1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/page_alloc.c
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
18 #include <linux/stddef.h>
20 #include <linux/highmem.h>
21 #include <linux/swap.h>
22 #include <linux/interrupt.h>
23 #include <linux/pagemap.h>
24 #include <linux/jiffies.h>
25 #include <linux/memblock.h>
26 #include <linux/compiler.h>
27 #include <linux/kernel.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/memremap.h>
46 #include <linux/stop_machine.h>
47 #include <linux/random.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <trace/events/oom.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/sched/mm.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
66 #include <linux/memcontrol.h>
67 #include <linux/ftrace.h>
68 #include <linux/lockdep.h>
69 #include <linux/nmi.h>
70 #include <linux/psi.h>
71 #include <linux/padata.h>
73 #include <asm/sections.h>
74 #include <asm/tlbflush.h>
75 #include <asm/div64.h>
78 #include "page_reporting.h"
80 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
81 static DEFINE_MUTEX(pcp_batch_high_lock);
82 #define MIN_PERCPU_PAGELIST_FRACTION (8)
84 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
85 DEFINE_PER_CPU(int, numa_node);
86 EXPORT_PER_CPU_SYMBOL(numa_node);
89 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
91 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
93 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
94 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
95 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
96 * defined in <linux/topology.h>.
98 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
99 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
102 /* work_structs for global per-cpu drains */
105 struct work_struct work;
107 static DEFINE_MUTEX(pcpu_drain_mutex);
108 static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
110 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
111 volatile unsigned long latent_entropy __latent_entropy;
112 EXPORT_SYMBOL(latent_entropy);
116 * Array of node states.
118 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
119 [N_POSSIBLE] = NODE_MASK_ALL,
120 [N_ONLINE] = { { [0] = 1UL } },
122 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
123 #ifdef CONFIG_HIGHMEM
124 [N_HIGH_MEMORY] = { { [0] = 1UL } },
126 [N_MEMORY] = { { [0] = 1UL } },
127 [N_CPU] = { { [0] = 1UL } },
130 EXPORT_SYMBOL(node_states);
132 atomic_long_t _totalram_pages __read_mostly;
133 EXPORT_SYMBOL(_totalram_pages);
134 unsigned long totalreserve_pages __read_mostly;
135 unsigned long totalcma_pages __read_mostly;
137 int percpu_pagelist_fraction;
138 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
139 #ifdef CONFIG_INIT_ON_ALLOC_DEFAULT_ON
140 DEFINE_STATIC_KEY_TRUE(init_on_alloc);
142 DEFINE_STATIC_KEY_FALSE(init_on_alloc);
144 EXPORT_SYMBOL(init_on_alloc);
146 #ifdef CONFIG_INIT_ON_FREE_DEFAULT_ON
147 DEFINE_STATIC_KEY_TRUE(init_on_free);
149 DEFINE_STATIC_KEY_FALSE(init_on_free);
151 EXPORT_SYMBOL(init_on_free);
153 static int __init early_init_on_alloc(char *buf)
160 ret = kstrtobool(buf, &bool_result);
161 if (bool_result && page_poisoning_enabled())
162 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_alloc\n");
164 static_branch_enable(&init_on_alloc);
166 static_branch_disable(&init_on_alloc);
169 early_param("init_on_alloc", early_init_on_alloc);
171 static int __init early_init_on_free(char *buf)
178 ret = kstrtobool(buf, &bool_result);
179 if (bool_result && page_poisoning_enabled())
180 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_free\n");
182 static_branch_enable(&init_on_free);
184 static_branch_disable(&init_on_free);
187 early_param("init_on_free", early_init_on_free);
190 * A cached value of the page's pageblock's migratetype, used when the page is
191 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
192 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
193 * Also the migratetype set in the page does not necessarily match the pcplist
194 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
195 * other index - this ensures that it will be put on the correct CMA freelist.
197 static inline int get_pcppage_migratetype(struct page *page)
202 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
204 page->index = migratetype;
207 #ifdef CONFIG_PM_SLEEP
209 * The following functions are used by the suspend/hibernate code to temporarily
210 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
211 * while devices are suspended. To avoid races with the suspend/hibernate code,
212 * they should always be called with system_transition_mutex held
213 * (gfp_allowed_mask also should only be modified with system_transition_mutex
214 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
215 * with that modification).
218 static gfp_t saved_gfp_mask;
220 void pm_restore_gfp_mask(void)
222 WARN_ON(!mutex_is_locked(&system_transition_mutex));
223 if (saved_gfp_mask) {
224 gfp_allowed_mask = saved_gfp_mask;
229 void pm_restrict_gfp_mask(void)
231 WARN_ON(!mutex_is_locked(&system_transition_mutex));
232 WARN_ON(saved_gfp_mask);
233 saved_gfp_mask = gfp_allowed_mask;
234 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
237 bool pm_suspended_storage(void)
239 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
243 #endif /* CONFIG_PM_SLEEP */
245 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
246 unsigned int pageblock_order __read_mostly;
249 static void __free_pages_ok(struct page *page, unsigned int order);
252 * results with 256, 32 in the lowmem_reserve sysctl:
253 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
254 * 1G machine -> (16M dma, 784M normal, 224M high)
255 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
256 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
257 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
259 * TBD: should special case ZONE_DMA32 machines here - in those we normally
260 * don't need any ZONE_NORMAL reservation
262 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
263 #ifdef CONFIG_ZONE_DMA
266 #ifdef CONFIG_ZONE_DMA32
270 #ifdef CONFIG_HIGHMEM
276 static char * const zone_names[MAX_NR_ZONES] = {
277 #ifdef CONFIG_ZONE_DMA
280 #ifdef CONFIG_ZONE_DMA32
284 #ifdef CONFIG_HIGHMEM
288 #ifdef CONFIG_ZONE_DEVICE
293 const char * const migratetype_names[MIGRATE_TYPES] = {
301 #ifdef CONFIG_MEMORY_ISOLATION
306 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
307 [NULL_COMPOUND_DTOR] = NULL,
308 [COMPOUND_PAGE_DTOR] = free_compound_page,
309 #ifdef CONFIG_HUGETLB_PAGE
310 [HUGETLB_PAGE_DTOR] = free_huge_page,
312 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
313 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
317 int min_free_kbytes = 1024;
318 int user_min_free_kbytes = -1;
319 #ifdef CONFIG_DISCONTIGMEM
321 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
322 * are not on separate NUMA nodes. Functionally this works but with
323 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
324 * quite small. By default, do not boost watermarks on discontigmem as in
325 * many cases very high-order allocations like THP are likely to be
326 * unsupported and the premature reclaim offsets the advantage of long-term
327 * fragmentation avoidance.
329 int watermark_boost_factor __read_mostly;
331 int watermark_boost_factor __read_mostly = 15000;
333 int watermark_scale_factor = 10;
335 static unsigned long nr_kernel_pages __initdata;
336 static unsigned long nr_all_pages __initdata;
337 static unsigned long dma_reserve __initdata;
339 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
340 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
341 static unsigned long required_kernelcore __initdata;
342 static unsigned long required_kernelcore_percent __initdata;
343 static unsigned long required_movablecore __initdata;
344 static unsigned long required_movablecore_percent __initdata;
345 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
346 static bool mirrored_kernelcore __meminitdata;
348 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
350 EXPORT_SYMBOL(movable_zone);
353 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
354 unsigned int nr_online_nodes __read_mostly = 1;
355 EXPORT_SYMBOL(nr_node_ids);
356 EXPORT_SYMBOL(nr_online_nodes);
359 int page_group_by_mobility_disabled __read_mostly;
361 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
363 * During boot we initialize deferred pages on-demand, as needed, but once
364 * page_alloc_init_late() has finished, the deferred pages are all initialized,
365 * and we can permanently disable that path.
367 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
370 * Calling kasan_free_pages() only after deferred memory initialization
371 * has completed. Poisoning pages during deferred memory init will greatly
372 * lengthen the process and cause problem in large memory systems as the
373 * deferred pages initialization is done with interrupt disabled.
375 * Assuming that there will be no reference to those newly initialized
376 * pages before they are ever allocated, this should have no effect on
377 * KASAN memory tracking as the poison will be properly inserted at page
378 * allocation time. The only corner case is when pages are allocated by
379 * on-demand allocation and then freed again before the deferred pages
380 * initialization is done, but this is not likely to happen.
382 static inline void kasan_free_nondeferred_pages(struct page *page, int order)
384 if (!static_branch_unlikely(&deferred_pages))
385 kasan_free_pages(page, order);
388 /* Returns true if the struct page for the pfn is uninitialised */
389 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
391 int nid = early_pfn_to_nid(pfn);
393 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
400 * Returns true when the remaining initialisation should be deferred until
401 * later in the boot cycle when it can be parallelised.
403 static bool __meminit
404 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
406 static unsigned long prev_end_pfn, nr_initialised;
409 * prev_end_pfn static that contains the end of previous zone
410 * No need to protect because called very early in boot before smp_init.
412 if (prev_end_pfn != end_pfn) {
413 prev_end_pfn = end_pfn;
417 /* Always populate low zones for address-constrained allocations */
418 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
422 * We start only with one section of pages, more pages are added as
423 * needed until the rest of deferred pages are initialized.
426 if ((nr_initialised > PAGES_PER_SECTION) &&
427 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
428 NODE_DATA(nid)->first_deferred_pfn = pfn;
434 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
436 static inline bool early_page_uninitialised(unsigned long pfn)
441 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
447 /* Return a pointer to the bitmap storing bits affecting a block of pages */
448 static inline unsigned long *get_pageblock_bitmap(struct page *page,
451 #ifdef CONFIG_SPARSEMEM
452 return section_to_usemap(__pfn_to_section(pfn));
454 return page_zone(page)->pageblock_flags;
455 #endif /* CONFIG_SPARSEMEM */
458 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
460 #ifdef CONFIG_SPARSEMEM
461 pfn &= (PAGES_PER_SECTION-1);
463 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
464 #endif /* CONFIG_SPARSEMEM */
465 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
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 * @mask: mask of bits that the caller is interested in
474 * Return: pageblock_bits flags
476 static __always_inline
477 unsigned long __get_pfnblock_flags_mask(struct page *page,
481 unsigned long *bitmap;
482 unsigned long bitidx, word_bitidx;
485 bitmap = get_pageblock_bitmap(page, pfn);
486 bitidx = pfn_to_bitidx(page, pfn);
487 word_bitidx = bitidx / BITS_PER_LONG;
488 bitidx &= (BITS_PER_LONG-1);
490 word = bitmap[word_bitidx];
491 return (word >> bitidx) & mask;
494 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
497 return __get_pfnblock_flags_mask(page, pfn, mask);
500 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
502 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
506 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
507 * @page: The page within the block of interest
508 * @flags: The flags to set
509 * @pfn: The target page frame number
510 * @mask: mask of bits that the caller is interested in
512 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
516 unsigned long *bitmap;
517 unsigned long bitidx, word_bitidx;
518 unsigned long old_word, word;
520 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
521 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
523 bitmap = get_pageblock_bitmap(page, pfn);
524 bitidx = pfn_to_bitidx(page, pfn);
525 word_bitidx = bitidx / BITS_PER_LONG;
526 bitidx &= (BITS_PER_LONG-1);
528 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
533 word = READ_ONCE(bitmap[word_bitidx]);
535 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
536 if (word == old_word)
542 void set_pageblock_migratetype(struct page *page, int migratetype)
544 if (unlikely(page_group_by_mobility_disabled &&
545 migratetype < MIGRATE_PCPTYPES))
546 migratetype = MIGRATE_UNMOVABLE;
548 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
549 page_to_pfn(page), MIGRATETYPE_MASK);
552 #ifdef CONFIG_DEBUG_VM
553 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
557 unsigned long pfn = page_to_pfn(page);
558 unsigned long sp, start_pfn;
561 seq = zone_span_seqbegin(zone);
562 start_pfn = zone->zone_start_pfn;
563 sp = zone->spanned_pages;
564 if (!zone_spans_pfn(zone, pfn))
566 } while (zone_span_seqretry(zone, seq));
569 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
570 pfn, zone_to_nid(zone), zone->name,
571 start_pfn, start_pfn + sp);
576 static int page_is_consistent(struct zone *zone, struct page *page)
578 if (!pfn_valid_within(page_to_pfn(page)))
580 if (zone != page_zone(page))
586 * Temporary debugging check for pages not lying within a given zone.
588 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
590 if (page_outside_zone_boundaries(zone, page))
592 if (!page_is_consistent(zone, page))
598 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
604 static void bad_page(struct page *page, const char *reason)
606 static unsigned long resume;
607 static unsigned long nr_shown;
608 static unsigned long nr_unshown;
611 * Allow a burst of 60 reports, then keep quiet for that minute;
612 * or allow a steady drip of one report per second.
614 if (nr_shown == 60) {
615 if (time_before(jiffies, resume)) {
621 "BUG: Bad page state: %lu messages suppressed\n",
628 resume = jiffies + 60 * HZ;
630 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
631 current->comm, page_to_pfn(page));
632 __dump_page(page, reason);
633 dump_page_owner(page);
638 /* Leave bad fields for debug, except PageBuddy could make trouble */
639 page_mapcount_reset(page); /* remove PageBuddy */
640 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
644 * Higher-order pages are called "compound pages". They are structured thusly:
646 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
648 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
649 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
651 * The first tail page's ->compound_dtor holds the offset in array of compound
652 * page destructors. See compound_page_dtors.
654 * The first tail page's ->compound_order holds the order of allocation.
655 * This usage means that zero-order pages may not be compound.
658 void free_compound_page(struct page *page)
660 mem_cgroup_uncharge(page);
661 __free_pages_ok(page, compound_order(page));
664 void prep_compound_page(struct page *page, unsigned int order)
667 int nr_pages = 1 << order;
669 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
670 set_compound_order(page, order);
672 for (i = 1; i < nr_pages; i++) {
673 struct page *p = page + i;
674 set_page_count(p, 0);
675 p->mapping = TAIL_MAPPING;
676 set_compound_head(p, page);
678 atomic_set(compound_mapcount_ptr(page), -1);
679 if (hpage_pincount_available(page))
680 atomic_set(compound_pincount_ptr(page), 0);
683 #ifdef CONFIG_DEBUG_PAGEALLOC
684 unsigned int _debug_guardpage_minorder;
686 bool _debug_pagealloc_enabled_early __read_mostly
687 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
688 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
689 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
690 EXPORT_SYMBOL(_debug_pagealloc_enabled);
692 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
694 static int __init early_debug_pagealloc(char *buf)
696 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
698 early_param("debug_pagealloc", early_debug_pagealloc);
700 void init_debug_pagealloc(void)
702 if (!debug_pagealloc_enabled())
705 static_branch_enable(&_debug_pagealloc_enabled);
707 if (!debug_guardpage_minorder())
710 static_branch_enable(&_debug_guardpage_enabled);
713 static int __init debug_guardpage_minorder_setup(char *buf)
717 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
718 pr_err("Bad debug_guardpage_minorder value\n");
721 _debug_guardpage_minorder = res;
722 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
725 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
727 static inline bool set_page_guard(struct zone *zone, struct page *page,
728 unsigned int order, int migratetype)
730 if (!debug_guardpage_enabled())
733 if (order >= debug_guardpage_minorder())
736 __SetPageGuard(page);
737 INIT_LIST_HEAD(&page->lru);
738 set_page_private(page, order);
739 /* Guard pages are not available for any usage */
740 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
745 static inline void clear_page_guard(struct zone *zone, struct page *page,
746 unsigned int order, int migratetype)
748 if (!debug_guardpage_enabled())
751 __ClearPageGuard(page);
753 set_page_private(page, 0);
754 if (!is_migrate_isolate(migratetype))
755 __mod_zone_freepage_state(zone, (1 << order), migratetype);
758 static inline bool set_page_guard(struct zone *zone, struct page *page,
759 unsigned int order, int migratetype) { return false; }
760 static inline void clear_page_guard(struct zone *zone, struct page *page,
761 unsigned int order, int migratetype) {}
764 static inline void set_page_order(struct page *page, unsigned int order)
766 set_page_private(page, order);
767 __SetPageBuddy(page);
771 * This function checks whether a page is free && is the buddy
772 * we can coalesce a page and its buddy if
773 * (a) the buddy is not in a hole (check before calling!) &&
774 * (b) the buddy is in the buddy system &&
775 * (c) a page and its buddy have the same order &&
776 * (d) a page and its buddy are in the same zone.
778 * For recording whether a page is in the buddy system, we set PageBuddy.
779 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
781 * For recording page's order, we use page_private(page).
783 static inline bool page_is_buddy(struct page *page, struct page *buddy,
786 if (!page_is_guard(buddy) && !PageBuddy(buddy))
789 if (page_order(buddy) != order)
793 * zone check is done late to avoid uselessly calculating
794 * zone/node ids for pages that could never merge.
796 if (page_zone_id(page) != page_zone_id(buddy))
799 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
804 #ifdef CONFIG_COMPACTION
805 static inline struct capture_control *task_capc(struct zone *zone)
807 struct capture_control *capc = current->capture_control;
809 return unlikely(capc) &&
810 !(current->flags & PF_KTHREAD) &&
812 capc->cc->zone == zone ? capc : NULL;
816 compaction_capture(struct capture_control *capc, struct page *page,
817 int order, int migratetype)
819 if (!capc || order != capc->cc->order)
822 /* Do not accidentally pollute CMA or isolated regions*/
823 if (is_migrate_cma(migratetype) ||
824 is_migrate_isolate(migratetype))
828 * Do not let lower order allocations polluate a movable pageblock.
829 * This might let an unmovable request use a reclaimable pageblock
830 * and vice-versa but no more than normal fallback logic which can
831 * have trouble finding a high-order free page.
833 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
841 static inline struct capture_control *task_capc(struct zone *zone)
847 compaction_capture(struct capture_control *capc, struct page *page,
848 int order, int migratetype)
852 #endif /* CONFIG_COMPACTION */
854 /* Used for pages not on another list */
855 static inline void add_to_free_list(struct page *page, struct zone *zone,
856 unsigned int order, int migratetype)
858 struct free_area *area = &zone->free_area[order];
860 list_add(&page->lru, &area->free_list[migratetype]);
864 /* Used for pages not on another list */
865 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
866 unsigned int order, int migratetype)
868 struct free_area *area = &zone->free_area[order];
870 list_add_tail(&page->lru, &area->free_list[migratetype]);
874 /* Used for pages which are on another list */
875 static inline void move_to_free_list(struct page *page, struct zone *zone,
876 unsigned int order, int migratetype)
878 struct free_area *area = &zone->free_area[order];
880 list_move(&page->lru, &area->free_list[migratetype]);
883 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
886 /* clear reported state and update reported page count */
887 if (page_reported(page))
888 __ClearPageReported(page);
890 list_del(&page->lru);
891 __ClearPageBuddy(page);
892 set_page_private(page, 0);
893 zone->free_area[order].nr_free--;
897 * If this is not the largest possible page, check if the buddy
898 * of the next-highest order is free. If it is, it's possible
899 * that pages are being freed that will coalesce soon. In case,
900 * that is happening, add the free page to the tail of the list
901 * so it's less likely to be used soon and more likely to be merged
902 * as a higher order page
905 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
906 struct page *page, unsigned int order)
908 struct page *higher_page, *higher_buddy;
909 unsigned long combined_pfn;
911 if (order >= MAX_ORDER - 2)
914 if (!pfn_valid_within(buddy_pfn))
917 combined_pfn = buddy_pfn & pfn;
918 higher_page = page + (combined_pfn - pfn);
919 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
920 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
922 return pfn_valid_within(buddy_pfn) &&
923 page_is_buddy(higher_page, higher_buddy, order + 1);
927 * Freeing function for a buddy system allocator.
929 * The concept of a buddy system is to maintain direct-mapped table
930 * (containing bit values) for memory blocks of various "orders".
931 * The bottom level table contains the map for the smallest allocatable
932 * units of memory (here, pages), and each level above it describes
933 * pairs of units from the levels below, hence, "buddies".
934 * At a high level, all that happens here is marking the table entry
935 * at the bottom level available, and propagating the changes upward
936 * as necessary, plus some accounting needed to play nicely with other
937 * parts of the VM system.
938 * At each level, we keep a list of pages, which are heads of continuous
939 * free pages of length of (1 << order) and marked with PageBuddy.
940 * Page's order is recorded in page_private(page) field.
941 * So when we are allocating or freeing one, we can derive the state of the
942 * other. That is, if we allocate a small block, and both were
943 * free, the remainder of the region must be split into blocks.
944 * If a block is freed, and its buddy is also free, then this
945 * triggers coalescing into a block of larger size.
950 static inline void __free_one_page(struct page *page,
952 struct zone *zone, unsigned int order,
953 int migratetype, bool report)
955 struct capture_control *capc = task_capc(zone);
956 unsigned long buddy_pfn;
957 unsigned long combined_pfn;
958 unsigned int max_order;
962 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
964 VM_BUG_ON(!zone_is_initialized(zone));
965 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
967 VM_BUG_ON(migratetype == -1);
968 if (likely(!is_migrate_isolate(migratetype)))
969 __mod_zone_freepage_state(zone, 1 << order, migratetype);
971 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
972 VM_BUG_ON_PAGE(bad_range(zone, page), page);
975 while (order < max_order - 1) {
976 if (compaction_capture(capc, page, order, migratetype)) {
977 __mod_zone_freepage_state(zone, -(1 << order),
981 buddy_pfn = __find_buddy_pfn(pfn, order);
982 buddy = page + (buddy_pfn - pfn);
984 if (!pfn_valid_within(buddy_pfn))
986 if (!page_is_buddy(page, buddy, order))
989 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
990 * merge with it and move up one order.
992 if (page_is_guard(buddy))
993 clear_page_guard(zone, buddy, order, migratetype);
995 del_page_from_free_list(buddy, zone, order);
996 combined_pfn = buddy_pfn & pfn;
997 page = page + (combined_pfn - pfn);
1001 if (max_order < MAX_ORDER) {
1002 /* If we are here, it means order is >= pageblock_order.
1003 * We want to prevent merge between freepages on isolate
1004 * pageblock and normal pageblock. Without this, pageblock
1005 * isolation could cause incorrect freepage or CMA accounting.
1007 * We don't want to hit this code for the more frequent
1008 * low-order merging.
1010 if (unlikely(has_isolate_pageblock(zone))) {
1013 buddy_pfn = __find_buddy_pfn(pfn, order);
1014 buddy = page + (buddy_pfn - pfn);
1015 buddy_mt = get_pageblock_migratetype(buddy);
1017 if (migratetype != buddy_mt
1018 && (is_migrate_isolate(migratetype) ||
1019 is_migrate_isolate(buddy_mt)))
1023 goto continue_merging;
1027 set_page_order(page, order);
1029 if (is_shuffle_order(order))
1030 to_tail = shuffle_pick_tail();
1032 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1035 add_to_free_list_tail(page, zone, order, migratetype);
1037 add_to_free_list(page, zone, order, migratetype);
1039 /* Notify page reporting subsystem of freed page */
1041 page_reporting_notify_free(order);
1045 * A bad page could be due to a number of fields. Instead of multiple branches,
1046 * try and check multiple fields with one check. The caller must do a detailed
1047 * check if necessary.
1049 static inline bool page_expected_state(struct page *page,
1050 unsigned long check_flags)
1052 if (unlikely(atomic_read(&page->_mapcount) != -1))
1055 if (unlikely((unsigned long)page->mapping |
1056 page_ref_count(page) |
1058 (unsigned long)page->mem_cgroup |
1060 (page->flags & check_flags)))
1066 static const char *page_bad_reason(struct page *page, unsigned long flags)
1068 const char *bad_reason = NULL;
1070 if (unlikely(atomic_read(&page->_mapcount) != -1))
1071 bad_reason = "nonzero mapcount";
1072 if (unlikely(page->mapping != NULL))
1073 bad_reason = "non-NULL mapping";
1074 if (unlikely(page_ref_count(page) != 0))
1075 bad_reason = "nonzero _refcount";
1076 if (unlikely(page->flags & flags)) {
1077 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1078 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1080 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1083 if (unlikely(page->mem_cgroup))
1084 bad_reason = "page still charged to cgroup";
1089 static void check_free_page_bad(struct page *page)
1092 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1095 static inline int check_free_page(struct page *page)
1097 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1100 /* Something has gone sideways, find it */
1101 check_free_page_bad(page);
1105 static int free_tail_pages_check(struct page *head_page, struct page *page)
1110 * We rely page->lru.next never has bit 0 set, unless the page
1111 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1113 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1115 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1119 switch (page - head_page) {
1121 /* the first tail page: ->mapping may be compound_mapcount() */
1122 if (unlikely(compound_mapcount(page))) {
1123 bad_page(page, "nonzero compound_mapcount");
1129 * the second tail page: ->mapping is
1130 * deferred_list.next -- ignore value.
1134 if (page->mapping != TAIL_MAPPING) {
1135 bad_page(page, "corrupted mapping in tail page");
1140 if (unlikely(!PageTail(page))) {
1141 bad_page(page, "PageTail not set");
1144 if (unlikely(compound_head(page) != head_page)) {
1145 bad_page(page, "compound_head not consistent");
1150 page->mapping = NULL;
1151 clear_compound_head(page);
1155 static void kernel_init_free_pages(struct page *page, int numpages)
1159 /* s390's use of memset() could override KASAN redzones. */
1160 kasan_disable_current();
1161 for (i = 0; i < numpages; i++)
1162 clear_highpage(page + i);
1163 kasan_enable_current();
1166 static __always_inline bool free_pages_prepare(struct page *page,
1167 unsigned int order, bool check_free)
1171 VM_BUG_ON_PAGE(PageTail(page), page);
1173 trace_mm_page_free(page, order);
1176 * Check tail pages before head page information is cleared to
1177 * avoid checking PageCompound for order-0 pages.
1179 if (unlikely(order)) {
1180 bool compound = PageCompound(page);
1183 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1186 ClearPageDoubleMap(page);
1187 for (i = 1; i < (1 << order); i++) {
1189 bad += free_tail_pages_check(page, page + i);
1190 if (unlikely(check_free_page(page + i))) {
1194 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1197 if (PageMappingFlags(page))
1198 page->mapping = NULL;
1199 if (memcg_kmem_enabled() && PageKmemcg(page))
1200 __memcg_kmem_uncharge_page(page, order);
1202 bad += check_free_page(page);
1206 page_cpupid_reset_last(page);
1207 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1208 reset_page_owner(page, order);
1210 if (!PageHighMem(page)) {
1211 debug_check_no_locks_freed(page_address(page),
1212 PAGE_SIZE << order);
1213 debug_check_no_obj_freed(page_address(page),
1214 PAGE_SIZE << order);
1216 if (want_init_on_free())
1217 kernel_init_free_pages(page, 1 << order);
1219 kernel_poison_pages(page, 1 << order, 0);
1221 * arch_free_page() can make the page's contents inaccessible. s390
1222 * does this. So nothing which can access the page's contents should
1223 * happen after this.
1225 arch_free_page(page, order);
1227 if (debug_pagealloc_enabled_static())
1228 kernel_map_pages(page, 1 << order, 0);
1230 kasan_free_nondeferred_pages(page, order);
1235 #ifdef CONFIG_DEBUG_VM
1237 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1238 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1239 * moved from pcp lists to free lists.
1241 static bool free_pcp_prepare(struct page *page)
1243 return free_pages_prepare(page, 0, true);
1246 static bool bulkfree_pcp_prepare(struct page *page)
1248 if (debug_pagealloc_enabled_static())
1249 return check_free_page(page);
1255 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1256 * moving from pcp lists to free list in order to reduce overhead. With
1257 * debug_pagealloc enabled, they are checked also immediately when being freed
1260 static bool free_pcp_prepare(struct page *page)
1262 if (debug_pagealloc_enabled_static())
1263 return free_pages_prepare(page, 0, true);
1265 return free_pages_prepare(page, 0, false);
1268 static bool bulkfree_pcp_prepare(struct page *page)
1270 return check_free_page(page);
1272 #endif /* CONFIG_DEBUG_VM */
1274 static inline void prefetch_buddy(struct page *page)
1276 unsigned long pfn = page_to_pfn(page);
1277 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1278 struct page *buddy = page + (buddy_pfn - pfn);
1284 * Frees a number of pages from the PCP lists
1285 * Assumes all pages on list are in same zone, and of same order.
1286 * count is the number of pages to free.
1288 * If the zone was previously in an "all pages pinned" state then look to
1289 * see if this freeing clears that state.
1291 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1292 * pinned" detection logic.
1294 static void free_pcppages_bulk(struct zone *zone, int count,
1295 struct per_cpu_pages *pcp)
1297 int migratetype = 0;
1299 int prefetch_nr = 0;
1300 bool isolated_pageblocks;
1301 struct page *page, *tmp;
1305 struct list_head *list;
1308 * Remove pages from lists in a round-robin fashion. A
1309 * batch_free count is maintained that is incremented when an
1310 * empty list is encountered. This is so more pages are freed
1311 * off fuller lists instead of spinning excessively around empty
1316 if (++migratetype == MIGRATE_PCPTYPES)
1318 list = &pcp->lists[migratetype];
1319 } while (list_empty(list));
1321 /* This is the only non-empty list. Free them all. */
1322 if (batch_free == MIGRATE_PCPTYPES)
1326 page = list_last_entry(list, struct page, lru);
1327 /* must delete to avoid corrupting pcp list */
1328 list_del(&page->lru);
1331 if (bulkfree_pcp_prepare(page))
1334 list_add_tail(&page->lru, &head);
1337 * We are going to put the page back to the global
1338 * pool, prefetch its buddy to speed up later access
1339 * under zone->lock. It is believed the overhead of
1340 * an additional test and calculating buddy_pfn here
1341 * can be offset by reduced memory latency later. To
1342 * avoid excessive prefetching due to large count, only
1343 * prefetch buddy for the first pcp->batch nr of pages.
1345 if (prefetch_nr++ < pcp->batch)
1346 prefetch_buddy(page);
1347 } while (--count && --batch_free && !list_empty(list));
1350 spin_lock(&zone->lock);
1351 isolated_pageblocks = has_isolate_pageblock(zone);
1354 * Use safe version since after __free_one_page(),
1355 * page->lru.next will not point to original list.
1357 list_for_each_entry_safe(page, tmp, &head, lru) {
1358 int mt = get_pcppage_migratetype(page);
1359 /* MIGRATE_ISOLATE page should not go to pcplists */
1360 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1361 /* Pageblock could have been isolated meanwhile */
1362 if (unlikely(isolated_pageblocks))
1363 mt = get_pageblock_migratetype(page);
1365 __free_one_page(page, page_to_pfn(page), zone, 0, mt, true);
1366 trace_mm_page_pcpu_drain(page, 0, mt);
1368 spin_unlock(&zone->lock);
1371 static void free_one_page(struct zone *zone,
1372 struct page *page, unsigned long pfn,
1376 spin_lock(&zone->lock);
1377 if (unlikely(has_isolate_pageblock(zone) ||
1378 is_migrate_isolate(migratetype))) {
1379 migratetype = get_pfnblock_migratetype(page, pfn);
1381 __free_one_page(page, pfn, zone, order, migratetype, true);
1382 spin_unlock(&zone->lock);
1385 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1386 unsigned long zone, int nid)
1388 mm_zero_struct_page(page);
1389 set_page_links(page, zone, nid, pfn);
1390 init_page_count(page);
1391 page_mapcount_reset(page);
1392 page_cpupid_reset_last(page);
1393 page_kasan_tag_reset(page);
1395 INIT_LIST_HEAD(&page->lru);
1396 #ifdef WANT_PAGE_VIRTUAL
1397 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1398 if (!is_highmem_idx(zone))
1399 set_page_address(page, __va(pfn << PAGE_SHIFT));
1403 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1404 static void __meminit init_reserved_page(unsigned long pfn)
1409 if (!early_page_uninitialised(pfn))
1412 nid = early_pfn_to_nid(pfn);
1413 pgdat = NODE_DATA(nid);
1415 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1416 struct zone *zone = &pgdat->node_zones[zid];
1418 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1421 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1424 static inline void init_reserved_page(unsigned long pfn)
1427 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1430 * Initialised pages do not have PageReserved set. This function is
1431 * called for each range allocated by the bootmem allocator and
1432 * marks the pages PageReserved. The remaining valid pages are later
1433 * sent to the buddy page allocator.
1435 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1437 unsigned long start_pfn = PFN_DOWN(start);
1438 unsigned long end_pfn = PFN_UP(end);
1440 for (; start_pfn < end_pfn; start_pfn++) {
1441 if (pfn_valid(start_pfn)) {
1442 struct page *page = pfn_to_page(start_pfn);
1444 init_reserved_page(start_pfn);
1446 /* Avoid false-positive PageTail() */
1447 INIT_LIST_HEAD(&page->lru);
1450 * no need for atomic set_bit because the struct
1451 * page is not visible yet so nobody should
1454 __SetPageReserved(page);
1459 static void __free_pages_ok(struct page *page, unsigned int order)
1461 unsigned long flags;
1463 unsigned long pfn = page_to_pfn(page);
1465 if (!free_pages_prepare(page, order, true))
1468 migratetype = get_pfnblock_migratetype(page, pfn);
1469 local_irq_save(flags);
1470 __count_vm_events(PGFREE, 1 << order);
1471 free_one_page(page_zone(page), page, pfn, order, migratetype);
1472 local_irq_restore(flags);
1475 void __free_pages_core(struct page *page, unsigned int order)
1477 unsigned int nr_pages = 1 << order;
1478 struct page *p = page;
1482 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1484 __ClearPageReserved(p);
1485 set_page_count(p, 0);
1487 __ClearPageReserved(p);
1488 set_page_count(p, 0);
1490 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1491 set_page_refcounted(page);
1492 __free_pages(page, order);
1495 #ifdef CONFIG_NEED_MULTIPLE_NODES
1497 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1499 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
1502 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1504 int __meminit __early_pfn_to_nid(unsigned long pfn,
1505 struct mminit_pfnnid_cache *state)
1507 unsigned long start_pfn, end_pfn;
1510 if (state->last_start <= pfn && pfn < state->last_end)
1511 return state->last_nid;
1513 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1514 if (nid != NUMA_NO_NODE) {
1515 state->last_start = start_pfn;
1516 state->last_end = end_pfn;
1517 state->last_nid = nid;
1522 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
1524 int __meminit early_pfn_to_nid(unsigned long pfn)
1526 static DEFINE_SPINLOCK(early_pfn_lock);
1529 spin_lock(&early_pfn_lock);
1530 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1532 nid = first_online_node;
1533 spin_unlock(&early_pfn_lock);
1537 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1539 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1542 if (early_page_uninitialised(pfn))
1544 __free_pages_core(page, order);
1548 * Check that the whole (or subset of) a pageblock given by the interval of
1549 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1550 * with the migration of free compaction scanner. The scanners then need to
1551 * use only pfn_valid_within() check for arches that allow holes within
1554 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1556 * It's possible on some configurations to have a setup like node0 node1 node0
1557 * i.e. it's possible that all pages within a zones range of pages do not
1558 * belong to a single zone. We assume that a border between node0 and node1
1559 * can occur within a single pageblock, but not a node0 node1 node0
1560 * interleaving within a single pageblock. It is therefore sufficient to check
1561 * the first and last page of a pageblock and avoid checking each individual
1562 * page in a pageblock.
1564 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1565 unsigned long end_pfn, struct zone *zone)
1567 struct page *start_page;
1568 struct page *end_page;
1570 /* end_pfn is one past the range we are checking */
1573 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1576 start_page = pfn_to_online_page(start_pfn);
1580 if (page_zone(start_page) != zone)
1583 end_page = pfn_to_page(end_pfn);
1585 /* This gives a shorter code than deriving page_zone(end_page) */
1586 if (page_zone_id(start_page) != page_zone_id(end_page))
1592 void set_zone_contiguous(struct zone *zone)
1594 unsigned long block_start_pfn = zone->zone_start_pfn;
1595 unsigned long block_end_pfn;
1597 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1598 for (; block_start_pfn < zone_end_pfn(zone);
1599 block_start_pfn = block_end_pfn,
1600 block_end_pfn += pageblock_nr_pages) {
1602 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1604 if (!__pageblock_pfn_to_page(block_start_pfn,
1605 block_end_pfn, zone))
1610 /* We confirm that there is no hole */
1611 zone->contiguous = true;
1614 void clear_zone_contiguous(struct zone *zone)
1616 zone->contiguous = false;
1619 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1620 static void __init deferred_free_range(unsigned long pfn,
1621 unsigned long nr_pages)
1629 page = pfn_to_page(pfn);
1631 /* Free a large naturally-aligned chunk if possible */
1632 if (nr_pages == pageblock_nr_pages &&
1633 (pfn & (pageblock_nr_pages - 1)) == 0) {
1634 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1635 __free_pages_core(page, pageblock_order);
1639 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1640 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1641 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1642 __free_pages_core(page, 0);
1646 /* Completion tracking for deferred_init_memmap() threads */
1647 static atomic_t pgdat_init_n_undone __initdata;
1648 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1650 static inline void __init pgdat_init_report_one_done(void)
1652 if (atomic_dec_and_test(&pgdat_init_n_undone))
1653 complete(&pgdat_init_all_done_comp);
1657 * Returns true if page needs to be initialized or freed to buddy allocator.
1659 * First we check if pfn is valid on architectures where it is possible to have
1660 * holes within pageblock_nr_pages. On systems where it is not possible, this
1661 * function is optimized out.
1663 * Then, we check if a current large page is valid by only checking the validity
1666 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1668 if (!pfn_valid_within(pfn))
1670 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1676 * Free pages to buddy allocator. Try to free aligned pages in
1677 * pageblock_nr_pages sizes.
1679 static void __init deferred_free_pages(unsigned long pfn,
1680 unsigned long end_pfn)
1682 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1683 unsigned long nr_free = 0;
1685 for (; pfn < end_pfn; pfn++) {
1686 if (!deferred_pfn_valid(pfn)) {
1687 deferred_free_range(pfn - nr_free, nr_free);
1689 } else if (!(pfn & nr_pgmask)) {
1690 deferred_free_range(pfn - nr_free, nr_free);
1696 /* Free the last block of pages to allocator */
1697 deferred_free_range(pfn - nr_free, nr_free);
1701 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1702 * by performing it only once every pageblock_nr_pages.
1703 * Return number of pages initialized.
1705 static unsigned long __init deferred_init_pages(struct zone *zone,
1707 unsigned long end_pfn)
1709 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1710 int nid = zone_to_nid(zone);
1711 unsigned long nr_pages = 0;
1712 int zid = zone_idx(zone);
1713 struct page *page = NULL;
1715 for (; pfn < end_pfn; pfn++) {
1716 if (!deferred_pfn_valid(pfn)) {
1719 } else if (!page || !(pfn & nr_pgmask)) {
1720 page = pfn_to_page(pfn);
1724 __init_single_page(page, pfn, zid, nid);
1731 * This function is meant to pre-load the iterator for the zone init.
1732 * Specifically it walks through the ranges until we are caught up to the
1733 * first_init_pfn value and exits there. If we never encounter the value we
1734 * return false indicating there are no valid ranges left.
1737 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1738 unsigned long *spfn, unsigned long *epfn,
1739 unsigned long first_init_pfn)
1744 * Start out by walking through the ranges in this zone that have
1745 * already been initialized. We don't need to do anything with them
1746 * so we just need to flush them out of the system.
1748 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1749 if (*epfn <= first_init_pfn)
1751 if (*spfn < first_init_pfn)
1752 *spfn = first_init_pfn;
1761 * Initialize and free pages. We do it in two loops: first we initialize
1762 * struct page, then free to buddy allocator, because while we are
1763 * freeing pages we can access pages that are ahead (computing buddy
1764 * page in __free_one_page()).
1766 * In order to try and keep some memory in the cache we have the loop
1767 * broken along max page order boundaries. This way we will not cause
1768 * any issues with the buddy page computation.
1770 static unsigned long __init
1771 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1772 unsigned long *end_pfn)
1774 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1775 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1776 unsigned long nr_pages = 0;
1779 /* First we loop through and initialize the page values */
1780 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1783 if (mo_pfn <= *start_pfn)
1786 t = min(mo_pfn, *end_pfn);
1787 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1789 if (mo_pfn < *end_pfn) {
1790 *start_pfn = mo_pfn;
1795 /* Reset values and now loop through freeing pages as needed */
1798 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1804 t = min(mo_pfn, epfn);
1805 deferred_free_pages(spfn, t);
1815 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1818 unsigned long spfn, epfn;
1819 struct zone *zone = arg;
1822 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1825 * Initialize and free pages in MAX_ORDER sized increments so that we
1826 * can avoid introducing any issues with the buddy allocator.
1828 while (spfn < end_pfn) {
1829 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1834 /* An arch may override for more concurrency. */
1836 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1841 /* Initialise remaining memory on a node */
1842 static int __init deferred_init_memmap(void *data)
1844 pg_data_t *pgdat = data;
1845 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1846 unsigned long spfn = 0, epfn = 0;
1847 unsigned long first_init_pfn, flags;
1848 unsigned long start = jiffies;
1850 int zid, max_threads;
1853 /* Bind memory initialisation thread to a local node if possible */
1854 if (!cpumask_empty(cpumask))
1855 set_cpus_allowed_ptr(current, cpumask);
1857 pgdat_resize_lock(pgdat, &flags);
1858 first_init_pfn = pgdat->first_deferred_pfn;
1859 if (first_init_pfn == ULONG_MAX) {
1860 pgdat_resize_unlock(pgdat, &flags);
1861 pgdat_init_report_one_done();
1865 /* Sanity check boundaries */
1866 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1867 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1868 pgdat->first_deferred_pfn = ULONG_MAX;
1871 * Once we unlock here, the zone cannot be grown anymore, thus if an
1872 * interrupt thread must allocate this early in boot, zone must be
1873 * pre-grown prior to start of deferred page initialization.
1875 pgdat_resize_unlock(pgdat, &flags);
1877 /* Only the highest zone is deferred so find it */
1878 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1879 zone = pgdat->node_zones + zid;
1880 if (first_init_pfn < zone_end_pfn(zone))
1884 /* If the zone is empty somebody else may have cleared out the zone */
1885 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1889 max_threads = deferred_page_init_max_threads(cpumask);
1891 while (spfn < epfn) {
1892 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
1893 struct padata_mt_job job = {
1894 .thread_fn = deferred_init_memmap_chunk,
1897 .size = epfn_align - spfn,
1898 .align = PAGES_PER_SECTION,
1899 .min_chunk = PAGES_PER_SECTION,
1900 .max_threads = max_threads,
1903 padata_do_multithreaded(&job);
1904 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1908 /* Sanity check that the next zone really is unpopulated */
1909 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1911 pr_info("node %d deferred pages initialised in %ums\n",
1912 pgdat->node_id, jiffies_to_msecs(jiffies - start));
1914 pgdat_init_report_one_done();
1919 * If this zone has deferred pages, try to grow it by initializing enough
1920 * deferred pages to satisfy the allocation specified by order, rounded up to
1921 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1922 * of SECTION_SIZE bytes by initializing struct pages in increments of
1923 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1925 * Return true when zone was grown, otherwise return false. We return true even
1926 * when we grow less than requested, to let the caller decide if there are
1927 * enough pages to satisfy the allocation.
1929 * Note: We use noinline because this function is needed only during boot, and
1930 * it is called from a __ref function _deferred_grow_zone. This way we are
1931 * making sure that it is not inlined into permanent text section.
1933 static noinline bool __init
1934 deferred_grow_zone(struct zone *zone, unsigned int order)
1936 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1937 pg_data_t *pgdat = zone->zone_pgdat;
1938 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1939 unsigned long spfn, epfn, flags;
1940 unsigned long nr_pages = 0;
1943 /* Only the last zone may have deferred pages */
1944 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1947 pgdat_resize_lock(pgdat, &flags);
1950 * If someone grew this zone while we were waiting for spinlock, return
1951 * true, as there might be enough pages already.
1953 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1954 pgdat_resize_unlock(pgdat, &flags);
1958 /* If the zone is empty somebody else may have cleared out the zone */
1959 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1960 first_deferred_pfn)) {
1961 pgdat->first_deferred_pfn = ULONG_MAX;
1962 pgdat_resize_unlock(pgdat, &flags);
1963 /* Retry only once. */
1964 return first_deferred_pfn != ULONG_MAX;
1968 * Initialize and free pages in MAX_ORDER sized increments so
1969 * that we can avoid introducing any issues with the buddy
1972 while (spfn < epfn) {
1973 /* update our first deferred PFN for this section */
1974 first_deferred_pfn = spfn;
1976 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1977 touch_nmi_watchdog();
1979 /* We should only stop along section boundaries */
1980 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
1983 /* If our quota has been met we can stop here */
1984 if (nr_pages >= nr_pages_needed)
1988 pgdat->first_deferred_pfn = spfn;
1989 pgdat_resize_unlock(pgdat, &flags);
1991 return nr_pages > 0;
1995 * deferred_grow_zone() is __init, but it is called from
1996 * get_page_from_freelist() during early boot until deferred_pages permanently
1997 * disables this call. This is why we have refdata wrapper to avoid warning,
1998 * and to ensure that the function body gets unloaded.
2001 _deferred_grow_zone(struct zone *zone, unsigned int order)
2003 return deferred_grow_zone(zone, order);
2006 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2008 void __init page_alloc_init_late(void)
2013 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2015 /* There will be num_node_state(N_MEMORY) threads */
2016 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2017 for_each_node_state(nid, N_MEMORY) {
2018 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2021 /* Block until all are initialised */
2022 wait_for_completion(&pgdat_init_all_done_comp);
2025 * The number of managed pages has changed due to the initialisation
2026 * so the pcpu batch and high limits needs to be updated or the limits
2027 * will be artificially small.
2029 for_each_populated_zone(zone)
2030 zone_pcp_update(zone);
2033 * We initialized the rest of the deferred pages. Permanently disable
2034 * on-demand struct page initialization.
2036 static_branch_disable(&deferred_pages);
2038 /* Reinit limits that are based on free pages after the kernel is up */
2039 files_maxfiles_init();
2042 /* Discard memblock private memory */
2045 for_each_node_state(nid, N_MEMORY)
2046 shuffle_free_memory(NODE_DATA(nid));
2048 for_each_populated_zone(zone)
2049 set_zone_contiguous(zone);
2053 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2054 void __init init_cma_reserved_pageblock(struct page *page)
2056 unsigned i = pageblock_nr_pages;
2057 struct page *p = page;
2060 __ClearPageReserved(p);
2061 set_page_count(p, 0);
2064 set_pageblock_migratetype(page, MIGRATE_CMA);
2066 if (pageblock_order >= MAX_ORDER) {
2067 i = pageblock_nr_pages;
2070 set_page_refcounted(p);
2071 __free_pages(p, MAX_ORDER - 1);
2072 p += MAX_ORDER_NR_PAGES;
2073 } while (i -= MAX_ORDER_NR_PAGES);
2075 set_page_refcounted(page);
2076 __free_pages(page, pageblock_order);
2079 adjust_managed_page_count(page, pageblock_nr_pages);
2084 * The order of subdivision here is critical for the IO subsystem.
2085 * Please do not alter this order without good reasons and regression
2086 * testing. Specifically, as large blocks of memory are subdivided,
2087 * the order in which smaller blocks are delivered depends on the order
2088 * they're subdivided in this function. This is the primary factor
2089 * influencing the order in which pages are delivered to the IO
2090 * subsystem according to empirical testing, and this is also justified
2091 * by considering the behavior of a buddy system containing a single
2092 * large block of memory acted on by a series of small allocations.
2093 * This behavior is a critical factor in sglist merging's success.
2097 static inline void expand(struct zone *zone, struct page *page,
2098 int low, int high, int migratetype)
2100 unsigned long size = 1 << high;
2102 while (high > low) {
2105 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2108 * Mark as guard pages (or page), that will allow to
2109 * merge back to allocator when buddy will be freed.
2110 * Corresponding page table entries will not be touched,
2111 * pages will stay not present in virtual address space
2113 if (set_page_guard(zone, &page[size], high, migratetype))
2116 add_to_free_list(&page[size], zone, high, migratetype);
2117 set_page_order(&page[size], high);
2121 static void check_new_page_bad(struct page *page)
2123 if (unlikely(page->flags & __PG_HWPOISON)) {
2124 /* Don't complain about hwpoisoned pages */
2125 page_mapcount_reset(page); /* remove PageBuddy */
2130 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2134 * This page is about to be returned from the page allocator
2136 static inline int check_new_page(struct page *page)
2138 if (likely(page_expected_state(page,
2139 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2142 check_new_page_bad(page);
2146 static inline bool free_pages_prezeroed(void)
2148 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2149 page_poisoning_enabled()) || want_init_on_free();
2152 #ifdef CONFIG_DEBUG_VM
2154 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2155 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2156 * also checked when pcp lists are refilled from the free lists.
2158 static inline bool check_pcp_refill(struct page *page)
2160 if (debug_pagealloc_enabled_static())
2161 return check_new_page(page);
2166 static inline bool check_new_pcp(struct page *page)
2168 return check_new_page(page);
2172 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2173 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2174 * enabled, they are also checked when being allocated from the pcp lists.
2176 static inline bool check_pcp_refill(struct page *page)
2178 return check_new_page(page);
2180 static inline bool check_new_pcp(struct page *page)
2182 if (debug_pagealloc_enabled_static())
2183 return check_new_page(page);
2187 #endif /* CONFIG_DEBUG_VM */
2189 static bool check_new_pages(struct page *page, unsigned int order)
2192 for (i = 0; i < (1 << order); i++) {
2193 struct page *p = page + i;
2195 if (unlikely(check_new_page(p)))
2202 inline void post_alloc_hook(struct page *page, unsigned int order,
2205 set_page_private(page, 0);
2206 set_page_refcounted(page);
2208 arch_alloc_page(page, order);
2209 if (debug_pagealloc_enabled_static())
2210 kernel_map_pages(page, 1 << order, 1);
2211 kasan_alloc_pages(page, order);
2212 kernel_poison_pages(page, 1 << order, 1);
2213 set_page_owner(page, order, gfp_flags);
2216 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2217 unsigned int alloc_flags)
2219 post_alloc_hook(page, order, gfp_flags);
2221 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2222 kernel_init_free_pages(page, 1 << order);
2224 if (order && (gfp_flags & __GFP_COMP))
2225 prep_compound_page(page, order);
2228 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2229 * allocate the page. The expectation is that the caller is taking
2230 * steps that will free more memory. The caller should avoid the page
2231 * being used for !PFMEMALLOC purposes.
2233 if (alloc_flags & ALLOC_NO_WATERMARKS)
2234 set_page_pfmemalloc(page);
2236 clear_page_pfmemalloc(page);
2240 * Go through the free lists for the given migratetype and remove
2241 * the smallest available page from the freelists
2243 static __always_inline
2244 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2247 unsigned int current_order;
2248 struct free_area *area;
2251 /* Find a page of the appropriate size in the preferred list */
2252 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2253 area = &(zone->free_area[current_order]);
2254 page = get_page_from_free_area(area, migratetype);
2257 del_page_from_free_list(page, zone, current_order);
2258 expand(zone, page, order, current_order, migratetype);
2259 set_pcppage_migratetype(page, migratetype);
2268 * This array describes the order lists are fallen back to when
2269 * the free lists for the desirable migrate type are depleted
2271 static int fallbacks[MIGRATE_TYPES][3] = {
2272 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2273 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2274 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2276 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2278 #ifdef CONFIG_MEMORY_ISOLATION
2279 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2284 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2287 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2290 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2291 unsigned int order) { return NULL; }
2295 * Move the free pages in a range to the free lists of the requested type.
2296 * Note that start_page and end_pages are not aligned on a pageblock
2297 * boundary. If alignment is required, use move_freepages_block()
2299 static int move_freepages(struct zone *zone,
2300 struct page *start_page, struct page *end_page,
2301 int migratetype, int *num_movable)
2305 int pages_moved = 0;
2307 for (page = start_page; page <= end_page;) {
2308 if (!pfn_valid_within(page_to_pfn(page))) {
2313 if (!PageBuddy(page)) {
2315 * We assume that pages that could be isolated for
2316 * migration are movable. But we don't actually try
2317 * isolating, as that would be expensive.
2320 (PageLRU(page) || __PageMovable(page)))
2327 /* Make sure we are not inadvertently changing nodes */
2328 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2329 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2331 order = page_order(page);
2332 move_to_free_list(page, zone, order, migratetype);
2334 pages_moved += 1 << order;
2340 int move_freepages_block(struct zone *zone, struct page *page,
2341 int migratetype, int *num_movable)
2343 unsigned long start_pfn, end_pfn;
2344 struct page *start_page, *end_page;
2349 start_pfn = page_to_pfn(page);
2350 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2351 start_page = pfn_to_page(start_pfn);
2352 end_page = start_page + pageblock_nr_pages - 1;
2353 end_pfn = start_pfn + pageblock_nr_pages - 1;
2355 /* Do not cross zone boundaries */
2356 if (!zone_spans_pfn(zone, start_pfn))
2358 if (!zone_spans_pfn(zone, end_pfn))
2361 return move_freepages(zone, start_page, end_page, migratetype,
2365 static void change_pageblock_range(struct page *pageblock_page,
2366 int start_order, int migratetype)
2368 int nr_pageblocks = 1 << (start_order - pageblock_order);
2370 while (nr_pageblocks--) {
2371 set_pageblock_migratetype(pageblock_page, migratetype);
2372 pageblock_page += pageblock_nr_pages;
2377 * When we are falling back to another migratetype during allocation, try to
2378 * steal extra free pages from the same pageblocks to satisfy further
2379 * allocations, instead of polluting multiple pageblocks.
2381 * If we are stealing a relatively large buddy page, it is likely there will
2382 * be more free pages in the pageblock, so try to steal them all. For
2383 * reclaimable and unmovable allocations, we steal regardless of page size,
2384 * as fragmentation caused by those allocations polluting movable pageblocks
2385 * is worse than movable allocations stealing from unmovable and reclaimable
2388 static bool can_steal_fallback(unsigned int order, int start_mt)
2391 * Leaving this order check is intended, although there is
2392 * relaxed order check in next check. The reason is that
2393 * we can actually steal whole pageblock if this condition met,
2394 * but, below check doesn't guarantee it and that is just heuristic
2395 * so could be changed anytime.
2397 if (order >= pageblock_order)
2400 if (order >= pageblock_order / 2 ||
2401 start_mt == MIGRATE_RECLAIMABLE ||
2402 start_mt == MIGRATE_UNMOVABLE ||
2403 page_group_by_mobility_disabled)
2409 static inline void boost_watermark(struct zone *zone)
2411 unsigned long max_boost;
2413 if (!watermark_boost_factor)
2416 * Don't bother in zones that are unlikely to produce results.
2417 * On small machines, including kdump capture kernels running
2418 * in a small area, boosting the watermark can cause an out of
2419 * memory situation immediately.
2421 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2424 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2425 watermark_boost_factor, 10000);
2428 * high watermark may be uninitialised if fragmentation occurs
2429 * very early in boot so do not boost. We do not fall
2430 * through and boost by pageblock_nr_pages as failing
2431 * allocations that early means that reclaim is not going
2432 * to help and it may even be impossible to reclaim the
2433 * boosted watermark resulting in a hang.
2438 max_boost = max(pageblock_nr_pages, max_boost);
2440 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2445 * This function implements actual steal behaviour. If order is large enough,
2446 * we can steal whole pageblock. If not, we first move freepages in this
2447 * pageblock to our migratetype and determine how many already-allocated pages
2448 * are there in the pageblock with a compatible migratetype. If at least half
2449 * of pages are free or compatible, we can change migratetype of the pageblock
2450 * itself, so pages freed in the future will be put on the correct free list.
2452 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2453 unsigned int alloc_flags, int start_type, bool whole_block)
2455 unsigned int current_order = page_order(page);
2456 int free_pages, movable_pages, alike_pages;
2459 old_block_type = get_pageblock_migratetype(page);
2462 * This can happen due to races and we want to prevent broken
2463 * highatomic accounting.
2465 if (is_migrate_highatomic(old_block_type))
2468 /* Take ownership for orders >= pageblock_order */
2469 if (current_order >= pageblock_order) {
2470 change_pageblock_range(page, current_order, start_type);
2475 * Boost watermarks to increase reclaim pressure to reduce the
2476 * likelihood of future fallbacks. Wake kswapd now as the node
2477 * may be balanced overall and kswapd will not wake naturally.
2479 boost_watermark(zone);
2480 if (alloc_flags & ALLOC_KSWAPD)
2481 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2483 /* We are not allowed to try stealing from the whole block */
2487 free_pages = move_freepages_block(zone, page, start_type,
2490 * Determine how many pages are compatible with our allocation.
2491 * For movable allocation, it's the number of movable pages which
2492 * we just obtained. For other types it's a bit more tricky.
2494 if (start_type == MIGRATE_MOVABLE) {
2495 alike_pages = movable_pages;
2498 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2499 * to MOVABLE pageblock, consider all non-movable pages as
2500 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2501 * vice versa, be conservative since we can't distinguish the
2502 * exact migratetype of non-movable pages.
2504 if (old_block_type == MIGRATE_MOVABLE)
2505 alike_pages = pageblock_nr_pages
2506 - (free_pages + movable_pages);
2511 /* moving whole block can fail due to zone boundary conditions */
2516 * If a sufficient number of pages in the block are either free or of
2517 * comparable migratability as our allocation, claim the whole block.
2519 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2520 page_group_by_mobility_disabled)
2521 set_pageblock_migratetype(page, start_type);
2526 move_to_free_list(page, zone, current_order, start_type);
2530 * Check whether there is a suitable fallback freepage with requested order.
2531 * If only_stealable is true, this function returns fallback_mt only if
2532 * we can steal other freepages all together. This would help to reduce
2533 * fragmentation due to mixed migratetype pages in one pageblock.
2535 int find_suitable_fallback(struct free_area *area, unsigned int order,
2536 int migratetype, bool only_stealable, bool *can_steal)
2541 if (area->nr_free == 0)
2546 fallback_mt = fallbacks[migratetype][i];
2547 if (fallback_mt == MIGRATE_TYPES)
2550 if (free_area_empty(area, fallback_mt))
2553 if (can_steal_fallback(order, migratetype))
2556 if (!only_stealable)
2567 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2568 * there are no empty page blocks that contain a page with a suitable order
2570 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2571 unsigned int alloc_order)
2574 unsigned long max_managed, flags;
2577 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2578 * Check is race-prone but harmless.
2580 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2581 if (zone->nr_reserved_highatomic >= max_managed)
2584 spin_lock_irqsave(&zone->lock, flags);
2586 /* Recheck the nr_reserved_highatomic limit under the lock */
2587 if (zone->nr_reserved_highatomic >= max_managed)
2591 mt = get_pageblock_migratetype(page);
2592 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2593 && !is_migrate_cma(mt)) {
2594 zone->nr_reserved_highatomic += pageblock_nr_pages;
2595 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2596 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2600 spin_unlock_irqrestore(&zone->lock, flags);
2604 * Used when an allocation is about to fail under memory pressure. This
2605 * potentially hurts the reliability of high-order allocations when under
2606 * intense memory pressure but failed atomic allocations should be easier
2607 * to recover from than an OOM.
2609 * If @force is true, try to unreserve a pageblock even though highatomic
2610 * pageblock is exhausted.
2612 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2615 struct zonelist *zonelist = ac->zonelist;
2616 unsigned long flags;
2623 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2626 * Preserve at least one pageblock unless memory pressure
2629 if (!force && zone->nr_reserved_highatomic <=
2633 spin_lock_irqsave(&zone->lock, flags);
2634 for (order = 0; order < MAX_ORDER; order++) {
2635 struct free_area *area = &(zone->free_area[order]);
2637 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2642 * In page freeing path, migratetype change is racy so
2643 * we can counter several free pages in a pageblock
2644 * in this loop althoug we changed the pageblock type
2645 * from highatomic to ac->migratetype. So we should
2646 * adjust the count once.
2648 if (is_migrate_highatomic_page(page)) {
2650 * It should never happen but changes to
2651 * locking could inadvertently allow a per-cpu
2652 * drain to add pages to MIGRATE_HIGHATOMIC
2653 * while unreserving so be safe and watch for
2656 zone->nr_reserved_highatomic -= min(
2658 zone->nr_reserved_highatomic);
2662 * Convert to ac->migratetype and avoid the normal
2663 * pageblock stealing heuristics. Minimally, the caller
2664 * is doing the work and needs the pages. More
2665 * importantly, if the block was always converted to
2666 * MIGRATE_UNMOVABLE or another type then the number
2667 * of pageblocks that cannot be completely freed
2670 set_pageblock_migratetype(page, ac->migratetype);
2671 ret = move_freepages_block(zone, page, ac->migratetype,
2674 spin_unlock_irqrestore(&zone->lock, flags);
2678 spin_unlock_irqrestore(&zone->lock, flags);
2685 * Try finding a free buddy page on the fallback list and put it on the free
2686 * list of requested migratetype, possibly along with other pages from the same
2687 * block, depending on fragmentation avoidance heuristics. Returns true if
2688 * fallback was found so that __rmqueue_smallest() can grab it.
2690 * The use of signed ints for order and current_order is a deliberate
2691 * deviation from the rest of this file, to make the for loop
2692 * condition simpler.
2694 static __always_inline bool
2695 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2696 unsigned int alloc_flags)
2698 struct free_area *area;
2700 int min_order = order;
2706 * Do not steal pages from freelists belonging to other pageblocks
2707 * i.e. orders < pageblock_order. If there are no local zones free,
2708 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2710 if (alloc_flags & ALLOC_NOFRAGMENT)
2711 min_order = pageblock_order;
2714 * Find the largest available free page in the other list. This roughly
2715 * approximates finding the pageblock with the most free pages, which
2716 * would be too costly to do exactly.
2718 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2720 area = &(zone->free_area[current_order]);
2721 fallback_mt = find_suitable_fallback(area, current_order,
2722 start_migratetype, false, &can_steal);
2723 if (fallback_mt == -1)
2727 * We cannot steal all free pages from the pageblock and the
2728 * requested migratetype is movable. In that case it's better to
2729 * steal and split the smallest available page instead of the
2730 * largest available page, because even if the next movable
2731 * allocation falls back into a different pageblock than this
2732 * one, it won't cause permanent fragmentation.
2734 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2735 && current_order > order)
2744 for (current_order = order; current_order < MAX_ORDER;
2746 area = &(zone->free_area[current_order]);
2747 fallback_mt = find_suitable_fallback(area, current_order,
2748 start_migratetype, false, &can_steal);
2749 if (fallback_mt != -1)
2754 * This should not happen - we already found a suitable fallback
2755 * when looking for the largest page.
2757 VM_BUG_ON(current_order == MAX_ORDER);
2760 page = get_page_from_free_area(area, fallback_mt);
2762 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2765 trace_mm_page_alloc_extfrag(page, order, current_order,
2766 start_migratetype, fallback_mt);
2773 * Do the hard work of removing an element from the buddy allocator.
2774 * Call me with the zone->lock already held.
2776 static __always_inline struct page *
2777 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2778 unsigned int alloc_flags)
2784 * Balance movable allocations between regular and CMA areas by
2785 * allocating from CMA when over half of the zone's free memory
2786 * is in the CMA area.
2788 if (alloc_flags & ALLOC_CMA &&
2789 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2790 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2791 page = __rmqueue_cma_fallback(zone, order);
2797 page = __rmqueue_smallest(zone, order, migratetype);
2798 if (unlikely(!page)) {
2799 if (alloc_flags & ALLOC_CMA)
2800 page = __rmqueue_cma_fallback(zone, order);
2802 if (!page && __rmqueue_fallback(zone, order, migratetype,
2807 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2812 * Obtain a specified number of elements from the buddy allocator, all under
2813 * a single hold of the lock, for efficiency. Add them to the supplied list.
2814 * Returns the number of new pages which were placed at *list.
2816 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2817 unsigned long count, struct list_head *list,
2818 int migratetype, unsigned int alloc_flags)
2822 spin_lock(&zone->lock);
2823 for (i = 0; i < count; ++i) {
2824 struct page *page = __rmqueue(zone, order, migratetype,
2826 if (unlikely(page == NULL))
2829 if (unlikely(check_pcp_refill(page)))
2833 * Split buddy pages returned by expand() are received here in
2834 * physical page order. The page is added to the tail of
2835 * caller's list. From the callers perspective, the linked list
2836 * is ordered by page number under some conditions. This is
2837 * useful for IO devices that can forward direction from the
2838 * head, thus also in the physical page order. This is useful
2839 * for IO devices that can merge IO requests if the physical
2840 * pages are ordered properly.
2842 list_add_tail(&page->lru, list);
2844 if (is_migrate_cma(get_pcppage_migratetype(page)))
2845 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2850 * i pages were removed from the buddy list even if some leak due
2851 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2852 * on i. Do not confuse with 'alloced' which is the number of
2853 * pages added to the pcp list.
2855 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2856 spin_unlock(&zone->lock);
2862 * Called from the vmstat counter updater to drain pagesets of this
2863 * currently executing processor on remote nodes after they have
2866 * Note that this function must be called with the thread pinned to
2867 * a single processor.
2869 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2871 unsigned long flags;
2872 int to_drain, batch;
2874 local_irq_save(flags);
2875 batch = READ_ONCE(pcp->batch);
2876 to_drain = min(pcp->count, batch);
2878 free_pcppages_bulk(zone, to_drain, pcp);
2879 local_irq_restore(flags);
2884 * Drain pcplists of the indicated processor and zone.
2886 * The processor must either be the current processor and the
2887 * thread pinned to the current processor or a processor that
2890 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2892 unsigned long flags;
2893 struct per_cpu_pageset *pset;
2894 struct per_cpu_pages *pcp;
2896 local_irq_save(flags);
2897 pset = per_cpu_ptr(zone->pageset, cpu);
2901 free_pcppages_bulk(zone, pcp->count, pcp);
2902 local_irq_restore(flags);
2906 * Drain pcplists of all zones on the indicated processor.
2908 * The processor must either be the current processor and the
2909 * thread pinned to the current processor or a processor that
2912 static void drain_pages(unsigned int cpu)
2916 for_each_populated_zone(zone) {
2917 drain_pages_zone(cpu, zone);
2922 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2924 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2925 * the single zone's pages.
2927 void drain_local_pages(struct zone *zone)
2929 int cpu = smp_processor_id();
2932 drain_pages_zone(cpu, zone);
2937 static void drain_local_pages_wq(struct work_struct *work)
2939 struct pcpu_drain *drain;
2941 drain = container_of(work, struct pcpu_drain, work);
2944 * drain_all_pages doesn't use proper cpu hotplug protection so
2945 * we can race with cpu offline when the WQ can move this from
2946 * a cpu pinned worker to an unbound one. We can operate on a different
2947 * cpu which is allright but we also have to make sure to not move to
2951 drain_local_pages(drain->zone);
2956 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2958 * When zone parameter is non-NULL, spill just the single zone's pages.
2960 * Note that this can be extremely slow as the draining happens in a workqueue.
2962 void drain_all_pages(struct zone *zone)
2967 * Allocate in the BSS so we wont require allocation in
2968 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2970 static cpumask_t cpus_with_pcps;
2973 * Make sure nobody triggers this path before mm_percpu_wq is fully
2976 if (WARN_ON_ONCE(!mm_percpu_wq))
2980 * Do not drain if one is already in progress unless it's specific to
2981 * a zone. Such callers are primarily CMA and memory hotplug and need
2982 * the drain to be complete when the call returns.
2984 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2987 mutex_lock(&pcpu_drain_mutex);
2991 * We don't care about racing with CPU hotplug event
2992 * as offline notification will cause the notified
2993 * cpu to drain that CPU pcps and on_each_cpu_mask
2994 * disables preemption as part of its processing
2996 for_each_online_cpu(cpu) {
2997 struct per_cpu_pageset *pcp;
2999 bool has_pcps = false;
3002 pcp = per_cpu_ptr(zone->pageset, cpu);
3006 for_each_populated_zone(z) {
3007 pcp = per_cpu_ptr(z->pageset, cpu);
3008 if (pcp->pcp.count) {
3016 cpumask_set_cpu(cpu, &cpus_with_pcps);
3018 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3021 for_each_cpu(cpu, &cpus_with_pcps) {
3022 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3025 INIT_WORK(&drain->work, drain_local_pages_wq);
3026 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3028 for_each_cpu(cpu, &cpus_with_pcps)
3029 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3031 mutex_unlock(&pcpu_drain_mutex);
3034 #ifdef CONFIG_HIBERNATION
3037 * Touch the watchdog for every WD_PAGE_COUNT pages.
3039 #define WD_PAGE_COUNT (128*1024)
3041 void mark_free_pages(struct zone *zone)
3043 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3044 unsigned long flags;
3045 unsigned int order, t;
3048 if (zone_is_empty(zone))
3051 spin_lock_irqsave(&zone->lock, flags);
3053 max_zone_pfn = zone_end_pfn(zone);
3054 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3055 if (pfn_valid(pfn)) {
3056 page = pfn_to_page(pfn);
3058 if (!--page_count) {
3059 touch_nmi_watchdog();
3060 page_count = WD_PAGE_COUNT;
3063 if (page_zone(page) != zone)
3066 if (!swsusp_page_is_forbidden(page))
3067 swsusp_unset_page_free(page);
3070 for_each_migratetype_order(order, t) {
3071 list_for_each_entry(page,
3072 &zone->free_area[order].free_list[t], lru) {
3075 pfn = page_to_pfn(page);
3076 for (i = 0; i < (1UL << order); i++) {
3077 if (!--page_count) {
3078 touch_nmi_watchdog();
3079 page_count = WD_PAGE_COUNT;
3081 swsusp_set_page_free(pfn_to_page(pfn + i));
3085 spin_unlock_irqrestore(&zone->lock, flags);
3087 #endif /* CONFIG_PM */
3089 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3093 if (!free_pcp_prepare(page))
3096 migratetype = get_pfnblock_migratetype(page, pfn);
3097 set_pcppage_migratetype(page, migratetype);
3101 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3103 struct zone *zone = page_zone(page);
3104 struct per_cpu_pages *pcp;
3107 migratetype = get_pcppage_migratetype(page);
3108 __count_vm_event(PGFREE);
3111 * We only track unmovable, reclaimable and movable on pcp lists.
3112 * Free ISOLATE pages back to the allocator because they are being
3113 * offlined but treat HIGHATOMIC as movable pages so we can get those
3114 * areas back if necessary. Otherwise, we may have to free
3115 * excessively into the page allocator
3117 if (migratetype >= MIGRATE_PCPTYPES) {
3118 if (unlikely(is_migrate_isolate(migratetype))) {
3119 free_one_page(zone, page, pfn, 0, migratetype);
3122 migratetype = MIGRATE_MOVABLE;
3125 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3126 list_add(&page->lru, &pcp->lists[migratetype]);
3128 if (pcp->count >= pcp->high) {
3129 unsigned long batch = READ_ONCE(pcp->batch);
3130 free_pcppages_bulk(zone, batch, pcp);
3135 * Free a 0-order page
3137 void free_unref_page(struct page *page)
3139 unsigned long flags;
3140 unsigned long pfn = page_to_pfn(page);
3142 if (!free_unref_page_prepare(page, pfn))
3145 local_irq_save(flags);
3146 free_unref_page_commit(page, pfn);
3147 local_irq_restore(flags);
3151 * Free a list of 0-order pages
3153 void free_unref_page_list(struct list_head *list)
3155 struct page *page, *next;
3156 unsigned long flags, pfn;
3157 int batch_count = 0;
3159 /* Prepare pages for freeing */
3160 list_for_each_entry_safe(page, next, list, lru) {
3161 pfn = page_to_pfn(page);
3162 if (!free_unref_page_prepare(page, pfn))
3163 list_del(&page->lru);
3164 set_page_private(page, pfn);
3167 local_irq_save(flags);
3168 list_for_each_entry_safe(page, next, list, lru) {
3169 unsigned long pfn = page_private(page);
3171 set_page_private(page, 0);
3172 trace_mm_page_free_batched(page);
3173 free_unref_page_commit(page, pfn);
3176 * Guard against excessive IRQ disabled times when we get
3177 * a large list of pages to free.
3179 if (++batch_count == SWAP_CLUSTER_MAX) {
3180 local_irq_restore(flags);
3182 local_irq_save(flags);
3185 local_irq_restore(flags);
3189 * split_page takes a non-compound higher-order page, and splits it into
3190 * n (1<<order) sub-pages: page[0..n]
3191 * Each sub-page must be freed individually.
3193 * Note: this is probably too low level an operation for use in drivers.
3194 * Please consult with lkml before using this in your driver.
3196 void split_page(struct page *page, unsigned int order)
3200 VM_BUG_ON_PAGE(PageCompound(page), page);
3201 VM_BUG_ON_PAGE(!page_count(page), page);
3203 for (i = 1; i < (1 << order); i++)
3204 set_page_refcounted(page + i);
3205 split_page_owner(page, order);
3207 EXPORT_SYMBOL_GPL(split_page);
3209 int __isolate_free_page(struct page *page, unsigned int order)
3211 unsigned long watermark;
3215 BUG_ON(!PageBuddy(page));
3217 zone = page_zone(page);
3218 mt = get_pageblock_migratetype(page);
3220 if (!is_migrate_isolate(mt)) {
3222 * Obey watermarks as if the page was being allocated. We can
3223 * emulate a high-order watermark check with a raised order-0
3224 * watermark, because we already know our high-order page
3227 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3228 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3231 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3234 /* Remove page from free list */
3236 del_page_from_free_list(page, zone, order);
3239 * Set the pageblock if the isolated page is at least half of a
3242 if (order >= pageblock_order - 1) {
3243 struct page *endpage = page + (1 << order) - 1;
3244 for (; page < endpage; page += pageblock_nr_pages) {
3245 int mt = get_pageblock_migratetype(page);
3246 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3247 && !is_migrate_highatomic(mt))
3248 set_pageblock_migratetype(page,
3254 return 1UL << order;
3258 * __putback_isolated_page - Return a now-isolated page back where we got it
3259 * @page: Page that was isolated
3260 * @order: Order of the isolated page
3261 * @mt: The page's pageblock's migratetype
3263 * This function is meant to return a page pulled from the free lists via
3264 * __isolate_free_page back to the free lists they were pulled from.
3266 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3268 struct zone *zone = page_zone(page);
3270 /* zone lock should be held when this function is called */
3271 lockdep_assert_held(&zone->lock);
3273 /* Return isolated page to tail of freelist. */
3274 __free_one_page(page, page_to_pfn(page), zone, order, mt, false);
3278 * Update NUMA hit/miss statistics
3280 * Must be called with interrupts disabled.
3282 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3285 enum numa_stat_item local_stat = NUMA_LOCAL;
3287 /* skip numa counters update if numa stats is disabled */
3288 if (!static_branch_likely(&vm_numa_stat_key))
3291 if (zone_to_nid(z) != numa_node_id())
3292 local_stat = NUMA_OTHER;
3294 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3295 __inc_numa_state(z, NUMA_HIT);
3297 __inc_numa_state(z, NUMA_MISS);
3298 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3300 __inc_numa_state(z, local_stat);
3304 /* Remove page from the per-cpu list, caller must protect the list */
3305 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3306 unsigned int alloc_flags,
3307 struct per_cpu_pages *pcp,
3308 struct list_head *list)
3313 if (list_empty(list)) {
3314 pcp->count += rmqueue_bulk(zone, 0,
3316 migratetype, alloc_flags);
3317 if (unlikely(list_empty(list)))
3321 page = list_first_entry(list, struct page, lru);
3322 list_del(&page->lru);
3324 } while (check_new_pcp(page));
3329 /* Lock and remove page from the per-cpu list */
3330 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3331 struct zone *zone, gfp_t gfp_flags,
3332 int migratetype, unsigned int alloc_flags)
3334 struct per_cpu_pages *pcp;
3335 struct list_head *list;
3337 unsigned long flags;
3339 local_irq_save(flags);
3340 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3341 list = &pcp->lists[migratetype];
3342 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3344 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3345 zone_statistics(preferred_zone, zone);
3347 local_irq_restore(flags);
3352 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3355 struct page *rmqueue(struct zone *preferred_zone,
3356 struct zone *zone, unsigned int order,
3357 gfp_t gfp_flags, unsigned int alloc_flags,
3360 unsigned long flags;
3363 if (likely(order == 0)) {
3364 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3365 migratetype, alloc_flags);
3370 * We most definitely don't want callers attempting to
3371 * allocate greater than order-1 page units with __GFP_NOFAIL.
3373 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3374 spin_lock_irqsave(&zone->lock, flags);
3378 if (alloc_flags & ALLOC_HARDER) {
3379 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3381 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3384 page = __rmqueue(zone, order, migratetype, alloc_flags);
3385 } while (page && check_new_pages(page, order));
3386 spin_unlock(&zone->lock);
3389 __mod_zone_freepage_state(zone, -(1 << order),
3390 get_pcppage_migratetype(page));
3392 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3393 zone_statistics(preferred_zone, zone);
3394 local_irq_restore(flags);
3397 /* Separate test+clear to avoid unnecessary atomics */
3398 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3399 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3400 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3403 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3407 local_irq_restore(flags);
3411 #ifdef CONFIG_FAIL_PAGE_ALLOC
3414 struct fault_attr attr;
3416 bool ignore_gfp_highmem;
3417 bool ignore_gfp_reclaim;
3419 } fail_page_alloc = {
3420 .attr = FAULT_ATTR_INITIALIZER,
3421 .ignore_gfp_reclaim = true,
3422 .ignore_gfp_highmem = true,
3426 static int __init setup_fail_page_alloc(char *str)
3428 return setup_fault_attr(&fail_page_alloc.attr, str);
3430 __setup("fail_page_alloc=", setup_fail_page_alloc);
3432 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3434 if (order < fail_page_alloc.min_order)
3436 if (gfp_mask & __GFP_NOFAIL)
3438 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3440 if (fail_page_alloc.ignore_gfp_reclaim &&
3441 (gfp_mask & __GFP_DIRECT_RECLAIM))
3444 return should_fail(&fail_page_alloc.attr, 1 << order);
3447 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3449 static int __init fail_page_alloc_debugfs(void)
3451 umode_t mode = S_IFREG | 0600;
3454 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3455 &fail_page_alloc.attr);
3457 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3458 &fail_page_alloc.ignore_gfp_reclaim);
3459 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3460 &fail_page_alloc.ignore_gfp_highmem);
3461 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3466 late_initcall(fail_page_alloc_debugfs);
3468 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3470 #else /* CONFIG_FAIL_PAGE_ALLOC */
3472 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3477 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3479 static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3481 return __should_fail_alloc_page(gfp_mask, order);
3483 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3485 static inline long __zone_watermark_unusable_free(struct zone *z,
3486 unsigned int order, unsigned int alloc_flags)
3488 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3489 long unusable_free = (1 << order) - 1;
3492 * If the caller does not have rights to ALLOC_HARDER then subtract
3493 * the high-atomic reserves. This will over-estimate the size of the
3494 * atomic reserve but it avoids a search.
3496 if (likely(!alloc_harder))
3497 unusable_free += z->nr_reserved_highatomic;
3500 /* If allocation can't use CMA areas don't use free CMA pages */
3501 if (!(alloc_flags & ALLOC_CMA))
3502 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3505 return unusable_free;
3509 * Return true if free base pages are above 'mark'. For high-order checks it
3510 * will return true of the order-0 watermark is reached and there is at least
3511 * one free page of a suitable size. Checking now avoids taking the zone lock
3512 * to check in the allocation paths if no pages are free.
3514 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3515 int highest_zoneidx, unsigned int alloc_flags,
3520 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3522 /* free_pages may go negative - that's OK */
3523 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3525 if (alloc_flags & ALLOC_HIGH)
3528 if (unlikely(alloc_harder)) {
3530 * OOM victims can try even harder than normal ALLOC_HARDER
3531 * users on the grounds that it's definitely going to be in
3532 * the exit path shortly and free memory. Any allocation it
3533 * makes during the free path will be small and short-lived.
3535 if (alloc_flags & ALLOC_OOM)
3542 * Check watermarks for an order-0 allocation request. If these
3543 * are not met, then a high-order request also cannot go ahead
3544 * even if a suitable page happened to be free.
3546 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3549 /* If this is an order-0 request then the watermark is fine */
3553 /* For a high-order request, check at least one suitable page is free */
3554 for (o = order; o < MAX_ORDER; o++) {
3555 struct free_area *area = &z->free_area[o];
3561 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3562 if (!free_area_empty(area, mt))
3567 if ((alloc_flags & ALLOC_CMA) &&
3568 !free_area_empty(area, MIGRATE_CMA)) {
3572 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3578 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3579 int highest_zoneidx, unsigned int alloc_flags)
3581 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3582 zone_page_state(z, NR_FREE_PAGES));
3585 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3586 unsigned long mark, int highest_zoneidx,
3587 unsigned int alloc_flags, gfp_t gfp_mask)
3591 free_pages = zone_page_state(z, NR_FREE_PAGES);
3594 * Fast check for order-0 only. If this fails then the reserves
3595 * need to be calculated.
3600 fast_free = free_pages;
3601 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3602 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3606 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3610 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3611 * when checking the min watermark. The min watermark is the
3612 * point where boosting is ignored so that kswapd is woken up
3613 * when below the low watermark.
3615 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3616 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3617 mark = z->_watermark[WMARK_MIN];
3618 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3619 alloc_flags, free_pages);
3625 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3626 unsigned long mark, int highest_zoneidx)
3628 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3630 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3631 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3633 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3638 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3640 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3641 node_reclaim_distance;
3643 #else /* CONFIG_NUMA */
3644 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3648 #endif /* CONFIG_NUMA */
3651 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3652 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3653 * premature use of a lower zone may cause lowmem pressure problems that
3654 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3655 * probably too small. It only makes sense to spread allocations to avoid
3656 * fragmentation between the Normal and DMA32 zones.
3658 static inline unsigned int
3659 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3661 unsigned int alloc_flags;
3664 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3667 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3669 #ifdef CONFIG_ZONE_DMA32
3673 if (zone_idx(zone) != ZONE_NORMAL)
3677 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3678 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3679 * on UMA that if Normal is populated then so is DMA32.
3681 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3682 if (nr_online_nodes > 1 && !populated_zone(--zone))
3685 alloc_flags |= ALLOC_NOFRAGMENT;
3686 #endif /* CONFIG_ZONE_DMA32 */
3690 static inline unsigned int current_alloc_flags(gfp_t gfp_mask,
3691 unsigned int alloc_flags)
3694 unsigned int pflags = current->flags;
3696 if (!(pflags & PF_MEMALLOC_NOCMA) &&
3697 gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3698 alloc_flags |= ALLOC_CMA;
3705 * get_page_from_freelist goes through the zonelist trying to allocate
3708 static struct page *
3709 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3710 const struct alloc_context *ac)
3714 struct pglist_data *last_pgdat_dirty_limit = NULL;
3719 * Scan zonelist, looking for a zone with enough free.
3720 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3722 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3723 z = ac->preferred_zoneref;
3724 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist,
3725 ac->highest_zoneidx, ac->nodemask) {
3729 if (cpusets_enabled() &&
3730 (alloc_flags & ALLOC_CPUSET) &&
3731 !__cpuset_zone_allowed(zone, gfp_mask))
3734 * When allocating a page cache page for writing, we
3735 * want to get it from a node that is within its dirty
3736 * limit, such that no single node holds more than its
3737 * proportional share of globally allowed dirty pages.
3738 * The dirty limits take into account the node's
3739 * lowmem reserves and high watermark so that kswapd
3740 * should be able to balance it without having to
3741 * write pages from its LRU list.
3743 * XXX: For now, allow allocations to potentially
3744 * exceed the per-node dirty limit in the slowpath
3745 * (spread_dirty_pages unset) before going into reclaim,
3746 * which is important when on a NUMA setup the allowed
3747 * nodes are together not big enough to reach the
3748 * global limit. The proper fix for these situations
3749 * will require awareness of nodes in the
3750 * dirty-throttling and the flusher threads.
3752 if (ac->spread_dirty_pages) {
3753 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3756 if (!node_dirty_ok(zone->zone_pgdat)) {
3757 last_pgdat_dirty_limit = zone->zone_pgdat;
3762 if (no_fallback && nr_online_nodes > 1 &&
3763 zone != ac->preferred_zoneref->zone) {
3767 * If moving to a remote node, retry but allow
3768 * fragmenting fallbacks. Locality is more important
3769 * than fragmentation avoidance.
3771 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3772 if (zone_to_nid(zone) != local_nid) {
3773 alloc_flags &= ~ALLOC_NOFRAGMENT;
3778 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3779 if (!zone_watermark_fast(zone, order, mark,
3780 ac->highest_zoneidx, alloc_flags,
3784 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3786 * Watermark failed for this zone, but see if we can
3787 * grow this zone if it contains deferred pages.
3789 if (static_branch_unlikely(&deferred_pages)) {
3790 if (_deferred_grow_zone(zone, order))
3794 /* Checked here to keep the fast path fast */
3795 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3796 if (alloc_flags & ALLOC_NO_WATERMARKS)
3799 if (node_reclaim_mode == 0 ||
3800 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3803 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3805 case NODE_RECLAIM_NOSCAN:
3808 case NODE_RECLAIM_FULL:
3809 /* scanned but unreclaimable */
3812 /* did we reclaim enough */
3813 if (zone_watermark_ok(zone, order, mark,
3814 ac->highest_zoneidx, alloc_flags))
3822 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3823 gfp_mask, alloc_flags, ac->migratetype);
3825 prep_new_page(page, order, gfp_mask, alloc_flags);
3828 * If this is a high-order atomic allocation then check
3829 * if the pageblock should be reserved for the future
3831 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3832 reserve_highatomic_pageblock(page, zone, order);
3836 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3837 /* Try again if zone has deferred pages */
3838 if (static_branch_unlikely(&deferred_pages)) {
3839 if (_deferred_grow_zone(zone, order))
3847 * It's possible on a UMA machine to get through all zones that are
3848 * fragmented. If avoiding fragmentation, reset and try again.
3851 alloc_flags &= ~ALLOC_NOFRAGMENT;
3858 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3860 unsigned int filter = SHOW_MEM_FILTER_NODES;
3863 * This documents exceptions given to allocations in certain
3864 * contexts that are allowed to allocate outside current's set
3867 if (!(gfp_mask & __GFP_NOMEMALLOC))
3868 if (tsk_is_oom_victim(current) ||
3869 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3870 filter &= ~SHOW_MEM_FILTER_NODES;
3871 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3872 filter &= ~SHOW_MEM_FILTER_NODES;
3874 show_mem(filter, nodemask);
3877 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3879 struct va_format vaf;
3881 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3883 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3886 va_start(args, fmt);
3889 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3890 current->comm, &vaf, gfp_mask, &gfp_mask,
3891 nodemask_pr_args(nodemask));
3894 cpuset_print_current_mems_allowed();
3897 warn_alloc_show_mem(gfp_mask, nodemask);
3900 static inline struct page *
3901 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3902 unsigned int alloc_flags,
3903 const struct alloc_context *ac)
3907 page = get_page_from_freelist(gfp_mask, order,
3908 alloc_flags|ALLOC_CPUSET, ac);
3910 * fallback to ignore cpuset restriction if our nodes
3914 page = get_page_from_freelist(gfp_mask, order,
3920 static inline struct page *
3921 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3922 const struct alloc_context *ac, unsigned long *did_some_progress)
3924 struct oom_control oc = {
3925 .zonelist = ac->zonelist,
3926 .nodemask = ac->nodemask,
3928 .gfp_mask = gfp_mask,
3933 *did_some_progress = 0;
3936 * Acquire the oom lock. If that fails, somebody else is
3937 * making progress for us.
3939 if (!mutex_trylock(&oom_lock)) {
3940 *did_some_progress = 1;
3941 schedule_timeout_uninterruptible(1);
3946 * Go through the zonelist yet one more time, keep very high watermark
3947 * here, this is only to catch a parallel oom killing, we must fail if
3948 * we're still under heavy pressure. But make sure that this reclaim
3949 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3950 * allocation which will never fail due to oom_lock already held.
3952 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3953 ~__GFP_DIRECT_RECLAIM, order,
3954 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3958 /* Coredumps can quickly deplete all memory reserves */
3959 if (current->flags & PF_DUMPCORE)
3961 /* The OOM killer will not help higher order allocs */
3962 if (order > PAGE_ALLOC_COSTLY_ORDER)
3965 * We have already exhausted all our reclaim opportunities without any
3966 * success so it is time to admit defeat. We will skip the OOM killer
3967 * because it is very likely that the caller has a more reasonable
3968 * fallback than shooting a random task.
3970 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3972 /* The OOM killer does not needlessly kill tasks for lowmem */
3973 if (ac->highest_zoneidx < ZONE_NORMAL)
3975 if (pm_suspended_storage())
3978 * XXX: GFP_NOFS allocations should rather fail than rely on
3979 * other request to make a forward progress.
3980 * We are in an unfortunate situation where out_of_memory cannot
3981 * do much for this context but let's try it to at least get
3982 * access to memory reserved if the current task is killed (see
3983 * out_of_memory). Once filesystems are ready to handle allocation
3984 * failures more gracefully we should just bail out here.
3987 /* The OOM killer may not free memory on a specific node */
3988 if (gfp_mask & __GFP_THISNODE)
3991 /* Exhausted what can be done so it's blame time */
3992 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3993 *did_some_progress = 1;
3996 * Help non-failing allocations by giving them access to memory
3999 if (gfp_mask & __GFP_NOFAIL)
4000 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4001 ALLOC_NO_WATERMARKS, ac);
4004 mutex_unlock(&oom_lock);
4009 * Maximum number of compaction retries wit a progress before OOM
4010 * killer is consider as the only way to move forward.
4012 #define MAX_COMPACT_RETRIES 16
4014 #ifdef CONFIG_COMPACTION
4015 /* Try memory compaction for high-order allocations before reclaim */
4016 static struct page *
4017 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4018 unsigned int alloc_flags, const struct alloc_context *ac,
4019 enum compact_priority prio, enum compact_result *compact_result)
4021 struct page *page = NULL;
4022 unsigned long pflags;
4023 unsigned int noreclaim_flag;
4028 psi_memstall_enter(&pflags);
4029 noreclaim_flag = memalloc_noreclaim_save();
4031 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4034 memalloc_noreclaim_restore(noreclaim_flag);
4035 psi_memstall_leave(&pflags);
4038 * At least in one zone compaction wasn't deferred or skipped, so let's
4039 * count a compaction stall
4041 count_vm_event(COMPACTSTALL);
4043 /* Prep a captured page if available */
4045 prep_new_page(page, order, gfp_mask, alloc_flags);
4047 /* Try get a page from the freelist if available */
4049 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4052 struct zone *zone = page_zone(page);
4054 zone->compact_blockskip_flush = false;
4055 compaction_defer_reset(zone, order, true);
4056 count_vm_event(COMPACTSUCCESS);
4061 * It's bad if compaction run occurs and fails. The most likely reason
4062 * is that pages exist, but not enough to satisfy watermarks.
4064 count_vm_event(COMPACTFAIL);
4072 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4073 enum compact_result compact_result,
4074 enum compact_priority *compact_priority,
4075 int *compaction_retries)
4077 int max_retries = MAX_COMPACT_RETRIES;
4080 int retries = *compaction_retries;
4081 enum compact_priority priority = *compact_priority;
4086 if (compaction_made_progress(compact_result))
4087 (*compaction_retries)++;
4090 * compaction considers all the zone as desperately out of memory
4091 * so it doesn't really make much sense to retry except when the
4092 * failure could be caused by insufficient priority
4094 if (compaction_failed(compact_result))
4095 goto check_priority;
4098 * compaction was skipped because there are not enough order-0 pages
4099 * to work with, so we retry only if it looks like reclaim can help.
4101 if (compaction_needs_reclaim(compact_result)) {
4102 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4107 * make sure the compaction wasn't deferred or didn't bail out early
4108 * due to locks contention before we declare that we should give up.
4109 * But the next retry should use a higher priority if allowed, so
4110 * we don't just keep bailing out endlessly.
4112 if (compaction_withdrawn(compact_result)) {
4113 goto check_priority;
4117 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4118 * costly ones because they are de facto nofail and invoke OOM
4119 * killer to move on while costly can fail and users are ready
4120 * to cope with that. 1/4 retries is rather arbitrary but we
4121 * would need much more detailed feedback from compaction to
4122 * make a better decision.
4124 if (order > PAGE_ALLOC_COSTLY_ORDER)
4126 if (*compaction_retries <= max_retries) {
4132 * Make sure there are attempts at the highest priority if we exhausted
4133 * all retries or failed at the lower priorities.
4136 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4137 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4139 if (*compact_priority > min_priority) {
4140 (*compact_priority)--;
4141 *compaction_retries = 0;
4145 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4149 static inline struct page *
4150 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4151 unsigned int alloc_flags, const struct alloc_context *ac,
4152 enum compact_priority prio, enum compact_result *compact_result)
4154 *compact_result = COMPACT_SKIPPED;
4159 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4160 enum compact_result compact_result,
4161 enum compact_priority *compact_priority,
4162 int *compaction_retries)
4167 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4171 * There are setups with compaction disabled which would prefer to loop
4172 * inside the allocator rather than hit the oom killer prematurely.
4173 * Let's give them a good hope and keep retrying while the order-0
4174 * watermarks are OK.
4176 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4177 ac->highest_zoneidx, ac->nodemask) {
4178 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4179 ac->highest_zoneidx, alloc_flags))
4184 #endif /* CONFIG_COMPACTION */
4186 #ifdef CONFIG_LOCKDEP
4187 static struct lockdep_map __fs_reclaim_map =
4188 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4190 static bool __need_fs_reclaim(gfp_t gfp_mask)
4192 gfp_mask = current_gfp_context(gfp_mask);
4194 /* no reclaim without waiting on it */
4195 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4198 /* this guy won't enter reclaim */
4199 if (current->flags & PF_MEMALLOC)
4202 /* We're only interested __GFP_FS allocations for now */
4203 if (!(gfp_mask & __GFP_FS))
4206 if (gfp_mask & __GFP_NOLOCKDEP)
4212 void __fs_reclaim_acquire(void)
4214 lock_map_acquire(&__fs_reclaim_map);
4217 void __fs_reclaim_release(void)
4219 lock_map_release(&__fs_reclaim_map);
4222 void fs_reclaim_acquire(gfp_t gfp_mask)
4224 if (__need_fs_reclaim(gfp_mask))
4225 __fs_reclaim_acquire();
4227 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4229 void fs_reclaim_release(gfp_t gfp_mask)
4231 if (__need_fs_reclaim(gfp_mask))
4232 __fs_reclaim_release();
4234 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4237 /* Perform direct synchronous page reclaim */
4239 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4240 const struct alloc_context *ac)
4243 unsigned int noreclaim_flag;
4244 unsigned long pflags;
4248 /* We now go into synchronous reclaim */
4249 cpuset_memory_pressure_bump();
4250 psi_memstall_enter(&pflags);
4251 fs_reclaim_acquire(gfp_mask);
4252 noreclaim_flag = memalloc_noreclaim_save();
4254 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4257 memalloc_noreclaim_restore(noreclaim_flag);
4258 fs_reclaim_release(gfp_mask);
4259 psi_memstall_leave(&pflags);
4266 /* The really slow allocator path where we enter direct reclaim */
4267 static inline struct page *
4268 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4269 unsigned int alloc_flags, const struct alloc_context *ac,
4270 unsigned long *did_some_progress)
4272 struct page *page = NULL;
4273 bool drained = false;
4275 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4276 if (unlikely(!(*did_some_progress)))
4280 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4283 * If an allocation failed after direct reclaim, it could be because
4284 * pages are pinned on the per-cpu lists or in high alloc reserves.
4285 * Shrink them them and try again
4287 if (!page && !drained) {
4288 unreserve_highatomic_pageblock(ac, false);
4289 drain_all_pages(NULL);
4297 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4298 const struct alloc_context *ac)
4302 pg_data_t *last_pgdat = NULL;
4303 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4305 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4307 if (last_pgdat != zone->zone_pgdat)
4308 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4309 last_pgdat = zone->zone_pgdat;
4313 static inline unsigned int
4314 gfp_to_alloc_flags(gfp_t gfp_mask)
4316 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4319 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4320 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4321 * to save two branches.
4323 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4324 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4327 * The caller may dip into page reserves a bit more if the caller
4328 * cannot run direct reclaim, or if the caller has realtime scheduling
4329 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4330 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4332 alloc_flags |= (__force int)
4333 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4335 if (gfp_mask & __GFP_ATOMIC) {
4337 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4338 * if it can't schedule.
4340 if (!(gfp_mask & __GFP_NOMEMALLOC))
4341 alloc_flags |= ALLOC_HARDER;
4343 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4344 * comment for __cpuset_node_allowed().
4346 alloc_flags &= ~ALLOC_CPUSET;
4347 } else if (unlikely(rt_task(current)) && !in_interrupt())
4348 alloc_flags |= ALLOC_HARDER;
4350 alloc_flags = current_alloc_flags(gfp_mask, alloc_flags);
4355 static bool oom_reserves_allowed(struct task_struct *tsk)
4357 if (!tsk_is_oom_victim(tsk))
4361 * !MMU doesn't have oom reaper so give access to memory reserves
4362 * only to the thread with TIF_MEMDIE set
4364 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4371 * Distinguish requests which really need access to full memory
4372 * reserves from oom victims which can live with a portion of it
4374 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4376 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4378 if (gfp_mask & __GFP_MEMALLOC)
4379 return ALLOC_NO_WATERMARKS;
4380 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4381 return ALLOC_NO_WATERMARKS;
4382 if (!in_interrupt()) {
4383 if (current->flags & PF_MEMALLOC)
4384 return ALLOC_NO_WATERMARKS;
4385 else if (oom_reserves_allowed(current))
4392 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4394 return !!__gfp_pfmemalloc_flags(gfp_mask);
4398 * Checks whether it makes sense to retry the reclaim to make a forward progress
4399 * for the given allocation request.
4401 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4402 * without success, or when we couldn't even meet the watermark if we
4403 * reclaimed all remaining pages on the LRU lists.
4405 * Returns true if a retry is viable or false to enter the oom path.
4408 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4409 struct alloc_context *ac, int alloc_flags,
4410 bool did_some_progress, int *no_progress_loops)
4417 * Costly allocations might have made a progress but this doesn't mean
4418 * their order will become available due to high fragmentation so
4419 * always increment the no progress counter for them
4421 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4422 *no_progress_loops = 0;
4424 (*no_progress_loops)++;
4427 * Make sure we converge to OOM if we cannot make any progress
4428 * several times in the row.
4430 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4431 /* Before OOM, exhaust highatomic_reserve */
4432 return unreserve_highatomic_pageblock(ac, true);
4436 * Keep reclaiming pages while there is a chance this will lead
4437 * somewhere. If none of the target zones can satisfy our allocation
4438 * request even if all reclaimable pages are considered then we are
4439 * screwed and have to go OOM.
4441 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4442 ac->highest_zoneidx, ac->nodemask) {
4443 unsigned long available;
4444 unsigned long reclaimable;
4445 unsigned long min_wmark = min_wmark_pages(zone);
4448 available = reclaimable = zone_reclaimable_pages(zone);
4449 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4452 * Would the allocation succeed if we reclaimed all
4453 * reclaimable pages?
4455 wmark = __zone_watermark_ok(zone, order, min_wmark,
4456 ac->highest_zoneidx, alloc_flags, available);
4457 trace_reclaim_retry_zone(z, order, reclaimable,
4458 available, min_wmark, *no_progress_loops, wmark);
4461 * If we didn't make any progress and have a lot of
4462 * dirty + writeback pages then we should wait for
4463 * an IO to complete to slow down the reclaim and
4464 * prevent from pre mature OOM
4466 if (!did_some_progress) {
4467 unsigned long write_pending;
4469 write_pending = zone_page_state_snapshot(zone,
4470 NR_ZONE_WRITE_PENDING);
4472 if (2 * write_pending > reclaimable) {
4473 congestion_wait(BLK_RW_ASYNC, HZ/10);
4485 * Memory allocation/reclaim might be called from a WQ context and the
4486 * current implementation of the WQ concurrency control doesn't
4487 * recognize that a particular WQ is congested if the worker thread is
4488 * looping without ever sleeping. Therefore we have to do a short sleep
4489 * here rather than calling cond_resched().
4491 if (current->flags & PF_WQ_WORKER)
4492 schedule_timeout_uninterruptible(1);
4499 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4502 * It's possible that cpuset's mems_allowed and the nodemask from
4503 * mempolicy don't intersect. This should be normally dealt with by
4504 * policy_nodemask(), but it's possible to race with cpuset update in
4505 * such a way the check therein was true, and then it became false
4506 * before we got our cpuset_mems_cookie here.
4507 * This assumes that for all allocations, ac->nodemask can come only
4508 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4509 * when it does not intersect with the cpuset restrictions) or the
4510 * caller can deal with a violated nodemask.
4512 if (cpusets_enabled() && ac->nodemask &&
4513 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4514 ac->nodemask = NULL;
4519 * When updating a task's mems_allowed or mempolicy nodemask, it is
4520 * possible to race with parallel threads in such a way that our
4521 * allocation can fail while the mask is being updated. If we are about
4522 * to fail, check if the cpuset changed during allocation and if so,
4525 if (read_mems_allowed_retry(cpuset_mems_cookie))
4531 static inline struct page *
4532 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4533 struct alloc_context *ac)
4535 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4536 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4537 struct page *page = NULL;
4538 unsigned int alloc_flags;
4539 unsigned long did_some_progress;
4540 enum compact_priority compact_priority;
4541 enum compact_result compact_result;
4542 int compaction_retries;
4543 int no_progress_loops;
4544 unsigned int cpuset_mems_cookie;
4548 * We also sanity check to catch abuse of atomic reserves being used by
4549 * callers that are not in atomic context.
4551 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4552 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4553 gfp_mask &= ~__GFP_ATOMIC;
4556 compaction_retries = 0;
4557 no_progress_loops = 0;
4558 compact_priority = DEF_COMPACT_PRIORITY;
4559 cpuset_mems_cookie = read_mems_allowed_begin();
4562 * The fast path uses conservative alloc_flags to succeed only until
4563 * kswapd needs to be woken up, and to avoid the cost of setting up
4564 * alloc_flags precisely. So we do that now.
4566 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4569 * We need to recalculate the starting point for the zonelist iterator
4570 * because we might have used different nodemask in the fast path, or
4571 * there was a cpuset modification and we are retrying - otherwise we
4572 * could end up iterating over non-eligible zones endlessly.
4574 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4575 ac->highest_zoneidx, ac->nodemask);
4576 if (!ac->preferred_zoneref->zone)
4579 if (alloc_flags & ALLOC_KSWAPD)
4580 wake_all_kswapds(order, gfp_mask, ac);
4583 * The adjusted alloc_flags might result in immediate success, so try
4586 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4591 * For costly allocations, try direct compaction first, as it's likely
4592 * that we have enough base pages and don't need to reclaim. For non-
4593 * movable high-order allocations, do that as well, as compaction will
4594 * try prevent permanent fragmentation by migrating from blocks of the
4596 * Don't try this for allocations that are allowed to ignore
4597 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4599 if (can_direct_reclaim &&
4601 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4602 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4603 page = __alloc_pages_direct_compact(gfp_mask, order,
4605 INIT_COMPACT_PRIORITY,
4611 * Checks for costly allocations with __GFP_NORETRY, which
4612 * includes some THP page fault allocations
4614 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4616 * If allocating entire pageblock(s) and compaction
4617 * failed because all zones are below low watermarks
4618 * or is prohibited because it recently failed at this
4619 * order, fail immediately unless the allocator has
4620 * requested compaction and reclaim retry.
4623 * - potentially very expensive because zones are far
4624 * below their low watermarks or this is part of very
4625 * bursty high order allocations,
4626 * - not guaranteed to help because isolate_freepages()
4627 * may not iterate over freed pages as part of its
4629 * - unlikely to make entire pageblocks free on its
4632 if (compact_result == COMPACT_SKIPPED ||
4633 compact_result == COMPACT_DEFERRED)
4637 * Looks like reclaim/compaction is worth trying, but
4638 * sync compaction could be very expensive, so keep
4639 * using async compaction.
4641 compact_priority = INIT_COMPACT_PRIORITY;
4646 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4647 if (alloc_flags & ALLOC_KSWAPD)
4648 wake_all_kswapds(order, gfp_mask, ac);
4650 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4652 alloc_flags = current_alloc_flags(gfp_mask, reserve_flags);
4655 * Reset the nodemask and zonelist iterators if memory policies can be
4656 * ignored. These allocations are high priority and system rather than
4659 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4660 ac->nodemask = NULL;
4661 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4662 ac->highest_zoneidx, ac->nodemask);
4665 /* Attempt with potentially adjusted zonelist and alloc_flags */
4666 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4670 /* Caller is not willing to reclaim, we can't balance anything */
4671 if (!can_direct_reclaim)
4674 /* Avoid recursion of direct reclaim */
4675 if (current->flags & PF_MEMALLOC)
4678 /* Try direct reclaim and then allocating */
4679 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4680 &did_some_progress);
4684 /* Try direct compaction and then allocating */
4685 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4686 compact_priority, &compact_result);
4690 /* Do not loop if specifically requested */
4691 if (gfp_mask & __GFP_NORETRY)
4695 * Do not retry costly high order allocations unless they are
4696 * __GFP_RETRY_MAYFAIL
4698 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4701 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4702 did_some_progress > 0, &no_progress_loops))
4706 * It doesn't make any sense to retry for the compaction if the order-0
4707 * reclaim is not able to make any progress because the current
4708 * implementation of the compaction depends on the sufficient amount
4709 * of free memory (see __compaction_suitable)
4711 if (did_some_progress > 0 &&
4712 should_compact_retry(ac, order, alloc_flags,
4713 compact_result, &compact_priority,
4714 &compaction_retries))
4718 /* Deal with possible cpuset update races before we start OOM killing */
4719 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4722 /* Reclaim has failed us, start killing things */
4723 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4727 /* Avoid allocations with no watermarks from looping endlessly */
4728 if (tsk_is_oom_victim(current) &&
4729 (alloc_flags & ALLOC_OOM ||
4730 (gfp_mask & __GFP_NOMEMALLOC)))
4733 /* Retry as long as the OOM killer is making progress */
4734 if (did_some_progress) {
4735 no_progress_loops = 0;
4740 /* Deal with possible cpuset update races before we fail */
4741 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4745 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4748 if (gfp_mask & __GFP_NOFAIL) {
4750 * All existing users of the __GFP_NOFAIL are blockable, so warn
4751 * of any new users that actually require GFP_NOWAIT
4753 if (WARN_ON_ONCE(!can_direct_reclaim))
4757 * PF_MEMALLOC request from this context is rather bizarre
4758 * because we cannot reclaim anything and only can loop waiting
4759 * for somebody to do a work for us
4761 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4764 * non failing costly orders are a hard requirement which we
4765 * are not prepared for much so let's warn about these users
4766 * so that we can identify them and convert them to something
4769 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4772 * Help non-failing allocations by giving them access to memory
4773 * reserves but do not use ALLOC_NO_WATERMARKS because this
4774 * could deplete whole memory reserves which would just make
4775 * the situation worse
4777 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4785 warn_alloc(gfp_mask, ac->nodemask,
4786 "page allocation failure: order:%u", order);
4791 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4792 int preferred_nid, nodemask_t *nodemask,
4793 struct alloc_context *ac, gfp_t *alloc_mask,
4794 unsigned int *alloc_flags)
4796 ac->highest_zoneidx = gfp_zone(gfp_mask);
4797 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4798 ac->nodemask = nodemask;
4799 ac->migratetype = gfp_migratetype(gfp_mask);
4801 if (cpusets_enabled()) {
4802 *alloc_mask |= __GFP_HARDWALL;
4804 * When we are in the interrupt context, it is irrelevant
4805 * to the current task context. It means that any node ok.
4807 if (!in_interrupt() && !ac->nodemask)
4808 ac->nodemask = &cpuset_current_mems_allowed;
4810 *alloc_flags |= ALLOC_CPUSET;
4813 fs_reclaim_acquire(gfp_mask);
4814 fs_reclaim_release(gfp_mask);
4816 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4818 if (should_fail_alloc_page(gfp_mask, order))
4821 *alloc_flags = current_alloc_flags(gfp_mask, *alloc_flags);
4826 /* Determine whether to spread dirty pages and what the first usable zone */
4827 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4829 /* Dirty zone balancing only done in the fast path */
4830 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4833 * The preferred zone is used for statistics but crucially it is
4834 * also used as the starting point for the zonelist iterator. It
4835 * may get reset for allocations that ignore memory policies.
4837 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4838 ac->highest_zoneidx, ac->nodemask);
4842 * This is the 'heart' of the zoned buddy allocator.
4845 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4846 nodemask_t *nodemask)
4849 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4850 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4851 struct alloc_context ac = { };
4854 * There are several places where we assume that the order value is sane
4855 * so bail out early if the request is out of bound.
4857 if (unlikely(order >= MAX_ORDER)) {
4858 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4862 gfp_mask &= gfp_allowed_mask;
4863 alloc_mask = gfp_mask;
4864 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4867 finalise_ac(gfp_mask, &ac);
4870 * Forbid the first pass from falling back to types that fragment
4871 * memory until all local zones are considered.
4873 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4875 /* First allocation attempt */
4876 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4881 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4882 * resp. GFP_NOIO which has to be inherited for all allocation requests
4883 * from a particular context which has been marked by
4884 * memalloc_no{fs,io}_{save,restore}.
4886 alloc_mask = current_gfp_context(gfp_mask);
4887 ac.spread_dirty_pages = false;
4890 * Restore the original nodemask if it was potentially replaced with
4891 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4893 ac.nodemask = nodemask;
4895 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4898 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4899 unlikely(__memcg_kmem_charge_page(page, gfp_mask, order) != 0)) {
4900 __free_pages(page, order);
4904 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4908 EXPORT_SYMBOL(__alloc_pages_nodemask);
4911 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4912 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4913 * you need to access high mem.
4915 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4919 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4922 return (unsigned long) page_address(page);
4924 EXPORT_SYMBOL(__get_free_pages);
4926 unsigned long get_zeroed_page(gfp_t gfp_mask)
4928 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4930 EXPORT_SYMBOL(get_zeroed_page);
4932 static inline void free_the_page(struct page *page, unsigned int order)
4934 if (order == 0) /* Via pcp? */
4935 free_unref_page(page);
4937 __free_pages_ok(page, order);
4940 void __free_pages(struct page *page, unsigned int order)
4942 if (put_page_testzero(page))
4943 free_the_page(page, order);
4945 EXPORT_SYMBOL(__free_pages);
4947 void free_pages(unsigned long addr, unsigned int order)
4950 VM_BUG_ON(!virt_addr_valid((void *)addr));
4951 __free_pages(virt_to_page((void *)addr), order);
4955 EXPORT_SYMBOL(free_pages);
4959 * An arbitrary-length arbitrary-offset area of memory which resides
4960 * within a 0 or higher order page. Multiple fragments within that page
4961 * are individually refcounted, in the page's reference counter.
4963 * The page_frag functions below provide a simple allocation framework for
4964 * page fragments. This is used by the network stack and network device
4965 * drivers to provide a backing region of memory for use as either an
4966 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4968 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4971 struct page *page = NULL;
4972 gfp_t gfp = gfp_mask;
4974 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4975 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4977 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4978 PAGE_FRAG_CACHE_MAX_ORDER);
4979 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4981 if (unlikely(!page))
4982 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4984 nc->va = page ? page_address(page) : NULL;
4989 void __page_frag_cache_drain(struct page *page, unsigned int count)
4991 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4993 if (page_ref_sub_and_test(page, count))
4994 free_the_page(page, compound_order(page));
4996 EXPORT_SYMBOL(__page_frag_cache_drain);
4998 void *page_frag_alloc(struct page_frag_cache *nc,
4999 unsigned int fragsz, gfp_t gfp_mask)
5001 unsigned int size = PAGE_SIZE;
5005 if (unlikely(!nc->va)) {
5007 page = __page_frag_cache_refill(nc, gfp_mask);
5011 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5012 /* if size can vary use size else just use PAGE_SIZE */
5015 /* Even if we own the page, we do not use atomic_set().
5016 * This would break get_page_unless_zero() users.
5018 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5020 /* reset page count bias and offset to start of new frag */
5021 nc->pfmemalloc = page_is_pfmemalloc(page);
5022 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5026 offset = nc->offset - fragsz;
5027 if (unlikely(offset < 0)) {
5028 page = virt_to_page(nc->va);
5030 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5033 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5034 /* if size can vary use size else just use PAGE_SIZE */
5037 /* OK, page count is 0, we can safely set it */
5038 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5040 /* reset page count bias and offset to start of new frag */
5041 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5042 offset = size - fragsz;
5046 nc->offset = offset;
5048 return nc->va + offset;
5050 EXPORT_SYMBOL(page_frag_alloc);
5053 * Frees a page fragment allocated out of either a compound or order 0 page.
5055 void page_frag_free(void *addr)
5057 struct page *page = virt_to_head_page(addr);
5059 if (unlikely(put_page_testzero(page)))
5060 free_the_page(page, compound_order(page));
5062 EXPORT_SYMBOL(page_frag_free);
5064 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5068 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5069 unsigned long used = addr + PAGE_ALIGN(size);
5071 split_page(virt_to_page((void *)addr), order);
5072 while (used < alloc_end) {
5077 return (void *)addr;
5081 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5082 * @size: the number of bytes to allocate
5083 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5085 * This function is similar to alloc_pages(), except that it allocates the
5086 * minimum number of pages to satisfy the request. alloc_pages() can only
5087 * allocate memory in power-of-two pages.
5089 * This function is also limited by MAX_ORDER.
5091 * Memory allocated by this function must be released by free_pages_exact().
5093 * Return: pointer to the allocated area or %NULL in case of error.
5095 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5097 unsigned int order = get_order(size);
5100 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5101 gfp_mask &= ~__GFP_COMP;
5103 addr = __get_free_pages(gfp_mask, order);
5104 return make_alloc_exact(addr, order, size);
5106 EXPORT_SYMBOL(alloc_pages_exact);
5109 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5111 * @nid: the preferred node ID where memory should be allocated
5112 * @size: the number of bytes to allocate
5113 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5115 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5118 * Return: pointer to the allocated area or %NULL in case of error.
5120 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5122 unsigned int order = get_order(size);
5125 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5126 gfp_mask &= ~__GFP_COMP;
5128 p = alloc_pages_node(nid, gfp_mask, order);
5131 return make_alloc_exact((unsigned long)page_address(p), order, size);
5135 * free_pages_exact - release memory allocated via alloc_pages_exact()
5136 * @virt: the value returned by alloc_pages_exact.
5137 * @size: size of allocation, same value as passed to alloc_pages_exact().
5139 * Release the memory allocated by a previous call to alloc_pages_exact.
5141 void free_pages_exact(void *virt, size_t size)
5143 unsigned long addr = (unsigned long)virt;
5144 unsigned long end = addr + PAGE_ALIGN(size);
5146 while (addr < end) {
5151 EXPORT_SYMBOL(free_pages_exact);
5154 * nr_free_zone_pages - count number of pages beyond high watermark
5155 * @offset: The zone index of the highest zone
5157 * nr_free_zone_pages() counts the number of pages which are beyond the
5158 * high watermark within all zones at or below a given zone index. For each
5159 * zone, the number of pages is calculated as:
5161 * nr_free_zone_pages = managed_pages - high_pages
5163 * Return: number of pages beyond high watermark.
5165 static unsigned long nr_free_zone_pages(int offset)
5170 /* Just pick one node, since fallback list is circular */
5171 unsigned long sum = 0;
5173 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5175 for_each_zone_zonelist(zone, z, zonelist, offset) {
5176 unsigned long size = zone_managed_pages(zone);
5177 unsigned long high = high_wmark_pages(zone);
5186 * nr_free_buffer_pages - count number of pages beyond high watermark
5188 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5189 * watermark within ZONE_DMA and ZONE_NORMAL.
5191 * Return: number of pages beyond high watermark within ZONE_DMA and
5194 unsigned long nr_free_buffer_pages(void)
5196 return nr_free_zone_pages(gfp_zone(GFP_USER));
5198 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5200 static inline void show_node(struct zone *zone)
5202 if (IS_ENABLED(CONFIG_NUMA))
5203 printk("Node %d ", zone_to_nid(zone));
5206 long si_mem_available(void)
5209 unsigned long pagecache;
5210 unsigned long wmark_low = 0;
5211 unsigned long pages[NR_LRU_LISTS];
5212 unsigned long reclaimable;
5216 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5217 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5220 wmark_low += low_wmark_pages(zone);
5223 * Estimate the amount of memory available for userspace allocations,
5224 * without causing swapping.
5226 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5229 * Not all the page cache can be freed, otherwise the system will
5230 * start swapping. Assume at least half of the page cache, or the
5231 * low watermark worth of cache, needs to stay.
5233 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5234 pagecache -= min(pagecache / 2, wmark_low);
5235 available += pagecache;
5238 * Part of the reclaimable slab and other kernel memory consists of
5239 * items that are in use, and cannot be freed. Cap this estimate at the
5242 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5243 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5244 available += reclaimable - min(reclaimable / 2, wmark_low);
5250 EXPORT_SYMBOL_GPL(si_mem_available);
5252 void si_meminfo(struct sysinfo *val)
5254 val->totalram = totalram_pages();
5255 val->sharedram = global_node_page_state(NR_SHMEM);
5256 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5257 val->bufferram = nr_blockdev_pages();
5258 val->totalhigh = totalhigh_pages();
5259 val->freehigh = nr_free_highpages();
5260 val->mem_unit = PAGE_SIZE;
5263 EXPORT_SYMBOL(si_meminfo);
5266 void si_meminfo_node(struct sysinfo *val, int nid)
5268 int zone_type; /* needs to be signed */
5269 unsigned long managed_pages = 0;
5270 unsigned long managed_highpages = 0;
5271 unsigned long free_highpages = 0;
5272 pg_data_t *pgdat = NODE_DATA(nid);
5274 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5275 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5276 val->totalram = managed_pages;
5277 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5278 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5279 #ifdef CONFIG_HIGHMEM
5280 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5281 struct zone *zone = &pgdat->node_zones[zone_type];
5283 if (is_highmem(zone)) {
5284 managed_highpages += zone_managed_pages(zone);
5285 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5288 val->totalhigh = managed_highpages;
5289 val->freehigh = free_highpages;
5291 val->totalhigh = managed_highpages;
5292 val->freehigh = free_highpages;
5294 val->mem_unit = PAGE_SIZE;
5299 * Determine whether the node should be displayed or not, depending on whether
5300 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5302 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5304 if (!(flags & SHOW_MEM_FILTER_NODES))
5308 * no node mask - aka implicit memory numa policy. Do not bother with
5309 * the synchronization - read_mems_allowed_begin - because we do not
5310 * have to be precise here.
5313 nodemask = &cpuset_current_mems_allowed;
5315 return !node_isset(nid, *nodemask);
5318 #define K(x) ((x) << (PAGE_SHIFT-10))
5320 static void show_migration_types(unsigned char type)
5322 static const char types[MIGRATE_TYPES] = {
5323 [MIGRATE_UNMOVABLE] = 'U',
5324 [MIGRATE_MOVABLE] = 'M',
5325 [MIGRATE_RECLAIMABLE] = 'E',
5326 [MIGRATE_HIGHATOMIC] = 'H',
5328 [MIGRATE_CMA] = 'C',
5330 #ifdef CONFIG_MEMORY_ISOLATION
5331 [MIGRATE_ISOLATE] = 'I',
5334 char tmp[MIGRATE_TYPES + 1];
5338 for (i = 0; i < MIGRATE_TYPES; i++) {
5339 if (type & (1 << i))
5344 printk(KERN_CONT "(%s) ", tmp);
5348 * Show free area list (used inside shift_scroll-lock stuff)
5349 * We also calculate the percentage fragmentation. We do this by counting the
5350 * memory on each free list with the exception of the first item on the list.
5353 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5356 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5358 unsigned long free_pcp = 0;
5363 for_each_populated_zone(zone) {
5364 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5367 for_each_online_cpu(cpu)
5368 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5371 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5372 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5373 " unevictable:%lu dirty:%lu writeback:%lu\n"
5374 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5375 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5376 " free:%lu free_pcp:%lu free_cma:%lu\n",
5377 global_node_page_state(NR_ACTIVE_ANON),
5378 global_node_page_state(NR_INACTIVE_ANON),
5379 global_node_page_state(NR_ISOLATED_ANON),
5380 global_node_page_state(NR_ACTIVE_FILE),
5381 global_node_page_state(NR_INACTIVE_FILE),
5382 global_node_page_state(NR_ISOLATED_FILE),
5383 global_node_page_state(NR_UNEVICTABLE),
5384 global_node_page_state(NR_FILE_DIRTY),
5385 global_node_page_state(NR_WRITEBACK),
5386 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5387 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5388 global_node_page_state(NR_FILE_MAPPED),
5389 global_node_page_state(NR_SHMEM),
5390 global_zone_page_state(NR_PAGETABLE),
5391 global_zone_page_state(NR_BOUNCE),
5392 global_zone_page_state(NR_FREE_PAGES),
5394 global_zone_page_state(NR_FREE_CMA_PAGES));
5396 for_each_online_pgdat(pgdat) {
5397 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5401 " active_anon:%lukB"
5402 " inactive_anon:%lukB"
5403 " active_file:%lukB"
5404 " inactive_file:%lukB"
5405 " unevictable:%lukB"
5406 " isolated(anon):%lukB"
5407 " isolated(file):%lukB"
5412 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5414 " shmem_pmdmapped: %lukB"
5417 " writeback_tmp:%lukB"
5418 " kernel_stack:%lukB"
5419 #ifdef CONFIG_SHADOW_CALL_STACK
5420 " shadow_call_stack:%lukB"
5422 " all_unreclaimable? %s"
5425 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5426 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5427 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5428 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5429 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5430 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5431 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5432 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5433 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5434 K(node_page_state(pgdat, NR_WRITEBACK)),
5435 K(node_page_state(pgdat, NR_SHMEM)),
5436 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5437 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5438 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5440 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5442 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5443 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5444 #ifdef CONFIG_SHADOW_CALL_STACK
5445 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5447 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5451 for_each_populated_zone(zone) {
5454 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5458 for_each_online_cpu(cpu)
5459 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5468 " reserved_highatomic:%luKB"
5469 " active_anon:%lukB"
5470 " inactive_anon:%lukB"
5471 " active_file:%lukB"
5472 " inactive_file:%lukB"
5473 " unevictable:%lukB"
5474 " writepending:%lukB"
5485 K(zone_page_state(zone, NR_FREE_PAGES)),
5486 K(min_wmark_pages(zone)),
5487 K(low_wmark_pages(zone)),
5488 K(high_wmark_pages(zone)),
5489 K(zone->nr_reserved_highatomic),
5490 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5491 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5492 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5493 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5494 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5495 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5496 K(zone->present_pages),
5497 K(zone_managed_pages(zone)),
5498 K(zone_page_state(zone, NR_MLOCK)),
5499 K(zone_page_state(zone, NR_PAGETABLE)),
5500 K(zone_page_state(zone, NR_BOUNCE)),
5502 K(this_cpu_read(zone->pageset->pcp.count)),
5503 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5504 printk("lowmem_reserve[]:");
5505 for (i = 0; i < MAX_NR_ZONES; i++)
5506 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5507 printk(KERN_CONT "\n");
5510 for_each_populated_zone(zone) {
5512 unsigned long nr[MAX_ORDER], flags, total = 0;
5513 unsigned char types[MAX_ORDER];
5515 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5518 printk(KERN_CONT "%s: ", zone->name);
5520 spin_lock_irqsave(&zone->lock, flags);
5521 for (order = 0; order < MAX_ORDER; order++) {
5522 struct free_area *area = &zone->free_area[order];
5525 nr[order] = area->nr_free;
5526 total += nr[order] << order;
5529 for (type = 0; type < MIGRATE_TYPES; type++) {
5530 if (!free_area_empty(area, type))
5531 types[order] |= 1 << type;
5534 spin_unlock_irqrestore(&zone->lock, flags);
5535 for (order = 0; order < MAX_ORDER; order++) {
5536 printk(KERN_CONT "%lu*%lukB ",
5537 nr[order], K(1UL) << order);
5539 show_migration_types(types[order]);
5541 printk(KERN_CONT "= %lukB\n", K(total));
5544 hugetlb_show_meminfo();
5546 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5548 show_swap_cache_info();
5551 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5553 zoneref->zone = zone;
5554 zoneref->zone_idx = zone_idx(zone);
5558 * Builds allocation fallback zone lists.
5560 * Add all populated zones of a node to the zonelist.
5562 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5565 enum zone_type zone_type = MAX_NR_ZONES;
5570 zone = pgdat->node_zones + zone_type;
5571 if (managed_zone(zone)) {
5572 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5573 check_highest_zone(zone_type);
5575 } while (zone_type);
5582 static int __parse_numa_zonelist_order(char *s)
5585 * We used to support different zonlists modes but they turned
5586 * out to be just not useful. Let's keep the warning in place
5587 * if somebody still use the cmd line parameter so that we do
5588 * not fail it silently
5590 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5591 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5597 char numa_zonelist_order[] = "Node";
5600 * sysctl handler for numa_zonelist_order
5602 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5603 void *buffer, size_t *length, loff_t *ppos)
5606 return __parse_numa_zonelist_order(buffer);
5607 return proc_dostring(table, write, buffer, length, ppos);
5611 #define MAX_NODE_LOAD (nr_online_nodes)
5612 static int node_load[MAX_NUMNODES];
5615 * find_next_best_node - find the next node that should appear in a given node's fallback list
5616 * @node: node whose fallback list we're appending
5617 * @used_node_mask: nodemask_t of already used nodes
5619 * We use a number of factors to determine which is the next node that should
5620 * appear on a given node's fallback list. The node should not have appeared
5621 * already in @node's fallback list, and it should be the next closest node
5622 * according to the distance array (which contains arbitrary distance values
5623 * from each node to each node in the system), and should also prefer nodes
5624 * with no CPUs, since presumably they'll have very little allocation pressure
5625 * on them otherwise.
5627 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5629 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5632 int min_val = INT_MAX;
5633 int best_node = NUMA_NO_NODE;
5634 const struct cpumask *tmp = cpumask_of_node(0);
5636 /* Use the local node if we haven't already */
5637 if (!node_isset(node, *used_node_mask)) {
5638 node_set(node, *used_node_mask);
5642 for_each_node_state(n, N_MEMORY) {
5644 /* Don't want a node to appear more than once */
5645 if (node_isset(n, *used_node_mask))
5648 /* Use the distance array to find the distance */
5649 val = node_distance(node, n);
5651 /* Penalize nodes under us ("prefer the next node") */
5654 /* Give preference to headless and unused nodes */
5655 tmp = cpumask_of_node(n);
5656 if (!cpumask_empty(tmp))
5657 val += PENALTY_FOR_NODE_WITH_CPUS;
5659 /* Slight preference for less loaded node */
5660 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5661 val += node_load[n];
5663 if (val < min_val) {
5670 node_set(best_node, *used_node_mask);
5677 * Build zonelists ordered by node and zones within node.
5678 * This results in maximum locality--normal zone overflows into local
5679 * DMA zone, if any--but risks exhausting DMA zone.
5681 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5684 struct zoneref *zonerefs;
5687 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5689 for (i = 0; i < nr_nodes; i++) {
5692 pg_data_t *node = NODE_DATA(node_order[i]);
5694 nr_zones = build_zonerefs_node(node, zonerefs);
5695 zonerefs += nr_zones;
5697 zonerefs->zone = NULL;
5698 zonerefs->zone_idx = 0;
5702 * Build gfp_thisnode zonelists
5704 static void build_thisnode_zonelists(pg_data_t *pgdat)
5706 struct zoneref *zonerefs;
5709 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5710 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5711 zonerefs += nr_zones;
5712 zonerefs->zone = NULL;
5713 zonerefs->zone_idx = 0;
5717 * Build zonelists ordered by zone and nodes within zones.
5718 * This results in conserving DMA zone[s] until all Normal memory is
5719 * exhausted, but results in overflowing to remote node while memory
5720 * may still exist in local DMA zone.
5723 static void build_zonelists(pg_data_t *pgdat)
5725 static int node_order[MAX_NUMNODES];
5726 int node, load, nr_nodes = 0;
5727 nodemask_t used_mask = NODE_MASK_NONE;
5728 int local_node, prev_node;
5730 /* NUMA-aware ordering of nodes */
5731 local_node = pgdat->node_id;
5732 load = nr_online_nodes;
5733 prev_node = local_node;
5735 memset(node_order, 0, sizeof(node_order));
5736 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5738 * We don't want to pressure a particular node.
5739 * So adding penalty to the first node in same
5740 * distance group to make it round-robin.
5742 if (node_distance(local_node, node) !=
5743 node_distance(local_node, prev_node))
5744 node_load[node] = load;
5746 node_order[nr_nodes++] = node;
5751 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5752 build_thisnode_zonelists(pgdat);
5755 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5757 * Return node id of node used for "local" allocations.
5758 * I.e., first node id of first zone in arg node's generic zonelist.
5759 * Used for initializing percpu 'numa_mem', which is used primarily
5760 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5762 int local_memory_node(int node)
5766 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5767 gfp_zone(GFP_KERNEL),
5769 return zone_to_nid(z->zone);
5773 static void setup_min_unmapped_ratio(void);
5774 static void setup_min_slab_ratio(void);
5775 #else /* CONFIG_NUMA */
5777 static void build_zonelists(pg_data_t *pgdat)
5779 int node, local_node;
5780 struct zoneref *zonerefs;
5783 local_node = pgdat->node_id;
5785 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5786 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5787 zonerefs += nr_zones;
5790 * Now we build the zonelist so that it contains the zones
5791 * of all the other nodes.
5792 * We don't want to pressure a particular node, so when
5793 * building the zones for node N, we make sure that the
5794 * zones coming right after the local ones are those from
5795 * node N+1 (modulo N)
5797 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5798 if (!node_online(node))
5800 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5801 zonerefs += nr_zones;
5803 for (node = 0; node < local_node; node++) {
5804 if (!node_online(node))
5806 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5807 zonerefs += nr_zones;
5810 zonerefs->zone = NULL;
5811 zonerefs->zone_idx = 0;
5814 #endif /* CONFIG_NUMA */
5817 * Boot pageset table. One per cpu which is going to be used for all
5818 * zones and all nodes. The parameters will be set in such a way
5819 * that an item put on a list will immediately be handed over to
5820 * the buddy list. This is safe since pageset manipulation is done
5821 * with interrupts disabled.
5823 * The boot_pagesets must be kept even after bootup is complete for
5824 * unused processors and/or zones. They do play a role for bootstrapping
5825 * hotplugged processors.
5827 * zoneinfo_show() and maybe other functions do
5828 * not check if the processor is online before following the pageset pointer.
5829 * Other parts of the kernel may not check if the zone is available.
5831 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5832 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5833 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5835 static void __build_all_zonelists(void *data)
5838 int __maybe_unused cpu;
5839 pg_data_t *self = data;
5840 static DEFINE_SPINLOCK(lock);
5845 memset(node_load, 0, sizeof(node_load));
5849 * This node is hotadded and no memory is yet present. So just
5850 * building zonelists is fine - no need to touch other nodes.
5852 if (self && !node_online(self->node_id)) {
5853 build_zonelists(self);
5855 for_each_online_node(nid) {
5856 pg_data_t *pgdat = NODE_DATA(nid);
5858 build_zonelists(pgdat);
5861 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5863 * We now know the "local memory node" for each node--
5864 * i.e., the node of the first zone in the generic zonelist.
5865 * Set up numa_mem percpu variable for on-line cpus. During
5866 * boot, only the boot cpu should be on-line; we'll init the
5867 * secondary cpus' numa_mem as they come on-line. During
5868 * node/memory hotplug, we'll fixup all on-line cpus.
5870 for_each_online_cpu(cpu)
5871 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5878 static noinline void __init
5879 build_all_zonelists_init(void)
5883 __build_all_zonelists(NULL);
5886 * Initialize the boot_pagesets that are going to be used
5887 * for bootstrapping processors. The real pagesets for
5888 * each zone will be allocated later when the per cpu
5889 * allocator is available.
5891 * boot_pagesets are used also for bootstrapping offline
5892 * cpus if the system is already booted because the pagesets
5893 * are needed to initialize allocators on a specific cpu too.
5894 * F.e. the percpu allocator needs the page allocator which
5895 * needs the percpu allocator in order to allocate its pagesets
5896 * (a chicken-egg dilemma).
5898 for_each_possible_cpu(cpu)
5899 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5901 mminit_verify_zonelist();
5902 cpuset_init_current_mems_allowed();
5906 * unless system_state == SYSTEM_BOOTING.
5908 * __ref due to call of __init annotated helper build_all_zonelists_init
5909 * [protected by SYSTEM_BOOTING].
5911 void __ref build_all_zonelists(pg_data_t *pgdat)
5913 unsigned long vm_total_pages;
5915 if (system_state == SYSTEM_BOOTING) {
5916 build_all_zonelists_init();
5918 __build_all_zonelists(pgdat);
5919 /* cpuset refresh routine should be here */
5921 /* Get the number of free pages beyond high watermark in all zones. */
5922 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5924 * Disable grouping by mobility if the number of pages in the
5925 * system is too low to allow the mechanism to work. It would be
5926 * more accurate, but expensive to check per-zone. This check is
5927 * made on memory-hotadd so a system can start with mobility
5928 * disabled and enable it later
5930 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5931 page_group_by_mobility_disabled = 1;
5933 page_group_by_mobility_disabled = 0;
5935 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5937 page_group_by_mobility_disabled ? "off" : "on",
5940 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5944 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5945 static bool __meminit
5946 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5948 static struct memblock_region *r;
5950 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5951 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5952 for_each_memblock(memory, r) {
5953 if (*pfn < memblock_region_memory_end_pfn(r))
5957 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5958 memblock_is_mirror(r)) {
5959 *pfn = memblock_region_memory_end_pfn(r);
5967 * Initially all pages are reserved - free ones are freed
5968 * up by memblock_free_all() once the early boot process is
5969 * done. Non-atomic initialization, single-pass.
5971 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5972 unsigned long start_pfn, enum memmap_context context,
5973 struct vmem_altmap *altmap)
5975 unsigned long pfn, end_pfn = start_pfn + size;
5978 if (highest_memmap_pfn < end_pfn - 1)
5979 highest_memmap_pfn = end_pfn - 1;
5981 #ifdef CONFIG_ZONE_DEVICE
5983 * Honor reservation requested by the driver for this ZONE_DEVICE
5984 * memory. We limit the total number of pages to initialize to just
5985 * those that might contain the memory mapping. We will defer the
5986 * ZONE_DEVICE page initialization until after we have released
5989 if (zone == ZONE_DEVICE) {
5993 if (start_pfn == altmap->base_pfn)
5994 start_pfn += altmap->reserve;
5995 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5999 for (pfn = start_pfn; pfn < end_pfn; ) {
6001 * There can be holes in boot-time mem_map[]s handed to this
6002 * function. They do not exist on hotplugged memory.
6004 if (context == MEMMAP_EARLY) {
6005 if (overlap_memmap_init(zone, &pfn))
6007 if (defer_init(nid, pfn, end_pfn))
6011 page = pfn_to_page(pfn);
6012 __init_single_page(page, pfn, zone, nid);
6013 if (context == MEMMAP_HOTPLUG)
6014 __SetPageReserved(page);
6017 * Mark the block movable so that blocks are reserved for
6018 * movable at startup. This will force kernel allocations
6019 * to reserve their blocks rather than leaking throughout
6020 * the address space during boot when many long-lived
6021 * kernel allocations are made.
6023 * bitmap is created for zone's valid pfn range. but memmap
6024 * can be created for invalid pages (for alignment)
6025 * check here not to call set_pageblock_migratetype() against
6028 if (!(pfn & (pageblock_nr_pages - 1))) {
6029 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6036 #ifdef CONFIG_ZONE_DEVICE
6037 void __ref memmap_init_zone_device(struct zone *zone,
6038 unsigned long start_pfn,
6039 unsigned long nr_pages,
6040 struct dev_pagemap *pgmap)
6042 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6043 struct pglist_data *pgdat = zone->zone_pgdat;
6044 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6045 unsigned long zone_idx = zone_idx(zone);
6046 unsigned long start = jiffies;
6047 int nid = pgdat->node_id;
6049 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6053 * The call to memmap_init_zone should have already taken care
6054 * of the pages reserved for the memmap, so we can just jump to
6055 * the end of that region and start processing the device pages.
6058 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6059 nr_pages = end_pfn - start_pfn;
6062 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6063 struct page *page = pfn_to_page(pfn);
6065 __init_single_page(page, pfn, zone_idx, nid);
6068 * Mark page reserved as it will need to wait for onlining
6069 * phase for it to be fully associated with a zone.
6071 * We can use the non-atomic __set_bit operation for setting
6072 * the flag as we are still initializing the pages.
6074 __SetPageReserved(page);
6077 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6078 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6079 * ever freed or placed on a driver-private list.
6081 page->pgmap = pgmap;
6082 page->zone_device_data = NULL;
6085 * Mark the block movable so that blocks are reserved for
6086 * movable at startup. This will force kernel allocations
6087 * to reserve their blocks rather than leaking throughout
6088 * the address space during boot when many long-lived
6089 * kernel allocations are made.
6091 * bitmap is created for zone's valid pfn range. but memmap
6092 * can be created for invalid pages (for alignment)
6093 * check here not to call set_pageblock_migratetype() against
6096 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
6097 * because this is done early in section_activate()
6099 if (!(pfn & (pageblock_nr_pages - 1))) {
6100 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6105 pr_info("%s initialised %lu pages in %ums\n", __func__,
6106 nr_pages, jiffies_to_msecs(jiffies - start));
6110 static void __meminit zone_init_free_lists(struct zone *zone)
6112 unsigned int order, t;
6113 for_each_migratetype_order(order, t) {
6114 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6115 zone->free_area[order].nr_free = 0;
6119 void __meminit __weak memmap_init(unsigned long size, int nid,
6121 unsigned long range_start_pfn)
6123 unsigned long start_pfn, end_pfn;
6124 unsigned long range_end_pfn = range_start_pfn + size;
6127 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6128 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6129 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6131 if (end_pfn > start_pfn) {
6132 size = end_pfn - start_pfn;
6133 memmap_init_zone(size, nid, zone, start_pfn,
6134 MEMMAP_EARLY, NULL);
6139 static int zone_batchsize(struct zone *zone)
6145 * The per-cpu-pages pools are set to around 1000th of the
6148 batch = zone_managed_pages(zone) / 1024;
6149 /* But no more than a meg. */
6150 if (batch * PAGE_SIZE > 1024 * 1024)
6151 batch = (1024 * 1024) / PAGE_SIZE;
6152 batch /= 4; /* We effectively *= 4 below */
6157 * Clamp the batch to a 2^n - 1 value. Having a power
6158 * of 2 value was found to be more likely to have
6159 * suboptimal cache aliasing properties in some cases.
6161 * For example if 2 tasks are alternately allocating
6162 * batches of pages, one task can end up with a lot
6163 * of pages of one half of the possible page colors
6164 * and the other with pages of the other colors.
6166 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6171 /* The deferral and batching of frees should be suppressed under NOMMU
6174 * The problem is that NOMMU needs to be able to allocate large chunks
6175 * of contiguous memory as there's no hardware page translation to
6176 * assemble apparent contiguous memory from discontiguous pages.
6178 * Queueing large contiguous runs of pages for batching, however,
6179 * causes the pages to actually be freed in smaller chunks. As there
6180 * can be a significant delay between the individual batches being
6181 * recycled, this leads to the once large chunks of space being
6182 * fragmented and becoming unavailable for high-order allocations.
6189 * pcp->high and pcp->batch values are related and dependent on one another:
6190 * ->batch must never be higher then ->high.
6191 * The following function updates them in a safe manner without read side
6194 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6195 * those fields changing asynchronously (acording the the above rule).
6197 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6198 * outside of boot time (or some other assurance that no concurrent updaters
6201 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6202 unsigned long batch)
6204 /* start with a fail safe value for batch */
6208 /* Update high, then batch, in order */
6215 /* a companion to pageset_set_high() */
6216 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6218 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6221 static void pageset_init(struct per_cpu_pageset *p)
6223 struct per_cpu_pages *pcp;
6226 memset(p, 0, sizeof(*p));
6229 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6230 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6233 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6236 pageset_set_batch(p, batch);
6240 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6241 * to the value high for the pageset p.
6243 static void pageset_set_high(struct per_cpu_pageset *p,
6246 unsigned long batch = max(1UL, high / 4);
6247 if ((high / 4) > (PAGE_SHIFT * 8))
6248 batch = PAGE_SHIFT * 8;
6250 pageset_update(&p->pcp, high, batch);
6253 static void pageset_set_high_and_batch(struct zone *zone,
6254 struct per_cpu_pageset *pcp)
6256 if (percpu_pagelist_fraction)
6257 pageset_set_high(pcp,
6258 (zone_managed_pages(zone) /
6259 percpu_pagelist_fraction));
6261 pageset_set_batch(pcp, zone_batchsize(zone));
6264 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6266 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6269 pageset_set_high_and_batch(zone, pcp);
6272 void __meminit setup_zone_pageset(struct zone *zone)
6275 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6276 for_each_possible_cpu(cpu)
6277 zone_pageset_init(zone, cpu);
6281 * Allocate per cpu pagesets and initialize them.
6282 * Before this call only boot pagesets were available.
6284 void __init setup_per_cpu_pageset(void)
6286 struct pglist_data *pgdat;
6288 int __maybe_unused cpu;
6290 for_each_populated_zone(zone)
6291 setup_zone_pageset(zone);
6295 * Unpopulated zones continue using the boot pagesets.
6296 * The numa stats for these pagesets need to be reset.
6297 * Otherwise, they will end up skewing the stats of
6298 * the nodes these zones are associated with.
6300 for_each_possible_cpu(cpu) {
6301 struct per_cpu_pageset *pcp = &per_cpu(boot_pageset, cpu);
6302 memset(pcp->vm_numa_stat_diff, 0,
6303 sizeof(pcp->vm_numa_stat_diff));
6307 for_each_online_pgdat(pgdat)
6308 pgdat->per_cpu_nodestats =
6309 alloc_percpu(struct per_cpu_nodestat);
6312 static __meminit void zone_pcp_init(struct zone *zone)
6315 * per cpu subsystem is not up at this point. The following code
6316 * relies on the ability of the linker to provide the
6317 * offset of a (static) per cpu variable into the per cpu area.
6319 zone->pageset = &boot_pageset;
6321 if (populated_zone(zone))
6322 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6323 zone->name, zone->present_pages,
6324 zone_batchsize(zone));
6327 void __meminit init_currently_empty_zone(struct zone *zone,
6328 unsigned long zone_start_pfn,
6331 struct pglist_data *pgdat = zone->zone_pgdat;
6332 int zone_idx = zone_idx(zone) + 1;
6334 if (zone_idx > pgdat->nr_zones)
6335 pgdat->nr_zones = zone_idx;
6337 zone->zone_start_pfn = zone_start_pfn;
6339 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6340 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6342 (unsigned long)zone_idx(zone),
6343 zone_start_pfn, (zone_start_pfn + size));
6345 zone_init_free_lists(zone);
6346 zone->initialized = 1;
6350 * get_pfn_range_for_nid - Return the start and end page frames for a node
6351 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6352 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6353 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6355 * It returns the start and end page frame of a node based on information
6356 * provided by memblock_set_node(). If called for a node
6357 * with no available memory, a warning is printed and the start and end
6360 void __init get_pfn_range_for_nid(unsigned int nid,
6361 unsigned long *start_pfn, unsigned long *end_pfn)
6363 unsigned long this_start_pfn, this_end_pfn;
6369 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6370 *start_pfn = min(*start_pfn, this_start_pfn);
6371 *end_pfn = max(*end_pfn, this_end_pfn);
6374 if (*start_pfn == -1UL)
6379 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6380 * assumption is made that zones within a node are ordered in monotonic
6381 * increasing memory addresses so that the "highest" populated zone is used
6383 static void __init find_usable_zone_for_movable(void)
6386 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6387 if (zone_index == ZONE_MOVABLE)
6390 if (arch_zone_highest_possible_pfn[zone_index] >
6391 arch_zone_lowest_possible_pfn[zone_index])
6395 VM_BUG_ON(zone_index == -1);
6396 movable_zone = zone_index;
6400 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6401 * because it is sized independent of architecture. Unlike the other zones,
6402 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6403 * in each node depending on the size of each node and how evenly kernelcore
6404 * is distributed. This helper function adjusts the zone ranges
6405 * provided by the architecture for a given node by using the end of the
6406 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6407 * zones within a node are in order of monotonic increases memory addresses
6409 static void __init adjust_zone_range_for_zone_movable(int nid,
6410 unsigned long zone_type,
6411 unsigned long node_start_pfn,
6412 unsigned long node_end_pfn,
6413 unsigned long *zone_start_pfn,
6414 unsigned long *zone_end_pfn)
6416 /* Only adjust if ZONE_MOVABLE is on this node */
6417 if (zone_movable_pfn[nid]) {
6418 /* Size ZONE_MOVABLE */
6419 if (zone_type == ZONE_MOVABLE) {
6420 *zone_start_pfn = zone_movable_pfn[nid];
6421 *zone_end_pfn = min(node_end_pfn,
6422 arch_zone_highest_possible_pfn[movable_zone]);
6424 /* Adjust for ZONE_MOVABLE starting within this range */
6425 } else if (!mirrored_kernelcore &&
6426 *zone_start_pfn < zone_movable_pfn[nid] &&
6427 *zone_end_pfn > zone_movable_pfn[nid]) {
6428 *zone_end_pfn = zone_movable_pfn[nid];
6430 /* Check if this whole range is within ZONE_MOVABLE */
6431 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6432 *zone_start_pfn = *zone_end_pfn;
6437 * Return the number of pages a zone spans in a node, including holes
6438 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6440 static unsigned long __init zone_spanned_pages_in_node(int nid,
6441 unsigned long zone_type,
6442 unsigned long node_start_pfn,
6443 unsigned long node_end_pfn,
6444 unsigned long *zone_start_pfn,
6445 unsigned long *zone_end_pfn)
6447 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6448 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6449 /* When hotadd a new node from cpu_up(), the node should be empty */
6450 if (!node_start_pfn && !node_end_pfn)
6453 /* Get the start and end of the zone */
6454 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6455 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6456 adjust_zone_range_for_zone_movable(nid, zone_type,
6457 node_start_pfn, node_end_pfn,
6458 zone_start_pfn, zone_end_pfn);
6460 /* Check that this node has pages within the zone's required range */
6461 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6464 /* Move the zone boundaries inside the node if necessary */
6465 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6466 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6468 /* Return the spanned pages */
6469 return *zone_end_pfn - *zone_start_pfn;
6473 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6474 * then all holes in the requested range will be accounted for.
6476 unsigned long __init __absent_pages_in_range(int nid,
6477 unsigned long range_start_pfn,
6478 unsigned long range_end_pfn)
6480 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6481 unsigned long start_pfn, end_pfn;
6484 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6485 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6486 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6487 nr_absent -= end_pfn - start_pfn;
6493 * absent_pages_in_range - Return number of page frames in holes within a range
6494 * @start_pfn: The start PFN to start searching for holes
6495 * @end_pfn: The end PFN to stop searching for holes
6497 * Return: the number of pages frames in memory holes within a range.
6499 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6500 unsigned long end_pfn)
6502 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6505 /* Return the number of page frames in holes in a zone on a node */
6506 static unsigned long __init zone_absent_pages_in_node(int nid,
6507 unsigned long zone_type,
6508 unsigned long node_start_pfn,
6509 unsigned long node_end_pfn)
6511 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6512 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6513 unsigned long zone_start_pfn, zone_end_pfn;
6514 unsigned long nr_absent;
6516 /* When hotadd a new node from cpu_up(), the node should be empty */
6517 if (!node_start_pfn && !node_end_pfn)
6520 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6521 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6523 adjust_zone_range_for_zone_movable(nid, zone_type,
6524 node_start_pfn, node_end_pfn,
6525 &zone_start_pfn, &zone_end_pfn);
6526 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6529 * ZONE_MOVABLE handling.
6530 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6533 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6534 unsigned long start_pfn, end_pfn;
6535 struct memblock_region *r;
6537 for_each_memblock(memory, r) {
6538 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6539 zone_start_pfn, zone_end_pfn);
6540 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6541 zone_start_pfn, zone_end_pfn);
6543 if (zone_type == ZONE_MOVABLE &&
6544 memblock_is_mirror(r))
6545 nr_absent += end_pfn - start_pfn;
6547 if (zone_type == ZONE_NORMAL &&
6548 !memblock_is_mirror(r))
6549 nr_absent += end_pfn - start_pfn;
6556 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6557 unsigned long node_start_pfn,
6558 unsigned long node_end_pfn)
6560 unsigned long realtotalpages = 0, totalpages = 0;
6563 for (i = 0; i < MAX_NR_ZONES; i++) {
6564 struct zone *zone = pgdat->node_zones + i;
6565 unsigned long zone_start_pfn, zone_end_pfn;
6566 unsigned long spanned, absent;
6567 unsigned long size, real_size;
6569 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
6574 absent = zone_absent_pages_in_node(pgdat->node_id, i,
6579 real_size = size - absent;
6582 zone->zone_start_pfn = zone_start_pfn;
6584 zone->zone_start_pfn = 0;
6585 zone->spanned_pages = size;
6586 zone->present_pages = real_size;
6589 realtotalpages += real_size;
6592 pgdat->node_spanned_pages = totalpages;
6593 pgdat->node_present_pages = realtotalpages;
6594 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6598 #ifndef CONFIG_SPARSEMEM
6600 * Calculate the size of the zone->blockflags rounded to an unsigned long
6601 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6602 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6603 * round what is now in bits to nearest long in bits, then return it in
6606 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6608 unsigned long usemapsize;
6610 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6611 usemapsize = roundup(zonesize, pageblock_nr_pages);
6612 usemapsize = usemapsize >> pageblock_order;
6613 usemapsize *= NR_PAGEBLOCK_BITS;
6614 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6616 return usemapsize / 8;
6619 static void __ref setup_usemap(struct pglist_data *pgdat,
6621 unsigned long zone_start_pfn,
6622 unsigned long zonesize)
6624 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6625 zone->pageblock_flags = NULL;
6627 zone->pageblock_flags =
6628 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6630 if (!zone->pageblock_flags)
6631 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6632 usemapsize, zone->name, pgdat->node_id);
6636 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6637 unsigned long zone_start_pfn, unsigned long zonesize) {}
6638 #endif /* CONFIG_SPARSEMEM */
6640 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6642 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6643 void __init set_pageblock_order(void)
6647 /* Check that pageblock_nr_pages has not already been setup */
6648 if (pageblock_order)
6651 if (HPAGE_SHIFT > PAGE_SHIFT)
6652 order = HUGETLB_PAGE_ORDER;
6654 order = MAX_ORDER - 1;
6657 * Assume the largest contiguous order of interest is a huge page.
6658 * This value may be variable depending on boot parameters on IA64 and
6661 pageblock_order = order;
6663 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6666 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6667 * is unused as pageblock_order is set at compile-time. See
6668 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6671 void __init set_pageblock_order(void)
6675 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6677 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6678 unsigned long present_pages)
6680 unsigned long pages = spanned_pages;
6683 * Provide a more accurate estimation if there are holes within
6684 * the zone and SPARSEMEM is in use. If there are holes within the
6685 * zone, each populated memory region may cost us one or two extra
6686 * memmap pages due to alignment because memmap pages for each
6687 * populated regions may not be naturally aligned on page boundary.
6688 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6690 if (spanned_pages > present_pages + (present_pages >> 4) &&
6691 IS_ENABLED(CONFIG_SPARSEMEM))
6692 pages = present_pages;
6694 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6697 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6698 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6700 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6702 spin_lock_init(&ds_queue->split_queue_lock);
6703 INIT_LIST_HEAD(&ds_queue->split_queue);
6704 ds_queue->split_queue_len = 0;
6707 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6710 #ifdef CONFIG_COMPACTION
6711 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6713 init_waitqueue_head(&pgdat->kcompactd_wait);
6716 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6719 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6721 pgdat_resize_init(pgdat);
6723 pgdat_init_split_queue(pgdat);
6724 pgdat_init_kcompactd(pgdat);
6726 init_waitqueue_head(&pgdat->kswapd_wait);
6727 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6729 pgdat_page_ext_init(pgdat);
6730 spin_lock_init(&pgdat->lru_lock);
6731 lruvec_init(&pgdat->__lruvec);
6734 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6735 unsigned long remaining_pages)
6737 atomic_long_set(&zone->managed_pages, remaining_pages);
6738 zone_set_nid(zone, nid);
6739 zone->name = zone_names[idx];
6740 zone->zone_pgdat = NODE_DATA(nid);
6741 spin_lock_init(&zone->lock);
6742 zone_seqlock_init(zone);
6743 zone_pcp_init(zone);
6747 * Set up the zone data structures
6748 * - init pgdat internals
6749 * - init all zones belonging to this node
6751 * NOTE: this function is only called during memory hotplug
6753 #ifdef CONFIG_MEMORY_HOTPLUG
6754 void __ref free_area_init_core_hotplug(int nid)
6757 pg_data_t *pgdat = NODE_DATA(nid);
6759 pgdat_init_internals(pgdat);
6760 for (z = 0; z < MAX_NR_ZONES; z++)
6761 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6766 * Set up the zone data structures:
6767 * - mark all pages reserved
6768 * - mark all memory queues empty
6769 * - clear the memory bitmaps
6771 * NOTE: pgdat should get zeroed by caller.
6772 * NOTE: this function is only called during early init.
6774 static void __init free_area_init_core(struct pglist_data *pgdat)
6777 int nid = pgdat->node_id;
6779 pgdat_init_internals(pgdat);
6780 pgdat->per_cpu_nodestats = &boot_nodestats;
6782 for (j = 0; j < MAX_NR_ZONES; j++) {
6783 struct zone *zone = pgdat->node_zones + j;
6784 unsigned long size, freesize, memmap_pages;
6785 unsigned long zone_start_pfn = zone->zone_start_pfn;
6787 size = zone->spanned_pages;
6788 freesize = zone->present_pages;
6791 * Adjust freesize so that it accounts for how much memory
6792 * is used by this zone for memmap. This affects the watermark
6793 * and per-cpu initialisations
6795 memmap_pages = calc_memmap_size(size, freesize);
6796 if (!is_highmem_idx(j)) {
6797 if (freesize >= memmap_pages) {
6798 freesize -= memmap_pages;
6801 " %s zone: %lu pages used for memmap\n",
6802 zone_names[j], memmap_pages);
6804 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6805 zone_names[j], memmap_pages, freesize);
6808 /* Account for reserved pages */
6809 if (j == 0 && freesize > dma_reserve) {
6810 freesize -= dma_reserve;
6811 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6812 zone_names[0], dma_reserve);
6815 if (!is_highmem_idx(j))
6816 nr_kernel_pages += freesize;
6817 /* Charge for highmem memmap if there are enough kernel pages */
6818 else if (nr_kernel_pages > memmap_pages * 2)
6819 nr_kernel_pages -= memmap_pages;
6820 nr_all_pages += freesize;
6823 * Set an approximate value for lowmem here, it will be adjusted
6824 * when the bootmem allocator frees pages into the buddy system.
6825 * And all highmem pages will be managed by the buddy system.
6827 zone_init_internals(zone, j, nid, freesize);
6832 set_pageblock_order();
6833 setup_usemap(pgdat, zone, zone_start_pfn, size);
6834 init_currently_empty_zone(zone, zone_start_pfn, size);
6835 memmap_init(size, nid, j, zone_start_pfn);
6839 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6840 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6842 unsigned long __maybe_unused start = 0;
6843 unsigned long __maybe_unused offset = 0;
6845 /* Skip empty nodes */
6846 if (!pgdat->node_spanned_pages)
6849 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6850 offset = pgdat->node_start_pfn - start;
6851 /* ia64 gets its own node_mem_map, before this, without bootmem */
6852 if (!pgdat->node_mem_map) {
6853 unsigned long size, end;
6857 * The zone's endpoints aren't required to be MAX_ORDER
6858 * aligned but the node_mem_map endpoints must be in order
6859 * for the buddy allocator to function correctly.
6861 end = pgdat_end_pfn(pgdat);
6862 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6863 size = (end - start) * sizeof(struct page);
6864 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6867 panic("Failed to allocate %ld bytes for node %d memory map\n",
6868 size, pgdat->node_id);
6869 pgdat->node_mem_map = map + offset;
6871 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6872 __func__, pgdat->node_id, (unsigned long)pgdat,
6873 (unsigned long)pgdat->node_mem_map);
6874 #ifndef CONFIG_NEED_MULTIPLE_NODES
6876 * With no DISCONTIG, the global mem_map is just set as node 0's
6878 if (pgdat == NODE_DATA(0)) {
6879 mem_map = NODE_DATA(0)->node_mem_map;
6880 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6886 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6887 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6889 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6890 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6892 pgdat->first_deferred_pfn = ULONG_MAX;
6895 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6898 static void __init free_area_init_node(int nid)
6900 pg_data_t *pgdat = NODE_DATA(nid);
6901 unsigned long start_pfn = 0;
6902 unsigned long end_pfn = 0;
6904 /* pg_data_t should be reset to zero when it's allocated */
6905 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
6907 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6909 pgdat->node_id = nid;
6910 pgdat->node_start_pfn = start_pfn;
6911 pgdat->per_cpu_nodestats = NULL;
6913 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6914 (u64)start_pfn << PAGE_SHIFT,
6915 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6916 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
6918 alloc_node_mem_map(pgdat);
6919 pgdat_set_deferred_range(pgdat);
6921 free_area_init_core(pgdat);
6924 void __init free_area_init_memoryless_node(int nid)
6926 free_area_init_node(nid);
6929 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6931 * Initialize all valid struct pages in the range [spfn, epfn) and mark them
6932 * PageReserved(). Return the number of struct pages that were initialized.
6934 static u64 __init init_unavailable_range(unsigned long spfn, unsigned long epfn)
6939 for (pfn = spfn; pfn < epfn; pfn++) {
6940 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6941 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6942 + pageblock_nr_pages - 1;
6946 * Use a fake node/zone (0) for now. Some of these pages
6947 * (in memblock.reserved but not in memblock.memory) will
6948 * get re-initialized via reserve_bootmem_region() later.
6950 __init_single_page(pfn_to_page(pfn), pfn, 0, 0);
6951 __SetPageReserved(pfn_to_page(pfn));
6959 * Only struct pages that are backed by physical memory are zeroed and
6960 * initialized by going through __init_single_page(). But, there are some
6961 * struct pages which are reserved in memblock allocator and their fields
6962 * may be accessed (for example page_to_pfn() on some configuration accesses
6963 * flags). We must explicitly initialize those struct pages.
6965 * This function also addresses a similar issue where struct pages are left
6966 * uninitialized because the physical address range is not covered by
6967 * memblock.memory or memblock.reserved. That could happen when memblock
6968 * layout is manually configured via memmap=, or when the highest physical
6969 * address (max_pfn) does not end on a section boundary.
6971 static void __init init_unavailable_mem(void)
6973 phys_addr_t start, end;
6975 phys_addr_t next = 0;
6978 * Loop through unavailable ranges not covered by memblock.memory.
6981 for_each_mem_range(i, &memblock.memory, NULL,
6982 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
6984 pgcnt += init_unavailable_range(PFN_DOWN(next),
6990 * Early sections always have a fully populated memmap for the whole
6991 * section - see pfn_valid(). If the last section has holes at the
6992 * end and that section is marked "online", the memmap will be
6993 * considered initialized. Make sure that memmap has a well defined
6996 pgcnt += init_unavailable_range(PFN_DOWN(next),
6997 round_up(max_pfn, PAGES_PER_SECTION));
7000 * Struct pages that do not have backing memory. This could be because
7001 * firmware is using some of this memory, or for some other reasons.
7004 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
7007 static inline void __init init_unavailable_mem(void)
7010 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
7012 #if MAX_NUMNODES > 1
7014 * Figure out the number of possible node ids.
7016 void __init setup_nr_node_ids(void)
7018 unsigned int highest;
7020 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7021 nr_node_ids = highest + 1;
7026 * node_map_pfn_alignment - determine the maximum internode alignment
7028 * This function should be called after node map is populated and sorted.
7029 * It calculates the maximum power of two alignment which can distinguish
7032 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7033 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7034 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7035 * shifted, 1GiB is enough and this function will indicate so.
7037 * This is used to test whether pfn -> nid mapping of the chosen memory
7038 * model has fine enough granularity to avoid incorrect mapping for the
7039 * populated node map.
7041 * Return: the determined alignment in pfn's. 0 if there is no alignment
7042 * requirement (single node).
7044 unsigned long __init node_map_pfn_alignment(void)
7046 unsigned long accl_mask = 0, last_end = 0;
7047 unsigned long start, end, mask;
7048 int last_nid = NUMA_NO_NODE;
7051 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7052 if (!start || last_nid < 0 || last_nid == nid) {
7059 * Start with a mask granular enough to pin-point to the
7060 * start pfn and tick off bits one-by-one until it becomes
7061 * too coarse to separate the current node from the last.
7063 mask = ~((1 << __ffs(start)) - 1);
7064 while (mask && last_end <= (start & (mask << 1)))
7067 /* accumulate all internode masks */
7071 /* convert mask to number of pages */
7072 return ~accl_mask + 1;
7076 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7078 * Return: the minimum PFN based on information provided via
7079 * memblock_set_node().
7081 unsigned long __init find_min_pfn_with_active_regions(void)
7083 return PHYS_PFN(memblock_start_of_DRAM());
7087 * early_calculate_totalpages()
7088 * Sum pages in active regions for movable zone.
7089 * Populate N_MEMORY for calculating usable_nodes.
7091 static unsigned long __init early_calculate_totalpages(void)
7093 unsigned long totalpages = 0;
7094 unsigned long start_pfn, end_pfn;
7097 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7098 unsigned long pages = end_pfn - start_pfn;
7100 totalpages += pages;
7102 node_set_state(nid, N_MEMORY);
7108 * Find the PFN the Movable zone begins in each node. Kernel memory
7109 * is spread evenly between nodes as long as the nodes have enough
7110 * memory. When they don't, some nodes will have more kernelcore than
7113 static void __init find_zone_movable_pfns_for_nodes(void)
7116 unsigned long usable_startpfn;
7117 unsigned long kernelcore_node, kernelcore_remaining;
7118 /* save the state before borrow the nodemask */
7119 nodemask_t saved_node_state = node_states[N_MEMORY];
7120 unsigned long totalpages = early_calculate_totalpages();
7121 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7122 struct memblock_region *r;
7124 /* Need to find movable_zone earlier when movable_node is specified. */
7125 find_usable_zone_for_movable();
7128 * If movable_node is specified, ignore kernelcore and movablecore
7131 if (movable_node_is_enabled()) {
7132 for_each_memblock(memory, r) {
7133 if (!memblock_is_hotpluggable(r))
7136 nid = memblock_get_region_node(r);
7138 usable_startpfn = PFN_DOWN(r->base);
7139 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7140 min(usable_startpfn, zone_movable_pfn[nid]) :
7148 * If kernelcore=mirror is specified, ignore movablecore option
7150 if (mirrored_kernelcore) {
7151 bool mem_below_4gb_not_mirrored = false;
7153 for_each_memblock(memory, r) {
7154 if (memblock_is_mirror(r))
7157 nid = memblock_get_region_node(r);
7159 usable_startpfn = memblock_region_memory_base_pfn(r);
7161 if (usable_startpfn < 0x100000) {
7162 mem_below_4gb_not_mirrored = true;
7166 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7167 min(usable_startpfn, zone_movable_pfn[nid]) :
7171 if (mem_below_4gb_not_mirrored)
7172 pr_warn("This configuration results in unmirrored kernel memory.\n");
7178 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7179 * amount of necessary memory.
7181 if (required_kernelcore_percent)
7182 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7184 if (required_movablecore_percent)
7185 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7189 * If movablecore= was specified, calculate what size of
7190 * kernelcore that corresponds so that memory usable for
7191 * any allocation type is evenly spread. If both kernelcore
7192 * and movablecore are specified, then the value of kernelcore
7193 * will be used for required_kernelcore if it's greater than
7194 * what movablecore would have allowed.
7196 if (required_movablecore) {
7197 unsigned long corepages;
7200 * Round-up so that ZONE_MOVABLE is at least as large as what
7201 * was requested by the user
7203 required_movablecore =
7204 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7205 required_movablecore = min(totalpages, required_movablecore);
7206 corepages = totalpages - required_movablecore;
7208 required_kernelcore = max(required_kernelcore, corepages);
7212 * If kernelcore was not specified or kernelcore size is larger
7213 * than totalpages, there is no ZONE_MOVABLE.
7215 if (!required_kernelcore || required_kernelcore >= totalpages)
7218 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7219 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7222 /* Spread kernelcore memory as evenly as possible throughout nodes */
7223 kernelcore_node = required_kernelcore / usable_nodes;
7224 for_each_node_state(nid, N_MEMORY) {
7225 unsigned long start_pfn, end_pfn;
7228 * Recalculate kernelcore_node if the division per node
7229 * now exceeds what is necessary to satisfy the requested
7230 * amount of memory for the kernel
7232 if (required_kernelcore < kernelcore_node)
7233 kernelcore_node = required_kernelcore / usable_nodes;
7236 * As the map is walked, we track how much memory is usable
7237 * by the kernel using kernelcore_remaining. When it is
7238 * 0, the rest of the node is usable by ZONE_MOVABLE
7240 kernelcore_remaining = kernelcore_node;
7242 /* Go through each range of PFNs within this node */
7243 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7244 unsigned long size_pages;
7246 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7247 if (start_pfn >= end_pfn)
7250 /* Account for what is only usable for kernelcore */
7251 if (start_pfn < usable_startpfn) {
7252 unsigned long kernel_pages;
7253 kernel_pages = min(end_pfn, usable_startpfn)
7256 kernelcore_remaining -= min(kernel_pages,
7257 kernelcore_remaining);
7258 required_kernelcore -= min(kernel_pages,
7259 required_kernelcore);
7261 /* Continue if range is now fully accounted */
7262 if (end_pfn <= usable_startpfn) {
7265 * Push zone_movable_pfn to the end so
7266 * that if we have to rebalance
7267 * kernelcore across nodes, we will
7268 * not double account here
7270 zone_movable_pfn[nid] = end_pfn;
7273 start_pfn = usable_startpfn;
7277 * The usable PFN range for ZONE_MOVABLE is from
7278 * start_pfn->end_pfn. Calculate size_pages as the
7279 * number of pages used as kernelcore
7281 size_pages = end_pfn - start_pfn;
7282 if (size_pages > kernelcore_remaining)
7283 size_pages = kernelcore_remaining;
7284 zone_movable_pfn[nid] = start_pfn + size_pages;
7287 * Some kernelcore has been met, update counts and
7288 * break if the kernelcore for this node has been
7291 required_kernelcore -= min(required_kernelcore,
7293 kernelcore_remaining -= size_pages;
7294 if (!kernelcore_remaining)
7300 * If there is still required_kernelcore, we do another pass with one
7301 * less node in the count. This will push zone_movable_pfn[nid] further
7302 * along on the nodes that still have memory until kernelcore is
7306 if (usable_nodes && required_kernelcore > usable_nodes)
7310 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7311 for (nid = 0; nid < MAX_NUMNODES; nid++)
7312 zone_movable_pfn[nid] =
7313 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7316 /* restore the node_state */
7317 node_states[N_MEMORY] = saved_node_state;
7320 /* Any regular or high memory on that node ? */
7321 static void check_for_memory(pg_data_t *pgdat, int nid)
7323 enum zone_type zone_type;
7325 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7326 struct zone *zone = &pgdat->node_zones[zone_type];
7327 if (populated_zone(zone)) {
7328 if (IS_ENABLED(CONFIG_HIGHMEM))
7329 node_set_state(nid, N_HIGH_MEMORY);
7330 if (zone_type <= ZONE_NORMAL)
7331 node_set_state(nid, N_NORMAL_MEMORY);
7338 * Some architecturs, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7339 * such cases we allow max_zone_pfn sorted in the descending order
7341 bool __weak arch_has_descending_max_zone_pfns(void)
7347 * free_area_init - Initialise all pg_data_t and zone data
7348 * @max_zone_pfn: an array of max PFNs for each zone
7350 * This will call free_area_init_node() for each active node in the system.
7351 * Using the page ranges provided by memblock_set_node(), the size of each
7352 * zone in each node and their holes is calculated. If the maximum PFN
7353 * between two adjacent zones match, it is assumed that the zone is empty.
7354 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7355 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7356 * starts where the previous one ended. For example, ZONE_DMA32 starts
7357 * at arch_max_dma_pfn.
7359 void __init free_area_init(unsigned long *max_zone_pfn)
7361 unsigned long start_pfn, end_pfn;
7365 /* Record where the zone boundaries are */
7366 memset(arch_zone_lowest_possible_pfn, 0,
7367 sizeof(arch_zone_lowest_possible_pfn));
7368 memset(arch_zone_highest_possible_pfn, 0,
7369 sizeof(arch_zone_highest_possible_pfn));
7371 start_pfn = find_min_pfn_with_active_regions();
7372 descending = arch_has_descending_max_zone_pfns();
7374 for (i = 0; i < MAX_NR_ZONES; i++) {
7376 zone = MAX_NR_ZONES - i - 1;
7380 if (zone == ZONE_MOVABLE)
7383 end_pfn = max(max_zone_pfn[zone], start_pfn);
7384 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7385 arch_zone_highest_possible_pfn[zone] = end_pfn;
7387 start_pfn = end_pfn;
7390 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7391 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7392 find_zone_movable_pfns_for_nodes();
7394 /* Print out the zone ranges */
7395 pr_info("Zone ranges:\n");
7396 for (i = 0; i < MAX_NR_ZONES; i++) {
7397 if (i == ZONE_MOVABLE)
7399 pr_info(" %-8s ", zone_names[i]);
7400 if (arch_zone_lowest_possible_pfn[i] ==
7401 arch_zone_highest_possible_pfn[i])
7404 pr_cont("[mem %#018Lx-%#018Lx]\n",
7405 (u64)arch_zone_lowest_possible_pfn[i]
7407 ((u64)arch_zone_highest_possible_pfn[i]
7408 << PAGE_SHIFT) - 1);
7411 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7412 pr_info("Movable zone start for each node\n");
7413 for (i = 0; i < MAX_NUMNODES; i++) {
7414 if (zone_movable_pfn[i])
7415 pr_info(" Node %d: %#018Lx\n", i,
7416 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7420 * Print out the early node map, and initialize the
7421 * subsection-map relative to active online memory ranges to
7422 * enable future "sub-section" extensions of the memory map.
7424 pr_info("Early memory node ranges\n");
7425 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7426 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7427 (u64)start_pfn << PAGE_SHIFT,
7428 ((u64)end_pfn << PAGE_SHIFT) - 1);
7429 subsection_map_init(start_pfn, end_pfn - start_pfn);
7432 /* Initialise every node */
7433 mminit_verify_pageflags_layout();
7434 setup_nr_node_ids();
7435 init_unavailable_mem();
7436 for_each_online_node(nid) {
7437 pg_data_t *pgdat = NODE_DATA(nid);
7438 free_area_init_node(nid);
7440 /* Any memory on that node */
7441 if (pgdat->node_present_pages)
7442 node_set_state(nid, N_MEMORY);
7443 check_for_memory(pgdat, nid);
7447 static int __init cmdline_parse_core(char *p, unsigned long *core,
7448 unsigned long *percent)
7450 unsigned long long coremem;
7456 /* Value may be a percentage of total memory, otherwise bytes */
7457 coremem = simple_strtoull(p, &endptr, 0);
7458 if (*endptr == '%') {
7459 /* Paranoid check for percent values greater than 100 */
7460 WARN_ON(coremem > 100);
7464 coremem = memparse(p, &p);
7465 /* Paranoid check that UL is enough for the coremem value */
7466 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7468 *core = coremem >> PAGE_SHIFT;
7475 * kernelcore=size sets the amount of memory for use for allocations that
7476 * cannot be reclaimed or migrated.
7478 static int __init cmdline_parse_kernelcore(char *p)
7480 /* parse kernelcore=mirror */
7481 if (parse_option_str(p, "mirror")) {
7482 mirrored_kernelcore = true;
7486 return cmdline_parse_core(p, &required_kernelcore,
7487 &required_kernelcore_percent);
7491 * movablecore=size sets the amount of memory for use for allocations that
7492 * can be reclaimed or migrated.
7494 static int __init cmdline_parse_movablecore(char *p)
7496 return cmdline_parse_core(p, &required_movablecore,
7497 &required_movablecore_percent);
7500 early_param("kernelcore", cmdline_parse_kernelcore);
7501 early_param("movablecore", cmdline_parse_movablecore);
7503 void adjust_managed_page_count(struct page *page, long count)
7505 atomic_long_add(count, &page_zone(page)->managed_pages);
7506 totalram_pages_add(count);
7507 #ifdef CONFIG_HIGHMEM
7508 if (PageHighMem(page))
7509 totalhigh_pages_add(count);
7512 EXPORT_SYMBOL(adjust_managed_page_count);
7514 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7517 unsigned long pages = 0;
7519 start = (void *)PAGE_ALIGN((unsigned long)start);
7520 end = (void *)((unsigned long)end & PAGE_MASK);
7521 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7522 struct page *page = virt_to_page(pos);
7523 void *direct_map_addr;
7526 * 'direct_map_addr' might be different from 'pos'
7527 * because some architectures' virt_to_page()
7528 * work with aliases. Getting the direct map
7529 * address ensures that we get a _writeable_
7530 * alias for the memset().
7532 direct_map_addr = page_address(page);
7533 if ((unsigned int)poison <= 0xFF)
7534 memset(direct_map_addr, poison, PAGE_SIZE);
7536 free_reserved_page(page);
7540 pr_info("Freeing %s memory: %ldK\n",
7541 s, pages << (PAGE_SHIFT - 10));
7546 #ifdef CONFIG_HIGHMEM
7547 void free_highmem_page(struct page *page)
7549 __free_reserved_page(page);
7550 totalram_pages_inc();
7551 atomic_long_inc(&page_zone(page)->managed_pages);
7552 totalhigh_pages_inc();
7557 void __init mem_init_print_info(const char *str)
7559 unsigned long physpages, codesize, datasize, rosize, bss_size;
7560 unsigned long init_code_size, init_data_size;
7562 physpages = get_num_physpages();
7563 codesize = _etext - _stext;
7564 datasize = _edata - _sdata;
7565 rosize = __end_rodata - __start_rodata;
7566 bss_size = __bss_stop - __bss_start;
7567 init_data_size = __init_end - __init_begin;
7568 init_code_size = _einittext - _sinittext;
7571 * Detect special cases and adjust section sizes accordingly:
7572 * 1) .init.* may be embedded into .data sections
7573 * 2) .init.text.* may be out of [__init_begin, __init_end],
7574 * please refer to arch/tile/kernel/vmlinux.lds.S.
7575 * 3) .rodata.* may be embedded into .text or .data sections.
7577 #define adj_init_size(start, end, size, pos, adj) \
7579 if (start <= pos && pos < end && size > adj) \
7583 adj_init_size(__init_begin, __init_end, init_data_size,
7584 _sinittext, init_code_size);
7585 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7586 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7587 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7588 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7590 #undef adj_init_size
7592 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7593 #ifdef CONFIG_HIGHMEM
7597 nr_free_pages() << (PAGE_SHIFT - 10),
7598 physpages << (PAGE_SHIFT - 10),
7599 codesize >> 10, datasize >> 10, rosize >> 10,
7600 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7601 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7602 totalcma_pages << (PAGE_SHIFT - 10),
7603 #ifdef CONFIG_HIGHMEM
7604 totalhigh_pages() << (PAGE_SHIFT - 10),
7606 str ? ", " : "", str ? str : "");
7610 * set_dma_reserve - set the specified number of pages reserved in the first zone
7611 * @new_dma_reserve: The number of pages to mark reserved
7613 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7614 * In the DMA zone, a significant percentage may be consumed by kernel image
7615 * and other unfreeable allocations which can skew the watermarks badly. This
7616 * function may optionally be used to account for unfreeable pages in the
7617 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7618 * smaller per-cpu batchsize.
7620 void __init set_dma_reserve(unsigned long new_dma_reserve)
7622 dma_reserve = new_dma_reserve;
7625 static int page_alloc_cpu_dead(unsigned int cpu)
7628 lru_add_drain_cpu(cpu);
7632 * Spill the event counters of the dead processor
7633 * into the current processors event counters.
7634 * This artificially elevates the count of the current
7637 vm_events_fold_cpu(cpu);
7640 * Zero the differential counters of the dead processor
7641 * so that the vm statistics are consistent.
7643 * This is only okay since the processor is dead and cannot
7644 * race with what we are doing.
7646 cpu_vm_stats_fold(cpu);
7651 int hashdist = HASHDIST_DEFAULT;
7653 static int __init set_hashdist(char *str)
7657 hashdist = simple_strtoul(str, &str, 0);
7660 __setup("hashdist=", set_hashdist);
7663 void __init page_alloc_init(void)
7668 if (num_node_state(N_MEMORY) == 1)
7672 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7673 "mm/page_alloc:dead", NULL,
7674 page_alloc_cpu_dead);
7679 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7680 * or min_free_kbytes changes.
7682 static void calculate_totalreserve_pages(void)
7684 struct pglist_data *pgdat;
7685 unsigned long reserve_pages = 0;
7686 enum zone_type i, j;
7688 for_each_online_pgdat(pgdat) {
7690 pgdat->totalreserve_pages = 0;
7692 for (i = 0; i < MAX_NR_ZONES; i++) {
7693 struct zone *zone = pgdat->node_zones + i;
7695 unsigned long managed_pages = zone_managed_pages(zone);
7697 /* Find valid and maximum lowmem_reserve in the zone */
7698 for (j = i; j < MAX_NR_ZONES; j++) {
7699 if (zone->lowmem_reserve[j] > max)
7700 max = zone->lowmem_reserve[j];
7703 /* we treat the high watermark as reserved pages. */
7704 max += high_wmark_pages(zone);
7706 if (max > managed_pages)
7707 max = managed_pages;
7709 pgdat->totalreserve_pages += max;
7711 reserve_pages += max;
7714 totalreserve_pages = reserve_pages;
7718 * setup_per_zone_lowmem_reserve - called whenever
7719 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7720 * has a correct pages reserved value, so an adequate number of
7721 * pages are left in the zone after a successful __alloc_pages().
7723 static void setup_per_zone_lowmem_reserve(void)
7725 struct pglist_data *pgdat;
7726 enum zone_type j, idx;
7728 for_each_online_pgdat(pgdat) {
7729 for (j = 0; j < MAX_NR_ZONES; j++) {
7730 struct zone *zone = pgdat->node_zones + j;
7731 unsigned long managed_pages = zone_managed_pages(zone);
7733 zone->lowmem_reserve[j] = 0;
7737 struct zone *lower_zone;
7740 lower_zone = pgdat->node_zones + idx;
7742 if (!sysctl_lowmem_reserve_ratio[idx] ||
7743 !zone_managed_pages(lower_zone)) {
7744 lower_zone->lowmem_reserve[j] = 0;
7747 lower_zone->lowmem_reserve[j] =
7748 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7750 managed_pages += zone_managed_pages(lower_zone);
7755 /* update totalreserve_pages */
7756 calculate_totalreserve_pages();
7759 static void __setup_per_zone_wmarks(void)
7761 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7762 unsigned long lowmem_pages = 0;
7764 unsigned long flags;
7766 /* Calculate total number of !ZONE_HIGHMEM pages */
7767 for_each_zone(zone) {
7768 if (!is_highmem(zone))
7769 lowmem_pages += zone_managed_pages(zone);
7772 for_each_zone(zone) {
7775 spin_lock_irqsave(&zone->lock, flags);
7776 tmp = (u64)pages_min * zone_managed_pages(zone);
7777 do_div(tmp, lowmem_pages);
7778 if (is_highmem(zone)) {
7780 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7781 * need highmem pages, so cap pages_min to a small
7784 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7785 * deltas control async page reclaim, and so should
7786 * not be capped for highmem.
7788 unsigned long min_pages;
7790 min_pages = zone_managed_pages(zone) / 1024;
7791 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7792 zone->_watermark[WMARK_MIN] = min_pages;
7795 * If it's a lowmem zone, reserve a number of pages
7796 * proportionate to the zone's size.
7798 zone->_watermark[WMARK_MIN] = tmp;
7802 * Set the kswapd watermarks distance according to the
7803 * scale factor in proportion to available memory, but
7804 * ensure a minimum size on small systems.
7806 tmp = max_t(u64, tmp >> 2,
7807 mult_frac(zone_managed_pages(zone),
7808 watermark_scale_factor, 10000));
7810 zone->watermark_boost = 0;
7811 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7812 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7814 spin_unlock_irqrestore(&zone->lock, flags);
7817 /* update totalreserve_pages */
7818 calculate_totalreserve_pages();
7822 * setup_per_zone_wmarks - called when min_free_kbytes changes
7823 * or when memory is hot-{added|removed}
7825 * Ensures that the watermark[min,low,high] values for each zone are set
7826 * correctly with respect to min_free_kbytes.
7828 void setup_per_zone_wmarks(void)
7830 static DEFINE_SPINLOCK(lock);
7833 __setup_per_zone_wmarks();
7838 * Initialise min_free_kbytes.
7840 * For small machines we want it small (128k min). For large machines
7841 * we want it large (256MB max). But it is not linear, because network
7842 * bandwidth does not increase linearly with machine size. We use
7844 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7845 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7861 int __meminit init_per_zone_wmark_min(void)
7863 unsigned long lowmem_kbytes;
7864 int new_min_free_kbytes;
7866 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7867 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7869 if (new_min_free_kbytes > user_min_free_kbytes) {
7870 min_free_kbytes = new_min_free_kbytes;
7871 if (min_free_kbytes < 128)
7872 min_free_kbytes = 128;
7873 if (min_free_kbytes > 262144)
7874 min_free_kbytes = 262144;
7876 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7877 new_min_free_kbytes, user_min_free_kbytes);
7879 setup_per_zone_wmarks();
7880 refresh_zone_stat_thresholds();
7881 setup_per_zone_lowmem_reserve();
7884 setup_min_unmapped_ratio();
7885 setup_min_slab_ratio();
7890 core_initcall(init_per_zone_wmark_min)
7893 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7894 * that we can call two helper functions whenever min_free_kbytes
7897 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7898 void *buffer, size_t *length, loff_t *ppos)
7902 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7907 user_min_free_kbytes = min_free_kbytes;
7908 setup_per_zone_wmarks();
7913 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7914 void *buffer, size_t *length, loff_t *ppos)
7918 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7923 setup_per_zone_wmarks();
7929 static void setup_min_unmapped_ratio(void)
7934 for_each_online_pgdat(pgdat)
7935 pgdat->min_unmapped_pages = 0;
7938 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7939 sysctl_min_unmapped_ratio) / 100;
7943 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7944 void *buffer, size_t *length, loff_t *ppos)
7948 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7952 setup_min_unmapped_ratio();
7957 static void setup_min_slab_ratio(void)
7962 for_each_online_pgdat(pgdat)
7963 pgdat->min_slab_pages = 0;
7966 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7967 sysctl_min_slab_ratio) / 100;
7970 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7971 void *buffer, size_t *length, loff_t *ppos)
7975 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7979 setup_min_slab_ratio();
7986 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7987 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7988 * whenever sysctl_lowmem_reserve_ratio changes.
7990 * The reserve ratio obviously has absolutely no relation with the
7991 * minimum watermarks. The lowmem reserve ratio can only make sense
7992 * if in function of the boot time zone sizes.
7994 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7995 void *buffer, size_t *length, loff_t *ppos)
7999 proc_dointvec_minmax(table, write, buffer, length, ppos);
8001 for (i = 0; i < MAX_NR_ZONES; i++) {
8002 if (sysctl_lowmem_reserve_ratio[i] < 1)
8003 sysctl_lowmem_reserve_ratio[i] = 0;
8006 setup_per_zone_lowmem_reserve();
8010 static void __zone_pcp_update(struct zone *zone)
8014 for_each_possible_cpu(cpu)
8015 pageset_set_high_and_batch(zone,
8016 per_cpu_ptr(zone->pageset, cpu));
8020 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8021 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8022 * pagelist can have before it gets flushed back to buddy allocator.
8024 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8025 void *buffer, size_t *length, loff_t *ppos)
8028 int old_percpu_pagelist_fraction;
8031 mutex_lock(&pcp_batch_high_lock);
8032 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8034 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8035 if (!write || ret < 0)
8038 /* Sanity checking to avoid pcp imbalance */
8039 if (percpu_pagelist_fraction &&
8040 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8041 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8047 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8050 for_each_populated_zone(zone)
8051 __zone_pcp_update(zone);
8053 mutex_unlock(&pcp_batch_high_lock);
8057 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8059 * Returns the number of pages that arch has reserved but
8060 * is not known to alloc_large_system_hash().
8062 static unsigned long __init arch_reserved_kernel_pages(void)
8069 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8070 * machines. As memory size is increased the scale is also increased but at
8071 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8072 * quadruples the scale is increased by one, which means the size of hash table
8073 * only doubles, instead of quadrupling as well.
8074 * Because 32-bit systems cannot have large physical memory, where this scaling
8075 * makes sense, it is disabled on such platforms.
8077 #if __BITS_PER_LONG > 32
8078 #define ADAPT_SCALE_BASE (64ul << 30)
8079 #define ADAPT_SCALE_SHIFT 2
8080 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8084 * allocate a large system hash table from bootmem
8085 * - it is assumed that the hash table must contain an exact power-of-2
8086 * quantity of entries
8087 * - limit is the number of hash buckets, not the total allocation size
8089 void *__init alloc_large_system_hash(const char *tablename,
8090 unsigned long bucketsize,
8091 unsigned long numentries,
8094 unsigned int *_hash_shift,
8095 unsigned int *_hash_mask,
8096 unsigned long low_limit,
8097 unsigned long high_limit)
8099 unsigned long long max = high_limit;
8100 unsigned long log2qty, size;
8105 /* allow the kernel cmdline to have a say */
8107 /* round applicable memory size up to nearest megabyte */
8108 numentries = nr_kernel_pages;
8109 numentries -= arch_reserved_kernel_pages();
8111 /* It isn't necessary when PAGE_SIZE >= 1MB */
8112 if (PAGE_SHIFT < 20)
8113 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8115 #if __BITS_PER_LONG > 32
8117 unsigned long adapt;
8119 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8120 adapt <<= ADAPT_SCALE_SHIFT)
8125 /* limit to 1 bucket per 2^scale bytes of low memory */
8126 if (scale > PAGE_SHIFT)
8127 numentries >>= (scale - PAGE_SHIFT);
8129 numentries <<= (PAGE_SHIFT - scale);
8131 /* Make sure we've got at least a 0-order allocation.. */
8132 if (unlikely(flags & HASH_SMALL)) {
8133 /* Makes no sense without HASH_EARLY */
8134 WARN_ON(!(flags & HASH_EARLY));
8135 if (!(numentries >> *_hash_shift)) {
8136 numentries = 1UL << *_hash_shift;
8137 BUG_ON(!numentries);
8139 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8140 numentries = PAGE_SIZE / bucketsize;
8142 numentries = roundup_pow_of_two(numentries);
8144 /* limit allocation size to 1/16 total memory by default */
8146 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8147 do_div(max, bucketsize);
8149 max = min(max, 0x80000000ULL);
8151 if (numentries < low_limit)
8152 numentries = low_limit;
8153 if (numentries > max)
8156 log2qty = ilog2(numentries);
8158 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8161 size = bucketsize << log2qty;
8162 if (flags & HASH_EARLY) {
8163 if (flags & HASH_ZERO)
8164 table = memblock_alloc(size, SMP_CACHE_BYTES);
8166 table = memblock_alloc_raw(size,
8168 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8169 table = __vmalloc(size, gfp_flags);
8173 * If bucketsize is not a power-of-two, we may free
8174 * some pages at the end of hash table which
8175 * alloc_pages_exact() automatically does
8177 table = alloc_pages_exact(size, gfp_flags);
8178 kmemleak_alloc(table, size, 1, gfp_flags);
8180 } while (!table && size > PAGE_SIZE && --log2qty);
8183 panic("Failed to allocate %s hash table\n", tablename);
8185 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8186 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8187 virt ? "vmalloc" : "linear");
8190 *_hash_shift = log2qty;
8192 *_hash_mask = (1 << log2qty) - 1;
8198 * This function checks whether pageblock includes unmovable pages or not.
8200 * PageLRU check without isolation or lru_lock could race so that
8201 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8202 * check without lock_page also may miss some movable non-lru pages at
8203 * race condition. So you can't expect this function should be exact.
8205 * Returns a page without holding a reference. If the caller wants to
8206 * dereference that page (e.g., dumping), it has to make sure that that it
8207 * cannot get removed (e.g., via memory unplug) concurrently.
8210 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8211 int migratetype, int flags)
8213 unsigned long iter = 0;
8214 unsigned long pfn = page_to_pfn(page);
8217 * TODO we could make this much more efficient by not checking every
8218 * page in the range if we know all of them are in MOVABLE_ZONE and
8219 * that the movable zone guarantees that pages are migratable but
8220 * the later is not the case right now unfortunatelly. E.g. movablecore
8221 * can still lead to having bootmem allocations in zone_movable.
8224 if (is_migrate_cma_page(page)) {
8226 * CMA allocations (alloc_contig_range) really need to mark
8227 * isolate CMA pageblocks even when they are not movable in fact
8228 * so consider them movable here.
8230 if (is_migrate_cma(migratetype))
8236 for (; iter < pageblock_nr_pages; iter++) {
8237 if (!pfn_valid_within(pfn + iter))
8240 page = pfn_to_page(pfn + iter);
8242 if (PageReserved(page))
8246 * If the zone is movable and we have ruled out all reserved
8247 * pages then it should be reasonably safe to assume the rest
8250 if (zone_idx(zone) == ZONE_MOVABLE)
8254 * Hugepages are not in LRU lists, but they're movable.
8255 * THPs are on the LRU, but need to be counted as #small pages.
8256 * We need not scan over tail pages because we don't
8257 * handle each tail page individually in migration.
8259 if (PageHuge(page) || PageTransCompound(page)) {
8260 struct page *head = compound_head(page);
8261 unsigned int skip_pages;
8263 if (PageHuge(page)) {
8264 if (!hugepage_migration_supported(page_hstate(head)))
8266 } else if (!PageLRU(head) && !__PageMovable(head)) {
8270 skip_pages = compound_nr(head) - (page - head);
8271 iter += skip_pages - 1;
8276 * We can't use page_count without pin a page
8277 * because another CPU can free compound page.
8278 * This check already skips compound tails of THP
8279 * because their page->_refcount is zero at all time.
8281 if (!page_ref_count(page)) {
8282 if (PageBuddy(page))
8283 iter += (1 << page_order(page)) - 1;
8288 * The HWPoisoned page may be not in buddy system, and
8289 * page_count() is not 0.
8291 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8295 * We treat all PageOffline() pages as movable when offlining
8296 * to give drivers a chance to decrement their reference count
8297 * in MEM_GOING_OFFLINE in order to indicate that these pages
8298 * can be offlined as there are no direct references anymore.
8299 * For actually unmovable PageOffline() where the driver does
8300 * not support this, we will fail later when trying to actually
8301 * move these pages that still have a reference count > 0.
8302 * (false negatives in this function only)
8304 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8307 if (__PageMovable(page) || PageLRU(page))
8311 * If there are RECLAIMABLE pages, we need to check
8312 * it. But now, memory offline itself doesn't call
8313 * shrink_node_slabs() and it still to be fixed.
8316 * If the page is not RAM, page_count()should be 0.
8317 * we don't need more check. This is an _used_ not-movable page.
8319 * The problematic thing here is PG_reserved pages. PG_reserved
8320 * is set to both of a memory hole page and a _used_ kernel
8328 #ifdef CONFIG_CONTIG_ALLOC
8329 static unsigned long pfn_max_align_down(unsigned long pfn)
8331 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8332 pageblock_nr_pages) - 1);
8335 static unsigned long pfn_max_align_up(unsigned long pfn)
8337 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8338 pageblock_nr_pages));
8341 /* [start, end) must belong to a single zone. */
8342 static int __alloc_contig_migrate_range(struct compact_control *cc,
8343 unsigned long start, unsigned long end)
8345 /* This function is based on compact_zone() from compaction.c. */
8346 unsigned int nr_reclaimed;
8347 unsigned long pfn = start;
8348 unsigned int tries = 0;
8353 while (pfn < end || !list_empty(&cc->migratepages)) {
8354 if (fatal_signal_pending(current)) {
8359 if (list_empty(&cc->migratepages)) {
8360 cc->nr_migratepages = 0;
8361 pfn = isolate_migratepages_range(cc, pfn, end);
8367 } else if (++tries == 5) {
8368 ret = ret < 0 ? ret : -EBUSY;
8372 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8374 cc->nr_migratepages -= nr_reclaimed;
8376 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
8377 NULL, 0, cc->mode, MR_CONTIG_RANGE);
8380 putback_movable_pages(&cc->migratepages);
8387 * alloc_contig_range() -- tries to allocate given range of pages
8388 * @start: start PFN to allocate
8389 * @end: one-past-the-last PFN to allocate
8390 * @migratetype: migratetype of the underlaying pageblocks (either
8391 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8392 * in range must have the same migratetype and it must
8393 * be either of the two.
8394 * @gfp_mask: GFP mask to use during compaction
8396 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8397 * aligned. The PFN range must belong to a single zone.
8399 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8400 * pageblocks in the range. Once isolated, the pageblocks should not
8401 * be modified by others.
8403 * Return: zero on success or negative error code. On success all
8404 * pages which PFN is in [start, end) are allocated for the caller and
8405 * need to be freed with free_contig_range().
8407 int alloc_contig_range(unsigned long start, unsigned long end,
8408 unsigned migratetype, gfp_t gfp_mask)
8410 unsigned long outer_start, outer_end;
8414 struct compact_control cc = {
8415 .nr_migratepages = 0,
8417 .zone = page_zone(pfn_to_page(start)),
8418 .mode = MIGRATE_SYNC,
8419 .ignore_skip_hint = true,
8420 .no_set_skip_hint = true,
8421 .gfp_mask = current_gfp_context(gfp_mask),
8422 .alloc_contig = true,
8424 INIT_LIST_HEAD(&cc.migratepages);
8427 * What we do here is we mark all pageblocks in range as
8428 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8429 * have different sizes, and due to the way page allocator
8430 * work, we align the range to biggest of the two pages so
8431 * that page allocator won't try to merge buddies from
8432 * different pageblocks and change MIGRATE_ISOLATE to some
8433 * other migration type.
8435 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8436 * migrate the pages from an unaligned range (ie. pages that
8437 * we are interested in). This will put all the pages in
8438 * range back to page allocator as MIGRATE_ISOLATE.
8440 * When this is done, we take the pages in range from page
8441 * allocator removing them from the buddy system. This way
8442 * page allocator will never consider using them.
8444 * This lets us mark the pageblocks back as
8445 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8446 * aligned range but not in the unaligned, original range are
8447 * put back to page allocator so that buddy can use them.
8450 ret = start_isolate_page_range(pfn_max_align_down(start),
8451 pfn_max_align_up(end), migratetype, 0);
8456 * In case of -EBUSY, we'd like to know which page causes problem.
8457 * So, just fall through. test_pages_isolated() has a tracepoint
8458 * which will report the busy page.
8460 * It is possible that busy pages could become available before
8461 * the call to test_pages_isolated, and the range will actually be
8462 * allocated. So, if we fall through be sure to clear ret so that
8463 * -EBUSY is not accidentally used or returned to caller.
8465 ret = __alloc_contig_migrate_range(&cc, start, end);
8466 if (ret && ret != -EBUSY)
8471 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8472 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8473 * more, all pages in [start, end) are free in page allocator.
8474 * What we are going to do is to allocate all pages from
8475 * [start, end) (that is remove them from page allocator).
8477 * The only problem is that pages at the beginning and at the
8478 * end of interesting range may be not aligned with pages that
8479 * page allocator holds, ie. they can be part of higher order
8480 * pages. Because of this, we reserve the bigger range and
8481 * once this is done free the pages we are not interested in.
8483 * We don't have to hold zone->lock here because the pages are
8484 * isolated thus they won't get removed from buddy.
8487 lru_add_drain_all();
8490 outer_start = start;
8491 while (!PageBuddy(pfn_to_page(outer_start))) {
8492 if (++order >= MAX_ORDER) {
8493 outer_start = start;
8496 outer_start &= ~0UL << order;
8499 if (outer_start != start) {
8500 order = page_order(pfn_to_page(outer_start));
8503 * outer_start page could be small order buddy page and
8504 * it doesn't include start page. Adjust outer_start
8505 * in this case to report failed page properly
8506 * on tracepoint in test_pages_isolated()
8508 if (outer_start + (1UL << order) <= start)
8509 outer_start = start;
8512 /* Make sure the range is really isolated. */
8513 if (test_pages_isolated(outer_start, end, 0)) {
8514 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8515 __func__, outer_start, end);
8520 /* Grab isolated pages from freelists. */
8521 outer_end = isolate_freepages_range(&cc, outer_start, end);
8527 /* Free head and tail (if any) */
8528 if (start != outer_start)
8529 free_contig_range(outer_start, start - outer_start);
8530 if (end != outer_end)
8531 free_contig_range(end, outer_end - end);
8534 undo_isolate_page_range(pfn_max_align_down(start),
8535 pfn_max_align_up(end), migratetype);
8538 EXPORT_SYMBOL(alloc_contig_range);
8540 static int __alloc_contig_pages(unsigned long start_pfn,
8541 unsigned long nr_pages, gfp_t gfp_mask)
8543 unsigned long end_pfn = start_pfn + nr_pages;
8545 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8549 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8550 unsigned long nr_pages)
8552 unsigned long i, end_pfn = start_pfn + nr_pages;
8555 for (i = start_pfn; i < end_pfn; i++) {
8556 page = pfn_to_online_page(i);
8560 if (page_zone(page) != z)
8563 if (PageReserved(page))
8566 if (page_count(page) > 0)
8575 static bool zone_spans_last_pfn(const struct zone *zone,
8576 unsigned long start_pfn, unsigned long nr_pages)
8578 unsigned long last_pfn = start_pfn + nr_pages - 1;
8580 return zone_spans_pfn(zone, last_pfn);
8584 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8585 * @nr_pages: Number of contiguous pages to allocate
8586 * @gfp_mask: GFP mask to limit search and used during compaction
8588 * @nodemask: Mask for other possible nodes
8590 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8591 * on an applicable zonelist to find a contiguous pfn range which can then be
8592 * tried for allocation with alloc_contig_range(). This routine is intended
8593 * for allocation requests which can not be fulfilled with the buddy allocator.
8595 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8596 * power of two then the alignment is guaranteed to be to the given nr_pages
8597 * (e.g. 1GB request would be aligned to 1GB).
8599 * Allocated pages can be freed with free_contig_range() or by manually calling
8600 * __free_page() on each allocated page.
8602 * Return: pointer to contiguous pages on success, or NULL if not successful.
8604 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8605 int nid, nodemask_t *nodemask)
8607 unsigned long ret, pfn, flags;
8608 struct zonelist *zonelist;
8612 zonelist = node_zonelist(nid, gfp_mask);
8613 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8614 gfp_zone(gfp_mask), nodemask) {
8615 spin_lock_irqsave(&zone->lock, flags);
8617 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8618 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8619 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8621 * We release the zone lock here because
8622 * alloc_contig_range() will also lock the zone
8623 * at some point. If there's an allocation
8624 * spinning on this lock, it may win the race
8625 * and cause alloc_contig_range() to fail...
8627 spin_unlock_irqrestore(&zone->lock, flags);
8628 ret = __alloc_contig_pages(pfn, nr_pages,
8631 return pfn_to_page(pfn);
8632 spin_lock_irqsave(&zone->lock, flags);
8636 spin_unlock_irqrestore(&zone->lock, flags);
8640 #endif /* CONFIG_CONTIG_ALLOC */
8642 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8644 unsigned int count = 0;
8646 for (; nr_pages--; pfn++) {
8647 struct page *page = pfn_to_page(pfn);
8649 count += page_count(page) != 1;
8652 WARN(count != 0, "%d pages are still in use!\n", count);
8654 EXPORT_SYMBOL(free_contig_range);
8657 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8658 * page high values need to be recalulated.
8660 void __meminit zone_pcp_update(struct zone *zone)
8662 mutex_lock(&pcp_batch_high_lock);
8663 __zone_pcp_update(zone);
8664 mutex_unlock(&pcp_batch_high_lock);
8667 void zone_pcp_reset(struct zone *zone)
8669 unsigned long flags;
8671 struct per_cpu_pageset *pset;
8673 /* avoid races with drain_pages() */
8674 local_irq_save(flags);
8675 if (zone->pageset != &boot_pageset) {
8676 for_each_online_cpu(cpu) {
8677 pset = per_cpu_ptr(zone->pageset, cpu);
8678 drain_zonestat(zone, pset);
8680 free_percpu(zone->pageset);
8681 zone->pageset = &boot_pageset;
8683 local_irq_restore(flags);
8686 #ifdef CONFIG_MEMORY_HOTREMOVE
8688 * All pages in the range must be in a single zone and isolated
8689 * before calling this.
8692 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8698 unsigned long flags;
8699 unsigned long offlined_pages = 0;
8701 /* find the first valid pfn */
8702 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8706 return offlined_pages;
8708 offline_mem_sections(pfn, end_pfn);
8709 zone = page_zone(pfn_to_page(pfn));
8710 spin_lock_irqsave(&zone->lock, flags);
8712 while (pfn < end_pfn) {
8713 if (!pfn_valid(pfn)) {
8717 page = pfn_to_page(pfn);
8719 * The HWPoisoned page may be not in buddy system, and
8720 * page_count() is not 0.
8722 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8728 * At this point all remaining PageOffline() pages have a
8729 * reference count of 0 and can simply be skipped.
8731 if (PageOffline(page)) {
8732 BUG_ON(page_count(page));
8733 BUG_ON(PageBuddy(page));
8739 BUG_ON(page_count(page));
8740 BUG_ON(!PageBuddy(page));
8741 order = page_order(page);
8742 offlined_pages += 1 << order;
8743 del_page_from_free_list(page, zone, order);
8744 pfn += (1 << order);
8746 spin_unlock_irqrestore(&zone->lock, flags);
8748 return offlined_pages;
8752 bool is_free_buddy_page(struct page *page)
8754 struct zone *zone = page_zone(page);
8755 unsigned long pfn = page_to_pfn(page);
8756 unsigned long flags;
8759 spin_lock_irqsave(&zone->lock, flags);
8760 for (order = 0; order < MAX_ORDER; order++) {
8761 struct page *page_head = page - (pfn & ((1 << order) - 1));
8763 if (PageBuddy(page_head) && page_order(page_head) >= order)
8766 spin_unlock_irqrestore(&zone->lock, flags);
8768 return order < MAX_ORDER;
8771 #ifdef CONFIG_MEMORY_FAILURE
8773 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8774 * test is performed under the zone lock to prevent a race against page
8777 bool set_hwpoison_free_buddy_page(struct page *page)
8779 struct zone *zone = page_zone(page);
8780 unsigned long pfn = page_to_pfn(page);
8781 unsigned long flags;
8783 bool hwpoisoned = false;
8785 spin_lock_irqsave(&zone->lock, flags);
8786 for (order = 0; order < MAX_ORDER; order++) {
8787 struct page *page_head = page - (pfn & ((1 << order) - 1));
8789 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8790 if (!TestSetPageHWPoison(page))
8795 spin_unlock_irqrestore(&zone->lock, flags);