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 if (want_init_on_free())
1179 kernel_init_free_pages(page, 1 << order);
1181 kernel_poison_pages(page, 1 << order, 0);
1183 * arch_free_page() can make the page's contents inaccessible. s390
1184 * does this. So nothing which can access the page's contents should
1185 * happen after this.
1187 arch_free_page(page, order);
1189 if (debug_pagealloc_enabled())
1190 kernel_map_pages(page, 1 << order, 0);
1192 kasan_free_nondeferred_pages(page, order);
1197 #ifdef CONFIG_DEBUG_VM
1199 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1200 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1201 * moved from pcp lists to free lists.
1203 static bool free_pcp_prepare(struct page *page)
1205 return free_pages_prepare(page, 0, true);
1208 static bool bulkfree_pcp_prepare(struct page *page)
1210 if (debug_pagealloc_enabled())
1211 return free_pages_check(page);
1217 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1218 * moving from pcp lists to free list in order to reduce overhead. With
1219 * debug_pagealloc enabled, they are checked also immediately when being freed
1222 static bool free_pcp_prepare(struct page *page)
1224 if (debug_pagealloc_enabled())
1225 return free_pages_prepare(page, 0, true);
1227 return free_pages_prepare(page, 0, false);
1230 static bool bulkfree_pcp_prepare(struct page *page)
1232 return free_pages_check(page);
1234 #endif /* CONFIG_DEBUG_VM */
1236 static inline void prefetch_buddy(struct page *page)
1238 unsigned long pfn = page_to_pfn(page);
1239 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1240 struct page *buddy = page + (buddy_pfn - pfn);
1246 * Frees a number of pages from the PCP lists
1247 * Assumes all pages on list are in same zone, and of same order.
1248 * count is the number of pages to free.
1250 * If the zone was previously in an "all pages pinned" state then look to
1251 * see if this freeing clears that state.
1253 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1254 * pinned" detection logic.
1256 static void free_pcppages_bulk(struct zone *zone, int count,
1257 struct per_cpu_pages *pcp)
1259 int migratetype = 0;
1261 int prefetch_nr = 0;
1262 bool isolated_pageblocks;
1263 struct page *page, *tmp;
1267 struct list_head *list;
1270 * Remove pages from lists in a round-robin fashion. A
1271 * batch_free count is maintained that is incremented when an
1272 * empty list is encountered. This is so more pages are freed
1273 * off fuller lists instead of spinning excessively around empty
1278 if (++migratetype == MIGRATE_PCPTYPES)
1280 list = &pcp->lists[migratetype];
1281 } while (list_empty(list));
1283 /* This is the only non-empty list. Free them all. */
1284 if (batch_free == MIGRATE_PCPTYPES)
1288 page = list_last_entry(list, struct page, lru);
1289 /* must delete to avoid corrupting pcp list */
1290 list_del(&page->lru);
1293 if (bulkfree_pcp_prepare(page))
1296 list_add_tail(&page->lru, &head);
1299 * We are going to put the page back to the global
1300 * pool, prefetch its buddy to speed up later access
1301 * under zone->lock. It is believed the overhead of
1302 * an additional test and calculating buddy_pfn here
1303 * can be offset by reduced memory latency later. To
1304 * avoid excessive prefetching due to large count, only
1305 * prefetch buddy for the first pcp->batch nr of pages.
1307 if (prefetch_nr++ < pcp->batch)
1308 prefetch_buddy(page);
1309 } while (--count && --batch_free && !list_empty(list));
1312 spin_lock(&zone->lock);
1313 isolated_pageblocks = has_isolate_pageblock(zone);
1316 * Use safe version since after __free_one_page(),
1317 * page->lru.next will not point to original list.
1319 list_for_each_entry_safe(page, tmp, &head, lru) {
1320 int mt = get_pcppage_migratetype(page);
1321 /* MIGRATE_ISOLATE page should not go to pcplists */
1322 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1323 /* Pageblock could have been isolated meanwhile */
1324 if (unlikely(isolated_pageblocks))
1325 mt = get_pageblock_migratetype(page);
1327 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1328 trace_mm_page_pcpu_drain(page, 0, mt);
1330 spin_unlock(&zone->lock);
1333 static void free_one_page(struct zone *zone,
1334 struct page *page, unsigned long pfn,
1338 spin_lock(&zone->lock);
1339 if (unlikely(has_isolate_pageblock(zone) ||
1340 is_migrate_isolate(migratetype))) {
1341 migratetype = get_pfnblock_migratetype(page, pfn);
1343 __free_one_page(page, pfn, zone, order, migratetype);
1344 spin_unlock(&zone->lock);
1347 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1348 unsigned long zone, int nid)
1350 mm_zero_struct_page(page);
1351 set_page_links(page, zone, nid, pfn);
1352 init_page_count(page);
1353 page_mapcount_reset(page);
1354 page_cpupid_reset_last(page);
1355 page_kasan_tag_reset(page);
1357 INIT_LIST_HEAD(&page->lru);
1358 #ifdef WANT_PAGE_VIRTUAL
1359 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1360 if (!is_highmem_idx(zone))
1361 set_page_address(page, __va(pfn << PAGE_SHIFT));
1365 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1366 static void __meminit init_reserved_page(unsigned long pfn)
1371 if (!early_page_uninitialised(pfn))
1374 nid = early_pfn_to_nid(pfn);
1375 pgdat = NODE_DATA(nid);
1377 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1378 struct zone *zone = &pgdat->node_zones[zid];
1380 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1383 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1386 static inline void init_reserved_page(unsigned long pfn)
1389 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1392 * Initialised pages do not have PageReserved set. This function is
1393 * called for each range allocated by the bootmem allocator and
1394 * marks the pages PageReserved. The remaining valid pages are later
1395 * sent to the buddy page allocator.
1397 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1399 unsigned long start_pfn = PFN_DOWN(start);
1400 unsigned long end_pfn = PFN_UP(end);
1402 for (; start_pfn < end_pfn; start_pfn++) {
1403 if (pfn_valid(start_pfn)) {
1404 struct page *page = pfn_to_page(start_pfn);
1406 init_reserved_page(start_pfn);
1408 /* Avoid false-positive PageTail() */
1409 INIT_LIST_HEAD(&page->lru);
1412 * no need for atomic set_bit because the struct
1413 * page is not visible yet so nobody should
1416 __SetPageReserved(page);
1421 static void __free_pages_ok(struct page *page, unsigned int order)
1423 unsigned long flags;
1425 unsigned long pfn = page_to_pfn(page);
1427 if (!free_pages_prepare(page, order, true))
1430 migratetype = get_pfnblock_migratetype(page, pfn);
1431 local_irq_save(flags);
1432 __count_vm_events(PGFREE, 1 << order);
1433 free_one_page(page_zone(page), page, pfn, order, migratetype);
1434 local_irq_restore(flags);
1437 void __free_pages_core(struct page *page, unsigned int order)
1439 unsigned int nr_pages = 1 << order;
1440 struct page *p = page;
1444 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1446 __ClearPageReserved(p);
1447 set_page_count(p, 0);
1449 __ClearPageReserved(p);
1450 set_page_count(p, 0);
1452 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1453 set_page_refcounted(page);
1454 __free_pages(page, order);
1457 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1458 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1460 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1462 int __meminit early_pfn_to_nid(unsigned long pfn)
1464 static DEFINE_SPINLOCK(early_pfn_lock);
1467 spin_lock(&early_pfn_lock);
1468 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1470 nid = first_online_node;
1471 spin_unlock(&early_pfn_lock);
1477 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1478 /* Only safe to use early in boot when initialisation is single-threaded */
1479 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1483 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1484 if (nid >= 0 && nid != node)
1490 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1497 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1500 if (early_page_uninitialised(pfn))
1502 __free_pages_core(page, order);
1506 * Check that the whole (or subset of) a pageblock given by the interval of
1507 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1508 * with the migration of free compaction scanner. The scanners then need to
1509 * use only pfn_valid_within() check for arches that allow holes within
1512 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1514 * It's possible on some configurations to have a setup like node0 node1 node0
1515 * i.e. it's possible that all pages within a zones range of pages do not
1516 * belong to a single zone. We assume that a border between node0 and node1
1517 * can occur within a single pageblock, but not a node0 node1 node0
1518 * interleaving within a single pageblock. It is therefore sufficient to check
1519 * the first and last page of a pageblock and avoid checking each individual
1520 * page in a pageblock.
1522 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1523 unsigned long end_pfn, struct zone *zone)
1525 struct page *start_page;
1526 struct page *end_page;
1528 /* end_pfn is one past the range we are checking */
1531 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1534 start_page = pfn_to_online_page(start_pfn);
1538 if (page_zone(start_page) != zone)
1541 end_page = pfn_to_page(end_pfn);
1543 /* This gives a shorter code than deriving page_zone(end_page) */
1544 if (page_zone_id(start_page) != page_zone_id(end_page))
1550 void set_zone_contiguous(struct zone *zone)
1552 unsigned long block_start_pfn = zone->zone_start_pfn;
1553 unsigned long block_end_pfn;
1555 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1556 for (; block_start_pfn < zone_end_pfn(zone);
1557 block_start_pfn = block_end_pfn,
1558 block_end_pfn += pageblock_nr_pages) {
1560 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1562 if (!__pageblock_pfn_to_page(block_start_pfn,
1563 block_end_pfn, zone))
1567 /* We confirm that there is no hole */
1568 zone->contiguous = true;
1571 void clear_zone_contiguous(struct zone *zone)
1573 zone->contiguous = false;
1576 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1577 static void __init deferred_free_range(unsigned long pfn,
1578 unsigned long nr_pages)
1586 page = pfn_to_page(pfn);
1588 /* Free a large naturally-aligned chunk if possible */
1589 if (nr_pages == pageblock_nr_pages &&
1590 (pfn & (pageblock_nr_pages - 1)) == 0) {
1591 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1592 __free_pages_core(page, pageblock_order);
1596 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1597 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1598 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1599 __free_pages_core(page, 0);
1603 /* Completion tracking for deferred_init_memmap() threads */
1604 static atomic_t pgdat_init_n_undone __initdata;
1605 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1607 static inline void __init pgdat_init_report_one_done(void)
1609 if (atomic_dec_and_test(&pgdat_init_n_undone))
1610 complete(&pgdat_init_all_done_comp);
1614 * Returns true if page needs to be initialized or freed to buddy allocator.
1616 * First we check if pfn is valid on architectures where it is possible to have
1617 * holes within pageblock_nr_pages. On systems where it is not possible, this
1618 * function is optimized out.
1620 * Then, we check if a current large page is valid by only checking the validity
1623 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1625 if (!pfn_valid_within(pfn))
1627 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1633 * Free pages to buddy allocator. Try to free aligned pages in
1634 * pageblock_nr_pages sizes.
1636 static void __init deferred_free_pages(unsigned long pfn,
1637 unsigned long end_pfn)
1639 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1640 unsigned long nr_free = 0;
1642 for (; pfn < end_pfn; pfn++) {
1643 if (!deferred_pfn_valid(pfn)) {
1644 deferred_free_range(pfn - nr_free, nr_free);
1646 } else if (!(pfn & nr_pgmask)) {
1647 deferred_free_range(pfn - nr_free, nr_free);
1649 touch_nmi_watchdog();
1654 /* Free the last block of pages to allocator */
1655 deferred_free_range(pfn - nr_free, nr_free);
1659 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1660 * by performing it only once every pageblock_nr_pages.
1661 * Return number of pages initialized.
1663 static unsigned long __init deferred_init_pages(struct zone *zone,
1665 unsigned long end_pfn)
1667 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1668 int nid = zone_to_nid(zone);
1669 unsigned long nr_pages = 0;
1670 int zid = zone_idx(zone);
1671 struct page *page = NULL;
1673 for (; pfn < end_pfn; pfn++) {
1674 if (!deferred_pfn_valid(pfn)) {
1677 } else if (!page || !(pfn & nr_pgmask)) {
1678 page = pfn_to_page(pfn);
1679 touch_nmi_watchdog();
1683 __init_single_page(page, pfn, zid, nid);
1690 * This function is meant to pre-load the iterator for the zone init.
1691 * Specifically it walks through the ranges until we are caught up to the
1692 * first_init_pfn value and exits there. If we never encounter the value we
1693 * return false indicating there are no valid ranges left.
1696 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1697 unsigned long *spfn, unsigned long *epfn,
1698 unsigned long first_init_pfn)
1703 * Start out by walking through the ranges in this zone that have
1704 * already been initialized. We don't need to do anything with them
1705 * so we just need to flush them out of the system.
1707 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1708 if (*epfn <= first_init_pfn)
1710 if (*spfn < first_init_pfn)
1711 *spfn = first_init_pfn;
1720 * Initialize and free pages. We do it in two loops: first we initialize
1721 * struct page, then free to buddy allocator, because while we are
1722 * freeing pages we can access pages that are ahead (computing buddy
1723 * page in __free_one_page()).
1725 * In order to try and keep some memory in the cache we have the loop
1726 * broken along max page order boundaries. This way we will not cause
1727 * any issues with the buddy page computation.
1729 static unsigned long __init
1730 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1731 unsigned long *end_pfn)
1733 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1734 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1735 unsigned long nr_pages = 0;
1738 /* First we loop through and initialize the page values */
1739 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1742 if (mo_pfn <= *start_pfn)
1745 t = min(mo_pfn, *end_pfn);
1746 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1748 if (mo_pfn < *end_pfn) {
1749 *start_pfn = mo_pfn;
1754 /* Reset values and now loop through freeing pages as needed */
1757 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1763 t = min(mo_pfn, epfn);
1764 deferred_free_pages(spfn, t);
1773 /* Initialise remaining memory on a node */
1774 static int __init deferred_init_memmap(void *data)
1776 pg_data_t *pgdat = data;
1777 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1778 unsigned long spfn = 0, epfn = 0, nr_pages = 0;
1779 unsigned long first_init_pfn, flags;
1780 unsigned long start = jiffies;
1785 /* Bind memory initialisation thread to a local node if possible */
1786 if (!cpumask_empty(cpumask))
1787 set_cpus_allowed_ptr(current, cpumask);
1789 pgdat_resize_lock(pgdat, &flags);
1790 first_init_pfn = pgdat->first_deferred_pfn;
1791 if (first_init_pfn == ULONG_MAX) {
1792 pgdat_resize_unlock(pgdat, &flags);
1793 pgdat_init_report_one_done();
1797 /* Sanity check boundaries */
1798 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1799 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1800 pgdat->first_deferred_pfn = ULONG_MAX;
1802 /* Only the highest zone is deferred so find it */
1803 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1804 zone = pgdat->node_zones + zid;
1805 if (first_init_pfn < zone_end_pfn(zone))
1809 /* If the zone is empty somebody else may have cleared out the zone */
1810 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1815 * Initialize and free pages in MAX_ORDER sized increments so
1816 * that we can avoid introducing any issues with the buddy
1820 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1822 pgdat_resize_unlock(pgdat, &flags);
1824 /* Sanity check that the next zone really is unpopulated */
1825 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1827 pr_info("node %d initialised, %lu pages in %ums\n",
1828 pgdat->node_id, nr_pages, jiffies_to_msecs(jiffies - start));
1830 pgdat_init_report_one_done();
1835 * If this zone has deferred pages, try to grow it by initializing enough
1836 * deferred pages to satisfy the allocation specified by order, rounded up to
1837 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1838 * of SECTION_SIZE bytes by initializing struct pages in increments of
1839 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1841 * Return true when zone was grown, otherwise return false. We return true even
1842 * when we grow less than requested, to let the caller decide if there are
1843 * enough pages to satisfy the allocation.
1845 * Note: We use noinline because this function is needed only during boot, and
1846 * it is called from a __ref function _deferred_grow_zone. This way we are
1847 * making sure that it is not inlined into permanent text section.
1849 static noinline bool __init
1850 deferred_grow_zone(struct zone *zone, unsigned int order)
1852 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1853 pg_data_t *pgdat = zone->zone_pgdat;
1854 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1855 unsigned long spfn, epfn, flags;
1856 unsigned long nr_pages = 0;
1859 /* Only the last zone may have deferred pages */
1860 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1863 pgdat_resize_lock(pgdat, &flags);
1866 * If deferred pages have been initialized while we were waiting for
1867 * the lock, return true, as the zone was grown. The caller will retry
1868 * this zone. We won't return to this function since the caller also
1869 * has this static branch.
1871 if (!static_branch_unlikely(&deferred_pages)) {
1872 pgdat_resize_unlock(pgdat, &flags);
1877 * If someone grew this zone while we were waiting for spinlock, return
1878 * true, as there might be enough pages already.
1880 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1881 pgdat_resize_unlock(pgdat, &flags);
1885 /* If the zone is empty somebody else may have cleared out the zone */
1886 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1887 first_deferred_pfn)) {
1888 pgdat->first_deferred_pfn = ULONG_MAX;
1889 pgdat_resize_unlock(pgdat, &flags);
1890 /* Retry only once. */
1891 return first_deferred_pfn != ULONG_MAX;
1895 * Initialize and free pages in MAX_ORDER sized increments so
1896 * that we can avoid introducing any issues with the buddy
1899 while (spfn < epfn) {
1900 /* update our first deferred PFN for this section */
1901 first_deferred_pfn = spfn;
1903 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1905 /* We should only stop along section boundaries */
1906 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
1909 /* If our quota has been met we can stop here */
1910 if (nr_pages >= nr_pages_needed)
1914 pgdat->first_deferred_pfn = spfn;
1915 pgdat_resize_unlock(pgdat, &flags);
1917 return nr_pages > 0;
1921 * deferred_grow_zone() is __init, but it is called from
1922 * get_page_from_freelist() during early boot until deferred_pages permanently
1923 * disables this call. This is why we have refdata wrapper to avoid warning,
1924 * and to ensure that the function body gets unloaded.
1927 _deferred_grow_zone(struct zone *zone, unsigned int order)
1929 return deferred_grow_zone(zone, order);
1932 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1934 void __init page_alloc_init_late(void)
1939 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1941 /* There will be num_node_state(N_MEMORY) threads */
1942 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1943 for_each_node_state(nid, N_MEMORY) {
1944 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1947 /* Block until all are initialised */
1948 wait_for_completion(&pgdat_init_all_done_comp);
1951 * The number of managed pages has changed due to the initialisation
1952 * so the pcpu batch and high limits needs to be updated or the limits
1953 * will be artificially small.
1955 for_each_populated_zone(zone)
1956 zone_pcp_update(zone);
1959 * We initialized the rest of the deferred pages. Permanently disable
1960 * on-demand struct page initialization.
1962 static_branch_disable(&deferred_pages);
1964 /* Reinit limits that are based on free pages after the kernel is up */
1965 files_maxfiles_init();
1968 /* Discard memblock private memory */
1971 for_each_node_state(nid, N_MEMORY)
1972 shuffle_free_memory(NODE_DATA(nid));
1974 for_each_populated_zone(zone)
1975 set_zone_contiguous(zone);
1977 #ifdef CONFIG_DEBUG_PAGEALLOC
1978 init_debug_guardpage();
1983 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1984 void __init init_cma_reserved_pageblock(struct page *page)
1986 unsigned i = pageblock_nr_pages;
1987 struct page *p = page;
1990 __ClearPageReserved(p);
1991 set_page_count(p, 0);
1994 set_pageblock_migratetype(page, MIGRATE_CMA);
1996 if (pageblock_order >= MAX_ORDER) {
1997 i = pageblock_nr_pages;
2000 set_page_refcounted(p);
2001 __free_pages(p, MAX_ORDER - 1);
2002 p += MAX_ORDER_NR_PAGES;
2003 } while (i -= MAX_ORDER_NR_PAGES);
2005 set_page_refcounted(page);
2006 __free_pages(page, pageblock_order);
2009 adjust_managed_page_count(page, pageblock_nr_pages);
2014 * The order of subdivision here is critical for the IO subsystem.
2015 * Please do not alter this order without good reasons and regression
2016 * testing. Specifically, as large blocks of memory are subdivided,
2017 * the order in which smaller blocks are delivered depends on the order
2018 * they're subdivided in this function. This is the primary factor
2019 * influencing the order in which pages are delivered to the IO
2020 * subsystem according to empirical testing, and this is also justified
2021 * by considering the behavior of a buddy system containing a single
2022 * large block of memory acted on by a series of small allocations.
2023 * This behavior is a critical factor in sglist merging's success.
2027 static inline void expand(struct zone *zone, struct page *page,
2028 int low, int high, struct free_area *area,
2031 unsigned long size = 1 << high;
2033 while (high > low) {
2037 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2040 * Mark as guard pages (or page), that will allow to
2041 * merge back to allocator when buddy will be freed.
2042 * Corresponding page table entries will not be touched,
2043 * pages will stay not present in virtual address space
2045 if (set_page_guard(zone, &page[size], high, migratetype))
2048 add_to_free_area(&page[size], area, migratetype);
2049 set_page_order(&page[size], high);
2053 static void check_new_page_bad(struct page *page)
2055 const char *bad_reason = NULL;
2056 unsigned long bad_flags = 0;
2058 if (unlikely(atomic_read(&page->_mapcount) != -1))
2059 bad_reason = "nonzero mapcount";
2060 if (unlikely(page->mapping != NULL))
2061 bad_reason = "non-NULL mapping";
2062 if (unlikely(page_ref_count(page) != 0))
2063 bad_reason = "nonzero _refcount";
2064 if (unlikely(page->flags & __PG_HWPOISON)) {
2065 bad_reason = "HWPoisoned (hardware-corrupted)";
2066 bad_flags = __PG_HWPOISON;
2067 /* Don't complain about hwpoisoned pages */
2068 page_mapcount_reset(page); /* remove PageBuddy */
2071 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
2072 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
2073 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
2076 if (unlikely(page->mem_cgroup))
2077 bad_reason = "page still charged to cgroup";
2079 bad_page(page, bad_reason, bad_flags);
2083 * This page is about to be returned from the page allocator
2085 static inline int check_new_page(struct page *page)
2087 if (likely(page_expected_state(page,
2088 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2091 check_new_page_bad(page);
2095 static inline bool free_pages_prezeroed(void)
2097 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2098 page_poisoning_enabled()) || want_init_on_free();
2101 #ifdef CONFIG_DEBUG_VM
2103 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2104 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2105 * also checked when pcp lists are refilled from the free lists.
2107 static inline bool check_pcp_refill(struct page *page)
2109 if (debug_pagealloc_enabled())
2110 return check_new_page(page);
2115 static inline bool check_new_pcp(struct page *page)
2117 return check_new_page(page);
2121 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2122 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2123 * enabled, they are also checked when being allocated from the pcp lists.
2125 static inline bool check_pcp_refill(struct page *page)
2127 return check_new_page(page);
2129 static inline bool check_new_pcp(struct page *page)
2131 if (debug_pagealloc_enabled())
2132 return check_new_page(page);
2136 #endif /* CONFIG_DEBUG_VM */
2138 static bool check_new_pages(struct page *page, unsigned int order)
2141 for (i = 0; i < (1 << order); i++) {
2142 struct page *p = page + i;
2144 if (unlikely(check_new_page(p)))
2151 inline void post_alloc_hook(struct page *page, unsigned int order,
2154 set_page_private(page, 0);
2155 set_page_refcounted(page);
2157 arch_alloc_page(page, order);
2158 if (debug_pagealloc_enabled())
2159 kernel_map_pages(page, 1 << order, 1);
2160 kasan_alloc_pages(page, order);
2161 kernel_poison_pages(page, 1 << order, 1);
2162 set_page_owner(page, order, gfp_flags);
2165 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2166 unsigned int alloc_flags)
2168 post_alloc_hook(page, order, gfp_flags);
2170 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2171 kernel_init_free_pages(page, 1 << order);
2173 if (order && (gfp_flags & __GFP_COMP))
2174 prep_compound_page(page, order);
2177 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2178 * allocate the page. The expectation is that the caller is taking
2179 * steps that will free more memory. The caller should avoid the page
2180 * being used for !PFMEMALLOC purposes.
2182 if (alloc_flags & ALLOC_NO_WATERMARKS)
2183 set_page_pfmemalloc(page);
2185 clear_page_pfmemalloc(page);
2189 * Go through the free lists for the given migratetype and remove
2190 * the smallest available page from the freelists
2192 static __always_inline
2193 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2196 unsigned int current_order;
2197 struct free_area *area;
2200 /* Find a page of the appropriate size in the preferred list */
2201 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2202 area = &(zone->free_area[current_order]);
2203 page = get_page_from_free_area(area, migratetype);
2206 del_page_from_free_area(page, area);
2207 expand(zone, page, order, current_order, area, migratetype);
2208 set_pcppage_migratetype(page, migratetype);
2217 * This array describes the order lists are fallen back to when
2218 * the free lists for the desirable migrate type are depleted
2220 static int fallbacks[MIGRATE_TYPES][4] = {
2221 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2222 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2223 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2225 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2227 #ifdef CONFIG_MEMORY_ISOLATION
2228 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2233 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2236 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2239 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2240 unsigned int order) { return NULL; }
2244 * Move the free pages in a range to the free lists of the requested type.
2245 * Note that start_page and end_pages are not aligned on a pageblock
2246 * boundary. If alignment is required, use move_freepages_block()
2248 static int move_freepages(struct zone *zone,
2249 struct page *start_page, struct page *end_page,
2250 int migratetype, int *num_movable)
2254 int pages_moved = 0;
2256 for (page = start_page; page <= end_page;) {
2257 if (!pfn_valid_within(page_to_pfn(page))) {
2262 if (!PageBuddy(page)) {
2264 * We assume that pages that could be isolated for
2265 * migration are movable. But we don't actually try
2266 * isolating, as that would be expensive.
2269 (PageLRU(page) || __PageMovable(page)))
2276 /* Make sure we are not inadvertently changing nodes */
2277 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2278 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2280 order = page_order(page);
2281 move_to_free_area(page, &zone->free_area[order], migratetype);
2283 pages_moved += 1 << order;
2289 int move_freepages_block(struct zone *zone, struct page *page,
2290 int migratetype, int *num_movable)
2292 unsigned long start_pfn, end_pfn;
2293 struct page *start_page, *end_page;
2298 start_pfn = page_to_pfn(page);
2299 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2300 start_page = pfn_to_page(start_pfn);
2301 end_page = start_page + pageblock_nr_pages - 1;
2302 end_pfn = start_pfn + pageblock_nr_pages - 1;
2304 /* Do not cross zone boundaries */
2305 if (!zone_spans_pfn(zone, start_pfn))
2307 if (!zone_spans_pfn(zone, end_pfn))
2310 return move_freepages(zone, start_page, end_page, migratetype,
2314 static void change_pageblock_range(struct page *pageblock_page,
2315 int start_order, int migratetype)
2317 int nr_pageblocks = 1 << (start_order - pageblock_order);
2319 while (nr_pageblocks--) {
2320 set_pageblock_migratetype(pageblock_page, migratetype);
2321 pageblock_page += pageblock_nr_pages;
2326 * When we are falling back to another migratetype during allocation, try to
2327 * steal extra free pages from the same pageblocks to satisfy further
2328 * allocations, instead of polluting multiple pageblocks.
2330 * If we are stealing a relatively large buddy page, it is likely there will
2331 * be more free pages in the pageblock, so try to steal them all. For
2332 * reclaimable and unmovable allocations, we steal regardless of page size,
2333 * as fragmentation caused by those allocations polluting movable pageblocks
2334 * is worse than movable allocations stealing from unmovable and reclaimable
2337 static bool can_steal_fallback(unsigned int order, int start_mt)
2340 * Leaving this order check is intended, although there is
2341 * relaxed order check in next check. The reason is that
2342 * we can actually steal whole pageblock if this condition met,
2343 * but, below check doesn't guarantee it and that is just heuristic
2344 * so could be changed anytime.
2346 if (order >= pageblock_order)
2349 if (order >= pageblock_order / 2 ||
2350 start_mt == MIGRATE_RECLAIMABLE ||
2351 start_mt == MIGRATE_UNMOVABLE ||
2352 page_group_by_mobility_disabled)
2358 static inline void boost_watermark(struct zone *zone)
2360 unsigned long max_boost;
2362 if (!watermark_boost_factor)
2365 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2366 watermark_boost_factor, 10000);
2369 * high watermark may be uninitialised if fragmentation occurs
2370 * very early in boot so do not boost. We do not fall
2371 * through and boost by pageblock_nr_pages as failing
2372 * allocations that early means that reclaim is not going
2373 * to help and it may even be impossible to reclaim the
2374 * boosted watermark resulting in a hang.
2379 max_boost = max(pageblock_nr_pages, max_boost);
2381 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2386 * This function implements actual steal behaviour. If order is large enough,
2387 * we can steal whole pageblock. If not, we first move freepages in this
2388 * pageblock to our migratetype and determine how many already-allocated pages
2389 * are there in the pageblock with a compatible migratetype. If at least half
2390 * of pages are free or compatible, we can change migratetype of the pageblock
2391 * itself, so pages freed in the future will be put on the correct free list.
2393 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2394 unsigned int alloc_flags, int start_type, bool whole_block)
2396 unsigned int current_order = page_order(page);
2397 struct free_area *area;
2398 int free_pages, movable_pages, alike_pages;
2401 old_block_type = get_pageblock_migratetype(page);
2404 * This can happen due to races and we want to prevent broken
2405 * highatomic accounting.
2407 if (is_migrate_highatomic(old_block_type))
2410 /* Take ownership for orders >= pageblock_order */
2411 if (current_order >= pageblock_order) {
2412 change_pageblock_range(page, current_order, start_type);
2417 * Boost watermarks to increase reclaim pressure to reduce the
2418 * likelihood of future fallbacks. Wake kswapd now as the node
2419 * may be balanced overall and kswapd will not wake naturally.
2421 boost_watermark(zone);
2422 if (alloc_flags & ALLOC_KSWAPD)
2423 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2425 /* We are not allowed to try stealing from the whole block */
2429 free_pages = move_freepages_block(zone, page, start_type,
2432 * Determine how many pages are compatible with our allocation.
2433 * For movable allocation, it's the number of movable pages which
2434 * we just obtained. For other types it's a bit more tricky.
2436 if (start_type == MIGRATE_MOVABLE) {
2437 alike_pages = movable_pages;
2440 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2441 * to MOVABLE pageblock, consider all non-movable pages as
2442 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2443 * vice versa, be conservative since we can't distinguish the
2444 * exact migratetype of non-movable pages.
2446 if (old_block_type == MIGRATE_MOVABLE)
2447 alike_pages = pageblock_nr_pages
2448 - (free_pages + movable_pages);
2453 /* moving whole block can fail due to zone boundary conditions */
2458 * If a sufficient number of pages in the block are either free or of
2459 * comparable migratability as our allocation, claim the whole block.
2461 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2462 page_group_by_mobility_disabled)
2463 set_pageblock_migratetype(page, start_type);
2468 area = &zone->free_area[current_order];
2469 move_to_free_area(page, area, start_type);
2473 * Check whether there is a suitable fallback freepage with requested order.
2474 * If only_stealable is true, this function returns fallback_mt only if
2475 * we can steal other freepages all together. This would help to reduce
2476 * fragmentation due to mixed migratetype pages in one pageblock.
2478 int find_suitable_fallback(struct free_area *area, unsigned int order,
2479 int migratetype, bool only_stealable, bool *can_steal)
2484 if (area->nr_free == 0)
2489 fallback_mt = fallbacks[migratetype][i];
2490 if (fallback_mt == MIGRATE_TYPES)
2493 if (free_area_empty(area, fallback_mt))
2496 if (can_steal_fallback(order, migratetype))
2499 if (!only_stealable)
2510 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2511 * there are no empty page blocks that contain a page with a suitable order
2513 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2514 unsigned int alloc_order)
2517 unsigned long max_managed, flags;
2520 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2521 * Check is race-prone but harmless.
2523 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2524 if (zone->nr_reserved_highatomic >= max_managed)
2527 spin_lock_irqsave(&zone->lock, flags);
2529 /* Recheck the nr_reserved_highatomic limit under the lock */
2530 if (zone->nr_reserved_highatomic >= max_managed)
2534 mt = get_pageblock_migratetype(page);
2535 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2536 && !is_migrate_cma(mt)) {
2537 zone->nr_reserved_highatomic += pageblock_nr_pages;
2538 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2539 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2543 spin_unlock_irqrestore(&zone->lock, flags);
2547 * Used when an allocation is about to fail under memory pressure. This
2548 * potentially hurts the reliability of high-order allocations when under
2549 * intense memory pressure but failed atomic allocations should be easier
2550 * to recover from than an OOM.
2552 * If @force is true, try to unreserve a pageblock even though highatomic
2553 * pageblock is exhausted.
2555 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2558 struct zonelist *zonelist = ac->zonelist;
2559 unsigned long flags;
2566 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2569 * Preserve at least one pageblock unless memory pressure
2572 if (!force && zone->nr_reserved_highatomic <=
2576 spin_lock_irqsave(&zone->lock, flags);
2577 for (order = 0; order < MAX_ORDER; order++) {
2578 struct free_area *area = &(zone->free_area[order]);
2580 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2585 * In page freeing path, migratetype change is racy so
2586 * we can counter several free pages in a pageblock
2587 * in this loop althoug we changed the pageblock type
2588 * from highatomic to ac->migratetype. So we should
2589 * adjust the count once.
2591 if (is_migrate_highatomic_page(page)) {
2593 * It should never happen but changes to
2594 * locking could inadvertently allow a per-cpu
2595 * drain to add pages to MIGRATE_HIGHATOMIC
2596 * while unreserving so be safe and watch for
2599 zone->nr_reserved_highatomic -= min(
2601 zone->nr_reserved_highatomic);
2605 * Convert to ac->migratetype and avoid the normal
2606 * pageblock stealing heuristics. Minimally, the caller
2607 * is doing the work and needs the pages. More
2608 * importantly, if the block was always converted to
2609 * MIGRATE_UNMOVABLE or another type then the number
2610 * of pageblocks that cannot be completely freed
2613 set_pageblock_migratetype(page, ac->migratetype);
2614 ret = move_freepages_block(zone, page, ac->migratetype,
2617 spin_unlock_irqrestore(&zone->lock, flags);
2621 spin_unlock_irqrestore(&zone->lock, flags);
2628 * Try finding a free buddy page on the fallback list and put it on the free
2629 * list of requested migratetype, possibly along with other pages from the same
2630 * block, depending on fragmentation avoidance heuristics. Returns true if
2631 * fallback was found so that __rmqueue_smallest() can grab it.
2633 * The use of signed ints for order and current_order is a deliberate
2634 * deviation from the rest of this file, to make the for loop
2635 * condition simpler.
2637 static __always_inline bool
2638 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2639 unsigned int alloc_flags)
2641 struct free_area *area;
2643 int min_order = order;
2649 * Do not steal pages from freelists belonging to other pageblocks
2650 * i.e. orders < pageblock_order. If there are no local zones free,
2651 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2653 if (alloc_flags & ALLOC_NOFRAGMENT)
2654 min_order = pageblock_order;
2657 * Find the largest available free page in the other list. This roughly
2658 * approximates finding the pageblock with the most free pages, which
2659 * would be too costly to do exactly.
2661 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2663 area = &(zone->free_area[current_order]);
2664 fallback_mt = find_suitable_fallback(area, current_order,
2665 start_migratetype, false, &can_steal);
2666 if (fallback_mt == -1)
2670 * We cannot steal all free pages from the pageblock and the
2671 * requested migratetype is movable. In that case it's better to
2672 * steal and split the smallest available page instead of the
2673 * largest available page, because even if the next movable
2674 * allocation falls back into a different pageblock than this
2675 * one, it won't cause permanent fragmentation.
2677 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2678 && current_order > order)
2687 for (current_order = order; current_order < MAX_ORDER;
2689 area = &(zone->free_area[current_order]);
2690 fallback_mt = find_suitable_fallback(area, current_order,
2691 start_migratetype, false, &can_steal);
2692 if (fallback_mt != -1)
2697 * This should not happen - we already found a suitable fallback
2698 * when looking for the largest page.
2700 VM_BUG_ON(current_order == MAX_ORDER);
2703 page = get_page_from_free_area(area, fallback_mt);
2705 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2708 trace_mm_page_alloc_extfrag(page, order, current_order,
2709 start_migratetype, fallback_mt);
2716 * Do the hard work of removing an element from the buddy allocator.
2717 * Call me with the zone->lock already held.
2719 static __always_inline struct page *
2720 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2721 unsigned int alloc_flags)
2726 page = __rmqueue_smallest(zone, order, migratetype);
2727 if (unlikely(!page)) {
2728 if (migratetype == MIGRATE_MOVABLE)
2729 page = __rmqueue_cma_fallback(zone, order);
2731 if (!page && __rmqueue_fallback(zone, order, migratetype,
2736 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2741 * Obtain a specified number of elements from the buddy allocator, all under
2742 * a single hold of the lock, for efficiency. Add them to the supplied list.
2743 * Returns the number of new pages which were placed at *list.
2745 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2746 unsigned long count, struct list_head *list,
2747 int migratetype, unsigned int alloc_flags)
2751 spin_lock(&zone->lock);
2752 for (i = 0; i < count; ++i) {
2753 struct page *page = __rmqueue(zone, order, migratetype,
2755 if (unlikely(page == NULL))
2758 if (unlikely(check_pcp_refill(page)))
2762 * Split buddy pages returned by expand() are received here in
2763 * physical page order. The page is added to the tail of
2764 * caller's list. From the callers perspective, the linked list
2765 * is ordered by page number under some conditions. This is
2766 * useful for IO devices that can forward direction from the
2767 * head, thus also in the physical page order. This is useful
2768 * for IO devices that can merge IO requests if the physical
2769 * pages are ordered properly.
2771 list_add_tail(&page->lru, list);
2773 if (is_migrate_cma(get_pcppage_migratetype(page)))
2774 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2779 * i pages were removed from the buddy list even if some leak due
2780 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2781 * on i. Do not confuse with 'alloced' which is the number of
2782 * pages added to the pcp list.
2784 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2785 spin_unlock(&zone->lock);
2791 * Called from the vmstat counter updater to drain pagesets of this
2792 * currently executing processor on remote nodes after they have
2795 * Note that this function must be called with the thread pinned to
2796 * a single processor.
2798 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2800 unsigned long flags;
2801 int to_drain, batch;
2803 local_irq_save(flags);
2804 batch = READ_ONCE(pcp->batch);
2805 to_drain = min(pcp->count, batch);
2807 free_pcppages_bulk(zone, to_drain, pcp);
2808 local_irq_restore(flags);
2813 * Drain pcplists of the indicated processor and zone.
2815 * The processor must either be the current processor and the
2816 * thread pinned to the current processor or a processor that
2819 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2821 unsigned long flags;
2822 struct per_cpu_pageset *pset;
2823 struct per_cpu_pages *pcp;
2825 local_irq_save(flags);
2826 pset = per_cpu_ptr(zone->pageset, cpu);
2830 free_pcppages_bulk(zone, pcp->count, pcp);
2831 local_irq_restore(flags);
2835 * Drain pcplists of all zones on the indicated processor.
2837 * The processor must either be the current processor and the
2838 * thread pinned to the current processor or a processor that
2841 static void drain_pages(unsigned int cpu)
2845 for_each_populated_zone(zone) {
2846 drain_pages_zone(cpu, zone);
2851 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2853 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2854 * the single zone's pages.
2856 void drain_local_pages(struct zone *zone)
2858 int cpu = smp_processor_id();
2861 drain_pages_zone(cpu, zone);
2866 static void drain_local_pages_wq(struct work_struct *work)
2868 struct pcpu_drain *drain;
2870 drain = container_of(work, struct pcpu_drain, work);
2873 * drain_all_pages doesn't use proper cpu hotplug protection so
2874 * we can race with cpu offline when the WQ can move this from
2875 * a cpu pinned worker to an unbound one. We can operate on a different
2876 * cpu which is allright but we also have to make sure to not move to
2880 drain_local_pages(drain->zone);
2885 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2887 * When zone parameter is non-NULL, spill just the single zone's pages.
2889 * Note that this can be extremely slow as the draining happens in a workqueue.
2891 void drain_all_pages(struct zone *zone)
2896 * Allocate in the BSS so we wont require allocation in
2897 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2899 static cpumask_t cpus_with_pcps;
2902 * Make sure nobody triggers this path before mm_percpu_wq is fully
2905 if (WARN_ON_ONCE(!mm_percpu_wq))
2909 * Do not drain if one is already in progress unless it's specific to
2910 * a zone. Such callers are primarily CMA and memory hotplug and need
2911 * the drain to be complete when the call returns.
2913 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2916 mutex_lock(&pcpu_drain_mutex);
2920 * We don't care about racing with CPU hotplug event
2921 * as offline notification will cause the notified
2922 * cpu to drain that CPU pcps and on_each_cpu_mask
2923 * disables preemption as part of its processing
2925 for_each_online_cpu(cpu) {
2926 struct per_cpu_pageset *pcp;
2928 bool has_pcps = false;
2931 pcp = per_cpu_ptr(zone->pageset, cpu);
2935 for_each_populated_zone(z) {
2936 pcp = per_cpu_ptr(z->pageset, cpu);
2937 if (pcp->pcp.count) {
2945 cpumask_set_cpu(cpu, &cpus_with_pcps);
2947 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2950 for_each_cpu(cpu, &cpus_with_pcps) {
2951 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
2954 INIT_WORK(&drain->work, drain_local_pages_wq);
2955 queue_work_on(cpu, mm_percpu_wq, &drain->work);
2957 for_each_cpu(cpu, &cpus_with_pcps)
2958 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
2960 mutex_unlock(&pcpu_drain_mutex);
2963 #ifdef CONFIG_HIBERNATION
2966 * Touch the watchdog for every WD_PAGE_COUNT pages.
2968 #define WD_PAGE_COUNT (128*1024)
2970 void mark_free_pages(struct zone *zone)
2972 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2973 unsigned long flags;
2974 unsigned int order, t;
2977 if (zone_is_empty(zone))
2980 spin_lock_irqsave(&zone->lock, flags);
2982 max_zone_pfn = zone_end_pfn(zone);
2983 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2984 if (pfn_valid(pfn)) {
2985 page = pfn_to_page(pfn);
2987 if (!--page_count) {
2988 touch_nmi_watchdog();
2989 page_count = WD_PAGE_COUNT;
2992 if (page_zone(page) != zone)
2995 if (!swsusp_page_is_forbidden(page))
2996 swsusp_unset_page_free(page);
2999 for_each_migratetype_order(order, t) {
3000 list_for_each_entry(page,
3001 &zone->free_area[order].free_list[t], lru) {
3004 pfn = page_to_pfn(page);
3005 for (i = 0; i < (1UL << order); i++) {
3006 if (!--page_count) {
3007 touch_nmi_watchdog();
3008 page_count = WD_PAGE_COUNT;
3010 swsusp_set_page_free(pfn_to_page(pfn + i));
3014 spin_unlock_irqrestore(&zone->lock, flags);
3016 #endif /* CONFIG_PM */
3018 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3022 if (!free_pcp_prepare(page))
3025 migratetype = get_pfnblock_migratetype(page, pfn);
3026 set_pcppage_migratetype(page, migratetype);
3030 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3032 struct zone *zone = page_zone(page);
3033 struct per_cpu_pages *pcp;
3036 migratetype = get_pcppage_migratetype(page);
3037 __count_vm_event(PGFREE);
3040 * We only track unmovable, reclaimable and movable on pcp lists.
3041 * Free ISOLATE pages back to the allocator because they are being
3042 * offlined but treat HIGHATOMIC as movable pages so we can get those
3043 * areas back if necessary. Otherwise, we may have to free
3044 * excessively into the page allocator
3046 if (migratetype >= MIGRATE_PCPTYPES) {
3047 if (unlikely(is_migrate_isolate(migratetype))) {
3048 free_one_page(zone, page, pfn, 0, migratetype);
3051 migratetype = MIGRATE_MOVABLE;
3054 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3055 list_add(&page->lru, &pcp->lists[migratetype]);
3057 if (pcp->count >= pcp->high) {
3058 unsigned long batch = READ_ONCE(pcp->batch);
3059 free_pcppages_bulk(zone, batch, pcp);
3064 * Free a 0-order page
3066 void free_unref_page(struct page *page)
3068 unsigned long flags;
3069 unsigned long pfn = page_to_pfn(page);
3071 if (!free_unref_page_prepare(page, pfn))
3074 local_irq_save(flags);
3075 free_unref_page_commit(page, pfn);
3076 local_irq_restore(flags);
3080 * Free a list of 0-order pages
3082 void free_unref_page_list(struct list_head *list)
3084 struct page *page, *next;
3085 unsigned long flags, pfn;
3086 int batch_count = 0;
3088 /* Prepare pages for freeing */
3089 list_for_each_entry_safe(page, next, list, lru) {
3090 pfn = page_to_pfn(page);
3091 if (!free_unref_page_prepare(page, pfn))
3092 list_del(&page->lru);
3093 set_page_private(page, pfn);
3096 local_irq_save(flags);
3097 list_for_each_entry_safe(page, next, list, lru) {
3098 unsigned long pfn = page_private(page);
3100 set_page_private(page, 0);
3101 trace_mm_page_free_batched(page);
3102 free_unref_page_commit(page, pfn);
3105 * Guard against excessive IRQ disabled times when we get
3106 * a large list of pages to free.
3108 if (++batch_count == SWAP_CLUSTER_MAX) {
3109 local_irq_restore(flags);
3111 local_irq_save(flags);
3114 local_irq_restore(flags);
3118 * split_page takes a non-compound higher-order page, and splits it into
3119 * n (1<<order) sub-pages: page[0..n]
3120 * Each sub-page must be freed individually.
3122 * Note: this is probably too low level an operation for use in drivers.
3123 * Please consult with lkml before using this in your driver.
3125 void split_page(struct page *page, unsigned int order)
3129 VM_BUG_ON_PAGE(PageCompound(page), page);
3130 VM_BUG_ON_PAGE(!page_count(page), page);
3132 for (i = 1; i < (1 << order); i++)
3133 set_page_refcounted(page + i);
3134 split_page_owner(page, order);
3136 EXPORT_SYMBOL_GPL(split_page);
3138 int __isolate_free_page(struct page *page, unsigned int order)
3140 struct free_area *area = &page_zone(page)->free_area[order];
3141 unsigned long watermark;
3145 BUG_ON(!PageBuddy(page));
3147 zone = page_zone(page);
3148 mt = get_pageblock_migratetype(page);
3150 if (!is_migrate_isolate(mt)) {
3152 * Obey watermarks as if the page was being allocated. We can
3153 * emulate a high-order watermark check with a raised order-0
3154 * watermark, because we already know our high-order page
3157 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3158 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3161 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3164 /* Remove page from free list */
3166 del_page_from_free_area(page, area);
3169 * Set the pageblock if the isolated page is at least half of a
3172 if (order >= pageblock_order - 1) {
3173 struct page *endpage = page + (1 << order) - 1;
3174 for (; page < endpage; page += pageblock_nr_pages) {
3175 int mt = get_pageblock_migratetype(page);
3176 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3177 && !is_migrate_highatomic(mt))
3178 set_pageblock_migratetype(page,
3184 return 1UL << order;
3188 * Update NUMA hit/miss statistics
3190 * Must be called with interrupts disabled.
3192 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3195 enum numa_stat_item local_stat = NUMA_LOCAL;
3197 /* skip numa counters update if numa stats is disabled */
3198 if (!static_branch_likely(&vm_numa_stat_key))
3201 if (zone_to_nid(z) != numa_node_id())
3202 local_stat = NUMA_OTHER;
3204 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3205 __inc_numa_state(z, NUMA_HIT);
3207 __inc_numa_state(z, NUMA_MISS);
3208 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3210 __inc_numa_state(z, local_stat);
3214 /* Remove page from the per-cpu list, caller must protect the list */
3215 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3216 unsigned int alloc_flags,
3217 struct per_cpu_pages *pcp,
3218 struct list_head *list)
3223 if (list_empty(list)) {
3224 pcp->count += rmqueue_bulk(zone, 0,
3226 migratetype, alloc_flags);
3227 if (unlikely(list_empty(list)))
3231 page = list_first_entry(list, struct page, lru);
3232 list_del(&page->lru);
3234 } while (check_new_pcp(page));
3239 /* Lock and remove page from the per-cpu list */
3240 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3241 struct zone *zone, gfp_t gfp_flags,
3242 int migratetype, unsigned int alloc_flags)
3244 struct per_cpu_pages *pcp;
3245 struct list_head *list;
3247 unsigned long flags;
3249 local_irq_save(flags);
3250 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3251 list = &pcp->lists[migratetype];
3252 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3254 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3255 zone_statistics(preferred_zone, zone);
3257 local_irq_restore(flags);
3262 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3265 struct page *rmqueue(struct zone *preferred_zone,
3266 struct zone *zone, unsigned int order,
3267 gfp_t gfp_flags, unsigned int alloc_flags,
3270 unsigned long flags;
3273 if (likely(order == 0)) {
3274 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3275 migratetype, alloc_flags);
3280 * We most definitely don't want callers attempting to
3281 * allocate greater than order-1 page units with __GFP_NOFAIL.
3283 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3284 spin_lock_irqsave(&zone->lock, flags);
3288 if (alloc_flags & ALLOC_HARDER) {
3289 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3291 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3294 page = __rmqueue(zone, order, migratetype, alloc_flags);
3295 } while (page && check_new_pages(page, order));
3296 spin_unlock(&zone->lock);
3299 __mod_zone_freepage_state(zone, -(1 << order),
3300 get_pcppage_migratetype(page));
3302 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3303 zone_statistics(preferred_zone, zone);
3304 local_irq_restore(flags);
3307 /* Separate test+clear to avoid unnecessary atomics */
3308 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3309 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3310 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3313 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3317 local_irq_restore(flags);
3321 #ifdef CONFIG_FAIL_PAGE_ALLOC
3324 struct fault_attr attr;
3326 bool ignore_gfp_highmem;
3327 bool ignore_gfp_reclaim;
3329 } fail_page_alloc = {
3330 .attr = FAULT_ATTR_INITIALIZER,
3331 .ignore_gfp_reclaim = true,
3332 .ignore_gfp_highmem = true,
3336 static int __init setup_fail_page_alloc(char *str)
3338 return setup_fault_attr(&fail_page_alloc.attr, str);
3340 __setup("fail_page_alloc=", setup_fail_page_alloc);
3342 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3344 if (order < fail_page_alloc.min_order)
3346 if (gfp_mask & __GFP_NOFAIL)
3348 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3350 if (fail_page_alloc.ignore_gfp_reclaim &&
3351 (gfp_mask & __GFP_DIRECT_RECLAIM))
3354 return should_fail(&fail_page_alloc.attr, 1 << order);
3357 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3359 static int __init fail_page_alloc_debugfs(void)
3361 umode_t mode = S_IFREG | 0600;
3364 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3365 &fail_page_alloc.attr);
3367 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3368 &fail_page_alloc.ignore_gfp_reclaim);
3369 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3370 &fail_page_alloc.ignore_gfp_highmem);
3371 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3376 late_initcall(fail_page_alloc_debugfs);
3378 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3380 #else /* CONFIG_FAIL_PAGE_ALLOC */
3382 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3387 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3389 static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3391 return __should_fail_alloc_page(gfp_mask, order);
3393 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3396 * Return true if free base pages are above 'mark'. For high-order checks it
3397 * will return true of the order-0 watermark is reached and there is at least
3398 * one free page of a suitable size. Checking now avoids taking the zone lock
3399 * to check in the allocation paths if no pages are free.
3401 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3402 int classzone_idx, unsigned int alloc_flags,
3407 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3409 /* free_pages may go negative - that's OK */
3410 free_pages -= (1 << order) - 1;
3412 if (alloc_flags & ALLOC_HIGH)
3416 * If the caller does not have rights to ALLOC_HARDER then subtract
3417 * the high-atomic reserves. This will over-estimate the size of the
3418 * atomic reserve but it avoids a search.
3420 if (likely(!alloc_harder)) {
3421 free_pages -= z->nr_reserved_highatomic;
3424 * OOM victims can try even harder than normal ALLOC_HARDER
3425 * users on the grounds that it's definitely going to be in
3426 * the exit path shortly and free memory. Any allocation it
3427 * makes during the free path will be small and short-lived.
3429 if (alloc_flags & ALLOC_OOM)
3437 /* If allocation can't use CMA areas don't use free CMA pages */
3438 if (!(alloc_flags & ALLOC_CMA))
3439 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3443 * Check watermarks for an order-0 allocation request. If these
3444 * are not met, then a high-order request also cannot go ahead
3445 * even if a suitable page happened to be free.
3447 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3450 /* If this is an order-0 request then the watermark is fine */
3454 /* For a high-order request, check at least one suitable page is free */
3455 for (o = order; o < MAX_ORDER; o++) {
3456 struct free_area *area = &z->free_area[o];
3462 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3463 if (!free_area_empty(area, mt))
3468 if ((alloc_flags & ALLOC_CMA) &&
3469 !free_area_empty(area, MIGRATE_CMA)) {
3474 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3480 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3481 int classzone_idx, unsigned int alloc_flags)
3483 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3484 zone_page_state(z, NR_FREE_PAGES));
3487 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3488 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3490 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3494 /* If allocation can't use CMA areas don't use free CMA pages */
3495 if (!(alloc_flags & ALLOC_CMA))
3496 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3500 * Fast check for order-0 only. If this fails then the reserves
3501 * need to be calculated. There is a corner case where the check
3502 * passes but only the high-order atomic reserve are free. If
3503 * the caller is !atomic then it'll uselessly search the free
3504 * list. That corner case is then slower but it is harmless.
3506 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3509 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3513 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3514 unsigned long mark, int classzone_idx)
3516 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3518 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3519 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3521 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3526 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3528 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3529 node_reclaim_distance;
3531 #else /* CONFIG_NUMA */
3532 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3536 #endif /* CONFIG_NUMA */
3539 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3540 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3541 * premature use of a lower zone may cause lowmem pressure problems that
3542 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3543 * probably too small. It only makes sense to spread allocations to avoid
3544 * fragmentation between the Normal and DMA32 zones.
3546 static inline unsigned int
3547 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3549 unsigned int alloc_flags = 0;
3551 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3552 alloc_flags |= ALLOC_KSWAPD;
3554 #ifdef CONFIG_ZONE_DMA32
3558 if (zone_idx(zone) != ZONE_NORMAL)
3562 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3563 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3564 * on UMA that if Normal is populated then so is DMA32.
3566 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3567 if (nr_online_nodes > 1 && !populated_zone(--zone))
3570 alloc_flags |= ALLOC_NOFRAGMENT;
3571 #endif /* CONFIG_ZONE_DMA32 */
3576 * get_page_from_freelist goes through the zonelist trying to allocate
3579 static struct page *
3580 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3581 const struct alloc_context *ac)
3585 struct pglist_data *last_pgdat_dirty_limit = NULL;
3590 * Scan zonelist, looking for a zone with enough free.
3591 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3593 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3594 z = ac->preferred_zoneref;
3595 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3600 if (cpusets_enabled() &&
3601 (alloc_flags & ALLOC_CPUSET) &&
3602 !__cpuset_zone_allowed(zone, gfp_mask))
3605 * When allocating a page cache page for writing, we
3606 * want to get it from a node that is within its dirty
3607 * limit, such that no single node holds more than its
3608 * proportional share of globally allowed dirty pages.
3609 * The dirty limits take into account the node's
3610 * lowmem reserves and high watermark so that kswapd
3611 * should be able to balance it without having to
3612 * write pages from its LRU list.
3614 * XXX: For now, allow allocations to potentially
3615 * exceed the per-node dirty limit in the slowpath
3616 * (spread_dirty_pages unset) before going into reclaim,
3617 * which is important when on a NUMA setup the allowed
3618 * nodes are together not big enough to reach the
3619 * global limit. The proper fix for these situations
3620 * will require awareness of nodes in the
3621 * dirty-throttling and the flusher threads.
3623 if (ac->spread_dirty_pages) {
3624 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3627 if (!node_dirty_ok(zone->zone_pgdat)) {
3628 last_pgdat_dirty_limit = zone->zone_pgdat;
3633 if (no_fallback && nr_online_nodes > 1 &&
3634 zone != ac->preferred_zoneref->zone) {
3638 * If moving to a remote node, retry but allow
3639 * fragmenting fallbacks. Locality is more important
3640 * than fragmentation avoidance.
3642 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3643 if (zone_to_nid(zone) != local_nid) {
3644 alloc_flags &= ~ALLOC_NOFRAGMENT;
3649 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3650 if (!zone_watermark_fast(zone, order, mark,
3651 ac_classzone_idx(ac), alloc_flags)) {
3654 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3656 * Watermark failed for this zone, but see if we can
3657 * grow this zone if it contains deferred pages.
3659 if (static_branch_unlikely(&deferred_pages)) {
3660 if (_deferred_grow_zone(zone, order))
3664 /* Checked here to keep the fast path fast */
3665 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3666 if (alloc_flags & ALLOC_NO_WATERMARKS)
3669 if (node_reclaim_mode == 0 ||
3670 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3673 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3675 case NODE_RECLAIM_NOSCAN:
3678 case NODE_RECLAIM_FULL:
3679 /* scanned but unreclaimable */
3682 /* did we reclaim enough */
3683 if (zone_watermark_ok(zone, order, mark,
3684 ac_classzone_idx(ac), alloc_flags))
3692 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3693 gfp_mask, alloc_flags, ac->migratetype);
3695 prep_new_page(page, order, gfp_mask, alloc_flags);
3698 * If this is a high-order atomic allocation then check
3699 * if the pageblock should be reserved for the future
3701 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3702 reserve_highatomic_pageblock(page, zone, order);
3706 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3707 /* Try again if zone has deferred pages */
3708 if (static_branch_unlikely(&deferred_pages)) {
3709 if (_deferred_grow_zone(zone, order))
3717 * It's possible on a UMA machine to get through all zones that are
3718 * fragmented. If avoiding fragmentation, reset and try again.
3721 alloc_flags &= ~ALLOC_NOFRAGMENT;
3728 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3730 unsigned int filter = SHOW_MEM_FILTER_NODES;
3733 * This documents exceptions given to allocations in certain
3734 * contexts that are allowed to allocate outside current's set
3737 if (!(gfp_mask & __GFP_NOMEMALLOC))
3738 if (tsk_is_oom_victim(current) ||
3739 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3740 filter &= ~SHOW_MEM_FILTER_NODES;
3741 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3742 filter &= ~SHOW_MEM_FILTER_NODES;
3744 show_mem(filter, nodemask);
3747 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3749 struct va_format vaf;
3751 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3753 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3756 va_start(args, fmt);
3759 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3760 current->comm, &vaf, gfp_mask, &gfp_mask,
3761 nodemask_pr_args(nodemask));
3764 cpuset_print_current_mems_allowed();
3767 warn_alloc_show_mem(gfp_mask, nodemask);
3770 static inline struct page *
3771 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3772 unsigned int alloc_flags,
3773 const struct alloc_context *ac)
3777 page = get_page_from_freelist(gfp_mask, order,
3778 alloc_flags|ALLOC_CPUSET, ac);
3780 * fallback to ignore cpuset restriction if our nodes
3784 page = get_page_from_freelist(gfp_mask, order,
3790 static inline struct page *
3791 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3792 const struct alloc_context *ac, unsigned long *did_some_progress)
3794 struct oom_control oc = {
3795 .zonelist = ac->zonelist,
3796 .nodemask = ac->nodemask,
3798 .gfp_mask = gfp_mask,
3803 *did_some_progress = 0;
3806 * Acquire the oom lock. If that fails, somebody else is
3807 * making progress for us.
3809 if (!mutex_trylock(&oom_lock)) {
3810 *did_some_progress = 1;
3811 schedule_timeout_uninterruptible(1);
3816 * Go through the zonelist yet one more time, keep very high watermark
3817 * here, this is only to catch a parallel oom killing, we must fail if
3818 * we're still under heavy pressure. But make sure that this reclaim
3819 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3820 * allocation which will never fail due to oom_lock already held.
3822 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3823 ~__GFP_DIRECT_RECLAIM, order,
3824 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3828 /* Coredumps can quickly deplete all memory reserves */
3829 if (current->flags & PF_DUMPCORE)
3831 /* The OOM killer will not help higher order allocs */
3832 if (order > PAGE_ALLOC_COSTLY_ORDER)
3835 * We have already exhausted all our reclaim opportunities without any
3836 * success so it is time to admit defeat. We will skip the OOM killer
3837 * because it is very likely that the caller has a more reasonable
3838 * fallback than shooting a random task.
3840 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3842 /* The OOM killer does not needlessly kill tasks for lowmem */
3843 if (ac->high_zoneidx < ZONE_NORMAL)
3845 if (pm_suspended_storage())
3848 * XXX: GFP_NOFS allocations should rather fail than rely on
3849 * other request to make a forward progress.
3850 * We are in an unfortunate situation where out_of_memory cannot
3851 * do much for this context but let's try it to at least get
3852 * access to memory reserved if the current task is killed (see
3853 * out_of_memory). Once filesystems are ready to handle allocation
3854 * failures more gracefully we should just bail out here.
3857 /* The OOM killer may not free memory on a specific node */
3858 if (gfp_mask & __GFP_THISNODE)
3861 /* Exhausted what can be done so it's blame time */
3862 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3863 *did_some_progress = 1;
3866 * Help non-failing allocations by giving them access to memory
3869 if (gfp_mask & __GFP_NOFAIL)
3870 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3871 ALLOC_NO_WATERMARKS, ac);
3874 mutex_unlock(&oom_lock);
3879 * Maximum number of compaction retries wit a progress before OOM
3880 * killer is consider as the only way to move forward.
3882 #define MAX_COMPACT_RETRIES 16
3884 #ifdef CONFIG_COMPACTION
3885 /* Try memory compaction for high-order allocations before reclaim */
3886 static struct page *
3887 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3888 unsigned int alloc_flags, const struct alloc_context *ac,
3889 enum compact_priority prio, enum compact_result *compact_result)
3891 struct page *page = NULL;
3892 unsigned long pflags;
3893 unsigned int noreclaim_flag;
3898 psi_memstall_enter(&pflags);
3899 noreclaim_flag = memalloc_noreclaim_save();
3901 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3904 memalloc_noreclaim_restore(noreclaim_flag);
3905 psi_memstall_leave(&pflags);
3908 * At least in one zone compaction wasn't deferred or skipped, so let's
3909 * count a compaction stall
3911 count_vm_event(COMPACTSTALL);
3913 /* Prep a captured page if available */
3915 prep_new_page(page, order, gfp_mask, alloc_flags);
3917 /* Try get a page from the freelist if available */
3919 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3922 struct zone *zone = page_zone(page);
3924 zone->compact_blockskip_flush = false;
3925 compaction_defer_reset(zone, order, true);
3926 count_vm_event(COMPACTSUCCESS);
3931 * It's bad if compaction run occurs and fails. The most likely reason
3932 * is that pages exist, but not enough to satisfy watermarks.
3934 count_vm_event(COMPACTFAIL);
3942 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3943 enum compact_result compact_result,
3944 enum compact_priority *compact_priority,
3945 int *compaction_retries)
3947 int max_retries = MAX_COMPACT_RETRIES;
3950 int retries = *compaction_retries;
3951 enum compact_priority priority = *compact_priority;
3956 if (compaction_made_progress(compact_result))
3957 (*compaction_retries)++;
3960 * compaction considers all the zone as desperately out of memory
3961 * so it doesn't really make much sense to retry except when the
3962 * failure could be caused by insufficient priority
3964 if (compaction_failed(compact_result))
3965 goto check_priority;
3968 * compaction was skipped because there are not enough order-0 pages
3969 * to work with, so we retry only if it looks like reclaim can help.
3971 if (compaction_needs_reclaim(compact_result)) {
3972 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3977 * make sure the compaction wasn't deferred or didn't bail out early
3978 * due to locks contention before we declare that we should give up.
3979 * But the next retry should use a higher priority if allowed, so
3980 * we don't just keep bailing out endlessly.
3982 if (compaction_withdrawn(compact_result)) {
3983 goto check_priority;
3987 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3988 * costly ones because they are de facto nofail and invoke OOM
3989 * killer to move on while costly can fail and users are ready
3990 * to cope with that. 1/4 retries is rather arbitrary but we
3991 * would need much more detailed feedback from compaction to
3992 * make a better decision.
3994 if (order > PAGE_ALLOC_COSTLY_ORDER)
3996 if (*compaction_retries <= max_retries) {
4002 * Make sure there are attempts at the highest priority if we exhausted
4003 * all retries or failed at the lower priorities.
4006 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4007 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4009 if (*compact_priority > min_priority) {
4010 (*compact_priority)--;
4011 *compaction_retries = 0;
4015 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4019 static inline struct page *
4020 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4021 unsigned int alloc_flags, const struct alloc_context *ac,
4022 enum compact_priority prio, enum compact_result *compact_result)
4024 *compact_result = COMPACT_SKIPPED;
4029 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4030 enum compact_result compact_result,
4031 enum compact_priority *compact_priority,
4032 int *compaction_retries)
4037 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4041 * There are setups with compaction disabled which would prefer to loop
4042 * inside the allocator rather than hit the oom killer prematurely.
4043 * Let's give them a good hope and keep retrying while the order-0
4044 * watermarks are OK.
4046 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4048 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4049 ac_classzone_idx(ac), alloc_flags))
4054 #endif /* CONFIG_COMPACTION */
4056 #ifdef CONFIG_LOCKDEP
4057 static struct lockdep_map __fs_reclaim_map =
4058 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4060 static bool __need_fs_reclaim(gfp_t gfp_mask)
4062 gfp_mask = current_gfp_context(gfp_mask);
4064 /* no reclaim without waiting on it */
4065 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4068 /* this guy won't enter reclaim */
4069 if (current->flags & PF_MEMALLOC)
4072 /* We're only interested __GFP_FS allocations for now */
4073 if (!(gfp_mask & __GFP_FS))
4076 if (gfp_mask & __GFP_NOLOCKDEP)
4082 void __fs_reclaim_acquire(void)
4084 lock_map_acquire(&__fs_reclaim_map);
4087 void __fs_reclaim_release(void)
4089 lock_map_release(&__fs_reclaim_map);
4092 void fs_reclaim_acquire(gfp_t gfp_mask)
4094 if (__need_fs_reclaim(gfp_mask))
4095 __fs_reclaim_acquire();
4097 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4099 void fs_reclaim_release(gfp_t gfp_mask)
4101 if (__need_fs_reclaim(gfp_mask))
4102 __fs_reclaim_release();
4104 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4107 /* Perform direct synchronous page reclaim */
4109 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4110 const struct alloc_context *ac)
4113 unsigned int noreclaim_flag;
4114 unsigned long pflags;
4118 /* We now go into synchronous reclaim */
4119 cpuset_memory_pressure_bump();
4120 psi_memstall_enter(&pflags);
4121 fs_reclaim_acquire(gfp_mask);
4122 noreclaim_flag = memalloc_noreclaim_save();
4124 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4127 memalloc_noreclaim_restore(noreclaim_flag);
4128 fs_reclaim_release(gfp_mask);
4129 psi_memstall_leave(&pflags);
4136 /* The really slow allocator path where we enter direct reclaim */
4137 static inline struct page *
4138 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4139 unsigned int alloc_flags, const struct alloc_context *ac,
4140 unsigned long *did_some_progress)
4142 struct page *page = NULL;
4143 bool drained = false;
4145 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4146 if (unlikely(!(*did_some_progress)))
4150 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4153 * If an allocation failed after direct reclaim, it could be because
4154 * pages are pinned on the per-cpu lists or in high alloc reserves.
4155 * Shrink them them and try again
4157 if (!page && !drained) {
4158 unreserve_highatomic_pageblock(ac, false);
4159 drain_all_pages(NULL);
4167 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4168 const struct alloc_context *ac)
4172 pg_data_t *last_pgdat = NULL;
4173 enum zone_type high_zoneidx = ac->high_zoneidx;
4175 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
4177 if (last_pgdat != zone->zone_pgdat)
4178 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
4179 last_pgdat = zone->zone_pgdat;
4183 static inline unsigned int
4184 gfp_to_alloc_flags(gfp_t gfp_mask)
4186 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4188 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
4189 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4192 * The caller may dip into page reserves a bit more if the caller
4193 * cannot run direct reclaim, or if the caller has realtime scheduling
4194 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4195 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4197 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
4199 if (gfp_mask & __GFP_ATOMIC) {
4201 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4202 * if it can't schedule.
4204 if (!(gfp_mask & __GFP_NOMEMALLOC))
4205 alloc_flags |= ALLOC_HARDER;
4207 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4208 * comment for __cpuset_node_allowed().
4210 alloc_flags &= ~ALLOC_CPUSET;
4211 } else if (unlikely(rt_task(current)) && !in_interrupt())
4212 alloc_flags |= ALLOC_HARDER;
4214 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4215 alloc_flags |= ALLOC_KSWAPD;
4218 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4219 alloc_flags |= ALLOC_CMA;
4224 static bool oom_reserves_allowed(struct task_struct *tsk)
4226 if (!tsk_is_oom_victim(tsk))
4230 * !MMU doesn't have oom reaper so give access to memory reserves
4231 * only to the thread with TIF_MEMDIE set
4233 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4240 * Distinguish requests which really need access to full memory
4241 * reserves from oom victims which can live with a portion of it
4243 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4245 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4247 if (gfp_mask & __GFP_MEMALLOC)
4248 return ALLOC_NO_WATERMARKS;
4249 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4250 return ALLOC_NO_WATERMARKS;
4251 if (!in_interrupt()) {
4252 if (current->flags & PF_MEMALLOC)
4253 return ALLOC_NO_WATERMARKS;
4254 else if (oom_reserves_allowed(current))
4261 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4263 return !!__gfp_pfmemalloc_flags(gfp_mask);
4267 * Checks whether it makes sense to retry the reclaim to make a forward progress
4268 * for the given allocation request.
4270 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4271 * without success, or when we couldn't even meet the watermark if we
4272 * reclaimed all remaining pages on the LRU lists.
4274 * Returns true if a retry is viable or false to enter the oom path.
4277 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4278 struct alloc_context *ac, int alloc_flags,
4279 bool did_some_progress, int *no_progress_loops)
4286 * Costly allocations might have made a progress but this doesn't mean
4287 * their order will become available due to high fragmentation so
4288 * always increment the no progress counter for them
4290 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4291 *no_progress_loops = 0;
4293 (*no_progress_loops)++;
4296 * Make sure we converge to OOM if we cannot make any progress
4297 * several times in the row.
4299 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4300 /* Before OOM, exhaust highatomic_reserve */
4301 return unreserve_highatomic_pageblock(ac, true);
4305 * Keep reclaiming pages while there is a chance this will lead
4306 * somewhere. If none of the target zones can satisfy our allocation
4307 * request even if all reclaimable pages are considered then we are
4308 * screwed and have to go OOM.
4310 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4312 unsigned long available;
4313 unsigned long reclaimable;
4314 unsigned long min_wmark = min_wmark_pages(zone);
4317 available = reclaimable = zone_reclaimable_pages(zone);
4318 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4321 * Would the allocation succeed if we reclaimed all
4322 * reclaimable pages?
4324 wmark = __zone_watermark_ok(zone, order, min_wmark,
4325 ac_classzone_idx(ac), alloc_flags, available);
4326 trace_reclaim_retry_zone(z, order, reclaimable,
4327 available, min_wmark, *no_progress_loops, wmark);
4330 * If we didn't make any progress and have a lot of
4331 * dirty + writeback pages then we should wait for
4332 * an IO to complete to slow down the reclaim and
4333 * prevent from pre mature OOM
4335 if (!did_some_progress) {
4336 unsigned long write_pending;
4338 write_pending = zone_page_state_snapshot(zone,
4339 NR_ZONE_WRITE_PENDING);
4341 if (2 * write_pending > reclaimable) {
4342 congestion_wait(BLK_RW_ASYNC, HZ/10);
4354 * Memory allocation/reclaim might be called from a WQ context and the
4355 * current implementation of the WQ concurrency control doesn't
4356 * recognize that a particular WQ is congested if the worker thread is
4357 * looping without ever sleeping. Therefore we have to do a short sleep
4358 * here rather than calling cond_resched().
4360 if (current->flags & PF_WQ_WORKER)
4361 schedule_timeout_uninterruptible(1);
4368 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4371 * It's possible that cpuset's mems_allowed and the nodemask from
4372 * mempolicy don't intersect. This should be normally dealt with by
4373 * policy_nodemask(), but it's possible to race with cpuset update in
4374 * such a way the check therein was true, and then it became false
4375 * before we got our cpuset_mems_cookie here.
4376 * This assumes that for all allocations, ac->nodemask can come only
4377 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4378 * when it does not intersect with the cpuset restrictions) or the
4379 * caller can deal with a violated nodemask.
4381 if (cpusets_enabled() && ac->nodemask &&
4382 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4383 ac->nodemask = NULL;
4388 * When updating a task's mems_allowed or mempolicy nodemask, it is
4389 * possible to race with parallel threads in such a way that our
4390 * allocation can fail while the mask is being updated. If we are about
4391 * to fail, check if the cpuset changed during allocation and if so,
4394 if (read_mems_allowed_retry(cpuset_mems_cookie))
4400 static inline struct page *
4401 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4402 struct alloc_context *ac)
4404 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4405 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4406 struct page *page = NULL;
4407 unsigned int alloc_flags;
4408 unsigned long did_some_progress;
4409 enum compact_priority compact_priority;
4410 enum compact_result compact_result;
4411 int compaction_retries;
4412 int no_progress_loops;
4413 unsigned int cpuset_mems_cookie;
4417 * We also sanity check to catch abuse of atomic reserves being used by
4418 * callers that are not in atomic context.
4420 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4421 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4422 gfp_mask &= ~__GFP_ATOMIC;
4425 compaction_retries = 0;
4426 no_progress_loops = 0;
4427 compact_priority = DEF_COMPACT_PRIORITY;
4428 cpuset_mems_cookie = read_mems_allowed_begin();
4431 * The fast path uses conservative alloc_flags to succeed only until
4432 * kswapd needs to be woken up, and to avoid the cost of setting up
4433 * alloc_flags precisely. So we do that now.
4435 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4438 * We need to recalculate the starting point for the zonelist iterator
4439 * because we might have used different nodemask in the fast path, or
4440 * there was a cpuset modification and we are retrying - otherwise we
4441 * could end up iterating over non-eligible zones endlessly.
4443 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4444 ac->high_zoneidx, ac->nodemask);
4445 if (!ac->preferred_zoneref->zone)
4448 if (alloc_flags & ALLOC_KSWAPD)
4449 wake_all_kswapds(order, gfp_mask, ac);
4452 * The adjusted alloc_flags might result in immediate success, so try
4455 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4460 * For costly allocations, try direct compaction first, as it's likely
4461 * that we have enough base pages and don't need to reclaim. For non-
4462 * movable high-order allocations, do that as well, as compaction will
4463 * try prevent permanent fragmentation by migrating from blocks of the
4465 * Don't try this for allocations that are allowed to ignore
4466 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4468 if (can_direct_reclaim &&
4470 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4471 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4472 page = __alloc_pages_direct_compact(gfp_mask, order,
4474 INIT_COMPACT_PRIORITY,
4480 * Checks for costly allocations with __GFP_NORETRY, which
4481 * includes some THP page fault allocations
4483 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4485 * If allocating entire pageblock(s) and compaction
4486 * failed because all zones are below low watermarks
4487 * or is prohibited because it recently failed at this
4488 * order, fail immediately unless the allocator has
4489 * requested compaction and reclaim retry.
4492 * - potentially very expensive because zones are far
4493 * below their low watermarks or this is part of very
4494 * bursty high order allocations,
4495 * - not guaranteed to help because isolate_freepages()
4496 * may not iterate over freed pages as part of its
4498 * - unlikely to make entire pageblocks free on its
4501 if (compact_result == COMPACT_SKIPPED ||
4502 compact_result == COMPACT_DEFERRED)
4506 * Looks like reclaim/compaction is worth trying, but
4507 * sync compaction could be very expensive, so keep
4508 * using async compaction.
4510 compact_priority = INIT_COMPACT_PRIORITY;
4515 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4516 if (alloc_flags & ALLOC_KSWAPD)
4517 wake_all_kswapds(order, gfp_mask, ac);
4519 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4521 alloc_flags = reserve_flags;
4524 * Reset the nodemask and zonelist iterators if memory policies can be
4525 * ignored. These allocations are high priority and system rather than
4528 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4529 ac->nodemask = NULL;
4530 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4531 ac->high_zoneidx, ac->nodemask);
4534 /* Attempt with potentially adjusted zonelist and alloc_flags */
4535 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4539 /* Caller is not willing to reclaim, we can't balance anything */
4540 if (!can_direct_reclaim)
4543 /* Avoid recursion of direct reclaim */
4544 if (current->flags & PF_MEMALLOC)
4547 /* Try direct reclaim and then allocating */
4548 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4549 &did_some_progress);
4553 /* Try direct compaction and then allocating */
4554 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4555 compact_priority, &compact_result);
4559 /* Do not loop if specifically requested */
4560 if (gfp_mask & __GFP_NORETRY)
4564 * Do not retry costly high order allocations unless they are
4565 * __GFP_RETRY_MAYFAIL
4567 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4570 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4571 did_some_progress > 0, &no_progress_loops))
4575 * It doesn't make any sense to retry for the compaction if the order-0
4576 * reclaim is not able to make any progress because the current
4577 * implementation of the compaction depends on the sufficient amount
4578 * of free memory (see __compaction_suitable)
4580 if (did_some_progress > 0 &&
4581 should_compact_retry(ac, order, alloc_flags,
4582 compact_result, &compact_priority,
4583 &compaction_retries))
4587 /* Deal with possible cpuset update races before we start OOM killing */
4588 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4591 /* Reclaim has failed us, start killing things */
4592 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4596 /* Avoid allocations with no watermarks from looping endlessly */
4597 if (tsk_is_oom_victim(current) &&
4598 (alloc_flags == ALLOC_OOM ||
4599 (gfp_mask & __GFP_NOMEMALLOC)))
4602 /* Retry as long as the OOM killer is making progress */
4603 if (did_some_progress) {
4604 no_progress_loops = 0;
4609 /* Deal with possible cpuset update races before we fail */
4610 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4614 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4617 if (gfp_mask & __GFP_NOFAIL) {
4619 * All existing users of the __GFP_NOFAIL are blockable, so warn
4620 * of any new users that actually require GFP_NOWAIT
4622 if (WARN_ON_ONCE(!can_direct_reclaim))
4626 * PF_MEMALLOC request from this context is rather bizarre
4627 * because we cannot reclaim anything and only can loop waiting
4628 * for somebody to do a work for us
4630 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4633 * non failing costly orders are a hard requirement which we
4634 * are not prepared for much so let's warn about these users
4635 * so that we can identify them and convert them to something
4638 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4641 * Help non-failing allocations by giving them access to memory
4642 * reserves but do not use ALLOC_NO_WATERMARKS because this
4643 * could deplete whole memory reserves which would just make
4644 * the situation worse
4646 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4654 warn_alloc(gfp_mask, ac->nodemask,
4655 "page allocation failure: order:%u", order);
4660 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4661 int preferred_nid, nodemask_t *nodemask,
4662 struct alloc_context *ac, gfp_t *alloc_mask,
4663 unsigned int *alloc_flags)
4665 ac->high_zoneidx = gfp_zone(gfp_mask);
4666 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4667 ac->nodemask = nodemask;
4668 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4670 if (cpusets_enabled()) {
4671 *alloc_mask |= __GFP_HARDWALL;
4673 ac->nodemask = &cpuset_current_mems_allowed;
4675 *alloc_flags |= ALLOC_CPUSET;
4678 fs_reclaim_acquire(gfp_mask);
4679 fs_reclaim_release(gfp_mask);
4681 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4683 if (should_fail_alloc_page(gfp_mask, order))
4686 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4687 *alloc_flags |= ALLOC_CMA;
4692 /* Determine whether to spread dirty pages and what the first usable zone */
4693 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4695 /* Dirty zone balancing only done in the fast path */
4696 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4699 * The preferred zone is used for statistics but crucially it is
4700 * also used as the starting point for the zonelist iterator. It
4701 * may get reset for allocations that ignore memory policies.
4703 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4704 ac->high_zoneidx, ac->nodemask);
4708 * This is the 'heart' of the zoned buddy allocator.
4711 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4712 nodemask_t *nodemask)
4715 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4716 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4717 struct alloc_context ac = { };
4720 * There are several places where we assume that the order value is sane
4721 * so bail out early if the request is out of bound.
4723 if (unlikely(order >= MAX_ORDER)) {
4724 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4728 gfp_mask &= gfp_allowed_mask;
4729 alloc_mask = gfp_mask;
4730 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4733 finalise_ac(gfp_mask, &ac);
4736 * Forbid the first pass from falling back to types that fragment
4737 * memory until all local zones are considered.
4739 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4741 /* First allocation attempt */
4742 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4747 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4748 * resp. GFP_NOIO which has to be inherited for all allocation requests
4749 * from a particular context which has been marked by
4750 * memalloc_no{fs,io}_{save,restore}.
4752 alloc_mask = current_gfp_context(gfp_mask);
4753 ac.spread_dirty_pages = false;
4756 * Restore the original nodemask if it was potentially replaced with
4757 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4759 if (unlikely(ac.nodemask != nodemask))
4760 ac.nodemask = nodemask;
4762 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4765 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4766 unlikely(__memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4767 __free_pages(page, order);
4771 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4775 EXPORT_SYMBOL(__alloc_pages_nodemask);
4778 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4779 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4780 * you need to access high mem.
4782 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4786 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4789 return (unsigned long) page_address(page);
4791 EXPORT_SYMBOL(__get_free_pages);
4793 unsigned long get_zeroed_page(gfp_t gfp_mask)
4795 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4797 EXPORT_SYMBOL(get_zeroed_page);
4799 static inline void free_the_page(struct page *page, unsigned int order)
4801 if (order == 0) /* Via pcp? */
4802 free_unref_page(page);
4804 __free_pages_ok(page, order);
4807 void __free_pages(struct page *page, unsigned int order)
4809 if (put_page_testzero(page))
4810 free_the_page(page, order);
4812 EXPORT_SYMBOL(__free_pages);
4814 void free_pages(unsigned long addr, unsigned int order)
4817 VM_BUG_ON(!virt_addr_valid((void *)addr));
4818 __free_pages(virt_to_page((void *)addr), order);
4822 EXPORT_SYMBOL(free_pages);
4826 * An arbitrary-length arbitrary-offset area of memory which resides
4827 * within a 0 or higher order page. Multiple fragments within that page
4828 * are individually refcounted, in the page's reference counter.
4830 * The page_frag functions below provide a simple allocation framework for
4831 * page fragments. This is used by the network stack and network device
4832 * drivers to provide a backing region of memory for use as either an
4833 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4835 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4838 struct page *page = NULL;
4839 gfp_t gfp = gfp_mask;
4841 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4842 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4844 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4845 PAGE_FRAG_CACHE_MAX_ORDER);
4846 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4848 if (unlikely(!page))
4849 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4851 nc->va = page ? page_address(page) : NULL;
4856 void __page_frag_cache_drain(struct page *page, unsigned int count)
4858 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4860 if (page_ref_sub_and_test(page, count))
4861 free_the_page(page, compound_order(page));
4863 EXPORT_SYMBOL(__page_frag_cache_drain);
4865 void *page_frag_alloc(struct page_frag_cache *nc,
4866 unsigned int fragsz, gfp_t gfp_mask)
4868 unsigned int size = PAGE_SIZE;
4872 if (unlikely(!nc->va)) {
4874 page = __page_frag_cache_refill(nc, gfp_mask);
4878 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4879 /* if size can vary use size else just use PAGE_SIZE */
4882 /* Even if we own the page, we do not use atomic_set().
4883 * This would break get_page_unless_zero() users.
4885 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4887 /* reset page count bias and offset to start of new frag */
4888 nc->pfmemalloc = page_is_pfmemalloc(page);
4889 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4893 offset = nc->offset - fragsz;
4894 if (unlikely(offset < 0)) {
4895 page = virt_to_page(nc->va);
4897 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4900 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4901 /* if size can vary use size else just use PAGE_SIZE */
4904 /* OK, page count is 0, we can safely set it */
4905 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4907 /* reset page count bias and offset to start of new frag */
4908 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4909 offset = size - fragsz;
4913 nc->offset = offset;
4915 return nc->va + offset;
4917 EXPORT_SYMBOL(page_frag_alloc);
4920 * Frees a page fragment allocated out of either a compound or order 0 page.
4922 void page_frag_free(void *addr)
4924 struct page *page = virt_to_head_page(addr);
4926 if (unlikely(put_page_testzero(page)))
4927 free_the_page(page, compound_order(page));
4929 EXPORT_SYMBOL(page_frag_free);
4931 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4935 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4936 unsigned long used = addr + PAGE_ALIGN(size);
4938 split_page(virt_to_page((void *)addr), order);
4939 while (used < alloc_end) {
4944 return (void *)addr;
4948 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4949 * @size: the number of bytes to allocate
4950 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4952 * This function is similar to alloc_pages(), except that it allocates the
4953 * minimum number of pages to satisfy the request. alloc_pages() can only
4954 * allocate memory in power-of-two pages.
4956 * This function is also limited by MAX_ORDER.
4958 * Memory allocated by this function must be released by free_pages_exact().
4960 * Return: pointer to the allocated area or %NULL in case of error.
4962 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4964 unsigned int order = get_order(size);
4967 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
4968 gfp_mask &= ~__GFP_COMP;
4970 addr = __get_free_pages(gfp_mask, order);
4971 return make_alloc_exact(addr, order, size);
4973 EXPORT_SYMBOL(alloc_pages_exact);
4976 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4978 * @nid: the preferred node ID where memory should be allocated
4979 * @size: the number of bytes to allocate
4980 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4982 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4985 * Return: pointer to the allocated area or %NULL in case of error.
4987 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4989 unsigned int order = get_order(size);
4992 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
4993 gfp_mask &= ~__GFP_COMP;
4995 p = alloc_pages_node(nid, gfp_mask, order);
4998 return make_alloc_exact((unsigned long)page_address(p), order, size);
5002 * free_pages_exact - release memory allocated via alloc_pages_exact()
5003 * @virt: the value returned by alloc_pages_exact.
5004 * @size: size of allocation, same value as passed to alloc_pages_exact().
5006 * Release the memory allocated by a previous call to alloc_pages_exact.
5008 void free_pages_exact(void *virt, size_t size)
5010 unsigned long addr = (unsigned long)virt;
5011 unsigned long end = addr + PAGE_ALIGN(size);
5013 while (addr < end) {
5018 EXPORT_SYMBOL(free_pages_exact);
5021 * nr_free_zone_pages - count number of pages beyond high watermark
5022 * @offset: The zone index of the highest zone
5024 * nr_free_zone_pages() counts the number of pages which are beyond the
5025 * high watermark within all zones at or below a given zone index. For each
5026 * zone, the number of pages is calculated as:
5028 * nr_free_zone_pages = managed_pages - high_pages
5030 * Return: number of pages beyond high watermark.
5032 static unsigned long nr_free_zone_pages(int offset)
5037 /* Just pick one node, since fallback list is circular */
5038 unsigned long sum = 0;
5040 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5042 for_each_zone_zonelist(zone, z, zonelist, offset) {
5043 unsigned long size = zone_managed_pages(zone);
5044 unsigned long high = high_wmark_pages(zone);
5053 * nr_free_buffer_pages - count number of pages beyond high watermark
5055 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5056 * watermark within ZONE_DMA and ZONE_NORMAL.
5058 * Return: number of pages beyond high watermark within ZONE_DMA and
5061 unsigned long nr_free_buffer_pages(void)
5063 return nr_free_zone_pages(gfp_zone(GFP_USER));
5065 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5068 * nr_free_pagecache_pages - count number of pages beyond high watermark
5070 * nr_free_pagecache_pages() counts the number of pages which are beyond the
5071 * high watermark within all zones.
5073 * Return: number of pages beyond high watermark within all zones.
5075 unsigned long nr_free_pagecache_pages(void)
5077 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5080 static inline void show_node(struct zone *zone)
5082 if (IS_ENABLED(CONFIG_NUMA))
5083 printk("Node %d ", zone_to_nid(zone));
5086 long si_mem_available(void)
5089 unsigned long pagecache;
5090 unsigned long wmark_low = 0;
5091 unsigned long pages[NR_LRU_LISTS];
5092 unsigned long reclaimable;
5096 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5097 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5100 wmark_low += low_wmark_pages(zone);
5103 * Estimate the amount of memory available for userspace allocations,
5104 * without causing swapping.
5106 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5109 * Not all the page cache can be freed, otherwise the system will
5110 * start swapping. Assume at least half of the page cache, or the
5111 * low watermark worth of cache, needs to stay.
5113 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5114 pagecache -= min(pagecache / 2, wmark_low);
5115 available += pagecache;
5118 * Part of the reclaimable slab and other kernel memory consists of
5119 * items that are in use, and cannot be freed. Cap this estimate at the
5122 reclaimable = global_node_page_state(NR_SLAB_RECLAIMABLE) +
5123 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5124 available += reclaimable - min(reclaimable / 2, wmark_low);
5130 EXPORT_SYMBOL_GPL(si_mem_available);
5132 void si_meminfo(struct sysinfo *val)
5134 val->totalram = totalram_pages();
5135 val->sharedram = global_node_page_state(NR_SHMEM);
5136 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5137 val->bufferram = nr_blockdev_pages();
5138 val->totalhigh = totalhigh_pages();
5139 val->freehigh = nr_free_highpages();
5140 val->mem_unit = PAGE_SIZE;
5143 EXPORT_SYMBOL(si_meminfo);
5146 void si_meminfo_node(struct sysinfo *val, int nid)
5148 int zone_type; /* needs to be signed */
5149 unsigned long managed_pages = 0;
5150 unsigned long managed_highpages = 0;
5151 unsigned long free_highpages = 0;
5152 pg_data_t *pgdat = NODE_DATA(nid);
5154 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5155 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5156 val->totalram = managed_pages;
5157 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5158 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5159 #ifdef CONFIG_HIGHMEM
5160 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5161 struct zone *zone = &pgdat->node_zones[zone_type];
5163 if (is_highmem(zone)) {
5164 managed_highpages += zone_managed_pages(zone);
5165 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5168 val->totalhigh = managed_highpages;
5169 val->freehigh = free_highpages;
5171 val->totalhigh = managed_highpages;
5172 val->freehigh = free_highpages;
5174 val->mem_unit = PAGE_SIZE;
5179 * Determine whether the node should be displayed or not, depending on whether
5180 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5182 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5184 if (!(flags & SHOW_MEM_FILTER_NODES))
5188 * no node mask - aka implicit memory numa policy. Do not bother with
5189 * the synchronization - read_mems_allowed_begin - because we do not
5190 * have to be precise here.
5193 nodemask = &cpuset_current_mems_allowed;
5195 return !node_isset(nid, *nodemask);
5198 #define K(x) ((x) << (PAGE_SHIFT-10))
5200 static void show_migration_types(unsigned char type)
5202 static const char types[MIGRATE_TYPES] = {
5203 [MIGRATE_UNMOVABLE] = 'U',
5204 [MIGRATE_MOVABLE] = 'M',
5205 [MIGRATE_RECLAIMABLE] = 'E',
5206 [MIGRATE_HIGHATOMIC] = 'H',
5208 [MIGRATE_CMA] = 'C',
5210 #ifdef CONFIG_MEMORY_ISOLATION
5211 [MIGRATE_ISOLATE] = 'I',
5214 char tmp[MIGRATE_TYPES + 1];
5218 for (i = 0; i < MIGRATE_TYPES; i++) {
5219 if (type & (1 << i))
5224 printk(KERN_CONT "(%s) ", tmp);
5228 * Show free area list (used inside shift_scroll-lock stuff)
5229 * We also calculate the percentage fragmentation. We do this by counting the
5230 * memory on each free list with the exception of the first item on the list.
5233 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5236 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5238 unsigned long free_pcp = 0;
5243 for_each_populated_zone(zone) {
5244 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5247 for_each_online_cpu(cpu)
5248 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5251 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5252 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5253 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
5254 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5255 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5256 " free:%lu free_pcp:%lu free_cma:%lu\n",
5257 global_node_page_state(NR_ACTIVE_ANON),
5258 global_node_page_state(NR_INACTIVE_ANON),
5259 global_node_page_state(NR_ISOLATED_ANON),
5260 global_node_page_state(NR_ACTIVE_FILE),
5261 global_node_page_state(NR_INACTIVE_FILE),
5262 global_node_page_state(NR_ISOLATED_FILE),
5263 global_node_page_state(NR_UNEVICTABLE),
5264 global_node_page_state(NR_FILE_DIRTY),
5265 global_node_page_state(NR_WRITEBACK),
5266 global_node_page_state(NR_UNSTABLE_NFS),
5267 global_node_page_state(NR_SLAB_RECLAIMABLE),
5268 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
5269 global_node_page_state(NR_FILE_MAPPED),
5270 global_node_page_state(NR_SHMEM),
5271 global_zone_page_state(NR_PAGETABLE),
5272 global_zone_page_state(NR_BOUNCE),
5273 global_zone_page_state(NR_FREE_PAGES),
5275 global_zone_page_state(NR_FREE_CMA_PAGES));
5277 for_each_online_pgdat(pgdat) {
5278 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5282 " active_anon:%lukB"
5283 " inactive_anon:%lukB"
5284 " active_file:%lukB"
5285 " inactive_file:%lukB"
5286 " unevictable:%lukB"
5287 " isolated(anon):%lukB"
5288 " isolated(file):%lukB"
5293 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5295 " shmem_pmdmapped: %lukB"
5298 " writeback_tmp:%lukB"
5300 " all_unreclaimable? %s"
5303 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5304 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5305 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5306 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5307 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5308 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5309 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5310 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5311 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5312 K(node_page_state(pgdat, NR_WRITEBACK)),
5313 K(node_page_state(pgdat, NR_SHMEM)),
5314 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5315 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5316 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5318 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5320 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5321 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
5322 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5326 for_each_populated_zone(zone) {
5329 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5333 for_each_online_cpu(cpu)
5334 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5343 " reserved_highatomic:%luKB"
5344 " active_anon:%lukB"
5345 " inactive_anon:%lukB"
5346 " active_file:%lukB"
5347 " inactive_file:%lukB"
5348 " unevictable:%lukB"
5349 " writepending:%lukB"
5353 " kernel_stack:%lukB"
5361 K(zone_page_state(zone, NR_FREE_PAGES)),
5362 K(min_wmark_pages(zone)),
5363 K(low_wmark_pages(zone)),
5364 K(high_wmark_pages(zone)),
5365 K(zone->nr_reserved_highatomic),
5366 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5367 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5368 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5369 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5370 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5371 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5372 K(zone->present_pages),
5373 K(zone_managed_pages(zone)),
5374 K(zone_page_state(zone, NR_MLOCK)),
5375 zone_page_state(zone, NR_KERNEL_STACK_KB),
5376 K(zone_page_state(zone, NR_PAGETABLE)),
5377 K(zone_page_state(zone, NR_BOUNCE)),
5379 K(this_cpu_read(zone->pageset->pcp.count)),
5380 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5381 printk("lowmem_reserve[]:");
5382 for (i = 0; i < MAX_NR_ZONES; i++)
5383 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5384 printk(KERN_CONT "\n");
5387 for_each_populated_zone(zone) {
5389 unsigned long nr[MAX_ORDER], flags, total = 0;
5390 unsigned char types[MAX_ORDER];
5392 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5395 printk(KERN_CONT "%s: ", zone->name);
5397 spin_lock_irqsave(&zone->lock, flags);
5398 for (order = 0; order < MAX_ORDER; order++) {
5399 struct free_area *area = &zone->free_area[order];
5402 nr[order] = area->nr_free;
5403 total += nr[order] << order;
5406 for (type = 0; type < MIGRATE_TYPES; type++) {
5407 if (!free_area_empty(area, type))
5408 types[order] |= 1 << type;
5411 spin_unlock_irqrestore(&zone->lock, flags);
5412 for (order = 0; order < MAX_ORDER; order++) {
5413 printk(KERN_CONT "%lu*%lukB ",
5414 nr[order], K(1UL) << order);
5416 show_migration_types(types[order]);
5418 printk(KERN_CONT "= %lukB\n", K(total));
5421 hugetlb_show_meminfo();
5423 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5425 show_swap_cache_info();
5428 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5430 zoneref->zone = zone;
5431 zoneref->zone_idx = zone_idx(zone);
5435 * Builds allocation fallback zone lists.
5437 * Add all populated zones of a node to the zonelist.
5439 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5442 enum zone_type zone_type = MAX_NR_ZONES;
5447 zone = pgdat->node_zones + zone_type;
5448 if (managed_zone(zone)) {
5449 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5450 check_highest_zone(zone_type);
5452 } while (zone_type);
5459 static int __parse_numa_zonelist_order(char *s)
5462 * We used to support different zonlists modes but they turned
5463 * out to be just not useful. Let's keep the warning in place
5464 * if somebody still use the cmd line parameter so that we do
5465 * not fail it silently
5467 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5468 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5474 static __init int setup_numa_zonelist_order(char *s)
5479 return __parse_numa_zonelist_order(s);
5481 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5483 char numa_zonelist_order[] = "Node";
5486 * sysctl handler for numa_zonelist_order
5488 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5489 void __user *buffer, size_t *length,
5496 return proc_dostring(table, write, buffer, length, ppos);
5497 str = memdup_user_nul(buffer, 16);
5499 return PTR_ERR(str);
5501 ret = __parse_numa_zonelist_order(str);
5507 #define MAX_NODE_LOAD (nr_online_nodes)
5508 static int node_load[MAX_NUMNODES];
5511 * find_next_best_node - find the next node that should appear in a given node's fallback list
5512 * @node: node whose fallback list we're appending
5513 * @used_node_mask: nodemask_t of already used nodes
5515 * We use a number of factors to determine which is the next node that should
5516 * appear on a given node's fallback list. The node should not have appeared
5517 * already in @node's fallback list, and it should be the next closest node
5518 * according to the distance array (which contains arbitrary distance values
5519 * from each node to each node in the system), and should also prefer nodes
5520 * with no CPUs, since presumably they'll have very little allocation pressure
5521 * on them otherwise.
5523 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5525 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5528 int min_val = INT_MAX;
5529 int best_node = NUMA_NO_NODE;
5530 const struct cpumask *tmp = cpumask_of_node(0);
5532 /* Use the local node if we haven't already */
5533 if (!node_isset(node, *used_node_mask)) {
5534 node_set(node, *used_node_mask);
5538 for_each_node_state(n, N_MEMORY) {
5540 /* Don't want a node to appear more than once */
5541 if (node_isset(n, *used_node_mask))
5544 /* Use the distance array to find the distance */
5545 val = node_distance(node, n);
5547 /* Penalize nodes under us ("prefer the next node") */
5550 /* Give preference to headless and unused nodes */
5551 tmp = cpumask_of_node(n);
5552 if (!cpumask_empty(tmp))
5553 val += PENALTY_FOR_NODE_WITH_CPUS;
5555 /* Slight preference for less loaded node */
5556 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5557 val += node_load[n];
5559 if (val < min_val) {
5566 node_set(best_node, *used_node_mask);
5573 * Build zonelists ordered by node and zones within node.
5574 * This results in maximum locality--normal zone overflows into local
5575 * DMA zone, if any--but risks exhausting DMA zone.
5577 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5580 struct zoneref *zonerefs;
5583 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5585 for (i = 0; i < nr_nodes; i++) {
5588 pg_data_t *node = NODE_DATA(node_order[i]);
5590 nr_zones = build_zonerefs_node(node, zonerefs);
5591 zonerefs += nr_zones;
5593 zonerefs->zone = NULL;
5594 zonerefs->zone_idx = 0;
5598 * Build gfp_thisnode zonelists
5600 static void build_thisnode_zonelists(pg_data_t *pgdat)
5602 struct zoneref *zonerefs;
5605 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5606 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5607 zonerefs += nr_zones;
5608 zonerefs->zone = NULL;
5609 zonerefs->zone_idx = 0;
5613 * Build zonelists ordered by zone and nodes within zones.
5614 * This results in conserving DMA zone[s] until all Normal memory is
5615 * exhausted, but results in overflowing to remote node while memory
5616 * may still exist in local DMA zone.
5619 static void build_zonelists(pg_data_t *pgdat)
5621 static int node_order[MAX_NUMNODES];
5622 int node, load, nr_nodes = 0;
5623 nodemask_t used_mask;
5624 int local_node, prev_node;
5626 /* NUMA-aware ordering of nodes */
5627 local_node = pgdat->node_id;
5628 load = nr_online_nodes;
5629 prev_node = local_node;
5630 nodes_clear(used_mask);
5632 memset(node_order, 0, sizeof(node_order));
5633 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5635 * We don't want to pressure a particular node.
5636 * So adding penalty to the first node in same
5637 * distance group to make it round-robin.
5639 if (node_distance(local_node, node) !=
5640 node_distance(local_node, prev_node))
5641 node_load[node] = load;
5643 node_order[nr_nodes++] = node;
5648 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5649 build_thisnode_zonelists(pgdat);
5652 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5654 * Return node id of node used for "local" allocations.
5655 * I.e., first node id of first zone in arg node's generic zonelist.
5656 * Used for initializing percpu 'numa_mem', which is used primarily
5657 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5659 int local_memory_node(int node)
5663 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5664 gfp_zone(GFP_KERNEL),
5666 return zone_to_nid(z->zone);
5670 static void setup_min_unmapped_ratio(void);
5671 static void setup_min_slab_ratio(void);
5672 #else /* CONFIG_NUMA */
5674 static void build_zonelists(pg_data_t *pgdat)
5676 int node, local_node;
5677 struct zoneref *zonerefs;
5680 local_node = pgdat->node_id;
5682 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5683 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5684 zonerefs += nr_zones;
5687 * Now we build the zonelist so that it contains the zones
5688 * of all the other nodes.
5689 * We don't want to pressure a particular node, so when
5690 * building the zones for node N, we make sure that the
5691 * zones coming right after the local ones are those from
5692 * node N+1 (modulo N)
5694 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5695 if (!node_online(node))
5697 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5698 zonerefs += nr_zones;
5700 for (node = 0; node < local_node; node++) {
5701 if (!node_online(node))
5703 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5704 zonerefs += nr_zones;
5707 zonerefs->zone = NULL;
5708 zonerefs->zone_idx = 0;
5711 #endif /* CONFIG_NUMA */
5714 * Boot pageset table. One per cpu which is going to be used for all
5715 * zones and all nodes. The parameters will be set in such a way
5716 * that an item put on a list will immediately be handed over to
5717 * the buddy list. This is safe since pageset manipulation is done
5718 * with interrupts disabled.
5720 * The boot_pagesets must be kept even after bootup is complete for
5721 * unused processors and/or zones. They do play a role for bootstrapping
5722 * hotplugged processors.
5724 * zoneinfo_show() and maybe other functions do
5725 * not check if the processor is online before following the pageset pointer.
5726 * Other parts of the kernel may not check if the zone is available.
5728 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5729 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5730 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5732 static void __build_all_zonelists(void *data)
5735 int __maybe_unused cpu;
5736 pg_data_t *self = data;
5737 static DEFINE_SPINLOCK(lock);
5742 memset(node_load, 0, sizeof(node_load));
5746 * This node is hotadded and no memory is yet present. So just
5747 * building zonelists is fine - no need to touch other nodes.
5749 if (self && !node_online(self->node_id)) {
5750 build_zonelists(self);
5752 for_each_online_node(nid) {
5753 pg_data_t *pgdat = NODE_DATA(nid);
5755 build_zonelists(pgdat);
5758 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5760 * We now know the "local memory node" for each node--
5761 * i.e., the node of the first zone in the generic zonelist.
5762 * Set up numa_mem percpu variable for on-line cpus. During
5763 * boot, only the boot cpu should be on-line; we'll init the
5764 * secondary cpus' numa_mem as they come on-line. During
5765 * node/memory hotplug, we'll fixup all on-line cpus.
5767 for_each_online_cpu(cpu)
5768 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5775 static noinline void __init
5776 build_all_zonelists_init(void)
5780 __build_all_zonelists(NULL);
5783 * Initialize the boot_pagesets that are going to be used
5784 * for bootstrapping processors. The real pagesets for
5785 * each zone will be allocated later when the per cpu
5786 * allocator is available.
5788 * boot_pagesets are used also for bootstrapping offline
5789 * cpus if the system is already booted because the pagesets
5790 * are needed to initialize allocators on a specific cpu too.
5791 * F.e. the percpu allocator needs the page allocator which
5792 * needs the percpu allocator in order to allocate its pagesets
5793 * (a chicken-egg dilemma).
5795 for_each_possible_cpu(cpu)
5796 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5798 mminit_verify_zonelist();
5799 cpuset_init_current_mems_allowed();
5803 * unless system_state == SYSTEM_BOOTING.
5805 * __ref due to call of __init annotated helper build_all_zonelists_init
5806 * [protected by SYSTEM_BOOTING].
5808 void __ref build_all_zonelists(pg_data_t *pgdat)
5810 if (system_state == SYSTEM_BOOTING) {
5811 build_all_zonelists_init();
5813 __build_all_zonelists(pgdat);
5814 /* cpuset refresh routine should be here */
5816 vm_total_pages = nr_free_pagecache_pages();
5818 * Disable grouping by mobility if the number of pages in the
5819 * system is too low to allow the mechanism to work. It would be
5820 * more accurate, but expensive to check per-zone. This check is
5821 * made on memory-hotadd so a system can start with mobility
5822 * disabled and enable it later
5824 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5825 page_group_by_mobility_disabled = 1;
5827 page_group_by_mobility_disabled = 0;
5829 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5831 page_group_by_mobility_disabled ? "off" : "on",
5834 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5838 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5839 static bool __meminit
5840 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5842 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5843 static struct memblock_region *r;
5845 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5846 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5847 for_each_memblock(memory, r) {
5848 if (*pfn < memblock_region_memory_end_pfn(r))
5852 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5853 memblock_is_mirror(r)) {
5854 *pfn = memblock_region_memory_end_pfn(r);
5863 * Initially all pages are reserved - free ones are freed
5864 * up by memblock_free_all() once the early boot process is
5865 * done. Non-atomic initialization, single-pass.
5867 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5868 unsigned long start_pfn, enum memmap_context context,
5869 struct vmem_altmap *altmap)
5871 unsigned long pfn, end_pfn = start_pfn + size;
5874 if (highest_memmap_pfn < end_pfn - 1)
5875 highest_memmap_pfn = end_pfn - 1;
5877 #ifdef CONFIG_ZONE_DEVICE
5879 * Honor reservation requested by the driver for this ZONE_DEVICE
5880 * memory. We limit the total number of pages to initialize to just
5881 * those that might contain the memory mapping. We will defer the
5882 * ZONE_DEVICE page initialization until after we have released
5885 if (zone == ZONE_DEVICE) {
5889 if (start_pfn == altmap->base_pfn)
5890 start_pfn += altmap->reserve;
5891 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5895 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5897 * There can be holes in boot-time mem_map[]s handed to this
5898 * function. They do not exist on hotplugged memory.
5900 if (context == MEMMAP_EARLY) {
5901 if (!early_pfn_valid(pfn))
5903 if (!early_pfn_in_nid(pfn, nid))
5905 if (overlap_memmap_init(zone, &pfn))
5907 if (defer_init(nid, pfn, end_pfn))
5911 page = pfn_to_page(pfn);
5912 __init_single_page(page, pfn, zone, nid);
5913 if (context == MEMMAP_HOTPLUG)
5914 __SetPageReserved(page);
5917 * Mark the block movable so that blocks are reserved for
5918 * movable at startup. This will force kernel allocations
5919 * to reserve their blocks rather than leaking throughout
5920 * the address space during boot when many long-lived
5921 * kernel allocations are made.
5923 * bitmap is created for zone's valid pfn range. but memmap
5924 * can be created for invalid pages (for alignment)
5925 * check here not to call set_pageblock_migratetype() against
5928 if (!(pfn & (pageblock_nr_pages - 1))) {
5929 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5935 #ifdef CONFIG_ZONE_DEVICE
5936 void __ref memmap_init_zone_device(struct zone *zone,
5937 unsigned long start_pfn,
5939 struct dev_pagemap *pgmap)
5941 unsigned long pfn, end_pfn = start_pfn + size;
5942 struct pglist_data *pgdat = zone->zone_pgdat;
5943 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
5944 unsigned long zone_idx = zone_idx(zone);
5945 unsigned long start = jiffies;
5946 int nid = pgdat->node_id;
5948 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
5952 * The call to memmap_init_zone should have already taken care
5953 * of the pages reserved for the memmap, so we can just jump to
5954 * the end of that region and start processing the device pages.
5957 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5958 size = end_pfn - start_pfn;
5961 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5962 struct page *page = pfn_to_page(pfn);
5964 __init_single_page(page, pfn, zone_idx, nid);
5967 * Mark page reserved as it will need to wait for onlining
5968 * phase for it to be fully associated with a zone.
5970 * We can use the non-atomic __set_bit operation for setting
5971 * the flag as we are still initializing the pages.
5973 __SetPageReserved(page);
5976 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
5977 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
5978 * ever freed or placed on a driver-private list.
5980 page->pgmap = pgmap;
5981 page->zone_device_data = NULL;
5984 * Mark the block movable so that blocks are reserved for
5985 * movable at startup. This will force kernel allocations
5986 * to reserve their blocks rather than leaking throughout
5987 * the address space during boot when many long-lived
5988 * kernel allocations are made.
5990 * bitmap is created for zone's valid pfn range. but memmap
5991 * can be created for invalid pages (for alignment)
5992 * check here not to call set_pageblock_migratetype() against
5995 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
5996 * because this is done early in section_activate()
5998 if (!(pfn & (pageblock_nr_pages - 1))) {
5999 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6004 pr_info("%s initialised %lu pages in %ums\n", __func__,
6005 size, jiffies_to_msecs(jiffies - start));
6009 static void __meminit zone_init_free_lists(struct zone *zone)
6011 unsigned int order, t;
6012 for_each_migratetype_order(order, t) {
6013 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6014 zone->free_area[order].nr_free = 0;
6018 void __meminit __weak memmap_init(unsigned long size, int nid,
6019 unsigned long zone, unsigned long start_pfn)
6021 memmap_init_zone(size, nid, zone, start_pfn, MEMMAP_EARLY, NULL);
6024 static int zone_batchsize(struct zone *zone)
6030 * The per-cpu-pages pools are set to around 1000th of the
6033 batch = zone_managed_pages(zone) / 1024;
6034 /* But no more than a meg. */
6035 if (batch * PAGE_SIZE > 1024 * 1024)
6036 batch = (1024 * 1024) / PAGE_SIZE;
6037 batch /= 4; /* We effectively *= 4 below */
6042 * Clamp the batch to a 2^n - 1 value. Having a power
6043 * of 2 value was found to be more likely to have
6044 * suboptimal cache aliasing properties in some cases.
6046 * For example if 2 tasks are alternately allocating
6047 * batches of pages, one task can end up with a lot
6048 * of pages of one half of the possible page colors
6049 * and the other with pages of the other colors.
6051 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6056 /* The deferral and batching of frees should be suppressed under NOMMU
6059 * The problem is that NOMMU needs to be able to allocate large chunks
6060 * of contiguous memory as there's no hardware page translation to
6061 * assemble apparent contiguous memory from discontiguous pages.
6063 * Queueing large contiguous runs of pages for batching, however,
6064 * causes the pages to actually be freed in smaller chunks. As there
6065 * can be a significant delay between the individual batches being
6066 * recycled, this leads to the once large chunks of space being
6067 * fragmented and becoming unavailable for high-order allocations.
6074 * pcp->high and pcp->batch values are related and dependent on one another:
6075 * ->batch must never be higher then ->high.
6076 * The following function updates them in a safe manner without read side
6079 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6080 * those fields changing asynchronously (acording the the above rule).
6082 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6083 * outside of boot time (or some other assurance that no concurrent updaters
6086 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6087 unsigned long batch)
6089 /* start with a fail safe value for batch */
6093 /* Update high, then batch, in order */
6100 /* a companion to pageset_set_high() */
6101 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6103 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6106 static void pageset_init(struct per_cpu_pageset *p)
6108 struct per_cpu_pages *pcp;
6111 memset(p, 0, sizeof(*p));
6114 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6115 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6118 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6121 pageset_set_batch(p, batch);
6125 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6126 * to the value high for the pageset p.
6128 static void pageset_set_high(struct per_cpu_pageset *p,
6131 unsigned long batch = max(1UL, high / 4);
6132 if ((high / 4) > (PAGE_SHIFT * 8))
6133 batch = PAGE_SHIFT * 8;
6135 pageset_update(&p->pcp, high, batch);
6138 static void pageset_set_high_and_batch(struct zone *zone,
6139 struct per_cpu_pageset *pcp)
6141 if (percpu_pagelist_fraction)
6142 pageset_set_high(pcp,
6143 (zone_managed_pages(zone) /
6144 percpu_pagelist_fraction));
6146 pageset_set_batch(pcp, zone_batchsize(zone));
6149 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6151 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6154 pageset_set_high_and_batch(zone, pcp);
6157 void __meminit setup_zone_pageset(struct zone *zone)
6160 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6161 for_each_possible_cpu(cpu)
6162 zone_pageset_init(zone, cpu);
6166 * Allocate per cpu pagesets and initialize them.
6167 * Before this call only boot pagesets were available.
6169 void __init setup_per_cpu_pageset(void)
6171 struct pglist_data *pgdat;
6174 for_each_populated_zone(zone)
6175 setup_zone_pageset(zone);
6177 for_each_online_pgdat(pgdat)
6178 pgdat->per_cpu_nodestats =
6179 alloc_percpu(struct per_cpu_nodestat);
6182 static __meminit void zone_pcp_init(struct zone *zone)
6185 * per cpu subsystem is not up at this point. The following code
6186 * relies on the ability of the linker to provide the
6187 * offset of a (static) per cpu variable into the per cpu area.
6189 zone->pageset = &boot_pageset;
6191 if (populated_zone(zone))
6192 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6193 zone->name, zone->present_pages,
6194 zone_batchsize(zone));
6197 void __meminit init_currently_empty_zone(struct zone *zone,
6198 unsigned long zone_start_pfn,
6201 struct pglist_data *pgdat = zone->zone_pgdat;
6202 int zone_idx = zone_idx(zone) + 1;
6204 if (zone_idx > pgdat->nr_zones)
6205 pgdat->nr_zones = zone_idx;
6207 zone->zone_start_pfn = zone_start_pfn;
6209 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6210 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6212 (unsigned long)zone_idx(zone),
6213 zone_start_pfn, (zone_start_pfn + size));
6215 zone_init_free_lists(zone);
6216 zone->initialized = 1;
6219 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6220 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
6223 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
6225 int __meminit __early_pfn_to_nid(unsigned long pfn,
6226 struct mminit_pfnnid_cache *state)
6228 unsigned long start_pfn, end_pfn;
6231 if (state->last_start <= pfn && pfn < state->last_end)
6232 return state->last_nid;
6234 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
6235 if (nid != NUMA_NO_NODE) {
6236 state->last_start = start_pfn;
6237 state->last_end = end_pfn;
6238 state->last_nid = nid;
6243 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
6246 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
6247 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
6248 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
6250 * If an architecture guarantees that all ranges registered contain no holes
6251 * and may be freed, this this function may be used instead of calling
6252 * memblock_free_early_nid() manually.
6254 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
6256 unsigned long start_pfn, end_pfn;
6259 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
6260 start_pfn = min(start_pfn, max_low_pfn);
6261 end_pfn = min(end_pfn, max_low_pfn);
6263 if (start_pfn < end_pfn)
6264 memblock_free_early_nid(PFN_PHYS(start_pfn),
6265 (end_pfn - start_pfn) << PAGE_SHIFT,
6271 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6272 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6274 * If an architecture guarantees that all ranges registered contain no holes and may
6275 * be freed, this function may be used instead of calling memory_present() manually.
6277 void __init sparse_memory_present_with_active_regions(int nid)
6279 unsigned long start_pfn, end_pfn;
6282 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
6283 memory_present(this_nid, start_pfn, end_pfn);
6287 * get_pfn_range_for_nid - Return the start and end page frames for a node
6288 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6289 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6290 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6292 * It returns the start and end page frame of a node based on information
6293 * provided by memblock_set_node(). If called for a node
6294 * with no available memory, a warning is printed and the start and end
6297 void __init get_pfn_range_for_nid(unsigned int nid,
6298 unsigned long *start_pfn, unsigned long *end_pfn)
6300 unsigned long this_start_pfn, this_end_pfn;
6306 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6307 *start_pfn = min(*start_pfn, this_start_pfn);
6308 *end_pfn = max(*end_pfn, this_end_pfn);
6311 if (*start_pfn == -1UL)
6316 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6317 * assumption is made that zones within a node are ordered in monotonic
6318 * increasing memory addresses so that the "highest" populated zone is used
6320 static void __init find_usable_zone_for_movable(void)
6323 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6324 if (zone_index == ZONE_MOVABLE)
6327 if (arch_zone_highest_possible_pfn[zone_index] >
6328 arch_zone_lowest_possible_pfn[zone_index])
6332 VM_BUG_ON(zone_index == -1);
6333 movable_zone = zone_index;
6337 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6338 * because it is sized independent of architecture. Unlike the other zones,
6339 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6340 * in each node depending on the size of each node and how evenly kernelcore
6341 * is distributed. This helper function adjusts the zone ranges
6342 * provided by the architecture for a given node by using the end of the
6343 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6344 * zones within a node are in order of monotonic increases memory addresses
6346 static void __init adjust_zone_range_for_zone_movable(int nid,
6347 unsigned long zone_type,
6348 unsigned long node_start_pfn,
6349 unsigned long node_end_pfn,
6350 unsigned long *zone_start_pfn,
6351 unsigned long *zone_end_pfn)
6353 /* Only adjust if ZONE_MOVABLE is on this node */
6354 if (zone_movable_pfn[nid]) {
6355 /* Size ZONE_MOVABLE */
6356 if (zone_type == ZONE_MOVABLE) {
6357 *zone_start_pfn = zone_movable_pfn[nid];
6358 *zone_end_pfn = min(node_end_pfn,
6359 arch_zone_highest_possible_pfn[movable_zone]);
6361 /* Adjust for ZONE_MOVABLE starting within this range */
6362 } else if (!mirrored_kernelcore &&
6363 *zone_start_pfn < zone_movable_pfn[nid] &&
6364 *zone_end_pfn > zone_movable_pfn[nid]) {
6365 *zone_end_pfn = zone_movable_pfn[nid];
6367 /* Check if this whole range is within ZONE_MOVABLE */
6368 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6369 *zone_start_pfn = *zone_end_pfn;
6374 * Return the number of pages a zone spans in a node, including holes
6375 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6377 static unsigned long __init zone_spanned_pages_in_node(int nid,
6378 unsigned long zone_type,
6379 unsigned long node_start_pfn,
6380 unsigned long node_end_pfn,
6381 unsigned long *zone_start_pfn,
6382 unsigned long *zone_end_pfn,
6383 unsigned long *ignored)
6385 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6386 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6387 /* When hotadd a new node from cpu_up(), the node should be empty */
6388 if (!node_start_pfn && !node_end_pfn)
6391 /* Get the start and end of the zone */
6392 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6393 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6394 adjust_zone_range_for_zone_movable(nid, zone_type,
6395 node_start_pfn, node_end_pfn,
6396 zone_start_pfn, zone_end_pfn);
6398 /* Check that this node has pages within the zone's required range */
6399 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6402 /* Move the zone boundaries inside the node if necessary */
6403 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6404 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6406 /* Return the spanned pages */
6407 return *zone_end_pfn - *zone_start_pfn;
6411 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6412 * then all holes in the requested range will be accounted for.
6414 unsigned long __init __absent_pages_in_range(int nid,
6415 unsigned long range_start_pfn,
6416 unsigned long range_end_pfn)
6418 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6419 unsigned long start_pfn, end_pfn;
6422 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6423 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6424 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6425 nr_absent -= end_pfn - start_pfn;
6431 * absent_pages_in_range - Return number of page frames in holes within a range
6432 * @start_pfn: The start PFN to start searching for holes
6433 * @end_pfn: The end PFN to stop searching for holes
6435 * Return: the number of pages frames in memory holes within a range.
6437 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6438 unsigned long end_pfn)
6440 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6443 /* Return the number of page frames in holes in a zone on a node */
6444 static unsigned long __init zone_absent_pages_in_node(int nid,
6445 unsigned long zone_type,
6446 unsigned long node_start_pfn,
6447 unsigned long node_end_pfn,
6448 unsigned long *ignored)
6450 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6451 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6452 unsigned long zone_start_pfn, zone_end_pfn;
6453 unsigned long nr_absent;
6455 /* When hotadd a new node from cpu_up(), the node should be empty */
6456 if (!node_start_pfn && !node_end_pfn)
6459 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6460 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6462 adjust_zone_range_for_zone_movable(nid, zone_type,
6463 node_start_pfn, node_end_pfn,
6464 &zone_start_pfn, &zone_end_pfn);
6465 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6468 * ZONE_MOVABLE handling.
6469 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6472 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6473 unsigned long start_pfn, end_pfn;
6474 struct memblock_region *r;
6476 for_each_memblock(memory, r) {
6477 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6478 zone_start_pfn, zone_end_pfn);
6479 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6480 zone_start_pfn, zone_end_pfn);
6482 if (zone_type == ZONE_MOVABLE &&
6483 memblock_is_mirror(r))
6484 nr_absent += end_pfn - start_pfn;
6486 if (zone_type == ZONE_NORMAL &&
6487 !memblock_is_mirror(r))
6488 nr_absent += end_pfn - start_pfn;
6495 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6496 static inline unsigned long __init zone_spanned_pages_in_node(int nid,
6497 unsigned long zone_type,
6498 unsigned long node_start_pfn,
6499 unsigned long node_end_pfn,
6500 unsigned long *zone_start_pfn,
6501 unsigned long *zone_end_pfn,
6502 unsigned long *zones_size)
6506 *zone_start_pfn = node_start_pfn;
6507 for (zone = 0; zone < zone_type; zone++)
6508 *zone_start_pfn += zones_size[zone];
6510 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6512 return zones_size[zone_type];
6515 static inline unsigned long __init zone_absent_pages_in_node(int nid,
6516 unsigned long zone_type,
6517 unsigned long node_start_pfn,
6518 unsigned long node_end_pfn,
6519 unsigned long *zholes_size)
6524 return zholes_size[zone_type];
6527 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6529 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6530 unsigned long node_start_pfn,
6531 unsigned long node_end_pfn,
6532 unsigned long *zones_size,
6533 unsigned long *zholes_size)
6535 unsigned long realtotalpages = 0, totalpages = 0;
6538 for (i = 0; i < MAX_NR_ZONES; i++) {
6539 struct zone *zone = pgdat->node_zones + i;
6540 unsigned long zone_start_pfn, zone_end_pfn;
6541 unsigned long size, real_size;
6543 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6549 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6550 node_start_pfn, node_end_pfn,
6553 zone->zone_start_pfn = zone_start_pfn;
6555 zone->zone_start_pfn = 0;
6556 zone->spanned_pages = size;
6557 zone->present_pages = real_size;
6560 realtotalpages += real_size;
6563 pgdat->node_spanned_pages = totalpages;
6564 pgdat->node_present_pages = realtotalpages;
6565 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6569 #ifndef CONFIG_SPARSEMEM
6571 * Calculate the size of the zone->blockflags rounded to an unsigned long
6572 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6573 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6574 * round what is now in bits to nearest long in bits, then return it in
6577 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6579 unsigned long usemapsize;
6581 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6582 usemapsize = roundup(zonesize, pageblock_nr_pages);
6583 usemapsize = usemapsize >> pageblock_order;
6584 usemapsize *= NR_PAGEBLOCK_BITS;
6585 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6587 return usemapsize / 8;
6590 static void __ref setup_usemap(struct pglist_data *pgdat,
6592 unsigned long zone_start_pfn,
6593 unsigned long zonesize)
6595 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6596 zone->pageblock_flags = NULL;
6598 zone->pageblock_flags =
6599 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6601 if (!zone->pageblock_flags)
6602 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6603 usemapsize, zone->name, pgdat->node_id);
6607 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6608 unsigned long zone_start_pfn, unsigned long zonesize) {}
6609 #endif /* CONFIG_SPARSEMEM */
6611 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6613 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6614 void __init set_pageblock_order(void)
6618 /* Check that pageblock_nr_pages has not already been setup */
6619 if (pageblock_order)
6622 if (HPAGE_SHIFT > PAGE_SHIFT)
6623 order = HUGETLB_PAGE_ORDER;
6625 order = MAX_ORDER - 1;
6628 * Assume the largest contiguous order of interest is a huge page.
6629 * This value may be variable depending on boot parameters on IA64 and
6632 pageblock_order = order;
6634 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6637 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6638 * is unused as pageblock_order is set at compile-time. See
6639 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6642 void __init set_pageblock_order(void)
6646 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6648 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6649 unsigned long present_pages)
6651 unsigned long pages = spanned_pages;
6654 * Provide a more accurate estimation if there are holes within
6655 * the zone and SPARSEMEM is in use. If there are holes within the
6656 * zone, each populated memory region may cost us one or two extra
6657 * memmap pages due to alignment because memmap pages for each
6658 * populated regions may not be naturally aligned on page boundary.
6659 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6661 if (spanned_pages > present_pages + (present_pages >> 4) &&
6662 IS_ENABLED(CONFIG_SPARSEMEM))
6663 pages = present_pages;
6665 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6668 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6669 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6671 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6673 spin_lock_init(&ds_queue->split_queue_lock);
6674 INIT_LIST_HEAD(&ds_queue->split_queue);
6675 ds_queue->split_queue_len = 0;
6678 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6681 #ifdef CONFIG_COMPACTION
6682 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6684 init_waitqueue_head(&pgdat->kcompactd_wait);
6687 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6690 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6692 pgdat_resize_init(pgdat);
6694 pgdat_init_split_queue(pgdat);
6695 pgdat_init_kcompactd(pgdat);
6697 init_waitqueue_head(&pgdat->kswapd_wait);
6698 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6700 pgdat_page_ext_init(pgdat);
6701 spin_lock_init(&pgdat->lru_lock);
6702 lruvec_init(&pgdat->__lruvec);
6705 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6706 unsigned long remaining_pages)
6708 atomic_long_set(&zone->managed_pages, remaining_pages);
6709 zone_set_nid(zone, nid);
6710 zone->name = zone_names[idx];
6711 zone->zone_pgdat = NODE_DATA(nid);
6712 spin_lock_init(&zone->lock);
6713 zone_seqlock_init(zone);
6714 zone_pcp_init(zone);
6718 * Set up the zone data structures
6719 * - init pgdat internals
6720 * - init all zones belonging to this node
6722 * NOTE: this function is only called during memory hotplug
6724 #ifdef CONFIG_MEMORY_HOTPLUG
6725 void __ref free_area_init_core_hotplug(int nid)
6728 pg_data_t *pgdat = NODE_DATA(nid);
6730 pgdat_init_internals(pgdat);
6731 for (z = 0; z < MAX_NR_ZONES; z++)
6732 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6737 * Set up the zone data structures:
6738 * - mark all pages reserved
6739 * - mark all memory queues empty
6740 * - clear the memory bitmaps
6742 * NOTE: pgdat should get zeroed by caller.
6743 * NOTE: this function is only called during early init.
6745 static void __init free_area_init_core(struct pglist_data *pgdat)
6748 int nid = pgdat->node_id;
6750 pgdat_init_internals(pgdat);
6751 pgdat->per_cpu_nodestats = &boot_nodestats;
6753 for (j = 0; j < MAX_NR_ZONES; j++) {
6754 struct zone *zone = pgdat->node_zones + j;
6755 unsigned long size, freesize, memmap_pages;
6756 unsigned long zone_start_pfn = zone->zone_start_pfn;
6758 size = zone->spanned_pages;
6759 freesize = zone->present_pages;
6762 * Adjust freesize so that it accounts for how much memory
6763 * is used by this zone for memmap. This affects the watermark
6764 * and per-cpu initialisations
6766 memmap_pages = calc_memmap_size(size, freesize);
6767 if (!is_highmem_idx(j)) {
6768 if (freesize >= memmap_pages) {
6769 freesize -= memmap_pages;
6772 " %s zone: %lu pages used for memmap\n",
6773 zone_names[j], memmap_pages);
6775 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6776 zone_names[j], memmap_pages, freesize);
6779 /* Account for reserved pages */
6780 if (j == 0 && freesize > dma_reserve) {
6781 freesize -= dma_reserve;
6782 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6783 zone_names[0], dma_reserve);
6786 if (!is_highmem_idx(j))
6787 nr_kernel_pages += freesize;
6788 /* Charge for highmem memmap if there are enough kernel pages */
6789 else if (nr_kernel_pages > memmap_pages * 2)
6790 nr_kernel_pages -= memmap_pages;
6791 nr_all_pages += freesize;
6794 * Set an approximate value for lowmem here, it will be adjusted
6795 * when the bootmem allocator frees pages into the buddy system.
6796 * And all highmem pages will be managed by the buddy system.
6798 zone_init_internals(zone, j, nid, freesize);
6803 set_pageblock_order();
6804 setup_usemap(pgdat, zone, zone_start_pfn, size);
6805 init_currently_empty_zone(zone, zone_start_pfn, size);
6806 memmap_init(size, nid, j, zone_start_pfn);
6810 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6811 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6813 unsigned long __maybe_unused start = 0;
6814 unsigned long __maybe_unused offset = 0;
6816 /* Skip empty nodes */
6817 if (!pgdat->node_spanned_pages)
6820 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6821 offset = pgdat->node_start_pfn - start;
6822 /* ia64 gets its own node_mem_map, before this, without bootmem */
6823 if (!pgdat->node_mem_map) {
6824 unsigned long size, end;
6828 * The zone's endpoints aren't required to be MAX_ORDER
6829 * aligned but the node_mem_map endpoints must be in order
6830 * for the buddy allocator to function correctly.
6832 end = pgdat_end_pfn(pgdat);
6833 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6834 size = (end - start) * sizeof(struct page);
6835 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6838 panic("Failed to allocate %ld bytes for node %d memory map\n",
6839 size, pgdat->node_id);
6840 pgdat->node_mem_map = map + offset;
6842 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6843 __func__, pgdat->node_id, (unsigned long)pgdat,
6844 (unsigned long)pgdat->node_mem_map);
6845 #ifndef CONFIG_NEED_MULTIPLE_NODES
6847 * With no DISCONTIG, the global mem_map is just set as node 0's
6849 if (pgdat == NODE_DATA(0)) {
6850 mem_map = NODE_DATA(0)->node_mem_map;
6851 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6852 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6854 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6859 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6860 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6862 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6863 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6865 pgdat->first_deferred_pfn = ULONG_MAX;
6868 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6871 void __init free_area_init_node(int nid, unsigned long *zones_size,
6872 unsigned long node_start_pfn,
6873 unsigned long *zholes_size)
6875 pg_data_t *pgdat = NODE_DATA(nid);
6876 unsigned long start_pfn = 0;
6877 unsigned long end_pfn = 0;
6879 /* pg_data_t should be reset to zero when it's allocated */
6880 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6882 pgdat->node_id = nid;
6883 pgdat->node_start_pfn = node_start_pfn;
6884 pgdat->per_cpu_nodestats = NULL;
6885 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6886 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6887 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6888 (u64)start_pfn << PAGE_SHIFT,
6889 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6891 start_pfn = node_start_pfn;
6893 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6894 zones_size, zholes_size);
6896 alloc_node_mem_map(pgdat);
6897 pgdat_set_deferred_range(pgdat);
6899 free_area_init_core(pgdat);
6902 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6904 * Zero all valid struct pages in range [spfn, epfn), return number of struct
6907 static u64 zero_pfn_range(unsigned long spfn, unsigned long epfn)
6912 for (pfn = spfn; pfn < epfn; pfn++) {
6913 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6914 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6915 + pageblock_nr_pages - 1;
6918 mm_zero_struct_page(pfn_to_page(pfn));
6926 * Only struct pages that are backed by physical memory are zeroed and
6927 * initialized by going through __init_single_page(). But, there are some
6928 * struct pages which are reserved in memblock allocator and their fields
6929 * may be accessed (for example page_to_pfn() on some configuration accesses
6930 * flags). We must explicitly zero those struct pages.
6932 * This function also addresses a similar issue where struct pages are left
6933 * uninitialized because the physical address range is not covered by
6934 * memblock.memory or memblock.reserved. That could happen when memblock
6935 * layout is manually configured via memmap=.
6937 void __init zero_resv_unavail(void)
6939 phys_addr_t start, end;
6941 phys_addr_t next = 0;
6944 * Loop through unavailable ranges not covered by memblock.memory.
6947 for_each_mem_range(i, &memblock.memory, NULL,
6948 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
6950 pgcnt += zero_pfn_range(PFN_DOWN(next), PFN_UP(start));
6953 pgcnt += zero_pfn_range(PFN_DOWN(next), max_pfn);
6956 * Struct pages that do not have backing memory. This could be because
6957 * firmware is using some of this memory, or for some other reasons.
6960 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
6962 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
6964 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6966 #if MAX_NUMNODES > 1
6968 * Figure out the number of possible node ids.
6970 void __init setup_nr_node_ids(void)
6972 unsigned int highest;
6974 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6975 nr_node_ids = highest + 1;
6980 * node_map_pfn_alignment - determine the maximum internode alignment
6982 * This function should be called after node map is populated and sorted.
6983 * It calculates the maximum power of two alignment which can distinguish
6986 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6987 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6988 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6989 * shifted, 1GiB is enough and this function will indicate so.
6991 * This is used to test whether pfn -> nid mapping of the chosen memory
6992 * model has fine enough granularity to avoid incorrect mapping for the
6993 * populated node map.
6995 * Return: the determined alignment in pfn's. 0 if there is no alignment
6996 * requirement (single node).
6998 unsigned long __init node_map_pfn_alignment(void)
7000 unsigned long accl_mask = 0, last_end = 0;
7001 unsigned long start, end, mask;
7002 int last_nid = NUMA_NO_NODE;
7005 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7006 if (!start || last_nid < 0 || last_nid == nid) {
7013 * Start with a mask granular enough to pin-point to the
7014 * start pfn and tick off bits one-by-one until it becomes
7015 * too coarse to separate the current node from the last.
7017 mask = ~((1 << __ffs(start)) - 1);
7018 while (mask && last_end <= (start & (mask << 1)))
7021 /* accumulate all internode masks */
7025 /* convert mask to number of pages */
7026 return ~accl_mask + 1;
7029 /* Find the lowest pfn for a node */
7030 static unsigned long __init find_min_pfn_for_node(int nid)
7032 unsigned long min_pfn = ULONG_MAX;
7033 unsigned long start_pfn;
7036 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
7037 min_pfn = min(min_pfn, start_pfn);
7039 if (min_pfn == ULONG_MAX) {
7040 pr_warn("Could not find start_pfn for node %d\n", nid);
7048 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7050 * Return: the minimum PFN based on information provided via
7051 * memblock_set_node().
7053 unsigned long __init find_min_pfn_with_active_regions(void)
7055 return find_min_pfn_for_node(MAX_NUMNODES);
7059 * early_calculate_totalpages()
7060 * Sum pages in active regions for movable zone.
7061 * Populate N_MEMORY for calculating usable_nodes.
7063 static unsigned long __init early_calculate_totalpages(void)
7065 unsigned long totalpages = 0;
7066 unsigned long start_pfn, end_pfn;
7069 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7070 unsigned long pages = end_pfn - start_pfn;
7072 totalpages += pages;
7074 node_set_state(nid, N_MEMORY);
7080 * Find the PFN the Movable zone begins in each node. Kernel memory
7081 * is spread evenly between nodes as long as the nodes have enough
7082 * memory. When they don't, some nodes will have more kernelcore than
7085 static void __init find_zone_movable_pfns_for_nodes(void)
7088 unsigned long usable_startpfn;
7089 unsigned long kernelcore_node, kernelcore_remaining;
7090 /* save the state before borrow the nodemask */
7091 nodemask_t saved_node_state = node_states[N_MEMORY];
7092 unsigned long totalpages = early_calculate_totalpages();
7093 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7094 struct memblock_region *r;
7096 /* Need to find movable_zone earlier when movable_node is specified. */
7097 find_usable_zone_for_movable();
7100 * If movable_node is specified, ignore kernelcore and movablecore
7103 if (movable_node_is_enabled()) {
7104 for_each_memblock(memory, r) {
7105 if (!memblock_is_hotpluggable(r))
7110 usable_startpfn = PFN_DOWN(r->base);
7111 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7112 min(usable_startpfn, zone_movable_pfn[nid]) :
7120 * If kernelcore=mirror is specified, ignore movablecore option
7122 if (mirrored_kernelcore) {
7123 bool mem_below_4gb_not_mirrored = false;
7125 for_each_memblock(memory, r) {
7126 if (memblock_is_mirror(r))
7131 usable_startpfn = memblock_region_memory_base_pfn(r);
7133 if (usable_startpfn < 0x100000) {
7134 mem_below_4gb_not_mirrored = true;
7138 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7139 min(usable_startpfn, zone_movable_pfn[nid]) :
7143 if (mem_below_4gb_not_mirrored)
7144 pr_warn("This configuration results in unmirrored kernel memory.");
7150 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7151 * amount of necessary memory.
7153 if (required_kernelcore_percent)
7154 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7156 if (required_movablecore_percent)
7157 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7161 * If movablecore= was specified, calculate what size of
7162 * kernelcore that corresponds so that memory usable for
7163 * any allocation type is evenly spread. If both kernelcore
7164 * and movablecore are specified, then the value of kernelcore
7165 * will be used for required_kernelcore if it's greater than
7166 * what movablecore would have allowed.
7168 if (required_movablecore) {
7169 unsigned long corepages;
7172 * Round-up so that ZONE_MOVABLE is at least as large as what
7173 * was requested by the user
7175 required_movablecore =
7176 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7177 required_movablecore = min(totalpages, required_movablecore);
7178 corepages = totalpages - required_movablecore;
7180 required_kernelcore = max(required_kernelcore, corepages);
7184 * If kernelcore was not specified or kernelcore size is larger
7185 * than totalpages, there is no ZONE_MOVABLE.
7187 if (!required_kernelcore || required_kernelcore >= totalpages)
7190 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7191 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7194 /* Spread kernelcore memory as evenly as possible throughout nodes */
7195 kernelcore_node = required_kernelcore / usable_nodes;
7196 for_each_node_state(nid, N_MEMORY) {
7197 unsigned long start_pfn, end_pfn;
7200 * Recalculate kernelcore_node if the division per node
7201 * now exceeds what is necessary to satisfy the requested
7202 * amount of memory for the kernel
7204 if (required_kernelcore < kernelcore_node)
7205 kernelcore_node = required_kernelcore / usable_nodes;
7208 * As the map is walked, we track how much memory is usable
7209 * by the kernel using kernelcore_remaining. When it is
7210 * 0, the rest of the node is usable by ZONE_MOVABLE
7212 kernelcore_remaining = kernelcore_node;
7214 /* Go through each range of PFNs within this node */
7215 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7216 unsigned long size_pages;
7218 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7219 if (start_pfn >= end_pfn)
7222 /* Account for what is only usable for kernelcore */
7223 if (start_pfn < usable_startpfn) {
7224 unsigned long kernel_pages;
7225 kernel_pages = min(end_pfn, usable_startpfn)
7228 kernelcore_remaining -= min(kernel_pages,
7229 kernelcore_remaining);
7230 required_kernelcore -= min(kernel_pages,
7231 required_kernelcore);
7233 /* Continue if range is now fully accounted */
7234 if (end_pfn <= usable_startpfn) {
7237 * Push zone_movable_pfn to the end so
7238 * that if we have to rebalance
7239 * kernelcore across nodes, we will
7240 * not double account here
7242 zone_movable_pfn[nid] = end_pfn;
7245 start_pfn = usable_startpfn;
7249 * The usable PFN range for ZONE_MOVABLE is from
7250 * start_pfn->end_pfn. Calculate size_pages as the
7251 * number of pages used as kernelcore
7253 size_pages = end_pfn - start_pfn;
7254 if (size_pages > kernelcore_remaining)
7255 size_pages = kernelcore_remaining;
7256 zone_movable_pfn[nid] = start_pfn + size_pages;
7259 * Some kernelcore has been met, update counts and
7260 * break if the kernelcore for this node has been
7263 required_kernelcore -= min(required_kernelcore,
7265 kernelcore_remaining -= size_pages;
7266 if (!kernelcore_remaining)
7272 * If there is still required_kernelcore, we do another pass with one
7273 * less node in the count. This will push zone_movable_pfn[nid] further
7274 * along on the nodes that still have memory until kernelcore is
7278 if (usable_nodes && required_kernelcore > usable_nodes)
7282 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7283 for (nid = 0; nid < MAX_NUMNODES; nid++)
7284 zone_movable_pfn[nid] =
7285 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7288 /* restore the node_state */
7289 node_states[N_MEMORY] = saved_node_state;
7292 /* Any regular or high memory on that node ? */
7293 static void check_for_memory(pg_data_t *pgdat, int nid)
7295 enum zone_type zone_type;
7297 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7298 struct zone *zone = &pgdat->node_zones[zone_type];
7299 if (populated_zone(zone)) {
7300 if (IS_ENABLED(CONFIG_HIGHMEM))
7301 node_set_state(nid, N_HIGH_MEMORY);
7302 if (zone_type <= ZONE_NORMAL)
7303 node_set_state(nid, N_NORMAL_MEMORY);
7310 * free_area_init_nodes - Initialise all pg_data_t and zone data
7311 * @max_zone_pfn: an array of max PFNs for each zone
7313 * This will call free_area_init_node() for each active node in the system.
7314 * Using the page ranges provided by memblock_set_node(), the size of each
7315 * zone in each node and their holes is calculated. If the maximum PFN
7316 * between two adjacent zones match, it is assumed that the zone is empty.
7317 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7318 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7319 * starts where the previous one ended. For example, ZONE_DMA32 starts
7320 * at arch_max_dma_pfn.
7322 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
7324 unsigned long start_pfn, end_pfn;
7327 /* Record where the zone boundaries are */
7328 memset(arch_zone_lowest_possible_pfn, 0,
7329 sizeof(arch_zone_lowest_possible_pfn));
7330 memset(arch_zone_highest_possible_pfn, 0,
7331 sizeof(arch_zone_highest_possible_pfn));
7333 start_pfn = find_min_pfn_with_active_regions();
7335 for (i = 0; i < MAX_NR_ZONES; i++) {
7336 if (i == ZONE_MOVABLE)
7339 end_pfn = max(max_zone_pfn[i], start_pfn);
7340 arch_zone_lowest_possible_pfn[i] = start_pfn;
7341 arch_zone_highest_possible_pfn[i] = end_pfn;
7343 start_pfn = end_pfn;
7346 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7347 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7348 find_zone_movable_pfns_for_nodes();
7350 /* Print out the zone ranges */
7351 pr_info("Zone ranges:\n");
7352 for (i = 0; i < MAX_NR_ZONES; i++) {
7353 if (i == ZONE_MOVABLE)
7355 pr_info(" %-8s ", zone_names[i]);
7356 if (arch_zone_lowest_possible_pfn[i] ==
7357 arch_zone_highest_possible_pfn[i])
7360 pr_cont("[mem %#018Lx-%#018Lx]\n",
7361 (u64)arch_zone_lowest_possible_pfn[i]
7363 ((u64)arch_zone_highest_possible_pfn[i]
7364 << PAGE_SHIFT) - 1);
7367 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7368 pr_info("Movable zone start for each node\n");
7369 for (i = 0; i < MAX_NUMNODES; i++) {
7370 if (zone_movable_pfn[i])
7371 pr_info(" Node %d: %#018Lx\n", i,
7372 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7376 * Print out the early node map, and initialize the
7377 * subsection-map relative to active online memory ranges to
7378 * enable future "sub-section" extensions of the memory map.
7380 pr_info("Early memory node ranges\n");
7381 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7382 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7383 (u64)start_pfn << PAGE_SHIFT,
7384 ((u64)end_pfn << PAGE_SHIFT) - 1);
7385 subsection_map_init(start_pfn, end_pfn - start_pfn);
7388 /* Initialise every node */
7389 mminit_verify_pageflags_layout();
7390 setup_nr_node_ids();
7391 zero_resv_unavail();
7392 for_each_online_node(nid) {
7393 pg_data_t *pgdat = NODE_DATA(nid);
7394 free_area_init_node(nid, NULL,
7395 find_min_pfn_for_node(nid), NULL);
7397 /* Any memory on that node */
7398 if (pgdat->node_present_pages)
7399 node_set_state(nid, N_MEMORY);
7400 check_for_memory(pgdat, nid);
7404 static int __init cmdline_parse_core(char *p, unsigned long *core,
7405 unsigned long *percent)
7407 unsigned long long coremem;
7413 /* Value may be a percentage of total memory, otherwise bytes */
7414 coremem = simple_strtoull(p, &endptr, 0);
7415 if (*endptr == '%') {
7416 /* Paranoid check for percent values greater than 100 */
7417 WARN_ON(coremem > 100);
7421 coremem = memparse(p, &p);
7422 /* Paranoid check that UL is enough for the coremem value */
7423 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7425 *core = coremem >> PAGE_SHIFT;
7432 * kernelcore=size sets the amount of memory for use for allocations that
7433 * cannot be reclaimed or migrated.
7435 static int __init cmdline_parse_kernelcore(char *p)
7437 /* parse kernelcore=mirror */
7438 if (parse_option_str(p, "mirror")) {
7439 mirrored_kernelcore = true;
7443 return cmdline_parse_core(p, &required_kernelcore,
7444 &required_kernelcore_percent);
7448 * movablecore=size sets the amount of memory for use for allocations that
7449 * can be reclaimed or migrated.
7451 static int __init cmdline_parse_movablecore(char *p)
7453 return cmdline_parse_core(p, &required_movablecore,
7454 &required_movablecore_percent);
7457 early_param("kernelcore", cmdline_parse_kernelcore);
7458 early_param("movablecore", cmdline_parse_movablecore);
7460 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
7462 void adjust_managed_page_count(struct page *page, long count)
7464 atomic_long_add(count, &page_zone(page)->managed_pages);
7465 totalram_pages_add(count);
7466 #ifdef CONFIG_HIGHMEM
7467 if (PageHighMem(page))
7468 totalhigh_pages_add(count);
7471 EXPORT_SYMBOL(adjust_managed_page_count);
7473 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7476 unsigned long pages = 0;
7478 start = (void *)PAGE_ALIGN((unsigned long)start);
7479 end = (void *)((unsigned long)end & PAGE_MASK);
7480 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7481 struct page *page = virt_to_page(pos);
7482 void *direct_map_addr;
7485 * 'direct_map_addr' might be different from 'pos'
7486 * because some architectures' virt_to_page()
7487 * work with aliases. Getting the direct map
7488 * address ensures that we get a _writeable_
7489 * alias for the memset().
7491 direct_map_addr = page_address(page);
7492 if ((unsigned int)poison <= 0xFF)
7493 memset(direct_map_addr, poison, PAGE_SIZE);
7495 free_reserved_page(page);
7499 pr_info("Freeing %s memory: %ldK\n",
7500 s, pages << (PAGE_SHIFT - 10));
7505 #ifdef CONFIG_HIGHMEM
7506 void free_highmem_page(struct page *page)
7508 __free_reserved_page(page);
7509 totalram_pages_inc();
7510 atomic_long_inc(&page_zone(page)->managed_pages);
7511 totalhigh_pages_inc();
7516 void __init mem_init_print_info(const char *str)
7518 unsigned long physpages, codesize, datasize, rosize, bss_size;
7519 unsigned long init_code_size, init_data_size;
7521 physpages = get_num_physpages();
7522 codesize = _etext - _stext;
7523 datasize = _edata - _sdata;
7524 rosize = __end_rodata - __start_rodata;
7525 bss_size = __bss_stop - __bss_start;
7526 init_data_size = __init_end - __init_begin;
7527 init_code_size = _einittext - _sinittext;
7530 * Detect special cases and adjust section sizes accordingly:
7531 * 1) .init.* may be embedded into .data sections
7532 * 2) .init.text.* may be out of [__init_begin, __init_end],
7533 * please refer to arch/tile/kernel/vmlinux.lds.S.
7534 * 3) .rodata.* may be embedded into .text or .data sections.
7536 #define adj_init_size(start, end, size, pos, adj) \
7538 if (start <= pos && pos < end && size > adj) \
7542 adj_init_size(__init_begin, __init_end, init_data_size,
7543 _sinittext, init_code_size);
7544 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7545 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7546 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7547 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7549 #undef adj_init_size
7551 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7552 #ifdef CONFIG_HIGHMEM
7556 nr_free_pages() << (PAGE_SHIFT - 10),
7557 physpages << (PAGE_SHIFT - 10),
7558 codesize >> 10, datasize >> 10, rosize >> 10,
7559 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7560 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7561 totalcma_pages << (PAGE_SHIFT - 10),
7562 #ifdef CONFIG_HIGHMEM
7563 totalhigh_pages() << (PAGE_SHIFT - 10),
7565 str ? ", " : "", str ? str : "");
7569 * set_dma_reserve - set the specified number of pages reserved in the first zone
7570 * @new_dma_reserve: The number of pages to mark reserved
7572 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7573 * In the DMA zone, a significant percentage may be consumed by kernel image
7574 * and other unfreeable allocations which can skew the watermarks badly. This
7575 * function may optionally be used to account for unfreeable pages in the
7576 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7577 * smaller per-cpu batchsize.
7579 void __init set_dma_reserve(unsigned long new_dma_reserve)
7581 dma_reserve = new_dma_reserve;
7584 void __init free_area_init(unsigned long *zones_size)
7586 zero_resv_unavail();
7587 free_area_init_node(0, zones_size,
7588 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
7591 static int page_alloc_cpu_dead(unsigned int cpu)
7594 lru_add_drain_cpu(cpu);
7598 * Spill the event counters of the dead processor
7599 * into the current processors event counters.
7600 * This artificially elevates the count of the current
7603 vm_events_fold_cpu(cpu);
7606 * Zero the differential counters of the dead processor
7607 * so that the vm statistics are consistent.
7609 * This is only okay since the processor is dead and cannot
7610 * race with what we are doing.
7612 cpu_vm_stats_fold(cpu);
7617 int hashdist = HASHDIST_DEFAULT;
7619 static int __init set_hashdist(char *str)
7623 hashdist = simple_strtoul(str, &str, 0);
7626 __setup("hashdist=", set_hashdist);
7629 void __init page_alloc_init(void)
7634 if (num_node_state(N_MEMORY) == 1)
7638 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7639 "mm/page_alloc:dead", NULL,
7640 page_alloc_cpu_dead);
7645 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7646 * or min_free_kbytes changes.
7648 static void calculate_totalreserve_pages(void)
7650 struct pglist_data *pgdat;
7651 unsigned long reserve_pages = 0;
7652 enum zone_type i, j;
7654 for_each_online_pgdat(pgdat) {
7656 pgdat->totalreserve_pages = 0;
7658 for (i = 0; i < MAX_NR_ZONES; i++) {
7659 struct zone *zone = pgdat->node_zones + i;
7661 unsigned long managed_pages = zone_managed_pages(zone);
7663 /* Find valid and maximum lowmem_reserve in the zone */
7664 for (j = i; j < MAX_NR_ZONES; j++) {
7665 if (zone->lowmem_reserve[j] > max)
7666 max = zone->lowmem_reserve[j];
7669 /* we treat the high watermark as reserved pages. */
7670 max += high_wmark_pages(zone);
7672 if (max > managed_pages)
7673 max = managed_pages;
7675 pgdat->totalreserve_pages += max;
7677 reserve_pages += max;
7680 totalreserve_pages = reserve_pages;
7684 * setup_per_zone_lowmem_reserve - called whenever
7685 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7686 * has a correct pages reserved value, so an adequate number of
7687 * pages are left in the zone after a successful __alloc_pages().
7689 static void setup_per_zone_lowmem_reserve(void)
7691 struct pglist_data *pgdat;
7692 enum zone_type j, idx;
7694 for_each_online_pgdat(pgdat) {
7695 for (j = 0; j < MAX_NR_ZONES; j++) {
7696 struct zone *zone = pgdat->node_zones + j;
7697 unsigned long managed_pages = zone_managed_pages(zone);
7699 zone->lowmem_reserve[j] = 0;
7703 struct zone *lower_zone;
7706 lower_zone = pgdat->node_zones + idx;
7708 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7709 sysctl_lowmem_reserve_ratio[idx] = 0;
7710 lower_zone->lowmem_reserve[j] = 0;
7712 lower_zone->lowmem_reserve[j] =
7713 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7715 managed_pages += zone_managed_pages(lower_zone);
7720 /* update totalreserve_pages */
7721 calculate_totalreserve_pages();
7724 static void __setup_per_zone_wmarks(void)
7726 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7727 unsigned long lowmem_pages = 0;
7729 unsigned long flags;
7731 /* Calculate total number of !ZONE_HIGHMEM pages */
7732 for_each_zone(zone) {
7733 if (!is_highmem(zone))
7734 lowmem_pages += zone_managed_pages(zone);
7737 for_each_zone(zone) {
7740 spin_lock_irqsave(&zone->lock, flags);
7741 tmp = (u64)pages_min * zone_managed_pages(zone);
7742 do_div(tmp, lowmem_pages);
7743 if (is_highmem(zone)) {
7745 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7746 * need highmem pages, so cap pages_min to a small
7749 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7750 * deltas control async page reclaim, and so should
7751 * not be capped for highmem.
7753 unsigned long min_pages;
7755 min_pages = zone_managed_pages(zone) / 1024;
7756 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7757 zone->_watermark[WMARK_MIN] = min_pages;
7760 * If it's a lowmem zone, reserve a number of pages
7761 * proportionate to the zone's size.
7763 zone->_watermark[WMARK_MIN] = tmp;
7767 * Set the kswapd watermarks distance according to the
7768 * scale factor in proportion to available memory, but
7769 * ensure a minimum size on small systems.
7771 tmp = max_t(u64, tmp >> 2,
7772 mult_frac(zone_managed_pages(zone),
7773 watermark_scale_factor, 10000));
7775 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7776 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7777 zone->watermark_boost = 0;
7779 spin_unlock_irqrestore(&zone->lock, flags);
7782 /* update totalreserve_pages */
7783 calculate_totalreserve_pages();
7787 * setup_per_zone_wmarks - called when min_free_kbytes changes
7788 * or when memory is hot-{added|removed}
7790 * Ensures that the watermark[min,low,high] values for each zone are set
7791 * correctly with respect to min_free_kbytes.
7793 void setup_per_zone_wmarks(void)
7795 static DEFINE_SPINLOCK(lock);
7798 __setup_per_zone_wmarks();
7803 * Initialise min_free_kbytes.
7805 * For small machines we want it small (128k min). For large machines
7806 * we want it large (64MB max). But it is not linear, because network
7807 * bandwidth does not increase linearly with machine size. We use
7809 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7810 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7826 int __meminit init_per_zone_wmark_min(void)
7828 unsigned long lowmem_kbytes;
7829 int new_min_free_kbytes;
7831 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7832 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7834 if (new_min_free_kbytes > user_min_free_kbytes) {
7835 min_free_kbytes = new_min_free_kbytes;
7836 if (min_free_kbytes < 128)
7837 min_free_kbytes = 128;
7838 if (min_free_kbytes > 65536)
7839 min_free_kbytes = 65536;
7841 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7842 new_min_free_kbytes, user_min_free_kbytes);
7844 setup_per_zone_wmarks();
7845 refresh_zone_stat_thresholds();
7846 setup_per_zone_lowmem_reserve();
7849 setup_min_unmapped_ratio();
7850 setup_min_slab_ratio();
7855 core_initcall(init_per_zone_wmark_min)
7858 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7859 * that we can call two helper functions whenever min_free_kbytes
7862 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7863 void __user *buffer, size_t *length, loff_t *ppos)
7867 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7872 user_min_free_kbytes = min_free_kbytes;
7873 setup_per_zone_wmarks();
7878 int watermark_boost_factor_sysctl_handler(struct ctl_table *table, int write,
7879 void __user *buffer, size_t *length, loff_t *ppos)
7883 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7890 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7891 void __user *buffer, size_t *length, loff_t *ppos)
7895 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7900 setup_per_zone_wmarks();
7906 static void setup_min_unmapped_ratio(void)
7911 for_each_online_pgdat(pgdat)
7912 pgdat->min_unmapped_pages = 0;
7915 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7916 sysctl_min_unmapped_ratio) / 100;
7920 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7921 void __user *buffer, size_t *length, loff_t *ppos)
7925 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7929 setup_min_unmapped_ratio();
7934 static void setup_min_slab_ratio(void)
7939 for_each_online_pgdat(pgdat)
7940 pgdat->min_slab_pages = 0;
7943 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7944 sysctl_min_slab_ratio) / 100;
7947 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7948 void __user *buffer, size_t *length, loff_t *ppos)
7952 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7956 setup_min_slab_ratio();
7963 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7964 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7965 * whenever sysctl_lowmem_reserve_ratio changes.
7967 * The reserve ratio obviously has absolutely no relation with the
7968 * minimum watermarks. The lowmem reserve ratio can only make sense
7969 * if in function of the boot time zone sizes.
7971 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7972 void __user *buffer, size_t *length, loff_t *ppos)
7974 proc_dointvec_minmax(table, write, buffer, length, ppos);
7975 setup_per_zone_lowmem_reserve();
7979 static void __zone_pcp_update(struct zone *zone)
7983 for_each_possible_cpu(cpu)
7984 pageset_set_high_and_batch(zone,
7985 per_cpu_ptr(zone->pageset, cpu));
7989 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7990 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7991 * pagelist can have before it gets flushed back to buddy allocator.
7993 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7994 void __user *buffer, size_t *length, loff_t *ppos)
7997 int old_percpu_pagelist_fraction;
8000 mutex_lock(&pcp_batch_high_lock);
8001 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8003 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8004 if (!write || ret < 0)
8007 /* Sanity checking to avoid pcp imbalance */
8008 if (percpu_pagelist_fraction &&
8009 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8010 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8016 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8019 for_each_populated_zone(zone)
8020 __zone_pcp_update(zone);
8022 mutex_unlock(&pcp_batch_high_lock);
8026 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8028 * Returns the number of pages that arch has reserved but
8029 * is not known to alloc_large_system_hash().
8031 static unsigned long __init arch_reserved_kernel_pages(void)
8038 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8039 * machines. As memory size is increased the scale is also increased but at
8040 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8041 * quadruples the scale is increased by one, which means the size of hash table
8042 * only doubles, instead of quadrupling as well.
8043 * Because 32-bit systems cannot have large physical memory, where this scaling
8044 * makes sense, it is disabled on such platforms.
8046 #if __BITS_PER_LONG > 32
8047 #define ADAPT_SCALE_BASE (64ul << 30)
8048 #define ADAPT_SCALE_SHIFT 2
8049 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8053 * allocate a large system hash table from bootmem
8054 * - it is assumed that the hash table must contain an exact power-of-2
8055 * quantity of entries
8056 * - limit is the number of hash buckets, not the total allocation size
8058 void *__init alloc_large_system_hash(const char *tablename,
8059 unsigned long bucketsize,
8060 unsigned long numentries,
8063 unsigned int *_hash_shift,
8064 unsigned int *_hash_mask,
8065 unsigned long low_limit,
8066 unsigned long high_limit)
8068 unsigned long long max = high_limit;
8069 unsigned long log2qty, size;
8074 /* allow the kernel cmdline to have a say */
8076 /* round applicable memory size up to nearest megabyte */
8077 numentries = nr_kernel_pages;
8078 numentries -= arch_reserved_kernel_pages();
8080 /* It isn't necessary when PAGE_SIZE >= 1MB */
8081 if (PAGE_SHIFT < 20)
8082 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8084 #if __BITS_PER_LONG > 32
8086 unsigned long adapt;
8088 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8089 adapt <<= ADAPT_SCALE_SHIFT)
8094 /* limit to 1 bucket per 2^scale bytes of low memory */
8095 if (scale > PAGE_SHIFT)
8096 numentries >>= (scale - PAGE_SHIFT);
8098 numentries <<= (PAGE_SHIFT - scale);
8100 /* Make sure we've got at least a 0-order allocation.. */
8101 if (unlikely(flags & HASH_SMALL)) {
8102 /* Makes no sense without HASH_EARLY */
8103 WARN_ON(!(flags & HASH_EARLY));
8104 if (!(numentries >> *_hash_shift)) {
8105 numentries = 1UL << *_hash_shift;
8106 BUG_ON(!numentries);
8108 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8109 numentries = PAGE_SIZE / bucketsize;
8111 numentries = roundup_pow_of_two(numentries);
8113 /* limit allocation size to 1/16 total memory by default */
8115 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8116 do_div(max, bucketsize);
8118 max = min(max, 0x80000000ULL);
8120 if (numentries < low_limit)
8121 numentries = low_limit;
8122 if (numentries > max)
8125 log2qty = ilog2(numentries);
8127 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8130 size = bucketsize << log2qty;
8131 if (flags & HASH_EARLY) {
8132 if (flags & HASH_ZERO)
8133 table = memblock_alloc(size, SMP_CACHE_BYTES);
8135 table = memblock_alloc_raw(size,
8137 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8138 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
8142 * If bucketsize is not a power-of-two, we may free
8143 * some pages at the end of hash table which
8144 * alloc_pages_exact() automatically does
8146 table = alloc_pages_exact(size, gfp_flags);
8147 kmemleak_alloc(table, size, 1, gfp_flags);
8149 } while (!table && size > PAGE_SIZE && --log2qty);
8152 panic("Failed to allocate %s hash table\n", tablename);
8154 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8155 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8156 virt ? "vmalloc" : "linear");
8159 *_hash_shift = log2qty;
8161 *_hash_mask = (1 << log2qty) - 1;
8167 * This function checks whether pageblock includes unmovable pages or not.
8168 * If @count is not zero, it is okay to include less @count unmovable pages
8170 * PageLRU check without isolation or lru_lock could race so that
8171 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8172 * check without lock_page also may miss some movable non-lru pages at
8173 * race condition. So you can't expect this function should be exact.
8175 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
8176 int migratetype, int flags)
8178 unsigned long found;
8179 unsigned long iter = 0;
8180 unsigned long pfn = page_to_pfn(page);
8181 const char *reason = "unmovable page";
8184 * TODO we could make this much more efficient by not checking every
8185 * page in the range if we know all of them are in MOVABLE_ZONE and
8186 * that the movable zone guarantees that pages are migratable but
8187 * the later is not the case right now unfortunatelly. E.g. movablecore
8188 * can still lead to having bootmem allocations in zone_movable.
8191 if (is_migrate_cma_page(page)) {
8193 * CMA allocations (alloc_contig_range) really need to mark
8194 * isolate CMA pageblocks even when they are not movable in fact
8195 * so consider them movable here.
8197 if (is_migrate_cma(migratetype))
8200 reason = "CMA page";
8204 for (found = 0; iter < pageblock_nr_pages; iter++) {
8205 unsigned long check = pfn + iter;
8207 if (!pfn_valid_within(check))
8210 page = pfn_to_page(check);
8212 if (PageReserved(page))
8216 * If the zone is movable and we have ruled out all reserved
8217 * pages then it should be reasonably safe to assume the rest
8220 if (zone_idx(zone) == ZONE_MOVABLE)
8224 * Hugepages are not in LRU lists, but they're movable.
8225 * We need not scan over tail pages because we don't
8226 * handle each tail page individually in migration.
8228 if (PageHuge(page)) {
8229 struct page *head = compound_head(page);
8230 unsigned int skip_pages;
8232 if (!hugepage_migration_supported(page_hstate(head)))
8235 skip_pages = compound_nr(head) - (page - head);
8236 iter += skip_pages - 1;
8241 * We can't use page_count without pin a page
8242 * because another CPU can free compound page.
8243 * This check already skips compound tails of THP
8244 * because their page->_refcount is zero at all time.
8246 if (!page_ref_count(page)) {
8247 if (PageBuddy(page))
8248 iter += (1 << page_order(page)) - 1;
8253 * The HWPoisoned page may be not in buddy system, and
8254 * page_count() is not 0.
8256 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8259 if (__PageMovable(page))
8265 * If there are RECLAIMABLE pages, we need to check
8266 * it. But now, memory offline itself doesn't call
8267 * shrink_node_slabs() and it still to be fixed.
8270 * If the page is not RAM, page_count()should be 0.
8271 * we don't need more check. This is an _used_ not-movable page.
8273 * The problematic thing here is PG_reserved pages. PG_reserved
8274 * is set to both of a memory hole page and a _used_ kernel
8282 WARN_ON_ONCE(zone_idx(zone) == ZONE_MOVABLE);
8283 if (flags & REPORT_FAILURE)
8284 dump_page(pfn_to_page(pfn + iter), reason);
8288 #ifdef CONFIG_CONTIG_ALLOC
8289 static unsigned long pfn_max_align_down(unsigned long pfn)
8291 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8292 pageblock_nr_pages) - 1);
8295 static unsigned long pfn_max_align_up(unsigned long pfn)
8297 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8298 pageblock_nr_pages));
8301 /* [start, end) must belong to a single zone. */
8302 static int __alloc_contig_migrate_range(struct compact_control *cc,
8303 unsigned long start, unsigned long end)
8305 /* This function is based on compact_zone() from compaction.c. */
8306 unsigned long nr_reclaimed;
8307 unsigned long pfn = start;
8308 unsigned int tries = 0;
8313 while (pfn < end || !list_empty(&cc->migratepages)) {
8314 if (fatal_signal_pending(current)) {
8319 if (list_empty(&cc->migratepages)) {
8320 cc->nr_migratepages = 0;
8321 pfn = isolate_migratepages_range(cc, pfn, end);
8327 } else if (++tries == 5) {
8328 ret = ret < 0 ? ret : -EBUSY;
8332 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8334 cc->nr_migratepages -= nr_reclaimed;
8336 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
8337 NULL, 0, cc->mode, MR_CONTIG_RANGE);
8340 putback_movable_pages(&cc->migratepages);
8347 * alloc_contig_range() -- tries to allocate given range of pages
8348 * @start: start PFN to allocate
8349 * @end: one-past-the-last PFN to allocate
8350 * @migratetype: migratetype of the underlaying pageblocks (either
8351 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8352 * in range must have the same migratetype and it must
8353 * be either of the two.
8354 * @gfp_mask: GFP mask to use during compaction
8356 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8357 * aligned. The PFN range must belong to a single zone.
8359 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8360 * pageblocks in the range. Once isolated, the pageblocks should not
8361 * be modified by others.
8363 * Return: zero on success or negative error code. On success all
8364 * pages which PFN is in [start, end) are allocated for the caller and
8365 * need to be freed with free_contig_range().
8367 int alloc_contig_range(unsigned long start, unsigned long end,
8368 unsigned migratetype, gfp_t gfp_mask)
8370 unsigned long outer_start, outer_end;
8374 struct compact_control cc = {
8375 .nr_migratepages = 0,
8377 .zone = page_zone(pfn_to_page(start)),
8378 .mode = MIGRATE_SYNC,
8379 .ignore_skip_hint = true,
8380 .no_set_skip_hint = true,
8381 .gfp_mask = current_gfp_context(gfp_mask),
8383 INIT_LIST_HEAD(&cc.migratepages);
8386 * What we do here is we mark all pageblocks in range as
8387 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8388 * have different sizes, and due to the way page allocator
8389 * work, we align the range to biggest of the two pages so
8390 * that page allocator won't try to merge buddies from
8391 * different pageblocks and change MIGRATE_ISOLATE to some
8392 * other migration type.
8394 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8395 * migrate the pages from an unaligned range (ie. pages that
8396 * we are interested in). This will put all the pages in
8397 * range back to page allocator as MIGRATE_ISOLATE.
8399 * When this is done, we take the pages in range from page
8400 * allocator removing them from the buddy system. This way
8401 * page allocator will never consider using them.
8403 * This lets us mark the pageblocks back as
8404 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8405 * aligned range but not in the unaligned, original range are
8406 * put back to page allocator so that buddy can use them.
8409 ret = start_isolate_page_range(pfn_max_align_down(start),
8410 pfn_max_align_up(end), migratetype, 0);
8415 * In case of -EBUSY, we'd like to know which page causes problem.
8416 * So, just fall through. test_pages_isolated() has a tracepoint
8417 * which will report the busy page.
8419 * It is possible that busy pages could become available before
8420 * the call to test_pages_isolated, and the range will actually be
8421 * allocated. So, if we fall through be sure to clear ret so that
8422 * -EBUSY is not accidentally used or returned to caller.
8424 ret = __alloc_contig_migrate_range(&cc, start, end);
8425 if (ret && ret != -EBUSY)
8430 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8431 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8432 * more, all pages in [start, end) are free in page allocator.
8433 * What we are going to do is to allocate all pages from
8434 * [start, end) (that is remove them from page allocator).
8436 * The only problem is that pages at the beginning and at the
8437 * end of interesting range may be not aligned with pages that
8438 * page allocator holds, ie. they can be part of higher order
8439 * pages. Because of this, we reserve the bigger range and
8440 * once this is done free the pages we are not interested in.
8442 * We don't have to hold zone->lock here because the pages are
8443 * isolated thus they won't get removed from buddy.
8446 lru_add_drain_all();
8449 outer_start = start;
8450 while (!PageBuddy(pfn_to_page(outer_start))) {
8451 if (++order >= MAX_ORDER) {
8452 outer_start = start;
8455 outer_start &= ~0UL << order;
8458 if (outer_start != start) {
8459 order = page_order(pfn_to_page(outer_start));
8462 * outer_start page could be small order buddy page and
8463 * it doesn't include start page. Adjust outer_start
8464 * in this case to report failed page properly
8465 * on tracepoint in test_pages_isolated()
8467 if (outer_start + (1UL << order) <= start)
8468 outer_start = start;
8471 /* Make sure the range is really isolated. */
8472 if (test_pages_isolated(outer_start, end, 0)) {
8473 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8474 __func__, outer_start, end);
8479 /* Grab isolated pages from freelists. */
8480 outer_end = isolate_freepages_range(&cc, outer_start, end);
8486 /* Free head and tail (if any) */
8487 if (start != outer_start)
8488 free_contig_range(outer_start, start - outer_start);
8489 if (end != outer_end)
8490 free_contig_range(end, outer_end - end);
8493 undo_isolate_page_range(pfn_max_align_down(start),
8494 pfn_max_align_up(end), migratetype);
8498 static int __alloc_contig_pages(unsigned long start_pfn,
8499 unsigned long nr_pages, gfp_t gfp_mask)
8501 unsigned long end_pfn = start_pfn + nr_pages;
8503 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8507 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8508 unsigned long nr_pages)
8510 unsigned long i, end_pfn = start_pfn + nr_pages;
8513 for (i = start_pfn; i < end_pfn; i++) {
8514 page = pfn_to_online_page(i);
8518 if (page_zone(page) != z)
8521 if (PageReserved(page))
8524 if (page_count(page) > 0)
8533 static bool zone_spans_last_pfn(const struct zone *zone,
8534 unsigned long start_pfn, unsigned long nr_pages)
8536 unsigned long last_pfn = start_pfn + nr_pages - 1;
8538 return zone_spans_pfn(zone, last_pfn);
8542 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8543 * @nr_pages: Number of contiguous pages to allocate
8544 * @gfp_mask: GFP mask to limit search and used during compaction
8546 * @nodemask: Mask for other possible nodes
8548 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8549 * on an applicable zonelist to find a contiguous pfn range which can then be
8550 * tried for allocation with alloc_contig_range(). This routine is intended
8551 * for allocation requests which can not be fulfilled with the buddy allocator.
8553 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8554 * power of two then the alignment is guaranteed to be to the given nr_pages
8555 * (e.g. 1GB request would be aligned to 1GB).
8557 * Allocated pages can be freed with free_contig_range() or by manually calling
8558 * __free_page() on each allocated page.
8560 * Return: pointer to contiguous pages on success, or NULL if not successful.
8562 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8563 int nid, nodemask_t *nodemask)
8565 unsigned long ret, pfn, flags;
8566 struct zonelist *zonelist;
8570 zonelist = node_zonelist(nid, gfp_mask);
8571 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8572 gfp_zone(gfp_mask), nodemask) {
8573 spin_lock_irqsave(&zone->lock, flags);
8575 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8576 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8577 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8579 * We release the zone lock here because
8580 * alloc_contig_range() will also lock the zone
8581 * at some point. If there's an allocation
8582 * spinning on this lock, it may win the race
8583 * and cause alloc_contig_range() to fail...
8585 spin_unlock_irqrestore(&zone->lock, flags);
8586 ret = __alloc_contig_pages(pfn, nr_pages,
8589 return pfn_to_page(pfn);
8590 spin_lock_irqsave(&zone->lock, flags);
8594 spin_unlock_irqrestore(&zone->lock, flags);
8598 #endif /* CONFIG_CONTIG_ALLOC */
8600 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8602 unsigned int count = 0;
8604 for (; nr_pages--; pfn++) {
8605 struct page *page = pfn_to_page(pfn);
8607 count += page_count(page) != 1;
8610 WARN(count != 0, "%d pages are still in use!\n", count);
8614 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8615 * page high values need to be recalulated.
8617 void __meminit zone_pcp_update(struct zone *zone)
8619 mutex_lock(&pcp_batch_high_lock);
8620 __zone_pcp_update(zone);
8621 mutex_unlock(&pcp_batch_high_lock);
8624 void zone_pcp_reset(struct zone *zone)
8626 unsigned long flags;
8628 struct per_cpu_pageset *pset;
8630 /* avoid races with drain_pages() */
8631 local_irq_save(flags);
8632 if (zone->pageset != &boot_pageset) {
8633 for_each_online_cpu(cpu) {
8634 pset = per_cpu_ptr(zone->pageset, cpu);
8635 drain_zonestat(zone, pset);
8637 free_percpu(zone->pageset);
8638 zone->pageset = &boot_pageset;
8640 local_irq_restore(flags);
8643 #ifdef CONFIG_MEMORY_HOTREMOVE
8645 * All pages in the range must be in a single zone and isolated
8646 * before calling this.
8649 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8655 unsigned long flags;
8656 unsigned long offlined_pages = 0;
8658 /* find the first valid pfn */
8659 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8663 return offlined_pages;
8665 offline_mem_sections(pfn, end_pfn);
8666 zone = page_zone(pfn_to_page(pfn));
8667 spin_lock_irqsave(&zone->lock, flags);
8669 while (pfn < end_pfn) {
8670 if (!pfn_valid(pfn)) {
8674 page = pfn_to_page(pfn);
8676 * The HWPoisoned page may be not in buddy system, and
8677 * page_count() is not 0.
8679 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8685 BUG_ON(page_count(page));
8686 BUG_ON(!PageBuddy(page));
8687 order = page_order(page);
8688 offlined_pages += 1 << order;
8689 #ifdef CONFIG_DEBUG_VM
8690 pr_info("remove from free list %lx %d %lx\n",
8691 pfn, 1 << order, end_pfn);
8693 del_page_from_free_area(page, &zone->free_area[order]);
8694 pfn += (1 << order);
8696 spin_unlock_irqrestore(&zone->lock, flags);
8698 return offlined_pages;
8702 bool is_free_buddy_page(struct page *page)
8704 struct zone *zone = page_zone(page);
8705 unsigned long pfn = page_to_pfn(page);
8706 unsigned long flags;
8709 spin_lock_irqsave(&zone->lock, flags);
8710 for (order = 0; order < MAX_ORDER; order++) {
8711 struct page *page_head = page - (pfn & ((1 << order) - 1));
8713 if (PageBuddy(page_head) && page_order(page_head) >= order)
8716 spin_unlock_irqrestore(&zone->lock, flags);
8718 return order < MAX_ORDER;
8721 #ifdef CONFIG_MEMORY_FAILURE
8723 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8724 * test is performed under the zone lock to prevent a race against page
8727 bool set_hwpoison_free_buddy_page(struct page *page)
8729 struct zone *zone = page_zone(page);
8730 unsigned long pfn = page_to_pfn(page);
8731 unsigned long flags;
8733 bool hwpoisoned = false;
8735 spin_lock_irqsave(&zone->lock, flags);
8736 for (order = 0; order < MAX_ORDER; order++) {
8737 struct page *page_head = page - (pfn & ((1 << order) - 1));
8739 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8740 if (!TestSetPageHWPoison(page))
8745 spin_unlock_irqrestore(&zone->lock, flags);