1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/page_alloc.c
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
18 #include <linux/stddef.h>
20 #include <linux/highmem.h>
21 #include <linux/swap.h>
22 #include <linux/interrupt.h>
23 #include <linux/pagemap.h>
24 #include <linux/jiffies.h>
25 #include <linux/memblock.h>
26 #include <linux/compiler.h>
27 #include <linux/kernel.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/memremap.h>
46 #include <linux/stop_machine.h>
47 #include <linux/random.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <trace/events/oom.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/sched/mm.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
66 #include <linux/memcontrol.h>
67 #include <linux/ftrace.h>
68 #include <linux/lockdep.h>
69 #include <linux/nmi.h>
70 #include <linux/psi.h>
72 #include <asm/sections.h>
73 #include <asm/tlbflush.h>
74 #include <asm/div64.h>
78 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
79 static DEFINE_MUTEX(pcp_batch_high_lock);
80 #define MIN_PERCPU_PAGELIST_FRACTION (8)
82 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
83 DEFINE_PER_CPU(int, numa_node);
84 EXPORT_PER_CPU_SYMBOL(numa_node);
87 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
89 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
91 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
92 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
93 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
94 * defined in <linux/topology.h>.
96 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
97 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
98 int _node_numa_mem_[MAX_NUMNODES];
101 /* work_structs for global per-cpu drains */
104 struct work_struct work;
106 DEFINE_MUTEX(pcpu_drain_mutex);
107 DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
109 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
110 volatile unsigned long latent_entropy __latent_entropy;
111 EXPORT_SYMBOL(latent_entropy);
115 * Array of node states.
117 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
118 [N_POSSIBLE] = NODE_MASK_ALL,
119 [N_ONLINE] = { { [0] = 1UL } },
121 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
122 #ifdef CONFIG_HIGHMEM
123 [N_HIGH_MEMORY] = { { [0] = 1UL } },
125 [N_MEMORY] = { { [0] = 1UL } },
126 [N_CPU] = { { [0] = 1UL } },
129 EXPORT_SYMBOL(node_states);
131 atomic_long_t _totalram_pages __read_mostly;
132 EXPORT_SYMBOL(_totalram_pages);
133 unsigned long totalreserve_pages __read_mostly;
134 unsigned long totalcma_pages __read_mostly;
136 int percpu_pagelist_fraction;
137 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
138 #ifdef CONFIG_INIT_ON_ALLOC_DEFAULT_ON
139 DEFINE_STATIC_KEY_TRUE(init_on_alloc);
141 DEFINE_STATIC_KEY_FALSE(init_on_alloc);
143 EXPORT_SYMBOL(init_on_alloc);
145 #ifdef CONFIG_INIT_ON_FREE_DEFAULT_ON
146 DEFINE_STATIC_KEY_TRUE(init_on_free);
148 DEFINE_STATIC_KEY_FALSE(init_on_free);
150 EXPORT_SYMBOL(init_on_free);
152 static int __init early_init_on_alloc(char *buf)
159 ret = kstrtobool(buf, &bool_result);
160 if (bool_result && page_poisoning_enabled())
161 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_alloc\n");
163 static_branch_enable(&init_on_alloc);
165 static_branch_disable(&init_on_alloc);
168 early_param("init_on_alloc", early_init_on_alloc);
170 static int __init early_init_on_free(char *buf)
177 ret = kstrtobool(buf, &bool_result);
178 if (bool_result && page_poisoning_enabled())
179 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_free\n");
181 static_branch_enable(&init_on_free);
183 static_branch_disable(&init_on_free);
186 early_param("init_on_free", early_init_on_free);
189 * A cached value of the page's pageblock's migratetype, used when the page is
190 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
191 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
192 * Also the migratetype set in the page does not necessarily match the pcplist
193 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
194 * other index - this ensures that it will be put on the correct CMA freelist.
196 static inline int get_pcppage_migratetype(struct page *page)
201 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
203 page->index = migratetype;
206 #ifdef CONFIG_PM_SLEEP
208 * The following functions are used by the suspend/hibernate code to temporarily
209 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
210 * while devices are suspended. To avoid races with the suspend/hibernate code,
211 * they should always be called with system_transition_mutex held
212 * (gfp_allowed_mask also should only be modified with system_transition_mutex
213 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
214 * with that modification).
217 static gfp_t saved_gfp_mask;
219 void pm_restore_gfp_mask(void)
221 WARN_ON(!mutex_is_locked(&system_transition_mutex));
222 if (saved_gfp_mask) {
223 gfp_allowed_mask = saved_gfp_mask;
228 void pm_restrict_gfp_mask(void)
230 WARN_ON(!mutex_is_locked(&system_transition_mutex));
231 WARN_ON(saved_gfp_mask);
232 saved_gfp_mask = gfp_allowed_mask;
233 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
236 bool pm_suspended_storage(void)
238 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
242 #endif /* CONFIG_PM_SLEEP */
244 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
245 unsigned int pageblock_order __read_mostly;
248 static void __free_pages_ok(struct page *page, unsigned int order);
251 * results with 256, 32 in the lowmem_reserve sysctl:
252 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
253 * 1G machine -> (16M dma, 784M normal, 224M high)
254 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
255 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
256 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
258 * TBD: should special case ZONE_DMA32 machines here - in those we normally
259 * don't need any ZONE_NORMAL reservation
261 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
262 #ifdef CONFIG_ZONE_DMA
265 #ifdef CONFIG_ZONE_DMA32
269 #ifdef CONFIG_HIGHMEM
275 static char * const zone_names[MAX_NR_ZONES] = {
276 #ifdef CONFIG_ZONE_DMA
279 #ifdef CONFIG_ZONE_DMA32
283 #ifdef CONFIG_HIGHMEM
287 #ifdef CONFIG_ZONE_DEVICE
292 const char * const migratetype_names[MIGRATE_TYPES] = {
300 #ifdef CONFIG_MEMORY_ISOLATION
305 compound_page_dtor * const compound_page_dtors[] = {
308 #ifdef CONFIG_HUGETLB_PAGE
311 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
316 int min_free_kbytes = 1024;
317 int user_min_free_kbytes = -1;
318 #ifdef CONFIG_DISCONTIGMEM
320 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
321 * are not on separate NUMA nodes. Functionally this works but with
322 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
323 * quite small. By default, do not boost watermarks on discontigmem as in
324 * many cases very high-order allocations like THP are likely to be
325 * unsupported and the premature reclaim offsets the advantage of long-term
326 * fragmentation avoidance.
328 int watermark_boost_factor __read_mostly;
330 int watermark_boost_factor __read_mostly = 15000;
332 int watermark_scale_factor = 10;
334 static unsigned long nr_kernel_pages __initdata;
335 static unsigned long nr_all_pages __initdata;
336 static unsigned long dma_reserve __initdata;
338 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
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);
351 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
354 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
355 unsigned int nr_online_nodes __read_mostly = 1;
356 EXPORT_SYMBOL(nr_node_ids);
357 EXPORT_SYMBOL(nr_online_nodes);
360 int page_group_by_mobility_disabled __read_mostly;
362 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
364 * During boot we initialize deferred pages on-demand, as needed, but once
365 * page_alloc_init_late() has finished, the deferred pages are all initialized,
366 * and we can permanently disable that path.
368 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
371 * Calling kasan_free_pages() only after deferred memory initialization
372 * has completed. Poisoning pages during deferred memory init will greatly
373 * lengthen the process and cause problem in large memory systems as the
374 * deferred pages initialization is done with interrupt disabled.
376 * Assuming that there will be no reference to those newly initialized
377 * pages before they are ever allocated, this should have no effect on
378 * KASAN memory tracking as the poison will be properly inserted at page
379 * allocation time. The only corner case is when pages are allocated by
380 * on-demand allocation and then freed again before the deferred pages
381 * initialization is done, but this is not likely to happen.
383 static inline void kasan_free_nondeferred_pages(struct page *page, int order)
385 if (!static_branch_unlikely(&deferred_pages))
386 kasan_free_pages(page, order);
389 /* Returns true if the struct page for the pfn is uninitialised */
390 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
392 int nid = early_pfn_to_nid(pfn);
394 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
401 * Returns true when the remaining initialisation should be deferred until
402 * later in the boot cycle when it can be parallelised.
404 static bool __meminit
405 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
407 static unsigned long prev_end_pfn, nr_initialised;
410 * prev_end_pfn static that contains the end of previous zone
411 * No need to protect because called very early in boot before smp_init.
413 if (prev_end_pfn != end_pfn) {
414 prev_end_pfn = end_pfn;
418 /* Always populate low zones for address-constrained allocations */
419 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
423 * We start only with one section of pages, more pages are added as
424 * needed until the rest of deferred pages are initialized.
427 if ((nr_initialised > PAGES_PER_SECTION) &&
428 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
429 NODE_DATA(nid)->first_deferred_pfn = pfn;
435 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
437 static inline bool early_page_uninitialised(unsigned long pfn)
442 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
448 /* Return a pointer to the bitmap storing bits affecting a block of pages */
449 static inline unsigned long *get_pageblock_bitmap(struct page *page,
452 #ifdef CONFIG_SPARSEMEM
453 return section_to_usemap(__pfn_to_section(pfn));
455 return page_zone(page)->pageblock_flags;
456 #endif /* CONFIG_SPARSEMEM */
459 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
461 #ifdef CONFIG_SPARSEMEM
462 pfn &= (PAGES_PER_SECTION-1);
463 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
465 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
466 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
467 #endif /* CONFIG_SPARSEMEM */
471 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
472 * @page: The page within the block of interest
473 * @pfn: The target page frame number
474 * @end_bitidx: The last bit of interest to retrieve
475 * @mask: mask of bits that the caller is interested in
477 * Return: pageblock_bits flags
479 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
481 unsigned long end_bitidx,
484 unsigned long *bitmap;
485 unsigned long bitidx, word_bitidx;
488 bitmap = get_pageblock_bitmap(page, pfn);
489 bitidx = pfn_to_bitidx(page, pfn);
490 word_bitidx = bitidx / BITS_PER_LONG;
491 bitidx &= (BITS_PER_LONG-1);
493 word = bitmap[word_bitidx];
494 bitidx += end_bitidx;
495 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
498 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
499 unsigned long end_bitidx,
502 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
505 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
507 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
511 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
512 * @page: The page within the block of interest
513 * @flags: The flags to set
514 * @pfn: The target page frame number
515 * @end_bitidx: The last bit of interest
516 * @mask: mask of bits that the caller is interested in
518 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
520 unsigned long end_bitidx,
523 unsigned long *bitmap;
524 unsigned long bitidx, word_bitidx;
525 unsigned long old_word, word;
527 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
528 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
530 bitmap = get_pageblock_bitmap(page, pfn);
531 bitidx = pfn_to_bitidx(page, pfn);
532 word_bitidx = bitidx / BITS_PER_LONG;
533 bitidx &= (BITS_PER_LONG-1);
535 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
537 bitidx += end_bitidx;
538 mask <<= (BITS_PER_LONG - bitidx - 1);
539 flags <<= (BITS_PER_LONG - bitidx - 1);
541 word = READ_ONCE(bitmap[word_bitidx]);
543 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
544 if (word == old_word)
550 void set_pageblock_migratetype(struct page *page, int migratetype)
552 if (unlikely(page_group_by_mobility_disabled &&
553 migratetype < MIGRATE_PCPTYPES))
554 migratetype = MIGRATE_UNMOVABLE;
556 set_pageblock_flags_group(page, (unsigned long)migratetype,
557 PB_migrate, PB_migrate_end);
560 #ifdef CONFIG_DEBUG_VM
561 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
565 unsigned long pfn = page_to_pfn(page);
566 unsigned long sp, start_pfn;
569 seq = zone_span_seqbegin(zone);
570 start_pfn = zone->zone_start_pfn;
571 sp = zone->spanned_pages;
572 if (!zone_spans_pfn(zone, pfn))
574 } while (zone_span_seqretry(zone, seq));
577 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
578 pfn, zone_to_nid(zone), zone->name,
579 start_pfn, start_pfn + sp);
584 static int page_is_consistent(struct zone *zone, struct page *page)
586 if (!pfn_valid_within(page_to_pfn(page)))
588 if (zone != page_zone(page))
594 * Temporary debugging check for pages not lying within a given zone.
596 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
598 if (page_outside_zone_boundaries(zone, page))
600 if (!page_is_consistent(zone, page))
606 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
612 static void bad_page(struct page *page, const char *reason,
613 unsigned long bad_flags)
615 static unsigned long resume;
616 static unsigned long nr_shown;
617 static unsigned long nr_unshown;
620 * Allow a burst of 60 reports, then keep quiet for that minute;
621 * or allow a steady drip of one report per second.
623 if (nr_shown == 60) {
624 if (time_before(jiffies, resume)) {
630 "BUG: Bad page state: %lu messages suppressed\n",
637 resume = jiffies + 60 * HZ;
639 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
640 current->comm, page_to_pfn(page));
641 __dump_page(page, reason);
642 bad_flags &= page->flags;
644 pr_alert("bad because of flags: %#lx(%pGp)\n",
645 bad_flags, &bad_flags);
646 dump_page_owner(page);
651 /* Leave bad fields for debug, except PageBuddy could make trouble */
652 page_mapcount_reset(page); /* remove PageBuddy */
653 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
657 * Higher-order pages are called "compound pages". They are structured thusly:
659 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
661 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
662 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
664 * The first tail page's ->compound_dtor holds the offset in array of compound
665 * page destructors. See compound_page_dtors.
667 * The first tail page's ->compound_order holds the order of allocation.
668 * This usage means that zero-order pages may not be compound.
671 void free_compound_page(struct page *page)
673 __free_pages_ok(page, compound_order(page));
676 void prep_compound_page(struct page *page, unsigned int order)
679 int nr_pages = 1 << order;
681 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
682 set_compound_order(page, order);
684 for (i = 1; i < nr_pages; i++) {
685 struct page *p = page + i;
686 set_page_count(p, 0);
687 p->mapping = TAIL_MAPPING;
688 set_compound_head(p, page);
690 atomic_set(compound_mapcount_ptr(page), -1);
693 #ifdef CONFIG_DEBUG_PAGEALLOC
694 unsigned int _debug_guardpage_minorder;
696 #ifdef CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
697 DEFINE_STATIC_KEY_TRUE(_debug_pagealloc_enabled);
699 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
701 EXPORT_SYMBOL(_debug_pagealloc_enabled);
703 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
705 static int __init early_debug_pagealloc(char *buf)
709 if (kstrtobool(buf, &enable))
713 static_branch_enable(&_debug_pagealloc_enabled);
717 early_param("debug_pagealloc", early_debug_pagealloc);
719 static void init_debug_guardpage(void)
721 if (!debug_pagealloc_enabled())
724 if (!debug_guardpage_minorder())
727 static_branch_enable(&_debug_guardpage_enabled);
730 static int __init debug_guardpage_minorder_setup(char *buf)
734 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
735 pr_err("Bad debug_guardpage_minorder value\n");
738 _debug_guardpage_minorder = res;
739 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
742 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
744 static inline bool set_page_guard(struct zone *zone, struct page *page,
745 unsigned int order, int migratetype)
747 if (!debug_guardpage_enabled())
750 if (order >= debug_guardpage_minorder())
753 __SetPageGuard(page);
754 INIT_LIST_HEAD(&page->lru);
755 set_page_private(page, order);
756 /* Guard pages are not available for any usage */
757 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
762 static inline void clear_page_guard(struct zone *zone, struct page *page,
763 unsigned int order, int migratetype)
765 if (!debug_guardpage_enabled())
768 __ClearPageGuard(page);
770 set_page_private(page, 0);
771 if (!is_migrate_isolate(migratetype))
772 __mod_zone_freepage_state(zone, (1 << order), migratetype);
775 static inline bool set_page_guard(struct zone *zone, struct page *page,
776 unsigned int order, int migratetype) { return false; }
777 static inline void clear_page_guard(struct zone *zone, struct page *page,
778 unsigned int order, int migratetype) {}
781 static inline void set_page_order(struct page *page, unsigned int order)
783 set_page_private(page, order);
784 __SetPageBuddy(page);
788 * This function checks whether a page is free && is the buddy
789 * we can coalesce a page and its buddy if
790 * (a) the buddy is not in a hole (check before calling!) &&
791 * (b) the buddy is in the buddy system &&
792 * (c) a page and its buddy have the same order &&
793 * (d) a page and its buddy are in the same zone.
795 * For recording whether a page is in the buddy system, we set PageBuddy.
796 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
798 * For recording page's order, we use page_private(page).
800 static inline int page_is_buddy(struct page *page, struct page *buddy,
803 if (page_is_guard(buddy) && page_order(buddy) == order) {
804 if (page_zone_id(page) != page_zone_id(buddy))
807 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
812 if (PageBuddy(buddy) && page_order(buddy) == order) {
814 * zone check is done late to avoid uselessly
815 * calculating zone/node ids for pages that could
818 if (page_zone_id(page) != page_zone_id(buddy))
821 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
828 #ifdef CONFIG_COMPACTION
829 static inline struct capture_control *task_capc(struct zone *zone)
831 struct capture_control *capc = current->capture_control;
834 !(current->flags & PF_KTHREAD) &&
836 capc->cc->zone == zone &&
837 capc->cc->direct_compaction ? capc : NULL;
841 compaction_capture(struct capture_control *capc, struct page *page,
842 int order, int migratetype)
844 if (!capc || order != capc->cc->order)
847 /* Do not accidentally pollute CMA or isolated regions*/
848 if (is_migrate_cma(migratetype) ||
849 is_migrate_isolate(migratetype))
853 * Do not let lower order allocations polluate a movable pageblock.
854 * This might let an unmovable request use a reclaimable pageblock
855 * and vice-versa but no more than normal fallback logic which can
856 * have trouble finding a high-order free page.
858 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
866 static inline struct capture_control *task_capc(struct zone *zone)
872 compaction_capture(struct capture_control *capc, struct page *page,
873 int order, int migratetype)
877 #endif /* CONFIG_COMPACTION */
880 * Freeing function for a buddy system allocator.
882 * The concept of a buddy system is to maintain direct-mapped table
883 * (containing bit values) for memory blocks of various "orders".
884 * The bottom level table contains the map for the smallest allocatable
885 * units of memory (here, pages), and each level above it describes
886 * pairs of units from the levels below, hence, "buddies".
887 * At a high level, all that happens here is marking the table entry
888 * at the bottom level available, and propagating the changes upward
889 * as necessary, plus some accounting needed to play nicely with other
890 * parts of the VM system.
891 * At each level, we keep a list of pages, which are heads of continuous
892 * free pages of length of (1 << order) and marked with PageBuddy.
893 * Page's order is recorded in page_private(page) field.
894 * So when we are allocating or freeing one, we can derive the state of the
895 * other. That is, if we allocate a small block, and both were
896 * free, the remainder of the region must be split into blocks.
897 * If a block is freed, and its buddy is also free, then this
898 * triggers coalescing into a block of larger size.
903 static inline void __free_one_page(struct page *page,
905 struct zone *zone, unsigned int order,
908 unsigned long combined_pfn;
909 unsigned long uninitialized_var(buddy_pfn);
911 unsigned int max_order;
912 struct capture_control *capc = task_capc(zone);
914 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
916 VM_BUG_ON(!zone_is_initialized(zone));
917 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
919 VM_BUG_ON(migratetype == -1);
920 if (likely(!is_migrate_isolate(migratetype)))
921 __mod_zone_freepage_state(zone, 1 << order, migratetype);
923 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
924 VM_BUG_ON_PAGE(bad_range(zone, page), page);
927 while (order < max_order - 1) {
928 if (compaction_capture(capc, page, order, migratetype)) {
929 __mod_zone_freepage_state(zone, -(1 << order),
933 buddy_pfn = __find_buddy_pfn(pfn, order);
934 buddy = page + (buddy_pfn - pfn);
936 if (!pfn_valid_within(buddy_pfn))
938 if (!page_is_buddy(page, buddy, order))
941 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
942 * merge with it and move up one order.
944 if (page_is_guard(buddy))
945 clear_page_guard(zone, buddy, order, migratetype);
947 del_page_from_free_area(buddy, &zone->free_area[order]);
948 combined_pfn = buddy_pfn & pfn;
949 page = page + (combined_pfn - pfn);
953 if (max_order < MAX_ORDER) {
954 /* If we are here, it means order is >= pageblock_order.
955 * We want to prevent merge between freepages on isolate
956 * pageblock and normal pageblock. Without this, pageblock
957 * isolation could cause incorrect freepage or CMA accounting.
959 * We don't want to hit this code for the more frequent
962 if (unlikely(has_isolate_pageblock(zone))) {
965 buddy_pfn = __find_buddy_pfn(pfn, order);
966 buddy = page + (buddy_pfn - pfn);
967 buddy_mt = get_pageblock_migratetype(buddy);
969 if (migratetype != buddy_mt
970 && (is_migrate_isolate(migratetype) ||
971 is_migrate_isolate(buddy_mt)))
975 goto continue_merging;
979 set_page_order(page, order);
982 * If this is not the largest possible page, check if the buddy
983 * of the next-highest order is free. If it is, it's possible
984 * that pages are being freed that will coalesce soon. In case,
985 * that is happening, add the free page to the tail of the list
986 * so it's less likely to be used soon and more likely to be merged
987 * as a higher order page
989 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)
990 && !is_shuffle_order(order)) {
991 struct page *higher_page, *higher_buddy;
992 combined_pfn = buddy_pfn & pfn;
993 higher_page = page + (combined_pfn - pfn);
994 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
995 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
996 if (pfn_valid_within(buddy_pfn) &&
997 page_is_buddy(higher_page, higher_buddy, order + 1)) {
998 add_to_free_area_tail(page, &zone->free_area[order],
1004 if (is_shuffle_order(order))
1005 add_to_free_area_random(page, &zone->free_area[order],
1008 add_to_free_area(page, &zone->free_area[order], migratetype);
1013 * A bad page could be due to a number of fields. Instead of multiple branches,
1014 * try and check multiple fields with one check. The caller must do a detailed
1015 * check if necessary.
1017 static inline bool page_expected_state(struct page *page,
1018 unsigned long check_flags)
1020 if (unlikely(atomic_read(&page->_mapcount) != -1))
1023 if (unlikely((unsigned long)page->mapping |
1024 page_ref_count(page) |
1026 (unsigned long)page->mem_cgroup |
1028 (page->flags & check_flags)))
1034 static void free_pages_check_bad(struct page *page)
1036 const char *bad_reason;
1037 unsigned long bad_flags;
1042 if (unlikely(atomic_read(&page->_mapcount) != -1))
1043 bad_reason = "nonzero mapcount";
1044 if (unlikely(page->mapping != NULL))
1045 bad_reason = "non-NULL mapping";
1046 if (unlikely(page_ref_count(page) != 0))
1047 bad_reason = "nonzero _refcount";
1048 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
1049 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1050 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
1053 if (unlikely(page->mem_cgroup))
1054 bad_reason = "page still charged to cgroup";
1056 bad_page(page, bad_reason, bad_flags);
1059 static inline int free_pages_check(struct page *page)
1061 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1064 /* Something has gone sideways, find it */
1065 free_pages_check_bad(page);
1069 static int free_tail_pages_check(struct page *head_page, struct page *page)
1074 * We rely page->lru.next never has bit 0 set, unless the page
1075 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1077 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1079 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1083 switch (page - head_page) {
1085 /* the first tail page: ->mapping may be compound_mapcount() */
1086 if (unlikely(compound_mapcount(page))) {
1087 bad_page(page, "nonzero compound_mapcount", 0);
1093 * the second tail page: ->mapping is
1094 * deferred_list.next -- ignore value.
1098 if (page->mapping != TAIL_MAPPING) {
1099 bad_page(page, "corrupted mapping in tail page", 0);
1104 if (unlikely(!PageTail(page))) {
1105 bad_page(page, "PageTail not set", 0);
1108 if (unlikely(compound_head(page) != head_page)) {
1109 bad_page(page, "compound_head not consistent", 0);
1114 page->mapping = NULL;
1115 clear_compound_head(page);
1119 static void kernel_init_free_pages(struct page *page, int numpages)
1123 for (i = 0; i < numpages; i++)
1124 clear_highpage(page + i);
1127 static __always_inline bool free_pages_prepare(struct page *page,
1128 unsigned int order, bool check_free)
1132 VM_BUG_ON_PAGE(PageTail(page), page);
1134 trace_mm_page_free(page, order);
1137 * Check tail pages before head page information is cleared to
1138 * avoid checking PageCompound for order-0 pages.
1140 if (unlikely(order)) {
1141 bool compound = PageCompound(page);
1144 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1147 ClearPageDoubleMap(page);
1148 for (i = 1; i < (1 << order); i++) {
1150 bad += free_tail_pages_check(page, page + i);
1151 if (unlikely(free_pages_check(page + i))) {
1155 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1158 if (PageMappingFlags(page))
1159 page->mapping = NULL;
1160 if (memcg_kmem_enabled() && PageKmemcg(page))
1161 __memcg_kmem_uncharge(page, order);
1163 bad += free_pages_check(page);
1167 page_cpupid_reset_last(page);
1168 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1169 reset_page_owner(page, order);
1171 if (!PageHighMem(page)) {
1172 debug_check_no_locks_freed(page_address(page),
1173 PAGE_SIZE << order);
1174 debug_check_no_obj_freed(page_address(page),
1175 PAGE_SIZE << order);
1177 arch_free_page(page, order);
1178 if (want_init_on_free())
1179 kernel_init_free_pages(page, 1 << order);
1181 kernel_poison_pages(page, 1 << order, 0);
1182 if (debug_pagealloc_enabled())
1183 kernel_map_pages(page, 1 << order, 0);
1185 kasan_free_nondeferred_pages(page, order);
1190 #ifdef CONFIG_DEBUG_VM
1192 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1193 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1194 * moved from pcp lists to free lists.
1196 static bool free_pcp_prepare(struct page *page)
1198 return free_pages_prepare(page, 0, true);
1201 static bool bulkfree_pcp_prepare(struct page *page)
1203 if (debug_pagealloc_enabled())
1204 return free_pages_check(page);
1210 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1211 * moving from pcp lists to free list in order to reduce overhead. With
1212 * debug_pagealloc enabled, they are checked also immediately when being freed
1215 static bool free_pcp_prepare(struct page *page)
1217 if (debug_pagealloc_enabled())
1218 return free_pages_prepare(page, 0, true);
1220 return free_pages_prepare(page, 0, false);
1223 static bool bulkfree_pcp_prepare(struct page *page)
1225 return free_pages_check(page);
1227 #endif /* CONFIG_DEBUG_VM */
1229 static inline void prefetch_buddy(struct page *page)
1231 unsigned long pfn = page_to_pfn(page);
1232 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1233 struct page *buddy = page + (buddy_pfn - pfn);
1239 * Frees a number of pages from the PCP lists
1240 * Assumes all pages on list are in same zone, and of same order.
1241 * count is the number of pages to free.
1243 * If the zone was previously in an "all pages pinned" state then look to
1244 * see if this freeing clears that state.
1246 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1247 * pinned" detection logic.
1249 static void free_pcppages_bulk(struct zone *zone, int count,
1250 struct per_cpu_pages *pcp)
1252 int migratetype = 0;
1254 int prefetch_nr = 0;
1255 bool isolated_pageblocks;
1256 struct page *page, *tmp;
1260 struct list_head *list;
1263 * Remove pages from lists in a round-robin fashion. A
1264 * batch_free count is maintained that is incremented when an
1265 * empty list is encountered. This is so more pages are freed
1266 * off fuller lists instead of spinning excessively around empty
1271 if (++migratetype == MIGRATE_PCPTYPES)
1273 list = &pcp->lists[migratetype];
1274 } while (list_empty(list));
1276 /* This is the only non-empty list. Free them all. */
1277 if (batch_free == MIGRATE_PCPTYPES)
1281 page = list_last_entry(list, struct page, lru);
1282 /* must delete to avoid corrupting pcp list */
1283 list_del(&page->lru);
1286 if (bulkfree_pcp_prepare(page))
1289 list_add_tail(&page->lru, &head);
1292 * We are going to put the page back to the global
1293 * pool, prefetch its buddy to speed up later access
1294 * under zone->lock. It is believed the overhead of
1295 * an additional test and calculating buddy_pfn here
1296 * can be offset by reduced memory latency later. To
1297 * avoid excessive prefetching due to large count, only
1298 * prefetch buddy for the first pcp->batch nr of pages.
1300 if (prefetch_nr++ < pcp->batch)
1301 prefetch_buddy(page);
1302 } while (--count && --batch_free && !list_empty(list));
1305 spin_lock(&zone->lock);
1306 isolated_pageblocks = has_isolate_pageblock(zone);
1309 * Use safe version since after __free_one_page(),
1310 * page->lru.next will not point to original list.
1312 list_for_each_entry_safe(page, tmp, &head, lru) {
1313 int mt = get_pcppage_migratetype(page);
1314 /* MIGRATE_ISOLATE page should not go to pcplists */
1315 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1316 /* Pageblock could have been isolated meanwhile */
1317 if (unlikely(isolated_pageblocks))
1318 mt = get_pageblock_migratetype(page);
1320 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1321 trace_mm_page_pcpu_drain(page, 0, mt);
1323 spin_unlock(&zone->lock);
1326 static void free_one_page(struct zone *zone,
1327 struct page *page, unsigned long pfn,
1331 spin_lock(&zone->lock);
1332 if (unlikely(has_isolate_pageblock(zone) ||
1333 is_migrate_isolate(migratetype))) {
1334 migratetype = get_pfnblock_migratetype(page, pfn);
1336 __free_one_page(page, pfn, zone, order, migratetype);
1337 spin_unlock(&zone->lock);
1340 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1341 unsigned long zone, int nid)
1343 mm_zero_struct_page(page);
1344 set_page_links(page, zone, nid, pfn);
1345 init_page_count(page);
1346 page_mapcount_reset(page);
1347 page_cpupid_reset_last(page);
1348 page_kasan_tag_reset(page);
1350 INIT_LIST_HEAD(&page->lru);
1351 #ifdef WANT_PAGE_VIRTUAL
1352 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1353 if (!is_highmem_idx(zone))
1354 set_page_address(page, __va(pfn << PAGE_SHIFT));
1358 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1359 static void __meminit init_reserved_page(unsigned long pfn)
1364 if (!early_page_uninitialised(pfn))
1367 nid = early_pfn_to_nid(pfn);
1368 pgdat = NODE_DATA(nid);
1370 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1371 struct zone *zone = &pgdat->node_zones[zid];
1373 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1376 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1379 static inline void init_reserved_page(unsigned long pfn)
1382 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1385 * Initialised pages do not have PageReserved set. This function is
1386 * called for each range allocated by the bootmem allocator and
1387 * marks the pages PageReserved. The remaining valid pages are later
1388 * sent to the buddy page allocator.
1390 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1392 unsigned long start_pfn = PFN_DOWN(start);
1393 unsigned long end_pfn = PFN_UP(end);
1395 for (; start_pfn < end_pfn; start_pfn++) {
1396 if (pfn_valid(start_pfn)) {
1397 struct page *page = pfn_to_page(start_pfn);
1399 init_reserved_page(start_pfn);
1401 /* Avoid false-positive PageTail() */
1402 INIT_LIST_HEAD(&page->lru);
1405 * no need for atomic set_bit because the struct
1406 * page is not visible yet so nobody should
1409 __SetPageReserved(page);
1414 static void __free_pages_ok(struct page *page, unsigned int order)
1416 unsigned long flags;
1418 unsigned long pfn = page_to_pfn(page);
1420 if (!free_pages_prepare(page, order, true))
1423 migratetype = get_pfnblock_migratetype(page, pfn);
1424 local_irq_save(flags);
1425 __count_vm_events(PGFREE, 1 << order);
1426 free_one_page(page_zone(page), page, pfn, order, migratetype);
1427 local_irq_restore(flags);
1430 void __free_pages_core(struct page *page, unsigned int order)
1432 unsigned int nr_pages = 1 << order;
1433 struct page *p = page;
1437 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1439 __ClearPageReserved(p);
1440 set_page_count(p, 0);
1442 __ClearPageReserved(p);
1443 set_page_count(p, 0);
1445 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1446 set_page_refcounted(page);
1447 __free_pages(page, order);
1450 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1451 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1453 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1455 int __meminit early_pfn_to_nid(unsigned long pfn)
1457 static DEFINE_SPINLOCK(early_pfn_lock);
1460 spin_lock(&early_pfn_lock);
1461 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1463 nid = first_online_node;
1464 spin_unlock(&early_pfn_lock);
1470 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1471 /* Only safe to use early in boot when initialisation is single-threaded */
1472 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1476 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1477 if (nid >= 0 && nid != node)
1483 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1490 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1493 if (early_page_uninitialised(pfn))
1495 __free_pages_core(page, order);
1499 * Check that the whole (or subset of) a pageblock given by the interval of
1500 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1501 * with the migration of free compaction scanner. The scanners then need to
1502 * use only pfn_valid_within() check for arches that allow holes within
1505 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1507 * It's possible on some configurations to have a setup like node0 node1 node0
1508 * i.e. it's possible that all pages within a zones range of pages do not
1509 * belong to a single zone. We assume that a border between node0 and node1
1510 * can occur within a single pageblock, but not a node0 node1 node0
1511 * interleaving within a single pageblock. It is therefore sufficient to check
1512 * the first and last page of a pageblock and avoid checking each individual
1513 * page in a pageblock.
1515 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1516 unsigned long end_pfn, struct zone *zone)
1518 struct page *start_page;
1519 struct page *end_page;
1521 /* end_pfn is one past the range we are checking */
1524 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1527 start_page = pfn_to_online_page(start_pfn);
1531 if (page_zone(start_page) != zone)
1534 end_page = pfn_to_page(end_pfn);
1536 /* This gives a shorter code than deriving page_zone(end_page) */
1537 if (page_zone_id(start_page) != page_zone_id(end_page))
1543 void set_zone_contiguous(struct zone *zone)
1545 unsigned long block_start_pfn = zone->zone_start_pfn;
1546 unsigned long block_end_pfn;
1548 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1549 for (; block_start_pfn < zone_end_pfn(zone);
1550 block_start_pfn = block_end_pfn,
1551 block_end_pfn += pageblock_nr_pages) {
1553 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1555 if (!__pageblock_pfn_to_page(block_start_pfn,
1556 block_end_pfn, zone))
1560 /* We confirm that there is no hole */
1561 zone->contiguous = true;
1564 void clear_zone_contiguous(struct zone *zone)
1566 zone->contiguous = false;
1569 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1570 static void __init deferred_free_range(unsigned long pfn,
1571 unsigned long nr_pages)
1579 page = pfn_to_page(pfn);
1581 /* Free a large naturally-aligned chunk if possible */
1582 if (nr_pages == pageblock_nr_pages &&
1583 (pfn & (pageblock_nr_pages - 1)) == 0) {
1584 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1585 __free_pages_core(page, pageblock_order);
1589 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1590 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1591 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1592 __free_pages_core(page, 0);
1596 /* Completion tracking for deferred_init_memmap() threads */
1597 static atomic_t pgdat_init_n_undone __initdata;
1598 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1600 static inline void __init pgdat_init_report_one_done(void)
1602 if (atomic_dec_and_test(&pgdat_init_n_undone))
1603 complete(&pgdat_init_all_done_comp);
1607 * Returns true if page needs to be initialized or freed to buddy allocator.
1609 * First we check if pfn is valid on architectures where it is possible to have
1610 * holes within pageblock_nr_pages. On systems where it is not possible, this
1611 * function is optimized out.
1613 * Then, we check if a current large page is valid by only checking the validity
1616 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1618 if (!pfn_valid_within(pfn))
1620 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1626 * Free pages to buddy allocator. Try to free aligned pages in
1627 * pageblock_nr_pages sizes.
1629 static void __init deferred_free_pages(unsigned long pfn,
1630 unsigned long end_pfn)
1632 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1633 unsigned long nr_free = 0;
1635 for (; pfn < end_pfn; pfn++) {
1636 if (!deferred_pfn_valid(pfn)) {
1637 deferred_free_range(pfn - nr_free, nr_free);
1639 } else if (!(pfn & nr_pgmask)) {
1640 deferred_free_range(pfn - nr_free, nr_free);
1642 touch_nmi_watchdog();
1647 /* Free the last block of pages to allocator */
1648 deferred_free_range(pfn - nr_free, nr_free);
1652 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1653 * by performing it only once every pageblock_nr_pages.
1654 * Return number of pages initialized.
1656 static unsigned long __init deferred_init_pages(struct zone *zone,
1658 unsigned long end_pfn)
1660 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1661 int nid = zone_to_nid(zone);
1662 unsigned long nr_pages = 0;
1663 int zid = zone_idx(zone);
1664 struct page *page = NULL;
1666 for (; pfn < end_pfn; pfn++) {
1667 if (!deferred_pfn_valid(pfn)) {
1670 } else if (!page || !(pfn & nr_pgmask)) {
1671 page = pfn_to_page(pfn);
1672 touch_nmi_watchdog();
1676 __init_single_page(page, pfn, zid, nid);
1683 * This function is meant to pre-load the iterator for the zone init.
1684 * Specifically it walks through the ranges until we are caught up to the
1685 * first_init_pfn value and exits there. If we never encounter the value we
1686 * return false indicating there are no valid ranges left.
1689 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1690 unsigned long *spfn, unsigned long *epfn,
1691 unsigned long first_init_pfn)
1696 * Start out by walking through the ranges in this zone that have
1697 * already been initialized. We don't need to do anything with them
1698 * so we just need to flush them out of the system.
1700 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1701 if (*epfn <= first_init_pfn)
1703 if (*spfn < first_init_pfn)
1704 *spfn = first_init_pfn;
1713 * Initialize and free pages. We do it in two loops: first we initialize
1714 * struct page, then free to buddy allocator, because while we are
1715 * freeing pages we can access pages that are ahead (computing buddy
1716 * page in __free_one_page()).
1718 * In order to try and keep some memory in the cache we have the loop
1719 * broken along max page order boundaries. This way we will not cause
1720 * any issues with the buddy page computation.
1722 static unsigned long __init
1723 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1724 unsigned long *end_pfn)
1726 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1727 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1728 unsigned long nr_pages = 0;
1731 /* First we loop through and initialize the page values */
1732 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1735 if (mo_pfn <= *start_pfn)
1738 t = min(mo_pfn, *end_pfn);
1739 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1741 if (mo_pfn < *end_pfn) {
1742 *start_pfn = mo_pfn;
1747 /* Reset values and now loop through freeing pages as needed */
1750 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1756 t = min(mo_pfn, epfn);
1757 deferred_free_pages(spfn, t);
1766 /* Initialise remaining memory on a node */
1767 static int __init deferred_init_memmap(void *data)
1769 pg_data_t *pgdat = data;
1770 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1771 unsigned long spfn = 0, epfn = 0, nr_pages = 0;
1772 unsigned long first_init_pfn, flags;
1773 unsigned long start = jiffies;
1778 /* Bind memory initialisation thread to a local node if possible */
1779 if (!cpumask_empty(cpumask))
1780 set_cpus_allowed_ptr(current, cpumask);
1782 pgdat_resize_lock(pgdat, &flags);
1783 first_init_pfn = pgdat->first_deferred_pfn;
1784 if (first_init_pfn == ULONG_MAX) {
1785 pgdat_resize_unlock(pgdat, &flags);
1786 pgdat_init_report_one_done();
1790 /* Sanity check boundaries */
1791 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1792 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1793 pgdat->first_deferred_pfn = ULONG_MAX;
1795 /* Only the highest zone is deferred so find it */
1796 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1797 zone = pgdat->node_zones + zid;
1798 if (first_init_pfn < zone_end_pfn(zone))
1802 /* If the zone is empty somebody else may have cleared out the zone */
1803 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1808 * Initialize and free pages in MAX_ORDER sized increments so
1809 * that we can avoid introducing any issues with the buddy
1813 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1815 pgdat_resize_unlock(pgdat, &flags);
1817 /* Sanity check that the next zone really is unpopulated */
1818 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1820 pr_info("node %d initialised, %lu pages in %ums\n",
1821 pgdat->node_id, nr_pages, jiffies_to_msecs(jiffies - start));
1823 pgdat_init_report_one_done();
1828 * If this zone has deferred pages, try to grow it by initializing enough
1829 * deferred pages to satisfy the allocation specified by order, rounded up to
1830 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1831 * of SECTION_SIZE bytes by initializing struct pages in increments of
1832 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1834 * Return true when zone was grown, otherwise return false. We return true even
1835 * when we grow less than requested, to let the caller decide if there are
1836 * enough pages to satisfy the allocation.
1838 * Note: We use noinline because this function is needed only during boot, and
1839 * it is called from a __ref function _deferred_grow_zone. This way we are
1840 * making sure that it is not inlined into permanent text section.
1842 static noinline bool __init
1843 deferred_grow_zone(struct zone *zone, unsigned int order)
1845 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1846 pg_data_t *pgdat = zone->zone_pgdat;
1847 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1848 unsigned long spfn, epfn, flags;
1849 unsigned long nr_pages = 0;
1852 /* Only the last zone may have deferred pages */
1853 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1856 pgdat_resize_lock(pgdat, &flags);
1859 * If deferred pages have been initialized while we were waiting for
1860 * the lock, return true, as the zone was grown. The caller will retry
1861 * this zone. We won't return to this function since the caller also
1862 * has this static branch.
1864 if (!static_branch_unlikely(&deferred_pages)) {
1865 pgdat_resize_unlock(pgdat, &flags);
1870 * If someone grew this zone while we were waiting for spinlock, return
1871 * true, as there might be enough pages already.
1873 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1874 pgdat_resize_unlock(pgdat, &flags);
1878 /* If the zone is empty somebody else may have cleared out the zone */
1879 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1880 first_deferred_pfn)) {
1881 pgdat->first_deferred_pfn = ULONG_MAX;
1882 pgdat_resize_unlock(pgdat, &flags);
1883 /* Retry only once. */
1884 return first_deferred_pfn != ULONG_MAX;
1888 * Initialize and free pages in MAX_ORDER sized increments so
1889 * that we can avoid introducing any issues with the buddy
1892 while (spfn < epfn) {
1893 /* update our first deferred PFN for this section */
1894 first_deferred_pfn = spfn;
1896 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1898 /* We should only stop along section boundaries */
1899 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
1902 /* If our quota has been met we can stop here */
1903 if (nr_pages >= nr_pages_needed)
1907 pgdat->first_deferred_pfn = spfn;
1908 pgdat_resize_unlock(pgdat, &flags);
1910 return nr_pages > 0;
1914 * deferred_grow_zone() is __init, but it is called from
1915 * get_page_from_freelist() during early boot until deferred_pages permanently
1916 * disables this call. This is why we have refdata wrapper to avoid warning,
1917 * and to ensure that the function body gets unloaded.
1920 _deferred_grow_zone(struct zone *zone, unsigned int order)
1922 return deferred_grow_zone(zone, order);
1925 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1927 void __init page_alloc_init_late(void)
1932 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1934 /* There will be num_node_state(N_MEMORY) threads */
1935 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1936 for_each_node_state(nid, N_MEMORY) {
1937 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1940 /* Block until all are initialised */
1941 wait_for_completion(&pgdat_init_all_done_comp);
1944 * We initialized the rest of the deferred pages. Permanently disable
1945 * on-demand struct page initialization.
1947 static_branch_disable(&deferred_pages);
1949 /* Reinit limits that are based on free pages after the kernel is up */
1950 files_maxfiles_init();
1953 /* Discard memblock private memory */
1956 for_each_node_state(nid, N_MEMORY)
1957 shuffle_free_memory(NODE_DATA(nid));
1959 for_each_populated_zone(zone)
1960 set_zone_contiguous(zone);
1962 #ifdef CONFIG_DEBUG_PAGEALLOC
1963 init_debug_guardpage();
1968 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1969 void __init init_cma_reserved_pageblock(struct page *page)
1971 unsigned i = pageblock_nr_pages;
1972 struct page *p = page;
1975 __ClearPageReserved(p);
1976 set_page_count(p, 0);
1979 set_pageblock_migratetype(page, MIGRATE_CMA);
1981 if (pageblock_order >= MAX_ORDER) {
1982 i = pageblock_nr_pages;
1985 set_page_refcounted(p);
1986 __free_pages(p, MAX_ORDER - 1);
1987 p += MAX_ORDER_NR_PAGES;
1988 } while (i -= MAX_ORDER_NR_PAGES);
1990 set_page_refcounted(page);
1991 __free_pages(page, pageblock_order);
1994 adjust_managed_page_count(page, pageblock_nr_pages);
1999 * The order of subdivision here is critical for the IO subsystem.
2000 * Please do not alter this order without good reasons and regression
2001 * testing. Specifically, as large blocks of memory are subdivided,
2002 * the order in which smaller blocks are delivered depends on the order
2003 * they're subdivided in this function. This is the primary factor
2004 * influencing the order in which pages are delivered to the IO
2005 * subsystem according to empirical testing, and this is also justified
2006 * by considering the behavior of a buddy system containing a single
2007 * large block of memory acted on by a series of small allocations.
2008 * This behavior is a critical factor in sglist merging's success.
2012 static inline void expand(struct zone *zone, struct page *page,
2013 int low, int high, struct free_area *area,
2016 unsigned long size = 1 << high;
2018 while (high > low) {
2022 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2025 * Mark as guard pages (or page), that will allow to
2026 * merge back to allocator when buddy will be freed.
2027 * Corresponding page table entries will not be touched,
2028 * pages will stay not present in virtual address space
2030 if (set_page_guard(zone, &page[size], high, migratetype))
2033 add_to_free_area(&page[size], area, migratetype);
2034 set_page_order(&page[size], high);
2038 static void check_new_page_bad(struct page *page)
2040 const char *bad_reason = NULL;
2041 unsigned long bad_flags = 0;
2043 if (unlikely(atomic_read(&page->_mapcount) != -1))
2044 bad_reason = "nonzero mapcount";
2045 if (unlikely(page->mapping != NULL))
2046 bad_reason = "non-NULL mapping";
2047 if (unlikely(page_ref_count(page) != 0))
2048 bad_reason = "nonzero _refcount";
2049 if (unlikely(page->flags & __PG_HWPOISON)) {
2050 bad_reason = "HWPoisoned (hardware-corrupted)";
2051 bad_flags = __PG_HWPOISON;
2052 /* Don't complain about hwpoisoned pages */
2053 page_mapcount_reset(page); /* remove PageBuddy */
2056 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
2057 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
2058 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
2061 if (unlikely(page->mem_cgroup))
2062 bad_reason = "page still charged to cgroup";
2064 bad_page(page, bad_reason, bad_flags);
2068 * This page is about to be returned from the page allocator
2070 static inline int check_new_page(struct page *page)
2072 if (likely(page_expected_state(page,
2073 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2076 check_new_page_bad(page);
2080 static inline bool free_pages_prezeroed(void)
2082 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2083 page_poisoning_enabled()) || want_init_on_free();
2086 #ifdef CONFIG_DEBUG_VM
2088 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2089 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2090 * also checked when pcp lists are refilled from the free lists.
2092 static inline bool check_pcp_refill(struct page *page)
2094 if (debug_pagealloc_enabled())
2095 return check_new_page(page);
2100 static inline bool check_new_pcp(struct page *page)
2102 return check_new_page(page);
2106 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2107 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2108 * enabled, they are also checked when being allocated from the pcp lists.
2110 static inline bool check_pcp_refill(struct page *page)
2112 return check_new_page(page);
2114 static inline bool check_new_pcp(struct page *page)
2116 if (debug_pagealloc_enabled())
2117 return check_new_page(page);
2121 #endif /* CONFIG_DEBUG_VM */
2123 static bool check_new_pages(struct page *page, unsigned int order)
2126 for (i = 0; i < (1 << order); i++) {
2127 struct page *p = page + i;
2129 if (unlikely(check_new_page(p)))
2136 inline void post_alloc_hook(struct page *page, unsigned int order,
2139 set_page_private(page, 0);
2140 set_page_refcounted(page);
2142 arch_alloc_page(page, order);
2143 if (debug_pagealloc_enabled())
2144 kernel_map_pages(page, 1 << order, 1);
2145 kasan_alloc_pages(page, order);
2146 kernel_poison_pages(page, 1 << order, 1);
2147 set_page_owner(page, order, gfp_flags);
2150 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2151 unsigned int alloc_flags)
2153 post_alloc_hook(page, order, gfp_flags);
2155 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2156 kernel_init_free_pages(page, 1 << order);
2158 if (order && (gfp_flags & __GFP_COMP))
2159 prep_compound_page(page, order);
2162 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2163 * allocate the page. The expectation is that the caller is taking
2164 * steps that will free more memory. The caller should avoid the page
2165 * being used for !PFMEMALLOC purposes.
2167 if (alloc_flags & ALLOC_NO_WATERMARKS)
2168 set_page_pfmemalloc(page);
2170 clear_page_pfmemalloc(page);
2174 * Go through the free lists for the given migratetype and remove
2175 * the smallest available page from the freelists
2177 static __always_inline
2178 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2181 unsigned int current_order;
2182 struct free_area *area;
2185 /* Find a page of the appropriate size in the preferred list */
2186 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2187 area = &(zone->free_area[current_order]);
2188 page = get_page_from_free_area(area, migratetype);
2191 del_page_from_free_area(page, area);
2192 expand(zone, page, order, current_order, area, migratetype);
2193 set_pcppage_migratetype(page, migratetype);
2202 * This array describes the order lists are fallen back to when
2203 * the free lists for the desirable migrate type are depleted
2205 static int fallbacks[MIGRATE_TYPES][4] = {
2206 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2207 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2208 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2210 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2212 #ifdef CONFIG_MEMORY_ISOLATION
2213 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2218 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2221 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2224 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2225 unsigned int order) { return NULL; }
2229 * Move the free pages in a range to the free lists of the requested type.
2230 * Note that start_page and end_pages are not aligned on a pageblock
2231 * boundary. If alignment is required, use move_freepages_block()
2233 static int move_freepages(struct zone *zone,
2234 struct page *start_page, struct page *end_page,
2235 int migratetype, int *num_movable)
2239 int pages_moved = 0;
2241 for (page = start_page; page <= end_page;) {
2242 if (!pfn_valid_within(page_to_pfn(page))) {
2247 if (!PageBuddy(page)) {
2249 * We assume that pages that could be isolated for
2250 * migration are movable. But we don't actually try
2251 * isolating, as that would be expensive.
2254 (PageLRU(page) || __PageMovable(page)))
2261 /* Make sure we are not inadvertently changing nodes */
2262 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2263 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2265 order = page_order(page);
2266 move_to_free_area(page, &zone->free_area[order], migratetype);
2268 pages_moved += 1 << order;
2274 int move_freepages_block(struct zone *zone, struct page *page,
2275 int migratetype, int *num_movable)
2277 unsigned long start_pfn, end_pfn;
2278 struct page *start_page, *end_page;
2283 start_pfn = page_to_pfn(page);
2284 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2285 start_page = pfn_to_page(start_pfn);
2286 end_page = start_page + pageblock_nr_pages - 1;
2287 end_pfn = start_pfn + pageblock_nr_pages - 1;
2289 /* Do not cross zone boundaries */
2290 if (!zone_spans_pfn(zone, start_pfn))
2292 if (!zone_spans_pfn(zone, end_pfn))
2295 return move_freepages(zone, start_page, end_page, migratetype,
2299 static void change_pageblock_range(struct page *pageblock_page,
2300 int start_order, int migratetype)
2302 int nr_pageblocks = 1 << (start_order - pageblock_order);
2304 while (nr_pageblocks--) {
2305 set_pageblock_migratetype(pageblock_page, migratetype);
2306 pageblock_page += pageblock_nr_pages;
2311 * When we are falling back to another migratetype during allocation, try to
2312 * steal extra free pages from the same pageblocks to satisfy further
2313 * allocations, instead of polluting multiple pageblocks.
2315 * If we are stealing a relatively large buddy page, it is likely there will
2316 * be more free pages in the pageblock, so try to steal them all. For
2317 * reclaimable and unmovable allocations, we steal regardless of page size,
2318 * as fragmentation caused by those allocations polluting movable pageblocks
2319 * is worse than movable allocations stealing from unmovable and reclaimable
2322 static bool can_steal_fallback(unsigned int order, int start_mt)
2325 * Leaving this order check is intended, although there is
2326 * relaxed order check in next check. The reason is that
2327 * we can actually steal whole pageblock if this condition met,
2328 * but, below check doesn't guarantee it and that is just heuristic
2329 * so could be changed anytime.
2331 if (order >= pageblock_order)
2334 if (order >= pageblock_order / 2 ||
2335 start_mt == MIGRATE_RECLAIMABLE ||
2336 start_mt == MIGRATE_UNMOVABLE ||
2337 page_group_by_mobility_disabled)
2343 static inline void boost_watermark(struct zone *zone)
2345 unsigned long max_boost;
2347 if (!watermark_boost_factor)
2350 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2351 watermark_boost_factor, 10000);
2354 * high watermark may be uninitialised if fragmentation occurs
2355 * very early in boot so do not boost. We do not fall
2356 * through and boost by pageblock_nr_pages as failing
2357 * allocations that early means that reclaim is not going
2358 * to help and it may even be impossible to reclaim the
2359 * boosted watermark resulting in a hang.
2364 max_boost = max(pageblock_nr_pages, max_boost);
2366 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2371 * This function implements actual steal behaviour. If order is large enough,
2372 * we can steal whole pageblock. If not, we first move freepages in this
2373 * pageblock to our migratetype and determine how many already-allocated pages
2374 * are there in the pageblock with a compatible migratetype. If at least half
2375 * of pages are free or compatible, we can change migratetype of the pageblock
2376 * itself, so pages freed in the future will be put on the correct free list.
2378 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2379 unsigned int alloc_flags, int start_type, bool whole_block)
2381 unsigned int current_order = page_order(page);
2382 struct free_area *area;
2383 int free_pages, movable_pages, alike_pages;
2386 old_block_type = get_pageblock_migratetype(page);
2389 * This can happen due to races and we want to prevent broken
2390 * highatomic accounting.
2392 if (is_migrate_highatomic(old_block_type))
2395 /* Take ownership for orders >= pageblock_order */
2396 if (current_order >= pageblock_order) {
2397 change_pageblock_range(page, current_order, start_type);
2402 * Boost watermarks to increase reclaim pressure to reduce the
2403 * likelihood of future fallbacks. Wake kswapd now as the node
2404 * may be balanced overall and kswapd will not wake naturally.
2406 boost_watermark(zone);
2407 if (alloc_flags & ALLOC_KSWAPD)
2408 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2410 /* We are not allowed to try stealing from the whole block */
2414 free_pages = move_freepages_block(zone, page, start_type,
2417 * Determine how many pages are compatible with our allocation.
2418 * For movable allocation, it's the number of movable pages which
2419 * we just obtained. For other types it's a bit more tricky.
2421 if (start_type == MIGRATE_MOVABLE) {
2422 alike_pages = movable_pages;
2425 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2426 * to MOVABLE pageblock, consider all non-movable pages as
2427 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2428 * vice versa, be conservative since we can't distinguish the
2429 * exact migratetype of non-movable pages.
2431 if (old_block_type == MIGRATE_MOVABLE)
2432 alike_pages = pageblock_nr_pages
2433 - (free_pages + movable_pages);
2438 /* moving whole block can fail due to zone boundary conditions */
2443 * If a sufficient number of pages in the block are either free or of
2444 * comparable migratability as our allocation, claim the whole block.
2446 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2447 page_group_by_mobility_disabled)
2448 set_pageblock_migratetype(page, start_type);
2453 area = &zone->free_area[current_order];
2454 move_to_free_area(page, area, start_type);
2458 * Check whether there is a suitable fallback freepage with requested order.
2459 * If only_stealable is true, this function returns fallback_mt only if
2460 * we can steal other freepages all together. This would help to reduce
2461 * fragmentation due to mixed migratetype pages in one pageblock.
2463 int find_suitable_fallback(struct free_area *area, unsigned int order,
2464 int migratetype, bool only_stealable, bool *can_steal)
2469 if (area->nr_free == 0)
2474 fallback_mt = fallbacks[migratetype][i];
2475 if (fallback_mt == MIGRATE_TYPES)
2478 if (free_area_empty(area, fallback_mt))
2481 if (can_steal_fallback(order, migratetype))
2484 if (!only_stealable)
2495 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2496 * there are no empty page blocks that contain a page with a suitable order
2498 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2499 unsigned int alloc_order)
2502 unsigned long max_managed, flags;
2505 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2506 * Check is race-prone but harmless.
2508 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2509 if (zone->nr_reserved_highatomic >= max_managed)
2512 spin_lock_irqsave(&zone->lock, flags);
2514 /* Recheck the nr_reserved_highatomic limit under the lock */
2515 if (zone->nr_reserved_highatomic >= max_managed)
2519 mt = get_pageblock_migratetype(page);
2520 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2521 && !is_migrate_cma(mt)) {
2522 zone->nr_reserved_highatomic += pageblock_nr_pages;
2523 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2524 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2528 spin_unlock_irqrestore(&zone->lock, flags);
2532 * Used when an allocation is about to fail under memory pressure. This
2533 * potentially hurts the reliability of high-order allocations when under
2534 * intense memory pressure but failed atomic allocations should be easier
2535 * to recover from than an OOM.
2537 * If @force is true, try to unreserve a pageblock even though highatomic
2538 * pageblock is exhausted.
2540 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2543 struct zonelist *zonelist = ac->zonelist;
2544 unsigned long flags;
2551 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2554 * Preserve at least one pageblock unless memory pressure
2557 if (!force && zone->nr_reserved_highatomic <=
2561 spin_lock_irqsave(&zone->lock, flags);
2562 for (order = 0; order < MAX_ORDER; order++) {
2563 struct free_area *area = &(zone->free_area[order]);
2565 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2570 * In page freeing path, migratetype change is racy so
2571 * we can counter several free pages in a pageblock
2572 * in this loop althoug we changed the pageblock type
2573 * from highatomic to ac->migratetype. So we should
2574 * adjust the count once.
2576 if (is_migrate_highatomic_page(page)) {
2578 * It should never happen but changes to
2579 * locking could inadvertently allow a per-cpu
2580 * drain to add pages to MIGRATE_HIGHATOMIC
2581 * while unreserving so be safe and watch for
2584 zone->nr_reserved_highatomic -= min(
2586 zone->nr_reserved_highatomic);
2590 * Convert to ac->migratetype and avoid the normal
2591 * pageblock stealing heuristics. Minimally, the caller
2592 * is doing the work and needs the pages. More
2593 * importantly, if the block was always converted to
2594 * MIGRATE_UNMOVABLE or another type then the number
2595 * of pageblocks that cannot be completely freed
2598 set_pageblock_migratetype(page, ac->migratetype);
2599 ret = move_freepages_block(zone, page, ac->migratetype,
2602 spin_unlock_irqrestore(&zone->lock, flags);
2606 spin_unlock_irqrestore(&zone->lock, flags);
2613 * Try finding a free buddy page on the fallback list and put it on the free
2614 * list of requested migratetype, possibly along with other pages from the same
2615 * block, depending on fragmentation avoidance heuristics. Returns true if
2616 * fallback was found so that __rmqueue_smallest() can grab it.
2618 * The use of signed ints for order and current_order is a deliberate
2619 * deviation from the rest of this file, to make the for loop
2620 * condition simpler.
2622 static __always_inline bool
2623 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2624 unsigned int alloc_flags)
2626 struct free_area *area;
2628 int min_order = order;
2634 * Do not steal pages from freelists belonging to other pageblocks
2635 * i.e. orders < pageblock_order. If there are no local zones free,
2636 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2638 if (alloc_flags & ALLOC_NOFRAGMENT)
2639 min_order = pageblock_order;
2642 * Find the largest available free page in the other list. This roughly
2643 * approximates finding the pageblock with the most free pages, which
2644 * would be too costly to do exactly.
2646 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2648 area = &(zone->free_area[current_order]);
2649 fallback_mt = find_suitable_fallback(area, current_order,
2650 start_migratetype, false, &can_steal);
2651 if (fallback_mt == -1)
2655 * We cannot steal all free pages from the pageblock and the
2656 * requested migratetype is movable. In that case it's better to
2657 * steal and split the smallest available page instead of the
2658 * largest available page, because even if the next movable
2659 * allocation falls back into a different pageblock than this
2660 * one, it won't cause permanent fragmentation.
2662 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2663 && current_order > order)
2672 for (current_order = order; current_order < MAX_ORDER;
2674 area = &(zone->free_area[current_order]);
2675 fallback_mt = find_suitable_fallback(area, current_order,
2676 start_migratetype, false, &can_steal);
2677 if (fallback_mt != -1)
2682 * This should not happen - we already found a suitable fallback
2683 * when looking for the largest page.
2685 VM_BUG_ON(current_order == MAX_ORDER);
2688 page = get_page_from_free_area(area, fallback_mt);
2690 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2693 trace_mm_page_alloc_extfrag(page, order, current_order,
2694 start_migratetype, fallback_mt);
2701 * Do the hard work of removing an element from the buddy allocator.
2702 * Call me with the zone->lock already held.
2704 static __always_inline struct page *
2705 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2706 unsigned int alloc_flags)
2711 page = __rmqueue_smallest(zone, order, migratetype);
2712 if (unlikely(!page)) {
2713 if (migratetype == MIGRATE_MOVABLE)
2714 page = __rmqueue_cma_fallback(zone, order);
2716 if (!page && __rmqueue_fallback(zone, order, migratetype,
2721 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2726 * Obtain a specified number of elements from the buddy allocator, all under
2727 * a single hold of the lock, for efficiency. Add them to the supplied list.
2728 * Returns the number of new pages which were placed at *list.
2730 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2731 unsigned long count, struct list_head *list,
2732 int migratetype, unsigned int alloc_flags)
2736 spin_lock(&zone->lock);
2737 for (i = 0; i < count; ++i) {
2738 struct page *page = __rmqueue(zone, order, migratetype,
2740 if (unlikely(page == NULL))
2743 if (unlikely(check_pcp_refill(page)))
2747 * Split buddy pages returned by expand() are received here in
2748 * physical page order. The page is added to the tail of
2749 * caller's list. From the callers perspective, the linked list
2750 * is ordered by page number under some conditions. This is
2751 * useful for IO devices that can forward direction from the
2752 * head, thus also in the physical page order. This is useful
2753 * for IO devices that can merge IO requests if the physical
2754 * pages are ordered properly.
2756 list_add_tail(&page->lru, list);
2758 if (is_migrate_cma(get_pcppage_migratetype(page)))
2759 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2764 * i pages were removed from the buddy list even if some leak due
2765 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2766 * on i. Do not confuse with 'alloced' which is the number of
2767 * pages added to the pcp list.
2769 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2770 spin_unlock(&zone->lock);
2776 * Called from the vmstat counter updater to drain pagesets of this
2777 * currently executing processor on remote nodes after they have
2780 * Note that this function must be called with the thread pinned to
2781 * a single processor.
2783 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2785 unsigned long flags;
2786 int to_drain, batch;
2788 local_irq_save(flags);
2789 batch = READ_ONCE(pcp->batch);
2790 to_drain = min(pcp->count, batch);
2792 free_pcppages_bulk(zone, to_drain, pcp);
2793 local_irq_restore(flags);
2798 * Drain pcplists of the indicated processor and zone.
2800 * The processor must either be the current processor and the
2801 * thread pinned to the current processor or a processor that
2804 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2806 unsigned long flags;
2807 struct per_cpu_pageset *pset;
2808 struct per_cpu_pages *pcp;
2810 local_irq_save(flags);
2811 pset = per_cpu_ptr(zone->pageset, cpu);
2815 free_pcppages_bulk(zone, pcp->count, pcp);
2816 local_irq_restore(flags);
2820 * Drain pcplists of all zones on the indicated processor.
2822 * The processor must either be the current processor and the
2823 * thread pinned to the current processor or a processor that
2826 static void drain_pages(unsigned int cpu)
2830 for_each_populated_zone(zone) {
2831 drain_pages_zone(cpu, zone);
2836 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2838 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2839 * the single zone's pages.
2841 void drain_local_pages(struct zone *zone)
2843 int cpu = smp_processor_id();
2846 drain_pages_zone(cpu, zone);
2851 static void drain_local_pages_wq(struct work_struct *work)
2853 struct pcpu_drain *drain;
2855 drain = container_of(work, struct pcpu_drain, work);
2858 * drain_all_pages doesn't use proper cpu hotplug protection so
2859 * we can race with cpu offline when the WQ can move this from
2860 * a cpu pinned worker to an unbound one. We can operate on a different
2861 * cpu which is allright but we also have to make sure to not move to
2865 drain_local_pages(drain->zone);
2870 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2872 * When zone parameter is non-NULL, spill just the single zone's pages.
2874 * Note that this can be extremely slow as the draining happens in a workqueue.
2876 void drain_all_pages(struct zone *zone)
2881 * Allocate in the BSS so we wont require allocation in
2882 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2884 static cpumask_t cpus_with_pcps;
2887 * Make sure nobody triggers this path before mm_percpu_wq is fully
2890 if (WARN_ON_ONCE(!mm_percpu_wq))
2894 * Do not drain if one is already in progress unless it's specific to
2895 * a zone. Such callers are primarily CMA and memory hotplug and need
2896 * the drain to be complete when the call returns.
2898 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2901 mutex_lock(&pcpu_drain_mutex);
2905 * We don't care about racing with CPU hotplug event
2906 * as offline notification will cause the notified
2907 * cpu to drain that CPU pcps and on_each_cpu_mask
2908 * disables preemption as part of its processing
2910 for_each_online_cpu(cpu) {
2911 struct per_cpu_pageset *pcp;
2913 bool has_pcps = false;
2916 pcp = per_cpu_ptr(zone->pageset, cpu);
2920 for_each_populated_zone(z) {
2921 pcp = per_cpu_ptr(z->pageset, cpu);
2922 if (pcp->pcp.count) {
2930 cpumask_set_cpu(cpu, &cpus_with_pcps);
2932 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2935 for_each_cpu(cpu, &cpus_with_pcps) {
2936 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
2939 INIT_WORK(&drain->work, drain_local_pages_wq);
2940 queue_work_on(cpu, mm_percpu_wq, &drain->work);
2942 for_each_cpu(cpu, &cpus_with_pcps)
2943 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
2945 mutex_unlock(&pcpu_drain_mutex);
2948 #ifdef CONFIG_HIBERNATION
2951 * Touch the watchdog for every WD_PAGE_COUNT pages.
2953 #define WD_PAGE_COUNT (128*1024)
2955 void mark_free_pages(struct zone *zone)
2957 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2958 unsigned long flags;
2959 unsigned int order, t;
2962 if (zone_is_empty(zone))
2965 spin_lock_irqsave(&zone->lock, flags);
2967 max_zone_pfn = zone_end_pfn(zone);
2968 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2969 if (pfn_valid(pfn)) {
2970 page = pfn_to_page(pfn);
2972 if (!--page_count) {
2973 touch_nmi_watchdog();
2974 page_count = WD_PAGE_COUNT;
2977 if (page_zone(page) != zone)
2980 if (!swsusp_page_is_forbidden(page))
2981 swsusp_unset_page_free(page);
2984 for_each_migratetype_order(order, t) {
2985 list_for_each_entry(page,
2986 &zone->free_area[order].free_list[t], lru) {
2989 pfn = page_to_pfn(page);
2990 for (i = 0; i < (1UL << order); i++) {
2991 if (!--page_count) {
2992 touch_nmi_watchdog();
2993 page_count = WD_PAGE_COUNT;
2995 swsusp_set_page_free(pfn_to_page(pfn + i));
2999 spin_unlock_irqrestore(&zone->lock, flags);
3001 #endif /* CONFIG_PM */
3003 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3007 if (!free_pcp_prepare(page))
3010 migratetype = get_pfnblock_migratetype(page, pfn);
3011 set_pcppage_migratetype(page, migratetype);
3015 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3017 struct zone *zone = page_zone(page);
3018 struct per_cpu_pages *pcp;
3021 migratetype = get_pcppage_migratetype(page);
3022 __count_vm_event(PGFREE);
3025 * We only track unmovable, reclaimable and movable on pcp lists.
3026 * Free ISOLATE pages back to the allocator because they are being
3027 * offlined but treat HIGHATOMIC as movable pages so we can get those
3028 * areas back if necessary. Otherwise, we may have to free
3029 * excessively into the page allocator
3031 if (migratetype >= MIGRATE_PCPTYPES) {
3032 if (unlikely(is_migrate_isolate(migratetype))) {
3033 free_one_page(zone, page, pfn, 0, migratetype);
3036 migratetype = MIGRATE_MOVABLE;
3039 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3040 list_add(&page->lru, &pcp->lists[migratetype]);
3042 if (pcp->count >= pcp->high) {
3043 unsigned long batch = READ_ONCE(pcp->batch);
3044 free_pcppages_bulk(zone, batch, pcp);
3049 * Free a 0-order page
3051 void free_unref_page(struct page *page)
3053 unsigned long flags;
3054 unsigned long pfn = page_to_pfn(page);
3056 if (!free_unref_page_prepare(page, pfn))
3059 local_irq_save(flags);
3060 free_unref_page_commit(page, pfn);
3061 local_irq_restore(flags);
3065 * Free a list of 0-order pages
3067 void free_unref_page_list(struct list_head *list)
3069 struct page *page, *next;
3070 unsigned long flags, pfn;
3071 int batch_count = 0;
3073 /* Prepare pages for freeing */
3074 list_for_each_entry_safe(page, next, list, lru) {
3075 pfn = page_to_pfn(page);
3076 if (!free_unref_page_prepare(page, pfn))
3077 list_del(&page->lru);
3078 set_page_private(page, pfn);
3081 local_irq_save(flags);
3082 list_for_each_entry_safe(page, next, list, lru) {
3083 unsigned long pfn = page_private(page);
3085 set_page_private(page, 0);
3086 trace_mm_page_free_batched(page);
3087 free_unref_page_commit(page, pfn);
3090 * Guard against excessive IRQ disabled times when we get
3091 * a large list of pages to free.
3093 if (++batch_count == SWAP_CLUSTER_MAX) {
3094 local_irq_restore(flags);
3096 local_irq_save(flags);
3099 local_irq_restore(flags);
3103 * split_page takes a non-compound higher-order page, and splits it into
3104 * n (1<<order) sub-pages: page[0..n]
3105 * Each sub-page must be freed individually.
3107 * Note: this is probably too low level an operation for use in drivers.
3108 * Please consult with lkml before using this in your driver.
3110 void split_page(struct page *page, unsigned int order)
3114 VM_BUG_ON_PAGE(PageCompound(page), page);
3115 VM_BUG_ON_PAGE(!page_count(page), page);
3117 for (i = 1; i < (1 << order); i++)
3118 set_page_refcounted(page + i);
3119 split_page_owner(page, order);
3121 EXPORT_SYMBOL_GPL(split_page);
3123 int __isolate_free_page(struct page *page, unsigned int order)
3125 struct free_area *area = &page_zone(page)->free_area[order];
3126 unsigned long watermark;
3130 BUG_ON(!PageBuddy(page));
3132 zone = page_zone(page);
3133 mt = get_pageblock_migratetype(page);
3135 if (!is_migrate_isolate(mt)) {
3137 * Obey watermarks as if the page was being allocated. We can
3138 * emulate a high-order watermark check with a raised order-0
3139 * watermark, because we already know our high-order page
3142 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3143 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3146 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3149 /* Remove page from free list */
3151 del_page_from_free_area(page, area);
3154 * Set the pageblock if the isolated page is at least half of a
3157 if (order >= pageblock_order - 1) {
3158 struct page *endpage = page + (1 << order) - 1;
3159 for (; page < endpage; page += pageblock_nr_pages) {
3160 int mt = get_pageblock_migratetype(page);
3161 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3162 && !is_migrate_highatomic(mt))
3163 set_pageblock_migratetype(page,
3169 return 1UL << order;
3173 * Update NUMA hit/miss statistics
3175 * Must be called with interrupts disabled.
3177 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3180 enum numa_stat_item local_stat = NUMA_LOCAL;
3182 /* skip numa counters update if numa stats is disabled */
3183 if (!static_branch_likely(&vm_numa_stat_key))
3186 if (zone_to_nid(z) != numa_node_id())
3187 local_stat = NUMA_OTHER;
3189 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3190 __inc_numa_state(z, NUMA_HIT);
3192 __inc_numa_state(z, NUMA_MISS);
3193 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3195 __inc_numa_state(z, local_stat);
3199 /* Remove page from the per-cpu list, caller must protect the list */
3200 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3201 unsigned int alloc_flags,
3202 struct per_cpu_pages *pcp,
3203 struct list_head *list)
3208 if (list_empty(list)) {
3209 pcp->count += rmqueue_bulk(zone, 0,
3211 migratetype, alloc_flags);
3212 if (unlikely(list_empty(list)))
3216 page = list_first_entry(list, struct page, lru);
3217 list_del(&page->lru);
3219 } while (check_new_pcp(page));
3224 /* Lock and remove page from the per-cpu list */
3225 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3226 struct zone *zone, gfp_t gfp_flags,
3227 int migratetype, unsigned int alloc_flags)
3229 struct per_cpu_pages *pcp;
3230 struct list_head *list;
3232 unsigned long flags;
3234 local_irq_save(flags);
3235 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3236 list = &pcp->lists[migratetype];
3237 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3239 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3240 zone_statistics(preferred_zone, zone);
3242 local_irq_restore(flags);
3247 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3250 struct page *rmqueue(struct zone *preferred_zone,
3251 struct zone *zone, unsigned int order,
3252 gfp_t gfp_flags, unsigned int alloc_flags,
3255 unsigned long flags;
3258 if (likely(order == 0)) {
3259 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3260 migratetype, alloc_flags);
3265 * We most definitely don't want callers attempting to
3266 * allocate greater than order-1 page units with __GFP_NOFAIL.
3268 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3269 spin_lock_irqsave(&zone->lock, flags);
3273 if (alloc_flags & ALLOC_HARDER) {
3274 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3276 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3279 page = __rmqueue(zone, order, migratetype, alloc_flags);
3280 } while (page && check_new_pages(page, order));
3281 spin_unlock(&zone->lock);
3284 __mod_zone_freepage_state(zone, -(1 << order),
3285 get_pcppage_migratetype(page));
3287 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3288 zone_statistics(preferred_zone, zone);
3289 local_irq_restore(flags);
3292 /* Separate test+clear to avoid unnecessary atomics */
3293 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3294 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3295 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3298 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3302 local_irq_restore(flags);
3306 #ifdef CONFIG_FAIL_PAGE_ALLOC
3309 struct fault_attr attr;
3311 bool ignore_gfp_highmem;
3312 bool ignore_gfp_reclaim;
3314 } fail_page_alloc = {
3315 .attr = FAULT_ATTR_INITIALIZER,
3316 .ignore_gfp_reclaim = true,
3317 .ignore_gfp_highmem = true,
3321 static int __init setup_fail_page_alloc(char *str)
3323 return setup_fault_attr(&fail_page_alloc.attr, str);
3325 __setup("fail_page_alloc=", setup_fail_page_alloc);
3327 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3329 if (order < fail_page_alloc.min_order)
3331 if (gfp_mask & __GFP_NOFAIL)
3333 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3335 if (fail_page_alloc.ignore_gfp_reclaim &&
3336 (gfp_mask & __GFP_DIRECT_RECLAIM))
3339 return should_fail(&fail_page_alloc.attr, 1 << order);
3342 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3344 static int __init fail_page_alloc_debugfs(void)
3346 umode_t mode = S_IFREG | 0600;
3349 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3350 &fail_page_alloc.attr);
3352 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3353 &fail_page_alloc.ignore_gfp_reclaim);
3354 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3355 &fail_page_alloc.ignore_gfp_highmem);
3356 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3361 late_initcall(fail_page_alloc_debugfs);
3363 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3365 #else /* CONFIG_FAIL_PAGE_ALLOC */
3367 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3372 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3374 static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3376 return __should_fail_alloc_page(gfp_mask, order);
3378 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3381 * Return true if free base pages are above 'mark'. For high-order checks it
3382 * will return true of the order-0 watermark is reached and there is at least
3383 * one free page of a suitable size. Checking now avoids taking the zone lock
3384 * to check in the allocation paths if no pages are free.
3386 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3387 int classzone_idx, unsigned int alloc_flags,
3392 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3394 /* free_pages may go negative - that's OK */
3395 free_pages -= (1 << order) - 1;
3397 if (alloc_flags & ALLOC_HIGH)
3401 * If the caller does not have rights to ALLOC_HARDER then subtract
3402 * the high-atomic reserves. This will over-estimate the size of the
3403 * atomic reserve but it avoids a search.
3405 if (likely(!alloc_harder)) {
3406 free_pages -= z->nr_reserved_highatomic;
3409 * OOM victims can try even harder than normal ALLOC_HARDER
3410 * users on the grounds that it's definitely going to be in
3411 * the exit path shortly and free memory. Any allocation it
3412 * makes during the free path will be small and short-lived.
3414 if (alloc_flags & ALLOC_OOM)
3422 /* If allocation can't use CMA areas don't use free CMA pages */
3423 if (!(alloc_flags & ALLOC_CMA))
3424 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3428 * Check watermarks for an order-0 allocation request. If these
3429 * are not met, then a high-order request also cannot go ahead
3430 * even if a suitable page happened to be free.
3432 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3435 /* If this is an order-0 request then the watermark is fine */
3439 /* For a high-order request, check at least one suitable page is free */
3440 for (o = order; o < MAX_ORDER; o++) {
3441 struct free_area *area = &z->free_area[o];
3447 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3448 if (!free_area_empty(area, mt))
3453 if ((alloc_flags & ALLOC_CMA) &&
3454 !free_area_empty(area, MIGRATE_CMA)) {
3459 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3465 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3466 int classzone_idx, unsigned int alloc_flags)
3468 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3469 zone_page_state(z, NR_FREE_PAGES));
3472 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3473 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3475 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3479 /* If allocation can't use CMA areas don't use free CMA pages */
3480 if (!(alloc_flags & ALLOC_CMA))
3481 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3485 * Fast check for order-0 only. If this fails then the reserves
3486 * need to be calculated. There is a corner case where the check
3487 * passes but only the high-order atomic reserve are free. If
3488 * the caller is !atomic then it'll uselessly search the free
3489 * list. That corner case is then slower but it is harmless.
3491 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3494 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3498 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3499 unsigned long mark, int classzone_idx)
3501 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3503 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3504 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3506 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3511 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3513 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3514 node_reclaim_distance;
3516 #else /* CONFIG_NUMA */
3517 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3521 #endif /* CONFIG_NUMA */
3524 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3525 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3526 * premature use of a lower zone may cause lowmem pressure problems that
3527 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3528 * probably too small. It only makes sense to spread allocations to avoid
3529 * fragmentation between the Normal and DMA32 zones.
3531 static inline unsigned int
3532 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3534 unsigned int alloc_flags = 0;
3536 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3537 alloc_flags |= ALLOC_KSWAPD;
3539 #ifdef CONFIG_ZONE_DMA32
3543 if (zone_idx(zone) != ZONE_NORMAL)
3547 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3548 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3549 * on UMA that if Normal is populated then so is DMA32.
3551 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3552 if (nr_online_nodes > 1 && !populated_zone(--zone))
3555 alloc_flags |= ALLOC_NOFRAGMENT;
3556 #endif /* CONFIG_ZONE_DMA32 */
3561 * get_page_from_freelist goes through the zonelist trying to allocate
3564 static struct page *
3565 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3566 const struct alloc_context *ac)
3570 struct pglist_data *last_pgdat_dirty_limit = NULL;
3575 * Scan zonelist, looking for a zone with enough free.
3576 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3578 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3579 z = ac->preferred_zoneref;
3580 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3585 if (cpusets_enabled() &&
3586 (alloc_flags & ALLOC_CPUSET) &&
3587 !__cpuset_zone_allowed(zone, gfp_mask))
3590 * When allocating a page cache page for writing, we
3591 * want to get it from a node that is within its dirty
3592 * limit, such that no single node holds more than its
3593 * proportional share of globally allowed dirty pages.
3594 * The dirty limits take into account the node's
3595 * lowmem reserves and high watermark so that kswapd
3596 * should be able to balance it without having to
3597 * write pages from its LRU list.
3599 * XXX: For now, allow allocations to potentially
3600 * exceed the per-node dirty limit in the slowpath
3601 * (spread_dirty_pages unset) before going into reclaim,
3602 * which is important when on a NUMA setup the allowed
3603 * nodes are together not big enough to reach the
3604 * global limit. The proper fix for these situations
3605 * will require awareness of nodes in the
3606 * dirty-throttling and the flusher threads.
3608 if (ac->spread_dirty_pages) {
3609 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3612 if (!node_dirty_ok(zone->zone_pgdat)) {
3613 last_pgdat_dirty_limit = zone->zone_pgdat;
3618 if (no_fallback && nr_online_nodes > 1 &&
3619 zone != ac->preferred_zoneref->zone) {
3623 * If moving to a remote node, retry but allow
3624 * fragmenting fallbacks. Locality is more important
3625 * than fragmentation avoidance.
3627 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3628 if (zone_to_nid(zone) != local_nid) {
3629 alloc_flags &= ~ALLOC_NOFRAGMENT;
3634 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3635 if (!zone_watermark_fast(zone, order, mark,
3636 ac_classzone_idx(ac), alloc_flags)) {
3639 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3641 * Watermark failed for this zone, but see if we can
3642 * grow this zone if it contains deferred pages.
3644 if (static_branch_unlikely(&deferred_pages)) {
3645 if (_deferred_grow_zone(zone, order))
3649 /* Checked here to keep the fast path fast */
3650 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3651 if (alloc_flags & ALLOC_NO_WATERMARKS)
3654 if (node_reclaim_mode == 0 ||
3655 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3658 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3660 case NODE_RECLAIM_NOSCAN:
3663 case NODE_RECLAIM_FULL:
3664 /* scanned but unreclaimable */
3667 /* did we reclaim enough */
3668 if (zone_watermark_ok(zone, order, mark,
3669 ac_classzone_idx(ac), alloc_flags))
3677 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3678 gfp_mask, alloc_flags, ac->migratetype);
3680 prep_new_page(page, order, gfp_mask, alloc_flags);
3683 * If this is a high-order atomic allocation then check
3684 * if the pageblock should be reserved for the future
3686 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3687 reserve_highatomic_pageblock(page, zone, order);
3691 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3692 /* Try again if zone has deferred pages */
3693 if (static_branch_unlikely(&deferred_pages)) {
3694 if (_deferred_grow_zone(zone, order))
3702 * It's possible on a UMA machine to get through all zones that are
3703 * fragmented. If avoiding fragmentation, reset and try again.
3706 alloc_flags &= ~ALLOC_NOFRAGMENT;
3713 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3715 unsigned int filter = SHOW_MEM_FILTER_NODES;
3716 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3718 if (!__ratelimit(&show_mem_rs))
3722 * This documents exceptions given to allocations in certain
3723 * contexts that are allowed to allocate outside current's set
3726 if (!(gfp_mask & __GFP_NOMEMALLOC))
3727 if (tsk_is_oom_victim(current) ||
3728 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3729 filter &= ~SHOW_MEM_FILTER_NODES;
3730 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3731 filter &= ~SHOW_MEM_FILTER_NODES;
3733 show_mem(filter, nodemask);
3736 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3738 struct va_format vaf;
3740 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3741 DEFAULT_RATELIMIT_BURST);
3743 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3746 va_start(args, fmt);
3749 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3750 current->comm, &vaf, gfp_mask, &gfp_mask,
3751 nodemask_pr_args(nodemask));
3754 cpuset_print_current_mems_allowed();
3757 warn_alloc_show_mem(gfp_mask, nodemask);
3760 static inline struct page *
3761 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3762 unsigned int alloc_flags,
3763 const struct alloc_context *ac)
3767 page = get_page_from_freelist(gfp_mask, order,
3768 alloc_flags|ALLOC_CPUSET, ac);
3770 * fallback to ignore cpuset restriction if our nodes
3774 page = get_page_from_freelist(gfp_mask, order,
3780 static inline struct page *
3781 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3782 const struct alloc_context *ac, unsigned long *did_some_progress)
3784 struct oom_control oc = {
3785 .zonelist = ac->zonelist,
3786 .nodemask = ac->nodemask,
3788 .gfp_mask = gfp_mask,
3793 *did_some_progress = 0;
3796 * Acquire the oom lock. If that fails, somebody else is
3797 * making progress for us.
3799 if (!mutex_trylock(&oom_lock)) {
3800 *did_some_progress = 1;
3801 schedule_timeout_uninterruptible(1);
3806 * Go through the zonelist yet one more time, keep very high watermark
3807 * here, this is only to catch a parallel oom killing, we must fail if
3808 * we're still under heavy pressure. But make sure that this reclaim
3809 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3810 * allocation which will never fail due to oom_lock already held.
3812 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3813 ~__GFP_DIRECT_RECLAIM, order,
3814 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3818 /* Coredumps can quickly deplete all memory reserves */
3819 if (current->flags & PF_DUMPCORE)
3821 /* The OOM killer will not help higher order allocs */
3822 if (order > PAGE_ALLOC_COSTLY_ORDER)
3825 * We have already exhausted all our reclaim opportunities without any
3826 * success so it is time to admit defeat. We will skip the OOM killer
3827 * because it is very likely that the caller has a more reasonable
3828 * fallback than shooting a random task.
3830 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3832 /* The OOM killer does not needlessly kill tasks for lowmem */
3833 if (ac->high_zoneidx < ZONE_NORMAL)
3835 if (pm_suspended_storage())
3838 * XXX: GFP_NOFS allocations should rather fail than rely on
3839 * other request to make a forward progress.
3840 * We are in an unfortunate situation where out_of_memory cannot
3841 * do much for this context but let's try it to at least get
3842 * access to memory reserved if the current task is killed (see
3843 * out_of_memory). Once filesystems are ready to handle allocation
3844 * failures more gracefully we should just bail out here.
3847 /* The OOM killer may not free memory on a specific node */
3848 if (gfp_mask & __GFP_THISNODE)
3851 /* Exhausted what can be done so it's blame time */
3852 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3853 *did_some_progress = 1;
3856 * Help non-failing allocations by giving them access to memory
3859 if (gfp_mask & __GFP_NOFAIL)
3860 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3861 ALLOC_NO_WATERMARKS, ac);
3864 mutex_unlock(&oom_lock);
3869 * Maximum number of compaction retries wit a progress before OOM
3870 * killer is consider as the only way to move forward.
3872 #define MAX_COMPACT_RETRIES 16
3874 #ifdef CONFIG_COMPACTION
3875 /* Try memory compaction for high-order allocations before reclaim */
3876 static struct page *
3877 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3878 unsigned int alloc_flags, const struct alloc_context *ac,
3879 enum compact_priority prio, enum compact_result *compact_result)
3881 struct page *page = NULL;
3882 unsigned long pflags;
3883 unsigned int noreclaim_flag;
3888 psi_memstall_enter(&pflags);
3889 noreclaim_flag = memalloc_noreclaim_save();
3891 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3894 memalloc_noreclaim_restore(noreclaim_flag);
3895 psi_memstall_leave(&pflags);
3898 * At least in one zone compaction wasn't deferred or skipped, so let's
3899 * count a compaction stall
3901 count_vm_event(COMPACTSTALL);
3903 /* Prep a captured page if available */
3905 prep_new_page(page, order, gfp_mask, alloc_flags);
3907 /* Try get a page from the freelist if available */
3909 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3912 struct zone *zone = page_zone(page);
3914 zone->compact_blockskip_flush = false;
3915 compaction_defer_reset(zone, order, true);
3916 count_vm_event(COMPACTSUCCESS);
3921 * It's bad if compaction run occurs and fails. The most likely reason
3922 * is that pages exist, but not enough to satisfy watermarks.
3924 count_vm_event(COMPACTFAIL);
3932 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3933 enum compact_result compact_result,
3934 enum compact_priority *compact_priority,
3935 int *compaction_retries)
3937 int max_retries = MAX_COMPACT_RETRIES;
3940 int retries = *compaction_retries;
3941 enum compact_priority priority = *compact_priority;
3946 if (compaction_made_progress(compact_result))
3947 (*compaction_retries)++;
3950 * compaction considers all the zone as desperately out of memory
3951 * so it doesn't really make much sense to retry except when the
3952 * failure could be caused by insufficient priority
3954 if (compaction_failed(compact_result))
3955 goto check_priority;
3958 * compaction was skipped because there are not enough order-0 pages
3959 * to work with, so we retry only if it looks like reclaim can help.
3961 if (compaction_needs_reclaim(compact_result)) {
3962 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3967 * make sure the compaction wasn't deferred or didn't bail out early
3968 * due to locks contention before we declare that we should give up.
3969 * But the next retry should use a higher priority if allowed, so
3970 * we don't just keep bailing out endlessly.
3972 if (compaction_withdrawn(compact_result)) {
3973 goto check_priority;
3977 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3978 * costly ones because they are de facto nofail and invoke OOM
3979 * killer to move on while costly can fail and users are ready
3980 * to cope with that. 1/4 retries is rather arbitrary but we
3981 * would need much more detailed feedback from compaction to
3982 * make a better decision.
3984 if (order > PAGE_ALLOC_COSTLY_ORDER)
3986 if (*compaction_retries <= max_retries) {
3992 * Make sure there are attempts at the highest priority if we exhausted
3993 * all retries or failed at the lower priorities.
3996 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3997 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3999 if (*compact_priority > min_priority) {
4000 (*compact_priority)--;
4001 *compaction_retries = 0;
4005 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4009 static inline struct page *
4010 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4011 unsigned int alloc_flags, const struct alloc_context *ac,
4012 enum compact_priority prio, enum compact_result *compact_result)
4014 *compact_result = COMPACT_SKIPPED;
4019 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4020 enum compact_result compact_result,
4021 enum compact_priority *compact_priority,
4022 int *compaction_retries)
4027 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4031 * There are setups with compaction disabled which would prefer to loop
4032 * inside the allocator rather than hit the oom killer prematurely.
4033 * Let's give them a good hope and keep retrying while the order-0
4034 * watermarks are OK.
4036 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4038 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4039 ac_classzone_idx(ac), alloc_flags))
4044 #endif /* CONFIG_COMPACTION */
4046 #ifdef CONFIG_LOCKDEP
4047 static struct lockdep_map __fs_reclaim_map =
4048 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4050 static bool __need_fs_reclaim(gfp_t gfp_mask)
4052 gfp_mask = current_gfp_context(gfp_mask);
4054 /* no reclaim without waiting on it */
4055 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4058 /* this guy won't enter reclaim */
4059 if (current->flags & PF_MEMALLOC)
4062 /* We're only interested __GFP_FS allocations for now */
4063 if (!(gfp_mask & __GFP_FS))
4066 if (gfp_mask & __GFP_NOLOCKDEP)
4072 void __fs_reclaim_acquire(void)
4074 lock_map_acquire(&__fs_reclaim_map);
4077 void __fs_reclaim_release(void)
4079 lock_map_release(&__fs_reclaim_map);
4082 void fs_reclaim_acquire(gfp_t gfp_mask)
4084 if (__need_fs_reclaim(gfp_mask))
4085 __fs_reclaim_acquire();
4087 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4089 void fs_reclaim_release(gfp_t gfp_mask)
4091 if (__need_fs_reclaim(gfp_mask))
4092 __fs_reclaim_release();
4094 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4097 /* Perform direct synchronous page reclaim */
4099 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4100 const struct alloc_context *ac)
4103 unsigned int noreclaim_flag;
4104 unsigned long pflags;
4108 /* We now go into synchronous reclaim */
4109 cpuset_memory_pressure_bump();
4110 psi_memstall_enter(&pflags);
4111 fs_reclaim_acquire(gfp_mask);
4112 noreclaim_flag = memalloc_noreclaim_save();
4114 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4117 memalloc_noreclaim_restore(noreclaim_flag);
4118 fs_reclaim_release(gfp_mask);
4119 psi_memstall_leave(&pflags);
4126 /* The really slow allocator path where we enter direct reclaim */
4127 static inline struct page *
4128 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4129 unsigned int alloc_flags, const struct alloc_context *ac,
4130 unsigned long *did_some_progress)
4132 struct page *page = NULL;
4133 bool drained = false;
4135 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4136 if (unlikely(!(*did_some_progress)))
4140 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4143 * If an allocation failed after direct reclaim, it could be because
4144 * pages are pinned on the per-cpu lists or in high alloc reserves.
4145 * Shrink them them and try again
4147 if (!page && !drained) {
4148 unreserve_highatomic_pageblock(ac, false);
4149 drain_all_pages(NULL);
4157 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4158 const struct alloc_context *ac)
4162 pg_data_t *last_pgdat = NULL;
4163 enum zone_type high_zoneidx = ac->high_zoneidx;
4165 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
4167 if (last_pgdat != zone->zone_pgdat)
4168 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
4169 last_pgdat = zone->zone_pgdat;
4173 static inline unsigned int
4174 gfp_to_alloc_flags(gfp_t gfp_mask)
4176 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4178 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
4179 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4182 * The caller may dip into page reserves a bit more if the caller
4183 * cannot run direct reclaim, or if the caller has realtime scheduling
4184 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4185 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4187 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
4189 if (gfp_mask & __GFP_ATOMIC) {
4191 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4192 * if it can't schedule.
4194 if (!(gfp_mask & __GFP_NOMEMALLOC))
4195 alloc_flags |= ALLOC_HARDER;
4197 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4198 * comment for __cpuset_node_allowed().
4200 alloc_flags &= ~ALLOC_CPUSET;
4201 } else if (unlikely(rt_task(current)) && !in_interrupt())
4202 alloc_flags |= ALLOC_HARDER;
4204 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4205 alloc_flags |= ALLOC_KSWAPD;
4208 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4209 alloc_flags |= ALLOC_CMA;
4214 static bool oom_reserves_allowed(struct task_struct *tsk)
4216 if (!tsk_is_oom_victim(tsk))
4220 * !MMU doesn't have oom reaper so give access to memory reserves
4221 * only to the thread with TIF_MEMDIE set
4223 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4230 * Distinguish requests which really need access to full memory
4231 * reserves from oom victims which can live with a portion of it
4233 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4235 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4237 if (gfp_mask & __GFP_MEMALLOC)
4238 return ALLOC_NO_WATERMARKS;
4239 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4240 return ALLOC_NO_WATERMARKS;
4241 if (!in_interrupt()) {
4242 if (current->flags & PF_MEMALLOC)
4243 return ALLOC_NO_WATERMARKS;
4244 else if (oom_reserves_allowed(current))
4251 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4253 return !!__gfp_pfmemalloc_flags(gfp_mask);
4257 * Checks whether it makes sense to retry the reclaim to make a forward progress
4258 * for the given allocation request.
4260 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4261 * without success, or when we couldn't even meet the watermark if we
4262 * reclaimed all remaining pages on the LRU lists.
4264 * Returns true if a retry is viable or false to enter the oom path.
4267 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4268 struct alloc_context *ac, int alloc_flags,
4269 bool did_some_progress, int *no_progress_loops)
4276 * Costly allocations might have made a progress but this doesn't mean
4277 * their order will become available due to high fragmentation so
4278 * always increment the no progress counter for them
4280 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4281 *no_progress_loops = 0;
4283 (*no_progress_loops)++;
4286 * Make sure we converge to OOM if we cannot make any progress
4287 * several times in the row.
4289 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4290 /* Before OOM, exhaust highatomic_reserve */
4291 return unreserve_highatomic_pageblock(ac, true);
4295 * Keep reclaiming pages while there is a chance this will lead
4296 * somewhere. If none of the target zones can satisfy our allocation
4297 * request even if all reclaimable pages are considered then we are
4298 * screwed and have to go OOM.
4300 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4302 unsigned long available;
4303 unsigned long reclaimable;
4304 unsigned long min_wmark = min_wmark_pages(zone);
4307 available = reclaimable = zone_reclaimable_pages(zone);
4308 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4311 * Would the allocation succeed if we reclaimed all
4312 * reclaimable pages?
4314 wmark = __zone_watermark_ok(zone, order, min_wmark,
4315 ac_classzone_idx(ac), alloc_flags, available);
4316 trace_reclaim_retry_zone(z, order, reclaimable,
4317 available, min_wmark, *no_progress_loops, wmark);
4320 * If we didn't make any progress and have a lot of
4321 * dirty + writeback pages then we should wait for
4322 * an IO to complete to slow down the reclaim and
4323 * prevent from pre mature OOM
4325 if (!did_some_progress) {
4326 unsigned long write_pending;
4328 write_pending = zone_page_state_snapshot(zone,
4329 NR_ZONE_WRITE_PENDING);
4331 if (2 * write_pending > reclaimable) {
4332 congestion_wait(BLK_RW_ASYNC, HZ/10);
4344 * Memory allocation/reclaim might be called from a WQ context and the
4345 * current implementation of the WQ concurrency control doesn't
4346 * recognize that a particular WQ is congested if the worker thread is
4347 * looping without ever sleeping. Therefore we have to do a short sleep
4348 * here rather than calling cond_resched().
4350 if (current->flags & PF_WQ_WORKER)
4351 schedule_timeout_uninterruptible(1);
4358 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4361 * It's possible that cpuset's mems_allowed and the nodemask from
4362 * mempolicy don't intersect. This should be normally dealt with by
4363 * policy_nodemask(), but it's possible to race with cpuset update in
4364 * such a way the check therein was true, and then it became false
4365 * before we got our cpuset_mems_cookie here.
4366 * This assumes that for all allocations, ac->nodemask can come only
4367 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4368 * when it does not intersect with the cpuset restrictions) or the
4369 * caller can deal with a violated nodemask.
4371 if (cpusets_enabled() && ac->nodemask &&
4372 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4373 ac->nodemask = NULL;
4378 * When updating a task's mems_allowed or mempolicy nodemask, it is
4379 * possible to race with parallel threads in such a way that our
4380 * allocation can fail while the mask is being updated. If we are about
4381 * to fail, check if the cpuset changed during allocation and if so,
4384 if (read_mems_allowed_retry(cpuset_mems_cookie))
4390 static inline struct page *
4391 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4392 struct alloc_context *ac)
4394 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4395 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4396 struct page *page = NULL;
4397 unsigned int alloc_flags;
4398 unsigned long did_some_progress;
4399 enum compact_priority compact_priority;
4400 enum compact_result compact_result;
4401 int compaction_retries;
4402 int no_progress_loops;
4403 unsigned int cpuset_mems_cookie;
4407 * We also sanity check to catch abuse of atomic reserves being used by
4408 * callers that are not in atomic context.
4410 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4411 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4412 gfp_mask &= ~__GFP_ATOMIC;
4415 compaction_retries = 0;
4416 no_progress_loops = 0;
4417 compact_priority = DEF_COMPACT_PRIORITY;
4418 cpuset_mems_cookie = read_mems_allowed_begin();
4421 * The fast path uses conservative alloc_flags to succeed only until
4422 * kswapd needs to be woken up, and to avoid the cost of setting up
4423 * alloc_flags precisely. So we do that now.
4425 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4428 * We need to recalculate the starting point for the zonelist iterator
4429 * because we might have used different nodemask in the fast path, or
4430 * there was a cpuset modification and we are retrying - otherwise we
4431 * could end up iterating over non-eligible zones endlessly.
4433 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4434 ac->high_zoneidx, ac->nodemask);
4435 if (!ac->preferred_zoneref->zone)
4438 if (alloc_flags & ALLOC_KSWAPD)
4439 wake_all_kswapds(order, gfp_mask, ac);
4442 * The adjusted alloc_flags might result in immediate success, so try
4445 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4450 * For costly allocations, try direct compaction first, as it's likely
4451 * that we have enough base pages and don't need to reclaim. For non-
4452 * movable high-order allocations, do that as well, as compaction will
4453 * try prevent permanent fragmentation by migrating from blocks of the
4455 * Don't try this for allocations that are allowed to ignore
4456 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4458 if (can_direct_reclaim &&
4460 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4461 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4462 page = __alloc_pages_direct_compact(gfp_mask, order,
4464 INIT_COMPACT_PRIORITY,
4470 * Checks for costly allocations with __GFP_NORETRY, which
4471 * includes THP page fault allocations
4473 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4475 * If compaction is deferred for high-order allocations,
4476 * it is because sync compaction recently failed. If
4477 * this is the case and the caller requested a THP
4478 * allocation, we do not want to heavily disrupt the
4479 * system, so we fail the allocation instead of entering
4482 if (compact_result == COMPACT_DEFERRED)
4486 * Looks like reclaim/compaction is worth trying, but
4487 * sync compaction could be very expensive, so keep
4488 * using async compaction.
4490 compact_priority = INIT_COMPACT_PRIORITY;
4495 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4496 if (alloc_flags & ALLOC_KSWAPD)
4497 wake_all_kswapds(order, gfp_mask, ac);
4499 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4501 alloc_flags = reserve_flags;
4504 * Reset the nodemask and zonelist iterators if memory policies can be
4505 * ignored. These allocations are high priority and system rather than
4508 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4509 ac->nodemask = NULL;
4510 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4511 ac->high_zoneidx, ac->nodemask);
4514 /* Attempt with potentially adjusted zonelist and alloc_flags */
4515 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4519 /* Caller is not willing to reclaim, we can't balance anything */
4520 if (!can_direct_reclaim)
4523 /* Avoid recursion of direct reclaim */
4524 if (current->flags & PF_MEMALLOC)
4527 /* Try direct reclaim and then allocating */
4528 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4529 &did_some_progress);
4533 /* Try direct compaction and then allocating */
4534 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4535 compact_priority, &compact_result);
4539 /* Do not loop if specifically requested */
4540 if (gfp_mask & __GFP_NORETRY)
4544 * Do not retry costly high order allocations unless they are
4545 * __GFP_RETRY_MAYFAIL
4547 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4550 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4551 did_some_progress > 0, &no_progress_loops))
4555 * It doesn't make any sense to retry for the compaction if the order-0
4556 * reclaim is not able to make any progress because the current
4557 * implementation of the compaction depends on the sufficient amount
4558 * of free memory (see __compaction_suitable)
4560 if (did_some_progress > 0 &&
4561 should_compact_retry(ac, order, alloc_flags,
4562 compact_result, &compact_priority,
4563 &compaction_retries))
4567 /* Deal with possible cpuset update races before we start OOM killing */
4568 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4571 /* Reclaim has failed us, start killing things */
4572 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4576 /* Avoid allocations with no watermarks from looping endlessly */
4577 if (tsk_is_oom_victim(current) &&
4578 (alloc_flags == ALLOC_OOM ||
4579 (gfp_mask & __GFP_NOMEMALLOC)))
4582 /* Retry as long as the OOM killer is making progress */
4583 if (did_some_progress) {
4584 no_progress_loops = 0;
4589 /* Deal with possible cpuset update races before we fail */
4590 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4594 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4597 if (gfp_mask & __GFP_NOFAIL) {
4599 * All existing users of the __GFP_NOFAIL are blockable, so warn
4600 * of any new users that actually require GFP_NOWAIT
4602 if (WARN_ON_ONCE(!can_direct_reclaim))
4606 * PF_MEMALLOC request from this context is rather bizarre
4607 * because we cannot reclaim anything and only can loop waiting
4608 * for somebody to do a work for us
4610 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4613 * non failing costly orders are a hard requirement which we
4614 * are not prepared for much so let's warn about these users
4615 * so that we can identify them and convert them to something
4618 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4621 * Help non-failing allocations by giving them access to memory
4622 * reserves but do not use ALLOC_NO_WATERMARKS because this
4623 * could deplete whole memory reserves which would just make
4624 * the situation worse
4626 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4634 warn_alloc(gfp_mask, ac->nodemask,
4635 "page allocation failure: order:%u", order);
4640 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4641 int preferred_nid, nodemask_t *nodemask,
4642 struct alloc_context *ac, gfp_t *alloc_mask,
4643 unsigned int *alloc_flags)
4645 ac->high_zoneidx = gfp_zone(gfp_mask);
4646 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4647 ac->nodemask = nodemask;
4648 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4650 if (cpusets_enabled()) {
4651 *alloc_mask |= __GFP_HARDWALL;
4653 ac->nodemask = &cpuset_current_mems_allowed;
4655 *alloc_flags |= ALLOC_CPUSET;
4658 fs_reclaim_acquire(gfp_mask);
4659 fs_reclaim_release(gfp_mask);
4661 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4663 if (should_fail_alloc_page(gfp_mask, order))
4666 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4667 *alloc_flags |= ALLOC_CMA;
4672 /* Determine whether to spread dirty pages and what the first usable zone */
4673 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4675 /* Dirty zone balancing only done in the fast path */
4676 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4679 * The preferred zone is used for statistics but crucially it is
4680 * also used as the starting point for the zonelist iterator. It
4681 * may get reset for allocations that ignore memory policies.
4683 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4684 ac->high_zoneidx, ac->nodemask);
4688 * This is the 'heart' of the zoned buddy allocator.
4691 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4692 nodemask_t *nodemask)
4695 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4696 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4697 struct alloc_context ac = { };
4700 * There are several places where we assume that the order value is sane
4701 * so bail out early if the request is out of bound.
4703 if (unlikely(order >= MAX_ORDER)) {
4704 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4708 gfp_mask &= gfp_allowed_mask;
4709 alloc_mask = gfp_mask;
4710 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4713 finalise_ac(gfp_mask, &ac);
4716 * Forbid the first pass from falling back to types that fragment
4717 * memory until all local zones are considered.
4719 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4721 /* First allocation attempt */
4722 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4727 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4728 * resp. GFP_NOIO which has to be inherited for all allocation requests
4729 * from a particular context which has been marked by
4730 * memalloc_no{fs,io}_{save,restore}.
4732 alloc_mask = current_gfp_context(gfp_mask);
4733 ac.spread_dirty_pages = false;
4736 * Restore the original nodemask if it was potentially replaced with
4737 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4739 if (unlikely(ac.nodemask != nodemask))
4740 ac.nodemask = nodemask;
4742 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4745 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4746 unlikely(__memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4747 __free_pages(page, order);
4751 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4755 EXPORT_SYMBOL(__alloc_pages_nodemask);
4758 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4759 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4760 * you need to access high mem.
4762 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4766 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4769 return (unsigned long) page_address(page);
4771 EXPORT_SYMBOL(__get_free_pages);
4773 unsigned long get_zeroed_page(gfp_t gfp_mask)
4775 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4777 EXPORT_SYMBOL(get_zeroed_page);
4779 static inline void free_the_page(struct page *page, unsigned int order)
4781 if (order == 0) /* Via pcp? */
4782 free_unref_page(page);
4784 __free_pages_ok(page, order);
4787 void __free_pages(struct page *page, unsigned int order)
4789 if (put_page_testzero(page))
4790 free_the_page(page, order);
4792 EXPORT_SYMBOL(__free_pages);
4794 void free_pages(unsigned long addr, unsigned int order)
4797 VM_BUG_ON(!virt_addr_valid((void *)addr));
4798 __free_pages(virt_to_page((void *)addr), order);
4802 EXPORT_SYMBOL(free_pages);
4806 * An arbitrary-length arbitrary-offset area of memory which resides
4807 * within a 0 or higher order page. Multiple fragments within that page
4808 * are individually refcounted, in the page's reference counter.
4810 * The page_frag functions below provide a simple allocation framework for
4811 * page fragments. This is used by the network stack and network device
4812 * drivers to provide a backing region of memory for use as either an
4813 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4815 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4818 struct page *page = NULL;
4819 gfp_t gfp = gfp_mask;
4821 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4822 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4824 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4825 PAGE_FRAG_CACHE_MAX_ORDER);
4826 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4828 if (unlikely(!page))
4829 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4831 nc->va = page ? page_address(page) : NULL;
4836 void __page_frag_cache_drain(struct page *page, unsigned int count)
4838 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4840 if (page_ref_sub_and_test(page, count))
4841 free_the_page(page, compound_order(page));
4843 EXPORT_SYMBOL(__page_frag_cache_drain);
4845 void *page_frag_alloc(struct page_frag_cache *nc,
4846 unsigned int fragsz, gfp_t gfp_mask)
4848 unsigned int size = PAGE_SIZE;
4852 if (unlikely(!nc->va)) {
4854 page = __page_frag_cache_refill(nc, gfp_mask);
4858 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4859 /* if size can vary use size else just use PAGE_SIZE */
4862 /* Even if we own the page, we do not use atomic_set().
4863 * This would break get_page_unless_zero() users.
4865 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4867 /* reset page count bias and offset to start of new frag */
4868 nc->pfmemalloc = page_is_pfmemalloc(page);
4869 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4873 offset = nc->offset - fragsz;
4874 if (unlikely(offset < 0)) {
4875 page = virt_to_page(nc->va);
4877 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4880 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4881 /* if size can vary use size else just use PAGE_SIZE */
4884 /* OK, page count is 0, we can safely set it */
4885 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4887 /* reset page count bias and offset to start of new frag */
4888 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4889 offset = size - fragsz;
4893 nc->offset = offset;
4895 return nc->va + offset;
4897 EXPORT_SYMBOL(page_frag_alloc);
4900 * Frees a page fragment allocated out of either a compound or order 0 page.
4902 void page_frag_free(void *addr)
4904 struct page *page = virt_to_head_page(addr);
4906 if (unlikely(put_page_testzero(page)))
4907 free_the_page(page, compound_order(page));
4909 EXPORT_SYMBOL(page_frag_free);
4911 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4915 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4916 unsigned long used = addr + PAGE_ALIGN(size);
4918 split_page(virt_to_page((void *)addr), order);
4919 while (used < alloc_end) {
4924 return (void *)addr;
4928 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4929 * @size: the number of bytes to allocate
4930 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4932 * This function is similar to alloc_pages(), except that it allocates the
4933 * minimum number of pages to satisfy the request. alloc_pages() can only
4934 * allocate memory in power-of-two pages.
4936 * This function is also limited by MAX_ORDER.
4938 * Memory allocated by this function must be released by free_pages_exact().
4940 * Return: pointer to the allocated area or %NULL in case of error.
4942 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4944 unsigned int order = get_order(size);
4947 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
4948 gfp_mask &= ~__GFP_COMP;
4950 addr = __get_free_pages(gfp_mask, order);
4951 return make_alloc_exact(addr, order, size);
4953 EXPORT_SYMBOL(alloc_pages_exact);
4956 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4958 * @nid: the preferred node ID where memory should be allocated
4959 * @size: the number of bytes to allocate
4960 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4962 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4965 * Return: pointer to the allocated area or %NULL in case of error.
4967 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4969 unsigned int order = get_order(size);
4972 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
4973 gfp_mask &= ~__GFP_COMP;
4975 p = alloc_pages_node(nid, gfp_mask, order);
4978 return make_alloc_exact((unsigned long)page_address(p), order, size);
4982 * free_pages_exact - release memory allocated via alloc_pages_exact()
4983 * @virt: the value returned by alloc_pages_exact.
4984 * @size: size of allocation, same value as passed to alloc_pages_exact().
4986 * Release the memory allocated by a previous call to alloc_pages_exact.
4988 void free_pages_exact(void *virt, size_t size)
4990 unsigned long addr = (unsigned long)virt;
4991 unsigned long end = addr + PAGE_ALIGN(size);
4993 while (addr < end) {
4998 EXPORT_SYMBOL(free_pages_exact);
5001 * nr_free_zone_pages - count number of pages beyond high watermark
5002 * @offset: The zone index of the highest zone
5004 * nr_free_zone_pages() counts the number of pages which are beyond the
5005 * high watermark within all zones at or below a given zone index. For each
5006 * zone, the number of pages is calculated as:
5008 * nr_free_zone_pages = managed_pages - high_pages
5010 * Return: number of pages beyond high watermark.
5012 static unsigned long nr_free_zone_pages(int offset)
5017 /* Just pick one node, since fallback list is circular */
5018 unsigned long sum = 0;
5020 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5022 for_each_zone_zonelist(zone, z, zonelist, offset) {
5023 unsigned long size = zone_managed_pages(zone);
5024 unsigned long high = high_wmark_pages(zone);
5033 * nr_free_buffer_pages - count number of pages beyond high watermark
5035 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5036 * watermark within ZONE_DMA and ZONE_NORMAL.
5038 * Return: number of pages beyond high watermark within ZONE_DMA and
5041 unsigned long nr_free_buffer_pages(void)
5043 return nr_free_zone_pages(gfp_zone(GFP_USER));
5045 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5048 * nr_free_pagecache_pages - count number of pages beyond high watermark
5050 * nr_free_pagecache_pages() counts the number of pages which are beyond the
5051 * high watermark within all zones.
5053 * Return: number of pages beyond high watermark within all zones.
5055 unsigned long nr_free_pagecache_pages(void)
5057 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5060 static inline void show_node(struct zone *zone)
5062 if (IS_ENABLED(CONFIG_NUMA))
5063 printk("Node %d ", zone_to_nid(zone));
5066 long si_mem_available(void)
5069 unsigned long pagecache;
5070 unsigned long wmark_low = 0;
5071 unsigned long pages[NR_LRU_LISTS];
5072 unsigned long reclaimable;
5076 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5077 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5080 wmark_low += low_wmark_pages(zone);
5083 * Estimate the amount of memory available for userspace allocations,
5084 * without causing swapping.
5086 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5089 * Not all the page cache can be freed, otherwise the system will
5090 * start swapping. Assume at least half of the page cache, or the
5091 * low watermark worth of cache, needs to stay.
5093 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5094 pagecache -= min(pagecache / 2, wmark_low);
5095 available += pagecache;
5098 * Part of the reclaimable slab and other kernel memory consists of
5099 * items that are in use, and cannot be freed. Cap this estimate at the
5102 reclaimable = global_node_page_state(NR_SLAB_RECLAIMABLE) +
5103 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5104 available += reclaimable - min(reclaimable / 2, wmark_low);
5110 EXPORT_SYMBOL_GPL(si_mem_available);
5112 void si_meminfo(struct sysinfo *val)
5114 val->totalram = totalram_pages();
5115 val->sharedram = global_node_page_state(NR_SHMEM);
5116 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5117 val->bufferram = nr_blockdev_pages();
5118 val->totalhigh = totalhigh_pages();
5119 val->freehigh = nr_free_highpages();
5120 val->mem_unit = PAGE_SIZE;
5123 EXPORT_SYMBOL(si_meminfo);
5126 void si_meminfo_node(struct sysinfo *val, int nid)
5128 int zone_type; /* needs to be signed */
5129 unsigned long managed_pages = 0;
5130 unsigned long managed_highpages = 0;
5131 unsigned long free_highpages = 0;
5132 pg_data_t *pgdat = NODE_DATA(nid);
5134 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5135 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5136 val->totalram = managed_pages;
5137 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5138 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5139 #ifdef CONFIG_HIGHMEM
5140 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5141 struct zone *zone = &pgdat->node_zones[zone_type];
5143 if (is_highmem(zone)) {
5144 managed_highpages += zone_managed_pages(zone);
5145 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5148 val->totalhigh = managed_highpages;
5149 val->freehigh = free_highpages;
5151 val->totalhigh = managed_highpages;
5152 val->freehigh = free_highpages;
5154 val->mem_unit = PAGE_SIZE;
5159 * Determine whether the node should be displayed or not, depending on whether
5160 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5162 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5164 if (!(flags & SHOW_MEM_FILTER_NODES))
5168 * no node mask - aka implicit memory numa policy. Do not bother with
5169 * the synchronization - read_mems_allowed_begin - because we do not
5170 * have to be precise here.
5173 nodemask = &cpuset_current_mems_allowed;
5175 return !node_isset(nid, *nodemask);
5178 #define K(x) ((x) << (PAGE_SHIFT-10))
5180 static void show_migration_types(unsigned char type)
5182 static const char types[MIGRATE_TYPES] = {
5183 [MIGRATE_UNMOVABLE] = 'U',
5184 [MIGRATE_MOVABLE] = 'M',
5185 [MIGRATE_RECLAIMABLE] = 'E',
5186 [MIGRATE_HIGHATOMIC] = 'H',
5188 [MIGRATE_CMA] = 'C',
5190 #ifdef CONFIG_MEMORY_ISOLATION
5191 [MIGRATE_ISOLATE] = 'I',
5194 char tmp[MIGRATE_TYPES + 1];
5198 for (i = 0; i < MIGRATE_TYPES; i++) {
5199 if (type & (1 << i))
5204 printk(KERN_CONT "(%s) ", tmp);
5208 * Show free area list (used inside shift_scroll-lock stuff)
5209 * We also calculate the percentage fragmentation. We do this by counting the
5210 * memory on each free list with the exception of the first item on the list.
5213 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5216 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5218 unsigned long free_pcp = 0;
5223 for_each_populated_zone(zone) {
5224 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5227 for_each_online_cpu(cpu)
5228 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5231 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5232 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5233 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
5234 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5235 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5236 " free:%lu free_pcp:%lu free_cma:%lu\n",
5237 global_node_page_state(NR_ACTIVE_ANON),
5238 global_node_page_state(NR_INACTIVE_ANON),
5239 global_node_page_state(NR_ISOLATED_ANON),
5240 global_node_page_state(NR_ACTIVE_FILE),
5241 global_node_page_state(NR_INACTIVE_FILE),
5242 global_node_page_state(NR_ISOLATED_FILE),
5243 global_node_page_state(NR_UNEVICTABLE),
5244 global_node_page_state(NR_FILE_DIRTY),
5245 global_node_page_state(NR_WRITEBACK),
5246 global_node_page_state(NR_UNSTABLE_NFS),
5247 global_node_page_state(NR_SLAB_RECLAIMABLE),
5248 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
5249 global_node_page_state(NR_FILE_MAPPED),
5250 global_node_page_state(NR_SHMEM),
5251 global_zone_page_state(NR_PAGETABLE),
5252 global_zone_page_state(NR_BOUNCE),
5253 global_zone_page_state(NR_FREE_PAGES),
5255 global_zone_page_state(NR_FREE_CMA_PAGES));
5257 for_each_online_pgdat(pgdat) {
5258 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5262 " active_anon:%lukB"
5263 " inactive_anon:%lukB"
5264 " active_file:%lukB"
5265 " inactive_file:%lukB"
5266 " unevictable:%lukB"
5267 " isolated(anon):%lukB"
5268 " isolated(file):%lukB"
5273 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5275 " shmem_pmdmapped: %lukB"
5278 " writeback_tmp:%lukB"
5280 " all_unreclaimable? %s"
5283 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5284 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5285 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5286 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5287 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5288 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5289 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5290 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5291 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5292 K(node_page_state(pgdat, NR_WRITEBACK)),
5293 K(node_page_state(pgdat, NR_SHMEM)),
5294 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5295 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5296 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5298 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5300 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5301 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
5302 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5306 for_each_populated_zone(zone) {
5309 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5313 for_each_online_cpu(cpu)
5314 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5323 " active_anon:%lukB"
5324 " inactive_anon:%lukB"
5325 " active_file:%lukB"
5326 " inactive_file:%lukB"
5327 " unevictable:%lukB"
5328 " writepending:%lukB"
5332 " kernel_stack:%lukB"
5340 K(zone_page_state(zone, NR_FREE_PAGES)),
5341 K(min_wmark_pages(zone)),
5342 K(low_wmark_pages(zone)),
5343 K(high_wmark_pages(zone)),
5344 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5345 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5346 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5347 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5348 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5349 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5350 K(zone->present_pages),
5351 K(zone_managed_pages(zone)),
5352 K(zone_page_state(zone, NR_MLOCK)),
5353 zone_page_state(zone, NR_KERNEL_STACK_KB),
5354 K(zone_page_state(zone, NR_PAGETABLE)),
5355 K(zone_page_state(zone, NR_BOUNCE)),
5357 K(this_cpu_read(zone->pageset->pcp.count)),
5358 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5359 printk("lowmem_reserve[]:");
5360 for (i = 0; i < MAX_NR_ZONES; i++)
5361 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5362 printk(KERN_CONT "\n");
5365 for_each_populated_zone(zone) {
5367 unsigned long nr[MAX_ORDER], flags, total = 0;
5368 unsigned char types[MAX_ORDER];
5370 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5373 printk(KERN_CONT "%s: ", zone->name);
5375 spin_lock_irqsave(&zone->lock, flags);
5376 for (order = 0; order < MAX_ORDER; order++) {
5377 struct free_area *area = &zone->free_area[order];
5380 nr[order] = area->nr_free;
5381 total += nr[order] << order;
5384 for (type = 0; type < MIGRATE_TYPES; type++) {
5385 if (!free_area_empty(area, type))
5386 types[order] |= 1 << type;
5389 spin_unlock_irqrestore(&zone->lock, flags);
5390 for (order = 0; order < MAX_ORDER; order++) {
5391 printk(KERN_CONT "%lu*%lukB ",
5392 nr[order], K(1UL) << order);
5394 show_migration_types(types[order]);
5396 printk(KERN_CONT "= %lukB\n", K(total));
5399 hugetlb_show_meminfo();
5401 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5403 show_swap_cache_info();
5406 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5408 zoneref->zone = zone;
5409 zoneref->zone_idx = zone_idx(zone);
5413 * Builds allocation fallback zone lists.
5415 * Add all populated zones of a node to the zonelist.
5417 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5420 enum zone_type zone_type = MAX_NR_ZONES;
5425 zone = pgdat->node_zones + zone_type;
5426 if (managed_zone(zone)) {
5427 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5428 check_highest_zone(zone_type);
5430 } while (zone_type);
5437 static int __parse_numa_zonelist_order(char *s)
5440 * We used to support different zonlists modes but they turned
5441 * out to be just not useful. Let's keep the warning in place
5442 * if somebody still use the cmd line parameter so that we do
5443 * not fail it silently
5445 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5446 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5452 static __init int setup_numa_zonelist_order(char *s)
5457 return __parse_numa_zonelist_order(s);
5459 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5461 char numa_zonelist_order[] = "Node";
5464 * sysctl handler for numa_zonelist_order
5466 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5467 void __user *buffer, size_t *length,
5474 return proc_dostring(table, write, buffer, length, ppos);
5475 str = memdup_user_nul(buffer, 16);
5477 return PTR_ERR(str);
5479 ret = __parse_numa_zonelist_order(str);
5485 #define MAX_NODE_LOAD (nr_online_nodes)
5486 static int node_load[MAX_NUMNODES];
5489 * find_next_best_node - find the next node that should appear in a given node's fallback list
5490 * @node: node whose fallback list we're appending
5491 * @used_node_mask: nodemask_t of already used nodes
5493 * We use a number of factors to determine which is the next node that should
5494 * appear on a given node's fallback list. The node should not have appeared
5495 * already in @node's fallback list, and it should be the next closest node
5496 * according to the distance array (which contains arbitrary distance values
5497 * from each node to each node in the system), and should also prefer nodes
5498 * with no CPUs, since presumably they'll have very little allocation pressure
5499 * on them otherwise.
5501 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5503 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5506 int min_val = INT_MAX;
5507 int best_node = NUMA_NO_NODE;
5508 const struct cpumask *tmp = cpumask_of_node(0);
5510 /* Use the local node if we haven't already */
5511 if (!node_isset(node, *used_node_mask)) {
5512 node_set(node, *used_node_mask);
5516 for_each_node_state(n, N_MEMORY) {
5518 /* Don't want a node to appear more than once */
5519 if (node_isset(n, *used_node_mask))
5522 /* Use the distance array to find the distance */
5523 val = node_distance(node, n);
5525 /* Penalize nodes under us ("prefer the next node") */
5528 /* Give preference to headless and unused nodes */
5529 tmp = cpumask_of_node(n);
5530 if (!cpumask_empty(tmp))
5531 val += PENALTY_FOR_NODE_WITH_CPUS;
5533 /* Slight preference for less loaded node */
5534 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5535 val += node_load[n];
5537 if (val < min_val) {
5544 node_set(best_node, *used_node_mask);
5551 * Build zonelists ordered by node and zones within node.
5552 * This results in maximum locality--normal zone overflows into local
5553 * DMA zone, if any--but risks exhausting DMA zone.
5555 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5558 struct zoneref *zonerefs;
5561 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5563 for (i = 0; i < nr_nodes; i++) {
5566 pg_data_t *node = NODE_DATA(node_order[i]);
5568 nr_zones = build_zonerefs_node(node, zonerefs);
5569 zonerefs += nr_zones;
5571 zonerefs->zone = NULL;
5572 zonerefs->zone_idx = 0;
5576 * Build gfp_thisnode zonelists
5578 static void build_thisnode_zonelists(pg_data_t *pgdat)
5580 struct zoneref *zonerefs;
5583 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5584 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5585 zonerefs += nr_zones;
5586 zonerefs->zone = NULL;
5587 zonerefs->zone_idx = 0;
5591 * Build zonelists ordered by zone and nodes within zones.
5592 * This results in conserving DMA zone[s] until all Normal memory is
5593 * exhausted, but results in overflowing to remote node while memory
5594 * may still exist in local DMA zone.
5597 static void build_zonelists(pg_data_t *pgdat)
5599 static int node_order[MAX_NUMNODES];
5600 int node, load, nr_nodes = 0;
5601 nodemask_t used_mask;
5602 int local_node, prev_node;
5604 /* NUMA-aware ordering of nodes */
5605 local_node = pgdat->node_id;
5606 load = nr_online_nodes;
5607 prev_node = local_node;
5608 nodes_clear(used_mask);
5610 memset(node_order, 0, sizeof(node_order));
5611 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5613 * We don't want to pressure a particular node.
5614 * So adding penalty to the first node in same
5615 * distance group to make it round-robin.
5617 if (node_distance(local_node, node) !=
5618 node_distance(local_node, prev_node))
5619 node_load[node] = load;
5621 node_order[nr_nodes++] = node;
5626 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5627 build_thisnode_zonelists(pgdat);
5630 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5632 * Return node id of node used for "local" allocations.
5633 * I.e., first node id of first zone in arg node's generic zonelist.
5634 * Used for initializing percpu 'numa_mem', which is used primarily
5635 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5637 int local_memory_node(int node)
5641 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5642 gfp_zone(GFP_KERNEL),
5644 return zone_to_nid(z->zone);
5648 static void setup_min_unmapped_ratio(void);
5649 static void setup_min_slab_ratio(void);
5650 #else /* CONFIG_NUMA */
5652 static void build_zonelists(pg_data_t *pgdat)
5654 int node, local_node;
5655 struct zoneref *zonerefs;
5658 local_node = pgdat->node_id;
5660 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5661 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5662 zonerefs += nr_zones;
5665 * Now we build the zonelist so that it contains the zones
5666 * of all the other nodes.
5667 * We don't want to pressure a particular node, so when
5668 * building the zones for node N, we make sure that the
5669 * zones coming right after the local ones are those from
5670 * node N+1 (modulo N)
5672 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5673 if (!node_online(node))
5675 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5676 zonerefs += nr_zones;
5678 for (node = 0; node < local_node; node++) {
5679 if (!node_online(node))
5681 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5682 zonerefs += nr_zones;
5685 zonerefs->zone = NULL;
5686 zonerefs->zone_idx = 0;
5689 #endif /* CONFIG_NUMA */
5692 * Boot pageset table. One per cpu which is going to be used for all
5693 * zones and all nodes. The parameters will be set in such a way
5694 * that an item put on a list will immediately be handed over to
5695 * the buddy list. This is safe since pageset manipulation is done
5696 * with interrupts disabled.
5698 * The boot_pagesets must be kept even after bootup is complete for
5699 * unused processors and/or zones. They do play a role for bootstrapping
5700 * hotplugged processors.
5702 * zoneinfo_show() and maybe other functions do
5703 * not check if the processor is online before following the pageset pointer.
5704 * Other parts of the kernel may not check if the zone is available.
5706 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5707 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5708 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5710 static void __build_all_zonelists(void *data)
5713 int __maybe_unused cpu;
5714 pg_data_t *self = data;
5715 static DEFINE_SPINLOCK(lock);
5720 memset(node_load, 0, sizeof(node_load));
5724 * This node is hotadded and no memory is yet present. So just
5725 * building zonelists is fine - no need to touch other nodes.
5727 if (self && !node_online(self->node_id)) {
5728 build_zonelists(self);
5730 for_each_online_node(nid) {
5731 pg_data_t *pgdat = NODE_DATA(nid);
5733 build_zonelists(pgdat);
5736 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5738 * We now know the "local memory node" for each node--
5739 * i.e., the node of the first zone in the generic zonelist.
5740 * Set up numa_mem percpu variable for on-line cpus. During
5741 * boot, only the boot cpu should be on-line; we'll init the
5742 * secondary cpus' numa_mem as they come on-line. During
5743 * node/memory hotplug, we'll fixup all on-line cpus.
5745 for_each_online_cpu(cpu)
5746 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5753 static noinline void __init
5754 build_all_zonelists_init(void)
5758 __build_all_zonelists(NULL);
5761 * Initialize the boot_pagesets that are going to be used
5762 * for bootstrapping processors. The real pagesets for
5763 * each zone will be allocated later when the per cpu
5764 * allocator is available.
5766 * boot_pagesets are used also for bootstrapping offline
5767 * cpus if the system is already booted because the pagesets
5768 * are needed to initialize allocators on a specific cpu too.
5769 * F.e. the percpu allocator needs the page allocator which
5770 * needs the percpu allocator in order to allocate its pagesets
5771 * (a chicken-egg dilemma).
5773 for_each_possible_cpu(cpu)
5774 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5776 mminit_verify_zonelist();
5777 cpuset_init_current_mems_allowed();
5781 * unless system_state == SYSTEM_BOOTING.
5783 * __ref due to call of __init annotated helper build_all_zonelists_init
5784 * [protected by SYSTEM_BOOTING].
5786 void __ref build_all_zonelists(pg_data_t *pgdat)
5788 if (system_state == SYSTEM_BOOTING) {
5789 build_all_zonelists_init();
5791 __build_all_zonelists(pgdat);
5792 /* cpuset refresh routine should be here */
5794 vm_total_pages = nr_free_pagecache_pages();
5796 * Disable grouping by mobility if the number of pages in the
5797 * system is too low to allow the mechanism to work. It would be
5798 * more accurate, but expensive to check per-zone. This check is
5799 * made on memory-hotadd so a system can start with mobility
5800 * disabled and enable it later
5802 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5803 page_group_by_mobility_disabled = 1;
5805 page_group_by_mobility_disabled = 0;
5807 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5809 page_group_by_mobility_disabled ? "off" : "on",
5812 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5816 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5817 static bool __meminit
5818 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5820 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5821 static struct memblock_region *r;
5823 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5824 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5825 for_each_memblock(memory, r) {
5826 if (*pfn < memblock_region_memory_end_pfn(r))
5830 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5831 memblock_is_mirror(r)) {
5832 *pfn = memblock_region_memory_end_pfn(r);
5841 * Initially all pages are reserved - free ones are freed
5842 * up by memblock_free_all() once the early boot process is
5843 * done. Non-atomic initialization, single-pass.
5845 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5846 unsigned long start_pfn, enum memmap_context context,
5847 struct vmem_altmap *altmap)
5849 unsigned long pfn, end_pfn = start_pfn + size;
5852 if (highest_memmap_pfn < end_pfn - 1)
5853 highest_memmap_pfn = end_pfn - 1;
5855 #ifdef CONFIG_ZONE_DEVICE
5857 * Honor reservation requested by the driver for this ZONE_DEVICE
5858 * memory. We limit the total number of pages to initialize to just
5859 * those that might contain the memory mapping. We will defer the
5860 * ZONE_DEVICE page initialization until after we have released
5863 if (zone == ZONE_DEVICE) {
5867 if (start_pfn == altmap->base_pfn)
5868 start_pfn += altmap->reserve;
5869 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5873 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5875 * There can be holes in boot-time mem_map[]s handed to this
5876 * function. They do not exist on hotplugged memory.
5878 if (context == MEMMAP_EARLY) {
5879 if (!early_pfn_valid(pfn))
5881 if (!early_pfn_in_nid(pfn, nid))
5883 if (overlap_memmap_init(zone, &pfn))
5885 if (defer_init(nid, pfn, end_pfn))
5889 page = pfn_to_page(pfn);
5890 __init_single_page(page, pfn, zone, nid);
5891 if (context == MEMMAP_HOTPLUG)
5892 __SetPageReserved(page);
5895 * Mark the block movable so that blocks are reserved for
5896 * movable at startup. This will force kernel allocations
5897 * to reserve their blocks rather than leaking throughout
5898 * the address space during boot when many long-lived
5899 * kernel allocations are made.
5901 * bitmap is created for zone's valid pfn range. but memmap
5902 * can be created for invalid pages (for alignment)
5903 * check here not to call set_pageblock_migratetype() against
5906 if (!(pfn & (pageblock_nr_pages - 1))) {
5907 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5913 #ifdef CONFIG_ZONE_DEVICE
5914 void __ref memmap_init_zone_device(struct zone *zone,
5915 unsigned long start_pfn,
5917 struct dev_pagemap *pgmap)
5919 unsigned long pfn, end_pfn = start_pfn + size;
5920 struct pglist_data *pgdat = zone->zone_pgdat;
5921 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
5922 unsigned long zone_idx = zone_idx(zone);
5923 unsigned long start = jiffies;
5924 int nid = pgdat->node_id;
5926 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
5930 * The call to memmap_init_zone should have already taken care
5931 * of the pages reserved for the memmap, so we can just jump to
5932 * the end of that region and start processing the device pages.
5935 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5936 size = end_pfn - start_pfn;
5939 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5940 struct page *page = pfn_to_page(pfn);
5942 __init_single_page(page, pfn, zone_idx, nid);
5945 * Mark page reserved as it will need to wait for onlining
5946 * phase for it to be fully associated with a zone.
5948 * We can use the non-atomic __set_bit operation for setting
5949 * the flag as we are still initializing the pages.
5951 __SetPageReserved(page);
5954 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
5955 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
5956 * ever freed or placed on a driver-private list.
5958 page->pgmap = pgmap;
5959 page->zone_device_data = NULL;
5962 * Mark the block movable so that blocks are reserved for
5963 * movable at startup. This will force kernel allocations
5964 * to reserve their blocks rather than leaking throughout
5965 * the address space during boot when many long-lived
5966 * kernel allocations are made.
5968 * bitmap is created for zone's valid pfn range. but memmap
5969 * can be created for invalid pages (for alignment)
5970 * check here not to call set_pageblock_migratetype() against
5973 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
5974 * because this is done early in section_activate()
5976 if (!(pfn & (pageblock_nr_pages - 1))) {
5977 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5982 pr_info("%s initialised %lu pages in %ums\n", __func__,
5983 size, jiffies_to_msecs(jiffies - start));
5987 static void __meminit zone_init_free_lists(struct zone *zone)
5989 unsigned int order, t;
5990 for_each_migratetype_order(order, t) {
5991 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5992 zone->free_area[order].nr_free = 0;
5996 void __meminit __weak memmap_init(unsigned long size, int nid,
5997 unsigned long zone, unsigned long start_pfn)
5999 memmap_init_zone(size, nid, zone, start_pfn, MEMMAP_EARLY, NULL);
6002 static int zone_batchsize(struct zone *zone)
6008 * The per-cpu-pages pools are set to around 1000th of the
6011 batch = zone_managed_pages(zone) / 1024;
6012 /* But no more than a meg. */
6013 if (batch * PAGE_SIZE > 1024 * 1024)
6014 batch = (1024 * 1024) / PAGE_SIZE;
6015 batch /= 4; /* We effectively *= 4 below */
6020 * Clamp the batch to a 2^n - 1 value. Having a power
6021 * of 2 value was found to be more likely to have
6022 * suboptimal cache aliasing properties in some cases.
6024 * For example if 2 tasks are alternately allocating
6025 * batches of pages, one task can end up with a lot
6026 * of pages of one half of the possible page colors
6027 * and the other with pages of the other colors.
6029 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6034 /* The deferral and batching of frees should be suppressed under NOMMU
6037 * The problem is that NOMMU needs to be able to allocate large chunks
6038 * of contiguous memory as there's no hardware page translation to
6039 * assemble apparent contiguous memory from discontiguous pages.
6041 * Queueing large contiguous runs of pages for batching, however,
6042 * causes the pages to actually be freed in smaller chunks. As there
6043 * can be a significant delay between the individual batches being
6044 * recycled, this leads to the once large chunks of space being
6045 * fragmented and becoming unavailable for high-order allocations.
6052 * pcp->high and pcp->batch values are related and dependent on one another:
6053 * ->batch must never be higher then ->high.
6054 * The following function updates them in a safe manner without read side
6057 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6058 * those fields changing asynchronously (acording the the above rule).
6060 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6061 * outside of boot time (or some other assurance that no concurrent updaters
6064 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6065 unsigned long batch)
6067 /* start with a fail safe value for batch */
6071 /* Update high, then batch, in order */
6078 /* a companion to pageset_set_high() */
6079 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6081 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6084 static void pageset_init(struct per_cpu_pageset *p)
6086 struct per_cpu_pages *pcp;
6089 memset(p, 0, sizeof(*p));
6092 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6093 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6096 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6099 pageset_set_batch(p, batch);
6103 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6104 * to the value high for the pageset p.
6106 static void pageset_set_high(struct per_cpu_pageset *p,
6109 unsigned long batch = max(1UL, high / 4);
6110 if ((high / 4) > (PAGE_SHIFT * 8))
6111 batch = PAGE_SHIFT * 8;
6113 pageset_update(&p->pcp, high, batch);
6116 static void pageset_set_high_and_batch(struct zone *zone,
6117 struct per_cpu_pageset *pcp)
6119 if (percpu_pagelist_fraction)
6120 pageset_set_high(pcp,
6121 (zone_managed_pages(zone) /
6122 percpu_pagelist_fraction));
6124 pageset_set_batch(pcp, zone_batchsize(zone));
6127 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6129 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6132 pageset_set_high_and_batch(zone, pcp);
6135 void __meminit setup_zone_pageset(struct zone *zone)
6138 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6139 for_each_possible_cpu(cpu)
6140 zone_pageset_init(zone, cpu);
6144 * Allocate per cpu pagesets and initialize them.
6145 * Before this call only boot pagesets were available.
6147 void __init setup_per_cpu_pageset(void)
6149 struct pglist_data *pgdat;
6152 for_each_populated_zone(zone)
6153 setup_zone_pageset(zone);
6155 for_each_online_pgdat(pgdat)
6156 pgdat->per_cpu_nodestats =
6157 alloc_percpu(struct per_cpu_nodestat);
6160 static __meminit void zone_pcp_init(struct zone *zone)
6163 * per cpu subsystem is not up at this point. The following code
6164 * relies on the ability of the linker to provide the
6165 * offset of a (static) per cpu variable into the per cpu area.
6167 zone->pageset = &boot_pageset;
6169 if (populated_zone(zone))
6170 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6171 zone->name, zone->present_pages,
6172 zone_batchsize(zone));
6175 void __meminit init_currently_empty_zone(struct zone *zone,
6176 unsigned long zone_start_pfn,
6179 struct pglist_data *pgdat = zone->zone_pgdat;
6180 int zone_idx = zone_idx(zone) + 1;
6182 if (zone_idx > pgdat->nr_zones)
6183 pgdat->nr_zones = zone_idx;
6185 zone->zone_start_pfn = zone_start_pfn;
6187 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6188 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6190 (unsigned long)zone_idx(zone),
6191 zone_start_pfn, (zone_start_pfn + size));
6193 zone_init_free_lists(zone);
6194 zone->initialized = 1;
6197 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6198 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
6201 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
6203 int __meminit __early_pfn_to_nid(unsigned long pfn,
6204 struct mminit_pfnnid_cache *state)
6206 unsigned long start_pfn, end_pfn;
6209 if (state->last_start <= pfn && pfn < state->last_end)
6210 return state->last_nid;
6212 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
6213 if (nid != NUMA_NO_NODE) {
6214 state->last_start = start_pfn;
6215 state->last_end = end_pfn;
6216 state->last_nid = nid;
6221 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
6224 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
6225 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
6226 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
6228 * If an architecture guarantees that all ranges registered contain no holes
6229 * and may be freed, this this function may be used instead of calling
6230 * memblock_free_early_nid() manually.
6232 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
6234 unsigned long start_pfn, end_pfn;
6237 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
6238 start_pfn = min(start_pfn, max_low_pfn);
6239 end_pfn = min(end_pfn, max_low_pfn);
6241 if (start_pfn < end_pfn)
6242 memblock_free_early_nid(PFN_PHYS(start_pfn),
6243 (end_pfn - start_pfn) << PAGE_SHIFT,
6249 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6250 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6252 * If an architecture guarantees that all ranges registered contain no holes and may
6253 * be freed, this function may be used instead of calling memory_present() manually.
6255 void __init sparse_memory_present_with_active_regions(int nid)
6257 unsigned long start_pfn, end_pfn;
6260 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
6261 memory_present(this_nid, start_pfn, end_pfn);
6265 * get_pfn_range_for_nid - Return the start and end page frames for a node
6266 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6267 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6268 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6270 * It returns the start and end page frame of a node based on information
6271 * provided by memblock_set_node(). If called for a node
6272 * with no available memory, a warning is printed and the start and end
6275 void __init get_pfn_range_for_nid(unsigned int nid,
6276 unsigned long *start_pfn, unsigned long *end_pfn)
6278 unsigned long this_start_pfn, this_end_pfn;
6284 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6285 *start_pfn = min(*start_pfn, this_start_pfn);
6286 *end_pfn = max(*end_pfn, this_end_pfn);
6289 if (*start_pfn == -1UL)
6294 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6295 * assumption is made that zones within a node are ordered in monotonic
6296 * increasing memory addresses so that the "highest" populated zone is used
6298 static void __init find_usable_zone_for_movable(void)
6301 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6302 if (zone_index == ZONE_MOVABLE)
6305 if (arch_zone_highest_possible_pfn[zone_index] >
6306 arch_zone_lowest_possible_pfn[zone_index])
6310 VM_BUG_ON(zone_index == -1);
6311 movable_zone = zone_index;
6315 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6316 * because it is sized independent of architecture. Unlike the other zones,
6317 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6318 * in each node depending on the size of each node and how evenly kernelcore
6319 * is distributed. This helper function adjusts the zone ranges
6320 * provided by the architecture for a given node by using the end of the
6321 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6322 * zones within a node are in order of monotonic increases memory addresses
6324 static void __init adjust_zone_range_for_zone_movable(int nid,
6325 unsigned long zone_type,
6326 unsigned long node_start_pfn,
6327 unsigned long node_end_pfn,
6328 unsigned long *zone_start_pfn,
6329 unsigned long *zone_end_pfn)
6331 /* Only adjust if ZONE_MOVABLE is on this node */
6332 if (zone_movable_pfn[nid]) {
6333 /* Size ZONE_MOVABLE */
6334 if (zone_type == ZONE_MOVABLE) {
6335 *zone_start_pfn = zone_movable_pfn[nid];
6336 *zone_end_pfn = min(node_end_pfn,
6337 arch_zone_highest_possible_pfn[movable_zone]);
6339 /* Adjust for ZONE_MOVABLE starting within this range */
6340 } else if (!mirrored_kernelcore &&
6341 *zone_start_pfn < zone_movable_pfn[nid] &&
6342 *zone_end_pfn > zone_movable_pfn[nid]) {
6343 *zone_end_pfn = zone_movable_pfn[nid];
6345 /* Check if this whole range is within ZONE_MOVABLE */
6346 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6347 *zone_start_pfn = *zone_end_pfn;
6352 * Return the number of pages a zone spans in a node, including holes
6353 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6355 static unsigned long __init zone_spanned_pages_in_node(int nid,
6356 unsigned long zone_type,
6357 unsigned long node_start_pfn,
6358 unsigned long node_end_pfn,
6359 unsigned long *zone_start_pfn,
6360 unsigned long *zone_end_pfn,
6361 unsigned long *ignored)
6363 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6364 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6365 /* When hotadd a new node from cpu_up(), the node should be empty */
6366 if (!node_start_pfn && !node_end_pfn)
6369 /* Get the start and end of the zone */
6370 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6371 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6372 adjust_zone_range_for_zone_movable(nid, zone_type,
6373 node_start_pfn, node_end_pfn,
6374 zone_start_pfn, zone_end_pfn);
6376 /* Check that this node has pages within the zone's required range */
6377 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6380 /* Move the zone boundaries inside the node if necessary */
6381 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6382 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6384 /* Return the spanned pages */
6385 return *zone_end_pfn - *zone_start_pfn;
6389 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6390 * then all holes in the requested range will be accounted for.
6392 unsigned long __init __absent_pages_in_range(int nid,
6393 unsigned long range_start_pfn,
6394 unsigned long range_end_pfn)
6396 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6397 unsigned long start_pfn, end_pfn;
6400 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6401 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6402 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6403 nr_absent -= end_pfn - start_pfn;
6409 * absent_pages_in_range - Return number of page frames in holes within a range
6410 * @start_pfn: The start PFN to start searching for holes
6411 * @end_pfn: The end PFN to stop searching for holes
6413 * Return: the number of pages frames in memory holes within a range.
6415 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6416 unsigned long end_pfn)
6418 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6421 /* Return the number of page frames in holes in a zone on a node */
6422 static unsigned long __init zone_absent_pages_in_node(int nid,
6423 unsigned long zone_type,
6424 unsigned long node_start_pfn,
6425 unsigned long node_end_pfn,
6426 unsigned long *ignored)
6428 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6429 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6430 unsigned long zone_start_pfn, zone_end_pfn;
6431 unsigned long nr_absent;
6433 /* When hotadd a new node from cpu_up(), the node should be empty */
6434 if (!node_start_pfn && !node_end_pfn)
6437 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6438 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6440 adjust_zone_range_for_zone_movable(nid, zone_type,
6441 node_start_pfn, node_end_pfn,
6442 &zone_start_pfn, &zone_end_pfn);
6443 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6446 * ZONE_MOVABLE handling.
6447 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6450 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6451 unsigned long start_pfn, end_pfn;
6452 struct memblock_region *r;
6454 for_each_memblock(memory, r) {
6455 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6456 zone_start_pfn, zone_end_pfn);
6457 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6458 zone_start_pfn, zone_end_pfn);
6460 if (zone_type == ZONE_MOVABLE &&
6461 memblock_is_mirror(r))
6462 nr_absent += end_pfn - start_pfn;
6464 if (zone_type == ZONE_NORMAL &&
6465 !memblock_is_mirror(r))
6466 nr_absent += end_pfn - start_pfn;
6473 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6474 static inline unsigned long __init zone_spanned_pages_in_node(int nid,
6475 unsigned long zone_type,
6476 unsigned long node_start_pfn,
6477 unsigned long node_end_pfn,
6478 unsigned long *zone_start_pfn,
6479 unsigned long *zone_end_pfn,
6480 unsigned long *zones_size)
6484 *zone_start_pfn = node_start_pfn;
6485 for (zone = 0; zone < zone_type; zone++)
6486 *zone_start_pfn += zones_size[zone];
6488 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6490 return zones_size[zone_type];
6493 static inline unsigned long __init zone_absent_pages_in_node(int nid,
6494 unsigned long zone_type,
6495 unsigned long node_start_pfn,
6496 unsigned long node_end_pfn,
6497 unsigned long *zholes_size)
6502 return zholes_size[zone_type];
6505 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6507 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6508 unsigned long node_start_pfn,
6509 unsigned long node_end_pfn,
6510 unsigned long *zones_size,
6511 unsigned long *zholes_size)
6513 unsigned long realtotalpages = 0, totalpages = 0;
6516 for (i = 0; i < MAX_NR_ZONES; i++) {
6517 struct zone *zone = pgdat->node_zones + i;
6518 unsigned long zone_start_pfn, zone_end_pfn;
6519 unsigned long size, real_size;
6521 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6527 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6528 node_start_pfn, node_end_pfn,
6531 zone->zone_start_pfn = zone_start_pfn;
6533 zone->zone_start_pfn = 0;
6534 zone->spanned_pages = size;
6535 zone->present_pages = real_size;
6538 realtotalpages += real_size;
6541 pgdat->node_spanned_pages = totalpages;
6542 pgdat->node_present_pages = realtotalpages;
6543 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6547 #ifndef CONFIG_SPARSEMEM
6549 * Calculate the size of the zone->blockflags rounded to an unsigned long
6550 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6551 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6552 * round what is now in bits to nearest long in bits, then return it in
6555 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6557 unsigned long usemapsize;
6559 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6560 usemapsize = roundup(zonesize, pageblock_nr_pages);
6561 usemapsize = usemapsize >> pageblock_order;
6562 usemapsize *= NR_PAGEBLOCK_BITS;
6563 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6565 return usemapsize / 8;
6568 static void __ref setup_usemap(struct pglist_data *pgdat,
6570 unsigned long zone_start_pfn,
6571 unsigned long zonesize)
6573 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6574 zone->pageblock_flags = NULL;
6576 zone->pageblock_flags =
6577 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6579 if (!zone->pageblock_flags)
6580 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6581 usemapsize, zone->name, pgdat->node_id);
6585 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6586 unsigned long zone_start_pfn, unsigned long zonesize) {}
6587 #endif /* CONFIG_SPARSEMEM */
6589 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6591 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6592 void __init set_pageblock_order(void)
6596 /* Check that pageblock_nr_pages has not already been setup */
6597 if (pageblock_order)
6600 if (HPAGE_SHIFT > PAGE_SHIFT)
6601 order = HUGETLB_PAGE_ORDER;
6603 order = MAX_ORDER - 1;
6606 * Assume the largest contiguous order of interest is a huge page.
6607 * This value may be variable depending on boot parameters on IA64 and
6610 pageblock_order = order;
6612 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6615 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6616 * is unused as pageblock_order is set at compile-time. See
6617 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6620 void __init set_pageblock_order(void)
6624 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6626 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6627 unsigned long present_pages)
6629 unsigned long pages = spanned_pages;
6632 * Provide a more accurate estimation if there are holes within
6633 * the zone and SPARSEMEM is in use. If there are holes within the
6634 * zone, each populated memory region may cost us one or two extra
6635 * memmap pages due to alignment because memmap pages for each
6636 * populated regions may not be naturally aligned on page boundary.
6637 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6639 if (spanned_pages > present_pages + (present_pages >> 4) &&
6640 IS_ENABLED(CONFIG_SPARSEMEM))
6641 pages = present_pages;
6643 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6646 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6647 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6649 spin_lock_init(&pgdat->split_queue_lock);
6650 INIT_LIST_HEAD(&pgdat->split_queue);
6651 pgdat->split_queue_len = 0;
6654 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6657 #ifdef CONFIG_COMPACTION
6658 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6660 init_waitqueue_head(&pgdat->kcompactd_wait);
6663 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6666 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6668 pgdat_resize_init(pgdat);
6670 pgdat_init_split_queue(pgdat);
6671 pgdat_init_kcompactd(pgdat);
6673 init_waitqueue_head(&pgdat->kswapd_wait);
6674 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6676 pgdat_page_ext_init(pgdat);
6677 spin_lock_init(&pgdat->lru_lock);
6678 lruvec_init(node_lruvec(pgdat));
6681 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6682 unsigned long remaining_pages)
6684 atomic_long_set(&zone->managed_pages, remaining_pages);
6685 zone_set_nid(zone, nid);
6686 zone->name = zone_names[idx];
6687 zone->zone_pgdat = NODE_DATA(nid);
6688 spin_lock_init(&zone->lock);
6689 zone_seqlock_init(zone);
6690 zone_pcp_init(zone);
6694 * Set up the zone data structures
6695 * - init pgdat internals
6696 * - init all zones belonging to this node
6698 * NOTE: this function is only called during memory hotplug
6700 #ifdef CONFIG_MEMORY_HOTPLUG
6701 void __ref free_area_init_core_hotplug(int nid)
6704 pg_data_t *pgdat = NODE_DATA(nid);
6706 pgdat_init_internals(pgdat);
6707 for (z = 0; z < MAX_NR_ZONES; z++)
6708 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6713 * Set up the zone data structures:
6714 * - mark all pages reserved
6715 * - mark all memory queues empty
6716 * - clear the memory bitmaps
6718 * NOTE: pgdat should get zeroed by caller.
6719 * NOTE: this function is only called during early init.
6721 static void __init free_area_init_core(struct pglist_data *pgdat)
6724 int nid = pgdat->node_id;
6726 pgdat_init_internals(pgdat);
6727 pgdat->per_cpu_nodestats = &boot_nodestats;
6729 for (j = 0; j < MAX_NR_ZONES; j++) {
6730 struct zone *zone = pgdat->node_zones + j;
6731 unsigned long size, freesize, memmap_pages;
6732 unsigned long zone_start_pfn = zone->zone_start_pfn;
6734 size = zone->spanned_pages;
6735 freesize = zone->present_pages;
6738 * Adjust freesize so that it accounts for how much memory
6739 * is used by this zone for memmap. This affects the watermark
6740 * and per-cpu initialisations
6742 memmap_pages = calc_memmap_size(size, freesize);
6743 if (!is_highmem_idx(j)) {
6744 if (freesize >= memmap_pages) {
6745 freesize -= memmap_pages;
6748 " %s zone: %lu pages used for memmap\n",
6749 zone_names[j], memmap_pages);
6751 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6752 zone_names[j], memmap_pages, freesize);
6755 /* Account for reserved pages */
6756 if (j == 0 && freesize > dma_reserve) {
6757 freesize -= dma_reserve;
6758 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6759 zone_names[0], dma_reserve);
6762 if (!is_highmem_idx(j))
6763 nr_kernel_pages += freesize;
6764 /* Charge for highmem memmap if there are enough kernel pages */
6765 else if (nr_kernel_pages > memmap_pages * 2)
6766 nr_kernel_pages -= memmap_pages;
6767 nr_all_pages += freesize;
6770 * Set an approximate value for lowmem here, it will be adjusted
6771 * when the bootmem allocator frees pages into the buddy system.
6772 * And all highmem pages will be managed by the buddy system.
6774 zone_init_internals(zone, j, nid, freesize);
6779 set_pageblock_order();
6780 setup_usemap(pgdat, zone, zone_start_pfn, size);
6781 init_currently_empty_zone(zone, zone_start_pfn, size);
6782 memmap_init(size, nid, j, zone_start_pfn);
6786 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6787 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6789 unsigned long __maybe_unused start = 0;
6790 unsigned long __maybe_unused offset = 0;
6792 /* Skip empty nodes */
6793 if (!pgdat->node_spanned_pages)
6796 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6797 offset = pgdat->node_start_pfn - start;
6798 /* ia64 gets its own node_mem_map, before this, without bootmem */
6799 if (!pgdat->node_mem_map) {
6800 unsigned long size, end;
6804 * The zone's endpoints aren't required to be MAX_ORDER
6805 * aligned but the node_mem_map endpoints must be in order
6806 * for the buddy allocator to function correctly.
6808 end = pgdat_end_pfn(pgdat);
6809 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6810 size = (end - start) * sizeof(struct page);
6811 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6814 panic("Failed to allocate %ld bytes for node %d memory map\n",
6815 size, pgdat->node_id);
6816 pgdat->node_mem_map = map + offset;
6818 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6819 __func__, pgdat->node_id, (unsigned long)pgdat,
6820 (unsigned long)pgdat->node_mem_map);
6821 #ifndef CONFIG_NEED_MULTIPLE_NODES
6823 * With no DISCONTIG, the global mem_map is just set as node 0's
6825 if (pgdat == NODE_DATA(0)) {
6826 mem_map = NODE_DATA(0)->node_mem_map;
6827 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6828 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6830 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6835 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6836 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6838 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6839 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6841 pgdat->first_deferred_pfn = ULONG_MAX;
6844 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6847 void __init free_area_init_node(int nid, unsigned long *zones_size,
6848 unsigned long node_start_pfn,
6849 unsigned long *zholes_size)
6851 pg_data_t *pgdat = NODE_DATA(nid);
6852 unsigned long start_pfn = 0;
6853 unsigned long end_pfn = 0;
6855 /* pg_data_t should be reset to zero when it's allocated */
6856 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6858 pgdat->node_id = nid;
6859 pgdat->node_start_pfn = node_start_pfn;
6860 pgdat->per_cpu_nodestats = NULL;
6861 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6862 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6863 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6864 (u64)start_pfn << PAGE_SHIFT,
6865 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6867 start_pfn = node_start_pfn;
6869 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6870 zones_size, zholes_size);
6872 alloc_node_mem_map(pgdat);
6873 pgdat_set_deferred_range(pgdat);
6875 free_area_init_core(pgdat);
6878 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6880 * Zero all valid struct pages in range [spfn, epfn), return number of struct
6883 static u64 zero_pfn_range(unsigned long spfn, unsigned long epfn)
6888 for (pfn = spfn; pfn < epfn; pfn++) {
6889 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6890 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6891 + pageblock_nr_pages - 1;
6894 mm_zero_struct_page(pfn_to_page(pfn));
6902 * Only struct pages that are backed by physical memory are zeroed and
6903 * initialized by going through __init_single_page(). But, there are some
6904 * struct pages which are reserved in memblock allocator and their fields
6905 * may be accessed (for example page_to_pfn() on some configuration accesses
6906 * flags). We must explicitly zero those struct pages.
6908 * This function also addresses a similar issue where struct pages are left
6909 * uninitialized because the physical address range is not covered by
6910 * memblock.memory or memblock.reserved. That could happen when memblock
6911 * layout is manually configured via memmap=.
6913 void __init zero_resv_unavail(void)
6915 phys_addr_t start, end;
6917 phys_addr_t next = 0;
6920 * Loop through unavailable ranges not covered by memblock.memory.
6923 for_each_mem_range(i, &memblock.memory, NULL,
6924 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
6926 pgcnt += zero_pfn_range(PFN_DOWN(next), PFN_UP(start));
6929 pgcnt += zero_pfn_range(PFN_DOWN(next), max_pfn);
6932 * Struct pages that do not have backing memory. This could be because
6933 * firmware is using some of this memory, or for some other reasons.
6936 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
6938 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
6940 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6942 #if MAX_NUMNODES > 1
6944 * Figure out the number of possible node ids.
6946 void __init setup_nr_node_ids(void)
6948 unsigned int highest;
6950 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6951 nr_node_ids = highest + 1;
6956 * node_map_pfn_alignment - determine the maximum internode alignment
6958 * This function should be called after node map is populated and sorted.
6959 * It calculates the maximum power of two alignment which can distinguish
6962 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6963 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6964 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6965 * shifted, 1GiB is enough and this function will indicate so.
6967 * This is used to test whether pfn -> nid mapping of the chosen memory
6968 * model has fine enough granularity to avoid incorrect mapping for the
6969 * populated node map.
6971 * Return: the determined alignment in pfn's. 0 if there is no alignment
6972 * requirement (single node).
6974 unsigned long __init node_map_pfn_alignment(void)
6976 unsigned long accl_mask = 0, last_end = 0;
6977 unsigned long start, end, mask;
6978 int last_nid = NUMA_NO_NODE;
6981 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6982 if (!start || last_nid < 0 || last_nid == nid) {
6989 * Start with a mask granular enough to pin-point to the
6990 * start pfn and tick off bits one-by-one until it becomes
6991 * too coarse to separate the current node from the last.
6993 mask = ~((1 << __ffs(start)) - 1);
6994 while (mask && last_end <= (start & (mask << 1)))
6997 /* accumulate all internode masks */
7001 /* convert mask to number of pages */
7002 return ~accl_mask + 1;
7005 /* Find the lowest pfn for a node */
7006 static unsigned long __init find_min_pfn_for_node(int nid)
7008 unsigned long min_pfn = ULONG_MAX;
7009 unsigned long start_pfn;
7012 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
7013 min_pfn = min(min_pfn, start_pfn);
7015 if (min_pfn == ULONG_MAX) {
7016 pr_warn("Could not find start_pfn for node %d\n", nid);
7024 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7026 * Return: the minimum PFN based on information provided via
7027 * memblock_set_node().
7029 unsigned long __init find_min_pfn_with_active_regions(void)
7031 return find_min_pfn_for_node(MAX_NUMNODES);
7035 * early_calculate_totalpages()
7036 * Sum pages in active regions for movable zone.
7037 * Populate N_MEMORY for calculating usable_nodes.
7039 static unsigned long __init early_calculate_totalpages(void)
7041 unsigned long totalpages = 0;
7042 unsigned long start_pfn, end_pfn;
7045 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7046 unsigned long pages = end_pfn - start_pfn;
7048 totalpages += pages;
7050 node_set_state(nid, N_MEMORY);
7056 * Find the PFN the Movable zone begins in each node. Kernel memory
7057 * is spread evenly between nodes as long as the nodes have enough
7058 * memory. When they don't, some nodes will have more kernelcore than
7061 static void __init find_zone_movable_pfns_for_nodes(void)
7064 unsigned long usable_startpfn;
7065 unsigned long kernelcore_node, kernelcore_remaining;
7066 /* save the state before borrow the nodemask */
7067 nodemask_t saved_node_state = node_states[N_MEMORY];
7068 unsigned long totalpages = early_calculate_totalpages();
7069 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7070 struct memblock_region *r;
7072 /* Need to find movable_zone earlier when movable_node is specified. */
7073 find_usable_zone_for_movable();
7076 * If movable_node is specified, ignore kernelcore and movablecore
7079 if (movable_node_is_enabled()) {
7080 for_each_memblock(memory, r) {
7081 if (!memblock_is_hotpluggable(r))
7086 usable_startpfn = PFN_DOWN(r->base);
7087 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7088 min(usable_startpfn, zone_movable_pfn[nid]) :
7096 * If kernelcore=mirror is specified, ignore movablecore option
7098 if (mirrored_kernelcore) {
7099 bool mem_below_4gb_not_mirrored = false;
7101 for_each_memblock(memory, r) {
7102 if (memblock_is_mirror(r))
7107 usable_startpfn = memblock_region_memory_base_pfn(r);
7109 if (usable_startpfn < 0x100000) {
7110 mem_below_4gb_not_mirrored = true;
7114 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7115 min(usable_startpfn, zone_movable_pfn[nid]) :
7119 if (mem_below_4gb_not_mirrored)
7120 pr_warn("This configuration results in unmirrored kernel memory.");
7126 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7127 * amount of necessary memory.
7129 if (required_kernelcore_percent)
7130 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7132 if (required_movablecore_percent)
7133 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7137 * If movablecore= was specified, calculate what size of
7138 * kernelcore that corresponds so that memory usable for
7139 * any allocation type is evenly spread. If both kernelcore
7140 * and movablecore are specified, then the value of kernelcore
7141 * will be used for required_kernelcore if it's greater than
7142 * what movablecore would have allowed.
7144 if (required_movablecore) {
7145 unsigned long corepages;
7148 * Round-up so that ZONE_MOVABLE is at least as large as what
7149 * was requested by the user
7151 required_movablecore =
7152 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7153 required_movablecore = min(totalpages, required_movablecore);
7154 corepages = totalpages - required_movablecore;
7156 required_kernelcore = max(required_kernelcore, corepages);
7160 * If kernelcore was not specified or kernelcore size is larger
7161 * than totalpages, there is no ZONE_MOVABLE.
7163 if (!required_kernelcore || required_kernelcore >= totalpages)
7166 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7167 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7170 /* Spread kernelcore memory as evenly as possible throughout nodes */
7171 kernelcore_node = required_kernelcore / usable_nodes;
7172 for_each_node_state(nid, N_MEMORY) {
7173 unsigned long start_pfn, end_pfn;
7176 * Recalculate kernelcore_node if the division per node
7177 * now exceeds what is necessary to satisfy the requested
7178 * amount of memory for the kernel
7180 if (required_kernelcore < kernelcore_node)
7181 kernelcore_node = required_kernelcore / usable_nodes;
7184 * As the map is walked, we track how much memory is usable
7185 * by the kernel using kernelcore_remaining. When it is
7186 * 0, the rest of the node is usable by ZONE_MOVABLE
7188 kernelcore_remaining = kernelcore_node;
7190 /* Go through each range of PFNs within this node */
7191 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7192 unsigned long size_pages;
7194 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7195 if (start_pfn >= end_pfn)
7198 /* Account for what is only usable for kernelcore */
7199 if (start_pfn < usable_startpfn) {
7200 unsigned long kernel_pages;
7201 kernel_pages = min(end_pfn, usable_startpfn)
7204 kernelcore_remaining -= min(kernel_pages,
7205 kernelcore_remaining);
7206 required_kernelcore -= min(kernel_pages,
7207 required_kernelcore);
7209 /* Continue if range is now fully accounted */
7210 if (end_pfn <= usable_startpfn) {
7213 * Push zone_movable_pfn to the end so
7214 * that if we have to rebalance
7215 * kernelcore across nodes, we will
7216 * not double account here
7218 zone_movable_pfn[nid] = end_pfn;
7221 start_pfn = usable_startpfn;
7225 * The usable PFN range for ZONE_MOVABLE is from
7226 * start_pfn->end_pfn. Calculate size_pages as the
7227 * number of pages used as kernelcore
7229 size_pages = end_pfn - start_pfn;
7230 if (size_pages > kernelcore_remaining)
7231 size_pages = kernelcore_remaining;
7232 zone_movable_pfn[nid] = start_pfn + size_pages;
7235 * Some kernelcore has been met, update counts and
7236 * break if the kernelcore for this node has been
7239 required_kernelcore -= min(required_kernelcore,
7241 kernelcore_remaining -= size_pages;
7242 if (!kernelcore_remaining)
7248 * If there is still required_kernelcore, we do another pass with one
7249 * less node in the count. This will push zone_movable_pfn[nid] further
7250 * along on the nodes that still have memory until kernelcore is
7254 if (usable_nodes && required_kernelcore > usable_nodes)
7258 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7259 for (nid = 0; nid < MAX_NUMNODES; nid++)
7260 zone_movable_pfn[nid] =
7261 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7264 /* restore the node_state */
7265 node_states[N_MEMORY] = saved_node_state;
7268 /* Any regular or high memory on that node ? */
7269 static void check_for_memory(pg_data_t *pgdat, int nid)
7271 enum zone_type zone_type;
7273 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7274 struct zone *zone = &pgdat->node_zones[zone_type];
7275 if (populated_zone(zone)) {
7276 if (IS_ENABLED(CONFIG_HIGHMEM))
7277 node_set_state(nid, N_HIGH_MEMORY);
7278 if (zone_type <= ZONE_NORMAL)
7279 node_set_state(nid, N_NORMAL_MEMORY);
7286 * free_area_init_nodes - Initialise all pg_data_t and zone data
7287 * @max_zone_pfn: an array of max PFNs for each zone
7289 * This will call free_area_init_node() for each active node in the system.
7290 * Using the page ranges provided by memblock_set_node(), the size of each
7291 * zone in each node and their holes is calculated. If the maximum PFN
7292 * between two adjacent zones match, it is assumed that the zone is empty.
7293 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7294 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7295 * starts where the previous one ended. For example, ZONE_DMA32 starts
7296 * at arch_max_dma_pfn.
7298 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
7300 unsigned long start_pfn, end_pfn;
7303 /* Record where the zone boundaries are */
7304 memset(arch_zone_lowest_possible_pfn, 0,
7305 sizeof(arch_zone_lowest_possible_pfn));
7306 memset(arch_zone_highest_possible_pfn, 0,
7307 sizeof(arch_zone_highest_possible_pfn));
7309 start_pfn = find_min_pfn_with_active_regions();
7311 for (i = 0; i < MAX_NR_ZONES; i++) {
7312 if (i == ZONE_MOVABLE)
7315 end_pfn = max(max_zone_pfn[i], start_pfn);
7316 arch_zone_lowest_possible_pfn[i] = start_pfn;
7317 arch_zone_highest_possible_pfn[i] = end_pfn;
7319 start_pfn = end_pfn;
7322 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7323 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7324 find_zone_movable_pfns_for_nodes();
7326 /* Print out the zone ranges */
7327 pr_info("Zone ranges:\n");
7328 for (i = 0; i < MAX_NR_ZONES; i++) {
7329 if (i == ZONE_MOVABLE)
7331 pr_info(" %-8s ", zone_names[i]);
7332 if (arch_zone_lowest_possible_pfn[i] ==
7333 arch_zone_highest_possible_pfn[i])
7336 pr_cont("[mem %#018Lx-%#018Lx]\n",
7337 (u64)arch_zone_lowest_possible_pfn[i]
7339 ((u64)arch_zone_highest_possible_pfn[i]
7340 << PAGE_SHIFT) - 1);
7343 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7344 pr_info("Movable zone start for each node\n");
7345 for (i = 0; i < MAX_NUMNODES; i++) {
7346 if (zone_movable_pfn[i])
7347 pr_info(" Node %d: %#018Lx\n", i,
7348 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7352 * Print out the early node map, and initialize the
7353 * subsection-map relative to active online memory ranges to
7354 * enable future "sub-section" extensions of the memory map.
7356 pr_info("Early memory node ranges\n");
7357 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7358 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7359 (u64)start_pfn << PAGE_SHIFT,
7360 ((u64)end_pfn << PAGE_SHIFT) - 1);
7361 subsection_map_init(start_pfn, end_pfn - start_pfn);
7364 /* Initialise every node */
7365 mminit_verify_pageflags_layout();
7366 setup_nr_node_ids();
7367 zero_resv_unavail();
7368 for_each_online_node(nid) {
7369 pg_data_t *pgdat = NODE_DATA(nid);
7370 free_area_init_node(nid, NULL,
7371 find_min_pfn_for_node(nid), NULL);
7373 /* Any memory on that node */
7374 if (pgdat->node_present_pages)
7375 node_set_state(nid, N_MEMORY);
7376 check_for_memory(pgdat, nid);
7380 static int __init cmdline_parse_core(char *p, unsigned long *core,
7381 unsigned long *percent)
7383 unsigned long long coremem;
7389 /* Value may be a percentage of total memory, otherwise bytes */
7390 coremem = simple_strtoull(p, &endptr, 0);
7391 if (*endptr == '%') {
7392 /* Paranoid check for percent values greater than 100 */
7393 WARN_ON(coremem > 100);
7397 coremem = memparse(p, &p);
7398 /* Paranoid check that UL is enough for the coremem value */
7399 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7401 *core = coremem >> PAGE_SHIFT;
7408 * kernelcore=size sets the amount of memory for use for allocations that
7409 * cannot be reclaimed or migrated.
7411 static int __init cmdline_parse_kernelcore(char *p)
7413 /* parse kernelcore=mirror */
7414 if (parse_option_str(p, "mirror")) {
7415 mirrored_kernelcore = true;
7419 return cmdline_parse_core(p, &required_kernelcore,
7420 &required_kernelcore_percent);
7424 * movablecore=size sets the amount of memory for use for allocations that
7425 * can be reclaimed or migrated.
7427 static int __init cmdline_parse_movablecore(char *p)
7429 return cmdline_parse_core(p, &required_movablecore,
7430 &required_movablecore_percent);
7433 early_param("kernelcore", cmdline_parse_kernelcore);
7434 early_param("movablecore", cmdline_parse_movablecore);
7436 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
7438 void adjust_managed_page_count(struct page *page, long count)
7440 atomic_long_add(count, &page_zone(page)->managed_pages);
7441 totalram_pages_add(count);
7442 #ifdef CONFIG_HIGHMEM
7443 if (PageHighMem(page))
7444 totalhigh_pages_add(count);
7447 EXPORT_SYMBOL(adjust_managed_page_count);
7449 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7452 unsigned long pages = 0;
7454 start = (void *)PAGE_ALIGN((unsigned long)start);
7455 end = (void *)((unsigned long)end & PAGE_MASK);
7456 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7457 struct page *page = virt_to_page(pos);
7458 void *direct_map_addr;
7461 * 'direct_map_addr' might be different from 'pos'
7462 * because some architectures' virt_to_page()
7463 * work with aliases. Getting the direct map
7464 * address ensures that we get a _writeable_
7465 * alias for the memset().
7467 direct_map_addr = page_address(page);
7468 if ((unsigned int)poison <= 0xFF)
7469 memset(direct_map_addr, poison, PAGE_SIZE);
7471 free_reserved_page(page);
7475 pr_info("Freeing %s memory: %ldK\n",
7476 s, pages << (PAGE_SHIFT - 10));
7481 #ifdef CONFIG_HIGHMEM
7482 void free_highmem_page(struct page *page)
7484 __free_reserved_page(page);
7485 totalram_pages_inc();
7486 atomic_long_inc(&page_zone(page)->managed_pages);
7487 totalhigh_pages_inc();
7492 void __init mem_init_print_info(const char *str)
7494 unsigned long physpages, codesize, datasize, rosize, bss_size;
7495 unsigned long init_code_size, init_data_size;
7497 physpages = get_num_physpages();
7498 codesize = _etext - _stext;
7499 datasize = _edata - _sdata;
7500 rosize = __end_rodata - __start_rodata;
7501 bss_size = __bss_stop - __bss_start;
7502 init_data_size = __init_end - __init_begin;
7503 init_code_size = _einittext - _sinittext;
7506 * Detect special cases and adjust section sizes accordingly:
7507 * 1) .init.* may be embedded into .data sections
7508 * 2) .init.text.* may be out of [__init_begin, __init_end],
7509 * please refer to arch/tile/kernel/vmlinux.lds.S.
7510 * 3) .rodata.* may be embedded into .text or .data sections.
7512 #define adj_init_size(start, end, size, pos, adj) \
7514 if (start <= pos && pos < end && size > adj) \
7518 adj_init_size(__init_begin, __init_end, init_data_size,
7519 _sinittext, init_code_size);
7520 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7521 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7522 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7523 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7525 #undef adj_init_size
7527 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7528 #ifdef CONFIG_HIGHMEM
7532 nr_free_pages() << (PAGE_SHIFT - 10),
7533 physpages << (PAGE_SHIFT - 10),
7534 codesize >> 10, datasize >> 10, rosize >> 10,
7535 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7536 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7537 totalcma_pages << (PAGE_SHIFT - 10),
7538 #ifdef CONFIG_HIGHMEM
7539 totalhigh_pages() << (PAGE_SHIFT - 10),
7541 str ? ", " : "", str ? str : "");
7545 * set_dma_reserve - set the specified number of pages reserved in the first zone
7546 * @new_dma_reserve: The number of pages to mark reserved
7548 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7549 * In the DMA zone, a significant percentage may be consumed by kernel image
7550 * and other unfreeable allocations which can skew the watermarks badly. This
7551 * function may optionally be used to account for unfreeable pages in the
7552 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7553 * smaller per-cpu batchsize.
7555 void __init set_dma_reserve(unsigned long new_dma_reserve)
7557 dma_reserve = new_dma_reserve;
7560 void __init free_area_init(unsigned long *zones_size)
7562 zero_resv_unavail();
7563 free_area_init_node(0, zones_size,
7564 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
7567 static int page_alloc_cpu_dead(unsigned int cpu)
7570 lru_add_drain_cpu(cpu);
7574 * Spill the event counters of the dead processor
7575 * into the current processors event counters.
7576 * This artificially elevates the count of the current
7579 vm_events_fold_cpu(cpu);
7582 * Zero the differential counters of the dead processor
7583 * so that the vm statistics are consistent.
7585 * This is only okay since the processor is dead and cannot
7586 * race with what we are doing.
7588 cpu_vm_stats_fold(cpu);
7593 int hashdist = HASHDIST_DEFAULT;
7595 static int __init set_hashdist(char *str)
7599 hashdist = simple_strtoul(str, &str, 0);
7602 __setup("hashdist=", set_hashdist);
7605 void __init page_alloc_init(void)
7610 if (num_node_state(N_MEMORY) == 1)
7614 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7615 "mm/page_alloc:dead", NULL,
7616 page_alloc_cpu_dead);
7621 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7622 * or min_free_kbytes changes.
7624 static void calculate_totalreserve_pages(void)
7626 struct pglist_data *pgdat;
7627 unsigned long reserve_pages = 0;
7628 enum zone_type i, j;
7630 for_each_online_pgdat(pgdat) {
7632 pgdat->totalreserve_pages = 0;
7634 for (i = 0; i < MAX_NR_ZONES; i++) {
7635 struct zone *zone = pgdat->node_zones + i;
7637 unsigned long managed_pages = zone_managed_pages(zone);
7639 /* Find valid and maximum lowmem_reserve in the zone */
7640 for (j = i; j < MAX_NR_ZONES; j++) {
7641 if (zone->lowmem_reserve[j] > max)
7642 max = zone->lowmem_reserve[j];
7645 /* we treat the high watermark as reserved pages. */
7646 max += high_wmark_pages(zone);
7648 if (max > managed_pages)
7649 max = managed_pages;
7651 pgdat->totalreserve_pages += max;
7653 reserve_pages += max;
7656 totalreserve_pages = reserve_pages;
7660 * setup_per_zone_lowmem_reserve - called whenever
7661 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7662 * has a correct pages reserved value, so an adequate number of
7663 * pages are left in the zone after a successful __alloc_pages().
7665 static void setup_per_zone_lowmem_reserve(void)
7667 struct pglist_data *pgdat;
7668 enum zone_type j, idx;
7670 for_each_online_pgdat(pgdat) {
7671 for (j = 0; j < MAX_NR_ZONES; j++) {
7672 struct zone *zone = pgdat->node_zones + j;
7673 unsigned long managed_pages = zone_managed_pages(zone);
7675 zone->lowmem_reserve[j] = 0;
7679 struct zone *lower_zone;
7682 lower_zone = pgdat->node_zones + idx;
7684 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7685 sysctl_lowmem_reserve_ratio[idx] = 0;
7686 lower_zone->lowmem_reserve[j] = 0;
7688 lower_zone->lowmem_reserve[j] =
7689 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7691 managed_pages += zone_managed_pages(lower_zone);
7696 /* update totalreserve_pages */
7697 calculate_totalreserve_pages();
7700 static void __setup_per_zone_wmarks(void)
7702 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7703 unsigned long lowmem_pages = 0;
7705 unsigned long flags;
7707 /* Calculate total number of !ZONE_HIGHMEM pages */
7708 for_each_zone(zone) {
7709 if (!is_highmem(zone))
7710 lowmem_pages += zone_managed_pages(zone);
7713 for_each_zone(zone) {
7716 spin_lock_irqsave(&zone->lock, flags);
7717 tmp = (u64)pages_min * zone_managed_pages(zone);
7718 do_div(tmp, lowmem_pages);
7719 if (is_highmem(zone)) {
7721 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7722 * need highmem pages, so cap pages_min to a small
7725 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7726 * deltas control async page reclaim, and so should
7727 * not be capped for highmem.
7729 unsigned long min_pages;
7731 min_pages = zone_managed_pages(zone) / 1024;
7732 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7733 zone->_watermark[WMARK_MIN] = min_pages;
7736 * If it's a lowmem zone, reserve a number of pages
7737 * proportionate to the zone's size.
7739 zone->_watermark[WMARK_MIN] = tmp;
7743 * Set the kswapd watermarks distance according to the
7744 * scale factor in proportion to available memory, but
7745 * ensure a minimum size on small systems.
7747 tmp = max_t(u64, tmp >> 2,
7748 mult_frac(zone_managed_pages(zone),
7749 watermark_scale_factor, 10000));
7751 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7752 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7753 zone->watermark_boost = 0;
7755 spin_unlock_irqrestore(&zone->lock, flags);
7758 /* update totalreserve_pages */
7759 calculate_totalreserve_pages();
7763 * setup_per_zone_wmarks - called when min_free_kbytes changes
7764 * or when memory is hot-{added|removed}
7766 * Ensures that the watermark[min,low,high] values for each zone are set
7767 * correctly with respect to min_free_kbytes.
7769 void setup_per_zone_wmarks(void)
7771 static DEFINE_SPINLOCK(lock);
7774 __setup_per_zone_wmarks();
7779 * Initialise min_free_kbytes.
7781 * For small machines we want it small (128k min). For large machines
7782 * we want it large (64MB max). But it is not linear, because network
7783 * bandwidth does not increase linearly with machine size. We use
7785 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7786 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7802 int __meminit init_per_zone_wmark_min(void)
7804 unsigned long lowmem_kbytes;
7805 int new_min_free_kbytes;
7807 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7808 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7810 if (new_min_free_kbytes > user_min_free_kbytes) {
7811 min_free_kbytes = new_min_free_kbytes;
7812 if (min_free_kbytes < 128)
7813 min_free_kbytes = 128;
7814 if (min_free_kbytes > 65536)
7815 min_free_kbytes = 65536;
7817 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7818 new_min_free_kbytes, user_min_free_kbytes);
7820 setup_per_zone_wmarks();
7821 refresh_zone_stat_thresholds();
7822 setup_per_zone_lowmem_reserve();
7825 setup_min_unmapped_ratio();
7826 setup_min_slab_ratio();
7831 core_initcall(init_per_zone_wmark_min)
7834 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7835 * that we can call two helper functions whenever min_free_kbytes
7838 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7839 void __user *buffer, size_t *length, loff_t *ppos)
7843 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7848 user_min_free_kbytes = min_free_kbytes;
7849 setup_per_zone_wmarks();
7854 int watermark_boost_factor_sysctl_handler(struct ctl_table *table, int write,
7855 void __user *buffer, size_t *length, loff_t *ppos)
7859 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7866 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7867 void __user *buffer, size_t *length, loff_t *ppos)
7871 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7876 setup_per_zone_wmarks();
7882 static void setup_min_unmapped_ratio(void)
7887 for_each_online_pgdat(pgdat)
7888 pgdat->min_unmapped_pages = 0;
7891 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7892 sysctl_min_unmapped_ratio) / 100;
7896 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7897 void __user *buffer, size_t *length, loff_t *ppos)
7901 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7905 setup_min_unmapped_ratio();
7910 static void setup_min_slab_ratio(void)
7915 for_each_online_pgdat(pgdat)
7916 pgdat->min_slab_pages = 0;
7919 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7920 sysctl_min_slab_ratio) / 100;
7923 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7924 void __user *buffer, size_t *length, loff_t *ppos)
7928 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7932 setup_min_slab_ratio();
7939 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7940 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7941 * whenever sysctl_lowmem_reserve_ratio changes.
7943 * The reserve ratio obviously has absolutely no relation with the
7944 * minimum watermarks. The lowmem reserve ratio can only make sense
7945 * if in function of the boot time zone sizes.
7947 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7948 void __user *buffer, size_t *length, loff_t *ppos)
7950 proc_dointvec_minmax(table, write, buffer, length, ppos);
7951 setup_per_zone_lowmem_reserve();
7956 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7957 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7958 * pagelist can have before it gets flushed back to buddy allocator.
7960 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7961 void __user *buffer, size_t *length, loff_t *ppos)
7964 int old_percpu_pagelist_fraction;
7967 mutex_lock(&pcp_batch_high_lock);
7968 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7970 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7971 if (!write || ret < 0)
7974 /* Sanity checking to avoid pcp imbalance */
7975 if (percpu_pagelist_fraction &&
7976 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7977 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7983 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7986 for_each_populated_zone(zone) {
7989 for_each_possible_cpu(cpu)
7990 pageset_set_high_and_batch(zone,
7991 per_cpu_ptr(zone->pageset, cpu));
7994 mutex_unlock(&pcp_batch_high_lock);
7998 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8000 * Returns the number of pages that arch has reserved but
8001 * is not known to alloc_large_system_hash().
8003 static unsigned long __init arch_reserved_kernel_pages(void)
8010 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8011 * machines. As memory size is increased the scale is also increased but at
8012 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8013 * quadruples the scale is increased by one, which means the size of hash table
8014 * only doubles, instead of quadrupling as well.
8015 * Because 32-bit systems cannot have large physical memory, where this scaling
8016 * makes sense, it is disabled on such platforms.
8018 #if __BITS_PER_LONG > 32
8019 #define ADAPT_SCALE_BASE (64ul << 30)
8020 #define ADAPT_SCALE_SHIFT 2
8021 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8025 * allocate a large system hash table from bootmem
8026 * - it is assumed that the hash table must contain an exact power-of-2
8027 * quantity of entries
8028 * - limit is the number of hash buckets, not the total allocation size
8030 void *__init alloc_large_system_hash(const char *tablename,
8031 unsigned long bucketsize,
8032 unsigned long numentries,
8035 unsigned int *_hash_shift,
8036 unsigned int *_hash_mask,
8037 unsigned long low_limit,
8038 unsigned long high_limit)
8040 unsigned long long max = high_limit;
8041 unsigned long log2qty, size;
8046 /* allow the kernel cmdline to have a say */
8048 /* round applicable memory size up to nearest megabyte */
8049 numentries = nr_kernel_pages;
8050 numentries -= arch_reserved_kernel_pages();
8052 /* It isn't necessary when PAGE_SIZE >= 1MB */
8053 if (PAGE_SHIFT < 20)
8054 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8056 #if __BITS_PER_LONG > 32
8058 unsigned long adapt;
8060 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8061 adapt <<= ADAPT_SCALE_SHIFT)
8066 /* limit to 1 bucket per 2^scale bytes of low memory */
8067 if (scale > PAGE_SHIFT)
8068 numentries >>= (scale - PAGE_SHIFT);
8070 numentries <<= (PAGE_SHIFT - scale);
8072 /* Make sure we've got at least a 0-order allocation.. */
8073 if (unlikely(flags & HASH_SMALL)) {
8074 /* Makes no sense without HASH_EARLY */
8075 WARN_ON(!(flags & HASH_EARLY));
8076 if (!(numentries >> *_hash_shift)) {
8077 numentries = 1UL << *_hash_shift;
8078 BUG_ON(!numentries);
8080 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8081 numentries = PAGE_SIZE / bucketsize;
8083 numentries = roundup_pow_of_two(numentries);
8085 /* limit allocation size to 1/16 total memory by default */
8087 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8088 do_div(max, bucketsize);
8090 max = min(max, 0x80000000ULL);
8092 if (numentries < low_limit)
8093 numentries = low_limit;
8094 if (numentries > max)
8097 log2qty = ilog2(numentries);
8099 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8102 size = bucketsize << log2qty;
8103 if (flags & HASH_EARLY) {
8104 if (flags & HASH_ZERO)
8105 table = memblock_alloc(size, SMP_CACHE_BYTES);
8107 table = memblock_alloc_raw(size,
8109 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8110 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
8114 * If bucketsize is not a power-of-two, we may free
8115 * some pages at the end of hash table which
8116 * alloc_pages_exact() automatically does
8118 table = alloc_pages_exact(size, gfp_flags);
8119 kmemleak_alloc(table, size, 1, gfp_flags);
8121 } while (!table && size > PAGE_SIZE && --log2qty);
8124 panic("Failed to allocate %s hash table\n", tablename);
8126 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8127 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8128 virt ? "vmalloc" : "linear");
8131 *_hash_shift = log2qty;
8133 *_hash_mask = (1 << log2qty) - 1;
8139 * This function checks whether pageblock includes unmovable pages or not.
8140 * If @count is not zero, it is okay to include less @count unmovable pages
8142 * PageLRU check without isolation or lru_lock could race so that
8143 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8144 * check without lock_page also may miss some movable non-lru pages at
8145 * race condition. So you can't expect this function should be exact.
8147 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
8148 int migratetype, int flags)
8150 unsigned long found;
8151 unsigned long iter = 0;
8152 unsigned long pfn = page_to_pfn(page);
8153 const char *reason = "unmovable page";
8156 * TODO we could make this much more efficient by not checking every
8157 * page in the range if we know all of them are in MOVABLE_ZONE and
8158 * that the movable zone guarantees that pages are migratable but
8159 * the later is not the case right now unfortunatelly. E.g. movablecore
8160 * can still lead to having bootmem allocations in zone_movable.
8163 if (is_migrate_cma_page(page)) {
8165 * CMA allocations (alloc_contig_range) really need to mark
8166 * isolate CMA pageblocks even when they are not movable in fact
8167 * so consider them movable here.
8169 if (is_migrate_cma(migratetype))
8172 reason = "CMA page";
8176 for (found = 0; iter < pageblock_nr_pages; iter++) {
8177 unsigned long check = pfn + iter;
8179 if (!pfn_valid_within(check))
8182 page = pfn_to_page(check);
8184 if (PageReserved(page))
8188 * If the zone is movable and we have ruled out all reserved
8189 * pages then it should be reasonably safe to assume the rest
8192 if (zone_idx(zone) == ZONE_MOVABLE)
8196 * Hugepages are not in LRU lists, but they're movable.
8197 * We need not scan over tail pages because we don't
8198 * handle each tail page individually in migration.
8200 if (PageHuge(page)) {
8201 struct page *head = compound_head(page);
8202 unsigned int skip_pages;
8204 if (!hugepage_migration_supported(page_hstate(head)))
8207 skip_pages = compound_nr(head) - (page - head);
8208 iter += skip_pages - 1;
8213 * We can't use page_count without pin a page
8214 * because another CPU can free compound page.
8215 * This check already skips compound tails of THP
8216 * because their page->_refcount is zero at all time.
8218 if (!page_ref_count(page)) {
8219 if (PageBuddy(page))
8220 iter += (1 << page_order(page)) - 1;
8225 * The HWPoisoned page may be not in buddy system, and
8226 * page_count() is not 0.
8228 if ((flags & SKIP_HWPOISON) && PageHWPoison(page))
8231 if (__PageMovable(page))
8237 * If there are RECLAIMABLE pages, we need to check
8238 * it. But now, memory offline itself doesn't call
8239 * shrink_node_slabs() and it still to be fixed.
8242 * If the page is not RAM, page_count()should be 0.
8243 * we don't need more check. This is an _used_ not-movable page.
8245 * The problematic thing here is PG_reserved pages. PG_reserved
8246 * is set to both of a memory hole page and a _used_ kernel
8254 WARN_ON_ONCE(zone_idx(zone) == ZONE_MOVABLE);
8255 if (flags & REPORT_FAILURE)
8256 dump_page(pfn_to_page(pfn + iter), reason);
8260 #ifdef CONFIG_CONTIG_ALLOC
8261 static unsigned long pfn_max_align_down(unsigned long pfn)
8263 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8264 pageblock_nr_pages) - 1);
8267 static unsigned long pfn_max_align_up(unsigned long pfn)
8269 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8270 pageblock_nr_pages));
8273 /* [start, end) must belong to a single zone. */
8274 static int __alloc_contig_migrate_range(struct compact_control *cc,
8275 unsigned long start, unsigned long end)
8277 /* This function is based on compact_zone() from compaction.c. */
8278 unsigned long nr_reclaimed;
8279 unsigned long pfn = start;
8280 unsigned int tries = 0;
8285 while (pfn < end || !list_empty(&cc->migratepages)) {
8286 if (fatal_signal_pending(current)) {
8291 if (list_empty(&cc->migratepages)) {
8292 cc->nr_migratepages = 0;
8293 pfn = isolate_migratepages_range(cc, pfn, end);
8299 } else if (++tries == 5) {
8300 ret = ret < 0 ? ret : -EBUSY;
8304 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8306 cc->nr_migratepages -= nr_reclaimed;
8308 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
8309 NULL, 0, cc->mode, MR_CONTIG_RANGE);
8312 putback_movable_pages(&cc->migratepages);
8319 * alloc_contig_range() -- tries to allocate given range of pages
8320 * @start: start PFN to allocate
8321 * @end: one-past-the-last PFN to allocate
8322 * @migratetype: migratetype of the underlaying pageblocks (either
8323 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8324 * in range must have the same migratetype and it must
8325 * be either of the two.
8326 * @gfp_mask: GFP mask to use during compaction
8328 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8329 * aligned. The PFN range must belong to a single zone.
8331 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8332 * pageblocks in the range. Once isolated, the pageblocks should not
8333 * be modified by others.
8335 * Return: zero on success or negative error code. On success all
8336 * pages which PFN is in [start, end) are allocated for the caller and
8337 * need to be freed with free_contig_range().
8339 int alloc_contig_range(unsigned long start, unsigned long end,
8340 unsigned migratetype, gfp_t gfp_mask)
8342 unsigned long outer_start, outer_end;
8346 struct compact_control cc = {
8347 .nr_migratepages = 0,
8349 .zone = page_zone(pfn_to_page(start)),
8350 .mode = MIGRATE_SYNC,
8351 .ignore_skip_hint = true,
8352 .no_set_skip_hint = true,
8353 .gfp_mask = current_gfp_context(gfp_mask),
8355 INIT_LIST_HEAD(&cc.migratepages);
8358 * What we do here is we mark all pageblocks in range as
8359 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8360 * have different sizes, and due to the way page allocator
8361 * work, we align the range to biggest of the two pages so
8362 * that page allocator won't try to merge buddies from
8363 * different pageblocks and change MIGRATE_ISOLATE to some
8364 * other migration type.
8366 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8367 * migrate the pages from an unaligned range (ie. pages that
8368 * we are interested in). This will put all the pages in
8369 * range back to page allocator as MIGRATE_ISOLATE.
8371 * When this is done, we take the pages in range from page
8372 * allocator removing them from the buddy system. This way
8373 * page allocator will never consider using them.
8375 * This lets us mark the pageblocks back as
8376 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8377 * aligned range but not in the unaligned, original range are
8378 * put back to page allocator so that buddy can use them.
8381 ret = start_isolate_page_range(pfn_max_align_down(start),
8382 pfn_max_align_up(end), migratetype, 0);
8387 * In case of -EBUSY, we'd like to know which page causes problem.
8388 * So, just fall through. test_pages_isolated() has a tracepoint
8389 * which will report the busy page.
8391 * It is possible that busy pages could become available before
8392 * the call to test_pages_isolated, and the range will actually be
8393 * allocated. So, if we fall through be sure to clear ret so that
8394 * -EBUSY is not accidentally used or returned to caller.
8396 ret = __alloc_contig_migrate_range(&cc, start, end);
8397 if (ret && ret != -EBUSY)
8402 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8403 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8404 * more, all pages in [start, end) are free in page allocator.
8405 * What we are going to do is to allocate all pages from
8406 * [start, end) (that is remove them from page allocator).
8408 * The only problem is that pages at the beginning and at the
8409 * end of interesting range may be not aligned with pages that
8410 * page allocator holds, ie. they can be part of higher order
8411 * pages. Because of this, we reserve the bigger range and
8412 * once this is done free the pages we are not interested in.
8414 * We don't have to hold zone->lock here because the pages are
8415 * isolated thus they won't get removed from buddy.
8418 lru_add_drain_all();
8421 outer_start = start;
8422 while (!PageBuddy(pfn_to_page(outer_start))) {
8423 if (++order >= MAX_ORDER) {
8424 outer_start = start;
8427 outer_start &= ~0UL << order;
8430 if (outer_start != start) {
8431 order = page_order(pfn_to_page(outer_start));
8434 * outer_start page could be small order buddy page and
8435 * it doesn't include start page. Adjust outer_start
8436 * in this case to report failed page properly
8437 * on tracepoint in test_pages_isolated()
8439 if (outer_start + (1UL << order) <= start)
8440 outer_start = start;
8443 /* Make sure the range is really isolated. */
8444 if (test_pages_isolated(outer_start, end, false)) {
8445 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8446 __func__, outer_start, end);
8451 /* Grab isolated pages from freelists. */
8452 outer_end = isolate_freepages_range(&cc, outer_start, end);
8458 /* Free head and tail (if any) */
8459 if (start != outer_start)
8460 free_contig_range(outer_start, start - outer_start);
8461 if (end != outer_end)
8462 free_contig_range(end, outer_end - end);
8465 undo_isolate_page_range(pfn_max_align_down(start),
8466 pfn_max_align_up(end), migratetype);
8469 #endif /* CONFIG_CONTIG_ALLOC */
8471 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8473 unsigned int count = 0;
8475 for (; nr_pages--; pfn++) {
8476 struct page *page = pfn_to_page(pfn);
8478 count += page_count(page) != 1;
8481 WARN(count != 0, "%d pages are still in use!\n", count);
8484 #ifdef CONFIG_MEMORY_HOTPLUG
8486 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8487 * page high values need to be recalulated.
8489 void __meminit zone_pcp_update(struct zone *zone)
8492 mutex_lock(&pcp_batch_high_lock);
8493 for_each_possible_cpu(cpu)
8494 pageset_set_high_and_batch(zone,
8495 per_cpu_ptr(zone->pageset, cpu));
8496 mutex_unlock(&pcp_batch_high_lock);
8500 void zone_pcp_reset(struct zone *zone)
8502 unsigned long flags;
8504 struct per_cpu_pageset *pset;
8506 /* avoid races with drain_pages() */
8507 local_irq_save(flags);
8508 if (zone->pageset != &boot_pageset) {
8509 for_each_online_cpu(cpu) {
8510 pset = per_cpu_ptr(zone->pageset, cpu);
8511 drain_zonestat(zone, pset);
8513 free_percpu(zone->pageset);
8514 zone->pageset = &boot_pageset;
8516 local_irq_restore(flags);
8519 #ifdef CONFIG_MEMORY_HOTREMOVE
8521 * All pages in the range must be in a single zone and isolated
8522 * before calling this.
8525 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8529 unsigned int order, i;
8531 unsigned long flags;
8532 unsigned long offlined_pages = 0;
8534 /* find the first valid pfn */
8535 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8539 return offlined_pages;
8541 offline_mem_sections(pfn, end_pfn);
8542 zone = page_zone(pfn_to_page(pfn));
8543 spin_lock_irqsave(&zone->lock, flags);
8545 while (pfn < end_pfn) {
8546 if (!pfn_valid(pfn)) {
8550 page = pfn_to_page(pfn);
8552 * The HWPoisoned page may be not in buddy system, and
8553 * page_count() is not 0.
8555 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8557 SetPageReserved(page);
8562 BUG_ON(page_count(page));
8563 BUG_ON(!PageBuddy(page));
8564 order = page_order(page);
8565 offlined_pages += 1 << order;
8566 #ifdef CONFIG_DEBUG_VM
8567 pr_info("remove from free list %lx %d %lx\n",
8568 pfn, 1 << order, end_pfn);
8570 del_page_from_free_area(page, &zone->free_area[order]);
8571 for (i = 0; i < (1 << order); i++)
8572 SetPageReserved((page+i));
8573 pfn += (1 << order);
8575 spin_unlock_irqrestore(&zone->lock, flags);
8577 return offlined_pages;
8581 bool is_free_buddy_page(struct page *page)
8583 struct zone *zone = page_zone(page);
8584 unsigned long pfn = page_to_pfn(page);
8585 unsigned long flags;
8588 spin_lock_irqsave(&zone->lock, flags);
8589 for (order = 0; order < MAX_ORDER; order++) {
8590 struct page *page_head = page - (pfn & ((1 << order) - 1));
8592 if (PageBuddy(page_head) && page_order(page_head) >= order)
8595 spin_unlock_irqrestore(&zone->lock, flags);
8597 return order < MAX_ORDER;
8600 #ifdef CONFIG_MEMORY_FAILURE
8602 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8603 * test is performed under the zone lock to prevent a race against page
8606 bool set_hwpoison_free_buddy_page(struct page *page)
8608 struct zone *zone = page_zone(page);
8609 unsigned long pfn = page_to_pfn(page);
8610 unsigned long flags;
8612 bool hwpoisoned = false;
8614 spin_lock_irqsave(&zone->lock, flags);
8615 for (order = 0; order < MAX_ORDER; order++) {
8616 struct page *page_head = page - (pfn & ((1 << order) - 1));
8618 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8619 if (!TestSetPageHWPoison(page))
8624 spin_unlock_irqrestore(&zone->lock, flags);