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 const char *page_bad_reason(struct page *page, unsigned long flags)
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 & flags)) {
1084 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1085 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1087 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1090 if (unlikely(page->mem_cgroup))
1091 bad_reason = "page still charged to cgroup";
1096 static void check_free_page_bad(struct page *page)
1099 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1102 static inline int check_free_page(struct page *page)
1104 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1107 /* Something has gone sideways, find it */
1108 check_free_page_bad(page);
1112 static int free_tail_pages_check(struct page *head_page, struct page *page)
1117 * We rely page->lru.next never has bit 0 set, unless the page
1118 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1120 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1122 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1126 switch (page - head_page) {
1128 /* the first tail page: ->mapping may be compound_mapcount() */
1129 if (unlikely(compound_mapcount(page))) {
1130 bad_page(page, "nonzero compound_mapcount");
1136 * the second tail page: ->mapping is
1137 * deferred_list.next -- ignore value.
1141 if (page->mapping != TAIL_MAPPING) {
1142 bad_page(page, "corrupted mapping in tail page");
1147 if (unlikely(!PageTail(page))) {
1148 bad_page(page, "PageTail not set");
1151 if (unlikely(compound_head(page) != head_page)) {
1152 bad_page(page, "compound_head not consistent");
1157 page->mapping = NULL;
1158 clear_compound_head(page);
1162 static void kernel_init_free_pages(struct page *page, int numpages)
1166 for (i = 0; i < numpages; i++)
1167 clear_highpage(page + i);
1170 static __always_inline bool free_pages_prepare(struct page *page,
1171 unsigned int order, bool check_free)
1175 VM_BUG_ON_PAGE(PageTail(page), page);
1177 trace_mm_page_free(page, order);
1180 * Check tail pages before head page information is cleared to
1181 * avoid checking PageCompound for order-0 pages.
1183 if (unlikely(order)) {
1184 bool compound = PageCompound(page);
1187 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1190 ClearPageDoubleMap(page);
1191 for (i = 1; i < (1 << order); i++) {
1193 bad += free_tail_pages_check(page, page + i);
1194 if (unlikely(check_free_page(page + i))) {
1198 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1201 if (PageMappingFlags(page))
1202 page->mapping = NULL;
1203 if (memcg_kmem_enabled() && PageKmemcg(page))
1204 __memcg_kmem_uncharge_page(page, order);
1206 bad += check_free_page(page);
1210 page_cpupid_reset_last(page);
1211 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1212 reset_page_owner(page, order);
1214 if (!PageHighMem(page)) {
1215 debug_check_no_locks_freed(page_address(page),
1216 PAGE_SIZE << order);
1217 debug_check_no_obj_freed(page_address(page),
1218 PAGE_SIZE << order);
1220 if (want_init_on_free())
1221 kernel_init_free_pages(page, 1 << order);
1223 kernel_poison_pages(page, 1 << order, 0);
1225 * arch_free_page() can make the page's contents inaccessible. s390
1226 * does this. So nothing which can access the page's contents should
1227 * happen after this.
1229 arch_free_page(page, order);
1231 if (debug_pagealloc_enabled_static())
1232 kernel_map_pages(page, 1 << order, 0);
1234 kasan_free_nondeferred_pages(page, order);
1239 #ifdef CONFIG_DEBUG_VM
1241 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1242 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1243 * moved from pcp lists to free lists.
1245 static bool free_pcp_prepare(struct page *page)
1247 return free_pages_prepare(page, 0, true);
1250 static bool bulkfree_pcp_prepare(struct page *page)
1252 if (debug_pagealloc_enabled_static())
1253 return check_free_page(page);
1259 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1260 * moving from pcp lists to free list in order to reduce overhead. With
1261 * debug_pagealloc enabled, they are checked also immediately when being freed
1264 static bool free_pcp_prepare(struct page *page)
1266 if (debug_pagealloc_enabled_static())
1267 return free_pages_prepare(page, 0, true);
1269 return free_pages_prepare(page, 0, false);
1272 static bool bulkfree_pcp_prepare(struct page *page)
1274 return check_free_page(page);
1276 #endif /* CONFIG_DEBUG_VM */
1278 static inline void prefetch_buddy(struct page *page)
1280 unsigned long pfn = page_to_pfn(page);
1281 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1282 struct page *buddy = page + (buddy_pfn - pfn);
1288 * Frees a number of pages from the PCP lists
1289 * Assumes all pages on list are in same zone, and of same order.
1290 * count is the number of pages to free.
1292 * If the zone was previously in an "all pages pinned" state then look to
1293 * see if this freeing clears that state.
1295 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1296 * pinned" detection logic.
1298 static void free_pcppages_bulk(struct zone *zone, int count,
1299 struct per_cpu_pages *pcp)
1301 int migratetype = 0;
1303 int prefetch_nr = 0;
1304 bool isolated_pageblocks;
1305 struct page *page, *tmp;
1309 struct list_head *list;
1312 * Remove pages from lists in a round-robin fashion. A
1313 * batch_free count is maintained that is incremented when an
1314 * empty list is encountered. This is so more pages are freed
1315 * off fuller lists instead of spinning excessively around empty
1320 if (++migratetype == MIGRATE_PCPTYPES)
1322 list = &pcp->lists[migratetype];
1323 } while (list_empty(list));
1325 /* This is the only non-empty list. Free them all. */
1326 if (batch_free == MIGRATE_PCPTYPES)
1330 page = list_last_entry(list, struct page, lru);
1331 /* must delete to avoid corrupting pcp list */
1332 list_del(&page->lru);
1335 if (bulkfree_pcp_prepare(page))
1338 list_add_tail(&page->lru, &head);
1341 * We are going to put the page back to the global
1342 * pool, prefetch its buddy to speed up later access
1343 * under zone->lock. It is believed the overhead of
1344 * an additional test and calculating buddy_pfn here
1345 * can be offset by reduced memory latency later. To
1346 * avoid excessive prefetching due to large count, only
1347 * prefetch buddy for the first pcp->batch nr of pages.
1349 if (prefetch_nr++ < pcp->batch)
1350 prefetch_buddy(page);
1351 } while (--count && --batch_free && !list_empty(list));
1354 spin_lock(&zone->lock);
1355 isolated_pageblocks = has_isolate_pageblock(zone);
1358 * Use safe version since after __free_one_page(),
1359 * page->lru.next will not point to original list.
1361 list_for_each_entry_safe(page, tmp, &head, lru) {
1362 int mt = get_pcppage_migratetype(page);
1363 /* MIGRATE_ISOLATE page should not go to pcplists */
1364 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1365 /* Pageblock could have been isolated meanwhile */
1366 if (unlikely(isolated_pageblocks))
1367 mt = get_pageblock_migratetype(page);
1369 __free_one_page(page, page_to_pfn(page), zone, 0, mt, true);
1370 trace_mm_page_pcpu_drain(page, 0, mt);
1372 spin_unlock(&zone->lock);
1375 static void free_one_page(struct zone *zone,
1376 struct page *page, unsigned long pfn,
1380 spin_lock(&zone->lock);
1381 if (unlikely(has_isolate_pageblock(zone) ||
1382 is_migrate_isolate(migratetype))) {
1383 migratetype = get_pfnblock_migratetype(page, pfn);
1385 __free_one_page(page, pfn, zone, order, migratetype, true);
1386 spin_unlock(&zone->lock);
1389 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1390 unsigned long zone, int nid)
1392 mm_zero_struct_page(page);
1393 set_page_links(page, zone, nid, pfn);
1394 init_page_count(page);
1395 page_mapcount_reset(page);
1396 page_cpupid_reset_last(page);
1397 page_kasan_tag_reset(page);
1399 INIT_LIST_HEAD(&page->lru);
1400 #ifdef WANT_PAGE_VIRTUAL
1401 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1402 if (!is_highmem_idx(zone))
1403 set_page_address(page, __va(pfn << PAGE_SHIFT));
1407 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1408 static void __meminit init_reserved_page(unsigned long pfn)
1413 if (!early_page_uninitialised(pfn))
1416 nid = early_pfn_to_nid(pfn);
1417 pgdat = NODE_DATA(nid);
1419 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1420 struct zone *zone = &pgdat->node_zones[zid];
1422 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1425 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1428 static inline void init_reserved_page(unsigned long pfn)
1431 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1434 * Initialised pages do not have PageReserved set. This function is
1435 * called for each range allocated by the bootmem allocator and
1436 * marks the pages PageReserved. The remaining valid pages are later
1437 * sent to the buddy page allocator.
1439 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1441 unsigned long start_pfn = PFN_DOWN(start);
1442 unsigned long end_pfn = PFN_UP(end);
1444 for (; start_pfn < end_pfn; start_pfn++) {
1445 if (pfn_valid(start_pfn)) {
1446 struct page *page = pfn_to_page(start_pfn);
1448 init_reserved_page(start_pfn);
1450 /* Avoid false-positive PageTail() */
1451 INIT_LIST_HEAD(&page->lru);
1454 * no need for atomic set_bit because the struct
1455 * page is not visible yet so nobody should
1458 __SetPageReserved(page);
1463 static void __free_pages_ok(struct page *page, unsigned int order)
1465 unsigned long flags;
1467 unsigned long pfn = page_to_pfn(page);
1469 if (!free_pages_prepare(page, order, true))
1472 migratetype = get_pfnblock_migratetype(page, pfn);
1473 local_irq_save(flags);
1474 __count_vm_events(PGFREE, 1 << order);
1475 free_one_page(page_zone(page), page, pfn, order, migratetype);
1476 local_irq_restore(flags);
1479 void __free_pages_core(struct page *page, unsigned int order)
1481 unsigned int nr_pages = 1 << order;
1482 struct page *p = page;
1486 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1488 __ClearPageReserved(p);
1489 set_page_count(p, 0);
1491 __ClearPageReserved(p);
1492 set_page_count(p, 0);
1494 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1495 set_page_refcounted(page);
1496 __free_pages(page, order);
1499 #ifdef CONFIG_NEED_MULTIPLE_NODES
1501 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1503 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
1506 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1508 int __meminit __early_pfn_to_nid(unsigned long pfn,
1509 struct mminit_pfnnid_cache *state)
1511 unsigned long start_pfn, end_pfn;
1514 if (state->last_start <= pfn && pfn < state->last_end)
1515 return state->last_nid;
1517 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1518 if (nid != NUMA_NO_NODE) {
1519 state->last_start = start_pfn;
1520 state->last_end = end_pfn;
1521 state->last_nid = nid;
1526 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
1528 int __meminit early_pfn_to_nid(unsigned long pfn)
1530 static DEFINE_SPINLOCK(early_pfn_lock);
1533 spin_lock(&early_pfn_lock);
1534 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1536 nid = first_online_node;
1537 spin_unlock(&early_pfn_lock);
1541 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1543 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1546 if (early_page_uninitialised(pfn))
1548 __free_pages_core(page, order);
1552 * Check that the whole (or subset of) a pageblock given by the interval of
1553 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1554 * with the migration of free compaction scanner. The scanners then need to
1555 * use only pfn_valid_within() check for arches that allow holes within
1558 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1560 * It's possible on some configurations to have a setup like node0 node1 node0
1561 * i.e. it's possible that all pages within a zones range of pages do not
1562 * belong to a single zone. We assume that a border between node0 and node1
1563 * can occur within a single pageblock, but not a node0 node1 node0
1564 * interleaving within a single pageblock. It is therefore sufficient to check
1565 * the first and last page of a pageblock and avoid checking each individual
1566 * page in a pageblock.
1568 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1569 unsigned long end_pfn, struct zone *zone)
1571 struct page *start_page;
1572 struct page *end_page;
1574 /* end_pfn is one past the range we are checking */
1577 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1580 start_page = pfn_to_online_page(start_pfn);
1584 if (page_zone(start_page) != zone)
1587 end_page = pfn_to_page(end_pfn);
1589 /* This gives a shorter code than deriving page_zone(end_page) */
1590 if (page_zone_id(start_page) != page_zone_id(end_page))
1596 void set_zone_contiguous(struct zone *zone)
1598 unsigned long block_start_pfn = zone->zone_start_pfn;
1599 unsigned long block_end_pfn;
1601 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1602 for (; block_start_pfn < zone_end_pfn(zone);
1603 block_start_pfn = block_end_pfn,
1604 block_end_pfn += pageblock_nr_pages) {
1606 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1608 if (!__pageblock_pfn_to_page(block_start_pfn,
1609 block_end_pfn, zone))
1614 /* We confirm that there is no hole */
1615 zone->contiguous = true;
1618 void clear_zone_contiguous(struct zone *zone)
1620 zone->contiguous = false;
1623 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1624 static void __init deferred_free_range(unsigned long pfn,
1625 unsigned long nr_pages)
1633 page = pfn_to_page(pfn);
1635 /* Free a large naturally-aligned chunk if possible */
1636 if (nr_pages == pageblock_nr_pages &&
1637 (pfn & (pageblock_nr_pages - 1)) == 0) {
1638 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1639 __free_pages_core(page, pageblock_order);
1643 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1644 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1645 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1646 __free_pages_core(page, 0);
1650 /* Completion tracking for deferred_init_memmap() threads */
1651 static atomic_t pgdat_init_n_undone __initdata;
1652 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1654 static inline void __init pgdat_init_report_one_done(void)
1656 if (atomic_dec_and_test(&pgdat_init_n_undone))
1657 complete(&pgdat_init_all_done_comp);
1661 * Returns true if page needs to be initialized or freed to buddy allocator.
1663 * First we check if pfn is valid on architectures where it is possible to have
1664 * holes within pageblock_nr_pages. On systems where it is not possible, this
1665 * function is optimized out.
1667 * Then, we check if a current large page is valid by only checking the validity
1670 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1672 if (!pfn_valid_within(pfn))
1674 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1680 * Free pages to buddy allocator. Try to free aligned pages in
1681 * pageblock_nr_pages sizes.
1683 static void __init deferred_free_pages(unsigned long pfn,
1684 unsigned long end_pfn)
1686 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1687 unsigned long nr_free = 0;
1689 for (; pfn < end_pfn; pfn++) {
1690 if (!deferred_pfn_valid(pfn)) {
1691 deferred_free_range(pfn - nr_free, nr_free);
1693 } else if (!(pfn & nr_pgmask)) {
1694 deferred_free_range(pfn - nr_free, nr_free);
1696 touch_nmi_watchdog();
1701 /* Free the last block of pages to allocator */
1702 deferred_free_range(pfn - nr_free, nr_free);
1706 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1707 * by performing it only once every pageblock_nr_pages.
1708 * Return number of pages initialized.
1710 static unsigned long __init deferred_init_pages(struct zone *zone,
1712 unsigned long end_pfn)
1714 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1715 int nid = zone_to_nid(zone);
1716 unsigned long nr_pages = 0;
1717 int zid = zone_idx(zone);
1718 struct page *page = NULL;
1720 for (; pfn < end_pfn; pfn++) {
1721 if (!deferred_pfn_valid(pfn)) {
1724 } else if (!page || !(pfn & nr_pgmask)) {
1725 page = pfn_to_page(pfn);
1726 touch_nmi_watchdog();
1730 __init_single_page(page, pfn, zid, nid);
1737 * This function is meant to pre-load the iterator for the zone init.
1738 * Specifically it walks through the ranges until we are caught up to the
1739 * first_init_pfn value and exits there. If we never encounter the value we
1740 * return false indicating there are no valid ranges left.
1743 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1744 unsigned long *spfn, unsigned long *epfn,
1745 unsigned long first_init_pfn)
1750 * Start out by walking through the ranges in this zone that have
1751 * already been initialized. We don't need to do anything with them
1752 * so we just need to flush them out of the system.
1754 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1755 if (*epfn <= first_init_pfn)
1757 if (*spfn < first_init_pfn)
1758 *spfn = first_init_pfn;
1767 * Initialize and free pages. We do it in two loops: first we initialize
1768 * struct page, then free to buddy allocator, because while we are
1769 * freeing pages we can access pages that are ahead (computing buddy
1770 * page in __free_one_page()).
1772 * In order to try and keep some memory in the cache we have the loop
1773 * broken along max page order boundaries. This way we will not cause
1774 * any issues with the buddy page computation.
1776 static unsigned long __init
1777 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1778 unsigned long *end_pfn)
1780 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1781 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1782 unsigned long nr_pages = 0;
1785 /* First we loop through and initialize the page values */
1786 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1789 if (mo_pfn <= *start_pfn)
1792 t = min(mo_pfn, *end_pfn);
1793 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1795 if (mo_pfn < *end_pfn) {
1796 *start_pfn = mo_pfn;
1801 /* Reset values and now loop through freeing pages as needed */
1804 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1810 t = min(mo_pfn, epfn);
1811 deferred_free_pages(spfn, t);
1820 /* Initialise remaining memory on a node */
1821 static int __init deferred_init_memmap(void *data)
1823 pg_data_t *pgdat = data;
1824 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1825 unsigned long spfn = 0, epfn = 0, nr_pages = 0;
1826 unsigned long first_init_pfn, flags;
1827 unsigned long start = jiffies;
1832 /* Bind memory initialisation thread to a local node if possible */
1833 if (!cpumask_empty(cpumask))
1834 set_cpus_allowed_ptr(current, cpumask);
1836 pgdat_resize_lock(pgdat, &flags);
1837 first_init_pfn = pgdat->first_deferred_pfn;
1838 if (first_init_pfn == ULONG_MAX) {
1839 pgdat_resize_unlock(pgdat, &flags);
1840 pgdat_init_report_one_done();
1844 /* Sanity check boundaries */
1845 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1846 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1847 pgdat->first_deferred_pfn = ULONG_MAX;
1849 /* Only the highest zone is deferred so find it */
1850 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1851 zone = pgdat->node_zones + zid;
1852 if (first_init_pfn < zone_end_pfn(zone))
1856 /* If the zone is empty somebody else may have cleared out the zone */
1857 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1862 * Initialize and free pages in MAX_ORDER sized increments so
1863 * that we can avoid introducing any issues with the buddy
1867 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1869 pgdat_resize_unlock(pgdat, &flags);
1871 /* Sanity check that the next zone really is unpopulated */
1872 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1874 pr_info("node %d initialised, %lu pages in %ums\n",
1875 pgdat->node_id, nr_pages, jiffies_to_msecs(jiffies - start));
1877 pgdat_init_report_one_done();
1882 * If this zone has deferred pages, try to grow it by initializing enough
1883 * deferred pages to satisfy the allocation specified by order, rounded up to
1884 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1885 * of SECTION_SIZE bytes by initializing struct pages in increments of
1886 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1888 * Return true when zone was grown, otherwise return false. We return true even
1889 * when we grow less than requested, to let the caller decide if there are
1890 * enough pages to satisfy the allocation.
1892 * Note: We use noinline because this function is needed only during boot, and
1893 * it is called from a __ref function _deferred_grow_zone. This way we are
1894 * making sure that it is not inlined into permanent text section.
1896 static noinline bool __init
1897 deferred_grow_zone(struct zone *zone, unsigned int order)
1899 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1900 pg_data_t *pgdat = zone->zone_pgdat;
1901 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1902 unsigned long spfn, epfn, flags;
1903 unsigned long nr_pages = 0;
1906 /* Only the last zone may have deferred pages */
1907 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1910 pgdat_resize_lock(pgdat, &flags);
1913 * If deferred pages have been initialized while we were waiting for
1914 * the lock, return true, as the zone was grown. The caller will retry
1915 * this zone. We won't return to this function since the caller also
1916 * has this static branch.
1918 if (!static_branch_unlikely(&deferred_pages)) {
1919 pgdat_resize_unlock(pgdat, &flags);
1924 * If someone grew this zone while we were waiting for spinlock, return
1925 * true, as there might be enough pages already.
1927 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1928 pgdat_resize_unlock(pgdat, &flags);
1932 /* If the zone is empty somebody else may have cleared out the zone */
1933 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1934 first_deferred_pfn)) {
1935 pgdat->first_deferred_pfn = ULONG_MAX;
1936 pgdat_resize_unlock(pgdat, &flags);
1937 /* Retry only once. */
1938 return first_deferred_pfn != ULONG_MAX;
1942 * Initialize and free pages in MAX_ORDER sized increments so
1943 * that we can avoid introducing any issues with the buddy
1946 while (spfn < epfn) {
1947 /* update our first deferred PFN for this section */
1948 first_deferred_pfn = spfn;
1950 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1952 /* We should only stop along section boundaries */
1953 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
1956 /* If our quota has been met we can stop here */
1957 if (nr_pages >= nr_pages_needed)
1961 pgdat->first_deferred_pfn = spfn;
1962 pgdat_resize_unlock(pgdat, &flags);
1964 return nr_pages > 0;
1968 * deferred_grow_zone() is __init, but it is called from
1969 * get_page_from_freelist() during early boot until deferred_pages permanently
1970 * disables this call. This is why we have refdata wrapper to avoid warning,
1971 * and to ensure that the function body gets unloaded.
1974 _deferred_grow_zone(struct zone *zone, unsigned int order)
1976 return deferred_grow_zone(zone, order);
1979 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1981 void __init page_alloc_init_late(void)
1986 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1988 /* There will be num_node_state(N_MEMORY) threads */
1989 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1990 for_each_node_state(nid, N_MEMORY) {
1991 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1994 /* Block until all are initialised */
1995 wait_for_completion(&pgdat_init_all_done_comp);
1998 * The number of managed pages has changed due to the initialisation
1999 * so the pcpu batch and high limits needs to be updated or the limits
2000 * will be artificially small.
2002 for_each_populated_zone(zone)
2003 zone_pcp_update(zone);
2006 * We initialized the rest of the deferred pages. Permanently disable
2007 * on-demand struct page initialization.
2009 static_branch_disable(&deferred_pages);
2011 /* Reinit limits that are based on free pages after the kernel is up */
2012 files_maxfiles_init();
2015 /* Discard memblock private memory */
2018 for_each_node_state(nid, N_MEMORY)
2019 shuffle_free_memory(NODE_DATA(nid));
2021 for_each_populated_zone(zone)
2022 set_zone_contiguous(zone);
2026 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2027 void __init init_cma_reserved_pageblock(struct page *page)
2029 unsigned i = pageblock_nr_pages;
2030 struct page *p = page;
2033 __ClearPageReserved(p);
2034 set_page_count(p, 0);
2037 set_pageblock_migratetype(page, MIGRATE_CMA);
2039 if (pageblock_order >= MAX_ORDER) {
2040 i = pageblock_nr_pages;
2043 set_page_refcounted(p);
2044 __free_pages(p, MAX_ORDER - 1);
2045 p += MAX_ORDER_NR_PAGES;
2046 } while (i -= MAX_ORDER_NR_PAGES);
2048 set_page_refcounted(page);
2049 __free_pages(page, pageblock_order);
2052 adjust_managed_page_count(page, pageblock_nr_pages);
2057 * The order of subdivision here is critical for the IO subsystem.
2058 * Please do not alter this order without good reasons and regression
2059 * testing. Specifically, as large blocks of memory are subdivided,
2060 * the order in which smaller blocks are delivered depends on the order
2061 * they're subdivided in this function. This is the primary factor
2062 * influencing the order in which pages are delivered to the IO
2063 * subsystem according to empirical testing, and this is also justified
2064 * by considering the behavior of a buddy system containing a single
2065 * large block of memory acted on by a series of small allocations.
2066 * This behavior is a critical factor in sglist merging's success.
2070 static inline void expand(struct zone *zone, struct page *page,
2071 int low, int high, int migratetype)
2073 unsigned long size = 1 << high;
2075 while (high > low) {
2078 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2081 * Mark as guard pages (or page), that will allow to
2082 * merge back to allocator when buddy will be freed.
2083 * Corresponding page table entries will not be touched,
2084 * pages will stay not present in virtual address space
2086 if (set_page_guard(zone, &page[size], high, migratetype))
2089 add_to_free_list(&page[size], zone, high, migratetype);
2090 set_page_order(&page[size], high);
2094 static void check_new_page_bad(struct page *page)
2096 if (unlikely(page->flags & __PG_HWPOISON)) {
2097 /* Don't complain about hwpoisoned pages */
2098 page_mapcount_reset(page); /* remove PageBuddy */
2103 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2107 * This page is about to be returned from the page allocator
2109 static inline int check_new_page(struct page *page)
2111 if (likely(page_expected_state(page,
2112 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2115 check_new_page_bad(page);
2119 static inline bool free_pages_prezeroed(void)
2121 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2122 page_poisoning_enabled()) || want_init_on_free();
2125 #ifdef CONFIG_DEBUG_VM
2127 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2128 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2129 * also checked when pcp lists are refilled from the free lists.
2131 static inline bool check_pcp_refill(struct page *page)
2133 if (debug_pagealloc_enabled_static())
2134 return check_new_page(page);
2139 static inline bool check_new_pcp(struct page *page)
2141 return check_new_page(page);
2145 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2146 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2147 * enabled, they are also checked when being allocated from the pcp lists.
2149 static inline bool check_pcp_refill(struct page *page)
2151 return check_new_page(page);
2153 static inline bool check_new_pcp(struct page *page)
2155 if (debug_pagealloc_enabled_static())
2156 return check_new_page(page);
2160 #endif /* CONFIG_DEBUG_VM */
2162 static bool check_new_pages(struct page *page, unsigned int order)
2165 for (i = 0; i < (1 << order); i++) {
2166 struct page *p = page + i;
2168 if (unlikely(check_new_page(p)))
2175 inline void post_alloc_hook(struct page *page, unsigned int order,
2178 set_page_private(page, 0);
2179 set_page_refcounted(page);
2181 arch_alloc_page(page, order);
2182 if (debug_pagealloc_enabled_static())
2183 kernel_map_pages(page, 1 << order, 1);
2184 kasan_alloc_pages(page, order);
2185 kernel_poison_pages(page, 1 << order, 1);
2186 set_page_owner(page, order, gfp_flags);
2189 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2190 unsigned int alloc_flags)
2192 post_alloc_hook(page, order, gfp_flags);
2194 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2195 kernel_init_free_pages(page, 1 << order);
2197 if (order && (gfp_flags & __GFP_COMP))
2198 prep_compound_page(page, order);
2201 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2202 * allocate the page. The expectation is that the caller is taking
2203 * steps that will free more memory. The caller should avoid the page
2204 * being used for !PFMEMALLOC purposes.
2206 if (alloc_flags & ALLOC_NO_WATERMARKS)
2207 set_page_pfmemalloc(page);
2209 clear_page_pfmemalloc(page);
2213 * Go through the free lists for the given migratetype and remove
2214 * the smallest available page from the freelists
2216 static __always_inline
2217 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2220 unsigned int current_order;
2221 struct free_area *area;
2224 /* Find a page of the appropriate size in the preferred list */
2225 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2226 area = &(zone->free_area[current_order]);
2227 page = get_page_from_free_area(area, migratetype);
2230 del_page_from_free_list(page, zone, current_order);
2231 expand(zone, page, order, current_order, migratetype);
2232 set_pcppage_migratetype(page, migratetype);
2241 * This array describes the order lists are fallen back to when
2242 * the free lists for the desirable migrate type are depleted
2244 static int fallbacks[MIGRATE_TYPES][4] = {
2245 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2246 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2247 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2249 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2251 #ifdef CONFIG_MEMORY_ISOLATION
2252 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2257 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2260 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2263 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2264 unsigned int order) { return NULL; }
2268 * Move the free pages in a range to the free lists of the requested type.
2269 * Note that start_page and end_pages are not aligned on a pageblock
2270 * boundary. If alignment is required, use move_freepages_block()
2272 static int move_freepages(struct zone *zone,
2273 struct page *start_page, struct page *end_page,
2274 int migratetype, int *num_movable)
2278 int pages_moved = 0;
2280 for (page = start_page; page <= end_page;) {
2281 if (!pfn_valid_within(page_to_pfn(page))) {
2286 if (!PageBuddy(page)) {
2288 * We assume that pages that could be isolated for
2289 * migration are movable. But we don't actually try
2290 * isolating, as that would be expensive.
2293 (PageLRU(page) || __PageMovable(page)))
2300 /* Make sure we are not inadvertently changing nodes */
2301 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2302 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2304 order = page_order(page);
2305 move_to_free_list(page, zone, order, migratetype);
2307 pages_moved += 1 << order;
2313 int move_freepages_block(struct zone *zone, struct page *page,
2314 int migratetype, int *num_movable)
2316 unsigned long start_pfn, end_pfn;
2317 struct page *start_page, *end_page;
2322 start_pfn = page_to_pfn(page);
2323 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2324 start_page = pfn_to_page(start_pfn);
2325 end_page = start_page + pageblock_nr_pages - 1;
2326 end_pfn = start_pfn + pageblock_nr_pages - 1;
2328 /* Do not cross zone boundaries */
2329 if (!zone_spans_pfn(zone, start_pfn))
2331 if (!zone_spans_pfn(zone, end_pfn))
2334 return move_freepages(zone, start_page, end_page, migratetype,
2338 static void change_pageblock_range(struct page *pageblock_page,
2339 int start_order, int migratetype)
2341 int nr_pageblocks = 1 << (start_order - pageblock_order);
2343 while (nr_pageblocks--) {
2344 set_pageblock_migratetype(pageblock_page, migratetype);
2345 pageblock_page += pageblock_nr_pages;
2350 * When we are falling back to another migratetype during allocation, try to
2351 * steal extra free pages from the same pageblocks to satisfy further
2352 * allocations, instead of polluting multiple pageblocks.
2354 * If we are stealing a relatively large buddy page, it is likely there will
2355 * be more free pages in the pageblock, so try to steal them all. For
2356 * reclaimable and unmovable allocations, we steal regardless of page size,
2357 * as fragmentation caused by those allocations polluting movable pageblocks
2358 * is worse than movable allocations stealing from unmovable and reclaimable
2361 static bool can_steal_fallback(unsigned int order, int start_mt)
2364 * Leaving this order check is intended, although there is
2365 * relaxed order check in next check. The reason is that
2366 * we can actually steal whole pageblock if this condition met,
2367 * but, below check doesn't guarantee it and that is just heuristic
2368 * so could be changed anytime.
2370 if (order >= pageblock_order)
2373 if (order >= pageblock_order / 2 ||
2374 start_mt == MIGRATE_RECLAIMABLE ||
2375 start_mt == MIGRATE_UNMOVABLE ||
2376 page_group_by_mobility_disabled)
2382 static inline void boost_watermark(struct zone *zone)
2384 unsigned long max_boost;
2386 if (!watermark_boost_factor)
2389 * Don't bother in zones that are unlikely to produce results.
2390 * On small machines, including kdump capture kernels running
2391 * in a small area, boosting the watermark can cause an out of
2392 * memory situation immediately.
2394 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2397 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2398 watermark_boost_factor, 10000);
2401 * high watermark may be uninitialised if fragmentation occurs
2402 * very early in boot so do not boost. We do not fall
2403 * through and boost by pageblock_nr_pages as failing
2404 * allocations that early means that reclaim is not going
2405 * to help and it may even be impossible to reclaim the
2406 * boosted watermark resulting in a hang.
2411 max_boost = max(pageblock_nr_pages, max_boost);
2413 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2418 * This function implements actual steal behaviour. If order is large enough,
2419 * we can steal whole pageblock. If not, we first move freepages in this
2420 * pageblock to our migratetype and determine how many already-allocated pages
2421 * are there in the pageblock with a compatible migratetype. If at least half
2422 * of pages are free or compatible, we can change migratetype of the pageblock
2423 * itself, so pages freed in the future will be put on the correct free list.
2425 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2426 unsigned int alloc_flags, int start_type, bool whole_block)
2428 unsigned int current_order = page_order(page);
2429 int free_pages, movable_pages, alike_pages;
2432 old_block_type = get_pageblock_migratetype(page);
2435 * This can happen due to races and we want to prevent broken
2436 * highatomic accounting.
2438 if (is_migrate_highatomic(old_block_type))
2441 /* Take ownership for orders >= pageblock_order */
2442 if (current_order >= pageblock_order) {
2443 change_pageblock_range(page, current_order, start_type);
2448 * Boost watermarks to increase reclaim pressure to reduce the
2449 * likelihood of future fallbacks. Wake kswapd now as the node
2450 * may be balanced overall and kswapd will not wake naturally.
2452 boost_watermark(zone);
2453 if (alloc_flags & ALLOC_KSWAPD)
2454 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2456 /* We are not allowed to try stealing from the whole block */
2460 free_pages = move_freepages_block(zone, page, start_type,
2463 * Determine how many pages are compatible with our allocation.
2464 * For movable allocation, it's the number of movable pages which
2465 * we just obtained. For other types it's a bit more tricky.
2467 if (start_type == MIGRATE_MOVABLE) {
2468 alike_pages = movable_pages;
2471 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2472 * to MOVABLE pageblock, consider all non-movable pages as
2473 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2474 * vice versa, be conservative since we can't distinguish the
2475 * exact migratetype of non-movable pages.
2477 if (old_block_type == MIGRATE_MOVABLE)
2478 alike_pages = pageblock_nr_pages
2479 - (free_pages + movable_pages);
2484 /* moving whole block can fail due to zone boundary conditions */
2489 * If a sufficient number of pages in the block are either free or of
2490 * comparable migratability as our allocation, claim the whole block.
2492 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2493 page_group_by_mobility_disabled)
2494 set_pageblock_migratetype(page, start_type);
2499 move_to_free_list(page, zone, current_order, start_type);
2503 * Check whether there is a suitable fallback freepage with requested order.
2504 * If only_stealable is true, this function returns fallback_mt only if
2505 * we can steal other freepages all together. This would help to reduce
2506 * fragmentation due to mixed migratetype pages in one pageblock.
2508 int find_suitable_fallback(struct free_area *area, unsigned int order,
2509 int migratetype, bool only_stealable, bool *can_steal)
2514 if (area->nr_free == 0)
2519 fallback_mt = fallbacks[migratetype][i];
2520 if (fallback_mt == MIGRATE_TYPES)
2523 if (free_area_empty(area, fallback_mt))
2526 if (can_steal_fallback(order, migratetype))
2529 if (!only_stealable)
2540 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2541 * there are no empty page blocks that contain a page with a suitable order
2543 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2544 unsigned int alloc_order)
2547 unsigned long max_managed, flags;
2550 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2551 * Check is race-prone but harmless.
2553 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2554 if (zone->nr_reserved_highatomic >= max_managed)
2557 spin_lock_irqsave(&zone->lock, flags);
2559 /* Recheck the nr_reserved_highatomic limit under the lock */
2560 if (zone->nr_reserved_highatomic >= max_managed)
2564 mt = get_pageblock_migratetype(page);
2565 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2566 && !is_migrate_cma(mt)) {
2567 zone->nr_reserved_highatomic += pageblock_nr_pages;
2568 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2569 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2573 spin_unlock_irqrestore(&zone->lock, flags);
2577 * Used when an allocation is about to fail under memory pressure. This
2578 * potentially hurts the reliability of high-order allocations when under
2579 * intense memory pressure but failed atomic allocations should be easier
2580 * to recover from than an OOM.
2582 * If @force is true, try to unreserve a pageblock even though highatomic
2583 * pageblock is exhausted.
2585 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2588 struct zonelist *zonelist = ac->zonelist;
2589 unsigned long flags;
2596 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2599 * Preserve at least one pageblock unless memory pressure
2602 if (!force && zone->nr_reserved_highatomic <=
2606 spin_lock_irqsave(&zone->lock, flags);
2607 for (order = 0; order < MAX_ORDER; order++) {
2608 struct free_area *area = &(zone->free_area[order]);
2610 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2615 * In page freeing path, migratetype change is racy so
2616 * we can counter several free pages in a pageblock
2617 * in this loop althoug we changed the pageblock type
2618 * from highatomic to ac->migratetype. So we should
2619 * adjust the count once.
2621 if (is_migrate_highatomic_page(page)) {
2623 * It should never happen but changes to
2624 * locking could inadvertently allow a per-cpu
2625 * drain to add pages to MIGRATE_HIGHATOMIC
2626 * while unreserving so be safe and watch for
2629 zone->nr_reserved_highatomic -= min(
2631 zone->nr_reserved_highatomic);
2635 * Convert to ac->migratetype and avoid the normal
2636 * pageblock stealing heuristics. Minimally, the caller
2637 * is doing the work and needs the pages. More
2638 * importantly, if the block was always converted to
2639 * MIGRATE_UNMOVABLE or another type then the number
2640 * of pageblocks that cannot be completely freed
2643 set_pageblock_migratetype(page, ac->migratetype);
2644 ret = move_freepages_block(zone, page, ac->migratetype,
2647 spin_unlock_irqrestore(&zone->lock, flags);
2651 spin_unlock_irqrestore(&zone->lock, flags);
2658 * Try finding a free buddy page on the fallback list and put it on the free
2659 * list of requested migratetype, possibly along with other pages from the same
2660 * block, depending on fragmentation avoidance heuristics. Returns true if
2661 * fallback was found so that __rmqueue_smallest() can grab it.
2663 * The use of signed ints for order and current_order is a deliberate
2664 * deviation from the rest of this file, to make the for loop
2665 * condition simpler.
2667 static __always_inline bool
2668 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2669 unsigned int alloc_flags)
2671 struct free_area *area;
2673 int min_order = order;
2679 * Do not steal pages from freelists belonging to other pageblocks
2680 * i.e. orders < pageblock_order. If there are no local zones free,
2681 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2683 if (alloc_flags & ALLOC_NOFRAGMENT)
2684 min_order = pageblock_order;
2687 * Find the largest available free page in the other list. This roughly
2688 * approximates finding the pageblock with the most free pages, which
2689 * would be too costly to do exactly.
2691 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2693 area = &(zone->free_area[current_order]);
2694 fallback_mt = find_suitable_fallback(area, current_order,
2695 start_migratetype, false, &can_steal);
2696 if (fallback_mt == -1)
2700 * We cannot steal all free pages from the pageblock and the
2701 * requested migratetype is movable. In that case it's better to
2702 * steal and split the smallest available page instead of the
2703 * largest available page, because even if the next movable
2704 * allocation falls back into a different pageblock than this
2705 * one, it won't cause permanent fragmentation.
2707 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2708 && current_order > order)
2717 for (current_order = order; current_order < MAX_ORDER;
2719 area = &(zone->free_area[current_order]);
2720 fallback_mt = find_suitable_fallback(area, current_order,
2721 start_migratetype, false, &can_steal);
2722 if (fallback_mt != -1)
2727 * This should not happen - we already found a suitable fallback
2728 * when looking for the largest page.
2730 VM_BUG_ON(current_order == MAX_ORDER);
2733 page = get_page_from_free_area(area, fallback_mt);
2735 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2738 trace_mm_page_alloc_extfrag(page, order, current_order,
2739 start_migratetype, fallback_mt);
2746 * Do the hard work of removing an element from the buddy allocator.
2747 * Call me with the zone->lock already held.
2749 static __always_inline struct page *
2750 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2751 unsigned int alloc_flags)
2757 * Balance movable allocations between regular and CMA areas by
2758 * allocating from CMA when over half of the zone's free memory
2759 * is in the CMA area.
2761 if (migratetype == MIGRATE_MOVABLE &&
2762 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2763 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2764 page = __rmqueue_cma_fallback(zone, order);
2770 page = __rmqueue_smallest(zone, order, migratetype);
2771 if (unlikely(!page)) {
2772 if (migratetype == MIGRATE_MOVABLE)
2773 page = __rmqueue_cma_fallback(zone, order);
2775 if (!page && __rmqueue_fallback(zone, order, migratetype,
2780 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2785 * Obtain a specified number of elements from the buddy allocator, all under
2786 * a single hold of the lock, for efficiency. Add them to the supplied list.
2787 * Returns the number of new pages which were placed at *list.
2789 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2790 unsigned long count, struct list_head *list,
2791 int migratetype, unsigned int alloc_flags)
2795 spin_lock(&zone->lock);
2796 for (i = 0; i < count; ++i) {
2797 struct page *page = __rmqueue(zone, order, migratetype,
2799 if (unlikely(page == NULL))
2802 if (unlikely(check_pcp_refill(page)))
2806 * Split buddy pages returned by expand() are received here in
2807 * physical page order. The page is added to the tail of
2808 * caller's list. From the callers perspective, the linked list
2809 * is ordered by page number under some conditions. This is
2810 * useful for IO devices that can forward direction from the
2811 * head, thus also in the physical page order. This is useful
2812 * for IO devices that can merge IO requests if the physical
2813 * pages are ordered properly.
2815 list_add_tail(&page->lru, list);
2817 if (is_migrate_cma(get_pcppage_migratetype(page)))
2818 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2823 * i pages were removed from the buddy list even if some leak due
2824 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2825 * on i. Do not confuse with 'alloced' which is the number of
2826 * pages added to the pcp list.
2828 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2829 spin_unlock(&zone->lock);
2835 * Called from the vmstat counter updater to drain pagesets of this
2836 * currently executing processor on remote nodes after they have
2839 * Note that this function must be called with the thread pinned to
2840 * a single processor.
2842 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2844 unsigned long flags;
2845 int to_drain, batch;
2847 local_irq_save(flags);
2848 batch = READ_ONCE(pcp->batch);
2849 to_drain = min(pcp->count, batch);
2851 free_pcppages_bulk(zone, to_drain, pcp);
2852 local_irq_restore(flags);
2857 * Drain pcplists of the indicated processor and zone.
2859 * The processor must either be the current processor and the
2860 * thread pinned to the current processor or a processor that
2863 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2865 unsigned long flags;
2866 struct per_cpu_pageset *pset;
2867 struct per_cpu_pages *pcp;
2869 local_irq_save(flags);
2870 pset = per_cpu_ptr(zone->pageset, cpu);
2874 free_pcppages_bulk(zone, pcp->count, pcp);
2875 local_irq_restore(flags);
2879 * Drain pcplists of all zones on the indicated processor.
2881 * The processor must either be the current processor and the
2882 * thread pinned to the current processor or a processor that
2885 static void drain_pages(unsigned int cpu)
2889 for_each_populated_zone(zone) {
2890 drain_pages_zone(cpu, zone);
2895 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2897 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2898 * the single zone's pages.
2900 void drain_local_pages(struct zone *zone)
2902 int cpu = smp_processor_id();
2905 drain_pages_zone(cpu, zone);
2910 static void drain_local_pages_wq(struct work_struct *work)
2912 struct pcpu_drain *drain;
2914 drain = container_of(work, struct pcpu_drain, work);
2917 * drain_all_pages doesn't use proper cpu hotplug protection so
2918 * we can race with cpu offline when the WQ can move this from
2919 * a cpu pinned worker to an unbound one. We can operate on a different
2920 * cpu which is allright but we also have to make sure to not move to
2924 drain_local_pages(drain->zone);
2929 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2931 * When zone parameter is non-NULL, spill just the single zone's pages.
2933 * Note that this can be extremely slow as the draining happens in a workqueue.
2935 void drain_all_pages(struct zone *zone)
2940 * Allocate in the BSS so we wont require allocation in
2941 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2943 static cpumask_t cpus_with_pcps;
2946 * Make sure nobody triggers this path before mm_percpu_wq is fully
2949 if (WARN_ON_ONCE(!mm_percpu_wq))
2953 * Do not drain if one is already in progress unless it's specific to
2954 * a zone. Such callers are primarily CMA and memory hotplug and need
2955 * the drain to be complete when the call returns.
2957 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2960 mutex_lock(&pcpu_drain_mutex);
2964 * We don't care about racing with CPU hotplug event
2965 * as offline notification will cause the notified
2966 * cpu to drain that CPU pcps and on_each_cpu_mask
2967 * disables preemption as part of its processing
2969 for_each_online_cpu(cpu) {
2970 struct per_cpu_pageset *pcp;
2972 bool has_pcps = false;
2975 pcp = per_cpu_ptr(zone->pageset, cpu);
2979 for_each_populated_zone(z) {
2980 pcp = per_cpu_ptr(z->pageset, cpu);
2981 if (pcp->pcp.count) {
2989 cpumask_set_cpu(cpu, &cpus_with_pcps);
2991 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2994 for_each_cpu(cpu, &cpus_with_pcps) {
2995 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
2998 INIT_WORK(&drain->work, drain_local_pages_wq);
2999 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3001 for_each_cpu(cpu, &cpus_with_pcps)
3002 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3004 mutex_unlock(&pcpu_drain_mutex);
3007 #ifdef CONFIG_HIBERNATION
3010 * Touch the watchdog for every WD_PAGE_COUNT pages.
3012 #define WD_PAGE_COUNT (128*1024)
3014 void mark_free_pages(struct zone *zone)
3016 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3017 unsigned long flags;
3018 unsigned int order, t;
3021 if (zone_is_empty(zone))
3024 spin_lock_irqsave(&zone->lock, flags);
3026 max_zone_pfn = zone_end_pfn(zone);
3027 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3028 if (pfn_valid(pfn)) {
3029 page = pfn_to_page(pfn);
3031 if (!--page_count) {
3032 touch_nmi_watchdog();
3033 page_count = WD_PAGE_COUNT;
3036 if (page_zone(page) != zone)
3039 if (!swsusp_page_is_forbidden(page))
3040 swsusp_unset_page_free(page);
3043 for_each_migratetype_order(order, t) {
3044 list_for_each_entry(page,
3045 &zone->free_area[order].free_list[t], lru) {
3048 pfn = page_to_pfn(page);
3049 for (i = 0; i < (1UL << order); i++) {
3050 if (!--page_count) {
3051 touch_nmi_watchdog();
3052 page_count = WD_PAGE_COUNT;
3054 swsusp_set_page_free(pfn_to_page(pfn + i));
3058 spin_unlock_irqrestore(&zone->lock, flags);
3060 #endif /* CONFIG_PM */
3062 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3066 if (!free_pcp_prepare(page))
3069 migratetype = get_pfnblock_migratetype(page, pfn);
3070 set_pcppage_migratetype(page, migratetype);
3074 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3076 struct zone *zone = page_zone(page);
3077 struct per_cpu_pages *pcp;
3080 migratetype = get_pcppage_migratetype(page);
3081 __count_vm_event(PGFREE);
3084 * We only track unmovable, reclaimable and movable on pcp lists.
3085 * Free ISOLATE pages back to the allocator because they are being
3086 * offlined but treat HIGHATOMIC as movable pages so we can get those
3087 * areas back if necessary. Otherwise, we may have to free
3088 * excessively into the page allocator
3090 if (migratetype >= MIGRATE_PCPTYPES) {
3091 if (unlikely(is_migrate_isolate(migratetype))) {
3092 free_one_page(zone, page, pfn, 0, migratetype);
3095 migratetype = MIGRATE_MOVABLE;
3098 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3099 list_add(&page->lru, &pcp->lists[migratetype]);
3101 if (pcp->count >= pcp->high) {
3102 unsigned long batch = READ_ONCE(pcp->batch);
3103 free_pcppages_bulk(zone, batch, pcp);
3108 * Free a 0-order page
3110 void free_unref_page(struct page *page)
3112 unsigned long flags;
3113 unsigned long pfn = page_to_pfn(page);
3115 if (!free_unref_page_prepare(page, pfn))
3118 local_irq_save(flags);
3119 free_unref_page_commit(page, pfn);
3120 local_irq_restore(flags);
3124 * Free a list of 0-order pages
3126 void free_unref_page_list(struct list_head *list)
3128 struct page *page, *next;
3129 unsigned long flags, pfn;
3130 int batch_count = 0;
3132 /* Prepare pages for freeing */
3133 list_for_each_entry_safe(page, next, list, lru) {
3134 pfn = page_to_pfn(page);
3135 if (!free_unref_page_prepare(page, pfn))
3136 list_del(&page->lru);
3137 set_page_private(page, pfn);
3140 local_irq_save(flags);
3141 list_for_each_entry_safe(page, next, list, lru) {
3142 unsigned long pfn = page_private(page);
3144 set_page_private(page, 0);
3145 trace_mm_page_free_batched(page);
3146 free_unref_page_commit(page, pfn);
3149 * Guard against excessive IRQ disabled times when we get
3150 * a large list of pages to free.
3152 if (++batch_count == SWAP_CLUSTER_MAX) {
3153 local_irq_restore(flags);
3155 local_irq_save(flags);
3158 local_irq_restore(flags);
3162 * split_page takes a non-compound higher-order page, and splits it into
3163 * n (1<<order) sub-pages: page[0..n]
3164 * Each sub-page must be freed individually.
3166 * Note: this is probably too low level an operation for use in drivers.
3167 * Please consult with lkml before using this in your driver.
3169 void split_page(struct page *page, unsigned int order)
3173 VM_BUG_ON_PAGE(PageCompound(page), page);
3174 VM_BUG_ON_PAGE(!page_count(page), page);
3176 for (i = 1; i < (1 << order); i++)
3177 set_page_refcounted(page + i);
3178 split_page_owner(page, order);
3180 EXPORT_SYMBOL_GPL(split_page);
3182 int __isolate_free_page(struct page *page, unsigned int order)
3184 unsigned long watermark;
3188 BUG_ON(!PageBuddy(page));
3190 zone = page_zone(page);
3191 mt = get_pageblock_migratetype(page);
3193 if (!is_migrate_isolate(mt)) {
3195 * Obey watermarks as if the page was being allocated. We can
3196 * emulate a high-order watermark check with a raised order-0
3197 * watermark, because we already know our high-order page
3200 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3201 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3204 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3207 /* Remove page from free list */
3209 del_page_from_free_list(page, zone, order);
3212 * Set the pageblock if the isolated page is at least half of a
3215 if (order >= pageblock_order - 1) {
3216 struct page *endpage = page + (1 << order) - 1;
3217 for (; page < endpage; page += pageblock_nr_pages) {
3218 int mt = get_pageblock_migratetype(page);
3219 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3220 && !is_migrate_highatomic(mt))
3221 set_pageblock_migratetype(page,
3227 return 1UL << order;
3231 * __putback_isolated_page - Return a now-isolated page back where we got it
3232 * @page: Page that was isolated
3233 * @order: Order of the isolated page
3234 * @mt: The page's pageblock's migratetype
3236 * This function is meant to return a page pulled from the free lists via
3237 * __isolate_free_page back to the free lists they were pulled from.
3239 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3241 struct zone *zone = page_zone(page);
3243 /* zone lock should be held when this function is called */
3244 lockdep_assert_held(&zone->lock);
3246 /* Return isolated page to tail of freelist. */
3247 __free_one_page(page, page_to_pfn(page), zone, order, mt, false);
3251 * Update NUMA hit/miss statistics
3253 * Must be called with interrupts disabled.
3255 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3258 enum numa_stat_item local_stat = NUMA_LOCAL;
3260 /* skip numa counters update if numa stats is disabled */
3261 if (!static_branch_likely(&vm_numa_stat_key))
3264 if (zone_to_nid(z) != numa_node_id())
3265 local_stat = NUMA_OTHER;
3267 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3268 __inc_numa_state(z, NUMA_HIT);
3270 __inc_numa_state(z, NUMA_MISS);
3271 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3273 __inc_numa_state(z, local_stat);
3277 /* Remove page from the per-cpu list, caller must protect the list */
3278 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3279 unsigned int alloc_flags,
3280 struct per_cpu_pages *pcp,
3281 struct list_head *list)
3286 if (list_empty(list)) {
3287 pcp->count += rmqueue_bulk(zone, 0,
3289 migratetype, alloc_flags);
3290 if (unlikely(list_empty(list)))
3294 page = list_first_entry(list, struct page, lru);
3295 list_del(&page->lru);
3297 } while (check_new_pcp(page));
3302 /* Lock and remove page from the per-cpu list */
3303 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3304 struct zone *zone, gfp_t gfp_flags,
3305 int migratetype, unsigned int alloc_flags)
3307 struct per_cpu_pages *pcp;
3308 struct list_head *list;
3310 unsigned long flags;
3312 local_irq_save(flags);
3313 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3314 list = &pcp->lists[migratetype];
3315 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3317 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3318 zone_statistics(preferred_zone, zone);
3320 local_irq_restore(flags);
3325 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3328 struct page *rmqueue(struct zone *preferred_zone,
3329 struct zone *zone, unsigned int order,
3330 gfp_t gfp_flags, unsigned int alloc_flags,
3333 unsigned long flags;
3336 if (likely(order == 0)) {
3337 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3338 migratetype, alloc_flags);
3343 * We most definitely don't want callers attempting to
3344 * allocate greater than order-1 page units with __GFP_NOFAIL.
3346 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3347 spin_lock_irqsave(&zone->lock, flags);
3351 if (alloc_flags & ALLOC_HARDER) {
3352 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3354 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3357 page = __rmqueue(zone, order, migratetype, alloc_flags);
3358 } while (page && check_new_pages(page, order));
3359 spin_unlock(&zone->lock);
3362 __mod_zone_freepage_state(zone, -(1 << order),
3363 get_pcppage_migratetype(page));
3365 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3366 zone_statistics(preferred_zone, zone);
3367 local_irq_restore(flags);
3370 /* Separate test+clear to avoid unnecessary atomics */
3371 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3372 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3373 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3376 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3380 local_irq_restore(flags);
3384 #ifdef CONFIG_FAIL_PAGE_ALLOC
3387 struct fault_attr attr;
3389 bool ignore_gfp_highmem;
3390 bool ignore_gfp_reclaim;
3392 } fail_page_alloc = {
3393 .attr = FAULT_ATTR_INITIALIZER,
3394 .ignore_gfp_reclaim = true,
3395 .ignore_gfp_highmem = true,
3399 static int __init setup_fail_page_alloc(char *str)
3401 return setup_fault_attr(&fail_page_alloc.attr, str);
3403 __setup("fail_page_alloc=", setup_fail_page_alloc);
3405 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3407 if (order < fail_page_alloc.min_order)
3409 if (gfp_mask & __GFP_NOFAIL)
3411 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3413 if (fail_page_alloc.ignore_gfp_reclaim &&
3414 (gfp_mask & __GFP_DIRECT_RECLAIM))
3417 return should_fail(&fail_page_alloc.attr, 1 << order);
3420 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3422 static int __init fail_page_alloc_debugfs(void)
3424 umode_t mode = S_IFREG | 0600;
3427 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3428 &fail_page_alloc.attr);
3430 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3431 &fail_page_alloc.ignore_gfp_reclaim);
3432 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3433 &fail_page_alloc.ignore_gfp_highmem);
3434 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3439 late_initcall(fail_page_alloc_debugfs);
3441 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3443 #else /* CONFIG_FAIL_PAGE_ALLOC */
3445 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3450 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3452 static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3454 return __should_fail_alloc_page(gfp_mask, order);
3456 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3459 * Return true if free base pages are above 'mark'. For high-order checks it
3460 * will return true of the order-0 watermark is reached and there is at least
3461 * one free page of a suitable size. Checking now avoids taking the zone lock
3462 * to check in the allocation paths if no pages are free.
3464 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3465 int classzone_idx, unsigned int alloc_flags,
3470 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3472 /* free_pages may go negative - that's OK */
3473 free_pages -= (1 << order) - 1;
3475 if (alloc_flags & ALLOC_HIGH)
3479 * If the caller does not have rights to ALLOC_HARDER then subtract
3480 * the high-atomic reserves. This will over-estimate the size of the
3481 * atomic reserve but it avoids a search.
3483 if (likely(!alloc_harder)) {
3484 free_pages -= z->nr_reserved_highatomic;
3487 * OOM victims can try even harder than normal ALLOC_HARDER
3488 * users on the grounds that it's definitely going to be in
3489 * the exit path shortly and free memory. Any allocation it
3490 * makes during the free path will be small and short-lived.
3492 if (alloc_flags & ALLOC_OOM)
3500 /* If allocation can't use CMA areas don't use free CMA pages */
3501 if (!(alloc_flags & ALLOC_CMA))
3502 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3506 * Check watermarks for an order-0 allocation request. If these
3507 * are not met, then a high-order request also cannot go ahead
3508 * even if a suitable page happened to be free.
3510 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3513 /* If this is an order-0 request then the watermark is fine */
3517 /* For a high-order request, check at least one suitable page is free */
3518 for (o = order; o < MAX_ORDER; o++) {
3519 struct free_area *area = &z->free_area[o];
3525 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3526 if (!free_area_empty(area, mt))
3531 if ((alloc_flags & ALLOC_CMA) &&
3532 !free_area_empty(area, MIGRATE_CMA)) {
3536 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3542 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3543 int classzone_idx, unsigned int alloc_flags)
3545 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3546 zone_page_state(z, NR_FREE_PAGES));
3549 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3550 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3552 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3556 /* If allocation can't use CMA areas don't use free CMA pages */
3557 if (!(alloc_flags & ALLOC_CMA))
3558 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3562 * Fast check for order-0 only. If this fails then the reserves
3563 * need to be calculated. There is a corner case where the check
3564 * passes but only the high-order atomic reserve are free. If
3565 * the caller is !atomic then it'll uselessly search the free
3566 * list. That corner case is then slower but it is harmless.
3568 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3571 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3575 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3576 unsigned long mark, int classzone_idx)
3578 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3580 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3581 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3583 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3588 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3590 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3591 node_reclaim_distance;
3593 #else /* CONFIG_NUMA */
3594 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3598 #endif /* CONFIG_NUMA */
3601 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3602 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3603 * premature use of a lower zone may cause lowmem pressure problems that
3604 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3605 * probably too small. It only makes sense to spread allocations to avoid
3606 * fragmentation between the Normal and DMA32 zones.
3608 static inline unsigned int
3609 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3611 unsigned int alloc_flags;
3614 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3617 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3619 #ifdef CONFIG_ZONE_DMA32
3623 if (zone_idx(zone) != ZONE_NORMAL)
3627 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3628 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3629 * on UMA that if Normal is populated then so is DMA32.
3631 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3632 if (nr_online_nodes > 1 && !populated_zone(--zone))
3635 alloc_flags |= ALLOC_NOFRAGMENT;
3636 #endif /* CONFIG_ZONE_DMA32 */
3641 * get_page_from_freelist goes through the zonelist trying to allocate
3644 static struct page *
3645 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3646 const struct alloc_context *ac)
3650 struct pglist_data *last_pgdat_dirty_limit = NULL;
3655 * Scan zonelist, looking for a zone with enough free.
3656 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3658 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3659 z = ac->preferred_zoneref;
3660 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3665 if (cpusets_enabled() &&
3666 (alloc_flags & ALLOC_CPUSET) &&
3667 !__cpuset_zone_allowed(zone, gfp_mask))
3670 * When allocating a page cache page for writing, we
3671 * want to get it from a node that is within its dirty
3672 * limit, such that no single node holds more than its
3673 * proportional share of globally allowed dirty pages.
3674 * The dirty limits take into account the node's
3675 * lowmem reserves and high watermark so that kswapd
3676 * should be able to balance it without having to
3677 * write pages from its LRU list.
3679 * XXX: For now, allow allocations to potentially
3680 * exceed the per-node dirty limit in the slowpath
3681 * (spread_dirty_pages unset) before going into reclaim,
3682 * which is important when on a NUMA setup the allowed
3683 * nodes are together not big enough to reach the
3684 * global limit. The proper fix for these situations
3685 * will require awareness of nodes in the
3686 * dirty-throttling and the flusher threads.
3688 if (ac->spread_dirty_pages) {
3689 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3692 if (!node_dirty_ok(zone->zone_pgdat)) {
3693 last_pgdat_dirty_limit = zone->zone_pgdat;
3698 if (no_fallback && nr_online_nodes > 1 &&
3699 zone != ac->preferred_zoneref->zone) {
3703 * If moving to a remote node, retry but allow
3704 * fragmenting fallbacks. Locality is more important
3705 * than fragmentation avoidance.
3707 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3708 if (zone_to_nid(zone) != local_nid) {
3709 alloc_flags &= ~ALLOC_NOFRAGMENT;
3714 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3715 if (!zone_watermark_fast(zone, order, mark,
3716 ac_classzone_idx(ac), alloc_flags)) {
3719 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3721 * Watermark failed for this zone, but see if we can
3722 * grow this zone if it contains deferred pages.
3724 if (static_branch_unlikely(&deferred_pages)) {
3725 if (_deferred_grow_zone(zone, order))
3729 /* Checked here to keep the fast path fast */
3730 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3731 if (alloc_flags & ALLOC_NO_WATERMARKS)
3734 if (node_reclaim_mode == 0 ||
3735 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3738 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3740 case NODE_RECLAIM_NOSCAN:
3743 case NODE_RECLAIM_FULL:
3744 /* scanned but unreclaimable */
3747 /* did we reclaim enough */
3748 if (zone_watermark_ok(zone, order, mark,
3749 ac_classzone_idx(ac), alloc_flags))
3757 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3758 gfp_mask, alloc_flags, ac->migratetype);
3760 prep_new_page(page, order, gfp_mask, alloc_flags);
3763 * If this is a high-order atomic allocation then check
3764 * if the pageblock should be reserved for the future
3766 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3767 reserve_highatomic_pageblock(page, zone, order);
3771 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3772 /* Try again if zone has deferred pages */
3773 if (static_branch_unlikely(&deferred_pages)) {
3774 if (_deferred_grow_zone(zone, order))
3782 * It's possible on a UMA machine to get through all zones that are
3783 * fragmented. If avoiding fragmentation, reset and try again.
3786 alloc_flags &= ~ALLOC_NOFRAGMENT;
3793 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3795 unsigned int filter = SHOW_MEM_FILTER_NODES;
3798 * This documents exceptions given to allocations in certain
3799 * contexts that are allowed to allocate outside current's set
3802 if (!(gfp_mask & __GFP_NOMEMALLOC))
3803 if (tsk_is_oom_victim(current) ||
3804 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3805 filter &= ~SHOW_MEM_FILTER_NODES;
3806 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3807 filter &= ~SHOW_MEM_FILTER_NODES;
3809 show_mem(filter, nodemask);
3812 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3814 struct va_format vaf;
3816 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3818 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3821 va_start(args, fmt);
3824 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3825 current->comm, &vaf, gfp_mask, &gfp_mask,
3826 nodemask_pr_args(nodemask));
3829 cpuset_print_current_mems_allowed();
3832 warn_alloc_show_mem(gfp_mask, nodemask);
3835 static inline struct page *
3836 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3837 unsigned int alloc_flags,
3838 const struct alloc_context *ac)
3842 page = get_page_from_freelist(gfp_mask, order,
3843 alloc_flags|ALLOC_CPUSET, ac);
3845 * fallback to ignore cpuset restriction if our nodes
3849 page = get_page_from_freelist(gfp_mask, order,
3855 static inline struct page *
3856 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3857 const struct alloc_context *ac, unsigned long *did_some_progress)
3859 struct oom_control oc = {
3860 .zonelist = ac->zonelist,
3861 .nodemask = ac->nodemask,
3863 .gfp_mask = gfp_mask,
3868 *did_some_progress = 0;
3871 * Acquire the oom lock. If that fails, somebody else is
3872 * making progress for us.
3874 if (!mutex_trylock(&oom_lock)) {
3875 *did_some_progress = 1;
3876 schedule_timeout_uninterruptible(1);
3881 * Go through the zonelist yet one more time, keep very high watermark
3882 * here, this is only to catch a parallel oom killing, we must fail if
3883 * we're still under heavy pressure. But make sure that this reclaim
3884 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3885 * allocation which will never fail due to oom_lock already held.
3887 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3888 ~__GFP_DIRECT_RECLAIM, order,
3889 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3893 /* Coredumps can quickly deplete all memory reserves */
3894 if (current->flags & PF_DUMPCORE)
3896 /* The OOM killer will not help higher order allocs */
3897 if (order > PAGE_ALLOC_COSTLY_ORDER)
3900 * We have already exhausted all our reclaim opportunities without any
3901 * success so it is time to admit defeat. We will skip the OOM killer
3902 * because it is very likely that the caller has a more reasonable
3903 * fallback than shooting a random task.
3905 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3907 /* The OOM killer does not needlessly kill tasks for lowmem */
3908 if (ac->high_zoneidx < ZONE_NORMAL)
3910 if (pm_suspended_storage())
3913 * XXX: GFP_NOFS allocations should rather fail than rely on
3914 * other request to make a forward progress.
3915 * We are in an unfortunate situation where out_of_memory cannot
3916 * do much for this context but let's try it to at least get
3917 * access to memory reserved if the current task is killed (see
3918 * out_of_memory). Once filesystems are ready to handle allocation
3919 * failures more gracefully we should just bail out here.
3922 /* The OOM killer may not free memory on a specific node */
3923 if (gfp_mask & __GFP_THISNODE)
3926 /* Exhausted what can be done so it's blame time */
3927 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3928 *did_some_progress = 1;
3931 * Help non-failing allocations by giving them access to memory
3934 if (gfp_mask & __GFP_NOFAIL)
3935 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3936 ALLOC_NO_WATERMARKS, ac);
3939 mutex_unlock(&oom_lock);
3944 * Maximum number of compaction retries wit a progress before OOM
3945 * killer is consider as the only way to move forward.
3947 #define MAX_COMPACT_RETRIES 16
3949 #ifdef CONFIG_COMPACTION
3950 /* Try memory compaction for high-order allocations before reclaim */
3951 static struct page *
3952 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3953 unsigned int alloc_flags, const struct alloc_context *ac,
3954 enum compact_priority prio, enum compact_result *compact_result)
3956 struct page *page = NULL;
3957 unsigned long pflags;
3958 unsigned int noreclaim_flag;
3963 psi_memstall_enter(&pflags);
3964 noreclaim_flag = memalloc_noreclaim_save();
3966 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3969 memalloc_noreclaim_restore(noreclaim_flag);
3970 psi_memstall_leave(&pflags);
3973 * At least in one zone compaction wasn't deferred or skipped, so let's
3974 * count a compaction stall
3976 count_vm_event(COMPACTSTALL);
3978 /* Prep a captured page if available */
3980 prep_new_page(page, order, gfp_mask, alloc_flags);
3982 /* Try get a page from the freelist if available */
3984 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3987 struct zone *zone = page_zone(page);
3989 zone->compact_blockskip_flush = false;
3990 compaction_defer_reset(zone, order, true);
3991 count_vm_event(COMPACTSUCCESS);
3996 * It's bad if compaction run occurs and fails. The most likely reason
3997 * is that pages exist, but not enough to satisfy watermarks.
3999 count_vm_event(COMPACTFAIL);
4007 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4008 enum compact_result compact_result,
4009 enum compact_priority *compact_priority,
4010 int *compaction_retries)
4012 int max_retries = MAX_COMPACT_RETRIES;
4015 int retries = *compaction_retries;
4016 enum compact_priority priority = *compact_priority;
4021 if (compaction_made_progress(compact_result))
4022 (*compaction_retries)++;
4025 * compaction considers all the zone as desperately out of memory
4026 * so it doesn't really make much sense to retry except when the
4027 * failure could be caused by insufficient priority
4029 if (compaction_failed(compact_result))
4030 goto check_priority;
4033 * compaction was skipped because there are not enough order-0 pages
4034 * to work with, so we retry only if it looks like reclaim can help.
4036 if (compaction_needs_reclaim(compact_result)) {
4037 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4042 * make sure the compaction wasn't deferred or didn't bail out early
4043 * due to locks contention before we declare that we should give up.
4044 * But the next retry should use a higher priority if allowed, so
4045 * we don't just keep bailing out endlessly.
4047 if (compaction_withdrawn(compact_result)) {
4048 goto check_priority;
4052 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4053 * costly ones because they are de facto nofail and invoke OOM
4054 * killer to move on while costly can fail and users are ready
4055 * to cope with that. 1/4 retries is rather arbitrary but we
4056 * would need much more detailed feedback from compaction to
4057 * make a better decision.
4059 if (order > PAGE_ALLOC_COSTLY_ORDER)
4061 if (*compaction_retries <= max_retries) {
4067 * Make sure there are attempts at the highest priority if we exhausted
4068 * all retries or failed at the lower priorities.
4071 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4072 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4074 if (*compact_priority > min_priority) {
4075 (*compact_priority)--;
4076 *compaction_retries = 0;
4080 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4084 static inline struct page *
4085 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4086 unsigned int alloc_flags, const struct alloc_context *ac,
4087 enum compact_priority prio, enum compact_result *compact_result)
4089 *compact_result = COMPACT_SKIPPED;
4094 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4095 enum compact_result compact_result,
4096 enum compact_priority *compact_priority,
4097 int *compaction_retries)
4102 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4106 * There are setups with compaction disabled which would prefer to loop
4107 * inside the allocator rather than hit the oom killer prematurely.
4108 * Let's give them a good hope and keep retrying while the order-0
4109 * watermarks are OK.
4111 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4113 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4114 ac_classzone_idx(ac), alloc_flags))
4119 #endif /* CONFIG_COMPACTION */
4121 #ifdef CONFIG_LOCKDEP
4122 static struct lockdep_map __fs_reclaim_map =
4123 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4125 static bool __need_fs_reclaim(gfp_t gfp_mask)
4127 gfp_mask = current_gfp_context(gfp_mask);
4129 /* no reclaim without waiting on it */
4130 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4133 /* this guy won't enter reclaim */
4134 if (current->flags & PF_MEMALLOC)
4137 /* We're only interested __GFP_FS allocations for now */
4138 if (!(gfp_mask & __GFP_FS))
4141 if (gfp_mask & __GFP_NOLOCKDEP)
4147 void __fs_reclaim_acquire(void)
4149 lock_map_acquire(&__fs_reclaim_map);
4152 void __fs_reclaim_release(void)
4154 lock_map_release(&__fs_reclaim_map);
4157 void fs_reclaim_acquire(gfp_t gfp_mask)
4159 if (__need_fs_reclaim(gfp_mask))
4160 __fs_reclaim_acquire();
4162 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4164 void fs_reclaim_release(gfp_t gfp_mask)
4166 if (__need_fs_reclaim(gfp_mask))
4167 __fs_reclaim_release();
4169 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4172 /* Perform direct synchronous page reclaim */
4174 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4175 const struct alloc_context *ac)
4178 unsigned int noreclaim_flag;
4179 unsigned long pflags;
4183 /* We now go into synchronous reclaim */
4184 cpuset_memory_pressure_bump();
4185 psi_memstall_enter(&pflags);
4186 fs_reclaim_acquire(gfp_mask);
4187 noreclaim_flag = memalloc_noreclaim_save();
4189 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4192 memalloc_noreclaim_restore(noreclaim_flag);
4193 fs_reclaim_release(gfp_mask);
4194 psi_memstall_leave(&pflags);
4201 /* The really slow allocator path where we enter direct reclaim */
4202 static inline struct page *
4203 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4204 unsigned int alloc_flags, const struct alloc_context *ac,
4205 unsigned long *did_some_progress)
4207 struct page *page = NULL;
4208 bool drained = false;
4210 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4211 if (unlikely(!(*did_some_progress)))
4215 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4218 * If an allocation failed after direct reclaim, it could be because
4219 * pages are pinned on the per-cpu lists or in high alloc reserves.
4220 * Shrink them them and try again
4222 if (!page && !drained) {
4223 unreserve_highatomic_pageblock(ac, false);
4224 drain_all_pages(NULL);
4232 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4233 const struct alloc_context *ac)
4237 pg_data_t *last_pgdat = NULL;
4238 enum zone_type high_zoneidx = ac->high_zoneidx;
4240 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
4242 if (last_pgdat != zone->zone_pgdat)
4243 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
4244 last_pgdat = zone->zone_pgdat;
4248 static inline unsigned int
4249 gfp_to_alloc_flags(gfp_t gfp_mask)
4251 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4254 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4255 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4256 * to save two branches.
4258 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4259 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4262 * The caller may dip into page reserves a bit more if the caller
4263 * cannot run direct reclaim, or if the caller has realtime scheduling
4264 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4265 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4267 alloc_flags |= (__force int)
4268 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4270 if (gfp_mask & __GFP_ATOMIC) {
4272 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4273 * if it can't schedule.
4275 if (!(gfp_mask & __GFP_NOMEMALLOC))
4276 alloc_flags |= ALLOC_HARDER;
4278 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4279 * comment for __cpuset_node_allowed().
4281 alloc_flags &= ~ALLOC_CPUSET;
4282 } else if (unlikely(rt_task(current)) && !in_interrupt())
4283 alloc_flags |= ALLOC_HARDER;
4286 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4287 alloc_flags |= ALLOC_CMA;
4292 static bool oom_reserves_allowed(struct task_struct *tsk)
4294 if (!tsk_is_oom_victim(tsk))
4298 * !MMU doesn't have oom reaper so give access to memory reserves
4299 * only to the thread with TIF_MEMDIE set
4301 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4308 * Distinguish requests which really need access to full memory
4309 * reserves from oom victims which can live with a portion of it
4311 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4313 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4315 if (gfp_mask & __GFP_MEMALLOC)
4316 return ALLOC_NO_WATERMARKS;
4317 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4318 return ALLOC_NO_WATERMARKS;
4319 if (!in_interrupt()) {
4320 if (current->flags & PF_MEMALLOC)
4321 return ALLOC_NO_WATERMARKS;
4322 else if (oom_reserves_allowed(current))
4329 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4331 return !!__gfp_pfmemalloc_flags(gfp_mask);
4335 * Checks whether it makes sense to retry the reclaim to make a forward progress
4336 * for the given allocation request.
4338 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4339 * without success, or when we couldn't even meet the watermark if we
4340 * reclaimed all remaining pages on the LRU lists.
4342 * Returns true if a retry is viable or false to enter the oom path.
4345 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4346 struct alloc_context *ac, int alloc_flags,
4347 bool did_some_progress, int *no_progress_loops)
4354 * Costly allocations might have made a progress but this doesn't mean
4355 * their order will become available due to high fragmentation so
4356 * always increment the no progress counter for them
4358 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4359 *no_progress_loops = 0;
4361 (*no_progress_loops)++;
4364 * Make sure we converge to OOM if we cannot make any progress
4365 * several times in the row.
4367 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4368 /* Before OOM, exhaust highatomic_reserve */
4369 return unreserve_highatomic_pageblock(ac, true);
4373 * Keep reclaiming pages while there is a chance this will lead
4374 * somewhere. If none of the target zones can satisfy our allocation
4375 * request even if all reclaimable pages are considered then we are
4376 * screwed and have to go OOM.
4378 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4380 unsigned long available;
4381 unsigned long reclaimable;
4382 unsigned long min_wmark = min_wmark_pages(zone);
4385 available = reclaimable = zone_reclaimable_pages(zone);
4386 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4389 * Would the allocation succeed if we reclaimed all
4390 * reclaimable pages?
4392 wmark = __zone_watermark_ok(zone, order, min_wmark,
4393 ac_classzone_idx(ac), alloc_flags, available);
4394 trace_reclaim_retry_zone(z, order, reclaimable,
4395 available, min_wmark, *no_progress_loops, wmark);
4398 * If we didn't make any progress and have a lot of
4399 * dirty + writeback pages then we should wait for
4400 * an IO to complete to slow down the reclaim and
4401 * prevent from pre mature OOM
4403 if (!did_some_progress) {
4404 unsigned long write_pending;
4406 write_pending = zone_page_state_snapshot(zone,
4407 NR_ZONE_WRITE_PENDING);
4409 if (2 * write_pending > reclaimable) {
4410 congestion_wait(BLK_RW_ASYNC, HZ/10);
4422 * Memory allocation/reclaim might be called from a WQ context and the
4423 * current implementation of the WQ concurrency control doesn't
4424 * recognize that a particular WQ is congested if the worker thread is
4425 * looping without ever sleeping. Therefore we have to do a short sleep
4426 * here rather than calling cond_resched().
4428 if (current->flags & PF_WQ_WORKER)
4429 schedule_timeout_uninterruptible(1);
4436 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4439 * It's possible that cpuset's mems_allowed and the nodemask from
4440 * mempolicy don't intersect. This should be normally dealt with by
4441 * policy_nodemask(), but it's possible to race with cpuset update in
4442 * such a way the check therein was true, and then it became false
4443 * before we got our cpuset_mems_cookie here.
4444 * This assumes that for all allocations, ac->nodemask can come only
4445 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4446 * when it does not intersect with the cpuset restrictions) or the
4447 * caller can deal with a violated nodemask.
4449 if (cpusets_enabled() && ac->nodemask &&
4450 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4451 ac->nodemask = NULL;
4456 * When updating a task's mems_allowed or mempolicy nodemask, it is
4457 * possible to race with parallel threads in such a way that our
4458 * allocation can fail while the mask is being updated. If we are about
4459 * to fail, check if the cpuset changed during allocation and if so,
4462 if (read_mems_allowed_retry(cpuset_mems_cookie))
4468 static inline struct page *
4469 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4470 struct alloc_context *ac)
4472 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4473 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4474 struct page *page = NULL;
4475 unsigned int alloc_flags;
4476 unsigned long did_some_progress;
4477 enum compact_priority compact_priority;
4478 enum compact_result compact_result;
4479 int compaction_retries;
4480 int no_progress_loops;
4481 unsigned int cpuset_mems_cookie;
4485 * We also sanity check to catch abuse of atomic reserves being used by
4486 * callers that are not in atomic context.
4488 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4489 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4490 gfp_mask &= ~__GFP_ATOMIC;
4493 compaction_retries = 0;
4494 no_progress_loops = 0;
4495 compact_priority = DEF_COMPACT_PRIORITY;
4496 cpuset_mems_cookie = read_mems_allowed_begin();
4499 * The fast path uses conservative alloc_flags to succeed only until
4500 * kswapd needs to be woken up, and to avoid the cost of setting up
4501 * alloc_flags precisely. So we do that now.
4503 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4506 * We need to recalculate the starting point for the zonelist iterator
4507 * because we might have used different nodemask in the fast path, or
4508 * there was a cpuset modification and we are retrying - otherwise we
4509 * could end up iterating over non-eligible zones endlessly.
4511 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4512 ac->high_zoneidx, ac->nodemask);
4513 if (!ac->preferred_zoneref->zone)
4516 if (alloc_flags & ALLOC_KSWAPD)
4517 wake_all_kswapds(order, gfp_mask, ac);
4520 * The adjusted alloc_flags might result in immediate success, so try
4523 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4528 * For costly allocations, try direct compaction first, as it's likely
4529 * that we have enough base pages and don't need to reclaim. For non-
4530 * movable high-order allocations, do that as well, as compaction will
4531 * try prevent permanent fragmentation by migrating from blocks of the
4533 * Don't try this for allocations that are allowed to ignore
4534 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4536 if (can_direct_reclaim &&
4538 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4539 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4540 page = __alloc_pages_direct_compact(gfp_mask, order,
4542 INIT_COMPACT_PRIORITY,
4548 * Checks for costly allocations with __GFP_NORETRY, which
4549 * includes some THP page fault allocations
4551 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4553 * If allocating entire pageblock(s) and compaction
4554 * failed because all zones are below low watermarks
4555 * or is prohibited because it recently failed at this
4556 * order, fail immediately unless the allocator has
4557 * requested compaction and reclaim retry.
4560 * - potentially very expensive because zones are far
4561 * below their low watermarks or this is part of very
4562 * bursty high order allocations,
4563 * - not guaranteed to help because isolate_freepages()
4564 * may not iterate over freed pages as part of its
4566 * - unlikely to make entire pageblocks free on its
4569 if (compact_result == COMPACT_SKIPPED ||
4570 compact_result == COMPACT_DEFERRED)
4574 * Looks like reclaim/compaction is worth trying, but
4575 * sync compaction could be very expensive, so keep
4576 * using async compaction.
4578 compact_priority = INIT_COMPACT_PRIORITY;
4583 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4584 if (alloc_flags & ALLOC_KSWAPD)
4585 wake_all_kswapds(order, gfp_mask, ac);
4587 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4589 alloc_flags = reserve_flags;
4592 * Reset the nodemask and zonelist iterators if memory policies can be
4593 * ignored. These allocations are high priority and system rather than
4596 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4597 ac->nodemask = NULL;
4598 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4599 ac->high_zoneidx, ac->nodemask);
4602 /* Attempt with potentially adjusted zonelist and alloc_flags */
4603 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4607 /* Caller is not willing to reclaim, we can't balance anything */
4608 if (!can_direct_reclaim)
4611 /* Avoid recursion of direct reclaim */
4612 if (current->flags & PF_MEMALLOC)
4615 /* Try direct reclaim and then allocating */
4616 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4617 &did_some_progress);
4621 /* Try direct compaction and then allocating */
4622 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4623 compact_priority, &compact_result);
4627 /* Do not loop if specifically requested */
4628 if (gfp_mask & __GFP_NORETRY)
4632 * Do not retry costly high order allocations unless they are
4633 * __GFP_RETRY_MAYFAIL
4635 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4638 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4639 did_some_progress > 0, &no_progress_loops))
4643 * It doesn't make any sense to retry for the compaction if the order-0
4644 * reclaim is not able to make any progress because the current
4645 * implementation of the compaction depends on the sufficient amount
4646 * of free memory (see __compaction_suitable)
4648 if (did_some_progress > 0 &&
4649 should_compact_retry(ac, order, alloc_flags,
4650 compact_result, &compact_priority,
4651 &compaction_retries))
4655 /* Deal with possible cpuset update races before we start OOM killing */
4656 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4659 /* Reclaim has failed us, start killing things */
4660 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4664 /* Avoid allocations with no watermarks from looping endlessly */
4665 if (tsk_is_oom_victim(current) &&
4666 (alloc_flags == ALLOC_OOM ||
4667 (gfp_mask & __GFP_NOMEMALLOC)))
4670 /* Retry as long as the OOM killer is making progress */
4671 if (did_some_progress) {
4672 no_progress_loops = 0;
4677 /* Deal with possible cpuset update races before we fail */
4678 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4682 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4685 if (gfp_mask & __GFP_NOFAIL) {
4687 * All existing users of the __GFP_NOFAIL are blockable, so warn
4688 * of any new users that actually require GFP_NOWAIT
4690 if (WARN_ON_ONCE(!can_direct_reclaim))
4694 * PF_MEMALLOC request from this context is rather bizarre
4695 * because we cannot reclaim anything and only can loop waiting
4696 * for somebody to do a work for us
4698 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4701 * non failing costly orders are a hard requirement which we
4702 * are not prepared for much so let's warn about these users
4703 * so that we can identify them and convert them to something
4706 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4709 * Help non-failing allocations by giving them access to memory
4710 * reserves but do not use ALLOC_NO_WATERMARKS because this
4711 * could deplete whole memory reserves which would just make
4712 * the situation worse
4714 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4722 warn_alloc(gfp_mask, ac->nodemask,
4723 "page allocation failure: order:%u", order);
4728 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4729 int preferred_nid, nodemask_t *nodemask,
4730 struct alloc_context *ac, gfp_t *alloc_mask,
4731 unsigned int *alloc_flags)
4733 ac->high_zoneidx = gfp_zone(gfp_mask);
4734 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4735 ac->nodemask = nodemask;
4736 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4738 if (cpusets_enabled()) {
4739 *alloc_mask |= __GFP_HARDWALL;
4741 ac->nodemask = &cpuset_current_mems_allowed;
4743 *alloc_flags |= ALLOC_CPUSET;
4746 fs_reclaim_acquire(gfp_mask);
4747 fs_reclaim_release(gfp_mask);
4749 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4751 if (should_fail_alloc_page(gfp_mask, order))
4754 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4755 *alloc_flags |= ALLOC_CMA;
4760 /* Determine whether to spread dirty pages and what the first usable zone */
4761 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4763 /* Dirty zone balancing only done in the fast path */
4764 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4767 * The preferred zone is used for statistics but crucially it is
4768 * also used as the starting point for the zonelist iterator. It
4769 * may get reset for allocations that ignore memory policies.
4771 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4772 ac->high_zoneidx, ac->nodemask);
4776 * This is the 'heart' of the zoned buddy allocator.
4779 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4780 nodemask_t *nodemask)
4783 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4784 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4785 struct alloc_context ac = { };
4788 * There are several places where we assume that the order value is sane
4789 * so bail out early if the request is out of bound.
4791 if (unlikely(order >= MAX_ORDER)) {
4792 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4796 gfp_mask &= gfp_allowed_mask;
4797 alloc_mask = gfp_mask;
4798 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4801 finalise_ac(gfp_mask, &ac);
4804 * Forbid the first pass from falling back to types that fragment
4805 * memory until all local zones are considered.
4807 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4809 /* First allocation attempt */
4810 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4815 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4816 * resp. GFP_NOIO which has to be inherited for all allocation requests
4817 * from a particular context which has been marked by
4818 * memalloc_no{fs,io}_{save,restore}.
4820 alloc_mask = current_gfp_context(gfp_mask);
4821 ac.spread_dirty_pages = false;
4824 * Restore the original nodemask if it was potentially replaced with
4825 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4827 ac.nodemask = nodemask;
4829 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4832 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4833 unlikely(__memcg_kmem_charge_page(page, gfp_mask, order) != 0)) {
4834 __free_pages(page, order);
4838 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4842 EXPORT_SYMBOL(__alloc_pages_nodemask);
4845 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4846 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4847 * you need to access high mem.
4849 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4853 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4856 return (unsigned long) page_address(page);
4858 EXPORT_SYMBOL(__get_free_pages);
4860 unsigned long get_zeroed_page(gfp_t gfp_mask)
4862 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4864 EXPORT_SYMBOL(get_zeroed_page);
4866 static inline void free_the_page(struct page *page, unsigned int order)
4868 if (order == 0) /* Via pcp? */
4869 free_unref_page(page);
4871 __free_pages_ok(page, order);
4874 void __free_pages(struct page *page, unsigned int order)
4876 if (put_page_testzero(page))
4877 free_the_page(page, order);
4879 EXPORT_SYMBOL(__free_pages);
4881 void free_pages(unsigned long addr, unsigned int order)
4884 VM_BUG_ON(!virt_addr_valid((void *)addr));
4885 __free_pages(virt_to_page((void *)addr), order);
4889 EXPORT_SYMBOL(free_pages);
4893 * An arbitrary-length arbitrary-offset area of memory which resides
4894 * within a 0 or higher order page. Multiple fragments within that page
4895 * are individually refcounted, in the page's reference counter.
4897 * The page_frag functions below provide a simple allocation framework for
4898 * page fragments. This is used by the network stack and network device
4899 * drivers to provide a backing region of memory for use as either an
4900 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4902 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4905 struct page *page = NULL;
4906 gfp_t gfp = gfp_mask;
4908 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4909 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4911 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4912 PAGE_FRAG_CACHE_MAX_ORDER);
4913 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4915 if (unlikely(!page))
4916 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4918 nc->va = page ? page_address(page) : NULL;
4923 void __page_frag_cache_drain(struct page *page, unsigned int count)
4925 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4927 if (page_ref_sub_and_test(page, count))
4928 free_the_page(page, compound_order(page));
4930 EXPORT_SYMBOL(__page_frag_cache_drain);
4932 void *page_frag_alloc(struct page_frag_cache *nc,
4933 unsigned int fragsz, gfp_t gfp_mask)
4935 unsigned int size = PAGE_SIZE;
4939 if (unlikely(!nc->va)) {
4941 page = __page_frag_cache_refill(nc, gfp_mask);
4945 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4946 /* if size can vary use size else just use PAGE_SIZE */
4949 /* Even if we own the page, we do not use atomic_set().
4950 * This would break get_page_unless_zero() users.
4952 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4954 /* reset page count bias and offset to start of new frag */
4955 nc->pfmemalloc = page_is_pfmemalloc(page);
4956 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4960 offset = nc->offset - fragsz;
4961 if (unlikely(offset < 0)) {
4962 page = virt_to_page(nc->va);
4964 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4967 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4968 /* if size can vary use size else just use PAGE_SIZE */
4971 /* OK, page count is 0, we can safely set it */
4972 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4974 /* reset page count bias and offset to start of new frag */
4975 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4976 offset = size - fragsz;
4980 nc->offset = offset;
4982 return nc->va + offset;
4984 EXPORT_SYMBOL(page_frag_alloc);
4987 * Frees a page fragment allocated out of either a compound or order 0 page.
4989 void page_frag_free(void *addr)
4991 struct page *page = virt_to_head_page(addr);
4993 if (unlikely(put_page_testzero(page)))
4994 free_the_page(page, compound_order(page));
4996 EXPORT_SYMBOL(page_frag_free);
4998 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5002 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5003 unsigned long used = addr + PAGE_ALIGN(size);
5005 split_page(virt_to_page((void *)addr), order);
5006 while (used < alloc_end) {
5011 return (void *)addr;
5015 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5016 * @size: the number of bytes to allocate
5017 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5019 * This function is similar to alloc_pages(), except that it allocates the
5020 * minimum number of pages to satisfy the request. alloc_pages() can only
5021 * allocate memory in power-of-two pages.
5023 * This function is also limited by MAX_ORDER.
5025 * Memory allocated by this function must be released by free_pages_exact().
5027 * Return: pointer to the allocated area or %NULL in case of error.
5029 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5031 unsigned int order = get_order(size);
5034 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5035 gfp_mask &= ~__GFP_COMP;
5037 addr = __get_free_pages(gfp_mask, order);
5038 return make_alloc_exact(addr, order, size);
5040 EXPORT_SYMBOL(alloc_pages_exact);
5043 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5045 * @nid: the preferred node ID where memory should be allocated
5046 * @size: the number of bytes to allocate
5047 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5049 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5052 * Return: pointer to the allocated area or %NULL in case of error.
5054 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5056 unsigned int order = get_order(size);
5059 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5060 gfp_mask &= ~__GFP_COMP;
5062 p = alloc_pages_node(nid, gfp_mask, order);
5065 return make_alloc_exact((unsigned long)page_address(p), order, size);
5069 * free_pages_exact - release memory allocated via alloc_pages_exact()
5070 * @virt: the value returned by alloc_pages_exact.
5071 * @size: size of allocation, same value as passed to alloc_pages_exact().
5073 * Release the memory allocated by a previous call to alloc_pages_exact.
5075 void free_pages_exact(void *virt, size_t size)
5077 unsigned long addr = (unsigned long)virt;
5078 unsigned long end = addr + PAGE_ALIGN(size);
5080 while (addr < end) {
5085 EXPORT_SYMBOL(free_pages_exact);
5088 * nr_free_zone_pages - count number of pages beyond high watermark
5089 * @offset: The zone index of the highest zone
5091 * nr_free_zone_pages() counts the number of pages which are beyond the
5092 * high watermark within all zones at or below a given zone index. For each
5093 * zone, the number of pages is calculated as:
5095 * nr_free_zone_pages = managed_pages - high_pages
5097 * Return: number of pages beyond high watermark.
5099 static unsigned long nr_free_zone_pages(int offset)
5104 /* Just pick one node, since fallback list is circular */
5105 unsigned long sum = 0;
5107 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5109 for_each_zone_zonelist(zone, z, zonelist, offset) {
5110 unsigned long size = zone_managed_pages(zone);
5111 unsigned long high = high_wmark_pages(zone);
5120 * nr_free_buffer_pages - count number of pages beyond high watermark
5122 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5123 * watermark within ZONE_DMA and ZONE_NORMAL.
5125 * Return: number of pages beyond high watermark within ZONE_DMA and
5128 unsigned long nr_free_buffer_pages(void)
5130 return nr_free_zone_pages(gfp_zone(GFP_USER));
5132 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5135 * nr_free_pagecache_pages - count number of pages beyond high watermark
5137 * nr_free_pagecache_pages() counts the number of pages which are beyond the
5138 * high watermark within all zones.
5140 * Return: number of pages beyond high watermark within all zones.
5142 unsigned long nr_free_pagecache_pages(void)
5144 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5147 static inline void show_node(struct zone *zone)
5149 if (IS_ENABLED(CONFIG_NUMA))
5150 printk("Node %d ", zone_to_nid(zone));
5153 long si_mem_available(void)
5156 unsigned long pagecache;
5157 unsigned long wmark_low = 0;
5158 unsigned long pages[NR_LRU_LISTS];
5159 unsigned long reclaimable;
5163 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5164 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5167 wmark_low += low_wmark_pages(zone);
5170 * Estimate the amount of memory available for userspace allocations,
5171 * without causing swapping.
5173 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5176 * Not all the page cache can be freed, otherwise the system will
5177 * start swapping. Assume at least half of the page cache, or the
5178 * low watermark worth of cache, needs to stay.
5180 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5181 pagecache -= min(pagecache / 2, wmark_low);
5182 available += pagecache;
5185 * Part of the reclaimable slab and other kernel memory consists of
5186 * items that are in use, and cannot be freed. Cap this estimate at the
5189 reclaimable = global_node_page_state(NR_SLAB_RECLAIMABLE) +
5190 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5191 available += reclaimable - min(reclaimable / 2, wmark_low);
5197 EXPORT_SYMBOL_GPL(si_mem_available);
5199 void si_meminfo(struct sysinfo *val)
5201 val->totalram = totalram_pages();
5202 val->sharedram = global_node_page_state(NR_SHMEM);
5203 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5204 val->bufferram = nr_blockdev_pages();
5205 val->totalhigh = totalhigh_pages();
5206 val->freehigh = nr_free_highpages();
5207 val->mem_unit = PAGE_SIZE;
5210 EXPORT_SYMBOL(si_meminfo);
5213 void si_meminfo_node(struct sysinfo *val, int nid)
5215 int zone_type; /* needs to be signed */
5216 unsigned long managed_pages = 0;
5217 unsigned long managed_highpages = 0;
5218 unsigned long free_highpages = 0;
5219 pg_data_t *pgdat = NODE_DATA(nid);
5221 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5222 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5223 val->totalram = managed_pages;
5224 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5225 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5226 #ifdef CONFIG_HIGHMEM
5227 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5228 struct zone *zone = &pgdat->node_zones[zone_type];
5230 if (is_highmem(zone)) {
5231 managed_highpages += zone_managed_pages(zone);
5232 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5235 val->totalhigh = managed_highpages;
5236 val->freehigh = free_highpages;
5238 val->totalhigh = managed_highpages;
5239 val->freehigh = free_highpages;
5241 val->mem_unit = PAGE_SIZE;
5246 * Determine whether the node should be displayed or not, depending on whether
5247 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5249 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5251 if (!(flags & SHOW_MEM_FILTER_NODES))
5255 * no node mask - aka implicit memory numa policy. Do not bother with
5256 * the synchronization - read_mems_allowed_begin - because we do not
5257 * have to be precise here.
5260 nodemask = &cpuset_current_mems_allowed;
5262 return !node_isset(nid, *nodemask);
5265 #define K(x) ((x) << (PAGE_SHIFT-10))
5267 static void show_migration_types(unsigned char type)
5269 static const char types[MIGRATE_TYPES] = {
5270 [MIGRATE_UNMOVABLE] = 'U',
5271 [MIGRATE_MOVABLE] = 'M',
5272 [MIGRATE_RECLAIMABLE] = 'E',
5273 [MIGRATE_HIGHATOMIC] = 'H',
5275 [MIGRATE_CMA] = 'C',
5277 #ifdef CONFIG_MEMORY_ISOLATION
5278 [MIGRATE_ISOLATE] = 'I',
5281 char tmp[MIGRATE_TYPES + 1];
5285 for (i = 0; i < MIGRATE_TYPES; i++) {
5286 if (type & (1 << i))
5291 printk(KERN_CONT "(%s) ", tmp);
5295 * Show free area list (used inside shift_scroll-lock stuff)
5296 * We also calculate the percentage fragmentation. We do this by counting the
5297 * memory on each free list with the exception of the first item on the list.
5300 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5303 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5305 unsigned long free_pcp = 0;
5310 for_each_populated_zone(zone) {
5311 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5314 for_each_online_cpu(cpu)
5315 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5318 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5319 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5320 " unevictable:%lu dirty:%lu writeback:%lu\n"
5321 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5322 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5323 " free:%lu free_pcp:%lu free_cma:%lu\n",
5324 global_node_page_state(NR_ACTIVE_ANON),
5325 global_node_page_state(NR_INACTIVE_ANON),
5326 global_node_page_state(NR_ISOLATED_ANON),
5327 global_node_page_state(NR_ACTIVE_FILE),
5328 global_node_page_state(NR_INACTIVE_FILE),
5329 global_node_page_state(NR_ISOLATED_FILE),
5330 global_node_page_state(NR_UNEVICTABLE),
5331 global_node_page_state(NR_FILE_DIRTY),
5332 global_node_page_state(NR_WRITEBACK),
5333 global_node_page_state(NR_SLAB_RECLAIMABLE),
5334 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
5335 global_node_page_state(NR_FILE_MAPPED),
5336 global_node_page_state(NR_SHMEM),
5337 global_zone_page_state(NR_PAGETABLE),
5338 global_zone_page_state(NR_BOUNCE),
5339 global_zone_page_state(NR_FREE_PAGES),
5341 global_zone_page_state(NR_FREE_CMA_PAGES));
5343 for_each_online_pgdat(pgdat) {
5344 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5348 " active_anon:%lukB"
5349 " inactive_anon:%lukB"
5350 " active_file:%lukB"
5351 " inactive_file:%lukB"
5352 " unevictable:%lukB"
5353 " isolated(anon):%lukB"
5354 " isolated(file):%lukB"
5359 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5361 " shmem_pmdmapped: %lukB"
5364 " writeback_tmp:%lukB"
5365 " all_unreclaimable? %s"
5368 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5369 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5370 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5371 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5372 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5373 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5374 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5375 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5376 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5377 K(node_page_state(pgdat, NR_WRITEBACK)),
5378 K(node_page_state(pgdat, NR_SHMEM)),
5379 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5380 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5381 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5383 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5385 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5386 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5390 for_each_populated_zone(zone) {
5393 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5397 for_each_online_cpu(cpu)
5398 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5407 " reserved_highatomic:%luKB"
5408 " active_anon:%lukB"
5409 " inactive_anon:%lukB"
5410 " active_file:%lukB"
5411 " inactive_file:%lukB"
5412 " unevictable:%lukB"
5413 " writepending:%lukB"
5417 " kernel_stack:%lukB"
5418 #ifdef CONFIG_SHADOW_CALL_STACK
5419 " shadow_call_stack:%lukB"
5428 K(zone_page_state(zone, NR_FREE_PAGES)),
5429 K(min_wmark_pages(zone)),
5430 K(low_wmark_pages(zone)),
5431 K(high_wmark_pages(zone)),
5432 K(zone->nr_reserved_highatomic),
5433 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5434 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5435 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5436 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5437 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5438 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5439 K(zone->present_pages),
5440 K(zone_managed_pages(zone)),
5441 K(zone_page_state(zone, NR_MLOCK)),
5442 zone_page_state(zone, NR_KERNEL_STACK_KB),
5443 #ifdef CONFIG_SHADOW_CALL_STACK
5444 zone_page_state(zone, NR_KERNEL_SCS_KB),
5446 K(zone_page_state(zone, NR_PAGETABLE)),
5447 K(zone_page_state(zone, NR_BOUNCE)),
5449 K(this_cpu_read(zone->pageset->pcp.count)),
5450 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5451 printk("lowmem_reserve[]:");
5452 for (i = 0; i < MAX_NR_ZONES; i++)
5453 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5454 printk(KERN_CONT "\n");
5457 for_each_populated_zone(zone) {
5459 unsigned long nr[MAX_ORDER], flags, total = 0;
5460 unsigned char types[MAX_ORDER];
5462 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5465 printk(KERN_CONT "%s: ", zone->name);
5467 spin_lock_irqsave(&zone->lock, flags);
5468 for (order = 0; order < MAX_ORDER; order++) {
5469 struct free_area *area = &zone->free_area[order];
5472 nr[order] = area->nr_free;
5473 total += nr[order] << order;
5476 for (type = 0; type < MIGRATE_TYPES; type++) {
5477 if (!free_area_empty(area, type))
5478 types[order] |= 1 << type;
5481 spin_unlock_irqrestore(&zone->lock, flags);
5482 for (order = 0; order < MAX_ORDER; order++) {
5483 printk(KERN_CONT "%lu*%lukB ",
5484 nr[order], K(1UL) << order);
5486 show_migration_types(types[order]);
5488 printk(KERN_CONT "= %lukB\n", K(total));
5491 hugetlb_show_meminfo();
5493 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5495 show_swap_cache_info();
5498 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5500 zoneref->zone = zone;
5501 zoneref->zone_idx = zone_idx(zone);
5505 * Builds allocation fallback zone lists.
5507 * Add all populated zones of a node to the zonelist.
5509 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5512 enum zone_type zone_type = MAX_NR_ZONES;
5517 zone = pgdat->node_zones + zone_type;
5518 if (managed_zone(zone)) {
5519 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5520 check_highest_zone(zone_type);
5522 } while (zone_type);
5529 static int __parse_numa_zonelist_order(char *s)
5532 * We used to support different zonlists modes but they turned
5533 * out to be just not useful. Let's keep the warning in place
5534 * if somebody still use the cmd line parameter so that we do
5535 * not fail it silently
5537 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5538 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5544 static __init int setup_numa_zonelist_order(char *s)
5549 return __parse_numa_zonelist_order(s);
5551 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5553 char numa_zonelist_order[] = "Node";
5556 * sysctl handler for numa_zonelist_order
5558 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5559 void __user *buffer, size_t *length,
5566 return proc_dostring(table, write, buffer, length, ppos);
5567 str = memdup_user_nul(buffer, 16);
5569 return PTR_ERR(str);
5571 ret = __parse_numa_zonelist_order(str);
5577 #define MAX_NODE_LOAD (nr_online_nodes)
5578 static int node_load[MAX_NUMNODES];
5581 * find_next_best_node - find the next node that should appear in a given node's fallback list
5582 * @node: node whose fallback list we're appending
5583 * @used_node_mask: nodemask_t of already used nodes
5585 * We use a number of factors to determine which is the next node that should
5586 * appear on a given node's fallback list. The node should not have appeared
5587 * already in @node's fallback list, and it should be the next closest node
5588 * according to the distance array (which contains arbitrary distance values
5589 * from each node to each node in the system), and should also prefer nodes
5590 * with no CPUs, since presumably they'll have very little allocation pressure
5591 * on them otherwise.
5593 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5595 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5598 int min_val = INT_MAX;
5599 int best_node = NUMA_NO_NODE;
5600 const struct cpumask *tmp = cpumask_of_node(0);
5602 /* Use the local node if we haven't already */
5603 if (!node_isset(node, *used_node_mask)) {
5604 node_set(node, *used_node_mask);
5608 for_each_node_state(n, N_MEMORY) {
5610 /* Don't want a node to appear more than once */
5611 if (node_isset(n, *used_node_mask))
5614 /* Use the distance array to find the distance */
5615 val = node_distance(node, n);
5617 /* Penalize nodes under us ("prefer the next node") */
5620 /* Give preference to headless and unused nodes */
5621 tmp = cpumask_of_node(n);
5622 if (!cpumask_empty(tmp))
5623 val += PENALTY_FOR_NODE_WITH_CPUS;
5625 /* Slight preference for less loaded node */
5626 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5627 val += node_load[n];
5629 if (val < min_val) {
5636 node_set(best_node, *used_node_mask);
5643 * Build zonelists ordered by node and zones within node.
5644 * This results in maximum locality--normal zone overflows into local
5645 * DMA zone, if any--but risks exhausting DMA zone.
5647 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5650 struct zoneref *zonerefs;
5653 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5655 for (i = 0; i < nr_nodes; i++) {
5658 pg_data_t *node = NODE_DATA(node_order[i]);
5660 nr_zones = build_zonerefs_node(node, zonerefs);
5661 zonerefs += nr_zones;
5663 zonerefs->zone = NULL;
5664 zonerefs->zone_idx = 0;
5668 * Build gfp_thisnode zonelists
5670 static void build_thisnode_zonelists(pg_data_t *pgdat)
5672 struct zoneref *zonerefs;
5675 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5676 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5677 zonerefs += nr_zones;
5678 zonerefs->zone = NULL;
5679 zonerefs->zone_idx = 0;
5683 * Build zonelists ordered by zone and nodes within zones.
5684 * This results in conserving DMA zone[s] until all Normal memory is
5685 * exhausted, but results in overflowing to remote node while memory
5686 * may still exist in local DMA zone.
5689 static void build_zonelists(pg_data_t *pgdat)
5691 static int node_order[MAX_NUMNODES];
5692 int node, load, nr_nodes = 0;
5693 nodemask_t used_mask;
5694 int local_node, prev_node;
5696 /* NUMA-aware ordering of nodes */
5697 local_node = pgdat->node_id;
5698 load = nr_online_nodes;
5699 prev_node = local_node;
5700 nodes_clear(used_mask);
5702 memset(node_order, 0, sizeof(node_order));
5703 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5705 * We don't want to pressure a particular node.
5706 * So adding penalty to the first node in same
5707 * distance group to make it round-robin.
5709 if (node_distance(local_node, node) !=
5710 node_distance(local_node, prev_node))
5711 node_load[node] = load;
5713 node_order[nr_nodes++] = node;
5718 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5719 build_thisnode_zonelists(pgdat);
5722 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5724 * Return node id of node used for "local" allocations.
5725 * I.e., first node id of first zone in arg node's generic zonelist.
5726 * Used for initializing percpu 'numa_mem', which is used primarily
5727 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5729 int local_memory_node(int node)
5733 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5734 gfp_zone(GFP_KERNEL),
5736 return zone_to_nid(z->zone);
5740 static void setup_min_unmapped_ratio(void);
5741 static void setup_min_slab_ratio(void);
5742 #else /* CONFIG_NUMA */
5744 static void build_zonelists(pg_data_t *pgdat)
5746 int node, local_node;
5747 struct zoneref *zonerefs;
5750 local_node = pgdat->node_id;
5752 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5753 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5754 zonerefs += nr_zones;
5757 * Now we build the zonelist so that it contains the zones
5758 * of all the other nodes.
5759 * We don't want to pressure a particular node, so when
5760 * building the zones for node N, we make sure that the
5761 * zones coming right after the local ones are those from
5762 * node N+1 (modulo N)
5764 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5765 if (!node_online(node))
5767 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5768 zonerefs += nr_zones;
5770 for (node = 0; node < local_node; node++) {
5771 if (!node_online(node))
5773 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5774 zonerefs += nr_zones;
5777 zonerefs->zone = NULL;
5778 zonerefs->zone_idx = 0;
5781 #endif /* CONFIG_NUMA */
5784 * Boot pageset table. One per cpu which is going to be used for all
5785 * zones and all nodes. The parameters will be set in such a way
5786 * that an item put on a list will immediately be handed over to
5787 * the buddy list. This is safe since pageset manipulation is done
5788 * with interrupts disabled.
5790 * The boot_pagesets must be kept even after bootup is complete for
5791 * unused processors and/or zones. They do play a role for bootstrapping
5792 * hotplugged processors.
5794 * zoneinfo_show() and maybe other functions do
5795 * not check if the processor is online before following the pageset pointer.
5796 * Other parts of the kernel may not check if the zone is available.
5798 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5799 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5800 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5802 static void __build_all_zonelists(void *data)
5805 int __maybe_unused cpu;
5806 pg_data_t *self = data;
5807 static DEFINE_SPINLOCK(lock);
5812 memset(node_load, 0, sizeof(node_load));
5816 * This node is hotadded and no memory is yet present. So just
5817 * building zonelists is fine - no need to touch other nodes.
5819 if (self && !node_online(self->node_id)) {
5820 build_zonelists(self);
5822 for_each_online_node(nid) {
5823 pg_data_t *pgdat = NODE_DATA(nid);
5825 build_zonelists(pgdat);
5828 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5830 * We now know the "local memory node" for each node--
5831 * i.e., the node of the first zone in the generic zonelist.
5832 * Set up numa_mem percpu variable for on-line cpus. During
5833 * boot, only the boot cpu should be on-line; we'll init the
5834 * secondary cpus' numa_mem as they come on-line. During
5835 * node/memory hotplug, we'll fixup all on-line cpus.
5837 for_each_online_cpu(cpu)
5838 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5845 static noinline void __init
5846 build_all_zonelists_init(void)
5850 __build_all_zonelists(NULL);
5853 * Initialize the boot_pagesets that are going to be used
5854 * for bootstrapping processors. The real pagesets for
5855 * each zone will be allocated later when the per cpu
5856 * allocator is available.
5858 * boot_pagesets are used also for bootstrapping offline
5859 * cpus if the system is already booted because the pagesets
5860 * are needed to initialize allocators on a specific cpu too.
5861 * F.e. the percpu allocator needs the page allocator which
5862 * needs the percpu allocator in order to allocate its pagesets
5863 * (a chicken-egg dilemma).
5865 for_each_possible_cpu(cpu)
5866 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5868 mminit_verify_zonelist();
5869 cpuset_init_current_mems_allowed();
5873 * unless system_state == SYSTEM_BOOTING.
5875 * __ref due to call of __init annotated helper build_all_zonelists_init
5876 * [protected by SYSTEM_BOOTING].
5878 void __ref build_all_zonelists(pg_data_t *pgdat)
5880 if (system_state == SYSTEM_BOOTING) {
5881 build_all_zonelists_init();
5883 __build_all_zonelists(pgdat);
5884 /* cpuset refresh routine should be here */
5886 vm_total_pages = nr_free_pagecache_pages();
5888 * Disable grouping by mobility if the number of pages in the
5889 * system is too low to allow the mechanism to work. It would be
5890 * more accurate, but expensive to check per-zone. This check is
5891 * made on memory-hotadd so a system can start with mobility
5892 * disabled and enable it later
5894 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5895 page_group_by_mobility_disabled = 1;
5897 page_group_by_mobility_disabled = 0;
5899 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5901 page_group_by_mobility_disabled ? "off" : "on",
5904 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5908 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5909 static bool __meminit
5910 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5912 static struct memblock_region *r;
5914 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5915 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5916 for_each_memblock(memory, r) {
5917 if (*pfn < memblock_region_memory_end_pfn(r))
5921 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5922 memblock_is_mirror(r)) {
5923 *pfn = memblock_region_memory_end_pfn(r);
5931 * Initially all pages are reserved - free ones are freed
5932 * up by memblock_free_all() once the early boot process is
5933 * done. Non-atomic initialization, single-pass.
5935 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5936 unsigned long start_pfn, enum memmap_context context,
5937 struct vmem_altmap *altmap)
5939 unsigned long pfn, end_pfn = start_pfn + size;
5942 if (highest_memmap_pfn < end_pfn - 1)
5943 highest_memmap_pfn = end_pfn - 1;
5945 #ifdef CONFIG_ZONE_DEVICE
5947 * Honor reservation requested by the driver for this ZONE_DEVICE
5948 * memory. We limit the total number of pages to initialize to just
5949 * those that might contain the memory mapping. We will defer the
5950 * ZONE_DEVICE page initialization until after we have released
5953 if (zone == ZONE_DEVICE) {
5957 if (start_pfn == altmap->base_pfn)
5958 start_pfn += altmap->reserve;
5959 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5963 for (pfn = start_pfn; pfn < end_pfn; ) {
5965 * There can be holes in boot-time mem_map[]s handed to this
5966 * function. They do not exist on hotplugged memory.
5968 if (context == MEMMAP_EARLY) {
5969 if (overlap_memmap_init(zone, &pfn))
5971 if (defer_init(nid, pfn, end_pfn))
5975 page = pfn_to_page(pfn);
5976 __init_single_page(page, pfn, zone, nid);
5977 if (context == MEMMAP_HOTPLUG)
5978 __SetPageReserved(page);
5981 * Mark the block movable so that blocks are reserved for
5982 * movable at startup. This will force kernel allocations
5983 * to reserve their blocks rather than leaking throughout
5984 * the address space during boot when many long-lived
5985 * kernel allocations are made.
5987 * bitmap is created for zone's valid pfn range. but memmap
5988 * can be created for invalid pages (for alignment)
5989 * check here not to call set_pageblock_migratetype() against
5992 if (!(pfn & (pageblock_nr_pages - 1))) {
5993 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6000 #ifdef CONFIG_ZONE_DEVICE
6001 void __ref memmap_init_zone_device(struct zone *zone,
6002 unsigned long start_pfn,
6003 unsigned long nr_pages,
6004 struct dev_pagemap *pgmap)
6006 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6007 struct pglist_data *pgdat = zone->zone_pgdat;
6008 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6009 unsigned long zone_idx = zone_idx(zone);
6010 unsigned long start = jiffies;
6011 int nid = pgdat->node_id;
6013 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6017 * The call to memmap_init_zone should have already taken care
6018 * of the pages reserved for the memmap, so we can just jump to
6019 * the end of that region and start processing the device pages.
6022 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6023 nr_pages = end_pfn - start_pfn;
6026 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6027 struct page *page = pfn_to_page(pfn);
6029 __init_single_page(page, pfn, zone_idx, nid);
6032 * Mark page reserved as it will need to wait for onlining
6033 * phase for it to be fully associated with a zone.
6035 * We can use the non-atomic __set_bit operation for setting
6036 * the flag as we are still initializing the pages.
6038 __SetPageReserved(page);
6041 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6042 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6043 * ever freed or placed on a driver-private list.
6045 page->pgmap = pgmap;
6046 page->zone_device_data = NULL;
6049 * Mark the block movable so that blocks are reserved for
6050 * movable at startup. This will force kernel allocations
6051 * to reserve their blocks rather than leaking throughout
6052 * the address space during boot when many long-lived
6053 * kernel allocations are made.
6055 * bitmap is created for zone's valid pfn range. but memmap
6056 * can be created for invalid pages (for alignment)
6057 * check here not to call set_pageblock_migratetype() against
6060 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
6061 * because this is done early in section_activate()
6063 if (!(pfn & (pageblock_nr_pages - 1))) {
6064 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6069 pr_info("%s initialised %lu pages in %ums\n", __func__,
6070 nr_pages, jiffies_to_msecs(jiffies - start));
6074 static void __meminit zone_init_free_lists(struct zone *zone)
6076 unsigned int order, t;
6077 for_each_migratetype_order(order, t) {
6078 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6079 zone->free_area[order].nr_free = 0;
6083 void __meminit __weak memmap_init(unsigned long size, int nid,
6085 unsigned long range_start_pfn)
6087 unsigned long start_pfn, end_pfn;
6088 unsigned long range_end_pfn = range_start_pfn + size;
6091 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6092 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6093 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6095 if (end_pfn > start_pfn) {
6096 size = end_pfn - start_pfn;
6097 memmap_init_zone(size, nid, zone, start_pfn,
6098 MEMMAP_EARLY, NULL);
6103 static int zone_batchsize(struct zone *zone)
6109 * The per-cpu-pages pools are set to around 1000th of the
6112 batch = zone_managed_pages(zone) / 1024;
6113 /* But no more than a meg. */
6114 if (batch * PAGE_SIZE > 1024 * 1024)
6115 batch = (1024 * 1024) / PAGE_SIZE;
6116 batch /= 4; /* We effectively *= 4 below */
6121 * Clamp the batch to a 2^n - 1 value. Having a power
6122 * of 2 value was found to be more likely to have
6123 * suboptimal cache aliasing properties in some cases.
6125 * For example if 2 tasks are alternately allocating
6126 * batches of pages, one task can end up with a lot
6127 * of pages of one half of the possible page colors
6128 * and the other with pages of the other colors.
6130 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6135 /* The deferral and batching of frees should be suppressed under NOMMU
6138 * The problem is that NOMMU needs to be able to allocate large chunks
6139 * of contiguous memory as there's no hardware page translation to
6140 * assemble apparent contiguous memory from discontiguous pages.
6142 * Queueing large contiguous runs of pages for batching, however,
6143 * causes the pages to actually be freed in smaller chunks. As there
6144 * can be a significant delay between the individual batches being
6145 * recycled, this leads to the once large chunks of space being
6146 * fragmented and becoming unavailable for high-order allocations.
6153 * pcp->high and pcp->batch values are related and dependent on one another:
6154 * ->batch must never be higher then ->high.
6155 * The following function updates them in a safe manner without read side
6158 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6159 * those fields changing asynchronously (acording the the above rule).
6161 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6162 * outside of boot time (or some other assurance that no concurrent updaters
6165 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6166 unsigned long batch)
6168 /* start with a fail safe value for batch */
6172 /* Update high, then batch, in order */
6179 /* a companion to pageset_set_high() */
6180 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6182 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6185 static void pageset_init(struct per_cpu_pageset *p)
6187 struct per_cpu_pages *pcp;
6190 memset(p, 0, sizeof(*p));
6193 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6194 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6197 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6200 pageset_set_batch(p, batch);
6204 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6205 * to the value high for the pageset p.
6207 static void pageset_set_high(struct per_cpu_pageset *p,
6210 unsigned long batch = max(1UL, high / 4);
6211 if ((high / 4) > (PAGE_SHIFT * 8))
6212 batch = PAGE_SHIFT * 8;
6214 pageset_update(&p->pcp, high, batch);
6217 static void pageset_set_high_and_batch(struct zone *zone,
6218 struct per_cpu_pageset *pcp)
6220 if (percpu_pagelist_fraction)
6221 pageset_set_high(pcp,
6222 (zone_managed_pages(zone) /
6223 percpu_pagelist_fraction));
6225 pageset_set_batch(pcp, zone_batchsize(zone));
6228 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6230 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6233 pageset_set_high_and_batch(zone, pcp);
6236 void __meminit setup_zone_pageset(struct zone *zone)
6239 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6240 for_each_possible_cpu(cpu)
6241 zone_pageset_init(zone, cpu);
6245 * Allocate per cpu pagesets and initialize them.
6246 * Before this call only boot pagesets were available.
6248 void __init setup_per_cpu_pageset(void)
6250 struct pglist_data *pgdat;
6253 for_each_populated_zone(zone)
6254 setup_zone_pageset(zone);
6256 for_each_online_pgdat(pgdat)
6257 pgdat->per_cpu_nodestats =
6258 alloc_percpu(struct per_cpu_nodestat);
6261 static __meminit void zone_pcp_init(struct zone *zone)
6264 * per cpu subsystem is not up at this point. The following code
6265 * relies on the ability of the linker to provide the
6266 * offset of a (static) per cpu variable into the per cpu area.
6268 zone->pageset = &boot_pageset;
6270 if (populated_zone(zone))
6271 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6272 zone->name, zone->present_pages,
6273 zone_batchsize(zone));
6276 void __meminit init_currently_empty_zone(struct zone *zone,
6277 unsigned long zone_start_pfn,
6280 struct pglist_data *pgdat = zone->zone_pgdat;
6281 int zone_idx = zone_idx(zone) + 1;
6283 if (zone_idx > pgdat->nr_zones)
6284 pgdat->nr_zones = zone_idx;
6286 zone->zone_start_pfn = zone_start_pfn;
6288 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6289 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6291 (unsigned long)zone_idx(zone),
6292 zone_start_pfn, (zone_start_pfn + size));
6294 zone_init_free_lists(zone);
6295 zone->initialized = 1;
6299 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6300 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6302 * If an architecture guarantees that all ranges registered contain no holes and may
6303 * be freed, this function may be used instead of calling memory_present() manually.
6305 void __init sparse_memory_present_with_active_regions(int nid)
6307 unsigned long start_pfn, end_pfn;
6310 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
6311 memory_present(this_nid, start_pfn, end_pfn);
6315 * get_pfn_range_for_nid - Return the start and end page frames for a node
6316 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6317 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6318 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6320 * It returns the start and end page frame of a node based on information
6321 * provided by memblock_set_node(). If called for a node
6322 * with no available memory, a warning is printed and the start and end
6325 void __init get_pfn_range_for_nid(unsigned int nid,
6326 unsigned long *start_pfn, unsigned long *end_pfn)
6328 unsigned long this_start_pfn, this_end_pfn;
6334 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6335 *start_pfn = min(*start_pfn, this_start_pfn);
6336 *end_pfn = max(*end_pfn, this_end_pfn);
6339 if (*start_pfn == -1UL)
6344 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6345 * assumption is made that zones within a node are ordered in monotonic
6346 * increasing memory addresses so that the "highest" populated zone is used
6348 static void __init find_usable_zone_for_movable(void)
6351 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6352 if (zone_index == ZONE_MOVABLE)
6355 if (arch_zone_highest_possible_pfn[zone_index] >
6356 arch_zone_lowest_possible_pfn[zone_index])
6360 VM_BUG_ON(zone_index == -1);
6361 movable_zone = zone_index;
6365 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6366 * because it is sized independent of architecture. Unlike the other zones,
6367 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6368 * in each node depending on the size of each node and how evenly kernelcore
6369 * is distributed. This helper function adjusts the zone ranges
6370 * provided by the architecture for a given node by using the end of the
6371 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6372 * zones within a node are in order of monotonic increases memory addresses
6374 static void __init adjust_zone_range_for_zone_movable(int nid,
6375 unsigned long zone_type,
6376 unsigned long node_start_pfn,
6377 unsigned long node_end_pfn,
6378 unsigned long *zone_start_pfn,
6379 unsigned long *zone_end_pfn)
6381 /* Only adjust if ZONE_MOVABLE is on this node */
6382 if (zone_movable_pfn[nid]) {
6383 /* Size ZONE_MOVABLE */
6384 if (zone_type == ZONE_MOVABLE) {
6385 *zone_start_pfn = zone_movable_pfn[nid];
6386 *zone_end_pfn = min(node_end_pfn,
6387 arch_zone_highest_possible_pfn[movable_zone]);
6389 /* Adjust for ZONE_MOVABLE starting within this range */
6390 } else if (!mirrored_kernelcore &&
6391 *zone_start_pfn < zone_movable_pfn[nid] &&
6392 *zone_end_pfn > zone_movable_pfn[nid]) {
6393 *zone_end_pfn = zone_movable_pfn[nid];
6395 /* Check if this whole range is within ZONE_MOVABLE */
6396 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6397 *zone_start_pfn = *zone_end_pfn;
6402 * Return the number of pages a zone spans in a node, including holes
6403 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6405 static unsigned long __init zone_spanned_pages_in_node(int nid,
6406 unsigned long zone_type,
6407 unsigned long node_start_pfn,
6408 unsigned long node_end_pfn,
6409 unsigned long *zone_start_pfn,
6410 unsigned long *zone_end_pfn)
6412 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6413 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6414 /* When hotadd a new node from cpu_up(), the node should be empty */
6415 if (!node_start_pfn && !node_end_pfn)
6418 /* Get the start and end of the zone */
6419 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6420 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6421 adjust_zone_range_for_zone_movable(nid, zone_type,
6422 node_start_pfn, node_end_pfn,
6423 zone_start_pfn, zone_end_pfn);
6425 /* Check that this node has pages within the zone's required range */
6426 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6429 /* Move the zone boundaries inside the node if necessary */
6430 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6431 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6433 /* Return the spanned pages */
6434 return *zone_end_pfn - *zone_start_pfn;
6438 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6439 * then all holes in the requested range will be accounted for.
6441 unsigned long __init __absent_pages_in_range(int nid,
6442 unsigned long range_start_pfn,
6443 unsigned long range_end_pfn)
6445 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6446 unsigned long start_pfn, end_pfn;
6449 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6450 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6451 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6452 nr_absent -= end_pfn - start_pfn;
6458 * absent_pages_in_range - Return number of page frames in holes within a range
6459 * @start_pfn: The start PFN to start searching for holes
6460 * @end_pfn: The end PFN to stop searching for holes
6462 * Return: the number of pages frames in memory holes within a range.
6464 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6465 unsigned long end_pfn)
6467 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6470 /* Return the number of page frames in holes in a zone on a node */
6471 static unsigned long __init zone_absent_pages_in_node(int nid,
6472 unsigned long zone_type,
6473 unsigned long node_start_pfn,
6474 unsigned long node_end_pfn)
6476 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6477 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6478 unsigned long zone_start_pfn, zone_end_pfn;
6479 unsigned long nr_absent;
6481 /* When hotadd a new node from cpu_up(), the node should be empty */
6482 if (!node_start_pfn && !node_end_pfn)
6485 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6486 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6488 adjust_zone_range_for_zone_movable(nid, zone_type,
6489 node_start_pfn, node_end_pfn,
6490 &zone_start_pfn, &zone_end_pfn);
6491 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6494 * ZONE_MOVABLE handling.
6495 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6498 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6499 unsigned long start_pfn, end_pfn;
6500 struct memblock_region *r;
6502 for_each_memblock(memory, r) {
6503 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6504 zone_start_pfn, zone_end_pfn);
6505 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6506 zone_start_pfn, zone_end_pfn);
6508 if (zone_type == ZONE_MOVABLE &&
6509 memblock_is_mirror(r))
6510 nr_absent += end_pfn - start_pfn;
6512 if (zone_type == ZONE_NORMAL &&
6513 !memblock_is_mirror(r))
6514 nr_absent += end_pfn - start_pfn;
6521 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6522 unsigned long node_start_pfn,
6523 unsigned long node_end_pfn)
6525 unsigned long realtotalpages = 0, totalpages = 0;
6528 for (i = 0; i < MAX_NR_ZONES; i++) {
6529 struct zone *zone = pgdat->node_zones + i;
6530 unsigned long zone_start_pfn, zone_end_pfn;
6531 unsigned long spanned, absent;
6532 unsigned long size, real_size;
6534 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
6539 absent = zone_absent_pages_in_node(pgdat->node_id, i,
6544 real_size = size - absent;
6547 zone->zone_start_pfn = zone_start_pfn;
6549 zone->zone_start_pfn = 0;
6550 zone->spanned_pages = size;
6551 zone->present_pages = real_size;
6554 realtotalpages += real_size;
6557 pgdat->node_spanned_pages = totalpages;
6558 pgdat->node_present_pages = realtotalpages;
6559 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6563 #ifndef CONFIG_SPARSEMEM
6565 * Calculate the size of the zone->blockflags rounded to an unsigned long
6566 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6567 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6568 * round what is now in bits to nearest long in bits, then return it in
6571 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6573 unsigned long usemapsize;
6575 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6576 usemapsize = roundup(zonesize, pageblock_nr_pages);
6577 usemapsize = usemapsize >> pageblock_order;
6578 usemapsize *= NR_PAGEBLOCK_BITS;
6579 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6581 return usemapsize / 8;
6584 static void __ref setup_usemap(struct pglist_data *pgdat,
6586 unsigned long zone_start_pfn,
6587 unsigned long zonesize)
6589 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6590 zone->pageblock_flags = NULL;
6592 zone->pageblock_flags =
6593 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6595 if (!zone->pageblock_flags)
6596 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6597 usemapsize, zone->name, pgdat->node_id);
6601 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6602 unsigned long zone_start_pfn, unsigned long zonesize) {}
6603 #endif /* CONFIG_SPARSEMEM */
6605 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6607 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6608 void __init set_pageblock_order(void)
6612 /* Check that pageblock_nr_pages has not already been setup */
6613 if (pageblock_order)
6616 if (HPAGE_SHIFT > PAGE_SHIFT)
6617 order = HUGETLB_PAGE_ORDER;
6619 order = MAX_ORDER - 1;
6622 * Assume the largest contiguous order of interest is a huge page.
6623 * This value may be variable depending on boot parameters on IA64 and
6626 pageblock_order = order;
6628 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6631 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6632 * is unused as pageblock_order is set at compile-time. See
6633 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6636 void __init set_pageblock_order(void)
6640 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6642 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6643 unsigned long present_pages)
6645 unsigned long pages = spanned_pages;
6648 * Provide a more accurate estimation if there are holes within
6649 * the zone and SPARSEMEM is in use. If there are holes within the
6650 * zone, each populated memory region may cost us one or two extra
6651 * memmap pages due to alignment because memmap pages for each
6652 * populated regions may not be naturally aligned on page boundary.
6653 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6655 if (spanned_pages > present_pages + (present_pages >> 4) &&
6656 IS_ENABLED(CONFIG_SPARSEMEM))
6657 pages = present_pages;
6659 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6662 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6663 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6665 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6667 spin_lock_init(&ds_queue->split_queue_lock);
6668 INIT_LIST_HEAD(&ds_queue->split_queue);
6669 ds_queue->split_queue_len = 0;
6672 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6675 #ifdef CONFIG_COMPACTION
6676 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6678 init_waitqueue_head(&pgdat->kcompactd_wait);
6681 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6684 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6686 pgdat_resize_init(pgdat);
6688 pgdat_init_split_queue(pgdat);
6689 pgdat_init_kcompactd(pgdat);
6691 init_waitqueue_head(&pgdat->kswapd_wait);
6692 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6694 pgdat_page_ext_init(pgdat);
6695 spin_lock_init(&pgdat->lru_lock);
6696 lruvec_init(&pgdat->__lruvec);
6699 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6700 unsigned long remaining_pages)
6702 atomic_long_set(&zone->managed_pages, remaining_pages);
6703 zone_set_nid(zone, nid);
6704 zone->name = zone_names[idx];
6705 zone->zone_pgdat = NODE_DATA(nid);
6706 spin_lock_init(&zone->lock);
6707 zone_seqlock_init(zone);
6708 zone_pcp_init(zone);
6712 * Set up the zone data structures
6713 * - init pgdat internals
6714 * - init all zones belonging to this node
6716 * NOTE: this function is only called during memory hotplug
6718 #ifdef CONFIG_MEMORY_HOTPLUG
6719 void __ref free_area_init_core_hotplug(int nid)
6722 pg_data_t *pgdat = NODE_DATA(nid);
6724 pgdat_init_internals(pgdat);
6725 for (z = 0; z < MAX_NR_ZONES; z++)
6726 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6731 * Set up the zone data structures:
6732 * - mark all pages reserved
6733 * - mark all memory queues empty
6734 * - clear the memory bitmaps
6736 * NOTE: pgdat should get zeroed by caller.
6737 * NOTE: this function is only called during early init.
6739 static void __init free_area_init_core(struct pglist_data *pgdat)
6742 int nid = pgdat->node_id;
6744 pgdat_init_internals(pgdat);
6745 pgdat->per_cpu_nodestats = &boot_nodestats;
6747 for (j = 0; j < MAX_NR_ZONES; j++) {
6748 struct zone *zone = pgdat->node_zones + j;
6749 unsigned long size, freesize, memmap_pages;
6750 unsigned long zone_start_pfn = zone->zone_start_pfn;
6752 size = zone->spanned_pages;
6753 freesize = zone->present_pages;
6756 * Adjust freesize so that it accounts for how much memory
6757 * is used by this zone for memmap. This affects the watermark
6758 * and per-cpu initialisations
6760 memmap_pages = calc_memmap_size(size, freesize);
6761 if (!is_highmem_idx(j)) {
6762 if (freesize >= memmap_pages) {
6763 freesize -= memmap_pages;
6766 " %s zone: %lu pages used for memmap\n",
6767 zone_names[j], memmap_pages);
6769 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6770 zone_names[j], memmap_pages, freesize);
6773 /* Account for reserved pages */
6774 if (j == 0 && freesize > dma_reserve) {
6775 freesize -= dma_reserve;
6776 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6777 zone_names[0], dma_reserve);
6780 if (!is_highmem_idx(j))
6781 nr_kernel_pages += freesize;
6782 /* Charge for highmem memmap if there are enough kernel pages */
6783 else if (nr_kernel_pages > memmap_pages * 2)
6784 nr_kernel_pages -= memmap_pages;
6785 nr_all_pages += freesize;
6788 * Set an approximate value for lowmem here, it will be adjusted
6789 * when the bootmem allocator frees pages into the buddy system.
6790 * And all highmem pages will be managed by the buddy system.
6792 zone_init_internals(zone, j, nid, freesize);
6797 set_pageblock_order();
6798 setup_usemap(pgdat, zone, zone_start_pfn, size);
6799 init_currently_empty_zone(zone, zone_start_pfn, size);
6800 memmap_init(size, nid, j, zone_start_pfn);
6804 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6805 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6807 unsigned long __maybe_unused start = 0;
6808 unsigned long __maybe_unused offset = 0;
6810 /* Skip empty nodes */
6811 if (!pgdat->node_spanned_pages)
6814 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6815 offset = pgdat->node_start_pfn - start;
6816 /* ia64 gets its own node_mem_map, before this, without bootmem */
6817 if (!pgdat->node_mem_map) {
6818 unsigned long size, end;
6822 * The zone's endpoints aren't required to be MAX_ORDER
6823 * aligned but the node_mem_map endpoints must be in order
6824 * for the buddy allocator to function correctly.
6826 end = pgdat_end_pfn(pgdat);
6827 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6828 size = (end - start) * sizeof(struct page);
6829 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6832 panic("Failed to allocate %ld bytes for node %d memory map\n",
6833 size, pgdat->node_id);
6834 pgdat->node_mem_map = map + offset;
6836 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6837 __func__, pgdat->node_id, (unsigned long)pgdat,
6838 (unsigned long)pgdat->node_mem_map);
6839 #ifndef CONFIG_NEED_MULTIPLE_NODES
6841 * With no DISCONTIG, the global mem_map is just set as node 0's
6843 if (pgdat == NODE_DATA(0)) {
6844 mem_map = NODE_DATA(0)->node_mem_map;
6845 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6851 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6852 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6854 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6855 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6857 pgdat->first_deferred_pfn = ULONG_MAX;
6860 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6863 static void __init free_area_init_node(int nid)
6865 pg_data_t *pgdat = NODE_DATA(nid);
6866 unsigned long start_pfn = 0;
6867 unsigned long end_pfn = 0;
6869 /* pg_data_t should be reset to zero when it's allocated */
6870 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6872 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6874 pgdat->node_id = nid;
6875 pgdat->node_start_pfn = start_pfn;
6876 pgdat->per_cpu_nodestats = NULL;
6878 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6879 (u64)start_pfn << PAGE_SHIFT,
6880 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6881 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
6883 alloc_node_mem_map(pgdat);
6884 pgdat_set_deferred_range(pgdat);
6886 free_area_init_core(pgdat);
6889 void __init free_area_init_memoryless_node(int nid)
6891 free_area_init_node(nid);
6894 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6896 * Initialize all valid struct pages in the range [spfn, epfn) and mark them
6897 * PageReserved(). Return the number of struct pages that were initialized.
6899 static u64 __init init_unavailable_range(unsigned long spfn, unsigned long epfn)
6904 for (pfn = spfn; pfn < epfn; pfn++) {
6905 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6906 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6907 + pageblock_nr_pages - 1;
6911 * Use a fake node/zone (0) for now. Some of these pages
6912 * (in memblock.reserved but not in memblock.memory) will
6913 * get re-initialized via reserve_bootmem_region() later.
6915 __init_single_page(pfn_to_page(pfn), pfn, 0, 0);
6916 __SetPageReserved(pfn_to_page(pfn));
6924 * Only struct pages that are backed by physical memory are zeroed and
6925 * initialized by going through __init_single_page(). But, there are some
6926 * struct pages which are reserved in memblock allocator and their fields
6927 * may be accessed (for example page_to_pfn() on some configuration accesses
6928 * flags). We must explicitly initialize those struct pages.
6930 * This function also addresses a similar issue where struct pages are left
6931 * uninitialized because the physical address range is not covered by
6932 * memblock.memory or memblock.reserved. That could happen when memblock
6933 * layout is manually configured via memmap=, or when the highest physical
6934 * address (max_pfn) does not end on a section boundary.
6936 static void __init init_unavailable_mem(void)
6938 phys_addr_t start, end;
6940 phys_addr_t next = 0;
6943 * Loop through unavailable ranges not covered by memblock.memory.
6946 for_each_mem_range(i, &memblock.memory, NULL,
6947 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
6949 pgcnt += init_unavailable_range(PFN_DOWN(next),
6955 * Early sections always have a fully populated memmap for the whole
6956 * section - see pfn_valid(). If the last section has holes at the
6957 * end and that section is marked "online", the memmap will be
6958 * considered initialized. Make sure that memmap has a well defined
6961 pgcnt += init_unavailable_range(PFN_DOWN(next),
6962 round_up(max_pfn, PAGES_PER_SECTION));
6965 * Struct pages that do not have backing memory. This could be because
6966 * firmware is using some of this memory, or for some other reasons.
6969 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
6972 static inline void __init init_unavailable_mem(void)
6975 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
6977 #if MAX_NUMNODES > 1
6979 * Figure out the number of possible node ids.
6981 void __init setup_nr_node_ids(void)
6983 unsigned int highest;
6985 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6986 nr_node_ids = highest + 1;
6991 * node_map_pfn_alignment - determine the maximum internode alignment
6993 * This function should be called after node map is populated and sorted.
6994 * It calculates the maximum power of two alignment which can distinguish
6997 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6998 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6999 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7000 * shifted, 1GiB is enough and this function will indicate so.
7002 * This is used to test whether pfn -> nid mapping of the chosen memory
7003 * model has fine enough granularity to avoid incorrect mapping for the
7004 * populated node map.
7006 * Return: the determined alignment in pfn's. 0 if there is no alignment
7007 * requirement (single node).
7009 unsigned long __init node_map_pfn_alignment(void)
7011 unsigned long accl_mask = 0, last_end = 0;
7012 unsigned long start, end, mask;
7013 int last_nid = NUMA_NO_NODE;
7016 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7017 if (!start || last_nid < 0 || last_nid == nid) {
7024 * Start with a mask granular enough to pin-point to the
7025 * start pfn and tick off bits one-by-one until it becomes
7026 * too coarse to separate the current node from the last.
7028 mask = ~((1 << __ffs(start)) - 1);
7029 while (mask && last_end <= (start & (mask << 1)))
7032 /* accumulate all internode masks */
7036 /* convert mask to number of pages */
7037 return ~accl_mask + 1;
7041 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7043 * Return: the minimum PFN based on information provided via
7044 * memblock_set_node().
7046 unsigned long __init find_min_pfn_with_active_regions(void)
7048 return PHYS_PFN(memblock_start_of_DRAM());
7052 * early_calculate_totalpages()
7053 * Sum pages in active regions for movable zone.
7054 * Populate N_MEMORY for calculating usable_nodes.
7056 static unsigned long __init early_calculate_totalpages(void)
7058 unsigned long totalpages = 0;
7059 unsigned long start_pfn, end_pfn;
7062 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7063 unsigned long pages = end_pfn - start_pfn;
7065 totalpages += pages;
7067 node_set_state(nid, N_MEMORY);
7073 * Find the PFN the Movable zone begins in each node. Kernel memory
7074 * is spread evenly between nodes as long as the nodes have enough
7075 * memory. When they don't, some nodes will have more kernelcore than
7078 static void __init find_zone_movable_pfns_for_nodes(void)
7081 unsigned long usable_startpfn;
7082 unsigned long kernelcore_node, kernelcore_remaining;
7083 /* save the state before borrow the nodemask */
7084 nodemask_t saved_node_state = node_states[N_MEMORY];
7085 unsigned long totalpages = early_calculate_totalpages();
7086 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7087 struct memblock_region *r;
7089 /* Need to find movable_zone earlier when movable_node is specified. */
7090 find_usable_zone_for_movable();
7093 * If movable_node is specified, ignore kernelcore and movablecore
7096 if (movable_node_is_enabled()) {
7097 for_each_memblock(memory, r) {
7098 if (!memblock_is_hotpluggable(r))
7101 nid = memblock_get_region_node(r);
7103 usable_startpfn = PFN_DOWN(r->base);
7104 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7105 min(usable_startpfn, zone_movable_pfn[nid]) :
7113 * If kernelcore=mirror is specified, ignore movablecore option
7115 if (mirrored_kernelcore) {
7116 bool mem_below_4gb_not_mirrored = false;
7118 for_each_memblock(memory, r) {
7119 if (memblock_is_mirror(r))
7122 nid = memblock_get_region_node(r);
7124 usable_startpfn = memblock_region_memory_base_pfn(r);
7126 if (usable_startpfn < 0x100000) {
7127 mem_below_4gb_not_mirrored = true;
7131 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7132 min(usable_startpfn, zone_movable_pfn[nid]) :
7136 if (mem_below_4gb_not_mirrored)
7137 pr_warn("This configuration results in unmirrored kernel memory.");
7143 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7144 * amount of necessary memory.
7146 if (required_kernelcore_percent)
7147 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7149 if (required_movablecore_percent)
7150 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7154 * If movablecore= was specified, calculate what size of
7155 * kernelcore that corresponds so that memory usable for
7156 * any allocation type is evenly spread. If both kernelcore
7157 * and movablecore are specified, then the value of kernelcore
7158 * will be used for required_kernelcore if it's greater than
7159 * what movablecore would have allowed.
7161 if (required_movablecore) {
7162 unsigned long corepages;
7165 * Round-up so that ZONE_MOVABLE is at least as large as what
7166 * was requested by the user
7168 required_movablecore =
7169 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7170 required_movablecore = min(totalpages, required_movablecore);
7171 corepages = totalpages - required_movablecore;
7173 required_kernelcore = max(required_kernelcore, corepages);
7177 * If kernelcore was not specified or kernelcore size is larger
7178 * than totalpages, there is no ZONE_MOVABLE.
7180 if (!required_kernelcore || required_kernelcore >= totalpages)
7183 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7184 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7187 /* Spread kernelcore memory as evenly as possible throughout nodes */
7188 kernelcore_node = required_kernelcore / usable_nodes;
7189 for_each_node_state(nid, N_MEMORY) {
7190 unsigned long start_pfn, end_pfn;
7193 * Recalculate kernelcore_node if the division per node
7194 * now exceeds what is necessary to satisfy the requested
7195 * amount of memory for the kernel
7197 if (required_kernelcore < kernelcore_node)
7198 kernelcore_node = required_kernelcore / usable_nodes;
7201 * As the map is walked, we track how much memory is usable
7202 * by the kernel using kernelcore_remaining. When it is
7203 * 0, the rest of the node is usable by ZONE_MOVABLE
7205 kernelcore_remaining = kernelcore_node;
7207 /* Go through each range of PFNs within this node */
7208 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7209 unsigned long size_pages;
7211 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7212 if (start_pfn >= end_pfn)
7215 /* Account for what is only usable for kernelcore */
7216 if (start_pfn < usable_startpfn) {
7217 unsigned long kernel_pages;
7218 kernel_pages = min(end_pfn, usable_startpfn)
7221 kernelcore_remaining -= min(kernel_pages,
7222 kernelcore_remaining);
7223 required_kernelcore -= min(kernel_pages,
7224 required_kernelcore);
7226 /* Continue if range is now fully accounted */
7227 if (end_pfn <= usable_startpfn) {
7230 * Push zone_movable_pfn to the end so
7231 * that if we have to rebalance
7232 * kernelcore across nodes, we will
7233 * not double account here
7235 zone_movable_pfn[nid] = end_pfn;
7238 start_pfn = usable_startpfn;
7242 * The usable PFN range for ZONE_MOVABLE is from
7243 * start_pfn->end_pfn. Calculate size_pages as the
7244 * number of pages used as kernelcore
7246 size_pages = end_pfn - start_pfn;
7247 if (size_pages > kernelcore_remaining)
7248 size_pages = kernelcore_remaining;
7249 zone_movable_pfn[nid] = start_pfn + size_pages;
7252 * Some kernelcore has been met, update counts and
7253 * break if the kernelcore for this node has been
7256 required_kernelcore -= min(required_kernelcore,
7258 kernelcore_remaining -= size_pages;
7259 if (!kernelcore_remaining)
7265 * If there is still required_kernelcore, we do another pass with one
7266 * less node in the count. This will push zone_movable_pfn[nid] further
7267 * along on the nodes that still have memory until kernelcore is
7271 if (usable_nodes && required_kernelcore > usable_nodes)
7275 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7276 for (nid = 0; nid < MAX_NUMNODES; nid++)
7277 zone_movable_pfn[nid] =
7278 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7281 /* restore the node_state */
7282 node_states[N_MEMORY] = saved_node_state;
7285 /* Any regular or high memory on that node ? */
7286 static void check_for_memory(pg_data_t *pgdat, int nid)
7288 enum zone_type zone_type;
7290 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7291 struct zone *zone = &pgdat->node_zones[zone_type];
7292 if (populated_zone(zone)) {
7293 if (IS_ENABLED(CONFIG_HIGHMEM))
7294 node_set_state(nid, N_HIGH_MEMORY);
7295 if (zone_type <= ZONE_NORMAL)
7296 node_set_state(nid, N_NORMAL_MEMORY);
7303 * Some architecturs, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7304 * such cases we allow max_zone_pfn sorted in the descending order
7306 bool __weak arch_has_descending_max_zone_pfns(void)
7312 * free_area_init - Initialise all pg_data_t and zone data
7313 * @max_zone_pfn: an array of max PFNs for each zone
7315 * This will call free_area_init_node() for each active node in the system.
7316 * Using the page ranges provided by memblock_set_node(), the size of each
7317 * zone in each node and their holes is calculated. If the maximum PFN
7318 * between two adjacent zones match, it is assumed that the zone is empty.
7319 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7320 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7321 * starts where the previous one ended. For example, ZONE_DMA32 starts
7322 * at arch_max_dma_pfn.
7324 void __init free_area_init(unsigned long *max_zone_pfn)
7326 unsigned long start_pfn, end_pfn;
7330 /* Record where the zone boundaries are */
7331 memset(arch_zone_lowest_possible_pfn, 0,
7332 sizeof(arch_zone_lowest_possible_pfn));
7333 memset(arch_zone_highest_possible_pfn, 0,
7334 sizeof(arch_zone_highest_possible_pfn));
7336 start_pfn = find_min_pfn_with_active_regions();
7337 descending = arch_has_descending_max_zone_pfns();
7339 for (i = 0; i < MAX_NR_ZONES; i++) {
7341 zone = MAX_NR_ZONES - i - 1;
7345 if (zone == ZONE_MOVABLE)
7348 end_pfn = max(max_zone_pfn[zone], start_pfn);
7349 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7350 arch_zone_highest_possible_pfn[zone] = end_pfn;
7352 start_pfn = end_pfn;
7355 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7356 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7357 find_zone_movable_pfns_for_nodes();
7359 /* Print out the zone ranges */
7360 pr_info("Zone ranges:\n");
7361 for (i = 0; i < MAX_NR_ZONES; i++) {
7362 if (i == ZONE_MOVABLE)
7364 pr_info(" %-8s ", zone_names[i]);
7365 if (arch_zone_lowest_possible_pfn[i] ==
7366 arch_zone_highest_possible_pfn[i])
7369 pr_cont("[mem %#018Lx-%#018Lx]\n",
7370 (u64)arch_zone_lowest_possible_pfn[i]
7372 ((u64)arch_zone_highest_possible_pfn[i]
7373 << PAGE_SHIFT) - 1);
7376 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7377 pr_info("Movable zone start for each node\n");
7378 for (i = 0; i < MAX_NUMNODES; i++) {
7379 if (zone_movable_pfn[i])
7380 pr_info(" Node %d: %#018Lx\n", i,
7381 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7385 * Print out the early node map, and initialize the
7386 * subsection-map relative to active online memory ranges to
7387 * enable future "sub-section" extensions of the memory map.
7389 pr_info("Early memory node ranges\n");
7390 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7391 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7392 (u64)start_pfn << PAGE_SHIFT,
7393 ((u64)end_pfn << PAGE_SHIFT) - 1);
7394 subsection_map_init(start_pfn, end_pfn - start_pfn);
7397 /* Initialise every node */
7398 mminit_verify_pageflags_layout();
7399 setup_nr_node_ids();
7400 init_unavailable_mem();
7401 for_each_online_node(nid) {
7402 pg_data_t *pgdat = NODE_DATA(nid);
7403 free_area_init_node(nid);
7405 /* Any memory on that node */
7406 if (pgdat->node_present_pages)
7407 node_set_state(nid, N_MEMORY);
7408 check_for_memory(pgdat, nid);
7412 static int __init cmdline_parse_core(char *p, unsigned long *core,
7413 unsigned long *percent)
7415 unsigned long long coremem;
7421 /* Value may be a percentage of total memory, otherwise bytes */
7422 coremem = simple_strtoull(p, &endptr, 0);
7423 if (*endptr == '%') {
7424 /* Paranoid check for percent values greater than 100 */
7425 WARN_ON(coremem > 100);
7429 coremem = memparse(p, &p);
7430 /* Paranoid check that UL is enough for the coremem value */
7431 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7433 *core = coremem >> PAGE_SHIFT;
7440 * kernelcore=size sets the amount of memory for use for allocations that
7441 * cannot be reclaimed or migrated.
7443 static int __init cmdline_parse_kernelcore(char *p)
7445 /* parse kernelcore=mirror */
7446 if (parse_option_str(p, "mirror")) {
7447 mirrored_kernelcore = true;
7451 return cmdline_parse_core(p, &required_kernelcore,
7452 &required_kernelcore_percent);
7456 * movablecore=size sets the amount of memory for use for allocations that
7457 * can be reclaimed or migrated.
7459 static int __init cmdline_parse_movablecore(char *p)
7461 return cmdline_parse_core(p, &required_movablecore,
7462 &required_movablecore_percent);
7465 early_param("kernelcore", cmdline_parse_kernelcore);
7466 early_param("movablecore", cmdline_parse_movablecore);
7468 void adjust_managed_page_count(struct page *page, long count)
7470 atomic_long_add(count, &page_zone(page)->managed_pages);
7471 totalram_pages_add(count);
7472 #ifdef CONFIG_HIGHMEM
7473 if (PageHighMem(page))
7474 totalhigh_pages_add(count);
7477 EXPORT_SYMBOL(adjust_managed_page_count);
7479 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7482 unsigned long pages = 0;
7484 start = (void *)PAGE_ALIGN((unsigned long)start);
7485 end = (void *)((unsigned long)end & PAGE_MASK);
7486 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7487 struct page *page = virt_to_page(pos);
7488 void *direct_map_addr;
7491 * 'direct_map_addr' might be different from 'pos'
7492 * because some architectures' virt_to_page()
7493 * work with aliases. Getting the direct map
7494 * address ensures that we get a _writeable_
7495 * alias for the memset().
7497 direct_map_addr = page_address(page);
7498 if ((unsigned int)poison <= 0xFF)
7499 memset(direct_map_addr, poison, PAGE_SIZE);
7501 free_reserved_page(page);
7505 pr_info("Freeing %s memory: %ldK\n",
7506 s, pages << (PAGE_SHIFT - 10));
7511 #ifdef CONFIG_HIGHMEM
7512 void free_highmem_page(struct page *page)
7514 __free_reserved_page(page);
7515 totalram_pages_inc();
7516 atomic_long_inc(&page_zone(page)->managed_pages);
7517 totalhigh_pages_inc();
7522 void __init mem_init_print_info(const char *str)
7524 unsigned long physpages, codesize, datasize, rosize, bss_size;
7525 unsigned long init_code_size, init_data_size;
7527 physpages = get_num_physpages();
7528 codesize = _etext - _stext;
7529 datasize = _edata - _sdata;
7530 rosize = __end_rodata - __start_rodata;
7531 bss_size = __bss_stop - __bss_start;
7532 init_data_size = __init_end - __init_begin;
7533 init_code_size = _einittext - _sinittext;
7536 * Detect special cases and adjust section sizes accordingly:
7537 * 1) .init.* may be embedded into .data sections
7538 * 2) .init.text.* may be out of [__init_begin, __init_end],
7539 * please refer to arch/tile/kernel/vmlinux.lds.S.
7540 * 3) .rodata.* may be embedded into .text or .data sections.
7542 #define adj_init_size(start, end, size, pos, adj) \
7544 if (start <= pos && pos < end && size > adj) \
7548 adj_init_size(__init_begin, __init_end, init_data_size,
7549 _sinittext, init_code_size);
7550 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7551 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7552 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7553 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7555 #undef adj_init_size
7557 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7558 #ifdef CONFIG_HIGHMEM
7562 nr_free_pages() << (PAGE_SHIFT - 10),
7563 physpages << (PAGE_SHIFT - 10),
7564 codesize >> 10, datasize >> 10, rosize >> 10,
7565 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7566 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7567 totalcma_pages << (PAGE_SHIFT - 10),
7568 #ifdef CONFIG_HIGHMEM
7569 totalhigh_pages() << (PAGE_SHIFT - 10),
7571 str ? ", " : "", str ? str : "");
7575 * set_dma_reserve - set the specified number of pages reserved in the first zone
7576 * @new_dma_reserve: The number of pages to mark reserved
7578 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7579 * In the DMA zone, a significant percentage may be consumed by kernel image
7580 * and other unfreeable allocations which can skew the watermarks badly. This
7581 * function may optionally be used to account for unfreeable pages in the
7582 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7583 * smaller per-cpu batchsize.
7585 void __init set_dma_reserve(unsigned long new_dma_reserve)
7587 dma_reserve = new_dma_reserve;
7590 static int page_alloc_cpu_dead(unsigned int cpu)
7593 lru_add_drain_cpu(cpu);
7597 * Spill the event counters of the dead processor
7598 * into the current processors event counters.
7599 * This artificially elevates the count of the current
7602 vm_events_fold_cpu(cpu);
7605 * Zero the differential counters of the dead processor
7606 * so that the vm statistics are consistent.
7608 * This is only okay since the processor is dead and cannot
7609 * race with what we are doing.
7611 cpu_vm_stats_fold(cpu);
7616 int hashdist = HASHDIST_DEFAULT;
7618 static int __init set_hashdist(char *str)
7622 hashdist = simple_strtoul(str, &str, 0);
7625 __setup("hashdist=", set_hashdist);
7628 void __init page_alloc_init(void)
7633 if (num_node_state(N_MEMORY) == 1)
7637 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7638 "mm/page_alloc:dead", NULL,
7639 page_alloc_cpu_dead);
7644 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7645 * or min_free_kbytes changes.
7647 static void calculate_totalreserve_pages(void)
7649 struct pglist_data *pgdat;
7650 unsigned long reserve_pages = 0;
7651 enum zone_type i, j;
7653 for_each_online_pgdat(pgdat) {
7655 pgdat->totalreserve_pages = 0;
7657 for (i = 0; i < MAX_NR_ZONES; i++) {
7658 struct zone *zone = pgdat->node_zones + i;
7660 unsigned long managed_pages = zone_managed_pages(zone);
7662 /* Find valid and maximum lowmem_reserve in the zone */
7663 for (j = i; j < MAX_NR_ZONES; j++) {
7664 if (zone->lowmem_reserve[j] > max)
7665 max = zone->lowmem_reserve[j];
7668 /* we treat the high watermark as reserved pages. */
7669 max += high_wmark_pages(zone);
7671 if (max > managed_pages)
7672 max = managed_pages;
7674 pgdat->totalreserve_pages += max;
7676 reserve_pages += max;
7679 totalreserve_pages = reserve_pages;
7683 * setup_per_zone_lowmem_reserve - called whenever
7684 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7685 * has a correct pages reserved value, so an adequate number of
7686 * pages are left in the zone after a successful __alloc_pages().
7688 static void setup_per_zone_lowmem_reserve(void)
7690 struct pglist_data *pgdat;
7691 enum zone_type j, idx;
7693 for_each_online_pgdat(pgdat) {
7694 for (j = 0; j < MAX_NR_ZONES; j++) {
7695 struct zone *zone = pgdat->node_zones + j;
7696 unsigned long managed_pages = zone_managed_pages(zone);
7698 zone->lowmem_reserve[j] = 0;
7702 struct zone *lower_zone;
7705 lower_zone = pgdat->node_zones + idx;
7707 if (!sysctl_lowmem_reserve_ratio[idx] ||
7708 !zone_managed_pages(lower_zone)) {
7709 lower_zone->lowmem_reserve[j] = 0;
7712 lower_zone->lowmem_reserve[j] =
7713 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7715 managed_pages += zone_managed_pages(lower_zone);
7720 /* update totalreserve_pages */
7721 calculate_totalreserve_pages();
7724 static void __setup_per_zone_wmarks(void)
7726 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7727 unsigned long lowmem_pages = 0;
7729 unsigned long flags;
7731 /* Calculate total number of !ZONE_HIGHMEM pages */
7732 for_each_zone(zone) {
7733 if (!is_highmem(zone))
7734 lowmem_pages += zone_managed_pages(zone);
7737 for_each_zone(zone) {
7740 spin_lock_irqsave(&zone->lock, flags);
7741 tmp = (u64)pages_min * zone_managed_pages(zone);
7742 do_div(tmp, lowmem_pages);
7743 if (is_highmem(zone)) {
7745 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7746 * need highmem pages, so cap pages_min to a small
7749 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7750 * deltas control async page reclaim, and so should
7751 * not be capped for highmem.
7753 unsigned long min_pages;
7755 min_pages = zone_managed_pages(zone) / 1024;
7756 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7757 zone->_watermark[WMARK_MIN] = min_pages;
7760 * If it's a lowmem zone, reserve a number of pages
7761 * proportionate to the zone's size.
7763 zone->_watermark[WMARK_MIN] = tmp;
7767 * Set the kswapd watermarks distance according to the
7768 * scale factor in proportion to available memory, but
7769 * ensure a minimum size on small systems.
7771 tmp = max_t(u64, tmp >> 2,
7772 mult_frac(zone_managed_pages(zone),
7773 watermark_scale_factor, 10000));
7775 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7776 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7777 zone->watermark_boost = 0;
7779 spin_unlock_irqrestore(&zone->lock, flags);
7782 /* update totalreserve_pages */
7783 calculate_totalreserve_pages();
7787 * setup_per_zone_wmarks - called when min_free_kbytes changes
7788 * or when memory is hot-{added|removed}
7790 * Ensures that the watermark[min,low,high] values for each zone are set
7791 * correctly with respect to min_free_kbytes.
7793 void setup_per_zone_wmarks(void)
7795 static DEFINE_SPINLOCK(lock);
7798 __setup_per_zone_wmarks();
7803 * Initialise min_free_kbytes.
7805 * For small machines we want it small (128k min). For large machines
7806 * we want it large (64MB max). But it is not linear, because network
7807 * bandwidth does not increase linearly with machine size. We use
7809 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7810 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7826 int __meminit init_per_zone_wmark_min(void)
7828 unsigned long lowmem_kbytes;
7829 int new_min_free_kbytes;
7831 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7832 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7834 if (new_min_free_kbytes > user_min_free_kbytes) {
7835 min_free_kbytes = new_min_free_kbytes;
7836 if (min_free_kbytes < 128)
7837 min_free_kbytes = 128;
7838 if (min_free_kbytes > 262144)
7839 min_free_kbytes = 262144;
7841 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7842 new_min_free_kbytes, user_min_free_kbytes);
7844 setup_per_zone_wmarks();
7845 refresh_zone_stat_thresholds();
7846 setup_per_zone_lowmem_reserve();
7849 setup_min_unmapped_ratio();
7850 setup_min_slab_ratio();
7855 core_initcall(init_per_zone_wmark_min)
7858 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7859 * that we can call two helper functions whenever min_free_kbytes
7862 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7863 void __user *buffer, size_t *length, loff_t *ppos)
7867 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7872 user_min_free_kbytes = min_free_kbytes;
7873 setup_per_zone_wmarks();
7878 int watermark_boost_factor_sysctl_handler(struct ctl_table *table, int write,
7879 void __user *buffer, size_t *length, loff_t *ppos)
7883 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7890 int watermark_scale_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);
7900 setup_per_zone_wmarks();
7906 static void setup_min_unmapped_ratio(void)
7911 for_each_online_pgdat(pgdat)
7912 pgdat->min_unmapped_pages = 0;
7915 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7916 sysctl_min_unmapped_ratio) / 100;
7920 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7921 void __user *buffer, size_t *length, loff_t *ppos)
7925 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7929 setup_min_unmapped_ratio();
7934 static void setup_min_slab_ratio(void)
7939 for_each_online_pgdat(pgdat)
7940 pgdat->min_slab_pages = 0;
7943 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7944 sysctl_min_slab_ratio) / 100;
7947 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7948 void __user *buffer, size_t *length, loff_t *ppos)
7952 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7956 setup_min_slab_ratio();
7963 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7964 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7965 * whenever sysctl_lowmem_reserve_ratio changes.
7967 * The reserve ratio obviously has absolutely no relation with the
7968 * minimum watermarks. The lowmem reserve ratio can only make sense
7969 * if in function of the boot time zone sizes.
7971 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7972 void __user *buffer, size_t *length, loff_t *ppos)
7976 proc_dointvec_minmax(table, write, buffer, length, ppos);
7978 for (i = 0; i < MAX_NR_ZONES; i++) {
7979 if (sysctl_lowmem_reserve_ratio[i] < 1)
7980 sysctl_lowmem_reserve_ratio[i] = 0;
7983 setup_per_zone_lowmem_reserve();
7987 static void __zone_pcp_update(struct zone *zone)
7991 for_each_possible_cpu(cpu)
7992 pageset_set_high_and_batch(zone,
7993 per_cpu_ptr(zone->pageset, cpu));
7997 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7998 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7999 * pagelist can have before it gets flushed back to buddy allocator.
8001 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8002 void __user *buffer, size_t *length, loff_t *ppos)
8005 int old_percpu_pagelist_fraction;
8008 mutex_lock(&pcp_batch_high_lock);
8009 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8011 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8012 if (!write || ret < 0)
8015 /* Sanity checking to avoid pcp imbalance */
8016 if (percpu_pagelist_fraction &&
8017 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8018 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8024 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8027 for_each_populated_zone(zone)
8028 __zone_pcp_update(zone);
8030 mutex_unlock(&pcp_batch_high_lock);
8034 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8036 * Returns the number of pages that arch has reserved but
8037 * is not known to alloc_large_system_hash().
8039 static unsigned long __init arch_reserved_kernel_pages(void)
8046 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8047 * machines. As memory size is increased the scale is also increased but at
8048 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8049 * quadruples the scale is increased by one, which means the size of hash table
8050 * only doubles, instead of quadrupling as well.
8051 * Because 32-bit systems cannot have large physical memory, where this scaling
8052 * makes sense, it is disabled on such platforms.
8054 #if __BITS_PER_LONG > 32
8055 #define ADAPT_SCALE_BASE (64ul << 30)
8056 #define ADAPT_SCALE_SHIFT 2
8057 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8061 * allocate a large system hash table from bootmem
8062 * - it is assumed that the hash table must contain an exact power-of-2
8063 * quantity of entries
8064 * - limit is the number of hash buckets, not the total allocation size
8066 void *__init alloc_large_system_hash(const char *tablename,
8067 unsigned long bucketsize,
8068 unsigned long numentries,
8071 unsigned int *_hash_shift,
8072 unsigned int *_hash_mask,
8073 unsigned long low_limit,
8074 unsigned long high_limit)
8076 unsigned long long max = high_limit;
8077 unsigned long log2qty, size;
8082 /* allow the kernel cmdline to have a say */
8084 /* round applicable memory size up to nearest megabyte */
8085 numentries = nr_kernel_pages;
8086 numentries -= arch_reserved_kernel_pages();
8088 /* It isn't necessary when PAGE_SIZE >= 1MB */
8089 if (PAGE_SHIFT < 20)
8090 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8092 #if __BITS_PER_LONG > 32
8094 unsigned long adapt;
8096 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8097 adapt <<= ADAPT_SCALE_SHIFT)
8102 /* limit to 1 bucket per 2^scale bytes of low memory */
8103 if (scale > PAGE_SHIFT)
8104 numentries >>= (scale - PAGE_SHIFT);
8106 numentries <<= (PAGE_SHIFT - scale);
8108 /* Make sure we've got at least a 0-order allocation.. */
8109 if (unlikely(flags & HASH_SMALL)) {
8110 /* Makes no sense without HASH_EARLY */
8111 WARN_ON(!(flags & HASH_EARLY));
8112 if (!(numentries >> *_hash_shift)) {
8113 numentries = 1UL << *_hash_shift;
8114 BUG_ON(!numentries);
8116 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8117 numentries = PAGE_SIZE / bucketsize;
8119 numentries = roundup_pow_of_two(numentries);
8121 /* limit allocation size to 1/16 total memory by default */
8123 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8124 do_div(max, bucketsize);
8126 max = min(max, 0x80000000ULL);
8128 if (numentries < low_limit)
8129 numentries = low_limit;
8130 if (numentries > max)
8133 log2qty = ilog2(numentries);
8135 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8138 size = bucketsize << log2qty;
8139 if (flags & HASH_EARLY) {
8140 if (flags & HASH_ZERO)
8141 table = memblock_alloc(size, SMP_CACHE_BYTES);
8143 table = memblock_alloc_raw(size,
8145 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8146 table = __vmalloc(size, gfp_flags);
8150 * If bucketsize is not a power-of-two, we may free
8151 * some pages at the end of hash table which
8152 * alloc_pages_exact() automatically does
8154 table = alloc_pages_exact(size, gfp_flags);
8155 kmemleak_alloc(table, size, 1, gfp_flags);
8157 } while (!table && size > PAGE_SIZE && --log2qty);
8160 panic("Failed to allocate %s hash table\n", tablename);
8162 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8163 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8164 virt ? "vmalloc" : "linear");
8167 *_hash_shift = log2qty;
8169 *_hash_mask = (1 << log2qty) - 1;
8175 * This function checks whether pageblock includes unmovable pages or not.
8177 * PageLRU check without isolation or lru_lock could race so that
8178 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8179 * check without lock_page also may miss some movable non-lru pages at
8180 * race condition. So you can't expect this function should be exact.
8182 * Returns a page without holding a reference. If the caller wants to
8183 * dereference that page (e.g., dumping), it has to make sure that that it
8184 * cannot get removed (e.g., via memory unplug) concurrently.
8187 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8188 int migratetype, int flags)
8190 unsigned long iter = 0;
8191 unsigned long pfn = page_to_pfn(page);
8194 * TODO we could make this much more efficient by not checking every
8195 * page in the range if we know all of them are in MOVABLE_ZONE and
8196 * that the movable zone guarantees that pages are migratable but
8197 * the later is not the case right now unfortunatelly. E.g. movablecore
8198 * can still lead to having bootmem allocations in zone_movable.
8201 if (is_migrate_cma_page(page)) {
8203 * CMA allocations (alloc_contig_range) really need to mark
8204 * isolate CMA pageblocks even when they are not movable in fact
8205 * so consider them movable here.
8207 if (is_migrate_cma(migratetype))
8213 for (; iter < pageblock_nr_pages; iter++) {
8214 if (!pfn_valid_within(pfn + iter))
8217 page = pfn_to_page(pfn + iter);
8219 if (PageReserved(page))
8223 * If the zone is movable and we have ruled out all reserved
8224 * pages then it should be reasonably safe to assume the rest
8227 if (zone_idx(zone) == ZONE_MOVABLE)
8231 * Hugepages are not in LRU lists, but they're movable.
8232 * THPs are on the LRU, but need to be counted as #small pages.
8233 * We need not scan over tail pages because we don't
8234 * handle each tail page individually in migration.
8236 if (PageHuge(page) || PageTransCompound(page)) {
8237 struct page *head = compound_head(page);
8238 unsigned int skip_pages;
8240 if (PageHuge(page)) {
8241 if (!hugepage_migration_supported(page_hstate(head)))
8243 } else if (!PageLRU(head) && !__PageMovable(head)) {
8247 skip_pages = compound_nr(head) - (page - head);
8248 iter += skip_pages - 1;
8253 * We can't use page_count without pin a page
8254 * because another CPU can free compound page.
8255 * This check already skips compound tails of THP
8256 * because their page->_refcount is zero at all time.
8258 if (!page_ref_count(page)) {
8259 if (PageBuddy(page))
8260 iter += (1 << page_order(page)) - 1;
8265 * The HWPoisoned page may be not in buddy system, and
8266 * page_count() is not 0.
8268 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8271 if (__PageMovable(page) || PageLRU(page))
8275 * If there are RECLAIMABLE pages, we need to check
8276 * it. But now, memory offline itself doesn't call
8277 * shrink_node_slabs() and it still to be fixed.
8280 * If the page is not RAM, page_count()should be 0.
8281 * we don't need more check. This is an _used_ not-movable page.
8283 * The problematic thing here is PG_reserved pages. PG_reserved
8284 * is set to both of a memory hole page and a _used_ kernel
8292 #ifdef CONFIG_CONTIG_ALLOC
8293 static unsigned long pfn_max_align_down(unsigned long pfn)
8295 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8296 pageblock_nr_pages) - 1);
8299 static unsigned long pfn_max_align_up(unsigned long pfn)
8301 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8302 pageblock_nr_pages));
8305 /* [start, end) must belong to a single zone. */
8306 static int __alloc_contig_migrate_range(struct compact_control *cc,
8307 unsigned long start, unsigned long end)
8309 /* This function is based on compact_zone() from compaction.c. */
8310 unsigned long nr_reclaimed;
8311 unsigned long pfn = start;
8312 unsigned int tries = 0;
8317 while (pfn < end || !list_empty(&cc->migratepages)) {
8318 if (fatal_signal_pending(current)) {
8323 if (list_empty(&cc->migratepages)) {
8324 cc->nr_migratepages = 0;
8325 pfn = isolate_migratepages_range(cc, pfn, end);
8331 } else if (++tries == 5) {
8332 ret = ret < 0 ? ret : -EBUSY;
8336 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8338 cc->nr_migratepages -= nr_reclaimed;
8340 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
8341 NULL, 0, cc->mode, MR_CONTIG_RANGE);
8344 putback_movable_pages(&cc->migratepages);
8351 * alloc_contig_range() -- tries to allocate given range of pages
8352 * @start: start PFN to allocate
8353 * @end: one-past-the-last PFN to allocate
8354 * @migratetype: migratetype of the underlaying pageblocks (either
8355 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8356 * in range must have the same migratetype and it must
8357 * be either of the two.
8358 * @gfp_mask: GFP mask to use during compaction
8360 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8361 * aligned. The PFN range must belong to a single zone.
8363 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8364 * pageblocks in the range. Once isolated, the pageblocks should not
8365 * be modified by others.
8367 * Return: zero on success or negative error code. On success all
8368 * pages which PFN is in [start, end) are allocated for the caller and
8369 * need to be freed with free_contig_range().
8371 int alloc_contig_range(unsigned long start, unsigned long end,
8372 unsigned migratetype, gfp_t gfp_mask)
8374 unsigned long outer_start, outer_end;
8378 struct compact_control cc = {
8379 .nr_migratepages = 0,
8381 .zone = page_zone(pfn_to_page(start)),
8382 .mode = MIGRATE_SYNC,
8383 .ignore_skip_hint = true,
8384 .no_set_skip_hint = true,
8385 .gfp_mask = current_gfp_context(gfp_mask),
8386 .alloc_contig = true,
8388 INIT_LIST_HEAD(&cc.migratepages);
8391 * What we do here is we mark all pageblocks in range as
8392 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8393 * have different sizes, and due to the way page allocator
8394 * work, we align the range to biggest of the two pages so
8395 * that page allocator won't try to merge buddies from
8396 * different pageblocks and change MIGRATE_ISOLATE to some
8397 * other migration type.
8399 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8400 * migrate the pages from an unaligned range (ie. pages that
8401 * we are interested in). This will put all the pages in
8402 * range back to page allocator as MIGRATE_ISOLATE.
8404 * When this is done, we take the pages in range from page
8405 * allocator removing them from the buddy system. This way
8406 * page allocator will never consider using them.
8408 * This lets us mark the pageblocks back as
8409 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8410 * aligned range but not in the unaligned, original range are
8411 * put back to page allocator so that buddy can use them.
8414 ret = start_isolate_page_range(pfn_max_align_down(start),
8415 pfn_max_align_up(end), migratetype, 0);
8420 * In case of -EBUSY, we'd like to know which page causes problem.
8421 * So, just fall through. test_pages_isolated() has a tracepoint
8422 * which will report the busy page.
8424 * It is possible that busy pages could become available before
8425 * the call to test_pages_isolated, and the range will actually be
8426 * allocated. So, if we fall through be sure to clear ret so that
8427 * -EBUSY is not accidentally used or returned to caller.
8429 ret = __alloc_contig_migrate_range(&cc, start, end);
8430 if (ret && ret != -EBUSY)
8435 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8436 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8437 * more, all pages in [start, end) are free in page allocator.
8438 * What we are going to do is to allocate all pages from
8439 * [start, end) (that is remove them from page allocator).
8441 * The only problem is that pages at the beginning and at the
8442 * end of interesting range may be not aligned with pages that
8443 * page allocator holds, ie. they can be part of higher order
8444 * pages. Because of this, we reserve the bigger range and
8445 * once this is done free the pages we are not interested in.
8447 * We don't have to hold zone->lock here because the pages are
8448 * isolated thus they won't get removed from buddy.
8451 lru_add_drain_all();
8454 outer_start = start;
8455 while (!PageBuddy(pfn_to_page(outer_start))) {
8456 if (++order >= MAX_ORDER) {
8457 outer_start = start;
8460 outer_start &= ~0UL << order;
8463 if (outer_start != start) {
8464 order = page_order(pfn_to_page(outer_start));
8467 * outer_start page could be small order buddy page and
8468 * it doesn't include start page. Adjust outer_start
8469 * in this case to report failed page properly
8470 * on tracepoint in test_pages_isolated()
8472 if (outer_start + (1UL << order) <= start)
8473 outer_start = start;
8476 /* Make sure the range is really isolated. */
8477 if (test_pages_isolated(outer_start, end, 0)) {
8478 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8479 __func__, outer_start, end);
8484 /* Grab isolated pages from freelists. */
8485 outer_end = isolate_freepages_range(&cc, outer_start, end);
8491 /* Free head and tail (if any) */
8492 if (start != outer_start)
8493 free_contig_range(outer_start, start - outer_start);
8494 if (end != outer_end)
8495 free_contig_range(end, outer_end - end);
8498 undo_isolate_page_range(pfn_max_align_down(start),
8499 pfn_max_align_up(end), migratetype);
8503 static int __alloc_contig_pages(unsigned long start_pfn,
8504 unsigned long nr_pages, gfp_t gfp_mask)
8506 unsigned long end_pfn = start_pfn + nr_pages;
8508 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8512 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8513 unsigned long nr_pages)
8515 unsigned long i, end_pfn = start_pfn + nr_pages;
8518 for (i = start_pfn; i < end_pfn; i++) {
8519 page = pfn_to_online_page(i);
8523 if (page_zone(page) != z)
8526 if (PageReserved(page))
8529 if (page_count(page) > 0)
8538 static bool zone_spans_last_pfn(const struct zone *zone,
8539 unsigned long start_pfn, unsigned long nr_pages)
8541 unsigned long last_pfn = start_pfn + nr_pages - 1;
8543 return zone_spans_pfn(zone, last_pfn);
8547 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8548 * @nr_pages: Number of contiguous pages to allocate
8549 * @gfp_mask: GFP mask to limit search and used during compaction
8551 * @nodemask: Mask for other possible nodes
8553 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8554 * on an applicable zonelist to find a contiguous pfn range which can then be
8555 * tried for allocation with alloc_contig_range(). This routine is intended
8556 * for allocation requests which can not be fulfilled with the buddy allocator.
8558 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8559 * power of two then the alignment is guaranteed to be to the given nr_pages
8560 * (e.g. 1GB request would be aligned to 1GB).
8562 * Allocated pages can be freed with free_contig_range() or by manually calling
8563 * __free_page() on each allocated page.
8565 * Return: pointer to contiguous pages on success, or NULL if not successful.
8567 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8568 int nid, nodemask_t *nodemask)
8570 unsigned long ret, pfn, flags;
8571 struct zonelist *zonelist;
8575 zonelist = node_zonelist(nid, gfp_mask);
8576 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8577 gfp_zone(gfp_mask), nodemask) {
8578 spin_lock_irqsave(&zone->lock, flags);
8580 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8581 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8582 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8584 * We release the zone lock here because
8585 * alloc_contig_range() will also lock the zone
8586 * at some point. If there's an allocation
8587 * spinning on this lock, it may win the race
8588 * and cause alloc_contig_range() to fail...
8590 spin_unlock_irqrestore(&zone->lock, flags);
8591 ret = __alloc_contig_pages(pfn, nr_pages,
8594 return pfn_to_page(pfn);
8595 spin_lock_irqsave(&zone->lock, flags);
8599 spin_unlock_irqrestore(&zone->lock, flags);
8603 #endif /* CONFIG_CONTIG_ALLOC */
8605 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8607 unsigned int count = 0;
8609 for (; nr_pages--; pfn++) {
8610 struct page *page = pfn_to_page(pfn);
8612 count += page_count(page) != 1;
8615 WARN(count != 0, "%d pages are still in use!\n", count);
8619 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8620 * page high values need to be recalulated.
8622 void __meminit zone_pcp_update(struct zone *zone)
8624 mutex_lock(&pcp_batch_high_lock);
8625 __zone_pcp_update(zone);
8626 mutex_unlock(&pcp_batch_high_lock);
8629 void zone_pcp_reset(struct zone *zone)
8631 unsigned long flags;
8633 struct per_cpu_pageset *pset;
8635 /* avoid races with drain_pages() */
8636 local_irq_save(flags);
8637 if (zone->pageset != &boot_pageset) {
8638 for_each_online_cpu(cpu) {
8639 pset = per_cpu_ptr(zone->pageset, cpu);
8640 drain_zonestat(zone, pset);
8642 free_percpu(zone->pageset);
8643 zone->pageset = &boot_pageset;
8645 local_irq_restore(flags);
8648 #ifdef CONFIG_MEMORY_HOTREMOVE
8650 * All pages in the range must be in a single zone and isolated
8651 * before calling this.
8654 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8660 unsigned long flags;
8661 unsigned long offlined_pages = 0;
8663 /* find the first valid pfn */
8664 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8668 return offlined_pages;
8670 offline_mem_sections(pfn, end_pfn);
8671 zone = page_zone(pfn_to_page(pfn));
8672 spin_lock_irqsave(&zone->lock, flags);
8674 while (pfn < end_pfn) {
8675 if (!pfn_valid(pfn)) {
8679 page = pfn_to_page(pfn);
8681 * The HWPoisoned page may be not in buddy system, and
8682 * page_count() is not 0.
8684 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8690 BUG_ON(page_count(page));
8691 BUG_ON(!PageBuddy(page));
8692 order = page_order(page);
8693 offlined_pages += 1 << order;
8694 del_page_from_free_list(page, zone, order);
8695 pfn += (1 << order);
8697 spin_unlock_irqrestore(&zone->lock, flags);
8699 return offlined_pages;
8703 bool is_free_buddy_page(struct page *page)
8705 struct zone *zone = page_zone(page);
8706 unsigned long pfn = page_to_pfn(page);
8707 unsigned long flags;
8710 spin_lock_irqsave(&zone->lock, flags);
8711 for (order = 0; order < MAX_ORDER; order++) {
8712 struct page *page_head = page - (pfn & ((1 << order) - 1));
8714 if (PageBuddy(page_head) && page_order(page_head) >= order)
8717 spin_unlock_irqrestore(&zone->lock, flags);
8719 return order < MAX_ORDER;
8722 #ifdef CONFIG_MEMORY_FAILURE
8724 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8725 * test is performed under the zone lock to prevent a race against page
8728 bool set_hwpoison_free_buddy_page(struct page *page)
8730 struct zone *zone = page_zone(page);
8731 unsigned long pfn = page_to_pfn(page);
8732 unsigned long flags;
8734 bool hwpoisoned = false;
8736 spin_lock_irqsave(&zone->lock, flags);
8737 for (order = 0; order < MAX_ORDER; order++) {
8738 struct page *page_head = page - (pfn & ((1 << order) - 1));
8740 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8741 if (!TestSetPageHWPoison(page))
8746 spin_unlock_irqrestore(&zone->lock, flags);