6066 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6067 zone->free_area[order].nr_free = 0;
6071 void __meminit __weak memmap_init(unsigned long size, int nid,
6073 unsigned long range_start_pfn)
6075 unsigned long start_pfn, end_pfn;
6076 unsigned long range_end_pfn = range_start_pfn + size;
6079 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6080 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6081 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6083 if (end_pfn > start_pfn) {
6084 size = end_pfn - start_pfn;
6085 memmap_init_zone(size, nid, zone, start_pfn,
6086 MEMMAP_EARLY, NULL);
6091 static int zone_batchsize(struct zone *zone)
6097 * The per-cpu-pages pools are set to around 1000th of the
6100 batch = zone_managed_pages(zone) / 1024;
6101 /* But no more than a meg. */
6102 if (batch * PAGE_SIZE > 1024 * 1024)
6103 batch = (1024 * 1024) / PAGE_SIZE;
6104 batch /= 4; /* We effectively *= 4 below */
6109 * Clamp the batch to a 2^n - 1 value. Having a power
6110 * of 2 value was found to be more likely to have
6111 * suboptimal cache aliasing properties in some cases.
6113 * For example if 2 tasks are alternately allocating
6114 * batches of pages, one task can end up with a lot
6115 * of pages of one half of the possible page colors
6116 * and the other with pages of the other colors.
6118 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6123 /* The deferral and batching of frees should be suppressed under NOMMU
6126 * The problem is that NOMMU needs to be able to allocate large chunks
6127 * of contiguous memory as there's no hardware page translation to
6128 * assemble apparent contiguous memory from discontiguous pages.
6130 * Queueing large contiguous runs of pages for batching, however,
6131 * causes the pages to actually be freed in smaller chunks. As there
6132 * can be a significant delay between the individual batches being
6133 * recycled, this leads to the once large chunks of space being
6134 * fragmented and becoming unavailable for high-order allocations.
6141 * pcp->high and pcp->batch values are related and dependent on one another:
6142 * ->batch must never be higher then ->high.
6143 * The following function updates them in a safe manner without read side
6146 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6147 * those fields changing asynchronously (acording the the above rule).
6149 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6150 * outside of boot time (or some other assurance that no concurrent updaters
6153 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6154 unsigned long batch)
6156 /* start with a fail safe value for batch */
6160 /* Update high, then batch, in order */
6167 /* a companion to pageset_set_high() */
6168 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6170 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6173 static void pageset_init(struct per_cpu_pageset *p)
6175 struct per_cpu_pages *pcp;
6178 memset(p, 0, sizeof(*p));
6181 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6182 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6185 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6188 pageset_set_batch(p, batch);
6192 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6193 * to the value high for the pageset p.
6195 static void pageset_set_high(struct per_cpu_pageset *p,
6198 unsigned long batch = max(1UL, high / 4);
6199 if ((high / 4) > (PAGE_SHIFT * 8))
6200 batch = PAGE_SHIFT * 8;
6202 pageset_update(&p->pcp, high, batch);
6205 static void pageset_set_high_and_batch(struct zone *zone,
6206 struct per_cpu_pageset *pcp)
6208 if (percpu_pagelist_fraction)
6209 pageset_set_high(pcp,
6210 (zone_managed_pages(zone) /
6211 percpu_pagelist_fraction));
6213 pageset_set_batch(pcp, zone_batchsize(zone));
6216 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6218 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6221 pageset_set_high_and_batch(zone, pcp);
6224 void __meminit setup_zone_pageset(struct zone *zone)
6227 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6228 for_each_possible_cpu(cpu)
6229 zone_pageset_init(zone, cpu);
6233 * Allocate per cpu pagesets and initialize them.
6234 * Before this call only boot pagesets were available.
6236 void __init setup_per_cpu_pageset(void)
6238 struct pglist_data *pgdat;
6241 for_each_populated_zone(zone)
6242 setup_zone_pageset(zone);
6244 for_each_online_pgdat(pgdat)
6245 pgdat->per_cpu_nodestats =
6246 alloc_percpu(struct per_cpu_nodestat);
6249 static __meminit void zone_pcp_init(struct zone *zone)
6252 * per cpu subsystem is not up at this point. The following code
6253 * relies on the ability of the linker to provide the
6254 * offset of a (static) per cpu variable into the per cpu area.
6256 zone->pageset = &boot_pageset;
6258 if (populated_zone(zone))
6259 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6260 zone->name, zone->present_pages,
6261 zone_batchsize(zone));
6264 void __meminit init_currently_empty_zone(struct zone *zone,
6265 unsigned long zone_start_pfn,
6268 struct pglist_data *pgdat = zone->zone_pgdat;
6269 int zone_idx = zone_idx(zone) + 1;
6271 if (zone_idx > pgdat->nr_zones)
6272 pgdat->nr_zones = zone_idx;
6274 zone->zone_start_pfn = zone_start_pfn;
6276 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6277 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6279 (unsigned long)zone_idx(zone),
6280 zone_start_pfn, (zone_start_pfn + size));
6282 zone_init_free_lists(zone);
6283 zone->initialized = 1;
6287 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
6288 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
6289 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
6291 * If an architecture guarantees that all ranges registered contain no holes
6292 * and may be freed, this this function may be used instead of calling
6293 * memblock_free_early_nid() manually.
6295 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
6297 unsigned long start_pfn, end_pfn;
6300 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
6301 start_pfn = min(start_pfn, max_low_pfn);
6302 end_pfn = min(end_pfn, max_low_pfn);
6304 if (start_pfn < end_pfn)
6305 memblock_free_early_nid(PFN_PHYS(start_pfn),
6306 (end_pfn - start_pfn) << PAGE_SHIFT,
6312 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6313 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6315 * If an architecture guarantees that all ranges registered contain no holes and may
6316 * be freed, this function may be used instead of calling memory_present() manually.
6318 void __init sparse_memory_present_with_active_regions(int nid)
6320 unsigned long start_pfn, end_pfn;
6323 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
6324 memory_present(this_nid, start_pfn, end_pfn);
6328 * get_pfn_range_for_nid - Return the start and end page frames for a node
6329 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6330 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6331 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6333 * It returns the start and end page frame of a node based on information
6334 * provided by memblock_set_node(). If called for a node
6335 * with no available memory, a warning is printed and the start and end
6338 void __init get_pfn_range_for_nid(unsigned int nid,
6339 unsigned long *start_pfn, unsigned long *end_pfn)
6341 unsigned long this_start_pfn, this_end_pfn;
6347 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6348 *start_pfn = min(*start_pfn, this_start_pfn);
6349 *end_pfn = max(*end_pfn, this_end_pfn);
6352 if (*start_pfn == -1UL)
6357 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6358 * assumption is made that zones within a node are ordered in monotonic
6359 * increasing memory addresses so that the "highest" populated zone is used
6361 static void __init find_usable_zone_for_movable(void)
6364 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6365 if (zone_index == ZONE_MOVABLE)
6368 if (arch_zone_highest_possible_pfn[zone_index] >
6369 arch_zone_lowest_possible_pfn[zone_index])
6373 VM_BUG_ON(zone_index == -1);
6374 movable_zone = zone_index;
6378 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6379 * because it is sized independent of architecture. Unlike the other zones,
6380 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6381 * in each node depending on the size of each node and how evenly kernelcore
6382 * is distributed. This helper function adjusts the zone ranges
6383 * provided by the architecture for a given node by using the end of the
6384 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6385 * zones within a node are in order of monotonic increases memory addresses
6387 static void __init adjust_zone_range_for_zone_movable(int nid,
6388 unsigned long zone_type,
6389 unsigned long node_start_pfn,
6390 unsigned long node_end_pfn,
6391 unsigned long *zone_start_pfn,
6392 unsigned long *zone_end_pfn)
6394 /* Only adjust if ZONE_MOVABLE is on this node */
6395 if (zone_movable_pfn[nid]) {
6396 /* Size ZONE_MOVABLE */
6397 if (zone_type == ZONE_MOVABLE) {
6398 *zone_start_pfn = zone_movable_pfn[nid];
6399 *zone_end_pfn = min(node_end_pfn,
6400 arch_zone_highest_possible_pfn[movable_zone]);
6402 /* Adjust for ZONE_MOVABLE starting within this range */
6403 } else if (!mirrored_kernelcore &&
6404 *zone_start_pfn < zone_movable_pfn[nid] &&
6405 *zone_end_pfn > zone_movable_pfn[nid]) {
6406 *zone_end_pfn = zone_movable_pfn[nid];
6408 /* Check if this whole range is within ZONE_MOVABLE */
6409 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6410 *zone_start_pfn = *zone_end_pfn;
6415 * Return the number of pages a zone spans in a node, including holes
6416 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6418 static unsigned long __init zone_spanned_pages_in_node(int nid,
6419 unsigned long zone_type,
6420 unsigned long node_start_pfn,
6421 unsigned long node_end_pfn,
6422 unsigned long *zone_start_pfn,
6423 unsigned long *zone_end_pfn)
6425 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6426 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6427 /* When hotadd a new node from cpu_up(), the node should be empty */
6428 if (!node_start_pfn && !node_end_pfn)
6431 /* Get the start and end of the zone */
6432 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6433 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6434 adjust_zone_range_for_zone_movable(nid, zone_type,
6435 node_start_pfn, node_end_pfn,
6436 zone_start_pfn, zone_end_pfn);
6438 /* Check that this node has pages within the zone's required range */
6439 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6442 /* Move the zone boundaries inside the node if necessary */
6443 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6444 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6446 /* Return the spanned pages */
6447 return *zone_end_pfn - *zone_start_pfn;
6451 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6452 * then all holes in the requested range will be accounted for.
6454 unsigned long __init __absent_pages_in_range(int nid,
6455 unsigned long range_start_pfn,
6456 unsigned long range_end_pfn)
6458 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6459 unsigned long start_pfn, end_pfn;
6462 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6463 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6464 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6465 nr_absent -= end_pfn - start_pfn;
6471 * absent_pages_in_range - Return number of page frames in holes within a range
6472 * @start_pfn: The start PFN to start searching for holes
6473 * @end_pfn: The end PFN to stop searching for holes
6475 * Return: the number of pages frames in memory holes within a range.
6477 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6478 unsigned long end_pfn)
6480 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6483 /* Return the number of page frames in holes in a zone on a node */
6484 static unsigned long __init zone_absent_pages_in_node(int nid,
6485 unsigned long zone_type,
6486 unsigned long node_start_pfn,
6487 unsigned long node_end_pfn)
6489 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6490 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6491 unsigned long zone_start_pfn, zone_end_pfn;
6492 unsigned long nr_absent;
6494 /* When hotadd a new node from cpu_up(), the node should be empty */
6495 if (!node_start_pfn && !node_end_pfn)
6498 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6499 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6501 adjust_zone_range_for_zone_movable(nid, zone_type,
6502 node_start_pfn, node_end_pfn,
6503 &zone_start_pfn, &zone_end_pfn);
6504 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6507 * ZONE_MOVABLE handling.
6508 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6511 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6512 unsigned long start_pfn, end_pfn;
6513 struct memblock_region *r;
6515 for_each_memblock(memory, r) {
6516 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6517 zone_start_pfn, zone_end_pfn);
6518 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6519 zone_start_pfn, zone_end_pfn);
6521 if (zone_type == ZONE_MOVABLE &&
6522 memblock_is_mirror(r))
6523 nr_absent += end_pfn - start_pfn;
6525 if (zone_type == ZONE_NORMAL &&
6526 !memblock_is_mirror(r))
6527 nr_absent += end_pfn - start_pfn;
6534 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6535 unsigned long node_start_pfn,
6536 unsigned long node_end_pfn)
6538 unsigned long realtotalpages = 0, totalpages = 0;
6541 for (i = 0; i < MAX_NR_ZONES; i++) {
6542 struct zone *zone = pgdat->node_zones + i;
6543 unsigned long zone_start_pfn, zone_end_pfn;
6544 unsigned long spanned, absent;
6545 unsigned long size, real_size;
6547 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
6552 absent = zone_absent_pages_in_node(pgdat->node_id, i,
6557 real_size = size - absent;
6560 zone->zone_start_pfn = zone_start_pfn;
6562 zone->zone_start_pfn = 0;
6563 zone->spanned_pages = size;
6564 zone->present_pages = real_size;
6567 realtotalpages += real_size;
6570 pgdat->node_spanned_pages = totalpages;
6571 pgdat->node_present_pages = realtotalpages;
6572 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6576 #ifndef CONFIG_SPARSEMEM
6578 * Calculate the size of the zone->blockflags rounded to an unsigned long
6579 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6580 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6581 * round what is now in bits to nearest long in bits, then return it in
6584 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6586 unsigned long usemapsize;
6588 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6589 usemapsize = roundup(zonesize, pageblock_nr_pages);
6590 usemapsize = usemapsize >> pageblock_order;
6591 usemapsize *= NR_PAGEBLOCK_BITS;
6592 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6594 return usemapsize / 8;
6597 static void __ref setup_usemap(struct pglist_data *pgdat,
6599 unsigned long zone_start_pfn,
6600 unsigned long zonesize)
6602 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6603 zone->pageblock_flags = NULL;
6605 zone->pageblock_flags =
6606 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6608 if (!zone->pageblock_flags)
6609 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6610 usemapsize, zone->name, pgdat->node_id);
6614 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6615 unsigned long zone_start_pfn, unsigned long zonesize) {}
6616 #endif /* CONFIG_SPARSEMEM */
6618 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6620 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6621 void __init set_pageblock_order(void)
6625 /* Check that pageblock_nr_pages has not already been setup */
6626 if (pageblock_order)
6629 if (HPAGE_SHIFT > PAGE_SHIFT)
6630 order = HUGETLB_PAGE_ORDER;
6632 order = MAX_ORDER - 1;
6635 * Assume the largest contiguous order of interest is a huge page.
6636 * This value may be variable depending on boot parameters on IA64 and
6639 pageblock_order = order;
6641 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6644 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6645 * is unused as pageblock_order is set at compile-time. See
6646 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6649 void __init set_pageblock_order(void)
6653 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6655 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6656 unsigned long present_pages)
6658 unsigned long pages = spanned_pages;
6661 * Provide a more accurate estimation if there are holes within
6662 * the zone and SPARSEMEM is in use. If there are holes within the
6663 * zone, each populated memory region may cost us one or two extra
6664 * memmap pages due to alignment because memmap pages for each
6665 * populated regions may not be naturally aligned on page boundary.
6666 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6668 if (spanned_pages > present_pages + (present_pages >> 4) &&
6669 IS_ENABLED(CONFIG_SPARSEMEM))
6670 pages = present_pages;
6672 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6675 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6676 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6678 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6680 spin_lock_init(&ds_queue->split_queue_lock);
6681 INIT_LIST_HEAD(&ds_queue->split_queue);
6682 ds_queue->split_queue_len = 0;
6685 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6688 #ifdef CONFIG_COMPACTION
6689 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6691 init_waitqueue_head(&pgdat->kcompactd_wait);
6694 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6697 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6699 pgdat_resize_init(pgdat);
6701 pgdat_init_split_queue(pgdat);
6702 pgdat_init_kcompactd(pgdat);
6704 init_waitqueue_head(&pgdat->kswapd_wait);
6705 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6707 pgdat_page_ext_init(pgdat);
6708 spin_lock_init(&pgdat->lru_lock);
6709 lruvec_init(&pgdat->__lruvec);
6712 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6713 unsigned long remaining_pages)
6715 atomic_long_set(&zone->managed_pages, remaining_pages);
6716 zone_set_nid(zone, nid);
6717 zone->name = zone_names[idx];
6718 zone->zone_pgdat = NODE_DATA(nid);
6719 spin_lock_init(&zone->lock);
6720 zone_seqlock_init(zone);
6721 zone_pcp_init(zone);
6725 * Set up the zone data structures
6726 * - init pgdat internals
6727 * - init all zones belonging to this node
6729 * NOTE: this function is only called during memory hotplug
6731 #ifdef CONFIG_MEMORY_HOTPLUG
6732 void __ref free_area_init_core_hotplug(int nid)
6735 pg_data_t *pgdat = NODE_DATA(nid);
6737 pgdat_init_internals(pgdat);
6738 for (z = 0; z < MAX_NR_ZONES; z++)
6739 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6744 * Set up the zone data structures:
6745 * - mark all pages reserved
6746 * - mark all memory queues empty
6747 * - clear the memory bitmaps
6749 * NOTE: pgdat should get zeroed by caller.
6750 * NOTE: this function is only called during early init.
6752 static void __init free_area_init_core(struct pglist_data *pgdat)
6755 int nid = pgdat->node_id;
6757 pgdat_init_internals(pgdat);
6758 pgdat->per_cpu_nodestats = &boot_nodestats;
6760 for (j = 0; j < MAX_NR_ZONES; j++) {
6761 struct zone *zone = pgdat->node_zones + j;
6762 unsigned long size, freesize, memmap_pages;
6763 unsigned long zone_start_pfn = zone->zone_start_pfn;
6765 size = zone->spanned_pages;
6766 freesize = zone->present_pages;
6769 * Adjust freesize so that it accounts for how much memory
6770 * is used by this zone for memmap. This affects the watermark
6771 * and per-cpu initialisations
6773 memmap_pages = calc_memmap_size(size, freesize);
6774 if (!is_highmem_idx(j)) {
6775 if (freesize >= memmap_pages) {
6776 freesize -= memmap_pages;
6779 " %s zone: %lu pages used for memmap\n",
6780 zone_names[j], memmap_pages);
6782 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6783 zone_names[j], memmap_pages, freesize);
6786 /* Account for reserved pages */
6787 if (j == 0 && freesize > dma_reserve) {
6788 freesize -= dma_reserve;
6789 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6790 zone_names[0], dma_reserve);
6793 if (!is_highmem_idx(j))
6794 nr_kernel_pages += freesize;
6795 /* Charge for highmem memmap if there are enough kernel pages */
6796 else if (nr_kernel_pages > memmap_pages * 2)
6797 nr_kernel_pages -= memmap_pages;
6798 nr_all_pages += freesize;
6801 * Set an approximate value for lowmem here, it will be adjusted
6802 * when the bootmem allocator frees pages into the buddy system.
6803 * And all highmem pages will be managed by the buddy system.
6805 zone_init_internals(zone, j, nid, freesize);
6810 set_pageblock_order();
6811 setup_usemap(pgdat, zone, zone_start_pfn, size);
6812 init_currently_empty_zone(zone, zone_start_pfn, size);
6813 memmap_init(size, nid, j, zone_start_pfn);
6817 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6818 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6820 unsigned long __maybe_unused start = 0;
6821 unsigned long __maybe_unused offset = 0;
6823 /* Skip empty nodes */
6824 if (!pgdat->node_spanned_pages)
6827 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6828 offset = pgdat->node_start_pfn - start;
6829 /* ia64 gets its own node_mem_map, before this, without bootmem */
6830 if (!pgdat->node_mem_map) {
6831 unsigned long size, end;
6835 * The zone's endpoints aren't required to be MAX_ORDER
6836 * aligned but the node_mem_map endpoints must be in order
6837 * for the buddy allocator to function correctly.
6839 end = pgdat_end_pfn(pgdat);
6840 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6841 size = (end - start) * sizeof(struct page);
6842 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6845 panic("Failed to allocate %ld bytes for node %d memory map\n",
6846 size, pgdat->node_id);
6847 pgdat->node_mem_map = map + offset;
6849 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6850 __func__, pgdat->node_id, (unsigned long)pgdat,
6851 (unsigned long)pgdat->node_mem_map);
6852 #ifndef CONFIG_NEED_MULTIPLE_NODES
6854 * With no DISCONTIG, the global mem_map is just set as node 0's
6856 if (pgdat == NODE_DATA(0)) {
6857 mem_map = NODE_DATA(0)->node_mem_map;
6858 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6864 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6865 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6867 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6868 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6870 pgdat->first_deferred_pfn = ULONG_MAX;
6873 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6876 static void __init free_area_init_node(int nid)
6878 pg_data_t *pgdat = NODE_DATA(nid);
6879 unsigned long start_pfn = 0;
6880 unsigned long end_pfn = 0;
6882 /* pg_data_t should be reset to zero when it's allocated */
6883 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6885 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6887 pgdat->node_id = nid;
6888 pgdat->node_start_pfn = start_pfn;
6889 pgdat->per_cpu_nodestats = NULL;
6891 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6892 (u64)start_pfn << PAGE_SHIFT,
6893 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6894 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
6896 alloc_node_mem_map(pgdat);
6897 pgdat_set_deferred_range(pgdat);
6899 free_area_init_core(pgdat);
6902 void __init free_area_init_memoryless_node(int nid)
6904 free_area_init_node(nid);
6907 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6909 * Initialize all valid struct pages in the range [spfn, epfn) and mark them
6910 * PageReserved(). Return the number of struct pages that were initialized.
6912 static u64 __init init_unavailable_range(unsigned long spfn, unsigned long epfn)
6917 for (pfn = spfn; pfn < epfn; pfn++) {
6918 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6919 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6920 + pageblock_nr_pages - 1;
6924 * Use a fake node/zone (0) for now. Some of these pages
6925 * (in memblock.reserved but not in memblock.memory) will
6926 * get re-initialized via reserve_bootmem_region() later.
6928 __init_single_page(pfn_to_page(pfn), pfn, 0, 0);
6929 __SetPageReserved(pfn_to_page(pfn));
6937 * Only struct pages that are backed by physical memory are zeroed and
6938 * initialized by going through __init_single_page(). But, there are some
6939 * struct pages which are reserved in memblock allocator and their fields
6940 * may be accessed (for example page_to_pfn() on some configuration accesses
6941 * flags). We must explicitly initialize those struct pages.
6943 * This function also addresses a similar issue where struct pages are left
6944 * uninitialized because the physical address range is not covered by
6945 * memblock.memory or memblock.reserved. That could happen when memblock
6946 * layout is manually configured via memmap=, or when the highest physical
6947 * address (max_pfn) does not end on a section boundary.
6949 static void __init init_unavailable_mem(void)
6951 phys_addr_t start, end;
6953 phys_addr_t next = 0;
6956 * Loop through unavailable ranges not covered by memblock.memory.
6959 for_each_mem_range(i, &memblock.memory, NULL,
6960 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
6962 pgcnt += init_unavailable_range(PFN_DOWN(next),
6968 * Early sections always have a fully populated memmap for the whole
6969 * section - see pfn_valid(). If the last section has holes at the
6970 * end and that section is marked "online", the memmap will be
6971 * considered initialized. Make sure that memmap has a well defined
6974 pgcnt += init_unavailable_range(PFN_DOWN(next),
6975 round_up(max_pfn, PAGES_PER_SECTION));
6978 * Struct pages that do not have backing memory. This could be because
6979 * firmware is using some of this memory, or for some other reasons.
6982 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
6985 static inline void __init init_unavailable_mem(void)
6988 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
6990 #if MAX_NUMNODES > 1
6992 * Figure out the number of possible node ids.
6994 void __init setup_nr_node_ids(void)
6996 unsigned int highest;
6998 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6999 nr_node_ids = highest + 1;
7004 * node_map_pfn_alignment - determine the maximum internode alignment
7006 * This function should be called after node map is populated and sorted.
7007 * It calculates the maximum power of two alignment which can distinguish
7010 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7011 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7012 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7013 * shifted, 1GiB is enough and this function will indicate so.
7015 * This is used to test whether pfn -> nid mapping of the chosen memory
7016 * model has fine enough granularity to avoid incorrect mapping for the
7017 * populated node map.
7019 * Return: the determined alignment in pfn's. 0 if there is no alignment
7020 * requirement (single node).
7022 unsigned long __init node_map_pfn_alignment(void)
7024 unsigned long accl_mask = 0, last_end = 0;
7025 unsigned long start, end, mask;
7026 int last_nid = NUMA_NO_NODE;
7029 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7030 if (!start || last_nid < 0 || last_nid == nid) {
7037 * Start with a mask granular enough to pin-point to the
7038 * start pfn and tick off bits one-by-one until it becomes
7039 * too coarse to separate the current node from the last.
7041 mask = ~((1 << __ffs(start)) - 1);
7042 while (mask && last_end <= (start & (mask << 1)))
7045 /* accumulate all internode masks */
7049 /* convert mask to number of pages */
7050 return ~accl_mask + 1;
7054 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7056 * Return: the minimum PFN based on information provided via
7057 * memblock_set_node().
7059 unsigned long __init find_min_pfn_with_active_regions(void)
7061 return PHYS_PFN(memblock_start_of_DRAM());
7065 * early_calculate_totalpages()
7066 * Sum pages in active regions for movable zone.
7067 * Populate N_MEMORY for calculating usable_nodes.
7069 static unsigned long __init early_calculate_totalpages(void)
7071 unsigned long totalpages = 0;
7072 unsigned long start_pfn, end_pfn;
7075 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7076 unsigned long pages = end_pfn - start_pfn;
7078 totalpages += pages;
7080 node_set_state(nid, N_MEMORY);
7086 * Find the PFN the Movable zone begins in each node. Kernel memory
7087 * is spread evenly between nodes as long as the nodes have enough
7088 * memory. When they don't, some nodes will have more kernelcore than
7091 static void __init find_zone_movable_pfns_for_nodes(void)
7094 unsigned long usable_startpfn;
7095 unsigned long kernelcore_node, kernelcore_remaining;
7096 /* save the state before borrow the nodemask */
7097 nodemask_t saved_node_state = node_states[N_MEMORY];
7098 unsigned long totalpages = early_calculate_totalpages();
7099 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7100 struct memblock_region *r;
7102 /* Need to find movable_zone earlier when movable_node is specified. */
7103 find_usable_zone_for_movable();
7106 * If movable_node is specified, ignore kernelcore and movablecore
7109 if (movable_node_is_enabled()) {
7110 for_each_memblock(memory, r) {
7111 if (!memblock_is_hotpluggable(r))
7114 nid = memblock_get_region_node(r);
7116 usable_startpfn = PFN_DOWN(r->base);
7117 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7118 min(usable_startpfn, zone_movable_pfn[nid]) :
7126 * If kernelcore=mirror is specified, ignore movablecore option
7128 if (mirrored_kernelcore) {
7129 bool mem_below_4gb_not_mirrored = false;
7131 for_each_memblock(memory, r) {
7132 if (memblock_is_mirror(r))
7135 nid = memblock_get_region_node(r);
7137 usable_startpfn = memblock_region_memory_base_pfn(r);
7139 if (usable_startpfn < 0x100000) {
7140 mem_below_4gb_not_mirrored = true;
7144 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7145 min(usable_startpfn, zone_movable_pfn[nid]) :
7149 if (mem_below_4gb_not_mirrored)
7150 pr_warn("This configuration results in unmirrored kernel memory.");
7156 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7157 * amount of necessary memory.
7159 if (required_kernelcore_percent)
7160 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7162 if (required_movablecore_percent)
7163 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7167 * If movablecore= was specified, calculate what size of
7168 * kernelcore that corresponds so that memory usable for
7169 * any allocation type is evenly spread. If both kernelcore
7170 * and movablecore are specified, then the value of kernelcore
7171 * will be used for required_kernelcore if it's greater than
7172 * what movablecore would have allowed.
7174 if (required_movablecore) {
7175 unsigned long corepages;
7178 * Round-up so that ZONE_MOVABLE is at least as large as what
7179 * was requested by the user
7181 required_movablecore =
7182 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7183 required_movablecore = min(totalpages, required_movablecore);
7184 corepages = totalpages - required_movablecore;
7186 required_kernelcore = max(required_kernelcore, corepages);
7190 * If kernelcore was not specified or kernelcore size is larger
7191 * than totalpages, there is no ZONE_MOVABLE.
7193 if (!required_kernelcore || required_kernelcore >= totalpages)
7196 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7197 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7200 /* Spread kernelcore memory as evenly as possible throughout nodes */
7201 kernelcore_node = required_kernelcore / usable_nodes;
7202 for_each_node_state(nid, N_MEMORY) {
7203 unsigned long start_pfn, end_pfn;
7206 * Recalculate kernelcore_node if the division per node
7207 * now exceeds what is necessary to satisfy the requested
7208 * amount of memory for the kernel
7210 if (required_kernelcore < kernelcore_node)
7211 kernelcore_node = required_kernelcore / usable_nodes;
7214 * As the map is walked, we track how much memory is usable
7215 * by the kernel using kernelcore_remaining. When it is
7216 * 0, the rest of the node is usable by ZONE_MOVABLE
7218 kernelcore_remaining = kernelcore_node;
7220 /* Go through each range of PFNs within this node */
7221 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7222 unsigned long size_pages;
7224 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7225 if (start_pfn >= end_pfn)
7228 /* Account for what is only usable for kernelcore */
7229 if (start_pfn < usable_startpfn) {
7230 unsigned long kernel_pages;
7231 kernel_pages = min(end_pfn, usable_startpfn)
7234 kernelcore_remaining -= min(kernel_pages,
7235 kernelcore_remaining);
7236 required_kernelcore -= min(kernel_pages,
7237 required_kernelcore);
7239 /* Continue if range is now fully accounted */
7240 if (end_pfn <= usable_startpfn) {
7243 * Push zone_movable_pfn to the end so
7244 * that if we have to rebalance
7245 * kernelcore across nodes, we will
7246 * not double account here
7248 zone_movable_pfn[nid] = end_pfn;
7251 start_pfn = usable_startpfn;
7255 * The usable PFN range for ZONE_MOVABLE is from
7256 * start_pfn->end_pfn. Calculate size_pages as the
7257 * number of pages used as kernelcore
7259 size_pages = end_pfn - start_pfn;
7260 if (size_pages > kernelcore_remaining)
7261 size_pages = kernelcore_remaining;
7262 zone_movable_pfn[nid] = start_pfn + size_pages;
7265 * Some kernelcore has been met, update counts and
7266 * break if the kernelcore for this node has been
7269 required_kernelcore -= min(required_kernelcore,
7271 kernelcore_remaining -= size_pages;
7272 if (!kernelcore_remaining)
7278 * If there is still required_kernelcore, we do another pass with one
7279 * less node in the count. This will push zone_movable_pfn[nid] further
7280 * along on the nodes that still have memory until kernelcore is
7284 if (usable_nodes && required_kernelcore > usable_nodes)
7288 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7289 for (nid = 0; nid < MAX_NUMNODES; nid++)
7290 zone_movable_pfn[nid] =
7291 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7294 /* restore the node_state */
7295 node_states[N_MEMORY] = saved_node_state;
7298 /* Any regular or high memory on that node ? */
7299 static void check_for_memory(pg_data_t *pgdat, int nid)
7301 enum zone_type zone_type;
7303 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7304 struct zone *zone = &pgdat->node_zones[zone_type];
7305 if (populated_zone(zone)) {
7306 if (IS_ENABLED(CONFIG_HIGHMEM))
7307 node_set_state(nid, N_HIGH_MEMORY);
7308 if (zone_type <= ZONE_NORMAL)
7309 node_set_state(nid, N_NORMAL_MEMORY);
7316 * Some architecturs, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7317 * such cases we allow max_zone_pfn sorted in the descending order
7319 bool __weak arch_has_descending_max_zone_pfns(void)
7325 * free_area_init - Initialise all pg_data_t and zone data
7326 * @max_zone_pfn: an array of max PFNs for each zone
7328 * This will call free_area_init_node() for each active node in the system.
7329 * Using the page ranges provided by memblock_set_node(), the size of each
7330 * zone in each node and their holes is calculated. If the maximum PFN
7331 * between two adjacent zones match, it is assumed that the zone is empty.
7332 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7333 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7334 * starts where the previous one ended. For example, ZONE_DMA32 starts
7335 * at arch_max_dma_pfn.
7337 void __init free_area_init(unsigned long *max_zone_pfn)
7339 unsigned long start_pfn, end_pfn;
7343 /* Record where the zone boundaries are */
7344 memset(arch_zone_lowest_possible_pfn, 0,
7345 sizeof(arch_zone_lowest_possible_pfn));
7346 memset(arch_zone_highest_possible_pfn, 0,
7347 sizeof(arch_zone_highest_possible_pfn));
7349 start_pfn = find_min_pfn_with_active_regions();
7350 descending = arch_has_descending_max_zone_pfns();
7352 for (i = 0; i < MAX_NR_ZONES; i++) {
7354 zone = MAX_NR_ZONES - i - 1;
7358 if (zone == ZONE_MOVABLE)
7361 end_pfn = max(max_zone_pfn[zone], start_pfn);
7362 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7363 arch_zone_highest_possible_pfn[zone] = end_pfn;
7365 start_pfn = end_pfn;
7368 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7369 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7370 find_zone_movable_pfns_for_nodes();
7372 /* Print out the zone ranges */
7373 pr_info("Zone ranges:\n");
7374 for (i = 0; i < MAX_NR_ZONES; i++) {
7375 if (i == ZONE_MOVABLE)
7377 pr_info(" %-8s ", zone_names[i]);
7378 if (arch_zone_lowest_possible_pfn[i] ==
7379 arch_zone_highest_possible_pfn[i])
7382 pr_cont("[mem %#018Lx-%#018Lx]\n",
7383 (u64)arch_zone_lowest_possible_pfn[i]
7385 ((u64)arch_zone_highest_possible_pfn[i]
7386 << PAGE_SHIFT) - 1);
7389 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7390 pr_info("Movable zone start for each node\n");
7391 for (i = 0; i < MAX_NUMNODES; i++) {
7392 if (zone_movable_pfn[i])
7393 pr_info(" Node %d: %#018Lx\n", i,
7394 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7398 * Print out the early node map, and initialize the
7399 * subsection-map relative to active online memory ranges to
7400 * enable future "sub-section" extensions of the memory map.
7402 pr_info("Early memory node ranges\n");
7403 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7404 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7405 (u64)start_pfn << PAGE_SHIFT,
7406 ((u64)end_pfn << PAGE_SHIFT) - 1);
7407 subsection_map_init(start_pfn, end_pfn - start_pfn);
7410 /* Initialise every node */
7411 mminit_verify_pageflags_layout();
7412 setup_nr_node_ids();
7413 init_unavailable_mem();
7414 for_each_online_node(nid) {
7415 pg_data_t *pgdat = NODE_DATA(nid);
7416 free_area_init_node(nid);
7418 /* Any memory on that node */
7419 if (pgdat->node_present_pages)
7420 node_set_state(nid, N_MEMORY);
7421 check_for_memory(pgdat, nid);
7425 static int __init cmdline_parse_core(char *p, unsigned long *core,
7426 unsigned long *percent)
7428 unsigned long long coremem;
7434 /* Value may be a percentage of total memory, otherwise bytes */
7435 coremem = simple_strtoull(p, &endptr, 0);
7436 if (*endptr == '%') {
7437 /* Paranoid check for percent values greater than 100 */
7438 WARN_ON(coremem > 100);
7442 coremem = memparse(p, &p);
7443 /* Paranoid check that UL is enough for the coremem value */
7444 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7446 *core = coremem >> PAGE_SHIFT;
7453 * kernelcore=size sets the amount of memory for use for allocations that
7454 * cannot be reclaimed or migrated.
7456 static int __init cmdline_parse_kernelcore(char *p)
7458 /* parse kernelcore=mirror */
7459 if (parse_option_str(p, "mirror")) {
7460 mirrored_kernelcore = true;
7464 return cmdline_parse_core(p, &required_kernelcore,
7465 &required_kernelcore_percent);
7469 * movablecore=size sets the amount of memory for use for allocations that
7470 * can be reclaimed or migrated.
7472 static int __init cmdline_parse_movablecore(char *p)
7474 return cmdline_parse_core(p, &required_movablecore,
7475 &required_movablecore_percent);
7478 early_param("kernelcore", cmdline_parse_kernelcore);
7479 early_param("movablecore", cmdline_parse_movablecore);
7481 void adjust_managed_page_count(struct page *page, long count)
7483 atomic_long_add(count, &page_zone(page)->managed_pages);
7484 totalram_pages_add(count);
7485 #ifdef CONFIG_HIGHMEM
7486 if (PageHighMem(page))
7487 totalhigh_pages_add(count);
7490 EXPORT_SYMBOL(adjust_managed_page_count);
7492 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7495 unsigned long pages = 0;
7497 start = (void *)PAGE_ALIGN((unsigned long)start);
7498 end = (void *)((unsigned long)end & PAGE_MASK);
7499 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7500 struct page *page = virt_to_page(pos);
7501 void *direct_map_addr;
7504 * 'direct_map_addr' might be different from 'pos'
7505 * because some architectures' virt_to_page()
7506 * work with aliases. Getting the direct map
7507 * address ensures that we get a _writeable_
7508 * alias for the memset().
7510 direct_map_addr = page_address(page);
7511 if ((unsigned int)poison <= 0xFF)
7512 memset(direct_map_addr, poison, PAGE_SIZE);
7514 free_reserved_page(page);
7518 pr_info("Freeing %s memory: %ldK\n",
7519 s, pages << (PAGE_SHIFT - 10));
7524 #ifdef CONFIG_HIGHMEM
7525 void free_highmem_page(struct page *page)
7527 __free_reserved_page(page);
7528 totalram_pages_inc();
7529 atomic_long_inc(&page_zone(page)->managed_pages);
7530 totalhigh_pages_inc();
7535 void __init mem_init_print_info(const char *str)
7537 unsigned long physpages, codesize, datasize, rosize, bss_size;
7538 unsigned long init_code_size, init_data_size;
7540 physpages = get_num_physpages();
7541 codesize = _etext - _stext;
7542 datasize = _edata - _sdata;
7543 rosize = __end_rodata - __start_rodata;
7544 bss_size = __bss_stop - __bss_start;
7545 init_data_size = __init_end - __init_begin;
7546 init_code_size = _einittext - _sinittext;
7549 * Detect special cases and adjust section sizes accordingly:
7550 * 1) .init.* may be embedded into .data sections
7551 * 2) .init.text.* may be out of [__init_begin, __init_end],
7552 * please refer to arch/tile/kernel/vmlinux.lds.S.
7553 * 3) .rodata.* may be embedded into .text or .data sections.
7555 #define adj_init_size(start, end, size, pos, adj) \
7557 if (start <= pos && pos < end && size > adj) \
7561 adj_init_size(__init_begin, __init_end, init_data_size,
7562 _sinittext, init_code_size);
7563 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7564 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7565 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7566 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7568 #undef adj_init_size
7570 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7571 #ifdef CONFIG_HIGHMEM
7575 nr_free_pages() << (PAGE_SHIFT - 10),
7576 physpages << (PAGE_SHIFT - 10),
7577 codesize >> 10, datasize >> 10, rosize >> 10,
7578 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7579 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7580 totalcma_pages << (PAGE_SHIFT - 10),
7581 #ifdef CONFIG_HIGHMEM
7582 totalhigh_pages() << (PAGE_SHIFT - 10),
7584 str ? ", " : "", str ? str : "");
7588 * set_dma_reserve - set the specified number of pages reserved in the first zone
7589 * @new_dma_reserve: The number of pages to mark reserved
7591 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7592 * In the DMA zone, a significant percentage may be consumed by kernel image
7593 * and other unfreeable allocations which can skew the watermarks badly. This
7594 * function may optionally be used to account for unfreeable pages in the
7595 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7596 * smaller per-cpu batchsize.
7598 void __init set_dma_reserve(unsigned long new_dma_reserve)
7600 dma_reserve = new_dma_reserve;
7603 static int page_alloc_cpu_dead(unsigned int cpu)
7606 lru_add_drain_cpu(cpu);
7610 * Spill the event counters of the dead processor
7611 * into the current processors event counters.
7612 * This artificially elevates the count of the current
7615 vm_events_fold_cpu(cpu);
7618 * Zero the differential counters of the dead processor
7619 * so that the vm statistics are consistent.
7621 * This is only okay since the processor is dead and cannot
7622 * race with what we are doing.
7624 cpu_vm_stats_fold(cpu);
7629 int hashdist = HASHDIST_DEFAULT;
7631 static int __init set_hashdist(char *str)
7635 hashdist = simple_strtoul(str, &str, 0);
7638 __setup("hashdist=", set_hashdist);
7641 void __init page_alloc_init(void)
7646 if (num_node_state(N_MEMORY) == 1)
7650 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7651 "mm/page_alloc:dead", NULL,
7652 page_alloc_cpu_dead);
7657 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7658 * or min_free_kbytes changes.
7660 static void calculate_totalreserve_pages(void)
7662 struct pglist_data *pgdat;
7663 unsigned long reserve_pages = 0;
7664 enum zone_type i, j;
7666 for_each_online_pgdat(pgdat) {
7668 pgdat->totalreserve_pages = 0;
7670 for (i = 0; i < MAX_NR_ZONES; i++) {
7671 struct zone *zone = pgdat->node_zones + i;
7673 unsigned long managed_pages = zone_managed_pages(zone);
7675 /* Find valid and maximum lowmem_reserve in the zone */
7676 for (j = i; j < MAX_NR_ZONES; j++) {
7677 if (zone->lowmem_reserve[j] > max)
7678 max = zone->lowmem_reserve[j];
7681 /* we treat the high watermark as reserved pages. */
7682 max += high_wmark_pages(zone);
7684 if (max > managed_pages)
7685 max = managed_pages;
7687 pgdat->totalreserve_pages += max;
7689 reserve_pages += max;
7692 totalreserve_pages = reserve_pages;
7696 * setup_per_zone_lowmem_reserve - called whenever
7697 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7698 * has a correct pages reserved value, so an adequate number of
7699 * pages are left in the zone after a successful __alloc_pages().
7701 static void setup_per_zone_lowmem_reserve(void)
7703 struct pglist_data *pgdat;
7704 enum zone_type j, idx;
7706 for_each_online_pgdat(pgdat) {
7707 for (j = 0; j < MAX_NR_ZONES; j++) {
7708 struct zone *zone = pgdat->node_zones + j;
7709 unsigned long managed_pages = zone_managed_pages(zone);
7711 zone->lowmem_reserve[j] = 0;
7715 struct zone *lower_zone;
7718 lower_zone = pgdat->node_zones + idx;
7720 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7721 sysctl_lowmem_reserve_ratio[idx] = 0;
7722 lower_zone->lowmem_reserve[j] = 0;
7724 lower_zone->lowmem_reserve[j] =
7725 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7727 managed_pages += zone_managed_pages(lower_zone);
7732 /* update totalreserve_pages */
7733 calculate_totalreserve_pages();
7736 static void __setup_per_zone_wmarks(void)
7738 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7739 unsigned long lowmem_pages = 0;
7741 unsigned long flags;
7743 /* Calculate total number of !ZONE_HIGHMEM pages */
7744 for_each_zone(zone) {
7745 if (!is_highmem(zone))
7746 lowmem_pages += zone_managed_pages(zone);
7749 for_each_zone(zone) {
7752 spin_lock_irqsave(&zone->lock, flags);
7753 tmp = (u64)pages_min * zone_managed_pages(zone);
7754 do_div(tmp, lowmem_pages);
7755 if (is_highmem(zone)) {
7757 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7758 * need highmem pages, so cap pages_min to a small
7761 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7762 * deltas control async page reclaim, and so should
7763 * not be capped for highmem.
7765 unsigned long min_pages;
7767 min_pages = zone_managed_pages(zone) / 1024;
7768 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7769 zone->_watermark[WMARK_MIN] = min_pages;
7772 * If it's a lowmem zone, reserve a number of pages
7773 * proportionate to the zone's size.
7775 zone->_watermark[WMARK_MIN] = tmp;
7779 * Set the kswapd watermarks distance according to the
7780 * scale factor in proportion to available memory, but
7781 * ensure a minimum size on small systems.
7783 tmp = max_t(u64, tmp >> 2,
7784 mult_frac(zone_managed_pages(zone),
7785 watermark_scale_factor, 10000));
7787 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7788 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7789 zone->watermark_boost = 0;
7791 spin_unlock_irqrestore(&zone->lock, flags);
7794 /* update totalreserve_pages */
7795 calculate_totalreserve_pages();
7799 * setup_per_zone_wmarks - called when min_free_kbytes changes
7800 * or when memory is hot-{added|removed}
7802 * Ensures that the watermark[min,low,high] values for each zone are set
7803 * correctly with respect to min_free_kbytes.
7805 void setup_per_zone_wmarks(void)
7807 static DEFINE_SPINLOCK(lock);
7810 __setup_per_zone_wmarks();
7815 * Initialise min_free_kbytes.
7817 * For small machines we want it small (128k min). For large machines
7818 * we want it large (64MB max). But it is not linear, because network
7819 * bandwidth does not increase linearly with machine size. We use
7821 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7822 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7838 int __meminit init_per_zone_wmark_min(void)
7840 unsigned long lowmem_kbytes;
7841 int new_min_free_kbytes;
7843 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7844 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7846 if (new_min_free_kbytes > user_min_free_kbytes) {
7847 min_free_kbytes = new_min_free_kbytes;
7848 if (min_free_kbytes < 128)
7849 min_free_kbytes = 128;
7850 if (min_free_kbytes > 262144)
7851 min_free_kbytes = 262144;
7853 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7854 new_min_free_kbytes, user_min_free_kbytes);
7856 setup_per_zone_wmarks();
7857 refresh_zone_stat_thresholds();
7858 setup_per_zone_lowmem_reserve();
7861 setup_min_unmapped_ratio();
7862 setup_min_slab_ratio();
7867 core_initcall(init_per_zone_wmark_min)
7870 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7871 * that we can call two helper functions whenever min_free_kbytes
7874 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7875 void __user *buffer, size_t *length, loff_t *ppos)
7879 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7884 user_min_free_kbytes = min_free_kbytes;
7885 setup_per_zone_wmarks();
7890 int watermark_boost_factor_sysctl_handler(struct ctl_table *table, int write,
7891 void __user *buffer, size_t *length, loff_t *ppos)
7895 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7902 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7903 void __user *buffer, size_t *length, loff_t *ppos)
7907 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7912 setup_per_zone_wmarks();
7918 static void setup_min_unmapped_ratio(void)
7923 for_each_online_pgdat(pgdat)
7924 pgdat->min_unmapped_pages = 0;
7927 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7928 sysctl_min_unmapped_ratio) / 100;
7932 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7933 void __user *buffer, size_t *length, loff_t *ppos)
7937 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7941 setup_min_unmapped_ratio();
7946 static void setup_min_slab_ratio(void)
7951 for_each_online_pgdat(pgdat)
7952 pgdat->min_slab_pages = 0;
7955 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7956 sysctl_min_slab_ratio) / 100;
7959 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7960 void __user *buffer, size_t *length, loff_t *ppos)
7964 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7968 setup_min_slab_ratio();
7975 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7976 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7977 * whenever sysctl_lowmem_reserve_ratio changes.
7979 * The reserve ratio obviously has absolutely no relation with the
7980 * minimum watermarks. The lowmem reserve ratio can only make sense
7981 * if in function of the boot time zone sizes.
7983 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7984 void __user *buffer, size_t *length, loff_t *ppos)
7986 proc_dointvec_minmax(table, write, buffer, length, ppos);
7987 setup_per_zone_lowmem_reserve();
7991 static void __zone_pcp_update(struct zone *zone)
7995 for_each_possible_cpu(cpu)
7996 pageset_set_high_and_batch(zone,
7997 per_cpu_ptr(zone->pageset, cpu));
8001 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8002 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8003 * pagelist can have before it gets flushed back to buddy allocator.
8005 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8006 void __user *buffer, size_t *length, loff_t *ppos)
8009 int old_percpu_pagelist_fraction;
8012 mutex_lock(&pcp_batch_high_lock);
8013 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8015 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8016 if (!write || ret < 0)
8019 /* Sanity checking to avoid pcp imbalance */
8020 if (percpu_pagelist_fraction &&
8021 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8022 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8028 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8031 for_each_populated_zone(zone)
8032 __zone_pcp_update(zone);
8034 mutex_unlock(&pcp_batch_high_lock);
8038 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8040 * Returns the number of pages that arch has reserved but
8041 * is not known to alloc_large_system_hash().
8043 static unsigned long __init arch_reserved_kernel_pages(void)
8050 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8051 * machines. As memory size is increased the scale is also increased but at
8052 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8053 * quadruples the scale is increased by one, which means the size of hash table
8054 * only doubles, instead of quadrupling as well.
8055 * Because 32-bit systems cannot have large physical memory, where this scaling
8056 * makes sense, it is disabled on such platforms.
8058 #if __BITS_PER_LONG > 32
8059 #define ADAPT_SCALE_BASE (64ul << 30)
8060 #define ADAPT_SCALE_SHIFT 2
8061 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8065 * allocate a large system hash table from bootmem
8066 * - it is assumed that the hash table must contain an exact power-of-2
8067 * quantity of entries
8068 * - limit is the number of hash buckets, not the total allocation size
8070 void *__init alloc_large_system_hash(const char *tablename,
8071 unsigned long bucketsize,
8072 unsigned long numentries,
8075 unsigned int *_hash_shift,
8076 unsigned int *_hash_mask,
8077 unsigned long low_limit,
8078 unsigned long high_limit)
8080 unsigned long long max = high_limit;
8081 unsigned long log2qty, size;
8086 /* allow the kernel cmdline to have a say */
8088 /* round applicable memory size up to nearest megabyte */
8089 numentries = nr_kernel_pages;
8090 numentries -= arch_reserved_kernel_pages();
8092 /* It isn't necessary when PAGE_SIZE >= 1MB */
8093 if (PAGE_SHIFT < 20)
8094 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8096 #if __BITS_PER_LONG > 32
8098 unsigned long adapt;
8100 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8101 adapt <<= ADAPT_SCALE_SHIFT)
8106 /* limit to 1 bucket per 2^scale bytes of low memory */
8107 if (scale > PAGE_SHIFT)
8108 numentries >>= (scale - PAGE_SHIFT);
8110 numentries <<= (PAGE_SHIFT - scale);
8112 /* Make sure we've got at least a 0-order allocation.. */
8113 if (unlikely(flags & HASH_SMALL)) {
8114 /* Makes no sense without HASH_EARLY */
8115 WARN_ON(!(flags & HASH_EARLY));
8116 if (!(numentries >> *_hash_shift)) {
8117 numentries = 1UL << *_hash_shift;
8118 BUG_ON(!numentries);
8120 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8121 numentries = PAGE_SIZE / bucketsize;
8123 numentries = roundup_pow_of_two(numentries);
8125 /* limit allocation size to 1/16 total memory by default */
8127 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8128 do_div(max, bucketsize);
8130 max = min(max, 0x80000000ULL);
8132 if (numentries < low_limit)
8133 numentries = low_limit;
8134 if (numentries > max)
8137 log2qty = ilog2(numentries);
8139 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8142 size = bucketsize << log2qty;
8143 if (flags & HASH_EARLY) {
8144 if (flags & HASH_ZERO)
8145 table = memblock_alloc(size, SMP_CACHE_BYTES);
8147 table = memblock_alloc_raw(size,
8149 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8150 table = __vmalloc(size, gfp_flags);
8154 * If bucketsize is not a power-of-two, we may free
8155 * some pages at the end of hash table which
8156 * alloc_pages_exact() automatically does
8158 table = alloc_pages_exact(size, gfp_flags);
8159 kmemleak_alloc(table, size, 1, gfp_flags);
8161 } while (!table && size > PAGE_SIZE && --log2qty);
8164 panic("Failed to allocate %s hash table\n", tablename);
8166 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8167 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8168 virt ? "vmalloc" : "linear");
8171 *_hash_shift = log2qty;
8173 *_hash_mask = (1 << log2qty) - 1;
8179 * This function checks whether pageblock includes unmovable pages or not.
8181 * PageLRU check without isolation or lru_lock could race so that
8182 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8183 * check without lock_page also may miss some movable non-lru pages at
8184 * race condition. So you can't expect this function should be exact.
8186 * Returns a page without holding a reference. If the caller wants to
8187 * dereference that page (e.g., dumping), it has to make sure that that it
8188 * cannot get removed (e.g., via memory unplug) concurrently.
8191 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8192 int migratetype, int flags)
8194 unsigned long iter = 0;
8195 unsigned long pfn = page_to_pfn(page);
8198 * TODO we could make this much more efficient by not checking every
8199 * page in the range if we know all of them are in MOVABLE_ZONE and
8200 * that the movable zone guarantees that pages are migratable but
8201 * the later is not the case right now unfortunatelly. E.g. movablecore
8202 * can still lead to having bootmem allocations in zone_movable.
8205 if (is_migrate_cma_page(page)) {
8207 * CMA allocations (alloc_contig_range) really need to mark
8208 * isolate CMA pageblocks even when they are not movable in fact
8209 * so consider them movable here.
8211 if (is_migrate_cma(migratetype))
8217 for (; iter < pageblock_nr_pages; iter++) {
8218 if (!pfn_valid_within(pfn + iter))
8221 page = pfn_to_page(pfn + iter);
8223 if (PageReserved(page))
8227 * If the zone is movable and we have ruled out all reserved
8228 * pages then it should be reasonably safe to assume the rest
8231 if (zone_idx(zone) == ZONE_MOVABLE)
8235 * Hugepages are not in LRU lists, but they're movable.
8236 * THPs are on the LRU, but need to be counted as #small pages.
8237 * We need not scan over tail pages because we don't
8238 * handle each tail page individually in migration.
8240 if (PageHuge(page) || PageTransCompound(page)) {
8241 struct page *head = compound_head(page);
8242 unsigned int skip_pages;
8244 if (PageHuge(page)) {
8245 if (!hugepage_migration_supported(page_hstate(head)))
8247 } else if (!PageLRU(head) && !__PageMovable(head)) {
8251 skip_pages = compound_nr(head) - (page - head);
8252 iter += skip_pages - 1;
8257 * We can't use page_count without pin a page
8258 * because another CPU can free compound page.
8259 * This check already skips compound tails of THP
8260 * because their page->_refcount is zero at all time.
8262 if (!page_ref_count(page)) {
8263 if (PageBuddy(page))
8264 iter += (1 << page_order(page)) - 1;
8269 * The HWPoisoned page may be not in buddy system, and
8270 * page_count() is not 0.
8272 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8275 if (__PageMovable(page) || PageLRU(page))
8279 * If there are RECLAIMABLE pages, we need to check
8280 * it. But now, memory offline itself doesn't call
8281 * shrink_node_slabs() and it still to be fixed.
8284 * If the page is not RAM, page_count()should be 0.
8285 * we don't need more check. This is an _used_ not-movable page.
8287 * The problematic thing here is PG_reserved pages. PG_reserved
8288 * is set to both of a memory hole page and a _used_ kernel
8296 #ifdef CONFIG_CONTIG_ALLOC
8297 static unsigned long pfn_max_align_down(unsigned long pfn)
8299 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8300 pageblock_nr_pages) - 1);
8303 static unsigned long pfn_max_align_up(unsigned long pfn)
8305 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8306 pageblock_nr_pages));
8309 /* [start, end) must belong to a single zone. */
8310 static int __alloc_contig_migrate_range(struct compact_control *cc,
8311 unsigned long start, unsigned long end)
8313 /* This function is based on compact_zone() from compaction.c. */
8314 unsigned long nr_reclaimed;
8315 unsigned long pfn = start;
8316 unsigned int tries = 0;
8321 while (pfn < end || !list_empty(&cc->migratepages)) {
8322 if (fatal_signal_pending(current)) {
8327 if (list_empty(&cc->migratepages)) {
8328 cc->nr_migratepages = 0;
8329 pfn = isolate_migratepages_range(cc, pfn, end);
8335 } else if (++tries == 5) {
8336 ret = ret < 0 ? ret : -EBUSY;
8340 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8342 cc->nr_migratepages -= nr_reclaimed;
8344 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
8345 NULL, 0, cc->mode, MR_CONTIG_RANGE);
8348 putback_movable_pages(&cc->migratepages);
8355 * alloc_contig_range() -- tries to allocate given range of pages
8356 * @start: start PFN to allocate
8357 * @end: one-past-the-last PFN to allocate
8358 * @migratetype: migratetype of the underlaying pageblocks (either
8359 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8360 * in range must have the same migratetype and it must
8361 * be either of the two.
8362 * @gfp_mask: GFP mask to use during compaction
8364 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8365 * aligned. The PFN range must belong to a single zone.
8367 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8368 * pageblocks in the range. Once isolated, the pageblocks should not
8369 * be modified by others.
8371 * Return: zero on success or negative error code. On success all
8372 * pages which PFN is in [start, end) are allocated for the caller and
8373 * need to be freed with free_contig_range().
8375 int alloc_contig_range(unsigned long start, unsigned long end,
8376 unsigned migratetype, gfp_t gfp_mask)
8378 unsigned long outer_start, outer_end;
8382 struct compact_control cc = {
8383 .nr_migratepages = 0,
8385 .zone = page_zone(pfn_to_page(start)),
8386 .mode = MIGRATE_SYNC,
8387 .ignore_skip_hint = true,
8388 .no_set_skip_hint = true,
8389 .gfp_mask = current_gfp_context(gfp_mask),
8390 .alloc_contig = true,
8392 INIT_LIST_HEAD(&cc.migratepages);
8395 * What we do here is we mark all pageblocks in range as
8396 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8397 * have different sizes, and due to the way page allocator
8398 * work, we align the range to biggest of the two pages so
8399 * that page allocator won't try to merge buddies from
8400 * different pageblocks and change MIGRATE_ISOLATE to some
8401 * other migration type.
8403 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8404 * migrate the pages from an unaligned range (ie. pages that
8405 * we are interested in). This will put all the pages in
8406 * range back to page allocator as MIGRATE_ISOLATE.
8408 * When this is done, we take the pages in range from page
8409 * allocator removing them from the buddy system. This way
8410 * page allocator will never consider using them.
8412 * This lets us mark the pageblocks back as
8413 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8414 * aligned range but not in the unaligned, original range are
8415 * put back to page allocator so that buddy can use them.
8418 ret = start_isolate_page_range(pfn_max_align_down(start),
8419 pfn_max_align_up(end), migratetype, 0);
8424 * In case of -EBUSY, we'd like to know which page causes problem.
8425 * So, just fall through. test_pages_isolated() has a tracepoint
8426 * which will report the busy page.
8428 * It is possible that busy pages could become available before
8429 * the call to test_pages_isolated, and the range will actually be
8430 * allocated. So, if we fall through be sure to clear ret so that
8431 * -EBUSY is not accidentally used or returned to caller.
8433 ret = __alloc_contig_migrate_range(&cc, start, end);
8434 if (ret && ret != -EBUSY)
8439 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8440 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8441 * more, all pages in [start, end) are free in page allocator.
8442 * What we are going to do is to allocate all pages from
8443 * [start, end) (that is remove them from page allocator).
8445 * The only problem is that pages at the beginning and at the
8446 * end of interesting range may be not aligned with pages that
8447 * page allocator holds, ie. they can be part of higher order
8448 * pages. Because of this, we reserve the bigger range and
8449 * once this is done free the pages we are not interested in.
8451 * We don't have to hold zone->lock here because the pages are
8452 * isolated thus they won't get removed from buddy.
8455 lru_add_drain_all();
8458 outer_start = start;
8459 while (!PageBuddy(pfn_to_page(outer_start))) {
8460 if (++order >= MAX_ORDER) {
8461 outer_start = start;
8464 outer_start &= ~0UL << order;
8467 if (outer_start != start) {
8468 order = page_order(pfn_to_page(outer_start));
8471 * outer_start page could be small order buddy page and
8472 * it doesn't include start page. Adjust outer_start
8473 * in this case to report failed page properly
8474 * on tracepoint in test_pages_isolated()
8476 if (outer_start + (1UL << order) <= start)
8477 outer_start = start;
8480 /* Make sure the range is really isolated. */
8481 if (test_pages_isolated(outer_start, end, 0)) {
8482 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8483 __func__, outer_start, end);
8488 /* Grab isolated pages from freelists. */
8489 outer_end = isolate_freepages_range(&cc, outer_start, end);
8495 /* Free head and tail (if any) */
8496 if (start != outer_start)
8497 free_contig_range(outer_start, start - outer_start);
8498 if (end != outer_end)
8499 free_contig_range(end, outer_end - end);
8502 undo_isolate_page_range(pfn_max_align_down(start),
8503 pfn_max_align_up(end), migratetype);
8507 static int __alloc_contig_pages(unsigned long start_pfn,
8508 unsigned long nr_pages, gfp_t gfp_mask)
8510 unsigned long end_pfn = start_pfn + nr_pages;
8512 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8516 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8517 unsigned long nr_pages)
8519 unsigned long i, end_pfn = start_pfn + nr_pages;
8522 for (i = start_pfn; i < end_pfn; i++) {
8523 page = pfn_to_online_page(i);
8527 if (page_zone(page) != z)
8530 if (PageReserved(page))
8533 if (page_count(page) > 0)
8542 static bool zone_spans_last_pfn(const struct zone *zone,
8543 unsigned long start_pfn, unsigned long nr_pages)
8545 unsigned long last_pfn = start_pfn + nr_pages - 1;
8547 return zone_spans_pfn(zone, last_pfn);
8551 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8552 * @nr_pages: Number of contiguous pages to allocate
8553 * @gfp_mask: GFP mask to limit search and used during compaction
8555 * @nodemask: Mask for other possible nodes
8557 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8558 * on an applicable zonelist to find a contiguous pfn range which can then be
8559 * tried for allocation with alloc_contig_range(). This routine is intended
8560 * for allocation requests which can not be fulfilled with the buddy allocator.
8562 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8563 * power of two then the alignment is guaranteed to be to the given nr_pages
8564 * (e.g. 1GB request would be aligned to 1GB).
8566 * Allocated pages can be freed with free_contig_range() or by manually calling
8567 * __free_page() on each allocated page.
8569 * Return: pointer to contiguous pages on success, or NULL if not successful.
8571 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8572 int nid, nodemask_t *nodemask)
8574 unsigned long ret, pfn, flags;
8575 struct zonelist *zonelist;
8579 zonelist = node_zonelist(nid, gfp_mask);
8580 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8581 gfp_zone(gfp_mask), nodemask) {
8582 spin_lock_irqsave(&zone->lock, flags);
8584 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8585 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8586 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8588 * We release the zone lock here because
8589 * alloc_contig_range() will also lock the zone
8590 * at some point. If there's an allocation
8591 * spinning on this lock, it may win the race
8592 * and cause alloc_contig_range() to fail...
8594 spin_unlock_irqrestore(&zone->lock, flags);
8595 ret = __alloc_contig_pages(pfn, nr_pages,
8598 return pfn_to_page(pfn);
8599 spin_lock_irqsave(&zone->lock, flags);
8603 spin_unlock_irqrestore(&zone->lock, flags);
8607 #endif /* CONFIG_CONTIG_ALLOC */
8609 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8611 unsigned int count = 0;
8613 for (; nr_pages--; pfn++) {
8614 struct page *page = pfn_to_page(pfn);
8616 count += page_count(page) != 1;
8619 WARN(count != 0, "%d pages are still in use!\n", count);
8623 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8624 * page high values need to be recalulated.
8626 void __meminit zone_pcp_update(struct zone *zone)
8628 mutex_lock(&pcp_batch_high_lock);
8629 __zone_pcp_update(zone);
8630 mutex_unlock(&pcp_batch_high_lock);
8633 void zone_pcp_reset(struct zone *zone)
8635 unsigned long flags;
8637 struct per_cpu_pageset *pset;
8639 /* avoid races with drain_pages() */
8640 local_irq_save(flags);
8641 if (zone->pageset != &boot_pageset) {
8642 for_each_online_cpu(cpu) {
8643 pset = per_cpu_ptr(zone->pageset, cpu);
8644 drain_zonestat(zone, pset);
8646 free_percpu(zone->pageset);
8647 zone->pageset = &boot_pageset;
8649 local_irq_restore(flags);
8652 #ifdef CONFIG_MEMORY_HOTREMOVE
8654 * All pages in the range must be in a single zone and isolated
8655 * before calling this.
8658 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8664 unsigned long flags;
8665 unsigned long offlined_pages = 0;
8667 /* find the first valid pfn */
8668 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8672 return offlined_pages;
8674 offline_mem_sections(pfn, end_pfn);
8675 zone = page_zone(pfn_to_page(pfn));
8676 spin_lock_irqsave(&zone->lock, flags);
8678 while (pfn < end_pfn) {
8679 if (!pfn_valid(pfn)) {
8683 page = pfn_to_page(pfn);
8685 * The HWPoisoned page may be not in buddy system, and
8686 * page_count() is not 0.
8688 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8694 BUG_ON(page_count(page));
8695 BUG_ON(!PageBuddy(page));
8696 order = page_order(page);
8697 offlined_pages += 1 << order;
8698 del_page_from_free_list(page, zone, order);
8699 pfn += (1 << order);
8701 spin_unlock_irqrestore(&zone->lock, flags);
8703 return offlined_pages;
8707 bool is_free_buddy_page(struct page *page)
8709 struct zone *zone = page_zone(page);
8710 unsigned long pfn = page_to_pfn(page);
8711 unsigned long flags;
8714 spin_lock_irqsave(&zone->lock, flags);
8715 for (order = 0; order < MAX_ORDER; order++) {
8716 struct page *page_head = page - (pfn & ((1 << order) - 1));
8718 if (PageBuddy(page_head) && page_order(page_head) >= order)
8721 spin_unlock_irqrestore(&zone->lock, flags);
8723 return order < MAX_ORDER;
8726 #ifdef CONFIG_MEMORY_FAILURE
8728 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8729 * test is performed under the zone lock to prevent a race against page
8732 bool set_hwpoison_free_buddy_page(struct page *page)
8734 struct zone *zone = page_zone(page);
8735 unsigned long pfn = page_to_pfn(page);
8736 unsigned long flags;
8738 bool hwpoisoned = false;
8740 spin_lock_irqsave(&zone->lock, flags);
8741 for (order = 0; order < MAX_ORDER; order++) {
8742 struct page *page_head = page - (pfn & ((1 << order) - 1));
8744 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8745 if (!TestSetPageHWPoison(page))
8750 spin_unlock_irqrestore(&zone->lock, flags);