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 mem_cgroup_uncharge(page);
674 __free_pages_ok(page, compound_order(page));
677 void prep_compound_page(struct page *page, unsigned int order)
680 int nr_pages = 1 << order;
682 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
683 set_compound_order(page, order);
685 for (i = 1; i < nr_pages; i++) {
686 struct page *p = page + i;
687 set_page_count(p, 0);
688 p->mapping = TAIL_MAPPING;
689 set_compound_head(p, page);
691 atomic_set(compound_mapcount_ptr(page), -1);
694 #ifdef CONFIG_DEBUG_PAGEALLOC
695 unsigned int _debug_guardpage_minorder;
697 #ifdef CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
698 DEFINE_STATIC_KEY_TRUE(_debug_pagealloc_enabled);
700 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
702 EXPORT_SYMBOL(_debug_pagealloc_enabled);
704 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
706 static int __init early_debug_pagealloc(char *buf)
710 if (kstrtobool(buf, &enable))
714 static_branch_enable(&_debug_pagealloc_enabled);
718 early_param("debug_pagealloc", early_debug_pagealloc);
720 static void init_debug_guardpage(void)
722 if (!debug_pagealloc_enabled())
725 if (!debug_guardpage_minorder())
728 static_branch_enable(&_debug_guardpage_enabled);
731 static int __init debug_guardpage_minorder_setup(char *buf)
735 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
736 pr_err("Bad debug_guardpage_minorder value\n");
739 _debug_guardpage_minorder = res;
740 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
743 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
745 static inline bool set_page_guard(struct zone *zone, struct page *page,
746 unsigned int order, int migratetype)
748 if (!debug_guardpage_enabled())
751 if (order >= debug_guardpage_minorder())
754 __SetPageGuard(page);
755 INIT_LIST_HEAD(&page->lru);
756 set_page_private(page, order);
757 /* Guard pages are not available for any usage */
758 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
763 static inline void clear_page_guard(struct zone *zone, struct page *page,
764 unsigned int order, int migratetype)
766 if (!debug_guardpage_enabled())
769 __ClearPageGuard(page);
771 set_page_private(page, 0);
772 if (!is_migrate_isolate(migratetype))
773 __mod_zone_freepage_state(zone, (1 << order), migratetype);
776 static inline bool set_page_guard(struct zone *zone, struct page *page,
777 unsigned int order, int migratetype) { return false; }
778 static inline void clear_page_guard(struct zone *zone, struct page *page,
779 unsigned int order, int migratetype) {}
782 static inline void set_page_order(struct page *page, unsigned int order)
784 set_page_private(page, order);
785 __SetPageBuddy(page);
789 * This function checks whether a page is free && is the buddy
790 * we can coalesce a page and its buddy if
791 * (a) the buddy is not in a hole (check before calling!) &&
792 * (b) the buddy is in the buddy system &&
793 * (c) a page and its buddy have the same order &&
794 * (d) a page and its buddy are in the same zone.
796 * For recording whether a page is in the buddy system, we set PageBuddy.
797 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
799 * For recording page's order, we use page_private(page).
801 static inline int page_is_buddy(struct page *page, struct page *buddy,
804 if (page_is_guard(buddy) && page_order(buddy) == order) {
805 if (page_zone_id(page) != page_zone_id(buddy))
808 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
813 if (PageBuddy(buddy) && page_order(buddy) == order) {
815 * zone check is done late to avoid uselessly
816 * calculating zone/node ids for pages that could
819 if (page_zone_id(page) != page_zone_id(buddy))
822 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
829 #ifdef CONFIG_COMPACTION
830 static inline struct capture_control *task_capc(struct zone *zone)
832 struct capture_control *capc = current->capture_control;
835 !(current->flags & PF_KTHREAD) &&
837 capc->cc->zone == zone &&
838 capc->cc->direct_compaction ? capc : NULL;
842 compaction_capture(struct capture_control *capc, struct page *page,
843 int order, int migratetype)
845 if (!capc || order != capc->cc->order)
848 /* Do not accidentally pollute CMA or isolated regions*/
849 if (is_migrate_cma(migratetype) ||
850 is_migrate_isolate(migratetype))
854 * Do not let lower order allocations polluate a movable pageblock.
855 * This might let an unmovable request use a reclaimable pageblock
856 * and vice-versa but no more than normal fallback logic which can
857 * have trouble finding a high-order free page.
859 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
867 static inline struct capture_control *task_capc(struct zone *zone)
873 compaction_capture(struct capture_control *capc, struct page *page,
874 int order, int migratetype)
878 #endif /* CONFIG_COMPACTION */
881 * Freeing function for a buddy system allocator.
883 * The concept of a buddy system is to maintain direct-mapped table
884 * (containing bit values) for memory blocks of various "orders".
885 * The bottom level table contains the map for the smallest allocatable
886 * units of memory (here, pages), and each level above it describes
887 * pairs of units from the levels below, hence, "buddies".
888 * At a high level, all that happens here is marking the table entry
889 * at the bottom level available, and propagating the changes upward
890 * as necessary, plus some accounting needed to play nicely with other
891 * parts of the VM system.
892 * At each level, we keep a list of pages, which are heads of continuous
893 * free pages of length of (1 << order) and marked with PageBuddy.
894 * Page's order is recorded in page_private(page) field.
895 * So when we are allocating or freeing one, we can derive the state of the
896 * other. That is, if we allocate a small block, and both were
897 * free, the remainder of the region must be split into blocks.
898 * If a block is freed, and its buddy is also free, then this
899 * triggers coalescing into a block of larger size.
904 static inline void __free_one_page(struct page *page,
906 struct zone *zone, unsigned int order,
909 unsigned long combined_pfn;
910 unsigned long uninitialized_var(buddy_pfn);
912 unsigned int max_order;
913 struct capture_control *capc = task_capc(zone);
915 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
917 VM_BUG_ON(!zone_is_initialized(zone));
918 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
920 VM_BUG_ON(migratetype == -1);
921 if (likely(!is_migrate_isolate(migratetype)))
922 __mod_zone_freepage_state(zone, 1 << order, migratetype);
924 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
925 VM_BUG_ON_PAGE(bad_range(zone, page), page);
928 while (order < max_order - 1) {
929 if (compaction_capture(capc, page, order, migratetype)) {
930 __mod_zone_freepage_state(zone, -(1 << order),
934 buddy_pfn = __find_buddy_pfn(pfn, order);
935 buddy = page + (buddy_pfn - pfn);
937 if (!pfn_valid_within(buddy_pfn))
939 if (!page_is_buddy(page, buddy, order))
942 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
943 * merge with it and move up one order.
945 if (page_is_guard(buddy))
946 clear_page_guard(zone, buddy, order, migratetype);
948 del_page_from_free_area(buddy, &zone->free_area[order]);
949 combined_pfn = buddy_pfn & pfn;
950 page = page + (combined_pfn - pfn);
954 if (max_order < MAX_ORDER) {
955 /* If we are here, it means order is >= pageblock_order.
956 * We want to prevent merge between freepages on isolate
957 * pageblock and normal pageblock. Without this, pageblock
958 * isolation could cause incorrect freepage or CMA accounting.
960 * We don't want to hit this code for the more frequent
963 if (unlikely(has_isolate_pageblock(zone))) {
966 buddy_pfn = __find_buddy_pfn(pfn, order);
967 buddy = page + (buddy_pfn - pfn);
968 buddy_mt = get_pageblock_migratetype(buddy);
970 if (migratetype != buddy_mt
971 && (is_migrate_isolate(migratetype) ||
972 is_migrate_isolate(buddy_mt)))
976 goto continue_merging;
980 set_page_order(page, order);
983 * If this is not the largest possible page, check if the buddy
984 * of the next-highest order is free. If it is, it's possible
985 * that pages are being freed that will coalesce soon. In case,
986 * that is happening, add the free page to the tail of the list
987 * so it's less likely to be used soon and more likely to be merged
988 * as a higher order page
990 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)
991 && !is_shuffle_order(order)) {
992 struct page *higher_page, *higher_buddy;
993 combined_pfn = buddy_pfn & pfn;
994 higher_page = page + (combined_pfn - pfn);
995 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
996 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
997 if (pfn_valid_within(buddy_pfn) &&
998 page_is_buddy(higher_page, higher_buddy, order + 1)) {
999 add_to_free_area_tail(page, &zone->free_area[order],
1005 if (is_shuffle_order(order))
1006 add_to_free_area_random(page, &zone->free_area[order],
1009 add_to_free_area(page, &zone->free_area[order], migratetype);
1014 * A bad page could be due to a number of fields. Instead of multiple branches,
1015 * try and check multiple fields with one check. The caller must do a detailed
1016 * check if necessary.
1018 static inline bool page_expected_state(struct page *page,
1019 unsigned long check_flags)
1021 if (unlikely(atomic_read(&page->_mapcount) != -1))
1024 if (unlikely((unsigned long)page->mapping |
1025 page_ref_count(page) |
1027 (unsigned long)page->mem_cgroup |
1029 (page->flags & check_flags)))
1035 static void free_pages_check_bad(struct page *page)
1037 const char *bad_reason;
1038 unsigned long bad_flags;
1043 if (unlikely(atomic_read(&page->_mapcount) != -1))
1044 bad_reason = "nonzero mapcount";
1045 if (unlikely(page->mapping != NULL))
1046 bad_reason = "non-NULL mapping";
1047 if (unlikely(page_ref_count(page) != 0))
1048 bad_reason = "nonzero _refcount";
1049 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
1050 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1051 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
1054 if (unlikely(page->mem_cgroup))
1055 bad_reason = "page still charged to cgroup";
1057 bad_page(page, bad_reason, bad_flags);
1060 static inline int free_pages_check(struct page *page)
1062 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1065 /* Something has gone sideways, find it */
1066 free_pages_check_bad(page);
1070 static int free_tail_pages_check(struct page *head_page, struct page *page)
1075 * We rely page->lru.next never has bit 0 set, unless the page
1076 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1078 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1080 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1084 switch (page - head_page) {
1086 /* the first tail page: ->mapping may be compound_mapcount() */
1087 if (unlikely(compound_mapcount(page))) {
1088 bad_page(page, "nonzero compound_mapcount", 0);
1094 * the second tail page: ->mapping is
1095 * deferred_list.next -- ignore value.
1099 if (page->mapping != TAIL_MAPPING) {
1100 bad_page(page, "corrupted mapping in tail page", 0);
1105 if (unlikely(!PageTail(page))) {
1106 bad_page(page, "PageTail not set", 0);
1109 if (unlikely(compound_head(page) != head_page)) {
1110 bad_page(page, "compound_head not consistent", 0);
1115 page->mapping = NULL;
1116 clear_compound_head(page);
1120 static void kernel_init_free_pages(struct page *page, int numpages)
1124 for (i = 0; i < numpages; i++)
1125 clear_highpage(page + i);
1128 static __always_inline bool free_pages_prepare(struct page *page,
1129 unsigned int order, bool check_free)
1133 VM_BUG_ON_PAGE(PageTail(page), page);
1135 trace_mm_page_free(page, order);
1138 * Check tail pages before head page information is cleared to
1139 * avoid checking PageCompound for order-0 pages.
1141 if (unlikely(order)) {
1142 bool compound = PageCompound(page);
1145 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1148 ClearPageDoubleMap(page);
1149 for (i = 1; i < (1 << order); i++) {
1151 bad += free_tail_pages_check(page, page + i);
1152 if (unlikely(free_pages_check(page + i))) {
1156 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1159 if (PageMappingFlags(page))
1160 page->mapping = NULL;
1161 if (memcg_kmem_enabled() && PageKmemcg(page))
1162 __memcg_kmem_uncharge(page, order);
1164 bad += free_pages_check(page);
1168 page_cpupid_reset_last(page);
1169 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1170 reset_page_owner(page, order);
1172 if (!PageHighMem(page)) {
1173 debug_check_no_locks_freed(page_address(page),
1174 PAGE_SIZE << order);
1175 debug_check_no_obj_freed(page_address(page),
1176 PAGE_SIZE << order);
1178 arch_free_page(page, order);
1179 if (want_init_on_free())
1180 kernel_init_free_pages(page, 1 << order);
1182 kernel_poison_pages(page, 1 << order, 0);
1183 if (debug_pagealloc_enabled())
1184 kernel_map_pages(page, 1 << order, 0);
1186 kasan_free_nondeferred_pages(page, order);
1191 #ifdef CONFIG_DEBUG_VM
1193 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1194 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1195 * moved from pcp lists to free lists.
1197 static bool free_pcp_prepare(struct page *page)
1199 return free_pages_prepare(page, 0, true);
1202 static bool bulkfree_pcp_prepare(struct page *page)
1204 if (debug_pagealloc_enabled())
1205 return free_pages_check(page);
1211 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1212 * moving from pcp lists to free list in order to reduce overhead. With
1213 * debug_pagealloc enabled, they are checked also immediately when being freed
1216 static bool free_pcp_prepare(struct page *page)
1218 if (debug_pagealloc_enabled())
1219 return free_pages_prepare(page, 0, true);
1221 return free_pages_prepare(page, 0, false);
1224 static bool bulkfree_pcp_prepare(struct page *page)
1226 return free_pages_check(page);
1228 #endif /* CONFIG_DEBUG_VM */
1230 static inline void prefetch_buddy(struct page *page)
1232 unsigned long pfn = page_to_pfn(page);
1233 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1234 struct page *buddy = page + (buddy_pfn - pfn);
1240 * Frees a number of pages from the PCP lists
1241 * Assumes all pages on list are in same zone, and of same order.
1242 * count is the number of pages to free.
1244 * If the zone was previously in an "all pages pinned" state then look to
1245 * see if this freeing clears that state.
1247 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1248 * pinned" detection logic.
1250 static void free_pcppages_bulk(struct zone *zone, int count,
1251 struct per_cpu_pages *pcp)
1253 int migratetype = 0;
1255 int prefetch_nr = 0;
1256 bool isolated_pageblocks;
1257 struct page *page, *tmp;
1261 struct list_head *list;
1264 * Remove pages from lists in a round-robin fashion. A
1265 * batch_free count is maintained that is incremented when an
1266 * empty list is encountered. This is so more pages are freed
1267 * off fuller lists instead of spinning excessively around empty
1272 if (++migratetype == MIGRATE_PCPTYPES)
1274 list = &pcp->lists[migratetype];
1275 } while (list_empty(list));
1277 /* This is the only non-empty list. Free them all. */
1278 if (batch_free == MIGRATE_PCPTYPES)
1282 page = list_last_entry(list, struct page, lru);
1283 /* must delete to avoid corrupting pcp list */
1284 list_del(&page->lru);
1287 if (bulkfree_pcp_prepare(page))
1290 list_add_tail(&page->lru, &head);
1293 * We are going to put the page back to the global
1294 * pool, prefetch its buddy to speed up later access
1295 * under zone->lock. It is believed the overhead of
1296 * an additional test and calculating buddy_pfn here
1297 * can be offset by reduced memory latency later. To
1298 * avoid excessive prefetching due to large count, only
1299 * prefetch buddy for the first pcp->batch nr of pages.
1301 if (prefetch_nr++ < pcp->batch)
1302 prefetch_buddy(page);
1303 } while (--count && --batch_free && !list_empty(list));
1306 spin_lock(&zone->lock);
1307 isolated_pageblocks = has_isolate_pageblock(zone);
1310 * Use safe version since after __free_one_page(),
1311 * page->lru.next will not point to original list.
1313 list_for_each_entry_safe(page, tmp, &head, lru) {
1314 int mt = get_pcppage_migratetype(page);
1315 /* MIGRATE_ISOLATE page should not go to pcplists */
1316 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1317 /* Pageblock could have been isolated meanwhile */
1318 if (unlikely(isolated_pageblocks))
1319 mt = get_pageblock_migratetype(page);
1321 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1322 trace_mm_page_pcpu_drain(page, 0, mt);
1324 spin_unlock(&zone->lock);
1327 static void free_one_page(struct zone *zone,
1328 struct page *page, unsigned long pfn,
1332 spin_lock(&zone->lock);
1333 if (unlikely(has_isolate_pageblock(zone) ||
1334 is_migrate_isolate(migratetype))) {
1335 migratetype = get_pfnblock_migratetype(page, pfn);
1337 __free_one_page(page, pfn, zone, order, migratetype);
1338 spin_unlock(&zone->lock);
1341 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1342 unsigned long zone, int nid)
1344 mm_zero_struct_page(page);
1345 set_page_links(page, zone, nid, pfn);
1346 init_page_count(page);
1347 page_mapcount_reset(page);
1348 page_cpupid_reset_last(page);
1349 page_kasan_tag_reset(page);
1351 INIT_LIST_HEAD(&page->lru);
1352 #ifdef WANT_PAGE_VIRTUAL
1353 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1354 if (!is_highmem_idx(zone))
1355 set_page_address(page, __va(pfn << PAGE_SHIFT));
1359 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1360 static void __meminit init_reserved_page(unsigned long pfn)
1365 if (!early_page_uninitialised(pfn))
1368 nid = early_pfn_to_nid(pfn);
1369 pgdat = NODE_DATA(nid);
1371 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1372 struct zone *zone = &pgdat->node_zones[zid];
1374 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1377 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1380 static inline void init_reserved_page(unsigned long pfn)
1383 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1386 * Initialised pages do not have PageReserved set. This function is
1387 * called for each range allocated by the bootmem allocator and
1388 * marks the pages PageReserved. The remaining valid pages are later
1389 * sent to the buddy page allocator.
1391 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1393 unsigned long start_pfn = PFN_DOWN(start);
1394 unsigned long end_pfn = PFN_UP(end);
1396 for (; start_pfn < end_pfn; start_pfn++) {
1397 if (pfn_valid(start_pfn)) {
1398 struct page *page = pfn_to_page(start_pfn);
1400 init_reserved_page(start_pfn);
1402 /* Avoid false-positive PageTail() */
1403 INIT_LIST_HEAD(&page->lru);
1406 * no need for atomic set_bit because the struct
1407 * page is not visible yet so nobody should
1410 __SetPageReserved(page);
1415 static void __free_pages_ok(struct page *page, unsigned int order)
1417 unsigned long flags;
1419 unsigned long pfn = page_to_pfn(page);
1421 if (!free_pages_prepare(page, order, true))
1424 migratetype = get_pfnblock_migratetype(page, pfn);
1425 local_irq_save(flags);
1426 __count_vm_events(PGFREE, 1 << order);
1427 free_one_page(page_zone(page), page, pfn, order, migratetype);
1428 local_irq_restore(flags);
1431 void __free_pages_core(struct page *page, unsigned int order)
1433 unsigned int nr_pages = 1 << order;
1434 struct page *p = page;
1438 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1440 __ClearPageReserved(p);
1441 set_page_count(p, 0);
1443 __ClearPageReserved(p);
1444 set_page_count(p, 0);
1446 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1447 set_page_refcounted(page);
1448 __free_pages(page, order);
1451 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1452 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1454 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1456 int __meminit early_pfn_to_nid(unsigned long pfn)
1458 static DEFINE_SPINLOCK(early_pfn_lock);
1461 spin_lock(&early_pfn_lock);
1462 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1464 nid = first_online_node;
1465 spin_unlock(&early_pfn_lock);
1471 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1472 /* Only safe to use early in boot when initialisation is single-threaded */
1473 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1477 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1478 if (nid >= 0 && nid != node)
1484 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1491 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1494 if (early_page_uninitialised(pfn))
1496 __free_pages_core(page, order);
1500 * Check that the whole (or subset of) a pageblock given by the interval of
1501 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1502 * with the migration of free compaction scanner. The scanners then need to
1503 * use only pfn_valid_within() check for arches that allow holes within
1506 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1508 * It's possible on some configurations to have a setup like node0 node1 node0
1509 * i.e. it's possible that all pages within a zones range of pages do not
1510 * belong to a single zone. We assume that a border between node0 and node1
1511 * can occur within a single pageblock, but not a node0 node1 node0
1512 * interleaving within a single pageblock. It is therefore sufficient to check
1513 * the first and last page of a pageblock and avoid checking each individual
1514 * page in a pageblock.
1516 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1517 unsigned long end_pfn, struct zone *zone)
1519 struct page *start_page;
1520 struct page *end_page;
1522 /* end_pfn is one past the range we are checking */
1525 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1528 start_page = pfn_to_online_page(start_pfn);
1532 if (page_zone(start_page) != zone)
1535 end_page = pfn_to_page(end_pfn);
1537 /* This gives a shorter code than deriving page_zone(end_page) */
1538 if (page_zone_id(start_page) != page_zone_id(end_page))
1544 void set_zone_contiguous(struct zone *zone)
1546 unsigned long block_start_pfn = zone->zone_start_pfn;
1547 unsigned long block_end_pfn;
1549 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1550 for (; block_start_pfn < zone_end_pfn(zone);
1551 block_start_pfn = block_end_pfn,
1552 block_end_pfn += pageblock_nr_pages) {
1554 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1556 if (!__pageblock_pfn_to_page(block_start_pfn,
1557 block_end_pfn, zone))
1561 /* We confirm that there is no hole */
1562 zone->contiguous = true;
1565 void clear_zone_contiguous(struct zone *zone)
1567 zone->contiguous = false;
1570 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1571 static void __init deferred_free_range(unsigned long pfn,
1572 unsigned long nr_pages)
1580 page = pfn_to_page(pfn);
1582 /* Free a large naturally-aligned chunk if possible */
1583 if (nr_pages == pageblock_nr_pages &&
1584 (pfn & (pageblock_nr_pages - 1)) == 0) {
1585 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1586 __free_pages_core(page, pageblock_order);
1590 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1591 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1592 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1593 __free_pages_core(page, 0);
1597 /* Completion tracking for deferred_init_memmap() threads */
1598 static atomic_t pgdat_init_n_undone __initdata;
1599 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1601 static inline void __init pgdat_init_report_one_done(void)
1603 if (atomic_dec_and_test(&pgdat_init_n_undone))
1604 complete(&pgdat_init_all_done_comp);
1608 * Returns true if page needs to be initialized or freed to buddy allocator.
1610 * First we check if pfn is valid on architectures where it is possible to have
1611 * holes within pageblock_nr_pages. On systems where it is not possible, this
1612 * function is optimized out.
1614 * Then, we check if a current large page is valid by only checking the validity
1617 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1619 if (!pfn_valid_within(pfn))
1621 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1627 * Free pages to buddy allocator. Try to free aligned pages in
1628 * pageblock_nr_pages sizes.
1630 static void __init deferred_free_pages(unsigned long pfn,
1631 unsigned long end_pfn)
1633 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1634 unsigned long nr_free = 0;
1636 for (; pfn < end_pfn; pfn++) {
1637 if (!deferred_pfn_valid(pfn)) {
1638 deferred_free_range(pfn - nr_free, nr_free);
1640 } else if (!(pfn & nr_pgmask)) {
1641 deferred_free_range(pfn - nr_free, nr_free);
1643 touch_nmi_watchdog();
1648 /* Free the last block of pages to allocator */
1649 deferred_free_range(pfn - nr_free, nr_free);
1653 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1654 * by performing it only once every pageblock_nr_pages.
1655 * Return number of pages initialized.
1657 static unsigned long __init deferred_init_pages(struct zone *zone,
1659 unsigned long end_pfn)
1661 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1662 int nid = zone_to_nid(zone);
1663 unsigned long nr_pages = 0;
1664 int zid = zone_idx(zone);
1665 struct page *page = NULL;
1667 for (; pfn < end_pfn; pfn++) {
1668 if (!deferred_pfn_valid(pfn)) {
1671 } else if (!page || !(pfn & nr_pgmask)) {
1672 page = pfn_to_page(pfn);
1673 touch_nmi_watchdog();
1677 __init_single_page(page, pfn, zid, nid);
1684 * This function is meant to pre-load the iterator for the zone init.
1685 * Specifically it walks through the ranges until we are caught up to the
1686 * first_init_pfn value and exits there. If we never encounter the value we
1687 * return false indicating there are no valid ranges left.
1690 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1691 unsigned long *spfn, unsigned long *epfn,
1692 unsigned long first_init_pfn)
1697 * Start out by walking through the ranges in this zone that have
1698 * already been initialized. We don't need to do anything with them
1699 * so we just need to flush them out of the system.
1701 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1702 if (*epfn <= first_init_pfn)
1704 if (*spfn < first_init_pfn)
1705 *spfn = first_init_pfn;
1714 * Initialize and free pages. We do it in two loops: first we initialize
1715 * struct page, then free to buddy allocator, because while we are
1716 * freeing pages we can access pages that are ahead (computing buddy
1717 * page in __free_one_page()).
1719 * In order to try and keep some memory in the cache we have the loop
1720 * broken along max page order boundaries. This way we will not cause
1721 * any issues with the buddy page computation.
1723 static unsigned long __init
1724 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1725 unsigned long *end_pfn)
1727 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1728 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1729 unsigned long nr_pages = 0;
1732 /* First we loop through and initialize the page values */
1733 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1736 if (mo_pfn <= *start_pfn)
1739 t = min(mo_pfn, *end_pfn);
1740 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1742 if (mo_pfn < *end_pfn) {
1743 *start_pfn = mo_pfn;
1748 /* Reset values and now loop through freeing pages as needed */
1751 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1757 t = min(mo_pfn, epfn);
1758 deferred_free_pages(spfn, t);
1767 /* Initialise remaining memory on a node */
1768 static int __init deferred_init_memmap(void *data)
1770 pg_data_t *pgdat = data;
1771 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1772 unsigned long spfn = 0, epfn = 0, nr_pages = 0;
1773 unsigned long first_init_pfn, flags;
1774 unsigned long start = jiffies;
1779 /* Bind memory initialisation thread to a local node if possible */
1780 if (!cpumask_empty(cpumask))
1781 set_cpus_allowed_ptr(current, cpumask);
1783 pgdat_resize_lock(pgdat, &flags);
1784 first_init_pfn = pgdat->first_deferred_pfn;
1785 if (first_init_pfn == ULONG_MAX) {
1786 pgdat_resize_unlock(pgdat, &flags);
1787 pgdat_init_report_one_done();
1791 /* Sanity check boundaries */
1792 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1793 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1794 pgdat->first_deferred_pfn = ULONG_MAX;
1796 /* Only the highest zone is deferred so find it */
1797 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1798 zone = pgdat->node_zones + zid;
1799 if (first_init_pfn < zone_end_pfn(zone))
1803 /* If the zone is empty somebody else may have cleared out the zone */
1804 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1809 * Initialize and free pages in MAX_ORDER sized increments so
1810 * that we can avoid introducing any issues with the buddy
1814 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1816 pgdat_resize_unlock(pgdat, &flags);
1818 /* Sanity check that the next zone really is unpopulated */
1819 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1821 pr_info("node %d initialised, %lu pages in %ums\n",
1822 pgdat->node_id, nr_pages, jiffies_to_msecs(jiffies - start));
1824 pgdat_init_report_one_done();
1829 * If this zone has deferred pages, try to grow it by initializing enough
1830 * deferred pages to satisfy the allocation specified by order, rounded up to
1831 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1832 * of SECTION_SIZE bytes by initializing struct pages in increments of
1833 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1835 * Return true when zone was grown, otherwise return false. We return true even
1836 * when we grow less than requested, to let the caller decide if there are
1837 * enough pages to satisfy the allocation.
1839 * Note: We use noinline because this function is needed only during boot, and
1840 * it is called from a __ref function _deferred_grow_zone. This way we are
1841 * making sure that it is not inlined into permanent text section.
1843 static noinline bool __init
1844 deferred_grow_zone(struct zone *zone, unsigned int order)
1846 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1847 pg_data_t *pgdat = zone->zone_pgdat;
1848 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1849 unsigned long spfn, epfn, flags;
1850 unsigned long nr_pages = 0;
1853 /* Only the last zone may have deferred pages */
1854 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1857 pgdat_resize_lock(pgdat, &flags);
1860 * If deferred pages have been initialized while we were waiting for
1861 * the lock, return true, as the zone was grown. The caller will retry
1862 * this zone. We won't return to this function since the caller also
1863 * has this static branch.
1865 if (!static_branch_unlikely(&deferred_pages)) {
1866 pgdat_resize_unlock(pgdat, &flags);
1871 * If someone grew this zone while we were waiting for spinlock, return
1872 * true, as there might be enough pages already.
1874 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1875 pgdat_resize_unlock(pgdat, &flags);
1879 /* If the zone is empty somebody else may have cleared out the zone */
1880 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1881 first_deferred_pfn)) {
1882 pgdat->first_deferred_pfn = ULONG_MAX;
1883 pgdat_resize_unlock(pgdat, &flags);
1884 /* Retry only once. */
1885 return first_deferred_pfn != ULONG_MAX;
1889 * Initialize and free pages in MAX_ORDER sized increments so
1890 * that we can avoid introducing any issues with the buddy
1893 while (spfn < epfn) {
1894 /* update our first deferred PFN for this section */
1895 first_deferred_pfn = spfn;
1897 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1899 /* We should only stop along section boundaries */
1900 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
1903 /* If our quota has been met we can stop here */
1904 if (nr_pages >= nr_pages_needed)
1908 pgdat->first_deferred_pfn = spfn;
1909 pgdat_resize_unlock(pgdat, &flags);
1911 return nr_pages > 0;
1915 * deferred_grow_zone() is __init, but it is called from
1916 * get_page_from_freelist() during early boot until deferred_pages permanently
1917 * disables this call. This is why we have refdata wrapper to avoid warning,
1918 * and to ensure that the function body gets unloaded.
1921 _deferred_grow_zone(struct zone *zone, unsigned int order)
1923 return deferred_grow_zone(zone, order);
1926 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1928 void __init page_alloc_init_late(void)
1933 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1935 /* There will be num_node_state(N_MEMORY) threads */
1936 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1937 for_each_node_state(nid, N_MEMORY) {
1938 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1941 /* Block until all are initialised */
1942 wait_for_completion(&pgdat_init_all_done_comp);
1945 * We initialized the rest of the deferred pages. Permanently disable
1946 * on-demand struct page initialization.
1948 static_branch_disable(&deferred_pages);
1950 /* Reinit limits that are based on free pages after the kernel is up */
1951 files_maxfiles_init();
1954 /* Discard memblock private memory */
1957 for_each_node_state(nid, N_MEMORY)
1958 shuffle_free_memory(NODE_DATA(nid));
1960 for_each_populated_zone(zone)
1961 set_zone_contiguous(zone);
1963 #ifdef CONFIG_DEBUG_PAGEALLOC
1964 init_debug_guardpage();
1969 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1970 void __init init_cma_reserved_pageblock(struct page *page)
1972 unsigned i = pageblock_nr_pages;
1973 struct page *p = page;
1976 __ClearPageReserved(p);
1977 set_page_count(p, 0);
1980 set_pageblock_migratetype(page, MIGRATE_CMA);
1982 if (pageblock_order >= MAX_ORDER) {
1983 i = pageblock_nr_pages;
1986 set_page_refcounted(p);
1987 __free_pages(p, MAX_ORDER - 1);
1988 p += MAX_ORDER_NR_PAGES;
1989 } while (i -= MAX_ORDER_NR_PAGES);
1991 set_page_refcounted(page);
1992 __free_pages(page, pageblock_order);
1995 adjust_managed_page_count(page, pageblock_nr_pages);
2000 * The order of subdivision here is critical for the IO subsystem.
2001 * Please do not alter this order without good reasons and regression
2002 * testing. Specifically, as large blocks of memory are subdivided,
2003 * the order in which smaller blocks are delivered depends on the order
2004 * they're subdivided in this function. This is the primary factor
2005 * influencing the order in which pages are delivered to the IO
2006 * subsystem according to empirical testing, and this is also justified
2007 * by considering the behavior of a buddy system containing a single
2008 * large block of memory acted on by a series of small allocations.
2009 * This behavior is a critical factor in sglist merging's success.
2013 static inline void expand(struct zone *zone, struct page *page,
2014 int low, int high, struct free_area *area,
2017 unsigned long size = 1 << high;
2019 while (high > low) {
2023 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2026 * Mark as guard pages (or page), that will allow to
2027 * merge back to allocator when buddy will be freed.
2028 * Corresponding page table entries will not be touched,
2029 * pages will stay not present in virtual address space
2031 if (set_page_guard(zone, &page[size], high, migratetype))
2034 add_to_free_area(&page[size], area, migratetype);
2035 set_page_order(&page[size], high);
2039 static void check_new_page_bad(struct page *page)
2041 const char *bad_reason = NULL;
2042 unsigned long bad_flags = 0;
2044 if (unlikely(atomic_read(&page->_mapcount) != -1))
2045 bad_reason = "nonzero mapcount";
2046 if (unlikely(page->mapping != NULL))
2047 bad_reason = "non-NULL mapping";
2048 if (unlikely(page_ref_count(page) != 0))
2049 bad_reason = "nonzero _refcount";
2050 if (unlikely(page->flags & __PG_HWPOISON)) {
2051 bad_reason = "HWPoisoned (hardware-corrupted)";
2052 bad_flags = __PG_HWPOISON;
2053 /* Don't complain about hwpoisoned pages */
2054 page_mapcount_reset(page); /* remove PageBuddy */
2057 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
2058 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
2059 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
2062 if (unlikely(page->mem_cgroup))
2063 bad_reason = "page still charged to cgroup";
2065 bad_page(page, bad_reason, bad_flags);
2069 * This page is about to be returned from the page allocator
2071 static inline int check_new_page(struct page *page)
2073 if (likely(page_expected_state(page,
2074 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2077 check_new_page_bad(page);
2081 static inline bool free_pages_prezeroed(void)
2083 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2084 page_poisoning_enabled()) || want_init_on_free();
2087 #ifdef CONFIG_DEBUG_VM
2089 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2090 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2091 * also checked when pcp lists are refilled from the free lists.
2093 static inline bool check_pcp_refill(struct page *page)
2095 if (debug_pagealloc_enabled())
2096 return check_new_page(page);
2101 static inline bool check_new_pcp(struct page *page)
2103 return check_new_page(page);
2107 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2108 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2109 * enabled, they are also checked when being allocated from the pcp lists.
2111 static inline bool check_pcp_refill(struct page *page)
2113 return check_new_page(page);
2115 static inline bool check_new_pcp(struct page *page)
2117 if (debug_pagealloc_enabled())
2118 return check_new_page(page);
2122 #endif /* CONFIG_DEBUG_VM */
2124 static bool check_new_pages(struct page *page, unsigned int order)
2127 for (i = 0; i < (1 << order); i++) {
2128 struct page *p = page + i;
2130 if (unlikely(check_new_page(p)))
2137 inline void post_alloc_hook(struct page *page, unsigned int order,
2140 set_page_private(page, 0);
2141 set_page_refcounted(page);
2143 arch_alloc_page(page, order);
2144 if (debug_pagealloc_enabled())
2145 kernel_map_pages(page, 1 << order, 1);
2146 kasan_alloc_pages(page, order);
2147 kernel_poison_pages(page, 1 << order, 1);
2148 set_page_owner(page, order, gfp_flags);
2151 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2152 unsigned int alloc_flags)
2154 post_alloc_hook(page, order, gfp_flags);
2156 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2157 kernel_init_free_pages(page, 1 << order);
2159 if (order && (gfp_flags & __GFP_COMP))
2160 prep_compound_page(page, order);
2163 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2164 * allocate the page. The expectation is that the caller is taking
2165 * steps that will free more memory. The caller should avoid the page
2166 * being used for !PFMEMALLOC purposes.
2168 if (alloc_flags & ALLOC_NO_WATERMARKS)
2169 set_page_pfmemalloc(page);
2171 clear_page_pfmemalloc(page);
2175 * Go through the free lists for the given migratetype and remove
2176 * the smallest available page from the freelists
2178 static __always_inline
2179 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2182 unsigned int current_order;
2183 struct free_area *area;
2186 /* Find a page of the appropriate size in the preferred list */
2187 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2188 area = &(zone->free_area[current_order]);
2189 page = get_page_from_free_area(area, migratetype);
2192 del_page_from_free_area(page, area);
2193 expand(zone, page, order, current_order, area, migratetype);
2194 set_pcppage_migratetype(page, migratetype);
2203 * This array describes the order lists are fallen back to when
2204 * the free lists for the desirable migrate type are depleted
2206 static int fallbacks[MIGRATE_TYPES][4] = {
2207 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2208 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2209 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2211 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2213 #ifdef CONFIG_MEMORY_ISOLATION
2214 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2219 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2222 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2225 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2226 unsigned int order) { return NULL; }
2230 * Move the free pages in a range to the free lists of the requested type.
2231 * Note that start_page and end_pages are not aligned on a pageblock
2232 * boundary. If alignment is required, use move_freepages_block()
2234 static int move_freepages(struct zone *zone,
2235 struct page *start_page, struct page *end_page,
2236 int migratetype, int *num_movable)
2240 int pages_moved = 0;
2242 for (page = start_page; page <= end_page;) {
2243 if (!pfn_valid_within(page_to_pfn(page))) {
2248 if (!PageBuddy(page)) {
2250 * We assume that pages that could be isolated for
2251 * migration are movable. But we don't actually try
2252 * isolating, as that would be expensive.
2255 (PageLRU(page) || __PageMovable(page)))
2262 /* Make sure we are not inadvertently changing nodes */
2263 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2264 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2266 order = page_order(page);
2267 move_to_free_area(page, &zone->free_area[order], migratetype);
2269 pages_moved += 1 << order;
2275 int move_freepages_block(struct zone *zone, struct page *page,
2276 int migratetype, int *num_movable)
2278 unsigned long start_pfn, end_pfn;
2279 struct page *start_page, *end_page;
2284 start_pfn = page_to_pfn(page);
2285 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2286 start_page = pfn_to_page(start_pfn);
2287 end_page = start_page + pageblock_nr_pages - 1;
2288 end_pfn = start_pfn + pageblock_nr_pages - 1;
2290 /* Do not cross zone boundaries */
2291 if (!zone_spans_pfn(zone, start_pfn))
2293 if (!zone_spans_pfn(zone, end_pfn))
2296 return move_freepages(zone, start_page, end_page, migratetype,
2300 static void change_pageblock_range(struct page *pageblock_page,
2301 int start_order, int migratetype)
2303 int nr_pageblocks = 1 << (start_order - pageblock_order);
2305 while (nr_pageblocks--) {
2306 set_pageblock_migratetype(pageblock_page, migratetype);
2307 pageblock_page += pageblock_nr_pages;
2312 * When we are falling back to another migratetype during allocation, try to
2313 * steal extra free pages from the same pageblocks to satisfy further
2314 * allocations, instead of polluting multiple pageblocks.
2316 * If we are stealing a relatively large buddy page, it is likely there will
2317 * be more free pages in the pageblock, so try to steal them all. For
2318 * reclaimable and unmovable allocations, we steal regardless of page size,
2319 * as fragmentation caused by those allocations polluting movable pageblocks
2320 * is worse than movable allocations stealing from unmovable and reclaimable
2323 static bool can_steal_fallback(unsigned int order, int start_mt)
2326 * Leaving this order check is intended, although there is
2327 * relaxed order check in next check. The reason is that
2328 * we can actually steal whole pageblock if this condition met,
2329 * but, below check doesn't guarantee it and that is just heuristic
2330 * so could be changed anytime.
2332 if (order >= pageblock_order)
2335 if (order >= pageblock_order / 2 ||
2336 start_mt == MIGRATE_RECLAIMABLE ||
2337 start_mt == MIGRATE_UNMOVABLE ||
2338 page_group_by_mobility_disabled)
2344 static inline void boost_watermark(struct zone *zone)
2346 unsigned long max_boost;
2348 if (!watermark_boost_factor)
2351 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2352 watermark_boost_factor, 10000);
2355 * high watermark may be uninitialised if fragmentation occurs
2356 * very early in boot so do not boost. We do not fall
2357 * through and boost by pageblock_nr_pages as failing
2358 * allocations that early means that reclaim is not going
2359 * to help and it may even be impossible to reclaim the
2360 * boosted watermark resulting in a hang.
2365 max_boost = max(pageblock_nr_pages, max_boost);
2367 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2372 * This function implements actual steal behaviour. If order is large enough,
2373 * we can steal whole pageblock. If not, we first move freepages in this
2374 * pageblock to our migratetype and determine how many already-allocated pages
2375 * are there in the pageblock with a compatible migratetype. If at least half
2376 * of pages are free or compatible, we can change migratetype of the pageblock
2377 * itself, so pages freed in the future will be put on the correct free list.
2379 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2380 unsigned int alloc_flags, int start_type, bool whole_block)
2382 unsigned int current_order = page_order(page);
2383 struct free_area *area;
2384 int free_pages, movable_pages, alike_pages;
2387 old_block_type = get_pageblock_migratetype(page);
2390 * This can happen due to races and we want to prevent broken
2391 * highatomic accounting.
2393 if (is_migrate_highatomic(old_block_type))
2396 /* Take ownership for orders >= pageblock_order */
2397 if (current_order >= pageblock_order) {
2398 change_pageblock_range(page, current_order, start_type);
2403 * Boost watermarks to increase reclaim pressure to reduce the
2404 * likelihood of future fallbacks. Wake kswapd now as the node
2405 * may be balanced overall and kswapd will not wake naturally.
2407 boost_watermark(zone);
2408 if (alloc_flags & ALLOC_KSWAPD)
2409 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2411 /* We are not allowed to try stealing from the whole block */
2415 free_pages = move_freepages_block(zone, page, start_type,
2418 * Determine how many pages are compatible with our allocation.
2419 * For movable allocation, it's the number of movable pages which
2420 * we just obtained. For other types it's a bit more tricky.
2422 if (start_type == MIGRATE_MOVABLE) {
2423 alike_pages = movable_pages;
2426 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2427 * to MOVABLE pageblock, consider all non-movable pages as
2428 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2429 * vice versa, be conservative since we can't distinguish the
2430 * exact migratetype of non-movable pages.
2432 if (old_block_type == MIGRATE_MOVABLE)
2433 alike_pages = pageblock_nr_pages
2434 - (free_pages + movable_pages);
2439 /* moving whole block can fail due to zone boundary conditions */
2444 * If a sufficient number of pages in the block are either free or of
2445 * comparable migratability as our allocation, claim the whole block.
2447 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2448 page_group_by_mobility_disabled)
2449 set_pageblock_migratetype(page, start_type);
2454 area = &zone->free_area[current_order];
2455 move_to_free_area(page, area, start_type);
2459 * Check whether there is a suitable fallback freepage with requested order.
2460 * If only_stealable is true, this function returns fallback_mt only if
2461 * we can steal other freepages all together. This would help to reduce
2462 * fragmentation due to mixed migratetype pages in one pageblock.
2464 int find_suitable_fallback(struct free_area *area, unsigned int order,
2465 int migratetype, bool only_stealable, bool *can_steal)
2470 if (area->nr_free == 0)
2475 fallback_mt = fallbacks[migratetype][i];
2476 if (fallback_mt == MIGRATE_TYPES)
2479 if (free_area_empty(area, fallback_mt))
2482 if (can_steal_fallback(order, migratetype))
2485 if (!only_stealable)
2496 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2497 * there are no empty page blocks that contain a page with a suitable order
2499 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2500 unsigned int alloc_order)
2503 unsigned long max_managed, flags;
2506 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2507 * Check is race-prone but harmless.
2509 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2510 if (zone->nr_reserved_highatomic >= max_managed)
2513 spin_lock_irqsave(&zone->lock, flags);
2515 /* Recheck the nr_reserved_highatomic limit under the lock */
2516 if (zone->nr_reserved_highatomic >= max_managed)
2520 mt = get_pageblock_migratetype(page);
2521 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2522 && !is_migrate_cma(mt)) {
2523 zone->nr_reserved_highatomic += pageblock_nr_pages;
2524 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2525 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2529 spin_unlock_irqrestore(&zone->lock, flags);
2533 * Used when an allocation is about to fail under memory pressure. This
2534 * potentially hurts the reliability of high-order allocations when under
2535 * intense memory pressure but failed atomic allocations should be easier
2536 * to recover from than an OOM.
2538 * If @force is true, try to unreserve a pageblock even though highatomic
2539 * pageblock is exhausted.
2541 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2544 struct zonelist *zonelist = ac->zonelist;
2545 unsigned long flags;
2552 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2555 * Preserve at least one pageblock unless memory pressure
2558 if (!force && zone->nr_reserved_highatomic <=
2562 spin_lock_irqsave(&zone->lock, flags);
2563 for (order = 0; order < MAX_ORDER; order++) {
2564 struct free_area *area = &(zone->free_area[order]);
2566 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2571 * In page freeing path, migratetype change is racy so
2572 * we can counter several free pages in a pageblock
2573 * in this loop althoug we changed the pageblock type
2574 * from highatomic to ac->migratetype. So we should
2575 * adjust the count once.
2577 if (is_migrate_highatomic_page(page)) {
2579 * It should never happen but changes to
2580 * locking could inadvertently allow a per-cpu
2581 * drain to add pages to MIGRATE_HIGHATOMIC
2582 * while unreserving so be safe and watch for
2585 zone->nr_reserved_highatomic -= min(
2587 zone->nr_reserved_highatomic);
2591 * Convert to ac->migratetype and avoid the normal
2592 * pageblock stealing heuristics. Minimally, the caller
2593 * is doing the work and needs the pages. More
2594 * importantly, if the block was always converted to
2595 * MIGRATE_UNMOVABLE or another type then the number
2596 * of pageblocks that cannot be completely freed
2599 set_pageblock_migratetype(page, ac->migratetype);
2600 ret = move_freepages_block(zone, page, ac->migratetype,
2603 spin_unlock_irqrestore(&zone->lock, flags);
2607 spin_unlock_irqrestore(&zone->lock, flags);
2614 * Try finding a free buddy page on the fallback list and put it on the free
2615 * list of requested migratetype, possibly along with other pages from the same
2616 * block, depending on fragmentation avoidance heuristics. Returns true if
2617 * fallback was found so that __rmqueue_smallest() can grab it.
2619 * The use of signed ints for order and current_order is a deliberate
2620 * deviation from the rest of this file, to make the for loop
2621 * condition simpler.
2623 static __always_inline bool
2624 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2625 unsigned int alloc_flags)
2627 struct free_area *area;
2629 int min_order = order;
2635 * Do not steal pages from freelists belonging to other pageblocks
2636 * i.e. orders < pageblock_order. If there are no local zones free,
2637 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2639 if (alloc_flags & ALLOC_NOFRAGMENT)
2640 min_order = pageblock_order;
2643 * Find the largest available free page in the other list. This roughly
2644 * approximates finding the pageblock with the most free pages, which
2645 * would be too costly to do exactly.
2647 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2649 area = &(zone->free_area[current_order]);
2650 fallback_mt = find_suitable_fallback(area, current_order,
2651 start_migratetype, false, &can_steal);
2652 if (fallback_mt == -1)
2656 * We cannot steal all free pages from the pageblock and the
2657 * requested migratetype is movable. In that case it's better to
2658 * steal and split the smallest available page instead of the
2659 * largest available page, because even if the next movable
2660 * allocation falls back into a different pageblock than this
2661 * one, it won't cause permanent fragmentation.
2663 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2664 && current_order > order)
2673 for (current_order = order; current_order < MAX_ORDER;
2675 area = &(zone->free_area[current_order]);
2676 fallback_mt = find_suitable_fallback(area, current_order,
2677 start_migratetype, false, &can_steal);
2678 if (fallback_mt != -1)
2683 * This should not happen - we already found a suitable fallback
2684 * when looking for the largest page.
2686 VM_BUG_ON(current_order == MAX_ORDER);
2689 page = get_page_from_free_area(area, fallback_mt);
2691 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2694 trace_mm_page_alloc_extfrag(page, order, current_order,
2695 start_migratetype, fallback_mt);
2702 * Do the hard work of removing an element from the buddy allocator.
2703 * Call me with the zone->lock already held.
2705 static __always_inline struct page *
2706 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2707 unsigned int alloc_flags)
2712 page = __rmqueue_smallest(zone, order, migratetype);
2713 if (unlikely(!page)) {
2714 if (migratetype == MIGRATE_MOVABLE)
2715 page = __rmqueue_cma_fallback(zone, order);
2717 if (!page && __rmqueue_fallback(zone, order, migratetype,
2722 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2727 * Obtain a specified number of elements from the buddy allocator, all under
2728 * a single hold of the lock, for efficiency. Add them to the supplied list.
2729 * Returns the number of new pages which were placed at *list.
2731 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2732 unsigned long count, struct list_head *list,
2733 int migratetype, unsigned int alloc_flags)
2737 spin_lock(&zone->lock);
2738 for (i = 0; i < count; ++i) {
2739 struct page *page = __rmqueue(zone, order, migratetype,
2741 if (unlikely(page == NULL))
2744 if (unlikely(check_pcp_refill(page)))
2748 * Split buddy pages returned by expand() are received here in
2749 * physical page order. The page is added to the tail of
2750 * caller's list. From the callers perspective, the linked list
2751 * is ordered by page number under some conditions. This is
2752 * useful for IO devices that can forward direction from the
2753 * head, thus also in the physical page order. This is useful
2754 * for IO devices that can merge IO requests if the physical
2755 * pages are ordered properly.
2757 list_add_tail(&page->lru, list);
2759 if (is_migrate_cma(get_pcppage_migratetype(page)))
2760 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2765 * i pages were removed from the buddy list even if some leak due
2766 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2767 * on i. Do not confuse with 'alloced' which is the number of
2768 * pages added to the pcp list.
2770 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2771 spin_unlock(&zone->lock);
2777 * Called from the vmstat counter updater to drain pagesets of this
2778 * currently executing processor on remote nodes after they have
2781 * Note that this function must be called with the thread pinned to
2782 * a single processor.
2784 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2786 unsigned long flags;
2787 int to_drain, batch;
2789 local_irq_save(flags);
2790 batch = READ_ONCE(pcp->batch);
2791 to_drain = min(pcp->count, batch);
2793 free_pcppages_bulk(zone, to_drain, pcp);
2794 local_irq_restore(flags);
2799 * Drain pcplists of the indicated processor and zone.
2801 * The processor must either be the current processor and the
2802 * thread pinned to the current processor or a processor that
2805 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2807 unsigned long flags;
2808 struct per_cpu_pageset *pset;
2809 struct per_cpu_pages *pcp;
2811 local_irq_save(flags);
2812 pset = per_cpu_ptr(zone->pageset, cpu);
2816 free_pcppages_bulk(zone, pcp->count, pcp);
2817 local_irq_restore(flags);
2821 * Drain pcplists of all zones on the indicated processor.
2823 * The processor must either be the current processor and the
2824 * thread pinned to the current processor or a processor that
2827 static void drain_pages(unsigned int cpu)
2831 for_each_populated_zone(zone) {
2832 drain_pages_zone(cpu, zone);
2837 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2839 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2840 * the single zone's pages.
2842 void drain_local_pages(struct zone *zone)
2844 int cpu = smp_processor_id();
2847 drain_pages_zone(cpu, zone);
2852 static void drain_local_pages_wq(struct work_struct *work)
2854 struct pcpu_drain *drain;
2856 drain = container_of(work, struct pcpu_drain, work);
2859 * drain_all_pages doesn't use proper cpu hotplug protection so
2860 * we can race with cpu offline when the WQ can move this from
2861 * a cpu pinned worker to an unbound one. We can operate on a different
2862 * cpu which is allright but we also have to make sure to not move to
2866 drain_local_pages(drain->zone);
2871 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2873 * When zone parameter is non-NULL, spill just the single zone's pages.
2875 * Note that this can be extremely slow as the draining happens in a workqueue.
2877 void drain_all_pages(struct zone *zone)
2882 * Allocate in the BSS so we wont require allocation in
2883 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2885 static cpumask_t cpus_with_pcps;
2888 * Make sure nobody triggers this path before mm_percpu_wq is fully
2891 if (WARN_ON_ONCE(!mm_percpu_wq))
2895 * Do not drain if one is already in progress unless it's specific to
2896 * a zone. Such callers are primarily CMA and memory hotplug and need
2897 * the drain to be complete when the call returns.
2899 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2902 mutex_lock(&pcpu_drain_mutex);
2906 * We don't care about racing with CPU hotplug event
2907 * as offline notification will cause the notified
2908 * cpu to drain that CPU pcps and on_each_cpu_mask
2909 * disables preemption as part of its processing
2911 for_each_online_cpu(cpu) {
2912 struct per_cpu_pageset *pcp;
2914 bool has_pcps = false;
2917 pcp = per_cpu_ptr(zone->pageset, cpu);
2921 for_each_populated_zone(z) {
2922 pcp = per_cpu_ptr(z->pageset, cpu);
2923 if (pcp->pcp.count) {
2931 cpumask_set_cpu(cpu, &cpus_with_pcps);
2933 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2936 for_each_cpu(cpu, &cpus_with_pcps) {
2937 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
2940 INIT_WORK(&drain->work, drain_local_pages_wq);
2941 queue_work_on(cpu, mm_percpu_wq, &drain->work);
2943 for_each_cpu(cpu, &cpus_with_pcps)
2944 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
2946 mutex_unlock(&pcpu_drain_mutex);
2949 #ifdef CONFIG_HIBERNATION
2952 * Touch the watchdog for every WD_PAGE_COUNT pages.
2954 #define WD_PAGE_COUNT (128*1024)
2956 void mark_free_pages(struct zone *zone)
2958 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2959 unsigned long flags;
2960 unsigned int order, t;
2963 if (zone_is_empty(zone))
2966 spin_lock_irqsave(&zone->lock, flags);
2968 max_zone_pfn = zone_end_pfn(zone);
2969 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2970 if (pfn_valid(pfn)) {
2971 page = pfn_to_page(pfn);
2973 if (!--page_count) {
2974 touch_nmi_watchdog();
2975 page_count = WD_PAGE_COUNT;
2978 if (page_zone(page) != zone)
2981 if (!swsusp_page_is_forbidden(page))
2982 swsusp_unset_page_free(page);
2985 for_each_migratetype_order(order, t) {
2986 list_for_each_entry(page,
2987 &zone->free_area[order].free_list[t], lru) {
2990 pfn = page_to_pfn(page);
2991 for (i = 0; i < (1UL << order); i++) {
2992 if (!--page_count) {
2993 touch_nmi_watchdog();
2994 page_count = WD_PAGE_COUNT;
2996 swsusp_set_page_free(pfn_to_page(pfn + i));
3000 spin_unlock_irqrestore(&zone->lock, flags);
3002 #endif /* CONFIG_PM */
3004 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3008 if (!free_pcp_prepare(page))
3011 migratetype = get_pfnblock_migratetype(page, pfn);
3012 set_pcppage_migratetype(page, migratetype);
3016 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3018 struct zone *zone = page_zone(page);
3019 struct per_cpu_pages *pcp;
3022 migratetype = get_pcppage_migratetype(page);
3023 __count_vm_event(PGFREE);
3026 * We only track unmovable, reclaimable and movable on pcp lists.
3027 * Free ISOLATE pages back to the allocator because they are being
3028 * offlined but treat HIGHATOMIC as movable pages so we can get those
3029 * areas back if necessary. Otherwise, we may have to free
3030 * excessively into the page allocator
3032 if (migratetype >= MIGRATE_PCPTYPES) {
3033 if (unlikely(is_migrate_isolate(migratetype))) {
3034 free_one_page(zone, page, pfn, 0, migratetype);
3037 migratetype = MIGRATE_MOVABLE;
3040 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3041 list_add(&page->lru, &pcp->lists[migratetype]);
3043 if (pcp->count >= pcp->high) {
3044 unsigned long batch = READ_ONCE(pcp->batch);
3045 free_pcppages_bulk(zone, batch, pcp);
3050 * Free a 0-order page
3052 void free_unref_page(struct page *page)
3054 unsigned long flags;
3055 unsigned long pfn = page_to_pfn(page);
3057 if (!free_unref_page_prepare(page, pfn))
3060 local_irq_save(flags);
3061 free_unref_page_commit(page, pfn);
3062 local_irq_restore(flags);
3066 * Free a list of 0-order pages
3068 void free_unref_page_list(struct list_head *list)
3070 struct page *page, *next;
3071 unsigned long flags, pfn;
3072 int batch_count = 0;
3074 /* Prepare pages for freeing */
3075 list_for_each_entry_safe(page, next, list, lru) {
3076 pfn = page_to_pfn(page);
3077 if (!free_unref_page_prepare(page, pfn))
3078 list_del(&page->lru);
3079 set_page_private(page, pfn);
3082 local_irq_save(flags);
3083 list_for_each_entry_safe(page, next, list, lru) {
3084 unsigned long pfn = page_private(page);
3086 set_page_private(page, 0);
3087 trace_mm_page_free_batched(page);
3088 free_unref_page_commit(page, pfn);
3091 * Guard against excessive IRQ disabled times when we get
3092 * a large list of pages to free.
3094 if (++batch_count == SWAP_CLUSTER_MAX) {
3095 local_irq_restore(flags);
3097 local_irq_save(flags);
3100 local_irq_restore(flags);
3104 * split_page takes a non-compound higher-order page, and splits it into
3105 * n (1<<order) sub-pages: page[0..n]
3106 * Each sub-page must be freed individually.
3108 * Note: this is probably too low level an operation for use in drivers.
3109 * Please consult with lkml before using this in your driver.
3111 void split_page(struct page *page, unsigned int order)
3115 VM_BUG_ON_PAGE(PageCompound(page), page);
3116 VM_BUG_ON_PAGE(!page_count(page), page);
3118 for (i = 1; i < (1 << order); i++)
3119 set_page_refcounted(page + i);
3120 split_page_owner(page, order);
3122 EXPORT_SYMBOL_GPL(split_page);
3124 int __isolate_free_page(struct page *page, unsigned int order)
3126 struct free_area *area = &page_zone(page)->free_area[order];
3127 unsigned long watermark;
3131 BUG_ON(!PageBuddy(page));
3133 zone = page_zone(page);
3134 mt = get_pageblock_migratetype(page);
3136 if (!is_migrate_isolate(mt)) {
3138 * Obey watermarks as if the page was being allocated. We can
3139 * emulate a high-order watermark check with a raised order-0
3140 * watermark, because we already know our high-order page
3143 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3144 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3147 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3150 /* Remove page from free list */
3152 del_page_from_free_area(page, area);
3155 * Set the pageblock if the isolated page is at least half of a
3158 if (order >= pageblock_order - 1) {
3159 struct page *endpage = page + (1 << order) - 1;
3160 for (; page < endpage; page += pageblock_nr_pages) {
3161 int mt = get_pageblock_migratetype(page);
3162 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3163 && !is_migrate_highatomic(mt))
3164 set_pageblock_migratetype(page,
3170 return 1UL << order;
3174 * Update NUMA hit/miss statistics
3176 * Must be called with interrupts disabled.
3178 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3181 enum numa_stat_item local_stat = NUMA_LOCAL;
3183 /* skip numa counters update if numa stats is disabled */
3184 if (!static_branch_likely(&vm_numa_stat_key))
3187 if (zone_to_nid(z) != numa_node_id())
3188 local_stat = NUMA_OTHER;
3190 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3191 __inc_numa_state(z, NUMA_HIT);
3193 __inc_numa_state(z, NUMA_MISS);
3194 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3196 __inc_numa_state(z, local_stat);
3200 /* Remove page from the per-cpu list, caller must protect the list */
3201 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3202 unsigned int alloc_flags,
3203 struct per_cpu_pages *pcp,
3204 struct list_head *list)
3209 if (list_empty(list)) {
3210 pcp->count += rmqueue_bulk(zone, 0,
3212 migratetype, alloc_flags);
3213 if (unlikely(list_empty(list)))
3217 page = list_first_entry(list, struct page, lru);
3218 list_del(&page->lru);
3220 } while (check_new_pcp(page));
3225 /* Lock and remove page from the per-cpu list */
3226 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3227 struct zone *zone, gfp_t gfp_flags,
3228 int migratetype, unsigned int alloc_flags)
3230 struct per_cpu_pages *pcp;
3231 struct list_head *list;
3233 unsigned long flags;
3235 local_irq_save(flags);
3236 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3237 list = &pcp->lists[migratetype];
3238 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3240 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3241 zone_statistics(preferred_zone, zone);
3243 local_irq_restore(flags);
3248 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3251 struct page *rmqueue(struct zone *preferred_zone,
3252 struct zone *zone, unsigned int order,
3253 gfp_t gfp_flags, unsigned int alloc_flags,
3256 unsigned long flags;
3259 if (likely(order == 0)) {
3260 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3261 migratetype, alloc_flags);
3266 * We most definitely don't want callers attempting to
3267 * allocate greater than order-1 page units with __GFP_NOFAIL.
3269 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3270 spin_lock_irqsave(&zone->lock, flags);
3274 if (alloc_flags & ALLOC_HARDER) {
3275 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3277 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3280 page = __rmqueue(zone, order, migratetype, alloc_flags);
3281 } while (page && check_new_pages(page, order));
3282 spin_unlock(&zone->lock);
3285 __mod_zone_freepage_state(zone, -(1 << order),
3286 get_pcppage_migratetype(page));
3288 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3289 zone_statistics(preferred_zone, zone);
3290 local_irq_restore(flags);
3293 /* Separate test+clear to avoid unnecessary atomics */
3294 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3295 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3296 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3299 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3303 local_irq_restore(flags);
3307 #ifdef CONFIG_FAIL_PAGE_ALLOC
3310 struct fault_attr attr;
3312 bool ignore_gfp_highmem;
3313 bool ignore_gfp_reclaim;
3315 } fail_page_alloc = {
3316 .attr = FAULT_ATTR_INITIALIZER,
3317 .ignore_gfp_reclaim = true,
3318 .ignore_gfp_highmem = true,
3322 static int __init setup_fail_page_alloc(char *str)
3324 return setup_fault_attr(&fail_page_alloc.attr, str);
3326 __setup("fail_page_alloc=", setup_fail_page_alloc);
3328 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3330 if (order < fail_page_alloc.min_order)
3332 if (gfp_mask & __GFP_NOFAIL)
3334 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3336 if (fail_page_alloc.ignore_gfp_reclaim &&
3337 (gfp_mask & __GFP_DIRECT_RECLAIM))
3340 return should_fail(&fail_page_alloc.attr, 1 << order);
3343 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3345 static int __init fail_page_alloc_debugfs(void)
3347 umode_t mode = S_IFREG | 0600;
3350 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3351 &fail_page_alloc.attr);
3353 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3354 &fail_page_alloc.ignore_gfp_reclaim);
3355 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3356 &fail_page_alloc.ignore_gfp_highmem);
3357 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3362 late_initcall(fail_page_alloc_debugfs);
3364 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3366 #else /* CONFIG_FAIL_PAGE_ALLOC */
3368 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3373 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3375 static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3377 return __should_fail_alloc_page(gfp_mask, order);
3379 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3382 * Return true if free base pages are above 'mark'. For high-order checks it
3383 * will return true of the order-0 watermark is reached and there is at least
3384 * one free page of a suitable size. Checking now avoids taking the zone lock
3385 * to check in the allocation paths if no pages are free.
3387 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3388 int classzone_idx, unsigned int alloc_flags,
3393 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3395 /* free_pages may go negative - that's OK */
3396 free_pages -= (1 << order) - 1;
3398 if (alloc_flags & ALLOC_HIGH)
3402 * If the caller does not have rights to ALLOC_HARDER then subtract
3403 * the high-atomic reserves. This will over-estimate the size of the
3404 * atomic reserve but it avoids a search.
3406 if (likely(!alloc_harder)) {
3407 free_pages -= z->nr_reserved_highatomic;
3410 * OOM victims can try even harder than normal ALLOC_HARDER
3411 * users on the grounds that it's definitely going to be in
3412 * the exit path shortly and free memory. Any allocation it
3413 * makes during the free path will be small and short-lived.
3415 if (alloc_flags & ALLOC_OOM)
3423 /* If allocation can't use CMA areas don't use free CMA pages */
3424 if (!(alloc_flags & ALLOC_CMA))
3425 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3429 * Check watermarks for an order-0 allocation request. If these
3430 * are not met, then a high-order request also cannot go ahead
3431 * even if a suitable page happened to be free.
3433 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3436 /* If this is an order-0 request then the watermark is fine */
3440 /* For a high-order request, check at least one suitable page is free */
3441 for (o = order; o < MAX_ORDER; o++) {
3442 struct free_area *area = &z->free_area[o];
3448 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3449 if (!free_area_empty(area, mt))
3454 if ((alloc_flags & ALLOC_CMA) &&
3455 !free_area_empty(area, MIGRATE_CMA)) {
3460 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3466 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3467 int classzone_idx, unsigned int alloc_flags)
3469 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3470 zone_page_state(z, NR_FREE_PAGES));
3473 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3474 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3476 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3480 /* If allocation can't use CMA areas don't use free CMA pages */
3481 if (!(alloc_flags & ALLOC_CMA))
3482 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3486 * Fast check for order-0 only. If this fails then the reserves
3487 * need to be calculated. There is a corner case where the check
3488 * passes but only the high-order atomic reserve are free. If
3489 * the caller is !atomic then it'll uselessly search the free
3490 * list. That corner case is then slower but it is harmless.
3492 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3495 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3499 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3500 unsigned long mark, int classzone_idx)
3502 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3504 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3505 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3507 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3512 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3514 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3515 node_reclaim_distance;
3517 #else /* CONFIG_NUMA */
3518 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3522 #endif /* CONFIG_NUMA */
3525 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3526 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3527 * premature use of a lower zone may cause lowmem pressure problems that
3528 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3529 * probably too small. It only makes sense to spread allocations to avoid
3530 * fragmentation between the Normal and DMA32 zones.
3532 static inline unsigned int
3533 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3535 unsigned int alloc_flags = 0;
3537 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3538 alloc_flags |= ALLOC_KSWAPD;
3540 #ifdef CONFIG_ZONE_DMA32
3544 if (zone_idx(zone) != ZONE_NORMAL)
3548 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3549 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3550 * on UMA that if Normal is populated then so is DMA32.
3552 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3553 if (nr_online_nodes > 1 && !populated_zone(--zone))
3556 alloc_flags |= ALLOC_NOFRAGMENT;
3557 #endif /* CONFIG_ZONE_DMA32 */
3562 * get_page_from_freelist goes through the zonelist trying to allocate
3565 static struct page *
3566 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3567 const struct alloc_context *ac)
3571 struct pglist_data *last_pgdat_dirty_limit = NULL;
3576 * Scan zonelist, looking for a zone with enough free.
3577 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3579 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3580 z = ac->preferred_zoneref;
3581 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3586 if (cpusets_enabled() &&
3587 (alloc_flags & ALLOC_CPUSET) &&
3588 !__cpuset_zone_allowed(zone, gfp_mask))
3591 * When allocating a page cache page for writing, we
3592 * want to get it from a node that is within its dirty
3593 * limit, such that no single node holds more than its
3594 * proportional share of globally allowed dirty pages.
3595 * The dirty limits take into account the node's
3596 * lowmem reserves and high watermark so that kswapd
3597 * should be able to balance it without having to
3598 * write pages from its LRU list.
3600 * XXX: For now, allow allocations to potentially
3601 * exceed the per-node dirty limit in the slowpath
3602 * (spread_dirty_pages unset) before going into reclaim,
3603 * which is important when on a NUMA setup the allowed
3604 * nodes are together not big enough to reach the
3605 * global limit. The proper fix for these situations
3606 * will require awareness of nodes in the
3607 * dirty-throttling and the flusher threads.
3609 if (ac->spread_dirty_pages) {
3610 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3613 if (!node_dirty_ok(zone->zone_pgdat)) {
3614 last_pgdat_dirty_limit = zone->zone_pgdat;
3619 if (no_fallback && nr_online_nodes > 1 &&
3620 zone != ac->preferred_zoneref->zone) {
3624 * If moving to a remote node, retry but allow
3625 * fragmenting fallbacks. Locality is more important
3626 * than fragmentation avoidance.
3628 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3629 if (zone_to_nid(zone) != local_nid) {
3630 alloc_flags &= ~ALLOC_NOFRAGMENT;
3635 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3636 if (!zone_watermark_fast(zone, order, mark,
3637 ac_classzone_idx(ac), alloc_flags)) {
3640 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3642 * Watermark failed for this zone, but see if we can
3643 * grow this zone if it contains deferred pages.
3645 if (static_branch_unlikely(&deferred_pages)) {
3646 if (_deferred_grow_zone(zone, order))
3650 /* Checked here to keep the fast path fast */
3651 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3652 if (alloc_flags & ALLOC_NO_WATERMARKS)
3655 if (node_reclaim_mode == 0 ||
3656 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3659 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3661 case NODE_RECLAIM_NOSCAN:
3664 case NODE_RECLAIM_FULL:
3665 /* scanned but unreclaimable */
3668 /* did we reclaim enough */
3669 if (zone_watermark_ok(zone, order, mark,
3670 ac_classzone_idx(ac), alloc_flags))
3678 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3679 gfp_mask, alloc_flags, ac->migratetype);
3681 prep_new_page(page, order, gfp_mask, alloc_flags);
3684 * If this is a high-order atomic allocation then check
3685 * if the pageblock should be reserved for the future
3687 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3688 reserve_highatomic_pageblock(page, zone, order);
3692 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3693 /* Try again if zone has deferred pages */
3694 if (static_branch_unlikely(&deferred_pages)) {
3695 if (_deferred_grow_zone(zone, order))
3703 * It's possible on a UMA machine to get through all zones that are
3704 * fragmented. If avoiding fragmentation, reset and try again.
3707 alloc_flags &= ~ALLOC_NOFRAGMENT;
3714 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3716 unsigned int filter = SHOW_MEM_FILTER_NODES;
3717 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3719 if (!__ratelimit(&show_mem_rs))
3723 * This documents exceptions given to allocations in certain
3724 * contexts that are allowed to allocate outside current's set
3727 if (!(gfp_mask & __GFP_NOMEMALLOC))
3728 if (tsk_is_oom_victim(current) ||
3729 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3730 filter &= ~SHOW_MEM_FILTER_NODES;
3731 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3732 filter &= ~SHOW_MEM_FILTER_NODES;
3734 show_mem(filter, nodemask);
3737 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3739 struct va_format vaf;
3741 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3742 DEFAULT_RATELIMIT_BURST);
3744 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3747 va_start(args, fmt);
3750 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3751 current->comm, &vaf, gfp_mask, &gfp_mask,
3752 nodemask_pr_args(nodemask));
3755 cpuset_print_current_mems_allowed();
3758 warn_alloc_show_mem(gfp_mask, nodemask);
3761 static inline struct page *
3762 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3763 unsigned int alloc_flags,
3764 const struct alloc_context *ac)
3768 page = get_page_from_freelist(gfp_mask, order,
3769 alloc_flags|ALLOC_CPUSET, ac);
3771 * fallback to ignore cpuset restriction if our nodes
3775 page = get_page_from_freelist(gfp_mask, order,
3781 static inline struct page *
3782 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3783 const struct alloc_context *ac, unsigned long *did_some_progress)
3785 struct oom_control oc = {
3786 .zonelist = ac->zonelist,
3787 .nodemask = ac->nodemask,
3789 .gfp_mask = gfp_mask,
3794 *did_some_progress = 0;
3797 * Acquire the oom lock. If that fails, somebody else is
3798 * making progress for us.
3800 if (!mutex_trylock(&oom_lock)) {
3801 *did_some_progress = 1;
3802 schedule_timeout_uninterruptible(1);
3807 * Go through the zonelist yet one more time, keep very high watermark
3808 * here, this is only to catch a parallel oom killing, we must fail if
3809 * we're still under heavy pressure. But make sure that this reclaim
3810 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3811 * allocation which will never fail due to oom_lock already held.
3813 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3814 ~__GFP_DIRECT_RECLAIM, order,
3815 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3819 /* Coredumps can quickly deplete all memory reserves */
3820 if (current->flags & PF_DUMPCORE)
3822 /* The OOM killer will not help higher order allocs */
3823 if (order > PAGE_ALLOC_COSTLY_ORDER)
3826 * We have already exhausted all our reclaim opportunities without any
3827 * success so it is time to admit defeat. We will skip the OOM killer
3828 * because it is very likely that the caller has a more reasonable
3829 * fallback than shooting a random task.
3831 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3833 /* The OOM killer does not needlessly kill tasks for lowmem */
3834 if (ac->high_zoneidx < ZONE_NORMAL)
3836 if (pm_suspended_storage())
3839 * XXX: GFP_NOFS allocations should rather fail than rely on
3840 * other request to make a forward progress.
3841 * We are in an unfortunate situation where out_of_memory cannot
3842 * do much for this context but let's try it to at least get
3843 * access to memory reserved if the current task is killed (see
3844 * out_of_memory). Once filesystems are ready to handle allocation
3845 * failures more gracefully we should just bail out here.
3848 /* The OOM killer may not free memory on a specific node */
3849 if (gfp_mask & __GFP_THISNODE)
3852 /* Exhausted what can be done so it's blame time */
3853 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3854 *did_some_progress = 1;
3857 * Help non-failing allocations by giving them access to memory
3860 if (gfp_mask & __GFP_NOFAIL)
3861 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3862 ALLOC_NO_WATERMARKS, ac);
3865 mutex_unlock(&oom_lock);
3870 * Maximum number of compaction retries wit a progress before OOM
3871 * killer is consider as the only way to move forward.
3873 #define MAX_COMPACT_RETRIES 16
3875 #ifdef CONFIG_COMPACTION
3876 /* Try memory compaction for high-order allocations before reclaim */
3877 static struct page *
3878 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3879 unsigned int alloc_flags, const struct alloc_context *ac,
3880 enum compact_priority prio, enum compact_result *compact_result)
3882 struct page *page = NULL;
3883 unsigned long pflags;
3884 unsigned int noreclaim_flag;
3889 psi_memstall_enter(&pflags);
3890 noreclaim_flag = memalloc_noreclaim_save();
3892 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3895 memalloc_noreclaim_restore(noreclaim_flag);
3896 psi_memstall_leave(&pflags);
3899 * At least in one zone compaction wasn't deferred or skipped, so let's
3900 * count a compaction stall
3902 count_vm_event(COMPACTSTALL);
3904 /* Prep a captured page if available */
3906 prep_new_page(page, order, gfp_mask, alloc_flags);
3908 /* Try get a page from the freelist if available */
3910 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3913 struct zone *zone = page_zone(page);
3915 zone->compact_blockskip_flush = false;
3916 compaction_defer_reset(zone, order, true);
3917 count_vm_event(COMPACTSUCCESS);
3922 * It's bad if compaction run occurs and fails. The most likely reason
3923 * is that pages exist, but not enough to satisfy watermarks.
3925 count_vm_event(COMPACTFAIL);
3933 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3934 enum compact_result compact_result,
3935 enum compact_priority *compact_priority,
3936 int *compaction_retries)
3938 int max_retries = MAX_COMPACT_RETRIES;
3941 int retries = *compaction_retries;
3942 enum compact_priority priority = *compact_priority;
3947 if (compaction_made_progress(compact_result))
3948 (*compaction_retries)++;
3951 * compaction considers all the zone as desperately out of memory
3952 * so it doesn't really make much sense to retry except when the
3953 * failure could be caused by insufficient priority
3955 if (compaction_failed(compact_result))
3956 goto check_priority;
3959 * compaction was skipped because there are not enough order-0 pages
3960 * to work with, so we retry only if it looks like reclaim can help.
3962 if (compaction_needs_reclaim(compact_result)) {
3963 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3968 * make sure the compaction wasn't deferred or didn't bail out early
3969 * due to locks contention before we declare that we should give up.
3970 * But the next retry should use a higher priority if allowed, so
3971 * we don't just keep bailing out endlessly.
3973 if (compaction_withdrawn(compact_result)) {
3974 goto check_priority;
3978 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3979 * costly ones because they are de facto nofail and invoke OOM
3980 * killer to move on while costly can fail and users are ready
3981 * to cope with that. 1/4 retries is rather arbitrary but we
3982 * would need much more detailed feedback from compaction to
3983 * make a better decision.
3985 if (order > PAGE_ALLOC_COSTLY_ORDER)
3987 if (*compaction_retries <= max_retries) {
3993 * Make sure there are attempts at the highest priority if we exhausted
3994 * all retries or failed at the lower priorities.
3997 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3998 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4000 if (*compact_priority > min_priority) {
4001 (*compact_priority)--;
4002 *compaction_retries = 0;
4006 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4010 static inline struct page *
4011 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4012 unsigned int alloc_flags, const struct alloc_context *ac,
4013 enum compact_priority prio, enum compact_result *compact_result)
4015 *compact_result = COMPACT_SKIPPED;
4020 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4021 enum compact_result compact_result,
4022 enum compact_priority *compact_priority,
4023 int *compaction_retries)
4028 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4032 * There are setups with compaction disabled which would prefer to loop
4033 * inside the allocator rather than hit the oom killer prematurely.
4034 * Let's give them a good hope and keep retrying while the order-0
4035 * watermarks are OK.
4037 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4039 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4040 ac_classzone_idx(ac), alloc_flags))
4045 #endif /* CONFIG_COMPACTION */
4047 #ifdef CONFIG_LOCKDEP
4048 static struct lockdep_map __fs_reclaim_map =
4049 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4051 static bool __need_fs_reclaim(gfp_t gfp_mask)
4053 gfp_mask = current_gfp_context(gfp_mask);
4055 /* no reclaim without waiting on it */
4056 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4059 /* this guy won't enter reclaim */
4060 if (current->flags & PF_MEMALLOC)
4063 /* We're only interested __GFP_FS allocations for now */
4064 if (!(gfp_mask & __GFP_FS))
4067 if (gfp_mask & __GFP_NOLOCKDEP)
4073 void __fs_reclaim_acquire(void)
4075 lock_map_acquire(&__fs_reclaim_map);
4078 void __fs_reclaim_release(void)
4080 lock_map_release(&__fs_reclaim_map);
4083 void fs_reclaim_acquire(gfp_t gfp_mask)
4085 if (__need_fs_reclaim(gfp_mask))
4086 __fs_reclaim_acquire();
4088 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4090 void fs_reclaim_release(gfp_t gfp_mask)
4092 if (__need_fs_reclaim(gfp_mask))
4093 __fs_reclaim_release();
4095 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4098 /* Perform direct synchronous page reclaim */
4100 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4101 const struct alloc_context *ac)
4104 unsigned int noreclaim_flag;
4105 unsigned long pflags;
4109 /* We now go into synchronous reclaim */
4110 cpuset_memory_pressure_bump();
4111 psi_memstall_enter(&pflags);
4112 fs_reclaim_acquire(gfp_mask);
4113 noreclaim_flag = memalloc_noreclaim_save();
4115 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4118 memalloc_noreclaim_restore(noreclaim_flag);
4119 fs_reclaim_release(gfp_mask);
4120 psi_memstall_leave(&pflags);
4127 /* The really slow allocator path where we enter direct reclaim */
4128 static inline struct page *
4129 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4130 unsigned int alloc_flags, const struct alloc_context *ac,
4131 unsigned long *did_some_progress)
4133 struct page *page = NULL;
4134 bool drained = false;
4136 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4137 if (unlikely(!(*did_some_progress)))
4141 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4144 * If an allocation failed after direct reclaim, it could be because
4145 * pages are pinned on the per-cpu lists or in high alloc reserves.
4146 * Shrink them them and try again
4148 if (!page && !drained) {
4149 unreserve_highatomic_pageblock(ac, false);
4150 drain_all_pages(NULL);
4158 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4159 const struct alloc_context *ac)
4163 pg_data_t *last_pgdat = NULL;
4164 enum zone_type high_zoneidx = ac->high_zoneidx;
4166 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
4168 if (last_pgdat != zone->zone_pgdat)
4169 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
4170 last_pgdat = zone->zone_pgdat;
4174 static inline unsigned int
4175 gfp_to_alloc_flags(gfp_t gfp_mask)
4177 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4179 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
4180 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4183 * The caller may dip into page reserves a bit more if the caller
4184 * cannot run direct reclaim, or if the caller has realtime scheduling
4185 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4186 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4188 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
4190 if (gfp_mask & __GFP_ATOMIC) {
4192 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4193 * if it can't schedule.
4195 if (!(gfp_mask & __GFP_NOMEMALLOC))
4196 alloc_flags |= ALLOC_HARDER;
4198 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4199 * comment for __cpuset_node_allowed().
4201 alloc_flags &= ~ALLOC_CPUSET;
4202 } else if (unlikely(rt_task(current)) && !in_interrupt())
4203 alloc_flags |= ALLOC_HARDER;
4205 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4206 alloc_flags |= ALLOC_KSWAPD;
4209 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4210 alloc_flags |= ALLOC_CMA;
4215 static bool oom_reserves_allowed(struct task_struct *tsk)
4217 if (!tsk_is_oom_victim(tsk))
4221 * !MMU doesn't have oom reaper so give access to memory reserves
4222 * only to the thread with TIF_MEMDIE set
4224 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4231 * Distinguish requests which really need access to full memory
4232 * reserves from oom victims which can live with a portion of it
4234 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4236 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4238 if (gfp_mask & __GFP_MEMALLOC)
4239 return ALLOC_NO_WATERMARKS;
4240 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4241 return ALLOC_NO_WATERMARKS;
4242 if (!in_interrupt()) {
4243 if (current->flags & PF_MEMALLOC)
4244 return ALLOC_NO_WATERMARKS;
4245 else if (oom_reserves_allowed(current))
4252 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4254 return !!__gfp_pfmemalloc_flags(gfp_mask);
4258 * Checks whether it makes sense to retry the reclaim to make a forward progress
4259 * for the given allocation request.
4261 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4262 * without success, or when we couldn't even meet the watermark if we
4263 * reclaimed all remaining pages on the LRU lists.
4265 * Returns true if a retry is viable or false to enter the oom path.
4268 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4269 struct alloc_context *ac, int alloc_flags,
4270 bool did_some_progress, int *no_progress_loops)
4277 * Costly allocations might have made a progress but this doesn't mean
4278 * their order will become available due to high fragmentation so
4279 * always increment the no progress counter for them
4281 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4282 *no_progress_loops = 0;
4284 (*no_progress_loops)++;
4287 * Make sure we converge to OOM if we cannot make any progress
4288 * several times in the row.
4290 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4291 /* Before OOM, exhaust highatomic_reserve */
4292 return unreserve_highatomic_pageblock(ac, true);
4296 * Keep reclaiming pages while there is a chance this will lead
4297 * somewhere. If none of the target zones can satisfy our allocation
4298 * request even if all reclaimable pages are considered then we are
4299 * screwed and have to go OOM.
4301 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4303 unsigned long available;
4304 unsigned long reclaimable;
4305 unsigned long min_wmark = min_wmark_pages(zone);
4308 available = reclaimable = zone_reclaimable_pages(zone);
4309 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4312 * Would the allocation succeed if we reclaimed all
4313 * reclaimable pages?
4315 wmark = __zone_watermark_ok(zone, order, min_wmark,
4316 ac_classzone_idx(ac), alloc_flags, available);
4317 trace_reclaim_retry_zone(z, order, reclaimable,
4318 available, min_wmark, *no_progress_loops, wmark);
4321 * If we didn't make any progress and have a lot of
4322 * dirty + writeback pages then we should wait for
4323 * an IO to complete to slow down the reclaim and
4324 * prevent from pre mature OOM
4326 if (!did_some_progress) {
4327 unsigned long write_pending;
4329 write_pending = zone_page_state_snapshot(zone,
4330 NR_ZONE_WRITE_PENDING);
4332 if (2 * write_pending > reclaimable) {
4333 congestion_wait(BLK_RW_ASYNC, HZ/10);
4345 * Memory allocation/reclaim might be called from a WQ context and the
4346 * current implementation of the WQ concurrency control doesn't
4347 * recognize that a particular WQ is congested if the worker thread is
4348 * looping without ever sleeping. Therefore we have to do a short sleep
4349 * here rather than calling cond_resched().
4351 if (current->flags & PF_WQ_WORKER)
4352 schedule_timeout_uninterruptible(1);
4359 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4362 * It's possible that cpuset's mems_allowed and the nodemask from
4363 * mempolicy don't intersect. This should be normally dealt with by
4364 * policy_nodemask(), but it's possible to race with cpuset update in
4365 * such a way the check therein was true, and then it became false
4366 * before we got our cpuset_mems_cookie here.
4367 * This assumes that for all allocations, ac->nodemask can come only
4368 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4369 * when it does not intersect with the cpuset restrictions) or the
4370 * caller can deal with a violated nodemask.
4372 if (cpusets_enabled() && ac->nodemask &&
4373 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4374 ac->nodemask = NULL;
4379 * When updating a task's mems_allowed or mempolicy nodemask, it is
4380 * possible to race with parallel threads in such a way that our
4381 * allocation can fail while the mask is being updated. If we are about
4382 * to fail, check if the cpuset changed during allocation and if so,
4385 if (read_mems_allowed_retry(cpuset_mems_cookie))
4391 static inline struct page *
4392 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4393 struct alloc_context *ac)
4395 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4396 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4397 struct page *page = NULL;
4398 unsigned int alloc_flags;
4399 unsigned long did_some_progress;
4400 enum compact_priority compact_priority;
4401 enum compact_result compact_result;
4402 int compaction_retries;
4403 int no_progress_loops;
4404 unsigned int cpuset_mems_cookie;
4408 * We also sanity check to catch abuse of atomic reserves being used by
4409 * callers that are not in atomic context.
4411 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4412 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4413 gfp_mask &= ~__GFP_ATOMIC;
4416 compaction_retries = 0;
4417 no_progress_loops = 0;
4418 compact_priority = DEF_COMPACT_PRIORITY;
4419 cpuset_mems_cookie = read_mems_allowed_begin();
4422 * The fast path uses conservative alloc_flags to succeed only until
4423 * kswapd needs to be woken up, and to avoid the cost of setting up
4424 * alloc_flags precisely. So we do that now.
4426 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4429 * We need to recalculate the starting point for the zonelist iterator
4430 * because we might have used different nodemask in the fast path, or
4431 * there was a cpuset modification and we are retrying - otherwise we
4432 * could end up iterating over non-eligible zones endlessly.
4434 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4435 ac->high_zoneidx, ac->nodemask);
4436 if (!ac->preferred_zoneref->zone)
4439 if (alloc_flags & ALLOC_KSWAPD)
4440 wake_all_kswapds(order, gfp_mask, ac);
4443 * The adjusted alloc_flags might result in immediate success, so try
4446 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4451 * For costly allocations, try direct compaction first, as it's likely
4452 * that we have enough base pages and don't need to reclaim. For non-
4453 * movable high-order allocations, do that as well, as compaction will
4454 * try prevent permanent fragmentation by migrating from blocks of the
4456 * Don't try this for allocations that are allowed to ignore
4457 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4459 if (can_direct_reclaim &&
4461 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4462 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4463 page = __alloc_pages_direct_compact(gfp_mask, order,
4465 INIT_COMPACT_PRIORITY,
4471 * Checks for costly allocations with __GFP_NORETRY, which
4472 * includes THP page fault allocations
4474 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4476 * If compaction is deferred for high-order allocations,
4477 * it is because sync compaction recently failed. If
4478 * this is the case and the caller requested a THP
4479 * allocation, we do not want to heavily disrupt the
4480 * system, so we fail the allocation instead of entering
4483 if (compact_result == COMPACT_DEFERRED)
4487 * Looks like reclaim/compaction is worth trying, but
4488 * sync compaction could be very expensive, so keep
4489 * using async compaction.
4491 compact_priority = INIT_COMPACT_PRIORITY;
4496 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4497 if (alloc_flags & ALLOC_KSWAPD)
4498 wake_all_kswapds(order, gfp_mask, ac);
4500 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4502 alloc_flags = reserve_flags;
4505 * Reset the nodemask and zonelist iterators if memory policies can be
4506 * ignored. These allocations are high priority and system rather than
4509 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4510 ac->nodemask = NULL;
4511 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4512 ac->high_zoneidx, ac->nodemask);
4515 /* Attempt with potentially adjusted zonelist and alloc_flags */
4516 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4520 /* Caller is not willing to reclaim, we can't balance anything */
4521 if (!can_direct_reclaim)
4524 /* Avoid recursion of direct reclaim */
4525 if (current->flags & PF_MEMALLOC)
4528 /* Try direct reclaim and then allocating */
4529 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4530 &did_some_progress);
4534 /* Try direct compaction and then allocating */
4535 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4536 compact_priority, &compact_result);
4540 /* Do not loop if specifically requested */
4541 if (gfp_mask & __GFP_NORETRY)
4545 * Do not retry costly high order allocations unless they are
4546 * __GFP_RETRY_MAYFAIL
4548 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4551 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4552 did_some_progress > 0, &no_progress_loops))
4556 * It doesn't make any sense to retry for the compaction if the order-0
4557 * reclaim is not able to make any progress because the current
4558 * implementation of the compaction depends on the sufficient amount
4559 * of free memory (see __compaction_suitable)
4561 if (did_some_progress > 0 &&
4562 should_compact_retry(ac, order, alloc_flags,
4563 compact_result, &compact_priority,
4564 &compaction_retries))
4568 /* Deal with possible cpuset update races before we start OOM killing */
4569 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4572 /* Reclaim has failed us, start killing things */
4573 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4577 /* Avoid allocations with no watermarks from looping endlessly */
4578 if (tsk_is_oom_victim(current) &&
4579 (alloc_flags == ALLOC_OOM ||
4580 (gfp_mask & __GFP_NOMEMALLOC)))
4583 /* Retry as long as the OOM killer is making progress */
4584 if (did_some_progress) {
4585 no_progress_loops = 0;
4590 /* Deal with possible cpuset update races before we fail */
4591 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4595 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4598 if (gfp_mask & __GFP_NOFAIL) {
4600 * All existing users of the __GFP_NOFAIL are blockable, so warn
4601 * of any new users that actually require GFP_NOWAIT
4603 if (WARN_ON_ONCE(!can_direct_reclaim))
4607 * PF_MEMALLOC request from this context is rather bizarre
4608 * because we cannot reclaim anything and only can loop waiting
4609 * for somebody to do a work for us
4611 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4614 * non failing costly orders are a hard requirement which we
4615 * are not prepared for much so let's warn about these users
4616 * so that we can identify them and convert them to something
4619 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4622 * Help non-failing allocations by giving them access to memory
4623 * reserves but do not use ALLOC_NO_WATERMARKS because this
4624 * could deplete whole memory reserves which would just make
4625 * the situation worse
4627 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4635 warn_alloc(gfp_mask, ac->nodemask,
4636 "page allocation failure: order:%u", order);
4641 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4642 int preferred_nid, nodemask_t *nodemask,
4643 struct alloc_context *ac, gfp_t *alloc_mask,
4644 unsigned int *alloc_flags)
4646 ac->high_zoneidx = gfp_zone(gfp_mask);
4647 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4648 ac->nodemask = nodemask;
4649 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4651 if (cpusets_enabled()) {
4652 *alloc_mask |= __GFP_HARDWALL;
4654 ac->nodemask = &cpuset_current_mems_allowed;
4656 *alloc_flags |= ALLOC_CPUSET;
4659 fs_reclaim_acquire(gfp_mask);
4660 fs_reclaim_release(gfp_mask);
4662 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4664 if (should_fail_alloc_page(gfp_mask, order))
4667 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4668 *alloc_flags |= ALLOC_CMA;
4673 /* Determine whether to spread dirty pages and what the first usable zone */
4674 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4676 /* Dirty zone balancing only done in the fast path */
4677 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4680 * The preferred zone is used for statistics but crucially it is
4681 * also used as the starting point for the zonelist iterator. It
4682 * may get reset for allocations that ignore memory policies.
4684 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4685 ac->high_zoneidx, ac->nodemask);
4689 * This is the 'heart' of the zoned buddy allocator.
4692 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4693 nodemask_t *nodemask)
4696 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4697 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4698 struct alloc_context ac = { };
4701 * There are several places where we assume that the order value is sane
4702 * so bail out early if the request is out of bound.
4704 if (unlikely(order >= MAX_ORDER)) {
4705 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4709 gfp_mask &= gfp_allowed_mask;
4710 alloc_mask = gfp_mask;
4711 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4714 finalise_ac(gfp_mask, &ac);
4717 * Forbid the first pass from falling back to types that fragment
4718 * memory until all local zones are considered.
4720 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4722 /* First allocation attempt */
4723 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4728 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4729 * resp. GFP_NOIO which has to be inherited for all allocation requests
4730 * from a particular context which has been marked by
4731 * memalloc_no{fs,io}_{save,restore}.
4733 alloc_mask = current_gfp_context(gfp_mask);
4734 ac.spread_dirty_pages = false;
4737 * Restore the original nodemask if it was potentially replaced with
4738 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4740 if (unlikely(ac.nodemask != nodemask))
4741 ac.nodemask = nodemask;
4743 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4746 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4747 unlikely(__memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4748 __free_pages(page, order);
4752 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4756 EXPORT_SYMBOL(__alloc_pages_nodemask);
4759 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4760 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4761 * you need to access high mem.
4763 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4767 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4770 return (unsigned long) page_address(page);
4772 EXPORT_SYMBOL(__get_free_pages);
4774 unsigned long get_zeroed_page(gfp_t gfp_mask)
4776 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4778 EXPORT_SYMBOL(get_zeroed_page);
4780 static inline void free_the_page(struct page *page, unsigned int order)
4782 if (order == 0) /* Via pcp? */
4783 free_unref_page(page);
4785 __free_pages_ok(page, order);
4788 void __free_pages(struct page *page, unsigned int order)
4790 if (put_page_testzero(page))
4791 free_the_page(page, order);
4793 EXPORT_SYMBOL(__free_pages);
4795 void free_pages(unsigned long addr, unsigned int order)
4798 VM_BUG_ON(!virt_addr_valid((void *)addr));
4799 __free_pages(virt_to_page((void *)addr), order);
4803 EXPORT_SYMBOL(free_pages);
4807 * An arbitrary-length arbitrary-offset area of memory which resides
4808 * within a 0 or higher order page. Multiple fragments within that page
4809 * are individually refcounted, in the page's reference counter.
4811 * The page_frag functions below provide a simple allocation framework for
4812 * page fragments. This is used by the network stack and network device
4813 * drivers to provide a backing region of memory for use as either an
4814 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4816 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4819 struct page *page = NULL;
4820 gfp_t gfp = gfp_mask;
4822 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4823 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4825 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4826 PAGE_FRAG_CACHE_MAX_ORDER);
4827 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4829 if (unlikely(!page))
4830 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4832 nc->va = page ? page_address(page) : NULL;
4837 void __page_frag_cache_drain(struct page *page, unsigned int count)
4839 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4841 if (page_ref_sub_and_test(page, count))
4842 free_the_page(page, compound_order(page));
4844 EXPORT_SYMBOL(__page_frag_cache_drain);
4846 void *page_frag_alloc(struct page_frag_cache *nc,
4847 unsigned int fragsz, gfp_t gfp_mask)
4849 unsigned int size = PAGE_SIZE;
4853 if (unlikely(!nc->va)) {
4855 page = __page_frag_cache_refill(nc, gfp_mask);
4859 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4860 /* if size can vary use size else just use PAGE_SIZE */
4863 /* Even if we own the page, we do not use atomic_set().
4864 * This would break get_page_unless_zero() users.
4866 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4868 /* reset page count bias and offset to start of new frag */
4869 nc->pfmemalloc = page_is_pfmemalloc(page);
4870 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4874 offset = nc->offset - fragsz;
4875 if (unlikely(offset < 0)) {
4876 page = virt_to_page(nc->va);
4878 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4881 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4882 /* if size can vary use size else just use PAGE_SIZE */
4885 /* OK, page count is 0, we can safely set it */
4886 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4888 /* reset page count bias and offset to start of new frag */
4889 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4890 offset = size - fragsz;
4894 nc->offset = offset;
4896 return nc->va + offset;
4898 EXPORT_SYMBOL(page_frag_alloc);
4901 * Frees a page fragment allocated out of either a compound or order 0 page.
4903 void page_frag_free(void *addr)
4905 struct page *page = virt_to_head_page(addr);
4907 if (unlikely(put_page_testzero(page)))
4908 free_the_page(page, compound_order(page));
4910 EXPORT_SYMBOL(page_frag_free);
4912 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4916 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4917 unsigned long used = addr + PAGE_ALIGN(size);
4919 split_page(virt_to_page((void *)addr), order);
4920 while (used < alloc_end) {
4925 return (void *)addr;
4929 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4930 * @size: the number of bytes to allocate
4931 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4933 * This function is similar to alloc_pages(), except that it allocates the
4934 * minimum number of pages to satisfy the request. alloc_pages() can only
4935 * allocate memory in power-of-two pages.
4937 * This function is also limited by MAX_ORDER.
4939 * Memory allocated by this function must be released by free_pages_exact().
4941 * Return: pointer to the allocated area or %NULL in case of error.
4943 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4945 unsigned int order = get_order(size);
4948 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
4949 gfp_mask &= ~__GFP_COMP;
4951 addr = __get_free_pages(gfp_mask, order);
4952 return make_alloc_exact(addr, order, size);
4954 EXPORT_SYMBOL(alloc_pages_exact);
4957 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4959 * @nid: the preferred node ID where memory should be allocated
4960 * @size: the number of bytes to allocate
4961 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4963 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4966 * Return: pointer to the allocated area or %NULL in case of error.
4968 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4970 unsigned int order = get_order(size);
4973 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
4974 gfp_mask &= ~__GFP_COMP;
4976 p = alloc_pages_node(nid, gfp_mask, order);
4979 return make_alloc_exact((unsigned long)page_address(p), order, size);
4983 * free_pages_exact - release memory allocated via alloc_pages_exact()
4984 * @virt: the value returned by alloc_pages_exact.
4985 * @size: size of allocation, same value as passed to alloc_pages_exact().
4987 * Release the memory allocated by a previous call to alloc_pages_exact.
4989 void free_pages_exact(void *virt, size_t size)
4991 unsigned long addr = (unsigned long)virt;
4992 unsigned long end = addr + PAGE_ALIGN(size);
4994 while (addr < end) {
4999 EXPORT_SYMBOL(free_pages_exact);
5002 * nr_free_zone_pages - count number of pages beyond high watermark
5003 * @offset: The zone index of the highest zone
5005 * nr_free_zone_pages() counts the number of pages which are beyond the
5006 * high watermark within all zones at or below a given zone index. For each
5007 * zone, the number of pages is calculated as:
5009 * nr_free_zone_pages = managed_pages - high_pages
5011 * Return: number of pages beyond high watermark.
5013 static unsigned long nr_free_zone_pages(int offset)
5018 /* Just pick one node, since fallback list is circular */
5019 unsigned long sum = 0;
5021 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5023 for_each_zone_zonelist(zone, z, zonelist, offset) {
5024 unsigned long size = zone_managed_pages(zone);
5025 unsigned long high = high_wmark_pages(zone);
5034 * nr_free_buffer_pages - count number of pages beyond high watermark
5036 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5037 * watermark within ZONE_DMA and ZONE_NORMAL.
5039 * Return: number of pages beyond high watermark within ZONE_DMA and
5042 unsigned long nr_free_buffer_pages(void)
5044 return nr_free_zone_pages(gfp_zone(GFP_USER));
5046 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5049 * nr_free_pagecache_pages - count number of pages beyond high watermark
5051 * nr_free_pagecache_pages() counts the number of pages which are beyond the
5052 * high watermark within all zones.
5054 * Return: number of pages beyond high watermark within all zones.
5056 unsigned long nr_free_pagecache_pages(void)
5058 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5061 static inline void show_node(struct zone *zone)
5063 if (IS_ENABLED(CONFIG_NUMA))
5064 printk("Node %d ", zone_to_nid(zone));
5067 long si_mem_available(void)
5070 unsigned long pagecache;
5071 unsigned long wmark_low = 0;
5072 unsigned long pages[NR_LRU_LISTS];
5073 unsigned long reclaimable;
5077 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5078 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5081 wmark_low += low_wmark_pages(zone);
5084 * Estimate the amount of memory available for userspace allocations,
5085 * without causing swapping.
5087 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5090 * Not all the page cache can be freed, otherwise the system will
5091 * start swapping. Assume at least half of the page cache, or the
5092 * low watermark worth of cache, needs to stay.
5094 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5095 pagecache -= min(pagecache / 2, wmark_low);
5096 available += pagecache;
5099 * Part of the reclaimable slab and other kernel memory consists of
5100 * items that are in use, and cannot be freed. Cap this estimate at the
5103 reclaimable = global_node_page_state(NR_SLAB_RECLAIMABLE) +
5104 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5105 available += reclaimable - min(reclaimable / 2, wmark_low);
5111 EXPORT_SYMBOL_GPL(si_mem_available);
5113 void si_meminfo(struct sysinfo *val)
5115 val->totalram = totalram_pages();
5116 val->sharedram = global_node_page_state(NR_SHMEM);
5117 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5118 val->bufferram = nr_blockdev_pages();
5119 val->totalhigh = totalhigh_pages();
5120 val->freehigh = nr_free_highpages();
5121 val->mem_unit = PAGE_SIZE;
5124 EXPORT_SYMBOL(si_meminfo);
5127 void si_meminfo_node(struct sysinfo *val, int nid)
5129 int zone_type; /* needs to be signed */
5130 unsigned long managed_pages = 0;
5131 unsigned long managed_highpages = 0;
5132 unsigned long free_highpages = 0;
5133 pg_data_t *pgdat = NODE_DATA(nid);
5135 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5136 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5137 val->totalram = managed_pages;
5138 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5139 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5140 #ifdef CONFIG_HIGHMEM
5141 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5142 struct zone *zone = &pgdat->node_zones[zone_type];
5144 if (is_highmem(zone)) {
5145 managed_highpages += zone_managed_pages(zone);
5146 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5149 val->totalhigh = managed_highpages;
5150 val->freehigh = free_highpages;
5152 val->totalhigh = managed_highpages;
5153 val->freehigh = free_highpages;
5155 val->mem_unit = PAGE_SIZE;
5160 * Determine whether the node should be displayed or not, depending on whether
5161 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5163 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5165 if (!(flags & SHOW_MEM_FILTER_NODES))
5169 * no node mask - aka implicit memory numa policy. Do not bother with
5170 * the synchronization - read_mems_allowed_begin - because we do not
5171 * have to be precise here.
5174 nodemask = &cpuset_current_mems_allowed;
5176 return !node_isset(nid, *nodemask);
5179 #define K(x) ((x) << (PAGE_SHIFT-10))
5181 static void show_migration_types(unsigned char type)
5183 static const char types[MIGRATE_TYPES] = {
5184 [MIGRATE_UNMOVABLE] = 'U',
5185 [MIGRATE_MOVABLE] = 'M',
5186 [MIGRATE_RECLAIMABLE] = 'E',
5187 [MIGRATE_HIGHATOMIC] = 'H',
5189 [MIGRATE_CMA] = 'C',
5191 #ifdef CONFIG_MEMORY_ISOLATION
5192 [MIGRATE_ISOLATE] = 'I',
5195 char tmp[MIGRATE_TYPES + 1];
5199 for (i = 0; i < MIGRATE_TYPES; i++) {
5200 if (type & (1 << i))
5205 printk(KERN_CONT "(%s) ", tmp);
5209 * Show free area list (used inside shift_scroll-lock stuff)
5210 * We also calculate the percentage fragmentation. We do this by counting the
5211 * memory on each free list with the exception of the first item on the list.
5214 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5217 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5219 unsigned long free_pcp = 0;
5224 for_each_populated_zone(zone) {
5225 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5228 for_each_online_cpu(cpu)
5229 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5232 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5233 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5234 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
5235 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5236 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5237 " free:%lu free_pcp:%lu free_cma:%lu\n",
5238 global_node_page_state(NR_ACTIVE_ANON),
5239 global_node_page_state(NR_INACTIVE_ANON),
5240 global_node_page_state(NR_ISOLATED_ANON),
5241 global_node_page_state(NR_ACTIVE_FILE),
5242 global_node_page_state(NR_INACTIVE_FILE),
5243 global_node_page_state(NR_ISOLATED_FILE),
5244 global_node_page_state(NR_UNEVICTABLE),
5245 global_node_page_state(NR_FILE_DIRTY),
5246 global_node_page_state(NR_WRITEBACK),
5247 global_node_page_state(NR_UNSTABLE_NFS),
5248 global_node_page_state(NR_SLAB_RECLAIMABLE),
5249 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
5250 global_node_page_state(NR_FILE_MAPPED),
5251 global_node_page_state(NR_SHMEM),
5252 global_zone_page_state(NR_PAGETABLE),
5253 global_zone_page_state(NR_BOUNCE),
5254 global_zone_page_state(NR_FREE_PAGES),
5256 global_zone_page_state(NR_FREE_CMA_PAGES));
5258 for_each_online_pgdat(pgdat) {
5259 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5263 " active_anon:%lukB"
5264 " inactive_anon:%lukB"
5265 " active_file:%lukB"
5266 " inactive_file:%lukB"
5267 " unevictable:%lukB"
5268 " isolated(anon):%lukB"
5269 " isolated(file):%lukB"
5274 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5276 " shmem_pmdmapped: %lukB"
5279 " writeback_tmp:%lukB"
5281 " all_unreclaimable? %s"
5284 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5285 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5286 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5287 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5288 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5289 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5290 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5291 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5292 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5293 K(node_page_state(pgdat, NR_WRITEBACK)),
5294 K(node_page_state(pgdat, NR_SHMEM)),
5295 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5296 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5297 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5299 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5301 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5302 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
5303 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5307 for_each_populated_zone(zone) {
5310 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5314 for_each_online_cpu(cpu)
5315 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5324 " active_anon:%lukB"
5325 " inactive_anon:%lukB"
5326 " active_file:%lukB"
5327 " inactive_file:%lukB"
5328 " unevictable:%lukB"
5329 " writepending:%lukB"
5333 " kernel_stack:%lukB"
5341 K(zone_page_state(zone, NR_FREE_PAGES)),
5342 K(min_wmark_pages(zone)),
5343 K(low_wmark_pages(zone)),
5344 K(high_wmark_pages(zone)),
5345 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5346 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5347 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5348 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5349 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5350 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5351 K(zone->present_pages),
5352 K(zone_managed_pages(zone)),
5353 K(zone_page_state(zone, NR_MLOCK)),
5354 zone_page_state(zone, NR_KERNEL_STACK_KB),
5355 K(zone_page_state(zone, NR_PAGETABLE)),
5356 K(zone_page_state(zone, NR_BOUNCE)),
5358 K(this_cpu_read(zone->pageset->pcp.count)),
5359 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5360 printk("lowmem_reserve[]:");
5361 for (i = 0; i < MAX_NR_ZONES; i++)
5362 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5363 printk(KERN_CONT "\n");
5366 for_each_populated_zone(zone) {
5368 unsigned long nr[MAX_ORDER], flags, total = 0;
5369 unsigned char types[MAX_ORDER];
5371 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5374 printk(KERN_CONT "%s: ", zone->name);
5376 spin_lock_irqsave(&zone->lock, flags);
5377 for (order = 0; order < MAX_ORDER; order++) {
5378 struct free_area *area = &zone->free_area[order];
5381 nr[order] = area->nr_free;
5382 total += nr[order] << order;
5385 for (type = 0; type < MIGRATE_TYPES; type++) {
5386 if (!free_area_empty(area, type))
5387 types[order] |= 1 << type;
5390 spin_unlock_irqrestore(&zone->lock, flags);
5391 for (order = 0; order < MAX_ORDER; order++) {
5392 printk(KERN_CONT "%lu*%lukB ",
5393 nr[order], K(1UL) << order);
5395 show_migration_types(types[order]);
5397 printk(KERN_CONT "= %lukB\n", K(total));
5400 hugetlb_show_meminfo();
5402 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5404 show_swap_cache_info();
5407 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5409 zoneref->zone = zone;
5410 zoneref->zone_idx = zone_idx(zone);
5414 * Builds allocation fallback zone lists.
5416 * Add all populated zones of a node to the zonelist.
5418 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5421 enum zone_type zone_type = MAX_NR_ZONES;
5426 zone = pgdat->node_zones + zone_type;
5427 if (managed_zone(zone)) {
5428 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5429 check_highest_zone(zone_type);
5431 } while (zone_type);
5438 static int __parse_numa_zonelist_order(char *s)
5441 * We used to support different zonlists modes but they turned
5442 * out to be just not useful. Let's keep the warning in place
5443 * if somebody still use the cmd line parameter so that we do
5444 * not fail it silently
5446 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5447 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5453 static __init int setup_numa_zonelist_order(char *s)
5458 return __parse_numa_zonelist_order(s);
5460 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5462 char numa_zonelist_order[] = "Node";
5465 * sysctl handler for numa_zonelist_order
5467 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5468 void __user *buffer, size_t *length,
5475 return proc_dostring(table, write, buffer, length, ppos);
5476 str = memdup_user_nul(buffer, 16);
5478 return PTR_ERR(str);
5480 ret = __parse_numa_zonelist_order(str);
5486 #define MAX_NODE_LOAD (nr_online_nodes)
5487 static int node_load[MAX_NUMNODES];
5490 * find_next_best_node - find the next node that should appear in a given node's fallback list
5491 * @node: node whose fallback list we're appending
5492 * @used_node_mask: nodemask_t of already used nodes
5494 * We use a number of factors to determine which is the next node that should
5495 * appear on a given node's fallback list. The node should not have appeared
5496 * already in @node's fallback list, and it should be the next closest node
5497 * according to the distance array (which contains arbitrary distance values
5498 * from each node to each node in the system), and should also prefer nodes
5499 * with no CPUs, since presumably they'll have very little allocation pressure
5500 * on them otherwise.
5502 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5504 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5507 int min_val = INT_MAX;
5508 int best_node = NUMA_NO_NODE;
5509 const struct cpumask *tmp = cpumask_of_node(0);
5511 /* Use the local node if we haven't already */
5512 if (!node_isset(node, *used_node_mask)) {
5513 node_set(node, *used_node_mask);
5517 for_each_node_state(n, N_MEMORY) {
5519 /* Don't want a node to appear more than once */
5520 if (node_isset(n, *used_node_mask))
5523 /* Use the distance array to find the distance */
5524 val = node_distance(node, n);
5526 /* Penalize nodes under us ("prefer the next node") */
5529 /* Give preference to headless and unused nodes */
5530 tmp = cpumask_of_node(n);
5531 if (!cpumask_empty(tmp))
5532 val += PENALTY_FOR_NODE_WITH_CPUS;
5534 /* Slight preference for less loaded node */
5535 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5536 val += node_load[n];
5538 if (val < min_val) {
5545 node_set(best_node, *used_node_mask);
5552 * Build zonelists ordered by node and zones within node.
5553 * This results in maximum locality--normal zone overflows into local
5554 * DMA zone, if any--but risks exhausting DMA zone.
5556 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5559 struct zoneref *zonerefs;
5562 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5564 for (i = 0; i < nr_nodes; i++) {
5567 pg_data_t *node = NODE_DATA(node_order[i]);
5569 nr_zones = build_zonerefs_node(node, zonerefs);
5570 zonerefs += nr_zones;
5572 zonerefs->zone = NULL;
5573 zonerefs->zone_idx = 0;
5577 * Build gfp_thisnode zonelists
5579 static void build_thisnode_zonelists(pg_data_t *pgdat)
5581 struct zoneref *zonerefs;
5584 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5585 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5586 zonerefs += nr_zones;
5587 zonerefs->zone = NULL;
5588 zonerefs->zone_idx = 0;
5592 * Build zonelists ordered by zone and nodes within zones.
5593 * This results in conserving DMA zone[s] until all Normal memory is
5594 * exhausted, but results in overflowing to remote node while memory
5595 * may still exist in local DMA zone.
5598 static void build_zonelists(pg_data_t *pgdat)
5600 static int node_order[MAX_NUMNODES];
5601 int node, load, nr_nodes = 0;
5602 nodemask_t used_mask;
5603 int local_node, prev_node;
5605 /* NUMA-aware ordering of nodes */
5606 local_node = pgdat->node_id;
5607 load = nr_online_nodes;
5608 prev_node = local_node;
5609 nodes_clear(used_mask);
5611 memset(node_order, 0, sizeof(node_order));
5612 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5614 * We don't want to pressure a particular node.
5615 * So adding penalty to the first node in same
5616 * distance group to make it round-robin.
5618 if (node_distance(local_node, node) !=
5619 node_distance(local_node, prev_node))
5620 node_load[node] = load;
5622 node_order[nr_nodes++] = node;
5627 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5628 build_thisnode_zonelists(pgdat);
5631 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5633 * Return node id of node used for "local" allocations.
5634 * I.e., first node id of first zone in arg node's generic zonelist.
5635 * Used for initializing percpu 'numa_mem', which is used primarily
5636 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5638 int local_memory_node(int node)
5642 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5643 gfp_zone(GFP_KERNEL),
5645 return zone_to_nid(z->zone);
5649 static void setup_min_unmapped_ratio(void);
5650 static void setup_min_slab_ratio(void);
5651 #else /* CONFIG_NUMA */
5653 static void build_zonelists(pg_data_t *pgdat)
5655 int node, local_node;
5656 struct zoneref *zonerefs;
5659 local_node = pgdat->node_id;
5661 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5662 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5663 zonerefs += nr_zones;
5666 * Now we build the zonelist so that it contains the zones
5667 * of all the other nodes.
5668 * We don't want to pressure a particular node, so when
5669 * building the zones for node N, we make sure that the
5670 * zones coming right after the local ones are those from
5671 * node N+1 (modulo N)
5673 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5674 if (!node_online(node))
5676 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5677 zonerefs += nr_zones;
5679 for (node = 0; node < local_node; node++) {
5680 if (!node_online(node))
5682 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5683 zonerefs += nr_zones;
5686 zonerefs->zone = NULL;
5687 zonerefs->zone_idx = 0;
5690 #endif /* CONFIG_NUMA */
5693 * Boot pageset table. One per cpu which is going to be used for all
5694 * zones and all nodes. The parameters will be set in such a way
5695 * that an item put on a list will immediately be handed over to
5696 * the buddy list. This is safe since pageset manipulation is done
5697 * with interrupts disabled.
5699 * The boot_pagesets must be kept even after bootup is complete for
5700 * unused processors and/or zones. They do play a role for bootstrapping
5701 * hotplugged processors.
5703 * zoneinfo_show() and maybe other functions do
5704 * not check if the processor is online before following the pageset pointer.
5705 * Other parts of the kernel may not check if the zone is available.
5707 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5708 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5709 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5711 static void __build_all_zonelists(void *data)
5714 int __maybe_unused cpu;
5715 pg_data_t *self = data;
5716 static DEFINE_SPINLOCK(lock);
5721 memset(node_load, 0, sizeof(node_load));
5725 * This node is hotadded and no memory is yet present. So just
5726 * building zonelists is fine - no need to touch other nodes.
5728 if (self && !node_online(self->node_id)) {
5729 build_zonelists(self);
5731 for_each_online_node(nid) {
5732 pg_data_t *pgdat = NODE_DATA(nid);
5734 build_zonelists(pgdat);
5737 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5739 * We now know the "local memory node" for each node--
5740 * i.e., the node of the first zone in the generic zonelist.
5741 * Set up numa_mem percpu variable for on-line cpus. During
5742 * boot, only the boot cpu should be on-line; we'll init the
5743 * secondary cpus' numa_mem as they come on-line. During
5744 * node/memory hotplug, we'll fixup all on-line cpus.
5746 for_each_online_cpu(cpu)
5747 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5754 static noinline void __init
5755 build_all_zonelists_init(void)
5759 __build_all_zonelists(NULL);
5762 * Initialize the boot_pagesets that are going to be used
5763 * for bootstrapping processors. The real pagesets for
5764 * each zone will be allocated later when the per cpu
5765 * allocator is available.
5767 * boot_pagesets are used also for bootstrapping offline
5768 * cpus if the system is already booted because the pagesets
5769 * are needed to initialize allocators on a specific cpu too.
5770 * F.e. the percpu allocator needs the page allocator which
5771 * needs the percpu allocator in order to allocate its pagesets
5772 * (a chicken-egg dilemma).
5774 for_each_possible_cpu(cpu)
5775 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5777 mminit_verify_zonelist();
5778 cpuset_init_current_mems_allowed();
5782 * unless system_state == SYSTEM_BOOTING.
5784 * __ref due to call of __init annotated helper build_all_zonelists_init
5785 * [protected by SYSTEM_BOOTING].
5787 void __ref build_all_zonelists(pg_data_t *pgdat)
5789 if (system_state == SYSTEM_BOOTING) {
5790 build_all_zonelists_init();
5792 __build_all_zonelists(pgdat);
5793 /* cpuset refresh routine should be here */
5795 vm_total_pages = nr_free_pagecache_pages();
5797 * Disable grouping by mobility if the number of pages in the
5798 * system is too low to allow the mechanism to work. It would be
5799 * more accurate, but expensive to check per-zone. This check is
5800 * made on memory-hotadd so a system can start with mobility
5801 * disabled and enable it later
5803 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5804 page_group_by_mobility_disabled = 1;
5806 page_group_by_mobility_disabled = 0;
5808 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5810 page_group_by_mobility_disabled ? "off" : "on",
5813 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5817 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5818 static bool __meminit
5819 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5821 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5822 static struct memblock_region *r;
5824 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5825 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5826 for_each_memblock(memory, r) {
5827 if (*pfn < memblock_region_memory_end_pfn(r))
5831 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5832 memblock_is_mirror(r)) {
5833 *pfn = memblock_region_memory_end_pfn(r);
5842 * Initially all pages are reserved - free ones are freed
5843 * up by memblock_free_all() once the early boot process is
5844 * done. Non-atomic initialization, single-pass.
5846 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5847 unsigned long start_pfn, enum memmap_context context,
5848 struct vmem_altmap *altmap)
5850 unsigned long pfn, end_pfn = start_pfn + size;
5853 if (highest_memmap_pfn < end_pfn - 1)
5854 highest_memmap_pfn = end_pfn - 1;
5856 #ifdef CONFIG_ZONE_DEVICE
5858 * Honor reservation requested by the driver for this ZONE_DEVICE
5859 * memory. We limit the total number of pages to initialize to just
5860 * those that might contain the memory mapping. We will defer the
5861 * ZONE_DEVICE page initialization until after we have released
5864 if (zone == ZONE_DEVICE) {
5868 if (start_pfn == altmap->base_pfn)
5869 start_pfn += altmap->reserve;
5870 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5874 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5876 * There can be holes in boot-time mem_map[]s handed to this
5877 * function. They do not exist on hotplugged memory.
5879 if (context == MEMMAP_EARLY) {
5880 if (!early_pfn_valid(pfn))
5882 if (!early_pfn_in_nid(pfn, nid))
5884 if (overlap_memmap_init(zone, &pfn))
5886 if (defer_init(nid, pfn, end_pfn))
5890 page = pfn_to_page(pfn);
5891 __init_single_page(page, pfn, zone, nid);
5892 if (context == MEMMAP_HOTPLUG)
5893 __SetPageReserved(page);
5896 * Mark the block movable so that blocks are reserved for
5897 * movable at startup. This will force kernel allocations
5898 * to reserve their blocks rather than leaking throughout
5899 * the address space during boot when many long-lived
5900 * kernel allocations are made.
5902 * bitmap is created for zone's valid pfn range. but memmap
5903 * can be created for invalid pages (for alignment)
5904 * check here not to call set_pageblock_migratetype() against
5907 if (!(pfn & (pageblock_nr_pages - 1))) {
5908 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5914 #ifdef CONFIG_ZONE_DEVICE
5915 void __ref memmap_init_zone_device(struct zone *zone,
5916 unsigned long start_pfn,
5918 struct dev_pagemap *pgmap)
5920 unsigned long pfn, end_pfn = start_pfn + size;
5921 struct pglist_data *pgdat = zone->zone_pgdat;
5922 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
5923 unsigned long zone_idx = zone_idx(zone);
5924 unsigned long start = jiffies;
5925 int nid = pgdat->node_id;
5927 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
5931 * The call to memmap_init_zone should have already taken care
5932 * of the pages reserved for the memmap, so we can just jump to
5933 * the end of that region and start processing the device pages.
5936 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5937 size = end_pfn - start_pfn;
5940 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5941 struct page *page = pfn_to_page(pfn);
5943 __init_single_page(page, pfn, zone_idx, nid);
5946 * Mark page reserved as it will need to wait for onlining
5947 * phase for it to be fully associated with a zone.
5949 * We can use the non-atomic __set_bit operation for setting
5950 * the flag as we are still initializing the pages.
5952 __SetPageReserved(page);
5955 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
5956 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
5957 * ever freed or placed on a driver-private list.
5959 page->pgmap = pgmap;
5960 page->zone_device_data = NULL;
5963 * Mark the block movable so that blocks are reserved for
5964 * movable at startup. This will force kernel allocations
5965 * to reserve their blocks rather than leaking throughout
5966 * the address space during boot when many long-lived
5967 * kernel allocations are made.
5969 * bitmap is created for zone's valid pfn range. but memmap
5970 * can be created for invalid pages (for alignment)
5971 * check here not to call set_pageblock_migratetype() against
5974 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
5975 * because this is done early in section_activate()
5977 if (!(pfn & (pageblock_nr_pages - 1))) {
5978 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5983 pr_info("%s initialised %lu pages in %ums\n", __func__,
5984 size, jiffies_to_msecs(jiffies - start));
5988 static void __meminit zone_init_free_lists(struct zone *zone)
5990 unsigned int order, t;
5991 for_each_migratetype_order(order, t) {
5992 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5993 zone->free_area[order].nr_free = 0;
5997 void __meminit __weak memmap_init(unsigned long size, int nid,
5998 unsigned long zone, unsigned long start_pfn)
6000 memmap_init_zone(size, nid, zone, start_pfn, MEMMAP_EARLY, NULL);
6003 static int zone_batchsize(struct zone *zone)
6009 * The per-cpu-pages pools are set to around 1000th of the
6012 batch = zone_managed_pages(zone) / 1024;
6013 /* But no more than a meg. */
6014 if (batch * PAGE_SIZE > 1024 * 1024)
6015 batch = (1024 * 1024) / PAGE_SIZE;
6016 batch /= 4; /* We effectively *= 4 below */
6021 * Clamp the batch to a 2^n - 1 value. Having a power
6022 * of 2 value was found to be more likely to have
6023 * suboptimal cache aliasing properties in some cases.
6025 * For example if 2 tasks are alternately allocating
6026 * batches of pages, one task can end up with a lot
6027 * of pages of one half of the possible page colors
6028 * and the other with pages of the other colors.
6030 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6035 /* The deferral and batching of frees should be suppressed under NOMMU
6038 * The problem is that NOMMU needs to be able to allocate large chunks
6039 * of contiguous memory as there's no hardware page translation to
6040 * assemble apparent contiguous memory from discontiguous pages.
6042 * Queueing large contiguous runs of pages for batching, however,
6043 * causes the pages to actually be freed in smaller chunks. As there
6044 * can be a significant delay between the individual batches being
6045 * recycled, this leads to the once large chunks of space being
6046 * fragmented and becoming unavailable for high-order allocations.
6053 * pcp->high and pcp->batch values are related and dependent on one another:
6054 * ->batch must never be higher then ->high.
6055 * The following function updates them in a safe manner without read side
6058 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6059 * those fields changing asynchronously (acording the the above rule).
6061 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6062 * outside of boot time (or some other assurance that no concurrent updaters
6065 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6066 unsigned long batch)
6068 /* start with a fail safe value for batch */
6072 /* Update high, then batch, in order */
6079 /* a companion to pageset_set_high() */
6080 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6082 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6085 static void pageset_init(struct per_cpu_pageset *p)
6087 struct per_cpu_pages *pcp;
6090 memset(p, 0, sizeof(*p));
6093 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6094 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6097 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6100 pageset_set_batch(p, batch);
6104 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6105 * to the value high for the pageset p.
6107 static void pageset_set_high(struct per_cpu_pageset *p,
6110 unsigned long batch = max(1UL, high / 4);
6111 if ((high / 4) > (PAGE_SHIFT * 8))
6112 batch = PAGE_SHIFT * 8;
6114 pageset_update(&p->pcp, high, batch);
6117 static void pageset_set_high_and_batch(struct zone *zone,
6118 struct per_cpu_pageset *pcp)
6120 if (percpu_pagelist_fraction)
6121 pageset_set_high(pcp,
6122 (zone_managed_pages(zone) /
6123 percpu_pagelist_fraction));
6125 pageset_set_batch(pcp, zone_batchsize(zone));
6128 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6130 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6133 pageset_set_high_and_batch(zone, pcp);
6136 void __meminit setup_zone_pageset(struct zone *zone)
6139 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6140 for_each_possible_cpu(cpu)
6141 zone_pageset_init(zone, cpu);
6145 * Allocate per cpu pagesets and initialize them.
6146 * Before this call only boot pagesets were available.
6148 void __init setup_per_cpu_pageset(void)
6150 struct pglist_data *pgdat;
6153 for_each_populated_zone(zone)
6154 setup_zone_pageset(zone);
6156 for_each_online_pgdat(pgdat)
6157 pgdat->per_cpu_nodestats =
6158 alloc_percpu(struct per_cpu_nodestat);
6161 static __meminit void zone_pcp_init(struct zone *zone)
6164 * per cpu subsystem is not up at this point. The following code
6165 * relies on the ability of the linker to provide the
6166 * offset of a (static) per cpu variable into the per cpu area.
6168 zone->pageset = &boot_pageset;
6170 if (populated_zone(zone))
6171 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6172 zone->name, zone->present_pages,
6173 zone_batchsize(zone));
6176 void __meminit init_currently_empty_zone(struct zone *zone,
6177 unsigned long zone_start_pfn,
6180 struct pglist_data *pgdat = zone->zone_pgdat;
6181 int zone_idx = zone_idx(zone) + 1;
6183 if (zone_idx > pgdat->nr_zones)
6184 pgdat->nr_zones = zone_idx;
6186 zone->zone_start_pfn = zone_start_pfn;
6188 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6189 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6191 (unsigned long)zone_idx(zone),
6192 zone_start_pfn, (zone_start_pfn + size));
6194 zone_init_free_lists(zone);
6195 zone->initialized = 1;
6198 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6199 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
6202 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
6204 int __meminit __early_pfn_to_nid(unsigned long pfn,
6205 struct mminit_pfnnid_cache *state)
6207 unsigned long start_pfn, end_pfn;
6210 if (state->last_start <= pfn && pfn < state->last_end)
6211 return state->last_nid;
6213 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
6214 if (nid != NUMA_NO_NODE) {
6215 state->last_start = start_pfn;
6216 state->last_end = end_pfn;
6217 state->last_nid = nid;
6222 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
6225 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
6226 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
6227 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
6229 * If an architecture guarantees that all ranges registered contain no holes
6230 * and may be freed, this this function may be used instead of calling
6231 * memblock_free_early_nid() manually.
6233 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
6235 unsigned long start_pfn, end_pfn;
6238 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
6239 start_pfn = min(start_pfn, max_low_pfn);
6240 end_pfn = min(end_pfn, max_low_pfn);
6242 if (start_pfn < end_pfn)
6243 memblock_free_early_nid(PFN_PHYS(start_pfn),
6244 (end_pfn - start_pfn) << PAGE_SHIFT,
6250 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6251 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6253 * If an architecture guarantees that all ranges registered contain no holes and may
6254 * be freed, this function may be used instead of calling memory_present() manually.
6256 void __init sparse_memory_present_with_active_regions(int nid)
6258 unsigned long start_pfn, end_pfn;
6261 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
6262 memory_present(this_nid, start_pfn, end_pfn);
6266 * get_pfn_range_for_nid - Return the start and end page frames for a node
6267 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6268 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6269 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6271 * It returns the start and end page frame of a node based on information
6272 * provided by memblock_set_node(). If called for a node
6273 * with no available memory, a warning is printed and the start and end
6276 void __init get_pfn_range_for_nid(unsigned int nid,
6277 unsigned long *start_pfn, unsigned long *end_pfn)
6279 unsigned long this_start_pfn, this_end_pfn;
6285 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6286 *start_pfn = min(*start_pfn, this_start_pfn);
6287 *end_pfn = max(*end_pfn, this_end_pfn);
6290 if (*start_pfn == -1UL)
6295 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6296 * assumption is made that zones within a node are ordered in monotonic
6297 * increasing memory addresses so that the "highest" populated zone is used
6299 static void __init find_usable_zone_for_movable(void)
6302 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6303 if (zone_index == ZONE_MOVABLE)
6306 if (arch_zone_highest_possible_pfn[zone_index] >
6307 arch_zone_lowest_possible_pfn[zone_index])
6311 VM_BUG_ON(zone_index == -1);
6312 movable_zone = zone_index;
6316 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6317 * because it is sized independent of architecture. Unlike the other zones,
6318 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6319 * in each node depending on the size of each node and how evenly kernelcore
6320 * is distributed. This helper function adjusts the zone ranges
6321 * provided by the architecture for a given node by using the end of the
6322 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6323 * zones within a node are in order of monotonic increases memory addresses
6325 static void __init adjust_zone_range_for_zone_movable(int nid,
6326 unsigned long zone_type,
6327 unsigned long node_start_pfn,
6328 unsigned long node_end_pfn,
6329 unsigned long *zone_start_pfn,
6330 unsigned long *zone_end_pfn)
6332 /* Only adjust if ZONE_MOVABLE is on this node */
6333 if (zone_movable_pfn[nid]) {
6334 /* Size ZONE_MOVABLE */
6335 if (zone_type == ZONE_MOVABLE) {
6336 *zone_start_pfn = zone_movable_pfn[nid];
6337 *zone_end_pfn = min(node_end_pfn,
6338 arch_zone_highest_possible_pfn[movable_zone]);
6340 /* Adjust for ZONE_MOVABLE starting within this range */
6341 } else if (!mirrored_kernelcore &&
6342 *zone_start_pfn < zone_movable_pfn[nid] &&
6343 *zone_end_pfn > zone_movable_pfn[nid]) {
6344 *zone_end_pfn = zone_movable_pfn[nid];
6346 /* Check if this whole range is within ZONE_MOVABLE */
6347 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6348 *zone_start_pfn = *zone_end_pfn;
6353 * Return the number of pages a zone spans in a node, including holes
6354 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6356 static unsigned long __init zone_spanned_pages_in_node(int nid,
6357 unsigned long zone_type,
6358 unsigned long node_start_pfn,
6359 unsigned long node_end_pfn,
6360 unsigned long *zone_start_pfn,
6361 unsigned long *zone_end_pfn,
6362 unsigned long *ignored)
6364 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6365 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6366 /* When hotadd a new node from cpu_up(), the node should be empty */
6367 if (!node_start_pfn && !node_end_pfn)
6370 /* Get the start and end of the zone */
6371 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6372 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6373 adjust_zone_range_for_zone_movable(nid, zone_type,
6374 node_start_pfn, node_end_pfn,
6375 zone_start_pfn, zone_end_pfn);
6377 /* Check that this node has pages within the zone's required range */
6378 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6381 /* Move the zone boundaries inside the node if necessary */
6382 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6383 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6385 /* Return the spanned pages */
6386 return *zone_end_pfn - *zone_start_pfn;
6390 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6391 * then all holes in the requested range will be accounted for.
6393 unsigned long __init __absent_pages_in_range(int nid,
6394 unsigned long range_start_pfn,
6395 unsigned long range_end_pfn)
6397 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6398 unsigned long start_pfn, end_pfn;
6401 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6402 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6403 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6404 nr_absent -= end_pfn - start_pfn;
6410 * absent_pages_in_range - Return number of page frames in holes within a range
6411 * @start_pfn: The start PFN to start searching for holes
6412 * @end_pfn: The end PFN to stop searching for holes
6414 * Return: the number of pages frames in memory holes within a range.
6416 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6417 unsigned long end_pfn)
6419 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6422 /* Return the number of page frames in holes in a zone on a node */
6423 static unsigned long __init zone_absent_pages_in_node(int nid,
6424 unsigned long zone_type,
6425 unsigned long node_start_pfn,
6426 unsigned long node_end_pfn,
6427 unsigned long *ignored)
6429 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6430 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6431 unsigned long zone_start_pfn, zone_end_pfn;
6432 unsigned long nr_absent;
6434 /* When hotadd a new node from cpu_up(), the node should be empty */
6435 if (!node_start_pfn && !node_end_pfn)
6438 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6439 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6441 adjust_zone_range_for_zone_movable(nid, zone_type,
6442 node_start_pfn, node_end_pfn,
6443 &zone_start_pfn, &zone_end_pfn);
6444 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6447 * ZONE_MOVABLE handling.
6448 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6451 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6452 unsigned long start_pfn, end_pfn;
6453 struct memblock_region *r;
6455 for_each_memblock(memory, r) {
6456 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6457 zone_start_pfn, zone_end_pfn);
6458 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6459 zone_start_pfn, zone_end_pfn);
6461 if (zone_type == ZONE_MOVABLE &&
6462 memblock_is_mirror(r))
6463 nr_absent += end_pfn - start_pfn;
6465 if (zone_type == ZONE_NORMAL &&
6466 !memblock_is_mirror(r))
6467 nr_absent += end_pfn - start_pfn;
6474 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6475 static inline unsigned long __init zone_spanned_pages_in_node(int nid,
6476 unsigned long zone_type,
6477 unsigned long node_start_pfn,
6478 unsigned long node_end_pfn,
6479 unsigned long *zone_start_pfn,
6480 unsigned long *zone_end_pfn,
6481 unsigned long *zones_size)
6485 *zone_start_pfn = node_start_pfn;
6486 for (zone = 0; zone < zone_type; zone++)
6487 *zone_start_pfn += zones_size[zone];
6489 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6491 return zones_size[zone_type];
6494 static inline unsigned long __init zone_absent_pages_in_node(int nid,
6495 unsigned long zone_type,
6496 unsigned long node_start_pfn,
6497 unsigned long node_end_pfn,
6498 unsigned long *zholes_size)
6503 return zholes_size[zone_type];
6506 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6508 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6509 unsigned long node_start_pfn,
6510 unsigned long node_end_pfn,
6511 unsigned long *zones_size,
6512 unsigned long *zholes_size)
6514 unsigned long realtotalpages = 0, totalpages = 0;
6517 for (i = 0; i < MAX_NR_ZONES; i++) {
6518 struct zone *zone = pgdat->node_zones + i;
6519 unsigned long zone_start_pfn, zone_end_pfn;
6520 unsigned long size, real_size;
6522 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6528 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6529 node_start_pfn, node_end_pfn,
6532 zone->zone_start_pfn = zone_start_pfn;
6534 zone->zone_start_pfn = 0;
6535 zone->spanned_pages = size;
6536 zone->present_pages = real_size;
6539 realtotalpages += real_size;
6542 pgdat->node_spanned_pages = totalpages;
6543 pgdat->node_present_pages = realtotalpages;
6544 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6548 #ifndef CONFIG_SPARSEMEM
6550 * Calculate the size of the zone->blockflags rounded to an unsigned long
6551 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6552 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6553 * round what is now in bits to nearest long in bits, then return it in
6556 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6558 unsigned long usemapsize;
6560 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6561 usemapsize = roundup(zonesize, pageblock_nr_pages);
6562 usemapsize = usemapsize >> pageblock_order;
6563 usemapsize *= NR_PAGEBLOCK_BITS;
6564 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6566 return usemapsize / 8;
6569 static void __ref setup_usemap(struct pglist_data *pgdat,
6571 unsigned long zone_start_pfn,
6572 unsigned long zonesize)
6574 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6575 zone->pageblock_flags = NULL;
6577 zone->pageblock_flags =
6578 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6580 if (!zone->pageblock_flags)
6581 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6582 usemapsize, zone->name, pgdat->node_id);
6586 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6587 unsigned long zone_start_pfn, unsigned long zonesize) {}
6588 #endif /* CONFIG_SPARSEMEM */
6590 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6592 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6593 void __init set_pageblock_order(void)
6597 /* Check that pageblock_nr_pages has not already been setup */
6598 if (pageblock_order)
6601 if (HPAGE_SHIFT > PAGE_SHIFT)
6602 order = HUGETLB_PAGE_ORDER;
6604 order = MAX_ORDER - 1;
6607 * Assume the largest contiguous order of interest is a huge page.
6608 * This value may be variable depending on boot parameters on IA64 and
6611 pageblock_order = order;
6613 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6616 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6617 * is unused as pageblock_order is set at compile-time. See
6618 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6621 void __init set_pageblock_order(void)
6625 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6627 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6628 unsigned long present_pages)
6630 unsigned long pages = spanned_pages;
6633 * Provide a more accurate estimation if there are holes within
6634 * the zone and SPARSEMEM is in use. If there are holes within the
6635 * zone, each populated memory region may cost us one or two extra
6636 * memmap pages due to alignment because memmap pages for each
6637 * populated regions may not be naturally aligned on page boundary.
6638 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6640 if (spanned_pages > present_pages + (present_pages >> 4) &&
6641 IS_ENABLED(CONFIG_SPARSEMEM))
6642 pages = present_pages;
6644 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6647 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6648 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6650 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6652 spin_lock_init(&ds_queue->split_queue_lock);
6653 INIT_LIST_HEAD(&ds_queue->split_queue);
6654 ds_queue->split_queue_len = 0;
6657 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6660 #ifdef CONFIG_COMPACTION
6661 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6663 init_waitqueue_head(&pgdat->kcompactd_wait);
6666 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6669 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6671 pgdat_resize_init(pgdat);
6673 pgdat_init_split_queue(pgdat);
6674 pgdat_init_kcompactd(pgdat);
6676 init_waitqueue_head(&pgdat->kswapd_wait);
6677 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6679 pgdat_page_ext_init(pgdat);
6680 spin_lock_init(&pgdat->lru_lock);
6681 lruvec_init(node_lruvec(pgdat));
6684 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6685 unsigned long remaining_pages)
6687 atomic_long_set(&zone->managed_pages, remaining_pages);
6688 zone_set_nid(zone, nid);
6689 zone->name = zone_names[idx];
6690 zone->zone_pgdat = NODE_DATA(nid);
6691 spin_lock_init(&zone->lock);
6692 zone_seqlock_init(zone);
6693 zone_pcp_init(zone);
6697 * Set up the zone data structures
6698 * - init pgdat internals
6699 * - init all zones belonging to this node
6701 * NOTE: this function is only called during memory hotplug
6703 #ifdef CONFIG_MEMORY_HOTPLUG
6704 void __ref free_area_init_core_hotplug(int nid)
6707 pg_data_t *pgdat = NODE_DATA(nid);
6709 pgdat_init_internals(pgdat);
6710 for (z = 0; z < MAX_NR_ZONES; z++)
6711 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6716 * Set up the zone data structures:
6717 * - mark all pages reserved
6718 * - mark all memory queues empty
6719 * - clear the memory bitmaps
6721 * NOTE: pgdat should get zeroed by caller.
6722 * NOTE: this function is only called during early init.
6724 static void __init free_area_init_core(struct pglist_data *pgdat)
6727 int nid = pgdat->node_id;
6729 pgdat_init_internals(pgdat);
6730 pgdat->per_cpu_nodestats = &boot_nodestats;
6732 for (j = 0; j < MAX_NR_ZONES; j++) {
6733 struct zone *zone = pgdat->node_zones + j;
6734 unsigned long size, freesize, memmap_pages;
6735 unsigned long zone_start_pfn = zone->zone_start_pfn;
6737 size = zone->spanned_pages;
6738 freesize = zone->present_pages;
6741 * Adjust freesize so that it accounts for how much memory
6742 * is used by this zone for memmap. This affects the watermark
6743 * and per-cpu initialisations
6745 memmap_pages = calc_memmap_size(size, freesize);
6746 if (!is_highmem_idx(j)) {
6747 if (freesize >= memmap_pages) {
6748 freesize -= memmap_pages;
6751 " %s zone: %lu pages used for memmap\n",
6752 zone_names[j], memmap_pages);
6754 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6755 zone_names[j], memmap_pages, freesize);
6758 /* Account for reserved pages */
6759 if (j == 0 && freesize > dma_reserve) {
6760 freesize -= dma_reserve;
6761 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6762 zone_names[0], dma_reserve);
6765 if (!is_highmem_idx(j))
6766 nr_kernel_pages += freesize;
6767 /* Charge for highmem memmap if there are enough kernel pages */
6768 else if (nr_kernel_pages > memmap_pages * 2)
6769 nr_kernel_pages -= memmap_pages;
6770 nr_all_pages += freesize;
6773 * Set an approximate value for lowmem here, it will be adjusted
6774 * when the bootmem allocator frees pages into the buddy system.
6775 * And all highmem pages will be managed by the buddy system.
6777 zone_init_internals(zone, j, nid, freesize);
6782 set_pageblock_order();
6783 setup_usemap(pgdat, zone, zone_start_pfn, size);
6784 init_currently_empty_zone(zone, zone_start_pfn, size);
6785 memmap_init(size, nid, j, zone_start_pfn);
6789 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6790 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6792 unsigned long __maybe_unused start = 0;
6793 unsigned long __maybe_unused offset = 0;
6795 /* Skip empty nodes */
6796 if (!pgdat->node_spanned_pages)
6799 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6800 offset = pgdat->node_start_pfn - start;
6801 /* ia64 gets its own node_mem_map, before this, without bootmem */
6802 if (!pgdat->node_mem_map) {
6803 unsigned long size, end;
6807 * The zone's endpoints aren't required to be MAX_ORDER
6808 * aligned but the node_mem_map endpoints must be in order
6809 * for the buddy allocator to function correctly.
6811 end = pgdat_end_pfn(pgdat);
6812 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6813 size = (end - start) * sizeof(struct page);
6814 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6817 panic("Failed to allocate %ld bytes for node %d memory map\n",
6818 size, pgdat->node_id);
6819 pgdat->node_mem_map = map + offset;
6821 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6822 __func__, pgdat->node_id, (unsigned long)pgdat,
6823 (unsigned long)pgdat->node_mem_map);
6824 #ifndef CONFIG_NEED_MULTIPLE_NODES
6826 * With no DISCONTIG, the global mem_map is just set as node 0's
6828 if (pgdat == NODE_DATA(0)) {
6829 mem_map = NODE_DATA(0)->node_mem_map;
6830 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6831 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6833 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6838 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6839 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6841 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6842 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6844 pgdat->first_deferred_pfn = ULONG_MAX;
6847 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6850 void __init free_area_init_node(int nid, unsigned long *zones_size,
6851 unsigned long node_start_pfn,
6852 unsigned long *zholes_size)
6854 pg_data_t *pgdat = NODE_DATA(nid);
6855 unsigned long start_pfn = 0;
6856 unsigned long end_pfn = 0;
6858 /* pg_data_t should be reset to zero when it's allocated */
6859 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6861 pgdat->node_id = nid;
6862 pgdat->node_start_pfn = node_start_pfn;
6863 pgdat->per_cpu_nodestats = NULL;
6864 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6865 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6866 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6867 (u64)start_pfn << PAGE_SHIFT,
6868 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6870 start_pfn = node_start_pfn;
6872 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6873 zones_size, zholes_size);
6875 alloc_node_mem_map(pgdat);
6876 pgdat_set_deferred_range(pgdat);
6878 free_area_init_core(pgdat);
6881 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6883 * Zero all valid struct pages in range [spfn, epfn), return number of struct
6886 static u64 zero_pfn_range(unsigned long spfn, unsigned long epfn)
6891 for (pfn = spfn; pfn < epfn; pfn++) {
6892 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6893 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6894 + pageblock_nr_pages - 1;
6897 mm_zero_struct_page(pfn_to_page(pfn));
6905 * Only struct pages that are backed by physical memory are zeroed and
6906 * initialized by going through __init_single_page(). But, there are some
6907 * struct pages which are reserved in memblock allocator and their fields
6908 * may be accessed (for example page_to_pfn() on some configuration accesses
6909 * flags). We must explicitly zero those struct pages.
6911 * This function also addresses a similar issue where struct pages are left
6912 * uninitialized because the physical address range is not covered by
6913 * memblock.memory or memblock.reserved. That could happen when memblock
6914 * layout is manually configured via memmap=.
6916 void __init zero_resv_unavail(void)
6918 phys_addr_t start, end;
6920 phys_addr_t next = 0;
6923 * Loop through unavailable ranges not covered by memblock.memory.
6926 for_each_mem_range(i, &memblock.memory, NULL,
6927 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
6929 pgcnt += zero_pfn_range(PFN_DOWN(next), PFN_UP(start));
6932 pgcnt += zero_pfn_range(PFN_DOWN(next), max_pfn);
6935 * Struct pages that do not have backing memory. This could be because
6936 * firmware is using some of this memory, or for some other reasons.
6939 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
6941 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
6943 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6945 #if MAX_NUMNODES > 1
6947 * Figure out the number of possible node ids.
6949 void __init setup_nr_node_ids(void)
6951 unsigned int highest;
6953 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6954 nr_node_ids = highest + 1;
6959 * node_map_pfn_alignment - determine the maximum internode alignment
6961 * This function should be called after node map is populated and sorted.
6962 * It calculates the maximum power of two alignment which can distinguish
6965 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6966 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6967 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6968 * shifted, 1GiB is enough and this function will indicate so.
6970 * This is used to test whether pfn -> nid mapping of the chosen memory
6971 * model has fine enough granularity to avoid incorrect mapping for the
6972 * populated node map.
6974 * Return: the determined alignment in pfn's. 0 if there is no alignment
6975 * requirement (single node).
6977 unsigned long __init node_map_pfn_alignment(void)
6979 unsigned long accl_mask = 0, last_end = 0;
6980 unsigned long start, end, mask;
6981 int last_nid = NUMA_NO_NODE;
6984 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6985 if (!start || last_nid < 0 || last_nid == nid) {
6992 * Start with a mask granular enough to pin-point to the
6993 * start pfn and tick off bits one-by-one until it becomes
6994 * too coarse to separate the current node from the last.
6996 mask = ~((1 << __ffs(start)) - 1);
6997 while (mask && last_end <= (start & (mask << 1)))
7000 /* accumulate all internode masks */
7004 /* convert mask to number of pages */
7005 return ~accl_mask + 1;
7008 /* Find the lowest pfn for a node */
7009 static unsigned long __init find_min_pfn_for_node(int nid)
7011 unsigned long min_pfn = ULONG_MAX;
7012 unsigned long start_pfn;
7015 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
7016 min_pfn = min(min_pfn, start_pfn);
7018 if (min_pfn == ULONG_MAX) {
7019 pr_warn("Could not find start_pfn for node %d\n", nid);
7027 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7029 * Return: the minimum PFN based on information provided via
7030 * memblock_set_node().
7032 unsigned long __init find_min_pfn_with_active_regions(void)
7034 return find_min_pfn_for_node(MAX_NUMNODES);
7038 * early_calculate_totalpages()
7039 * Sum pages in active regions for movable zone.
7040 * Populate N_MEMORY for calculating usable_nodes.
7042 static unsigned long __init early_calculate_totalpages(void)
7044 unsigned long totalpages = 0;
7045 unsigned long start_pfn, end_pfn;
7048 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7049 unsigned long pages = end_pfn - start_pfn;
7051 totalpages += pages;
7053 node_set_state(nid, N_MEMORY);
7059 * Find the PFN the Movable zone begins in each node. Kernel memory
7060 * is spread evenly between nodes as long as the nodes have enough
7061 * memory. When they don't, some nodes will have more kernelcore than
7064 static void __init find_zone_movable_pfns_for_nodes(void)
7067 unsigned long usable_startpfn;
7068 unsigned long kernelcore_node, kernelcore_remaining;
7069 /* save the state before borrow the nodemask */
7070 nodemask_t saved_node_state = node_states[N_MEMORY];
7071 unsigned long totalpages = early_calculate_totalpages();
7072 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7073 struct memblock_region *r;
7075 /* Need to find movable_zone earlier when movable_node is specified. */
7076 find_usable_zone_for_movable();
7079 * If movable_node is specified, ignore kernelcore and movablecore
7082 if (movable_node_is_enabled()) {
7083 for_each_memblock(memory, r) {
7084 if (!memblock_is_hotpluggable(r))
7089 usable_startpfn = PFN_DOWN(r->base);
7090 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7091 min(usable_startpfn, zone_movable_pfn[nid]) :
7099 * If kernelcore=mirror is specified, ignore movablecore option
7101 if (mirrored_kernelcore) {
7102 bool mem_below_4gb_not_mirrored = false;
7104 for_each_memblock(memory, r) {
7105 if (memblock_is_mirror(r))
7110 usable_startpfn = memblock_region_memory_base_pfn(r);
7112 if (usable_startpfn < 0x100000) {
7113 mem_below_4gb_not_mirrored = true;
7117 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7118 min(usable_startpfn, zone_movable_pfn[nid]) :
7122 if (mem_below_4gb_not_mirrored)
7123 pr_warn("This configuration results in unmirrored kernel memory.");
7129 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7130 * amount of necessary memory.
7132 if (required_kernelcore_percent)
7133 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7135 if (required_movablecore_percent)
7136 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7140 * If movablecore= was specified, calculate what size of
7141 * kernelcore that corresponds so that memory usable for
7142 * any allocation type is evenly spread. If both kernelcore
7143 * and movablecore are specified, then the value of kernelcore
7144 * will be used for required_kernelcore if it's greater than
7145 * what movablecore would have allowed.
7147 if (required_movablecore) {
7148 unsigned long corepages;
7151 * Round-up so that ZONE_MOVABLE is at least as large as what
7152 * was requested by the user
7154 required_movablecore =
7155 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7156 required_movablecore = min(totalpages, required_movablecore);
7157 corepages = totalpages - required_movablecore;
7159 required_kernelcore = max(required_kernelcore, corepages);
7163 * If kernelcore was not specified or kernelcore size is larger
7164 * than totalpages, there is no ZONE_MOVABLE.
7166 if (!required_kernelcore || required_kernelcore >= totalpages)
7169 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7170 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7173 /* Spread kernelcore memory as evenly as possible throughout nodes */
7174 kernelcore_node = required_kernelcore / usable_nodes;
7175 for_each_node_state(nid, N_MEMORY) {
7176 unsigned long start_pfn, end_pfn;
7179 * Recalculate kernelcore_node if the division per node
7180 * now exceeds what is necessary to satisfy the requested
7181 * amount of memory for the kernel
7183 if (required_kernelcore < kernelcore_node)
7184 kernelcore_node = required_kernelcore / usable_nodes;
7187 * As the map is walked, we track how much memory is usable
7188 * by the kernel using kernelcore_remaining. When it is
7189 * 0, the rest of the node is usable by ZONE_MOVABLE
7191 kernelcore_remaining = kernelcore_node;
7193 /* Go through each range of PFNs within this node */
7194 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7195 unsigned long size_pages;
7197 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7198 if (start_pfn >= end_pfn)
7201 /* Account for what is only usable for kernelcore */
7202 if (start_pfn < usable_startpfn) {
7203 unsigned long kernel_pages;
7204 kernel_pages = min(end_pfn, usable_startpfn)
7207 kernelcore_remaining -= min(kernel_pages,
7208 kernelcore_remaining);
7209 required_kernelcore -= min(kernel_pages,
7210 required_kernelcore);
7212 /* Continue if range is now fully accounted */
7213 if (end_pfn <= usable_startpfn) {
7216 * Push zone_movable_pfn to the end so
7217 * that if we have to rebalance
7218 * kernelcore across nodes, we will
7219 * not double account here
7221 zone_movable_pfn[nid] = end_pfn;
7224 start_pfn = usable_startpfn;
7228 * The usable PFN range for ZONE_MOVABLE is from
7229 * start_pfn->end_pfn. Calculate size_pages as the
7230 * number of pages used as kernelcore
7232 size_pages = end_pfn - start_pfn;
7233 if (size_pages > kernelcore_remaining)
7234 size_pages = kernelcore_remaining;
7235 zone_movable_pfn[nid] = start_pfn + size_pages;
7238 * Some kernelcore has been met, update counts and
7239 * break if the kernelcore for this node has been
7242 required_kernelcore -= min(required_kernelcore,
7244 kernelcore_remaining -= size_pages;
7245 if (!kernelcore_remaining)
7251 * If there is still required_kernelcore, we do another pass with one
7252 * less node in the count. This will push zone_movable_pfn[nid] further
7253 * along on the nodes that still have memory until kernelcore is
7257 if (usable_nodes && required_kernelcore > usable_nodes)
7261 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7262 for (nid = 0; nid < MAX_NUMNODES; nid++)
7263 zone_movable_pfn[nid] =
7264 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7267 /* restore the node_state */
7268 node_states[N_MEMORY] = saved_node_state;
7271 /* Any regular or high memory on that node ? */
7272 static void check_for_memory(pg_data_t *pgdat, int nid)
7274 enum zone_type zone_type;
7276 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7277 struct zone *zone = &pgdat->node_zones[zone_type];
7278 if (populated_zone(zone)) {
7279 if (IS_ENABLED(CONFIG_HIGHMEM))
7280 node_set_state(nid, N_HIGH_MEMORY);
7281 if (zone_type <= ZONE_NORMAL)
7282 node_set_state(nid, N_NORMAL_MEMORY);
7289 * free_area_init_nodes - Initialise all pg_data_t and zone data
7290 * @max_zone_pfn: an array of max PFNs for each zone
7292 * This will call free_area_init_node() for each active node in the system.
7293 * Using the page ranges provided by memblock_set_node(), the size of each
7294 * zone in each node and their holes is calculated. If the maximum PFN
7295 * between two adjacent zones match, it is assumed that the zone is empty.
7296 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7297 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7298 * starts where the previous one ended. For example, ZONE_DMA32 starts
7299 * at arch_max_dma_pfn.
7301 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
7303 unsigned long start_pfn, end_pfn;
7306 /* Record where the zone boundaries are */
7307 memset(arch_zone_lowest_possible_pfn, 0,
7308 sizeof(arch_zone_lowest_possible_pfn));
7309 memset(arch_zone_highest_possible_pfn, 0,
7310 sizeof(arch_zone_highest_possible_pfn));
7312 start_pfn = find_min_pfn_with_active_regions();
7314 for (i = 0; i < MAX_NR_ZONES; i++) {
7315 if (i == ZONE_MOVABLE)
7318 end_pfn = max(max_zone_pfn[i], start_pfn);
7319 arch_zone_lowest_possible_pfn[i] = start_pfn;
7320 arch_zone_highest_possible_pfn[i] = end_pfn;
7322 start_pfn = end_pfn;
7325 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7326 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7327 find_zone_movable_pfns_for_nodes();
7329 /* Print out the zone ranges */
7330 pr_info("Zone ranges:\n");
7331 for (i = 0; i < MAX_NR_ZONES; i++) {
7332 if (i == ZONE_MOVABLE)
7334 pr_info(" %-8s ", zone_names[i]);
7335 if (arch_zone_lowest_possible_pfn[i] ==
7336 arch_zone_highest_possible_pfn[i])
7339 pr_cont("[mem %#018Lx-%#018Lx]\n",
7340 (u64)arch_zone_lowest_possible_pfn[i]
7342 ((u64)arch_zone_highest_possible_pfn[i]
7343 << PAGE_SHIFT) - 1);
7346 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7347 pr_info("Movable zone start for each node\n");
7348 for (i = 0; i < MAX_NUMNODES; i++) {
7349 if (zone_movable_pfn[i])
7350 pr_info(" Node %d: %#018Lx\n", i,
7351 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7355 * Print out the early node map, and initialize the
7356 * subsection-map relative to active online memory ranges to
7357 * enable future "sub-section" extensions of the memory map.
7359 pr_info("Early memory node ranges\n");
7360 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7361 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7362 (u64)start_pfn << PAGE_SHIFT,
7363 ((u64)end_pfn << PAGE_SHIFT) - 1);
7364 subsection_map_init(start_pfn, end_pfn - start_pfn);
7367 /* Initialise every node */
7368 mminit_verify_pageflags_layout();
7369 setup_nr_node_ids();
7370 zero_resv_unavail();
7371 for_each_online_node(nid) {
7372 pg_data_t *pgdat = NODE_DATA(nid);
7373 free_area_init_node(nid, NULL,
7374 find_min_pfn_for_node(nid), NULL);
7376 /* Any memory on that node */
7377 if (pgdat->node_present_pages)
7378 node_set_state(nid, N_MEMORY);
7379 check_for_memory(pgdat, nid);
7383 static int __init cmdline_parse_core(char *p, unsigned long *core,
7384 unsigned long *percent)
7386 unsigned long long coremem;
7392 /* Value may be a percentage of total memory, otherwise bytes */
7393 coremem = simple_strtoull(p, &endptr, 0);
7394 if (*endptr == '%') {
7395 /* Paranoid check for percent values greater than 100 */
7396 WARN_ON(coremem > 100);
7400 coremem = memparse(p, &p);
7401 /* Paranoid check that UL is enough for the coremem value */
7402 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7404 *core = coremem >> PAGE_SHIFT;
7411 * kernelcore=size sets the amount of memory for use for allocations that
7412 * cannot be reclaimed or migrated.
7414 static int __init cmdline_parse_kernelcore(char *p)
7416 /* parse kernelcore=mirror */
7417 if (parse_option_str(p, "mirror")) {
7418 mirrored_kernelcore = true;
7422 return cmdline_parse_core(p, &required_kernelcore,
7423 &required_kernelcore_percent);
7427 * movablecore=size sets the amount of memory for use for allocations that
7428 * can be reclaimed or migrated.
7430 static int __init cmdline_parse_movablecore(char *p)
7432 return cmdline_parse_core(p, &required_movablecore,
7433 &required_movablecore_percent);
7436 early_param("kernelcore", cmdline_parse_kernelcore);
7437 early_param("movablecore", cmdline_parse_movablecore);
7439 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
7441 void adjust_managed_page_count(struct page *page, long count)
7443 atomic_long_add(count, &page_zone(page)->managed_pages);
7444 totalram_pages_add(count);
7445 #ifdef CONFIG_HIGHMEM
7446 if (PageHighMem(page))
7447 totalhigh_pages_add(count);
7450 EXPORT_SYMBOL(adjust_managed_page_count);
7452 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7455 unsigned long pages = 0;
7457 start = (void *)PAGE_ALIGN((unsigned long)start);
7458 end = (void *)((unsigned long)end & PAGE_MASK);
7459 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7460 struct page *page = virt_to_page(pos);
7461 void *direct_map_addr;
7464 * 'direct_map_addr' might be different from 'pos'
7465 * because some architectures' virt_to_page()
7466 * work with aliases. Getting the direct map
7467 * address ensures that we get a _writeable_
7468 * alias for the memset().
7470 direct_map_addr = page_address(page);
7471 if ((unsigned int)poison <= 0xFF)
7472 memset(direct_map_addr, poison, PAGE_SIZE);
7474 free_reserved_page(page);
7478 pr_info("Freeing %s memory: %ldK\n",
7479 s, pages << (PAGE_SHIFT - 10));
7484 #ifdef CONFIG_HIGHMEM
7485 void free_highmem_page(struct page *page)
7487 __free_reserved_page(page);
7488 totalram_pages_inc();
7489 atomic_long_inc(&page_zone(page)->managed_pages);
7490 totalhigh_pages_inc();
7495 void __init mem_init_print_info(const char *str)
7497 unsigned long physpages, codesize, datasize, rosize, bss_size;
7498 unsigned long init_code_size, init_data_size;
7500 physpages = get_num_physpages();
7501 codesize = _etext - _stext;
7502 datasize = _edata - _sdata;
7503 rosize = __end_rodata - __start_rodata;
7504 bss_size = __bss_stop - __bss_start;
7505 init_data_size = __init_end - __init_begin;
7506 init_code_size = _einittext - _sinittext;
7509 * Detect special cases and adjust section sizes accordingly:
7510 * 1) .init.* may be embedded into .data sections
7511 * 2) .init.text.* may be out of [__init_begin, __init_end],
7512 * please refer to arch/tile/kernel/vmlinux.lds.S.
7513 * 3) .rodata.* may be embedded into .text or .data sections.
7515 #define adj_init_size(start, end, size, pos, adj) \
7517 if (start <= pos && pos < end && size > adj) \
7521 adj_init_size(__init_begin, __init_end, init_data_size,
7522 _sinittext, init_code_size);
7523 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7524 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7525 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7526 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7528 #undef adj_init_size
7530 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7531 #ifdef CONFIG_HIGHMEM
7535 nr_free_pages() << (PAGE_SHIFT - 10),
7536 physpages << (PAGE_SHIFT - 10),
7537 codesize >> 10, datasize >> 10, rosize >> 10,
7538 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7539 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7540 totalcma_pages << (PAGE_SHIFT - 10),
7541 #ifdef CONFIG_HIGHMEM
7542 totalhigh_pages() << (PAGE_SHIFT - 10),
7544 str ? ", " : "", str ? str : "");
7548 * set_dma_reserve - set the specified number of pages reserved in the first zone
7549 * @new_dma_reserve: The number of pages to mark reserved
7551 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7552 * In the DMA zone, a significant percentage may be consumed by kernel image
7553 * and other unfreeable allocations which can skew the watermarks badly. This
7554 * function may optionally be used to account for unfreeable pages in the
7555 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7556 * smaller per-cpu batchsize.
7558 void __init set_dma_reserve(unsigned long new_dma_reserve)
7560 dma_reserve = new_dma_reserve;
7563 void __init free_area_init(unsigned long *zones_size)
7565 zero_resv_unavail();
7566 free_area_init_node(0, zones_size,
7567 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
7570 static int page_alloc_cpu_dead(unsigned int cpu)
7573 lru_add_drain_cpu(cpu);
7577 * Spill the event counters of the dead processor
7578 * into the current processors event counters.
7579 * This artificially elevates the count of the current
7582 vm_events_fold_cpu(cpu);
7585 * Zero the differential counters of the dead processor
7586 * so that the vm statistics are consistent.
7588 * This is only okay since the processor is dead and cannot
7589 * race with what we are doing.
7591 cpu_vm_stats_fold(cpu);
7596 int hashdist = HASHDIST_DEFAULT;
7598 static int __init set_hashdist(char *str)
7602 hashdist = simple_strtoul(str, &str, 0);
7605 __setup("hashdist=", set_hashdist);
7608 void __init page_alloc_init(void)
7613 if (num_node_state(N_MEMORY) == 1)
7617 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7618 "mm/page_alloc:dead", NULL,
7619 page_alloc_cpu_dead);
7624 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7625 * or min_free_kbytes changes.
7627 static void calculate_totalreserve_pages(void)
7629 struct pglist_data *pgdat;
7630 unsigned long reserve_pages = 0;
7631 enum zone_type i, j;
7633 for_each_online_pgdat(pgdat) {
7635 pgdat->totalreserve_pages = 0;
7637 for (i = 0; i < MAX_NR_ZONES; i++) {
7638 struct zone *zone = pgdat->node_zones + i;
7640 unsigned long managed_pages = zone_managed_pages(zone);
7642 /* Find valid and maximum lowmem_reserve in the zone */
7643 for (j = i; j < MAX_NR_ZONES; j++) {
7644 if (zone->lowmem_reserve[j] > max)
7645 max = zone->lowmem_reserve[j];
7648 /* we treat the high watermark as reserved pages. */
7649 max += high_wmark_pages(zone);
7651 if (max > managed_pages)
7652 max = managed_pages;
7654 pgdat->totalreserve_pages += max;
7656 reserve_pages += max;
7659 totalreserve_pages = reserve_pages;
7663 * setup_per_zone_lowmem_reserve - called whenever
7664 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7665 * has a correct pages reserved value, so an adequate number of
7666 * pages are left in the zone after a successful __alloc_pages().
7668 static void setup_per_zone_lowmem_reserve(void)
7670 struct pglist_data *pgdat;
7671 enum zone_type j, idx;
7673 for_each_online_pgdat(pgdat) {
7674 for (j = 0; j < MAX_NR_ZONES; j++) {
7675 struct zone *zone = pgdat->node_zones + j;
7676 unsigned long managed_pages = zone_managed_pages(zone);
7678 zone->lowmem_reserve[j] = 0;
7682 struct zone *lower_zone;
7685 lower_zone = pgdat->node_zones + idx;
7687 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7688 sysctl_lowmem_reserve_ratio[idx] = 0;
7689 lower_zone->lowmem_reserve[j] = 0;
7691 lower_zone->lowmem_reserve[j] =
7692 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7694 managed_pages += zone_managed_pages(lower_zone);
7699 /* update totalreserve_pages */
7700 calculate_totalreserve_pages();
7703 static void __setup_per_zone_wmarks(void)
7705 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7706 unsigned long lowmem_pages = 0;
7708 unsigned long flags;
7710 /* Calculate total number of !ZONE_HIGHMEM pages */
7711 for_each_zone(zone) {
7712 if (!is_highmem(zone))
7713 lowmem_pages += zone_managed_pages(zone);
7716 for_each_zone(zone) {
7719 spin_lock_irqsave(&zone->lock, flags);
7720 tmp = (u64)pages_min * zone_managed_pages(zone);
7721 do_div(tmp, lowmem_pages);
7722 if (is_highmem(zone)) {
7724 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7725 * need highmem pages, so cap pages_min to a small
7728 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7729 * deltas control async page reclaim, and so should
7730 * not be capped for highmem.
7732 unsigned long min_pages;
7734 min_pages = zone_managed_pages(zone) / 1024;
7735 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7736 zone->_watermark[WMARK_MIN] = min_pages;
7739 * If it's a lowmem zone, reserve a number of pages
7740 * proportionate to the zone's size.
7742 zone->_watermark[WMARK_MIN] = tmp;
7746 * Set the kswapd watermarks distance according to the
7747 * scale factor in proportion to available memory, but
7748 * ensure a minimum size on small systems.
7750 tmp = max_t(u64, tmp >> 2,
7751 mult_frac(zone_managed_pages(zone),
7752 watermark_scale_factor, 10000));
7754 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7755 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7756 zone->watermark_boost = 0;
7758 spin_unlock_irqrestore(&zone->lock, flags);
7761 /* update totalreserve_pages */
7762 calculate_totalreserve_pages();
7766 * setup_per_zone_wmarks - called when min_free_kbytes changes
7767 * or when memory is hot-{added|removed}
7769 * Ensures that the watermark[min,low,high] values for each zone are set
7770 * correctly with respect to min_free_kbytes.
7772 void setup_per_zone_wmarks(void)
7774 static DEFINE_SPINLOCK(lock);
7777 __setup_per_zone_wmarks();
7782 * Initialise min_free_kbytes.
7784 * For small machines we want it small (128k min). For large machines
7785 * we want it large (64MB max). But it is not linear, because network
7786 * bandwidth does not increase linearly with machine size. We use
7788 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7789 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7805 int __meminit init_per_zone_wmark_min(void)
7807 unsigned long lowmem_kbytes;
7808 int new_min_free_kbytes;
7810 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7811 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7813 if (new_min_free_kbytes > user_min_free_kbytes) {
7814 min_free_kbytes = new_min_free_kbytes;
7815 if (min_free_kbytes < 128)
7816 min_free_kbytes = 128;
7817 if (min_free_kbytes > 65536)
7818 min_free_kbytes = 65536;
7820 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7821 new_min_free_kbytes, user_min_free_kbytes);
7823 setup_per_zone_wmarks();
7824 refresh_zone_stat_thresholds();
7825 setup_per_zone_lowmem_reserve();
7828 setup_min_unmapped_ratio();
7829 setup_min_slab_ratio();
7834 core_initcall(init_per_zone_wmark_min)
7837 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7838 * that we can call two helper functions whenever min_free_kbytes
7841 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7842 void __user *buffer, size_t *length, loff_t *ppos)
7846 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7851 user_min_free_kbytes = min_free_kbytes;
7852 setup_per_zone_wmarks();
7857 int watermark_boost_factor_sysctl_handler(struct ctl_table *table, int write,
7858 void __user *buffer, size_t *length, loff_t *ppos)
7862 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7869 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7870 void __user *buffer, size_t *length, loff_t *ppos)
7874 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7879 setup_per_zone_wmarks();
7885 static void setup_min_unmapped_ratio(void)
7890 for_each_online_pgdat(pgdat)
7891 pgdat->min_unmapped_pages = 0;
7894 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7895 sysctl_min_unmapped_ratio) / 100;
7899 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7900 void __user *buffer, size_t *length, loff_t *ppos)
7904 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7908 setup_min_unmapped_ratio();
7913 static void setup_min_slab_ratio(void)
7918 for_each_online_pgdat(pgdat)
7919 pgdat->min_slab_pages = 0;
7922 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7923 sysctl_min_slab_ratio) / 100;
7926 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7927 void __user *buffer, size_t *length, loff_t *ppos)
7931 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7935 setup_min_slab_ratio();
7942 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7943 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7944 * whenever sysctl_lowmem_reserve_ratio changes.
7946 * The reserve ratio obviously has absolutely no relation with the
7947 * minimum watermarks. The lowmem reserve ratio can only make sense
7948 * if in function of the boot time zone sizes.
7950 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7951 void __user *buffer, size_t *length, loff_t *ppos)
7953 proc_dointvec_minmax(table, write, buffer, length, ppos);
7954 setup_per_zone_lowmem_reserve();
7959 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7960 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7961 * pagelist can have before it gets flushed back to buddy allocator.
7963 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7964 void __user *buffer, size_t *length, loff_t *ppos)
7967 int old_percpu_pagelist_fraction;
7970 mutex_lock(&pcp_batch_high_lock);
7971 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7973 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7974 if (!write || ret < 0)
7977 /* Sanity checking to avoid pcp imbalance */
7978 if (percpu_pagelist_fraction &&
7979 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7980 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7986 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7989 for_each_populated_zone(zone) {
7992 for_each_possible_cpu(cpu)
7993 pageset_set_high_and_batch(zone,
7994 per_cpu_ptr(zone->pageset, cpu));
7997 mutex_unlock(&pcp_batch_high_lock);
8001 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8003 * Returns the number of pages that arch has reserved but
8004 * is not known to alloc_large_system_hash().
8006 static unsigned long __init arch_reserved_kernel_pages(void)
8013 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8014 * machines. As memory size is increased the scale is also increased but at
8015 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8016 * quadruples the scale is increased by one, which means the size of hash table
8017 * only doubles, instead of quadrupling as well.
8018 * Because 32-bit systems cannot have large physical memory, where this scaling
8019 * makes sense, it is disabled on such platforms.
8021 #if __BITS_PER_LONG > 32
8022 #define ADAPT_SCALE_BASE (64ul << 30)
8023 #define ADAPT_SCALE_SHIFT 2
8024 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8028 * allocate a large system hash table from bootmem
8029 * - it is assumed that the hash table must contain an exact power-of-2
8030 * quantity of entries
8031 * - limit is the number of hash buckets, not the total allocation size
8033 void *__init alloc_large_system_hash(const char *tablename,
8034 unsigned long bucketsize,
8035 unsigned long numentries,
8038 unsigned int *_hash_shift,
8039 unsigned int *_hash_mask,
8040 unsigned long low_limit,
8041 unsigned long high_limit)
8043 unsigned long long max = high_limit;
8044 unsigned long log2qty, size;
8049 /* allow the kernel cmdline to have a say */
8051 /* round applicable memory size up to nearest megabyte */
8052 numentries = nr_kernel_pages;
8053 numentries -= arch_reserved_kernel_pages();
8055 /* It isn't necessary when PAGE_SIZE >= 1MB */
8056 if (PAGE_SHIFT < 20)
8057 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8059 #if __BITS_PER_LONG > 32
8061 unsigned long adapt;
8063 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8064 adapt <<= ADAPT_SCALE_SHIFT)
8069 /* limit to 1 bucket per 2^scale bytes of low memory */
8070 if (scale > PAGE_SHIFT)
8071 numentries >>= (scale - PAGE_SHIFT);
8073 numentries <<= (PAGE_SHIFT - scale);
8075 /* Make sure we've got at least a 0-order allocation.. */
8076 if (unlikely(flags & HASH_SMALL)) {
8077 /* Makes no sense without HASH_EARLY */
8078 WARN_ON(!(flags & HASH_EARLY));
8079 if (!(numentries >> *_hash_shift)) {
8080 numentries = 1UL << *_hash_shift;
8081 BUG_ON(!numentries);
8083 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8084 numentries = PAGE_SIZE / bucketsize;
8086 numentries = roundup_pow_of_two(numentries);
8088 /* limit allocation size to 1/16 total memory by default */
8090 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8091 do_div(max, bucketsize);
8093 max = min(max, 0x80000000ULL);
8095 if (numentries < low_limit)
8096 numentries = low_limit;
8097 if (numentries > max)
8100 log2qty = ilog2(numentries);
8102 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8105 size = bucketsize << log2qty;
8106 if (flags & HASH_EARLY) {
8107 if (flags & HASH_ZERO)
8108 table = memblock_alloc(size, SMP_CACHE_BYTES);
8110 table = memblock_alloc_raw(size,
8112 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8113 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
8117 * If bucketsize is not a power-of-two, we may free
8118 * some pages at the end of hash table which
8119 * alloc_pages_exact() automatically does
8121 table = alloc_pages_exact(size, gfp_flags);
8122 kmemleak_alloc(table, size, 1, gfp_flags);
8124 } while (!table && size > PAGE_SIZE && --log2qty);
8127 panic("Failed to allocate %s hash table\n", tablename);
8129 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8130 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8131 virt ? "vmalloc" : "linear");
8134 *_hash_shift = log2qty;
8136 *_hash_mask = (1 << log2qty) - 1;
8142 * This function checks whether pageblock includes unmovable pages or not.
8143 * If @count is not zero, it is okay to include less @count unmovable pages
8145 * PageLRU check without isolation or lru_lock could race so that
8146 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8147 * check without lock_page also may miss some movable non-lru pages at
8148 * race condition. So you can't expect this function should be exact.
8150 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
8151 int migratetype, int flags)
8153 unsigned long found;
8154 unsigned long iter = 0;
8155 unsigned long pfn = page_to_pfn(page);
8156 const char *reason = "unmovable page";
8159 * TODO we could make this much more efficient by not checking every
8160 * page in the range if we know all of them are in MOVABLE_ZONE and
8161 * that the movable zone guarantees that pages are migratable but
8162 * the later is not the case right now unfortunatelly. E.g. movablecore
8163 * can still lead to having bootmem allocations in zone_movable.
8166 if (is_migrate_cma_page(page)) {
8168 * CMA allocations (alloc_contig_range) really need to mark
8169 * isolate CMA pageblocks even when they are not movable in fact
8170 * so consider them movable here.
8172 if (is_migrate_cma(migratetype))
8175 reason = "CMA page";
8179 for (found = 0; iter < pageblock_nr_pages; iter++) {
8180 unsigned long check = pfn + iter;
8182 if (!pfn_valid_within(check))
8185 page = pfn_to_page(check);
8187 if (PageReserved(page))
8191 * If the zone is movable and we have ruled out all reserved
8192 * pages then it should be reasonably safe to assume the rest
8195 if (zone_idx(zone) == ZONE_MOVABLE)
8199 * Hugepages are not in LRU lists, but they're movable.
8200 * We need not scan over tail pages because we don't
8201 * handle each tail page individually in migration.
8203 if (PageHuge(page)) {
8204 struct page *head = compound_head(page);
8205 unsigned int skip_pages;
8207 if (!hugepage_migration_supported(page_hstate(head)))
8210 skip_pages = compound_nr(head) - (page - head);
8211 iter += skip_pages - 1;
8216 * We can't use page_count without pin a page
8217 * because another CPU can free compound page.
8218 * This check already skips compound tails of THP
8219 * because their page->_refcount is zero at all time.
8221 if (!page_ref_count(page)) {
8222 if (PageBuddy(page))
8223 iter += (1 << page_order(page)) - 1;
8228 * The HWPoisoned page may be not in buddy system, and
8229 * page_count() is not 0.
8231 if ((flags & SKIP_HWPOISON) && PageHWPoison(page))
8234 if (__PageMovable(page))
8240 * If there are RECLAIMABLE pages, we need to check
8241 * it. But now, memory offline itself doesn't call
8242 * shrink_node_slabs() and it still to be fixed.
8245 * If the page is not RAM, page_count()should be 0.
8246 * we don't need more check. This is an _used_ not-movable page.
8248 * The problematic thing here is PG_reserved pages. PG_reserved
8249 * is set to both of a memory hole page and a _used_ kernel
8257 WARN_ON_ONCE(zone_idx(zone) == ZONE_MOVABLE);
8258 if (flags & REPORT_FAILURE)
8259 dump_page(pfn_to_page(pfn + iter), reason);
8263 #ifdef CONFIG_CONTIG_ALLOC
8264 static unsigned long pfn_max_align_down(unsigned long pfn)
8266 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8267 pageblock_nr_pages) - 1);
8270 static unsigned long pfn_max_align_up(unsigned long pfn)
8272 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8273 pageblock_nr_pages));
8276 /* [start, end) must belong to a single zone. */
8277 static int __alloc_contig_migrate_range(struct compact_control *cc,
8278 unsigned long start, unsigned long end)
8280 /* This function is based on compact_zone() from compaction.c. */
8281 unsigned long nr_reclaimed;
8282 unsigned long pfn = start;
8283 unsigned int tries = 0;
8288 while (pfn < end || !list_empty(&cc->migratepages)) {
8289 if (fatal_signal_pending(current)) {
8294 if (list_empty(&cc->migratepages)) {
8295 cc->nr_migratepages = 0;
8296 pfn = isolate_migratepages_range(cc, pfn, end);
8302 } else if (++tries == 5) {
8303 ret = ret < 0 ? ret : -EBUSY;
8307 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8309 cc->nr_migratepages -= nr_reclaimed;
8311 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
8312 NULL, 0, cc->mode, MR_CONTIG_RANGE);
8315 putback_movable_pages(&cc->migratepages);
8322 * alloc_contig_range() -- tries to allocate given range of pages
8323 * @start: start PFN to allocate
8324 * @end: one-past-the-last PFN to allocate
8325 * @migratetype: migratetype of the underlaying pageblocks (either
8326 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8327 * in range must have the same migratetype and it must
8328 * be either of the two.
8329 * @gfp_mask: GFP mask to use during compaction
8331 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8332 * aligned. The PFN range must belong to a single zone.
8334 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8335 * pageblocks in the range. Once isolated, the pageblocks should not
8336 * be modified by others.
8338 * Return: zero on success or negative error code. On success all
8339 * pages which PFN is in [start, end) are allocated for the caller and
8340 * need to be freed with free_contig_range().
8342 int alloc_contig_range(unsigned long start, unsigned long end,
8343 unsigned migratetype, gfp_t gfp_mask)
8345 unsigned long outer_start, outer_end;
8349 struct compact_control cc = {
8350 .nr_migratepages = 0,
8352 .zone = page_zone(pfn_to_page(start)),
8353 .mode = MIGRATE_SYNC,
8354 .ignore_skip_hint = true,
8355 .no_set_skip_hint = true,
8356 .gfp_mask = current_gfp_context(gfp_mask),
8358 INIT_LIST_HEAD(&cc.migratepages);
8361 * What we do here is we mark all pageblocks in range as
8362 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8363 * have different sizes, and due to the way page allocator
8364 * work, we align the range to biggest of the two pages so
8365 * that page allocator won't try to merge buddies from
8366 * different pageblocks and change MIGRATE_ISOLATE to some
8367 * other migration type.
8369 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8370 * migrate the pages from an unaligned range (ie. pages that
8371 * we are interested in). This will put all the pages in
8372 * range back to page allocator as MIGRATE_ISOLATE.
8374 * When this is done, we take the pages in range from page
8375 * allocator removing them from the buddy system. This way
8376 * page allocator will never consider using them.
8378 * This lets us mark the pageblocks back as
8379 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8380 * aligned range but not in the unaligned, original range are
8381 * put back to page allocator so that buddy can use them.
8384 ret = start_isolate_page_range(pfn_max_align_down(start),
8385 pfn_max_align_up(end), migratetype, 0);
8390 * In case of -EBUSY, we'd like to know which page causes problem.
8391 * So, just fall through. test_pages_isolated() has a tracepoint
8392 * which will report the busy page.
8394 * It is possible that busy pages could become available before
8395 * the call to test_pages_isolated, and the range will actually be
8396 * allocated. So, if we fall through be sure to clear ret so that
8397 * -EBUSY is not accidentally used or returned to caller.
8399 ret = __alloc_contig_migrate_range(&cc, start, end);
8400 if (ret && ret != -EBUSY)
8405 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8406 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8407 * more, all pages in [start, end) are free in page allocator.
8408 * What we are going to do is to allocate all pages from
8409 * [start, end) (that is remove them from page allocator).
8411 * The only problem is that pages at the beginning and at the
8412 * end of interesting range may be not aligned with pages that
8413 * page allocator holds, ie. they can be part of higher order
8414 * pages. Because of this, we reserve the bigger range and
8415 * once this is done free the pages we are not interested in.
8417 * We don't have to hold zone->lock here because the pages are
8418 * isolated thus they won't get removed from buddy.
8421 lru_add_drain_all();
8424 outer_start = start;
8425 while (!PageBuddy(pfn_to_page(outer_start))) {
8426 if (++order >= MAX_ORDER) {
8427 outer_start = start;
8430 outer_start &= ~0UL << order;
8433 if (outer_start != start) {
8434 order = page_order(pfn_to_page(outer_start));
8437 * outer_start page could be small order buddy page and
8438 * it doesn't include start page. Adjust outer_start
8439 * in this case to report failed page properly
8440 * on tracepoint in test_pages_isolated()
8442 if (outer_start + (1UL << order) <= start)
8443 outer_start = start;
8446 /* Make sure the range is really isolated. */
8447 if (test_pages_isolated(outer_start, end, false)) {
8448 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8449 __func__, outer_start, end);
8454 /* Grab isolated pages from freelists. */
8455 outer_end = isolate_freepages_range(&cc, outer_start, end);
8461 /* Free head and tail (if any) */
8462 if (start != outer_start)
8463 free_contig_range(outer_start, start - outer_start);
8464 if (end != outer_end)
8465 free_contig_range(end, outer_end - end);
8468 undo_isolate_page_range(pfn_max_align_down(start),
8469 pfn_max_align_up(end), migratetype);
8472 #endif /* CONFIG_CONTIG_ALLOC */
8474 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8476 unsigned int count = 0;
8478 for (; nr_pages--; pfn++) {
8479 struct page *page = pfn_to_page(pfn);
8481 count += page_count(page) != 1;
8484 WARN(count != 0, "%d pages are still in use!\n", count);
8487 #ifdef CONFIG_MEMORY_HOTPLUG
8489 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8490 * page high values need to be recalulated.
8492 void __meminit zone_pcp_update(struct zone *zone)
8495 mutex_lock(&pcp_batch_high_lock);
8496 for_each_possible_cpu(cpu)
8497 pageset_set_high_and_batch(zone,
8498 per_cpu_ptr(zone->pageset, cpu));
8499 mutex_unlock(&pcp_batch_high_lock);
8503 void zone_pcp_reset(struct zone *zone)
8505 unsigned long flags;
8507 struct per_cpu_pageset *pset;
8509 /* avoid races with drain_pages() */
8510 local_irq_save(flags);
8511 if (zone->pageset != &boot_pageset) {
8512 for_each_online_cpu(cpu) {
8513 pset = per_cpu_ptr(zone->pageset, cpu);
8514 drain_zonestat(zone, pset);
8516 free_percpu(zone->pageset);
8517 zone->pageset = &boot_pageset;
8519 local_irq_restore(flags);
8522 #ifdef CONFIG_MEMORY_HOTREMOVE
8524 * All pages in the range must be in a single zone and isolated
8525 * before calling this.
8528 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8532 unsigned int order, i;
8534 unsigned long flags;
8535 unsigned long offlined_pages = 0;
8537 /* find the first valid pfn */
8538 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8542 return offlined_pages;
8544 offline_mem_sections(pfn, end_pfn);
8545 zone = page_zone(pfn_to_page(pfn));
8546 spin_lock_irqsave(&zone->lock, flags);
8548 while (pfn < end_pfn) {
8549 if (!pfn_valid(pfn)) {
8553 page = pfn_to_page(pfn);
8555 * The HWPoisoned page may be not in buddy system, and
8556 * page_count() is not 0.
8558 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8560 SetPageReserved(page);
8565 BUG_ON(page_count(page));
8566 BUG_ON(!PageBuddy(page));
8567 order = page_order(page);
8568 offlined_pages += 1 << order;
8569 #ifdef CONFIG_DEBUG_VM
8570 pr_info("remove from free list %lx %d %lx\n",
8571 pfn, 1 << order, end_pfn);
8573 del_page_from_free_area(page, &zone->free_area[order]);
8574 for (i = 0; i < (1 << order); i++)
8575 SetPageReserved((page+i));
8576 pfn += (1 << order);
8578 spin_unlock_irqrestore(&zone->lock, flags);
8580 return offlined_pages;
8584 bool is_free_buddy_page(struct page *page)
8586 struct zone *zone = page_zone(page);
8587 unsigned long pfn = page_to_pfn(page);
8588 unsigned long flags;
8591 spin_lock_irqsave(&zone->lock, flags);
8592 for (order = 0; order < MAX_ORDER; order++) {
8593 struct page *page_head = page - (pfn & ((1 << order) - 1));
8595 if (PageBuddy(page_head) && page_order(page_head) >= order)
8598 spin_unlock_irqrestore(&zone->lock, flags);
8600 return order < MAX_ORDER;
8603 #ifdef CONFIG_MEMORY_FAILURE
8605 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8606 * test is performed under the zone lock to prevent a race against page
8609 bool set_hwpoison_free_buddy_page(struct page *page)
8611 struct zone *zone = page_zone(page);
8612 unsigned long pfn = page_to_pfn(page);
8613 unsigned long flags;
8615 bool hwpoisoned = false;
8617 spin_lock_irqsave(&zone->lock, flags);
8618 for (order = 0; order < MAX_ORDER; order++) {
8619 struct page *page_head = page - (pfn & ((1 << order) - 1));
8621 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8622 if (!TestSetPageHWPoison(page))
8627 spin_unlock_irqrestore(&zone->lock, flags);