2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/highmem.h>
20 #include <linux/swap.h>
21 #include <linux/interrupt.h>
22 #include <linux/pagemap.h>
23 #include <linux/jiffies.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kasan.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/topology.h>
36 #include <linux/sysctl.h>
37 #include <linux/cpu.h>
38 #include <linux/cpuset.h>
39 #include <linux/memory_hotplug.h>
40 #include <linux/nodemask.h>
41 #include <linux/vmalloc.h>
42 #include <linux/vmstat.h>
43 #include <linux/mempolicy.h>
44 #include <linux/memremap.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_ext.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <trace/events/oom.h>
57 #include <linux/prefetch.h>
58 #include <linux/mm_inline.h>
59 #include <linux/migrate.h>
60 #include <linux/hugetlb.h>
61 #include <linux/sched/rt.h>
62 #include <linux/sched/mm.h>
63 #include <linux/page_owner.h>
64 #include <linux/kthread.h>
65 #include <linux/memcontrol.h>
66 #include <linux/ftrace.h>
67 #include <linux/lockdep.h>
68 #include <linux/nmi.h>
69 #include <linux/psi.h>
71 #include <asm/sections.h>
72 #include <asm/tlbflush.h>
73 #include <asm/div64.h>
76 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
77 static DEFINE_MUTEX(pcp_batch_high_lock);
78 #define MIN_PERCPU_PAGELIST_FRACTION (8)
80 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
81 DEFINE_PER_CPU(int, numa_node);
82 EXPORT_PER_CPU_SYMBOL(numa_node);
85 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
87 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
89 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
90 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
91 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
92 * defined in <linux/topology.h>.
94 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
95 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
96 int _node_numa_mem_[MAX_NUMNODES];
99 /* work_structs for global per-cpu drains */
102 struct work_struct work;
104 DEFINE_MUTEX(pcpu_drain_mutex);
105 DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
107 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
108 volatile unsigned long latent_entropy __latent_entropy;
109 EXPORT_SYMBOL(latent_entropy);
113 * Array of node states.
115 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
116 [N_POSSIBLE] = NODE_MASK_ALL,
117 [N_ONLINE] = { { [0] = 1UL } },
119 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
120 #ifdef CONFIG_HIGHMEM
121 [N_HIGH_MEMORY] = { { [0] = 1UL } },
123 [N_MEMORY] = { { [0] = 1UL } },
124 [N_CPU] = { { [0] = 1UL } },
127 EXPORT_SYMBOL(node_states);
129 atomic_long_t _totalram_pages __read_mostly;
130 EXPORT_SYMBOL(_totalram_pages);
131 unsigned long totalreserve_pages __read_mostly;
132 unsigned long totalcma_pages __read_mostly;
134 int percpu_pagelist_fraction;
135 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
138 * A cached value of the page's pageblock's migratetype, used when the page is
139 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
140 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
141 * Also the migratetype set in the page does not necessarily match the pcplist
142 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
143 * other index - this ensures that it will be put on the correct CMA freelist.
145 static inline int get_pcppage_migratetype(struct page *page)
150 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
152 page->index = migratetype;
155 #ifdef CONFIG_PM_SLEEP
157 * The following functions are used by the suspend/hibernate code to temporarily
158 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
159 * while devices are suspended. To avoid races with the suspend/hibernate code,
160 * they should always be called with system_transition_mutex held
161 * (gfp_allowed_mask also should only be modified with system_transition_mutex
162 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
163 * with that modification).
166 static gfp_t saved_gfp_mask;
168 void pm_restore_gfp_mask(void)
170 WARN_ON(!mutex_is_locked(&system_transition_mutex));
171 if (saved_gfp_mask) {
172 gfp_allowed_mask = saved_gfp_mask;
177 void pm_restrict_gfp_mask(void)
179 WARN_ON(!mutex_is_locked(&system_transition_mutex));
180 WARN_ON(saved_gfp_mask);
181 saved_gfp_mask = gfp_allowed_mask;
182 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
185 bool pm_suspended_storage(void)
187 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
191 #endif /* CONFIG_PM_SLEEP */
193 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
194 unsigned int pageblock_order __read_mostly;
197 static void __free_pages_ok(struct page *page, unsigned int order);
200 * results with 256, 32 in the lowmem_reserve sysctl:
201 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
202 * 1G machine -> (16M dma, 784M normal, 224M high)
203 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
204 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
205 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
207 * TBD: should special case ZONE_DMA32 machines here - in those we normally
208 * don't need any ZONE_NORMAL reservation
210 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
211 #ifdef CONFIG_ZONE_DMA
214 #ifdef CONFIG_ZONE_DMA32
218 #ifdef CONFIG_HIGHMEM
224 EXPORT_SYMBOL(totalram_pages);
226 static char * const zone_names[MAX_NR_ZONES] = {
227 #ifdef CONFIG_ZONE_DMA
230 #ifdef CONFIG_ZONE_DMA32
234 #ifdef CONFIG_HIGHMEM
238 #ifdef CONFIG_ZONE_DEVICE
243 const char * const migratetype_names[MIGRATE_TYPES] = {
251 #ifdef CONFIG_MEMORY_ISOLATION
256 compound_page_dtor * const compound_page_dtors[] = {
259 #ifdef CONFIG_HUGETLB_PAGE
262 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
267 int min_free_kbytes = 1024;
268 int user_min_free_kbytes = -1;
269 int watermark_boost_factor __read_mostly = 15000;
270 int watermark_scale_factor = 10;
272 static unsigned long nr_kernel_pages __initdata;
273 static unsigned long nr_all_pages __initdata;
274 static unsigned long dma_reserve __initdata;
276 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
277 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
278 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
279 static unsigned long required_kernelcore __initdata;
280 static unsigned long required_kernelcore_percent __initdata;
281 static unsigned long required_movablecore __initdata;
282 static unsigned long required_movablecore_percent __initdata;
283 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
284 static bool mirrored_kernelcore __meminitdata;
286 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
288 EXPORT_SYMBOL(movable_zone);
289 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
292 int nr_node_ids __read_mostly = MAX_NUMNODES;
293 int nr_online_nodes __read_mostly = 1;
294 EXPORT_SYMBOL(nr_node_ids);
295 EXPORT_SYMBOL(nr_online_nodes);
298 int page_group_by_mobility_disabled __read_mostly;
300 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
302 * During boot we initialize deferred pages on-demand, as needed, but once
303 * page_alloc_init_late() has finished, the deferred pages are all initialized,
304 * and we can permanently disable that path.
306 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
309 * Calling kasan_free_pages() only after deferred memory initialization
310 * has completed. Poisoning pages during deferred memory init will greatly
311 * lengthen the process and cause problem in large memory systems as the
312 * deferred pages initialization is done with interrupt disabled.
314 * Assuming that there will be no reference to those newly initialized
315 * pages before they are ever allocated, this should have no effect on
316 * KASAN memory tracking as the poison will be properly inserted at page
317 * allocation time. The only corner case is when pages are allocated by
318 * on-demand allocation and then freed again before the deferred pages
319 * initialization is done, but this is not likely to happen.
321 static inline void kasan_free_nondeferred_pages(struct page *page, int order)
323 if (!static_branch_unlikely(&deferred_pages))
324 kasan_free_pages(page, order);
327 /* Returns true if the struct page for the pfn is uninitialised */
328 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
330 int nid = early_pfn_to_nid(pfn);
332 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
339 * Returns true when the remaining initialisation should be deferred until
340 * later in the boot cycle when it can be parallelised.
342 static bool __meminit
343 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
345 static unsigned long prev_end_pfn, nr_initialised;
348 * prev_end_pfn static that contains the end of previous zone
349 * No need to protect because called very early in boot before smp_init.
351 if (prev_end_pfn != end_pfn) {
352 prev_end_pfn = end_pfn;
356 /* Always populate low zones for address-constrained allocations */
357 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
361 * We start only with one section of pages, more pages are added as
362 * needed until the rest of deferred pages are initialized.
365 if ((nr_initialised > PAGES_PER_SECTION) &&
366 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
367 NODE_DATA(nid)->first_deferred_pfn = pfn;
373 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
375 static inline bool early_page_uninitialised(unsigned long pfn)
380 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
386 /* Return a pointer to the bitmap storing bits affecting a block of pages */
387 static inline unsigned long *get_pageblock_bitmap(struct page *page,
390 #ifdef CONFIG_SPARSEMEM
391 return __pfn_to_section(pfn)->pageblock_flags;
393 return page_zone(page)->pageblock_flags;
394 #endif /* CONFIG_SPARSEMEM */
397 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
399 #ifdef CONFIG_SPARSEMEM
400 pfn &= (PAGES_PER_SECTION-1);
401 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
403 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
404 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
405 #endif /* CONFIG_SPARSEMEM */
409 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
410 * @page: The page within the block of interest
411 * @pfn: The target page frame number
412 * @end_bitidx: The last bit of interest to retrieve
413 * @mask: mask of bits that the caller is interested in
415 * Return: pageblock_bits flags
417 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
419 unsigned long end_bitidx,
422 unsigned long *bitmap;
423 unsigned long bitidx, word_bitidx;
426 bitmap = get_pageblock_bitmap(page, pfn);
427 bitidx = pfn_to_bitidx(page, pfn);
428 word_bitidx = bitidx / BITS_PER_LONG;
429 bitidx &= (BITS_PER_LONG-1);
431 word = bitmap[word_bitidx];
432 bitidx += end_bitidx;
433 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
436 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
437 unsigned long end_bitidx,
440 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
443 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
445 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
449 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
450 * @page: The page within the block of interest
451 * @flags: The flags to set
452 * @pfn: The target page frame number
453 * @end_bitidx: The last bit of interest
454 * @mask: mask of bits that the caller is interested in
456 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
458 unsigned long end_bitidx,
461 unsigned long *bitmap;
462 unsigned long bitidx, word_bitidx;
463 unsigned long old_word, word;
465 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
466 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
468 bitmap = get_pageblock_bitmap(page, pfn);
469 bitidx = pfn_to_bitidx(page, pfn);
470 word_bitidx = bitidx / BITS_PER_LONG;
471 bitidx &= (BITS_PER_LONG-1);
473 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
475 bitidx += end_bitidx;
476 mask <<= (BITS_PER_LONG - bitidx - 1);
477 flags <<= (BITS_PER_LONG - bitidx - 1);
479 word = READ_ONCE(bitmap[word_bitidx]);
481 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
482 if (word == old_word)
488 void set_pageblock_migratetype(struct page *page, int migratetype)
490 if (unlikely(page_group_by_mobility_disabled &&
491 migratetype < MIGRATE_PCPTYPES))
492 migratetype = MIGRATE_UNMOVABLE;
494 set_pageblock_flags_group(page, (unsigned long)migratetype,
495 PB_migrate, PB_migrate_end);
498 #ifdef CONFIG_DEBUG_VM
499 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
503 unsigned long pfn = page_to_pfn(page);
504 unsigned long sp, start_pfn;
507 seq = zone_span_seqbegin(zone);
508 start_pfn = zone->zone_start_pfn;
509 sp = zone->spanned_pages;
510 if (!zone_spans_pfn(zone, pfn))
512 } while (zone_span_seqretry(zone, seq));
515 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
516 pfn, zone_to_nid(zone), zone->name,
517 start_pfn, start_pfn + sp);
522 static int page_is_consistent(struct zone *zone, struct page *page)
524 if (!pfn_valid_within(page_to_pfn(page)))
526 if (zone != page_zone(page))
532 * Temporary debugging check for pages not lying within a given zone.
534 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
536 if (page_outside_zone_boundaries(zone, page))
538 if (!page_is_consistent(zone, page))
544 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
550 static void bad_page(struct page *page, const char *reason,
551 unsigned long bad_flags)
553 static unsigned long resume;
554 static unsigned long nr_shown;
555 static unsigned long nr_unshown;
558 * Allow a burst of 60 reports, then keep quiet for that minute;
559 * or allow a steady drip of one report per second.
561 if (nr_shown == 60) {
562 if (time_before(jiffies, resume)) {
568 "BUG: Bad page state: %lu messages suppressed\n",
575 resume = jiffies + 60 * HZ;
577 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
578 current->comm, page_to_pfn(page));
579 __dump_page(page, reason);
580 bad_flags &= page->flags;
582 pr_alert("bad because of flags: %#lx(%pGp)\n",
583 bad_flags, &bad_flags);
584 dump_page_owner(page);
589 /* Leave bad fields for debug, except PageBuddy could make trouble */
590 page_mapcount_reset(page); /* remove PageBuddy */
591 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
595 * Higher-order pages are called "compound pages". They are structured thusly:
597 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
599 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
600 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
602 * The first tail page's ->compound_dtor holds the offset in array of compound
603 * page destructors. See compound_page_dtors.
605 * The first tail page's ->compound_order holds the order of allocation.
606 * This usage means that zero-order pages may not be compound.
609 void free_compound_page(struct page *page)
611 __free_pages_ok(page, compound_order(page));
614 void prep_compound_page(struct page *page, unsigned int order)
617 int nr_pages = 1 << order;
619 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
620 set_compound_order(page, order);
622 for (i = 1; i < nr_pages; i++) {
623 struct page *p = page + i;
624 set_page_count(p, 0);
625 p->mapping = TAIL_MAPPING;
626 set_compound_head(p, page);
628 atomic_set(compound_mapcount_ptr(page), -1);
631 #ifdef CONFIG_DEBUG_PAGEALLOC
632 unsigned int _debug_guardpage_minorder;
633 bool _debug_pagealloc_enabled __read_mostly
634 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
635 EXPORT_SYMBOL(_debug_pagealloc_enabled);
636 bool _debug_guardpage_enabled __read_mostly;
638 static int __init early_debug_pagealloc(char *buf)
642 return kstrtobool(buf, &_debug_pagealloc_enabled);
644 early_param("debug_pagealloc", early_debug_pagealloc);
646 static bool need_debug_guardpage(void)
648 /* If we don't use debug_pagealloc, we don't need guard page */
649 if (!debug_pagealloc_enabled())
652 if (!debug_guardpage_minorder())
658 static void init_debug_guardpage(void)
660 if (!debug_pagealloc_enabled())
663 if (!debug_guardpage_minorder())
666 _debug_guardpage_enabled = true;
669 struct page_ext_operations debug_guardpage_ops = {
670 .need = need_debug_guardpage,
671 .init = init_debug_guardpage,
674 static int __init debug_guardpage_minorder_setup(char *buf)
678 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
679 pr_err("Bad debug_guardpage_minorder value\n");
682 _debug_guardpage_minorder = res;
683 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
686 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
688 static inline bool set_page_guard(struct zone *zone, struct page *page,
689 unsigned int order, int migratetype)
691 struct page_ext *page_ext;
693 if (!debug_guardpage_enabled())
696 if (order >= debug_guardpage_minorder())
699 page_ext = lookup_page_ext(page);
700 if (unlikely(!page_ext))
703 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
705 INIT_LIST_HEAD(&page->lru);
706 set_page_private(page, order);
707 /* Guard pages are not available for any usage */
708 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
713 static inline void clear_page_guard(struct zone *zone, struct page *page,
714 unsigned int order, int migratetype)
716 struct page_ext *page_ext;
718 if (!debug_guardpage_enabled())
721 page_ext = lookup_page_ext(page);
722 if (unlikely(!page_ext))
725 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
727 set_page_private(page, 0);
728 if (!is_migrate_isolate(migratetype))
729 __mod_zone_freepage_state(zone, (1 << order), migratetype);
732 struct page_ext_operations debug_guardpage_ops;
733 static inline bool set_page_guard(struct zone *zone, struct page *page,
734 unsigned int order, int migratetype) { return false; }
735 static inline void clear_page_guard(struct zone *zone, struct page *page,
736 unsigned int order, int migratetype) {}
739 static inline void set_page_order(struct page *page, unsigned int order)
741 set_page_private(page, order);
742 __SetPageBuddy(page);
745 static inline void rmv_page_order(struct page *page)
747 __ClearPageBuddy(page);
748 set_page_private(page, 0);
752 * This function checks whether a page is free && is the buddy
753 * we can coalesce a page and its buddy if
754 * (a) the buddy is not in a hole (check before calling!) &&
755 * (b) the buddy is in the buddy system &&
756 * (c) a page and its buddy have the same order &&
757 * (d) a page and its buddy are in the same zone.
759 * For recording whether a page is in the buddy system, we set PageBuddy.
760 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
762 * For recording page's order, we use page_private(page).
764 static inline int page_is_buddy(struct page *page, struct page *buddy,
767 if (page_is_guard(buddy) && page_order(buddy) == order) {
768 if (page_zone_id(page) != page_zone_id(buddy))
771 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
776 if (PageBuddy(buddy) && page_order(buddy) == order) {
778 * zone check is done late to avoid uselessly
779 * calculating zone/node ids for pages that could
782 if (page_zone_id(page) != page_zone_id(buddy))
785 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
793 * Freeing function for a buddy system allocator.
795 * The concept of a buddy system is to maintain direct-mapped table
796 * (containing bit values) for memory blocks of various "orders".
797 * The bottom level table contains the map for the smallest allocatable
798 * units of memory (here, pages), and each level above it describes
799 * pairs of units from the levels below, hence, "buddies".
800 * At a high level, all that happens here is marking the table entry
801 * at the bottom level available, and propagating the changes upward
802 * as necessary, plus some accounting needed to play nicely with other
803 * parts of the VM system.
804 * At each level, we keep a list of pages, which are heads of continuous
805 * free pages of length of (1 << order) and marked with PageBuddy.
806 * Page's order is recorded in page_private(page) field.
807 * So when we are allocating or freeing one, we can derive the state of the
808 * other. That is, if we allocate a small block, and both were
809 * free, the remainder of the region must be split into blocks.
810 * If a block is freed, and its buddy is also free, then this
811 * triggers coalescing into a block of larger size.
816 static inline void __free_one_page(struct page *page,
818 struct zone *zone, unsigned int order,
821 unsigned long combined_pfn;
822 unsigned long uninitialized_var(buddy_pfn);
824 unsigned int max_order;
826 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
828 VM_BUG_ON(!zone_is_initialized(zone));
829 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
831 VM_BUG_ON(migratetype == -1);
832 if (likely(!is_migrate_isolate(migratetype)))
833 __mod_zone_freepage_state(zone, 1 << order, migratetype);
835 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
836 VM_BUG_ON_PAGE(bad_range(zone, page), page);
839 while (order < max_order - 1) {
840 buddy_pfn = __find_buddy_pfn(pfn, order);
841 buddy = page + (buddy_pfn - pfn);
843 if (!pfn_valid_within(buddy_pfn))
845 if (!page_is_buddy(page, buddy, order))
848 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
849 * merge with it and move up one order.
851 if (page_is_guard(buddy)) {
852 clear_page_guard(zone, buddy, order, migratetype);
854 list_del(&buddy->lru);
855 zone->free_area[order].nr_free--;
856 rmv_page_order(buddy);
858 combined_pfn = buddy_pfn & pfn;
859 page = page + (combined_pfn - pfn);
863 if (max_order < MAX_ORDER) {
864 /* If we are here, it means order is >= pageblock_order.
865 * We want to prevent merge between freepages on isolate
866 * pageblock and normal pageblock. Without this, pageblock
867 * isolation could cause incorrect freepage or CMA accounting.
869 * We don't want to hit this code for the more frequent
872 if (unlikely(has_isolate_pageblock(zone))) {
875 buddy_pfn = __find_buddy_pfn(pfn, order);
876 buddy = page + (buddy_pfn - pfn);
877 buddy_mt = get_pageblock_migratetype(buddy);
879 if (migratetype != buddy_mt
880 && (is_migrate_isolate(migratetype) ||
881 is_migrate_isolate(buddy_mt)))
885 goto continue_merging;
889 set_page_order(page, order);
892 * If this is not the largest possible page, check if the buddy
893 * of the next-highest order is free. If it is, it's possible
894 * that pages are being freed that will coalesce soon. In case,
895 * that is happening, add the free page to the tail of the list
896 * so it's less likely to be used soon and more likely to be merged
897 * as a higher order page
899 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
900 struct page *higher_page, *higher_buddy;
901 combined_pfn = buddy_pfn & pfn;
902 higher_page = page + (combined_pfn - pfn);
903 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
904 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
905 if (pfn_valid_within(buddy_pfn) &&
906 page_is_buddy(higher_page, higher_buddy, order + 1)) {
907 list_add_tail(&page->lru,
908 &zone->free_area[order].free_list[migratetype]);
913 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
915 zone->free_area[order].nr_free++;
919 * A bad page could be due to a number of fields. Instead of multiple branches,
920 * try and check multiple fields with one check. The caller must do a detailed
921 * check if necessary.
923 static inline bool page_expected_state(struct page *page,
924 unsigned long check_flags)
926 if (unlikely(atomic_read(&page->_mapcount) != -1))
929 if (unlikely((unsigned long)page->mapping |
930 page_ref_count(page) |
932 (unsigned long)page->mem_cgroup |
934 (page->flags & check_flags)))
940 static void free_pages_check_bad(struct page *page)
942 const char *bad_reason;
943 unsigned long bad_flags;
948 if (unlikely(atomic_read(&page->_mapcount) != -1))
949 bad_reason = "nonzero mapcount";
950 if (unlikely(page->mapping != NULL))
951 bad_reason = "non-NULL mapping";
952 if (unlikely(page_ref_count(page) != 0))
953 bad_reason = "nonzero _refcount";
954 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
955 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
956 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
959 if (unlikely(page->mem_cgroup))
960 bad_reason = "page still charged to cgroup";
962 bad_page(page, bad_reason, bad_flags);
965 static inline int free_pages_check(struct page *page)
967 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
970 /* Something has gone sideways, find it */
971 free_pages_check_bad(page);
975 static int free_tail_pages_check(struct page *head_page, struct page *page)
980 * We rely page->lru.next never has bit 0 set, unless the page
981 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
983 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
985 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
989 switch (page - head_page) {
991 /* the first tail page: ->mapping may be compound_mapcount() */
992 if (unlikely(compound_mapcount(page))) {
993 bad_page(page, "nonzero compound_mapcount", 0);
999 * the second tail page: ->mapping is
1000 * deferred_list.next -- ignore value.
1004 if (page->mapping != TAIL_MAPPING) {
1005 bad_page(page, "corrupted mapping in tail page", 0);
1010 if (unlikely(!PageTail(page))) {
1011 bad_page(page, "PageTail not set", 0);
1014 if (unlikely(compound_head(page) != head_page)) {
1015 bad_page(page, "compound_head not consistent", 0);
1020 page->mapping = NULL;
1021 clear_compound_head(page);
1025 static __always_inline bool free_pages_prepare(struct page *page,
1026 unsigned int order, bool check_free)
1030 VM_BUG_ON_PAGE(PageTail(page), page);
1032 trace_mm_page_free(page, order);
1035 * Check tail pages before head page information is cleared to
1036 * avoid checking PageCompound for order-0 pages.
1038 if (unlikely(order)) {
1039 bool compound = PageCompound(page);
1042 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1045 ClearPageDoubleMap(page);
1046 for (i = 1; i < (1 << order); i++) {
1048 bad += free_tail_pages_check(page, page + i);
1049 if (unlikely(free_pages_check(page + i))) {
1053 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1056 if (PageMappingFlags(page))
1057 page->mapping = NULL;
1058 if (memcg_kmem_enabled() && PageKmemcg(page))
1059 memcg_kmem_uncharge(page, order);
1061 bad += free_pages_check(page);
1065 page_cpupid_reset_last(page);
1066 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1067 reset_page_owner(page, order);
1069 if (!PageHighMem(page)) {
1070 debug_check_no_locks_freed(page_address(page),
1071 PAGE_SIZE << order);
1072 debug_check_no_obj_freed(page_address(page),
1073 PAGE_SIZE << order);
1075 arch_free_page(page, order);
1076 kernel_poison_pages(page, 1 << order, 0);
1077 kernel_map_pages(page, 1 << order, 0);
1078 kasan_free_nondeferred_pages(page, order);
1083 #ifdef CONFIG_DEBUG_VM
1084 static inline bool free_pcp_prepare(struct page *page)
1086 return free_pages_prepare(page, 0, true);
1089 static inline bool bulkfree_pcp_prepare(struct page *page)
1094 static bool free_pcp_prepare(struct page *page)
1096 return free_pages_prepare(page, 0, false);
1099 static bool bulkfree_pcp_prepare(struct page *page)
1101 return free_pages_check(page);
1103 #endif /* CONFIG_DEBUG_VM */
1105 static inline void prefetch_buddy(struct page *page)
1107 unsigned long pfn = page_to_pfn(page);
1108 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1109 struct page *buddy = page + (buddy_pfn - pfn);
1115 * Frees a number of pages from the PCP lists
1116 * Assumes all pages on list are in same zone, and of same order.
1117 * count is the number of pages to free.
1119 * If the zone was previously in an "all pages pinned" state then look to
1120 * see if this freeing clears that state.
1122 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1123 * pinned" detection logic.
1125 static void free_pcppages_bulk(struct zone *zone, int count,
1126 struct per_cpu_pages *pcp)
1128 int migratetype = 0;
1130 int prefetch_nr = 0;
1131 bool isolated_pageblocks;
1132 struct page *page, *tmp;
1136 struct list_head *list;
1139 * Remove pages from lists in a round-robin fashion. A
1140 * batch_free count is maintained that is incremented when an
1141 * empty list is encountered. This is so more pages are freed
1142 * off fuller lists instead of spinning excessively around empty
1147 if (++migratetype == MIGRATE_PCPTYPES)
1149 list = &pcp->lists[migratetype];
1150 } while (list_empty(list));
1152 /* This is the only non-empty list. Free them all. */
1153 if (batch_free == MIGRATE_PCPTYPES)
1157 page = list_last_entry(list, struct page, lru);
1158 /* must delete to avoid corrupting pcp list */
1159 list_del(&page->lru);
1162 if (bulkfree_pcp_prepare(page))
1165 list_add_tail(&page->lru, &head);
1168 * We are going to put the page back to the global
1169 * pool, prefetch its buddy to speed up later access
1170 * under zone->lock. It is believed the overhead of
1171 * an additional test and calculating buddy_pfn here
1172 * can be offset by reduced memory latency later. To
1173 * avoid excessive prefetching due to large count, only
1174 * prefetch buddy for the first pcp->batch nr of pages.
1176 if (prefetch_nr++ < pcp->batch)
1177 prefetch_buddy(page);
1178 } while (--count && --batch_free && !list_empty(list));
1181 spin_lock(&zone->lock);
1182 isolated_pageblocks = has_isolate_pageblock(zone);
1185 * Use safe version since after __free_one_page(),
1186 * page->lru.next will not point to original list.
1188 list_for_each_entry_safe(page, tmp, &head, lru) {
1189 int mt = get_pcppage_migratetype(page);
1190 /* MIGRATE_ISOLATE page should not go to pcplists */
1191 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1192 /* Pageblock could have been isolated meanwhile */
1193 if (unlikely(isolated_pageblocks))
1194 mt = get_pageblock_migratetype(page);
1196 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1197 trace_mm_page_pcpu_drain(page, 0, mt);
1199 spin_unlock(&zone->lock);
1202 static void free_one_page(struct zone *zone,
1203 struct page *page, unsigned long pfn,
1207 spin_lock(&zone->lock);
1208 if (unlikely(has_isolate_pageblock(zone) ||
1209 is_migrate_isolate(migratetype))) {
1210 migratetype = get_pfnblock_migratetype(page, pfn);
1212 __free_one_page(page, pfn, zone, order, migratetype);
1213 spin_unlock(&zone->lock);
1216 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1217 unsigned long zone, int nid)
1219 mm_zero_struct_page(page);
1220 set_page_links(page, zone, nid, pfn);
1221 init_page_count(page);
1222 page_mapcount_reset(page);
1223 page_cpupid_reset_last(page);
1224 page_kasan_tag_reset(page);
1226 INIT_LIST_HEAD(&page->lru);
1227 #ifdef WANT_PAGE_VIRTUAL
1228 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1229 if (!is_highmem_idx(zone))
1230 set_page_address(page, __va(pfn << PAGE_SHIFT));
1234 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1235 static void __meminit init_reserved_page(unsigned long pfn)
1240 if (!early_page_uninitialised(pfn))
1243 nid = early_pfn_to_nid(pfn);
1244 pgdat = NODE_DATA(nid);
1246 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1247 struct zone *zone = &pgdat->node_zones[zid];
1249 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1252 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1255 static inline void init_reserved_page(unsigned long pfn)
1258 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1261 * Initialised pages do not have PageReserved set. This function is
1262 * called for each range allocated by the bootmem allocator and
1263 * marks the pages PageReserved. The remaining valid pages are later
1264 * sent to the buddy page allocator.
1266 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1268 unsigned long start_pfn = PFN_DOWN(start);
1269 unsigned long end_pfn = PFN_UP(end);
1271 for (; start_pfn < end_pfn; start_pfn++) {
1272 if (pfn_valid(start_pfn)) {
1273 struct page *page = pfn_to_page(start_pfn);
1275 init_reserved_page(start_pfn);
1277 /* Avoid false-positive PageTail() */
1278 INIT_LIST_HEAD(&page->lru);
1281 * no need for atomic set_bit because the struct
1282 * page is not visible yet so nobody should
1285 __SetPageReserved(page);
1290 static void __free_pages_ok(struct page *page, unsigned int order)
1292 unsigned long flags;
1294 unsigned long pfn = page_to_pfn(page);
1296 if (!free_pages_prepare(page, order, true))
1299 migratetype = get_pfnblock_migratetype(page, pfn);
1300 local_irq_save(flags);
1301 __count_vm_events(PGFREE, 1 << order);
1302 free_one_page(page_zone(page), page, pfn, order, migratetype);
1303 local_irq_restore(flags);
1306 void __free_pages_core(struct page *page, unsigned int order)
1308 unsigned int nr_pages = 1 << order;
1309 struct page *p = page;
1313 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1315 __ClearPageReserved(p);
1316 set_page_count(p, 0);
1318 __ClearPageReserved(p);
1319 set_page_count(p, 0);
1321 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1322 set_page_refcounted(page);
1323 __free_pages(page, order);
1326 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1327 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1329 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1331 int __meminit early_pfn_to_nid(unsigned long pfn)
1333 static DEFINE_SPINLOCK(early_pfn_lock);
1336 spin_lock(&early_pfn_lock);
1337 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1339 nid = first_online_node;
1340 spin_unlock(&early_pfn_lock);
1346 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1347 static inline bool __meminit __maybe_unused
1348 meminit_pfn_in_nid(unsigned long pfn, int node,
1349 struct mminit_pfnnid_cache *state)
1353 nid = __early_pfn_to_nid(pfn, state);
1354 if (nid >= 0 && nid != node)
1359 /* Only safe to use early in boot when initialisation is single-threaded */
1360 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1362 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1367 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1371 static inline bool __meminit __maybe_unused
1372 meminit_pfn_in_nid(unsigned long pfn, int node,
1373 struct mminit_pfnnid_cache *state)
1380 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1383 if (early_page_uninitialised(pfn))
1385 __free_pages_core(page, order);
1389 * Check that the whole (or subset of) a pageblock given by the interval of
1390 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1391 * with the migration of free compaction scanner. The scanners then need to
1392 * use only pfn_valid_within() check for arches that allow holes within
1395 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1397 * It's possible on some configurations to have a setup like node0 node1 node0
1398 * i.e. it's possible that all pages within a zones range of pages do not
1399 * belong to a single zone. We assume that a border between node0 and node1
1400 * can occur within a single pageblock, but not a node0 node1 node0
1401 * interleaving within a single pageblock. It is therefore sufficient to check
1402 * the first and last page of a pageblock and avoid checking each individual
1403 * page in a pageblock.
1405 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1406 unsigned long end_pfn, struct zone *zone)
1408 struct page *start_page;
1409 struct page *end_page;
1411 /* end_pfn is one past the range we are checking */
1414 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1417 start_page = pfn_to_online_page(start_pfn);
1421 if (page_zone(start_page) != zone)
1424 end_page = pfn_to_page(end_pfn);
1426 /* This gives a shorter code than deriving page_zone(end_page) */
1427 if (page_zone_id(start_page) != page_zone_id(end_page))
1433 void set_zone_contiguous(struct zone *zone)
1435 unsigned long block_start_pfn = zone->zone_start_pfn;
1436 unsigned long block_end_pfn;
1438 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1439 for (; block_start_pfn < zone_end_pfn(zone);
1440 block_start_pfn = block_end_pfn,
1441 block_end_pfn += pageblock_nr_pages) {
1443 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1445 if (!__pageblock_pfn_to_page(block_start_pfn,
1446 block_end_pfn, zone))
1450 /* We confirm that there is no hole */
1451 zone->contiguous = true;
1454 void clear_zone_contiguous(struct zone *zone)
1456 zone->contiguous = false;
1459 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1460 static void __init deferred_free_range(unsigned long pfn,
1461 unsigned long nr_pages)
1469 page = pfn_to_page(pfn);
1471 /* Free a large naturally-aligned chunk if possible */
1472 if (nr_pages == pageblock_nr_pages &&
1473 (pfn & (pageblock_nr_pages - 1)) == 0) {
1474 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1475 __free_pages_core(page, pageblock_order);
1479 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1480 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1481 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1482 __free_pages_core(page, 0);
1486 /* Completion tracking for deferred_init_memmap() threads */
1487 static atomic_t pgdat_init_n_undone __initdata;
1488 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1490 static inline void __init pgdat_init_report_one_done(void)
1492 if (atomic_dec_and_test(&pgdat_init_n_undone))
1493 complete(&pgdat_init_all_done_comp);
1497 * Returns true if page needs to be initialized or freed to buddy allocator.
1499 * First we check if pfn is valid on architectures where it is possible to have
1500 * holes within pageblock_nr_pages. On systems where it is not possible, this
1501 * function is optimized out.
1503 * Then, we check if a current large page is valid by only checking the validity
1506 * Finally, meminit_pfn_in_nid is checked on systems where pfns can interleave
1507 * within a node: a pfn is between start and end of a node, but does not belong
1508 * to this memory node.
1510 static inline bool __init
1511 deferred_pfn_valid(int nid, unsigned long pfn,
1512 struct mminit_pfnnid_cache *nid_init_state)
1514 if (!pfn_valid_within(pfn))
1516 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1518 if (!meminit_pfn_in_nid(pfn, nid, nid_init_state))
1524 * Free pages to buddy allocator. Try to free aligned pages in
1525 * pageblock_nr_pages sizes.
1527 static void __init deferred_free_pages(int nid, int zid, unsigned long pfn,
1528 unsigned long end_pfn)
1530 struct mminit_pfnnid_cache nid_init_state = { };
1531 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1532 unsigned long nr_free = 0;
1534 for (; pfn < end_pfn; pfn++) {
1535 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1536 deferred_free_range(pfn - nr_free, nr_free);
1538 } else if (!(pfn & nr_pgmask)) {
1539 deferred_free_range(pfn - nr_free, nr_free);
1541 touch_nmi_watchdog();
1546 /* Free the last block of pages to allocator */
1547 deferred_free_range(pfn - nr_free, nr_free);
1551 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1552 * by performing it only once every pageblock_nr_pages.
1553 * Return number of pages initialized.
1555 static unsigned long __init deferred_init_pages(int nid, int zid,
1557 unsigned long end_pfn)
1559 struct mminit_pfnnid_cache nid_init_state = { };
1560 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1561 unsigned long nr_pages = 0;
1562 struct page *page = NULL;
1564 for (; pfn < end_pfn; pfn++) {
1565 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1568 } else if (!page || !(pfn & nr_pgmask)) {
1569 page = pfn_to_page(pfn);
1570 touch_nmi_watchdog();
1574 __init_single_page(page, pfn, zid, nid);
1580 /* Initialise remaining memory on a node */
1581 static int __init deferred_init_memmap(void *data)
1583 pg_data_t *pgdat = data;
1584 int nid = pgdat->node_id;
1585 unsigned long start = jiffies;
1586 unsigned long nr_pages = 0;
1587 unsigned long spfn, epfn, first_init_pfn, flags;
1588 phys_addr_t spa, epa;
1591 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1594 /* Bind memory initialisation thread to a local node if possible */
1595 if (!cpumask_empty(cpumask))
1596 set_cpus_allowed_ptr(current, cpumask);
1598 pgdat_resize_lock(pgdat, &flags);
1599 first_init_pfn = pgdat->first_deferred_pfn;
1600 if (first_init_pfn == ULONG_MAX) {
1601 pgdat_resize_unlock(pgdat, &flags);
1602 pgdat_init_report_one_done();
1606 /* Sanity check boundaries */
1607 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1608 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1609 pgdat->first_deferred_pfn = ULONG_MAX;
1611 /* Only the highest zone is deferred so find it */
1612 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1613 zone = pgdat->node_zones + zid;
1614 if (first_init_pfn < zone_end_pfn(zone))
1617 first_init_pfn = max(zone->zone_start_pfn, first_init_pfn);
1620 * Initialize and free pages. We do it in two loops: first we initialize
1621 * struct page, than free to buddy allocator, because while we are
1622 * freeing pages we can access pages that are ahead (computing buddy
1623 * page in __free_one_page()).
1625 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1626 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1627 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1628 nr_pages += deferred_init_pages(nid, zid, spfn, epfn);
1630 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1631 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1632 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1633 deferred_free_pages(nid, zid, spfn, epfn);
1635 pgdat_resize_unlock(pgdat, &flags);
1637 /* Sanity check that the next zone really is unpopulated */
1638 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1640 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1641 jiffies_to_msecs(jiffies - start));
1643 pgdat_init_report_one_done();
1648 * If this zone has deferred pages, try to grow it by initializing enough
1649 * deferred pages to satisfy the allocation specified by order, rounded up to
1650 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1651 * of SECTION_SIZE bytes by initializing struct pages in increments of
1652 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1654 * Return true when zone was grown, otherwise return false. We return true even
1655 * when we grow less than requested, to let the caller decide if there are
1656 * enough pages to satisfy the allocation.
1658 * Note: We use noinline because this function is needed only during boot, and
1659 * it is called from a __ref function _deferred_grow_zone. This way we are
1660 * making sure that it is not inlined into permanent text section.
1662 static noinline bool __init
1663 deferred_grow_zone(struct zone *zone, unsigned int order)
1665 int zid = zone_idx(zone);
1666 int nid = zone_to_nid(zone);
1667 pg_data_t *pgdat = NODE_DATA(nid);
1668 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1669 unsigned long nr_pages = 0;
1670 unsigned long first_init_pfn, spfn, epfn, t, flags;
1671 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1672 phys_addr_t spa, epa;
1675 /* Only the last zone may have deferred pages */
1676 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1679 pgdat_resize_lock(pgdat, &flags);
1682 * If deferred pages have been initialized while we were waiting for
1683 * the lock, return true, as the zone was grown. The caller will retry
1684 * this zone. We won't return to this function since the caller also
1685 * has this static branch.
1687 if (!static_branch_unlikely(&deferred_pages)) {
1688 pgdat_resize_unlock(pgdat, &flags);
1693 * If someone grew this zone while we were waiting for spinlock, return
1694 * true, as there might be enough pages already.
1696 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1697 pgdat_resize_unlock(pgdat, &flags);
1701 first_init_pfn = max(zone->zone_start_pfn, first_deferred_pfn);
1703 if (first_init_pfn >= pgdat_end_pfn(pgdat)) {
1704 pgdat_resize_unlock(pgdat, &flags);
1708 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1709 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1710 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1712 while (spfn < epfn && nr_pages < nr_pages_needed) {
1713 t = ALIGN(spfn + PAGES_PER_SECTION, PAGES_PER_SECTION);
1714 first_deferred_pfn = min(t, epfn);
1715 nr_pages += deferred_init_pages(nid, zid, spfn,
1716 first_deferred_pfn);
1717 spfn = first_deferred_pfn;
1720 if (nr_pages >= nr_pages_needed)
1724 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1725 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1726 epfn = min_t(unsigned long, first_deferred_pfn, PFN_DOWN(epa));
1727 deferred_free_pages(nid, zid, spfn, epfn);
1729 if (first_deferred_pfn == epfn)
1732 pgdat->first_deferred_pfn = first_deferred_pfn;
1733 pgdat_resize_unlock(pgdat, &flags);
1735 return nr_pages > 0;
1739 * deferred_grow_zone() is __init, but it is called from
1740 * get_page_from_freelist() during early boot until deferred_pages permanently
1741 * disables this call. This is why we have refdata wrapper to avoid warning,
1742 * and to ensure that the function body gets unloaded.
1745 _deferred_grow_zone(struct zone *zone, unsigned int order)
1747 return deferred_grow_zone(zone, order);
1750 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1752 void __init page_alloc_init_late(void)
1756 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1759 /* There will be num_node_state(N_MEMORY) threads */
1760 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1761 for_each_node_state(nid, N_MEMORY) {
1762 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1765 /* Block until all are initialised */
1766 wait_for_completion(&pgdat_init_all_done_comp);
1769 * We initialized the rest of the deferred pages. Permanently disable
1770 * on-demand struct page initialization.
1772 static_branch_disable(&deferred_pages);
1774 /* Reinit limits that are based on free pages after the kernel is up */
1775 files_maxfiles_init();
1777 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1778 /* Discard memblock private memory */
1782 for_each_populated_zone(zone)
1783 set_zone_contiguous(zone);
1787 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1788 void __init init_cma_reserved_pageblock(struct page *page)
1790 unsigned i = pageblock_nr_pages;
1791 struct page *p = page;
1794 __ClearPageReserved(p);
1795 set_page_count(p, 0);
1798 set_pageblock_migratetype(page, MIGRATE_CMA);
1800 if (pageblock_order >= MAX_ORDER) {
1801 i = pageblock_nr_pages;
1804 set_page_refcounted(p);
1805 __free_pages(p, MAX_ORDER - 1);
1806 p += MAX_ORDER_NR_PAGES;
1807 } while (i -= MAX_ORDER_NR_PAGES);
1809 set_page_refcounted(page);
1810 __free_pages(page, pageblock_order);
1813 adjust_managed_page_count(page, pageblock_nr_pages);
1818 * The order of subdivision here is critical for the IO subsystem.
1819 * Please do not alter this order without good reasons and regression
1820 * testing. Specifically, as large blocks of memory are subdivided,
1821 * the order in which smaller blocks are delivered depends on the order
1822 * they're subdivided in this function. This is the primary factor
1823 * influencing the order in which pages are delivered to the IO
1824 * subsystem according to empirical testing, and this is also justified
1825 * by considering the behavior of a buddy system containing a single
1826 * large block of memory acted on by a series of small allocations.
1827 * This behavior is a critical factor in sglist merging's success.
1831 static inline void expand(struct zone *zone, struct page *page,
1832 int low, int high, struct free_area *area,
1835 unsigned long size = 1 << high;
1837 while (high > low) {
1841 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1844 * Mark as guard pages (or page), that will allow to
1845 * merge back to allocator when buddy will be freed.
1846 * Corresponding page table entries will not be touched,
1847 * pages will stay not present in virtual address space
1849 if (set_page_guard(zone, &page[size], high, migratetype))
1852 list_add(&page[size].lru, &area->free_list[migratetype]);
1854 set_page_order(&page[size], high);
1858 static void check_new_page_bad(struct page *page)
1860 const char *bad_reason = NULL;
1861 unsigned long bad_flags = 0;
1863 if (unlikely(atomic_read(&page->_mapcount) != -1))
1864 bad_reason = "nonzero mapcount";
1865 if (unlikely(page->mapping != NULL))
1866 bad_reason = "non-NULL mapping";
1867 if (unlikely(page_ref_count(page) != 0))
1868 bad_reason = "nonzero _count";
1869 if (unlikely(page->flags & __PG_HWPOISON)) {
1870 bad_reason = "HWPoisoned (hardware-corrupted)";
1871 bad_flags = __PG_HWPOISON;
1872 /* Don't complain about hwpoisoned pages */
1873 page_mapcount_reset(page); /* remove PageBuddy */
1876 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1877 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1878 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1881 if (unlikely(page->mem_cgroup))
1882 bad_reason = "page still charged to cgroup";
1884 bad_page(page, bad_reason, bad_flags);
1888 * This page is about to be returned from the page allocator
1890 static inline int check_new_page(struct page *page)
1892 if (likely(page_expected_state(page,
1893 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1896 check_new_page_bad(page);
1900 static inline bool free_pages_prezeroed(void)
1902 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1903 page_poisoning_enabled();
1906 #ifdef CONFIG_DEBUG_VM
1907 static bool check_pcp_refill(struct page *page)
1912 static bool check_new_pcp(struct page *page)
1914 return check_new_page(page);
1917 static bool check_pcp_refill(struct page *page)
1919 return check_new_page(page);
1921 static bool check_new_pcp(struct page *page)
1925 #endif /* CONFIG_DEBUG_VM */
1927 static bool check_new_pages(struct page *page, unsigned int order)
1930 for (i = 0; i < (1 << order); i++) {
1931 struct page *p = page + i;
1933 if (unlikely(check_new_page(p)))
1940 inline void post_alloc_hook(struct page *page, unsigned int order,
1943 set_page_private(page, 0);
1944 set_page_refcounted(page);
1946 arch_alloc_page(page, order);
1947 kernel_map_pages(page, 1 << order, 1);
1948 kasan_alloc_pages(page, order);
1949 kernel_poison_pages(page, 1 << order, 1);
1950 set_page_owner(page, order, gfp_flags);
1953 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1954 unsigned int alloc_flags)
1958 post_alloc_hook(page, order, gfp_flags);
1960 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
1961 for (i = 0; i < (1 << order); i++)
1962 clear_highpage(page + i);
1964 if (order && (gfp_flags & __GFP_COMP))
1965 prep_compound_page(page, order);
1968 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1969 * allocate the page. The expectation is that the caller is taking
1970 * steps that will free more memory. The caller should avoid the page
1971 * being used for !PFMEMALLOC purposes.
1973 if (alloc_flags & ALLOC_NO_WATERMARKS)
1974 set_page_pfmemalloc(page);
1976 clear_page_pfmemalloc(page);
1980 * Go through the free lists for the given migratetype and remove
1981 * the smallest available page from the freelists
1983 static __always_inline
1984 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1987 unsigned int current_order;
1988 struct free_area *area;
1991 /* Find a page of the appropriate size in the preferred list */
1992 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1993 area = &(zone->free_area[current_order]);
1994 page = list_first_entry_or_null(&area->free_list[migratetype],
1998 list_del(&page->lru);
1999 rmv_page_order(page);
2001 expand(zone, page, order, current_order, area, migratetype);
2002 set_pcppage_migratetype(page, migratetype);
2011 * This array describes the order lists are fallen back to when
2012 * the free lists for the desirable migrate type are depleted
2014 static int fallbacks[MIGRATE_TYPES][4] = {
2015 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2016 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2017 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2019 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2021 #ifdef CONFIG_MEMORY_ISOLATION
2022 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2027 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2030 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2033 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2034 unsigned int order) { return NULL; }
2038 * Move the free pages in a range to the free lists of the requested type.
2039 * Note that start_page and end_pages are not aligned on a pageblock
2040 * boundary. If alignment is required, use move_freepages_block()
2042 static int move_freepages(struct zone *zone,
2043 struct page *start_page, struct page *end_page,
2044 int migratetype, int *num_movable)
2048 int pages_moved = 0;
2050 #ifndef CONFIG_HOLES_IN_ZONE
2052 * page_zone is not safe to call in this context when
2053 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
2054 * anyway as we check zone boundaries in move_freepages_block().
2055 * Remove at a later date when no bug reports exist related to
2056 * grouping pages by mobility
2058 VM_BUG_ON(pfn_valid(page_to_pfn(start_page)) &&
2059 pfn_valid(page_to_pfn(end_page)) &&
2060 page_zone(start_page) != page_zone(end_page));
2062 for (page = start_page; page <= end_page;) {
2063 if (!pfn_valid_within(page_to_pfn(page))) {
2068 /* Make sure we are not inadvertently changing nodes */
2069 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2071 if (!PageBuddy(page)) {
2073 * We assume that pages that could be isolated for
2074 * migration are movable. But we don't actually try
2075 * isolating, as that would be expensive.
2078 (PageLRU(page) || __PageMovable(page)))
2085 order = page_order(page);
2086 list_move(&page->lru,
2087 &zone->free_area[order].free_list[migratetype]);
2089 pages_moved += 1 << order;
2095 int move_freepages_block(struct zone *zone, struct page *page,
2096 int migratetype, int *num_movable)
2098 unsigned long start_pfn, end_pfn;
2099 struct page *start_page, *end_page;
2104 start_pfn = page_to_pfn(page);
2105 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2106 start_page = pfn_to_page(start_pfn);
2107 end_page = start_page + pageblock_nr_pages - 1;
2108 end_pfn = start_pfn + pageblock_nr_pages - 1;
2110 /* Do not cross zone boundaries */
2111 if (!zone_spans_pfn(zone, start_pfn))
2113 if (!zone_spans_pfn(zone, end_pfn))
2116 return move_freepages(zone, start_page, end_page, migratetype,
2120 static void change_pageblock_range(struct page *pageblock_page,
2121 int start_order, int migratetype)
2123 int nr_pageblocks = 1 << (start_order - pageblock_order);
2125 while (nr_pageblocks--) {
2126 set_pageblock_migratetype(pageblock_page, migratetype);
2127 pageblock_page += pageblock_nr_pages;
2132 * When we are falling back to another migratetype during allocation, try to
2133 * steal extra free pages from the same pageblocks to satisfy further
2134 * allocations, instead of polluting multiple pageblocks.
2136 * If we are stealing a relatively large buddy page, it is likely there will
2137 * be more free pages in the pageblock, so try to steal them all. For
2138 * reclaimable and unmovable allocations, we steal regardless of page size,
2139 * as fragmentation caused by those allocations polluting movable pageblocks
2140 * is worse than movable allocations stealing from unmovable and reclaimable
2143 static bool can_steal_fallback(unsigned int order, int start_mt)
2146 * Leaving this order check is intended, although there is
2147 * relaxed order check in next check. The reason is that
2148 * we can actually steal whole pageblock if this condition met,
2149 * but, below check doesn't guarantee it and that is just heuristic
2150 * so could be changed anytime.
2152 if (order >= pageblock_order)
2155 if (order >= pageblock_order / 2 ||
2156 start_mt == MIGRATE_RECLAIMABLE ||
2157 start_mt == MIGRATE_UNMOVABLE ||
2158 page_group_by_mobility_disabled)
2164 static inline void boost_watermark(struct zone *zone)
2166 unsigned long max_boost;
2168 if (!watermark_boost_factor)
2171 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2172 watermark_boost_factor, 10000);
2175 * high watermark may be uninitialised if fragmentation occurs
2176 * very early in boot so do not boost. We do not fall
2177 * through and boost by pageblock_nr_pages as failing
2178 * allocations that early means that reclaim is not going
2179 * to help and it may even be impossible to reclaim the
2180 * boosted watermark resulting in a hang.
2185 max_boost = max(pageblock_nr_pages, max_boost);
2187 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2192 * This function implements actual steal behaviour. If order is large enough,
2193 * we can steal whole pageblock. If not, we first move freepages in this
2194 * pageblock to our migratetype and determine how many already-allocated pages
2195 * are there in the pageblock with a compatible migratetype. If at least half
2196 * of pages are free or compatible, we can change migratetype of the pageblock
2197 * itself, so pages freed in the future will be put on the correct free list.
2199 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2200 unsigned int alloc_flags, int start_type, bool whole_block)
2202 unsigned int current_order = page_order(page);
2203 struct free_area *area;
2204 int free_pages, movable_pages, alike_pages;
2207 old_block_type = get_pageblock_migratetype(page);
2210 * This can happen due to races and we want to prevent broken
2211 * highatomic accounting.
2213 if (is_migrate_highatomic(old_block_type))
2216 /* Take ownership for orders >= pageblock_order */
2217 if (current_order >= pageblock_order) {
2218 change_pageblock_range(page, current_order, start_type);
2223 * Boost watermarks to increase reclaim pressure to reduce the
2224 * likelihood of future fallbacks. Wake kswapd now as the node
2225 * may be balanced overall and kswapd will not wake naturally.
2227 boost_watermark(zone);
2228 if (alloc_flags & ALLOC_KSWAPD)
2229 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2231 /* We are not allowed to try stealing from the whole block */
2235 free_pages = move_freepages_block(zone, page, start_type,
2238 * Determine how many pages are compatible with our allocation.
2239 * For movable allocation, it's the number of movable pages which
2240 * we just obtained. For other types it's a bit more tricky.
2242 if (start_type == MIGRATE_MOVABLE) {
2243 alike_pages = movable_pages;
2246 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2247 * to MOVABLE pageblock, consider all non-movable pages as
2248 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2249 * vice versa, be conservative since we can't distinguish the
2250 * exact migratetype of non-movable pages.
2252 if (old_block_type == MIGRATE_MOVABLE)
2253 alike_pages = pageblock_nr_pages
2254 - (free_pages + movable_pages);
2259 /* moving whole block can fail due to zone boundary conditions */
2264 * If a sufficient number of pages in the block are either free or of
2265 * comparable migratability as our allocation, claim the whole block.
2267 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2268 page_group_by_mobility_disabled)
2269 set_pageblock_migratetype(page, start_type);
2274 area = &zone->free_area[current_order];
2275 list_move(&page->lru, &area->free_list[start_type]);
2279 * Check whether there is a suitable fallback freepage with requested order.
2280 * If only_stealable is true, this function returns fallback_mt only if
2281 * we can steal other freepages all together. This would help to reduce
2282 * fragmentation due to mixed migratetype pages in one pageblock.
2284 int find_suitable_fallback(struct free_area *area, unsigned int order,
2285 int migratetype, bool only_stealable, bool *can_steal)
2290 if (area->nr_free == 0)
2295 fallback_mt = fallbacks[migratetype][i];
2296 if (fallback_mt == MIGRATE_TYPES)
2299 if (list_empty(&area->free_list[fallback_mt]))
2302 if (can_steal_fallback(order, migratetype))
2305 if (!only_stealable)
2316 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2317 * there are no empty page blocks that contain a page with a suitable order
2319 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2320 unsigned int alloc_order)
2323 unsigned long max_managed, flags;
2326 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2327 * Check is race-prone but harmless.
2329 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2330 if (zone->nr_reserved_highatomic >= max_managed)
2333 spin_lock_irqsave(&zone->lock, flags);
2335 /* Recheck the nr_reserved_highatomic limit under the lock */
2336 if (zone->nr_reserved_highatomic >= max_managed)
2340 mt = get_pageblock_migratetype(page);
2341 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2342 && !is_migrate_cma(mt)) {
2343 zone->nr_reserved_highatomic += pageblock_nr_pages;
2344 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2345 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2349 spin_unlock_irqrestore(&zone->lock, flags);
2353 * Used when an allocation is about to fail under memory pressure. This
2354 * potentially hurts the reliability of high-order allocations when under
2355 * intense memory pressure but failed atomic allocations should be easier
2356 * to recover from than an OOM.
2358 * If @force is true, try to unreserve a pageblock even though highatomic
2359 * pageblock is exhausted.
2361 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2364 struct zonelist *zonelist = ac->zonelist;
2365 unsigned long flags;
2372 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2375 * Preserve at least one pageblock unless memory pressure
2378 if (!force && zone->nr_reserved_highatomic <=
2382 spin_lock_irqsave(&zone->lock, flags);
2383 for (order = 0; order < MAX_ORDER; order++) {
2384 struct free_area *area = &(zone->free_area[order]);
2386 page = list_first_entry_or_null(
2387 &area->free_list[MIGRATE_HIGHATOMIC],
2393 * In page freeing path, migratetype change is racy so
2394 * we can counter several free pages in a pageblock
2395 * in this loop althoug we changed the pageblock type
2396 * from highatomic to ac->migratetype. So we should
2397 * adjust the count once.
2399 if (is_migrate_highatomic_page(page)) {
2401 * It should never happen but changes to
2402 * locking could inadvertently allow a per-cpu
2403 * drain to add pages to MIGRATE_HIGHATOMIC
2404 * while unreserving so be safe and watch for
2407 zone->nr_reserved_highatomic -= min(
2409 zone->nr_reserved_highatomic);
2413 * Convert to ac->migratetype and avoid the normal
2414 * pageblock stealing heuristics. Minimally, the caller
2415 * is doing the work and needs the pages. More
2416 * importantly, if the block was always converted to
2417 * MIGRATE_UNMOVABLE or another type then the number
2418 * of pageblocks that cannot be completely freed
2421 set_pageblock_migratetype(page, ac->migratetype);
2422 ret = move_freepages_block(zone, page, ac->migratetype,
2425 spin_unlock_irqrestore(&zone->lock, flags);
2429 spin_unlock_irqrestore(&zone->lock, flags);
2436 * Try finding a free buddy page on the fallback list and put it on the free
2437 * list of requested migratetype, possibly along with other pages from the same
2438 * block, depending on fragmentation avoidance heuristics. Returns true if
2439 * fallback was found so that __rmqueue_smallest() can grab it.
2441 * The use of signed ints for order and current_order is a deliberate
2442 * deviation from the rest of this file, to make the for loop
2443 * condition simpler.
2445 static __always_inline bool
2446 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2447 unsigned int alloc_flags)
2449 struct free_area *area;
2451 int min_order = order;
2457 * Do not steal pages from freelists belonging to other pageblocks
2458 * i.e. orders < pageblock_order. If there are no local zones free,
2459 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2461 if (alloc_flags & ALLOC_NOFRAGMENT)
2462 min_order = pageblock_order;
2465 * Find the largest available free page in the other list. This roughly
2466 * approximates finding the pageblock with the most free pages, which
2467 * would be too costly to do exactly.
2469 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2471 area = &(zone->free_area[current_order]);
2472 fallback_mt = find_suitable_fallback(area, current_order,
2473 start_migratetype, false, &can_steal);
2474 if (fallback_mt == -1)
2478 * We cannot steal all free pages from the pageblock and the
2479 * requested migratetype is movable. In that case it's better to
2480 * steal and split the smallest available page instead of the
2481 * largest available page, because even if the next movable
2482 * allocation falls back into a different pageblock than this
2483 * one, it won't cause permanent fragmentation.
2485 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2486 && current_order > order)
2495 for (current_order = order; current_order < MAX_ORDER;
2497 area = &(zone->free_area[current_order]);
2498 fallback_mt = find_suitable_fallback(area, current_order,
2499 start_migratetype, false, &can_steal);
2500 if (fallback_mt != -1)
2505 * This should not happen - we already found a suitable fallback
2506 * when looking for the largest page.
2508 VM_BUG_ON(current_order == MAX_ORDER);
2511 page = list_first_entry(&area->free_list[fallback_mt],
2514 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2517 trace_mm_page_alloc_extfrag(page, order, current_order,
2518 start_migratetype, fallback_mt);
2525 * Do the hard work of removing an element from the buddy allocator.
2526 * Call me with the zone->lock already held.
2528 static __always_inline struct page *
2529 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2530 unsigned int alloc_flags)
2535 page = __rmqueue_smallest(zone, order, migratetype);
2536 if (unlikely(!page)) {
2537 if (migratetype == MIGRATE_MOVABLE)
2538 page = __rmqueue_cma_fallback(zone, order);
2540 if (!page && __rmqueue_fallback(zone, order, migratetype,
2545 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2550 * Obtain a specified number of elements from the buddy allocator, all under
2551 * a single hold of the lock, for efficiency. Add them to the supplied list.
2552 * Returns the number of new pages which were placed at *list.
2554 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2555 unsigned long count, struct list_head *list,
2556 int migratetype, unsigned int alloc_flags)
2560 spin_lock(&zone->lock);
2561 for (i = 0; i < count; ++i) {
2562 struct page *page = __rmqueue(zone, order, migratetype,
2564 if (unlikely(page == NULL))
2567 if (unlikely(check_pcp_refill(page)))
2571 * Split buddy pages returned by expand() are received here in
2572 * physical page order. The page is added to the tail of
2573 * caller's list. From the callers perspective, the linked list
2574 * is ordered by page number under some conditions. This is
2575 * useful for IO devices that can forward direction from the
2576 * head, thus also in the physical page order. This is useful
2577 * for IO devices that can merge IO requests if the physical
2578 * pages are ordered properly.
2580 list_add_tail(&page->lru, list);
2582 if (is_migrate_cma(get_pcppage_migratetype(page)))
2583 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2588 * i pages were removed from the buddy list even if some leak due
2589 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2590 * on i. Do not confuse with 'alloced' which is the number of
2591 * pages added to the pcp list.
2593 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2594 spin_unlock(&zone->lock);
2600 * Called from the vmstat counter updater to drain pagesets of this
2601 * currently executing processor on remote nodes after they have
2604 * Note that this function must be called with the thread pinned to
2605 * a single processor.
2607 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2609 unsigned long flags;
2610 int to_drain, batch;
2612 local_irq_save(flags);
2613 batch = READ_ONCE(pcp->batch);
2614 to_drain = min(pcp->count, batch);
2616 free_pcppages_bulk(zone, to_drain, pcp);
2617 local_irq_restore(flags);
2622 * Drain pcplists of the indicated processor and zone.
2624 * The processor must either be the current processor and the
2625 * thread pinned to the current processor or a processor that
2628 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2630 unsigned long flags;
2631 struct per_cpu_pageset *pset;
2632 struct per_cpu_pages *pcp;
2634 local_irq_save(flags);
2635 pset = per_cpu_ptr(zone->pageset, cpu);
2639 free_pcppages_bulk(zone, pcp->count, pcp);
2640 local_irq_restore(flags);
2644 * Drain pcplists of all zones on the indicated processor.
2646 * The processor must either be the current processor and the
2647 * thread pinned to the current processor or a processor that
2650 static void drain_pages(unsigned int cpu)
2654 for_each_populated_zone(zone) {
2655 drain_pages_zone(cpu, zone);
2660 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2662 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2663 * the single zone's pages.
2665 void drain_local_pages(struct zone *zone)
2667 int cpu = smp_processor_id();
2670 drain_pages_zone(cpu, zone);
2675 static void drain_local_pages_wq(struct work_struct *work)
2677 struct pcpu_drain *drain;
2679 drain = container_of(work, struct pcpu_drain, work);
2682 * drain_all_pages doesn't use proper cpu hotplug protection so
2683 * we can race with cpu offline when the WQ can move this from
2684 * a cpu pinned worker to an unbound one. We can operate on a different
2685 * cpu which is allright but we also have to make sure to not move to
2689 drain_local_pages(drain->zone);
2694 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2696 * When zone parameter is non-NULL, spill just the single zone's pages.
2698 * Note that this can be extremely slow as the draining happens in a workqueue.
2700 void drain_all_pages(struct zone *zone)
2705 * Allocate in the BSS so we wont require allocation in
2706 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2708 static cpumask_t cpus_with_pcps;
2711 * Make sure nobody triggers this path before mm_percpu_wq is fully
2714 if (WARN_ON_ONCE(!mm_percpu_wq))
2718 * Do not drain if one is already in progress unless it's specific to
2719 * a zone. Such callers are primarily CMA and memory hotplug and need
2720 * the drain to be complete when the call returns.
2722 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2725 mutex_lock(&pcpu_drain_mutex);
2729 * We don't care about racing with CPU hotplug event
2730 * as offline notification will cause the notified
2731 * cpu to drain that CPU pcps and on_each_cpu_mask
2732 * disables preemption as part of its processing
2734 for_each_online_cpu(cpu) {
2735 struct per_cpu_pageset *pcp;
2737 bool has_pcps = false;
2740 pcp = per_cpu_ptr(zone->pageset, cpu);
2744 for_each_populated_zone(z) {
2745 pcp = per_cpu_ptr(z->pageset, cpu);
2746 if (pcp->pcp.count) {
2754 cpumask_set_cpu(cpu, &cpus_with_pcps);
2756 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2759 for_each_cpu(cpu, &cpus_with_pcps) {
2760 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
2763 INIT_WORK(&drain->work, drain_local_pages_wq);
2764 queue_work_on(cpu, mm_percpu_wq, &drain->work);
2766 for_each_cpu(cpu, &cpus_with_pcps)
2767 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
2769 mutex_unlock(&pcpu_drain_mutex);
2772 #ifdef CONFIG_HIBERNATION
2775 * Touch the watchdog for every WD_PAGE_COUNT pages.
2777 #define WD_PAGE_COUNT (128*1024)
2779 void mark_free_pages(struct zone *zone)
2781 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2782 unsigned long flags;
2783 unsigned int order, t;
2786 if (zone_is_empty(zone))
2789 spin_lock_irqsave(&zone->lock, flags);
2791 max_zone_pfn = zone_end_pfn(zone);
2792 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2793 if (pfn_valid(pfn)) {
2794 page = pfn_to_page(pfn);
2796 if (!--page_count) {
2797 touch_nmi_watchdog();
2798 page_count = WD_PAGE_COUNT;
2801 if (page_zone(page) != zone)
2804 if (!swsusp_page_is_forbidden(page))
2805 swsusp_unset_page_free(page);
2808 for_each_migratetype_order(order, t) {
2809 list_for_each_entry(page,
2810 &zone->free_area[order].free_list[t], lru) {
2813 pfn = page_to_pfn(page);
2814 for (i = 0; i < (1UL << order); i++) {
2815 if (!--page_count) {
2816 touch_nmi_watchdog();
2817 page_count = WD_PAGE_COUNT;
2819 swsusp_set_page_free(pfn_to_page(pfn + i));
2823 spin_unlock_irqrestore(&zone->lock, flags);
2825 #endif /* CONFIG_PM */
2827 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
2831 if (!free_pcp_prepare(page))
2834 migratetype = get_pfnblock_migratetype(page, pfn);
2835 set_pcppage_migratetype(page, migratetype);
2839 static void free_unref_page_commit(struct page *page, unsigned long pfn)
2841 struct zone *zone = page_zone(page);
2842 struct per_cpu_pages *pcp;
2845 migratetype = get_pcppage_migratetype(page);
2846 __count_vm_event(PGFREE);
2849 * We only track unmovable, reclaimable and movable on pcp lists.
2850 * Free ISOLATE pages back to the allocator because they are being
2851 * offlined but treat HIGHATOMIC as movable pages so we can get those
2852 * areas back if necessary. Otherwise, we may have to free
2853 * excessively into the page allocator
2855 if (migratetype >= MIGRATE_PCPTYPES) {
2856 if (unlikely(is_migrate_isolate(migratetype))) {
2857 free_one_page(zone, page, pfn, 0, migratetype);
2860 migratetype = MIGRATE_MOVABLE;
2863 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2864 list_add(&page->lru, &pcp->lists[migratetype]);
2866 if (pcp->count >= pcp->high) {
2867 unsigned long batch = READ_ONCE(pcp->batch);
2868 free_pcppages_bulk(zone, batch, pcp);
2873 * Free a 0-order page
2875 void free_unref_page(struct page *page)
2877 unsigned long flags;
2878 unsigned long pfn = page_to_pfn(page);
2880 if (!free_unref_page_prepare(page, pfn))
2883 local_irq_save(flags);
2884 free_unref_page_commit(page, pfn);
2885 local_irq_restore(flags);
2889 * Free a list of 0-order pages
2891 void free_unref_page_list(struct list_head *list)
2893 struct page *page, *next;
2894 unsigned long flags, pfn;
2895 int batch_count = 0;
2897 /* Prepare pages for freeing */
2898 list_for_each_entry_safe(page, next, list, lru) {
2899 pfn = page_to_pfn(page);
2900 if (!free_unref_page_prepare(page, pfn))
2901 list_del(&page->lru);
2902 set_page_private(page, pfn);
2905 local_irq_save(flags);
2906 list_for_each_entry_safe(page, next, list, lru) {
2907 unsigned long pfn = page_private(page);
2909 set_page_private(page, 0);
2910 trace_mm_page_free_batched(page);
2911 free_unref_page_commit(page, pfn);
2914 * Guard against excessive IRQ disabled times when we get
2915 * a large list of pages to free.
2917 if (++batch_count == SWAP_CLUSTER_MAX) {
2918 local_irq_restore(flags);
2920 local_irq_save(flags);
2923 local_irq_restore(flags);
2927 * split_page takes a non-compound higher-order page, and splits it into
2928 * n (1<<order) sub-pages: page[0..n]
2929 * Each sub-page must be freed individually.
2931 * Note: this is probably too low level an operation for use in drivers.
2932 * Please consult with lkml before using this in your driver.
2934 void split_page(struct page *page, unsigned int order)
2938 VM_BUG_ON_PAGE(PageCompound(page), page);
2939 VM_BUG_ON_PAGE(!page_count(page), page);
2941 for (i = 1; i < (1 << order); i++)
2942 set_page_refcounted(page + i);
2943 split_page_owner(page, order);
2945 EXPORT_SYMBOL_GPL(split_page);
2947 int __isolate_free_page(struct page *page, unsigned int order)
2949 unsigned long watermark;
2953 BUG_ON(!PageBuddy(page));
2955 zone = page_zone(page);
2956 mt = get_pageblock_migratetype(page);
2958 if (!is_migrate_isolate(mt)) {
2960 * Obey watermarks as if the page was being allocated. We can
2961 * emulate a high-order watermark check with a raised order-0
2962 * watermark, because we already know our high-order page
2965 watermark = min_wmark_pages(zone) + (1UL << order);
2966 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2969 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2972 /* Remove page from free list */
2973 list_del(&page->lru);
2974 zone->free_area[order].nr_free--;
2975 rmv_page_order(page);
2978 * Set the pageblock if the isolated page is at least half of a
2981 if (order >= pageblock_order - 1) {
2982 struct page *endpage = page + (1 << order) - 1;
2983 for (; page < endpage; page += pageblock_nr_pages) {
2984 int mt = get_pageblock_migratetype(page);
2985 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2986 && !is_migrate_highatomic(mt))
2987 set_pageblock_migratetype(page,
2993 return 1UL << order;
2997 * Update NUMA hit/miss statistics
2999 * Must be called with interrupts disabled.
3001 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3004 enum numa_stat_item local_stat = NUMA_LOCAL;
3006 /* skip numa counters update if numa stats is disabled */
3007 if (!static_branch_likely(&vm_numa_stat_key))
3010 if (zone_to_nid(z) != numa_node_id())
3011 local_stat = NUMA_OTHER;
3013 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3014 __inc_numa_state(z, NUMA_HIT);
3016 __inc_numa_state(z, NUMA_MISS);
3017 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3019 __inc_numa_state(z, local_stat);
3023 /* Remove page from the per-cpu list, caller must protect the list */
3024 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3025 unsigned int alloc_flags,
3026 struct per_cpu_pages *pcp,
3027 struct list_head *list)
3032 if (list_empty(list)) {
3033 pcp->count += rmqueue_bulk(zone, 0,
3035 migratetype, alloc_flags);
3036 if (unlikely(list_empty(list)))
3040 page = list_first_entry(list, struct page, lru);
3041 list_del(&page->lru);
3043 } while (check_new_pcp(page));
3048 /* Lock and remove page from the per-cpu list */
3049 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3050 struct zone *zone, unsigned int order,
3051 gfp_t gfp_flags, int migratetype,
3052 unsigned int alloc_flags)
3054 struct per_cpu_pages *pcp;
3055 struct list_head *list;
3057 unsigned long flags;
3059 local_irq_save(flags);
3060 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3061 list = &pcp->lists[migratetype];
3062 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3064 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3065 zone_statistics(preferred_zone, zone);
3067 local_irq_restore(flags);
3072 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3075 struct page *rmqueue(struct zone *preferred_zone,
3076 struct zone *zone, unsigned int order,
3077 gfp_t gfp_flags, unsigned int alloc_flags,
3080 unsigned long flags;
3083 if (likely(order == 0)) {
3084 page = rmqueue_pcplist(preferred_zone, zone, order,
3085 gfp_flags, migratetype, alloc_flags);
3090 * We most definitely don't want callers attempting to
3091 * allocate greater than order-1 page units with __GFP_NOFAIL.
3093 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3094 spin_lock_irqsave(&zone->lock, flags);
3098 if (alloc_flags & ALLOC_HARDER) {
3099 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3101 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3104 page = __rmqueue(zone, order, migratetype, alloc_flags);
3105 } while (page && check_new_pages(page, order));
3106 spin_unlock(&zone->lock);
3109 __mod_zone_freepage_state(zone, -(1 << order),
3110 get_pcppage_migratetype(page));
3112 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3113 zone_statistics(preferred_zone, zone);
3114 local_irq_restore(flags);
3117 /* Separate test+clear to avoid unnecessary atomics */
3118 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3119 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3120 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3123 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3127 local_irq_restore(flags);
3131 #ifdef CONFIG_FAIL_PAGE_ALLOC
3134 struct fault_attr attr;
3136 bool ignore_gfp_highmem;
3137 bool ignore_gfp_reclaim;
3139 } fail_page_alloc = {
3140 .attr = FAULT_ATTR_INITIALIZER,
3141 .ignore_gfp_reclaim = true,
3142 .ignore_gfp_highmem = true,
3146 static int __init setup_fail_page_alloc(char *str)
3148 return setup_fault_attr(&fail_page_alloc.attr, str);
3150 __setup("fail_page_alloc=", setup_fail_page_alloc);
3152 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3154 if (order < fail_page_alloc.min_order)
3156 if (gfp_mask & __GFP_NOFAIL)
3158 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3160 if (fail_page_alloc.ignore_gfp_reclaim &&
3161 (gfp_mask & __GFP_DIRECT_RECLAIM))
3164 return should_fail(&fail_page_alloc.attr, 1 << order);
3167 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3169 static int __init fail_page_alloc_debugfs(void)
3171 umode_t mode = S_IFREG | 0600;
3174 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3175 &fail_page_alloc.attr);
3177 return PTR_ERR(dir);
3179 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
3180 &fail_page_alloc.ignore_gfp_reclaim))
3182 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3183 &fail_page_alloc.ignore_gfp_highmem))
3185 if (!debugfs_create_u32("min-order", mode, dir,
3186 &fail_page_alloc.min_order))
3191 debugfs_remove_recursive(dir);
3196 late_initcall(fail_page_alloc_debugfs);
3198 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3200 #else /* CONFIG_FAIL_PAGE_ALLOC */
3202 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3207 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3209 static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3211 return __should_fail_alloc_page(gfp_mask, order);
3213 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3216 * Return true if free base pages are above 'mark'. For high-order checks it
3217 * will return true of the order-0 watermark is reached and there is at least
3218 * one free page of a suitable size. Checking now avoids taking the zone lock
3219 * to check in the allocation paths if no pages are free.
3221 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3222 int classzone_idx, unsigned int alloc_flags,
3227 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3229 /* free_pages may go negative - that's OK */
3230 free_pages -= (1 << order) - 1;
3232 if (alloc_flags & ALLOC_HIGH)
3236 * If the caller does not have rights to ALLOC_HARDER then subtract
3237 * the high-atomic reserves. This will over-estimate the size of the
3238 * atomic reserve but it avoids a search.
3240 if (likely(!alloc_harder)) {
3241 free_pages -= z->nr_reserved_highatomic;
3244 * OOM victims can try even harder than normal ALLOC_HARDER
3245 * users on the grounds that it's definitely going to be in
3246 * the exit path shortly and free memory. Any allocation it
3247 * makes during the free path will be small and short-lived.
3249 if (alloc_flags & ALLOC_OOM)
3257 /* If allocation can't use CMA areas don't use free CMA pages */
3258 if (!(alloc_flags & ALLOC_CMA))
3259 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3263 * Check watermarks for an order-0 allocation request. If these
3264 * are not met, then a high-order request also cannot go ahead
3265 * even if a suitable page happened to be free.
3267 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3270 /* If this is an order-0 request then the watermark is fine */
3274 /* For a high-order request, check at least one suitable page is free */
3275 for (o = order; o < MAX_ORDER; o++) {
3276 struct free_area *area = &z->free_area[o];
3282 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3283 if (!list_empty(&area->free_list[mt]))
3288 if ((alloc_flags & ALLOC_CMA) &&
3289 !list_empty(&area->free_list[MIGRATE_CMA])) {
3294 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3300 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3301 int classzone_idx, unsigned int alloc_flags)
3303 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3304 zone_page_state(z, NR_FREE_PAGES));
3307 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3308 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3310 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3314 /* If allocation can't use CMA areas don't use free CMA pages */
3315 if (!(alloc_flags & ALLOC_CMA))
3316 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3320 * Fast check for order-0 only. If this fails then the reserves
3321 * need to be calculated. There is a corner case where the check
3322 * passes but only the high-order atomic reserve are free. If
3323 * the caller is !atomic then it'll uselessly search the free
3324 * list. That corner case is then slower but it is harmless.
3326 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3329 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3333 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3334 unsigned long mark, int classzone_idx)
3336 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3338 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3339 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3341 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3346 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3348 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3351 #else /* CONFIG_NUMA */
3352 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3356 #endif /* CONFIG_NUMA */
3359 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3360 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3361 * premature use of a lower zone may cause lowmem pressure problems that
3362 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3363 * probably too small. It only makes sense to spread allocations to avoid
3364 * fragmentation between the Normal and DMA32 zones.
3366 static inline unsigned int
3367 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3369 unsigned int alloc_flags = 0;
3371 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3372 alloc_flags |= ALLOC_KSWAPD;
3374 #ifdef CONFIG_ZONE_DMA32
3375 if (zone_idx(zone) != ZONE_NORMAL)
3379 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3380 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3381 * on UMA that if Normal is populated then so is DMA32.
3383 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3384 if (nr_online_nodes > 1 && !populated_zone(--zone))
3388 #endif /* CONFIG_ZONE_DMA32 */
3393 * get_page_from_freelist goes through the zonelist trying to allocate
3396 static struct page *
3397 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3398 const struct alloc_context *ac)
3402 struct pglist_data *last_pgdat_dirty_limit = NULL;
3407 * Scan zonelist, looking for a zone with enough free.
3408 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3410 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3411 z = ac->preferred_zoneref;
3412 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3417 if (cpusets_enabled() &&
3418 (alloc_flags & ALLOC_CPUSET) &&
3419 !__cpuset_zone_allowed(zone, gfp_mask))
3422 * When allocating a page cache page for writing, we
3423 * want to get it from a node that is within its dirty
3424 * limit, such that no single node holds more than its
3425 * proportional share of globally allowed dirty pages.
3426 * The dirty limits take into account the node's
3427 * lowmem reserves and high watermark so that kswapd
3428 * should be able to balance it without having to
3429 * write pages from its LRU list.
3431 * XXX: For now, allow allocations to potentially
3432 * exceed the per-node dirty limit in the slowpath
3433 * (spread_dirty_pages unset) before going into reclaim,
3434 * which is important when on a NUMA setup the allowed
3435 * nodes are together not big enough to reach the
3436 * global limit. The proper fix for these situations
3437 * will require awareness of nodes in the
3438 * dirty-throttling and the flusher threads.
3440 if (ac->spread_dirty_pages) {
3441 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3444 if (!node_dirty_ok(zone->zone_pgdat)) {
3445 last_pgdat_dirty_limit = zone->zone_pgdat;
3450 if (no_fallback && nr_online_nodes > 1 &&
3451 zone != ac->preferred_zoneref->zone) {
3455 * If moving to a remote node, retry but allow
3456 * fragmenting fallbacks. Locality is more important
3457 * than fragmentation avoidance.
3459 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3460 if (zone_to_nid(zone) != local_nid) {
3461 alloc_flags &= ~ALLOC_NOFRAGMENT;
3466 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3467 if (!zone_watermark_fast(zone, order, mark,
3468 ac_classzone_idx(ac), alloc_flags)) {
3471 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3473 * Watermark failed for this zone, but see if we can
3474 * grow this zone if it contains deferred pages.
3476 if (static_branch_unlikely(&deferred_pages)) {
3477 if (_deferred_grow_zone(zone, order))
3481 /* Checked here to keep the fast path fast */
3482 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3483 if (alloc_flags & ALLOC_NO_WATERMARKS)
3486 if (node_reclaim_mode == 0 ||
3487 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3490 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3492 case NODE_RECLAIM_NOSCAN:
3495 case NODE_RECLAIM_FULL:
3496 /* scanned but unreclaimable */
3499 /* did we reclaim enough */
3500 if (zone_watermark_ok(zone, order, mark,
3501 ac_classzone_idx(ac), alloc_flags))
3509 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3510 gfp_mask, alloc_flags, ac->migratetype);
3512 prep_new_page(page, order, gfp_mask, alloc_flags);
3515 * If this is a high-order atomic allocation then check
3516 * if the pageblock should be reserved for the future
3518 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3519 reserve_highatomic_pageblock(page, zone, order);
3523 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3524 /* Try again if zone has deferred pages */
3525 if (static_branch_unlikely(&deferred_pages)) {
3526 if (_deferred_grow_zone(zone, order))
3534 * It's possible on a UMA machine to get through all zones that are
3535 * fragmented. If avoiding fragmentation, reset and try again.
3538 alloc_flags &= ~ALLOC_NOFRAGMENT;
3545 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3547 unsigned int filter = SHOW_MEM_FILTER_NODES;
3548 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3550 if (!__ratelimit(&show_mem_rs))
3554 * This documents exceptions given to allocations in certain
3555 * contexts that are allowed to allocate outside current's set
3558 if (!(gfp_mask & __GFP_NOMEMALLOC))
3559 if (tsk_is_oom_victim(current) ||
3560 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3561 filter &= ~SHOW_MEM_FILTER_NODES;
3562 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3563 filter &= ~SHOW_MEM_FILTER_NODES;
3565 show_mem(filter, nodemask);
3568 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3570 struct va_format vaf;
3572 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3573 DEFAULT_RATELIMIT_BURST);
3575 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3578 va_start(args, fmt);
3581 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3582 current->comm, &vaf, gfp_mask, &gfp_mask,
3583 nodemask_pr_args(nodemask));
3586 cpuset_print_current_mems_allowed();
3589 warn_alloc_show_mem(gfp_mask, nodemask);
3592 static inline struct page *
3593 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3594 unsigned int alloc_flags,
3595 const struct alloc_context *ac)
3599 page = get_page_from_freelist(gfp_mask, order,
3600 alloc_flags|ALLOC_CPUSET, ac);
3602 * fallback to ignore cpuset restriction if our nodes
3606 page = get_page_from_freelist(gfp_mask, order,
3612 static inline struct page *
3613 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3614 const struct alloc_context *ac, unsigned long *did_some_progress)
3616 struct oom_control oc = {
3617 .zonelist = ac->zonelist,
3618 .nodemask = ac->nodemask,
3620 .gfp_mask = gfp_mask,
3625 *did_some_progress = 0;
3628 * Acquire the oom lock. If that fails, somebody else is
3629 * making progress for us.
3631 if (!mutex_trylock(&oom_lock)) {
3632 *did_some_progress = 1;
3633 schedule_timeout_uninterruptible(1);
3638 * Go through the zonelist yet one more time, keep very high watermark
3639 * here, this is only to catch a parallel oom killing, we must fail if
3640 * we're still under heavy pressure. But make sure that this reclaim
3641 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3642 * allocation which will never fail due to oom_lock already held.
3644 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3645 ~__GFP_DIRECT_RECLAIM, order,
3646 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3650 /* Coredumps can quickly deplete all memory reserves */
3651 if (current->flags & PF_DUMPCORE)
3653 /* The OOM killer will not help higher order allocs */
3654 if (order > PAGE_ALLOC_COSTLY_ORDER)
3657 * We have already exhausted all our reclaim opportunities without any
3658 * success so it is time to admit defeat. We will skip the OOM killer
3659 * because it is very likely that the caller has a more reasonable
3660 * fallback than shooting a random task.
3662 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3664 /* The OOM killer does not needlessly kill tasks for lowmem */
3665 if (ac->high_zoneidx < ZONE_NORMAL)
3667 if (pm_suspended_storage())
3670 * XXX: GFP_NOFS allocations should rather fail than rely on
3671 * other request to make a forward progress.
3672 * We are in an unfortunate situation where out_of_memory cannot
3673 * do much for this context but let's try it to at least get
3674 * access to memory reserved if the current task is killed (see
3675 * out_of_memory). Once filesystems are ready to handle allocation
3676 * failures more gracefully we should just bail out here.
3679 /* The OOM killer may not free memory on a specific node */
3680 if (gfp_mask & __GFP_THISNODE)
3683 /* Exhausted what can be done so it's blame time */
3684 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3685 *did_some_progress = 1;
3688 * Help non-failing allocations by giving them access to memory
3691 if (gfp_mask & __GFP_NOFAIL)
3692 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3693 ALLOC_NO_WATERMARKS, ac);
3696 mutex_unlock(&oom_lock);
3701 * Maximum number of compaction retries wit a progress before OOM
3702 * killer is consider as the only way to move forward.
3704 #define MAX_COMPACT_RETRIES 16
3706 #ifdef CONFIG_COMPACTION
3707 /* Try memory compaction for high-order allocations before reclaim */
3708 static struct page *
3709 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3710 unsigned int alloc_flags, const struct alloc_context *ac,
3711 enum compact_priority prio, enum compact_result *compact_result)
3714 unsigned long pflags;
3715 unsigned int noreclaim_flag;
3720 psi_memstall_enter(&pflags);
3721 noreclaim_flag = memalloc_noreclaim_save();
3723 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3726 memalloc_noreclaim_restore(noreclaim_flag);
3727 psi_memstall_leave(&pflags);
3729 if (*compact_result <= COMPACT_INACTIVE)
3733 * At least in one zone compaction wasn't deferred or skipped, so let's
3734 * count a compaction stall
3736 count_vm_event(COMPACTSTALL);
3738 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3741 struct zone *zone = page_zone(page);
3743 zone->compact_blockskip_flush = false;
3744 compaction_defer_reset(zone, order, true);
3745 count_vm_event(COMPACTSUCCESS);
3750 * It's bad if compaction run occurs and fails. The most likely reason
3751 * is that pages exist, but not enough to satisfy watermarks.
3753 count_vm_event(COMPACTFAIL);
3761 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3762 enum compact_result compact_result,
3763 enum compact_priority *compact_priority,
3764 int *compaction_retries)
3766 int max_retries = MAX_COMPACT_RETRIES;
3769 int retries = *compaction_retries;
3770 enum compact_priority priority = *compact_priority;
3775 if (compaction_made_progress(compact_result))
3776 (*compaction_retries)++;
3779 * compaction considers all the zone as desperately out of memory
3780 * so it doesn't really make much sense to retry except when the
3781 * failure could be caused by insufficient priority
3783 if (compaction_failed(compact_result))
3784 goto check_priority;
3787 * make sure the compaction wasn't deferred or didn't bail out early
3788 * due to locks contention before we declare that we should give up.
3789 * But do not retry if the given zonelist is not suitable for
3792 if (compaction_withdrawn(compact_result)) {
3793 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3798 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3799 * costly ones because they are de facto nofail and invoke OOM
3800 * killer to move on while costly can fail and users are ready
3801 * to cope with that. 1/4 retries is rather arbitrary but we
3802 * would need much more detailed feedback from compaction to
3803 * make a better decision.
3805 if (order > PAGE_ALLOC_COSTLY_ORDER)
3807 if (*compaction_retries <= max_retries) {
3813 * Make sure there are attempts at the highest priority if we exhausted
3814 * all retries or failed at the lower priorities.
3817 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3818 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3820 if (*compact_priority > min_priority) {
3821 (*compact_priority)--;
3822 *compaction_retries = 0;
3826 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3830 static inline struct page *
3831 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3832 unsigned int alloc_flags, const struct alloc_context *ac,
3833 enum compact_priority prio, enum compact_result *compact_result)
3835 *compact_result = COMPACT_SKIPPED;
3840 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3841 enum compact_result compact_result,
3842 enum compact_priority *compact_priority,
3843 int *compaction_retries)
3848 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3852 * There are setups with compaction disabled which would prefer to loop
3853 * inside the allocator rather than hit the oom killer prematurely.
3854 * Let's give them a good hope and keep retrying while the order-0
3855 * watermarks are OK.
3857 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3859 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3860 ac_classzone_idx(ac), alloc_flags))
3865 #endif /* CONFIG_COMPACTION */
3867 #ifdef CONFIG_LOCKDEP
3868 static struct lockdep_map __fs_reclaim_map =
3869 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3871 static bool __need_fs_reclaim(gfp_t gfp_mask)
3873 gfp_mask = current_gfp_context(gfp_mask);
3875 /* no reclaim without waiting on it */
3876 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3879 /* this guy won't enter reclaim */
3880 if (current->flags & PF_MEMALLOC)
3883 /* We're only interested __GFP_FS allocations for now */
3884 if (!(gfp_mask & __GFP_FS))
3887 if (gfp_mask & __GFP_NOLOCKDEP)
3893 void __fs_reclaim_acquire(void)
3895 lock_map_acquire(&__fs_reclaim_map);
3898 void __fs_reclaim_release(void)
3900 lock_map_release(&__fs_reclaim_map);
3903 void fs_reclaim_acquire(gfp_t gfp_mask)
3905 if (__need_fs_reclaim(gfp_mask))
3906 __fs_reclaim_acquire();
3908 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3910 void fs_reclaim_release(gfp_t gfp_mask)
3912 if (__need_fs_reclaim(gfp_mask))
3913 __fs_reclaim_release();
3915 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3918 /* Perform direct synchronous page reclaim */
3920 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3921 const struct alloc_context *ac)
3923 struct reclaim_state reclaim_state;
3925 unsigned int noreclaim_flag;
3926 unsigned long pflags;
3930 /* We now go into synchronous reclaim */
3931 cpuset_memory_pressure_bump();
3932 psi_memstall_enter(&pflags);
3933 fs_reclaim_acquire(gfp_mask);
3934 noreclaim_flag = memalloc_noreclaim_save();
3935 reclaim_state.reclaimed_slab = 0;
3936 current->reclaim_state = &reclaim_state;
3938 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3941 current->reclaim_state = NULL;
3942 memalloc_noreclaim_restore(noreclaim_flag);
3943 fs_reclaim_release(gfp_mask);
3944 psi_memstall_leave(&pflags);
3951 /* The really slow allocator path where we enter direct reclaim */
3952 static inline struct page *
3953 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3954 unsigned int alloc_flags, const struct alloc_context *ac,
3955 unsigned long *did_some_progress)
3957 struct page *page = NULL;
3958 bool drained = false;
3960 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3961 if (unlikely(!(*did_some_progress)))
3965 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3968 * If an allocation failed after direct reclaim, it could be because
3969 * pages are pinned on the per-cpu lists or in high alloc reserves.
3970 * Shrink them them and try again
3972 if (!page && !drained) {
3973 unreserve_highatomic_pageblock(ac, false);
3974 drain_all_pages(NULL);
3982 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
3983 const struct alloc_context *ac)
3987 pg_data_t *last_pgdat = NULL;
3988 enum zone_type high_zoneidx = ac->high_zoneidx;
3990 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
3992 if (last_pgdat != zone->zone_pgdat)
3993 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
3994 last_pgdat = zone->zone_pgdat;
3998 static inline unsigned int
3999 gfp_to_alloc_flags(gfp_t gfp_mask)
4001 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4003 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
4004 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4007 * The caller may dip into page reserves a bit more if the caller
4008 * cannot run direct reclaim, or if the caller has realtime scheduling
4009 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4010 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4012 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
4014 if (gfp_mask & __GFP_ATOMIC) {
4016 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4017 * if it can't schedule.
4019 if (!(gfp_mask & __GFP_NOMEMALLOC))
4020 alloc_flags |= ALLOC_HARDER;
4022 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4023 * comment for __cpuset_node_allowed().
4025 alloc_flags &= ~ALLOC_CPUSET;
4026 } else if (unlikely(rt_task(current)) && !in_interrupt())
4027 alloc_flags |= ALLOC_HARDER;
4029 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4030 alloc_flags |= ALLOC_KSWAPD;
4033 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4034 alloc_flags |= ALLOC_CMA;
4039 static bool oom_reserves_allowed(struct task_struct *tsk)
4041 if (!tsk_is_oom_victim(tsk))
4045 * !MMU doesn't have oom reaper so give access to memory reserves
4046 * only to the thread with TIF_MEMDIE set
4048 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4055 * Distinguish requests which really need access to full memory
4056 * reserves from oom victims which can live with a portion of it
4058 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4060 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4062 if (gfp_mask & __GFP_MEMALLOC)
4063 return ALLOC_NO_WATERMARKS;
4064 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4065 return ALLOC_NO_WATERMARKS;
4066 if (!in_interrupt()) {
4067 if (current->flags & PF_MEMALLOC)
4068 return ALLOC_NO_WATERMARKS;
4069 else if (oom_reserves_allowed(current))
4076 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4078 return !!__gfp_pfmemalloc_flags(gfp_mask);
4082 * Checks whether it makes sense to retry the reclaim to make a forward progress
4083 * for the given allocation request.
4085 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4086 * without success, or when we couldn't even meet the watermark if we
4087 * reclaimed all remaining pages on the LRU lists.
4089 * Returns true if a retry is viable or false to enter the oom path.
4092 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4093 struct alloc_context *ac, int alloc_flags,
4094 bool did_some_progress, int *no_progress_loops)
4101 * Costly allocations might have made a progress but this doesn't mean
4102 * their order will become available due to high fragmentation so
4103 * always increment the no progress counter for them
4105 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4106 *no_progress_loops = 0;
4108 (*no_progress_loops)++;
4111 * Make sure we converge to OOM if we cannot make any progress
4112 * several times in the row.
4114 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4115 /* Before OOM, exhaust highatomic_reserve */
4116 return unreserve_highatomic_pageblock(ac, true);
4120 * Keep reclaiming pages while there is a chance this will lead
4121 * somewhere. If none of the target zones can satisfy our allocation
4122 * request even if all reclaimable pages are considered then we are
4123 * screwed and have to go OOM.
4125 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4127 unsigned long available;
4128 unsigned long reclaimable;
4129 unsigned long min_wmark = min_wmark_pages(zone);
4132 available = reclaimable = zone_reclaimable_pages(zone);
4133 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4136 * Would the allocation succeed if we reclaimed all
4137 * reclaimable pages?
4139 wmark = __zone_watermark_ok(zone, order, min_wmark,
4140 ac_classzone_idx(ac), alloc_flags, available);
4141 trace_reclaim_retry_zone(z, order, reclaimable,
4142 available, min_wmark, *no_progress_loops, wmark);
4145 * If we didn't make any progress and have a lot of
4146 * dirty + writeback pages then we should wait for
4147 * an IO to complete to slow down the reclaim and
4148 * prevent from pre mature OOM
4150 if (!did_some_progress) {
4151 unsigned long write_pending;
4153 write_pending = zone_page_state_snapshot(zone,
4154 NR_ZONE_WRITE_PENDING);
4156 if (2 * write_pending > reclaimable) {
4157 congestion_wait(BLK_RW_ASYNC, HZ/10);
4169 * Memory allocation/reclaim might be called from a WQ context and the
4170 * current implementation of the WQ concurrency control doesn't
4171 * recognize that a particular WQ is congested if the worker thread is
4172 * looping without ever sleeping. Therefore we have to do a short sleep
4173 * here rather than calling cond_resched().
4175 if (current->flags & PF_WQ_WORKER)
4176 schedule_timeout_uninterruptible(1);
4183 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4186 * It's possible that cpuset's mems_allowed and the nodemask from
4187 * mempolicy don't intersect. This should be normally dealt with by
4188 * policy_nodemask(), but it's possible to race with cpuset update in
4189 * such a way the check therein was true, and then it became false
4190 * before we got our cpuset_mems_cookie here.
4191 * This assumes that for all allocations, ac->nodemask can come only
4192 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4193 * when it does not intersect with the cpuset restrictions) or the
4194 * caller can deal with a violated nodemask.
4196 if (cpusets_enabled() && ac->nodemask &&
4197 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4198 ac->nodemask = NULL;
4203 * When updating a task's mems_allowed or mempolicy nodemask, it is
4204 * possible to race with parallel threads in such a way that our
4205 * allocation can fail while the mask is being updated. If we are about
4206 * to fail, check if the cpuset changed during allocation and if so,
4209 if (read_mems_allowed_retry(cpuset_mems_cookie))
4215 static inline struct page *
4216 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4217 struct alloc_context *ac)
4219 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4220 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4221 struct page *page = NULL;
4222 unsigned int alloc_flags;
4223 unsigned long did_some_progress;
4224 enum compact_priority compact_priority;
4225 enum compact_result compact_result;
4226 int compaction_retries;
4227 int no_progress_loops;
4228 unsigned int cpuset_mems_cookie;
4232 * We also sanity check to catch abuse of atomic reserves being used by
4233 * callers that are not in atomic context.
4235 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4236 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4237 gfp_mask &= ~__GFP_ATOMIC;
4240 compaction_retries = 0;
4241 no_progress_loops = 0;
4242 compact_priority = DEF_COMPACT_PRIORITY;
4243 cpuset_mems_cookie = read_mems_allowed_begin();
4246 * The fast path uses conservative alloc_flags to succeed only until
4247 * kswapd needs to be woken up, and to avoid the cost of setting up
4248 * alloc_flags precisely. So we do that now.
4250 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4253 * We need to recalculate the starting point for the zonelist iterator
4254 * because we might have used different nodemask in the fast path, or
4255 * there was a cpuset modification and we are retrying - otherwise we
4256 * could end up iterating over non-eligible zones endlessly.
4258 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4259 ac->high_zoneidx, ac->nodemask);
4260 if (!ac->preferred_zoneref->zone)
4263 if (alloc_flags & ALLOC_KSWAPD)
4264 wake_all_kswapds(order, gfp_mask, ac);
4267 * The adjusted alloc_flags might result in immediate success, so try
4270 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4275 * For costly allocations, try direct compaction first, as it's likely
4276 * that we have enough base pages and don't need to reclaim. For non-
4277 * movable high-order allocations, do that as well, as compaction will
4278 * try prevent permanent fragmentation by migrating from blocks of the
4280 * Don't try this for allocations that are allowed to ignore
4281 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4283 if (can_direct_reclaim &&
4285 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4286 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4287 page = __alloc_pages_direct_compact(gfp_mask, order,
4289 INIT_COMPACT_PRIORITY,
4295 * Checks for costly allocations with __GFP_NORETRY, which
4296 * includes THP page fault allocations
4298 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4300 * If compaction is deferred for high-order allocations,
4301 * it is because sync compaction recently failed. If
4302 * this is the case and the caller requested a THP
4303 * allocation, we do not want to heavily disrupt the
4304 * system, so we fail the allocation instead of entering
4307 if (compact_result == COMPACT_DEFERRED)
4311 * Looks like reclaim/compaction is worth trying, but
4312 * sync compaction could be very expensive, so keep
4313 * using async compaction.
4315 compact_priority = INIT_COMPACT_PRIORITY;
4320 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4321 if (alloc_flags & ALLOC_KSWAPD)
4322 wake_all_kswapds(order, gfp_mask, ac);
4324 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4326 alloc_flags = reserve_flags;
4329 * Reset the nodemask and zonelist iterators if memory policies can be
4330 * ignored. These allocations are high priority and system rather than
4333 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4334 ac->nodemask = NULL;
4335 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4336 ac->high_zoneidx, ac->nodemask);
4339 /* Attempt with potentially adjusted zonelist and alloc_flags */
4340 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4344 /* Caller is not willing to reclaim, we can't balance anything */
4345 if (!can_direct_reclaim)
4348 /* Avoid recursion of direct reclaim */
4349 if (current->flags & PF_MEMALLOC)
4352 /* Try direct reclaim and then allocating */
4353 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4354 &did_some_progress);
4358 /* Try direct compaction and then allocating */
4359 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4360 compact_priority, &compact_result);
4364 /* Do not loop if specifically requested */
4365 if (gfp_mask & __GFP_NORETRY)
4369 * Do not retry costly high order allocations unless they are
4370 * __GFP_RETRY_MAYFAIL
4372 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4375 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4376 did_some_progress > 0, &no_progress_loops))
4380 * It doesn't make any sense to retry for the compaction if the order-0
4381 * reclaim is not able to make any progress because the current
4382 * implementation of the compaction depends on the sufficient amount
4383 * of free memory (see __compaction_suitable)
4385 if (did_some_progress > 0 &&
4386 should_compact_retry(ac, order, alloc_flags,
4387 compact_result, &compact_priority,
4388 &compaction_retries))
4392 /* Deal with possible cpuset update races before we start OOM killing */
4393 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4396 /* Reclaim has failed us, start killing things */
4397 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4401 /* Avoid allocations with no watermarks from looping endlessly */
4402 if (tsk_is_oom_victim(current) &&
4403 (alloc_flags == ALLOC_OOM ||
4404 (gfp_mask & __GFP_NOMEMALLOC)))
4407 /* Retry as long as the OOM killer is making progress */
4408 if (did_some_progress) {
4409 no_progress_loops = 0;
4414 /* Deal with possible cpuset update races before we fail */
4415 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4419 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4422 if (gfp_mask & __GFP_NOFAIL) {
4424 * All existing users of the __GFP_NOFAIL are blockable, so warn
4425 * of any new users that actually require GFP_NOWAIT
4427 if (WARN_ON_ONCE(!can_direct_reclaim))
4431 * PF_MEMALLOC request from this context is rather bizarre
4432 * because we cannot reclaim anything and only can loop waiting
4433 * for somebody to do a work for us
4435 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4438 * non failing costly orders are a hard requirement which we
4439 * are not prepared for much so let's warn about these users
4440 * so that we can identify them and convert them to something
4443 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4446 * Help non-failing allocations by giving them access to memory
4447 * reserves but do not use ALLOC_NO_WATERMARKS because this
4448 * could deplete whole memory reserves which would just make
4449 * the situation worse
4451 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4459 warn_alloc(gfp_mask, ac->nodemask,
4460 "page allocation failure: order:%u", order);
4465 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4466 int preferred_nid, nodemask_t *nodemask,
4467 struct alloc_context *ac, gfp_t *alloc_mask,
4468 unsigned int *alloc_flags)
4470 ac->high_zoneidx = gfp_zone(gfp_mask);
4471 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4472 ac->nodemask = nodemask;
4473 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4475 if (cpusets_enabled()) {
4476 *alloc_mask |= __GFP_HARDWALL;
4478 ac->nodemask = &cpuset_current_mems_allowed;
4480 *alloc_flags |= ALLOC_CPUSET;
4483 fs_reclaim_acquire(gfp_mask);
4484 fs_reclaim_release(gfp_mask);
4486 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4488 if (should_fail_alloc_page(gfp_mask, order))
4491 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4492 *alloc_flags |= ALLOC_CMA;
4497 /* Determine whether to spread dirty pages and what the first usable zone */
4498 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4500 /* Dirty zone balancing only done in the fast path */
4501 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4504 * The preferred zone is used for statistics but crucially it is
4505 * also used as the starting point for the zonelist iterator. It
4506 * may get reset for allocations that ignore memory policies.
4508 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4509 ac->high_zoneidx, ac->nodemask);
4513 * This is the 'heart' of the zoned buddy allocator.
4516 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4517 nodemask_t *nodemask)
4520 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4521 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4522 struct alloc_context ac = { };
4525 * There are several places where we assume that the order value is sane
4526 * so bail out early if the request is out of bound.
4528 if (unlikely(order >= MAX_ORDER)) {
4529 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4533 gfp_mask &= gfp_allowed_mask;
4534 alloc_mask = gfp_mask;
4535 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4538 finalise_ac(gfp_mask, &ac);
4541 * Forbid the first pass from falling back to types that fragment
4542 * memory until all local zones are considered.
4544 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4546 /* First allocation attempt */
4547 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4552 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4553 * resp. GFP_NOIO which has to be inherited for all allocation requests
4554 * from a particular context which has been marked by
4555 * memalloc_no{fs,io}_{save,restore}.
4557 alloc_mask = current_gfp_context(gfp_mask);
4558 ac.spread_dirty_pages = false;
4561 * Restore the original nodemask if it was potentially replaced with
4562 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4564 if (unlikely(ac.nodemask != nodemask))
4565 ac.nodemask = nodemask;
4567 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4570 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4571 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4572 __free_pages(page, order);
4576 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4580 EXPORT_SYMBOL(__alloc_pages_nodemask);
4583 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4584 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4585 * you need to access high mem.
4587 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4591 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4594 return (unsigned long) page_address(page);
4596 EXPORT_SYMBOL(__get_free_pages);
4598 unsigned long get_zeroed_page(gfp_t gfp_mask)
4600 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4602 EXPORT_SYMBOL(get_zeroed_page);
4604 static inline void free_the_page(struct page *page, unsigned int order)
4606 if (order == 0) /* Via pcp? */
4607 free_unref_page(page);
4609 __free_pages_ok(page, order);
4612 void __free_pages(struct page *page, unsigned int order)
4614 if (put_page_testzero(page))
4615 free_the_page(page, order);
4617 EXPORT_SYMBOL(__free_pages);
4619 void free_pages(unsigned long addr, unsigned int order)
4622 VM_BUG_ON(!virt_addr_valid((void *)addr));
4623 __free_pages(virt_to_page((void *)addr), order);
4627 EXPORT_SYMBOL(free_pages);
4631 * An arbitrary-length arbitrary-offset area of memory which resides
4632 * within a 0 or higher order page. Multiple fragments within that page
4633 * are individually refcounted, in the page's reference counter.
4635 * The page_frag functions below provide a simple allocation framework for
4636 * page fragments. This is used by the network stack and network device
4637 * drivers to provide a backing region of memory for use as either an
4638 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4640 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4643 struct page *page = NULL;
4644 gfp_t gfp = gfp_mask;
4646 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4647 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4649 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4650 PAGE_FRAG_CACHE_MAX_ORDER);
4651 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4653 if (unlikely(!page))
4654 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4656 nc->va = page ? page_address(page) : NULL;
4661 void __page_frag_cache_drain(struct page *page, unsigned int count)
4663 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4665 if (page_ref_sub_and_test(page, count))
4666 free_the_page(page, compound_order(page));
4668 EXPORT_SYMBOL(__page_frag_cache_drain);
4670 void *page_frag_alloc(struct page_frag_cache *nc,
4671 unsigned int fragsz, gfp_t gfp_mask)
4673 unsigned int size = PAGE_SIZE;
4677 if (unlikely(!nc->va)) {
4679 page = __page_frag_cache_refill(nc, gfp_mask);
4683 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4684 /* if size can vary use size else just use PAGE_SIZE */
4687 /* Even if we own the page, we do not use atomic_set().
4688 * This would break get_page_unless_zero() users.
4690 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4692 /* reset page count bias and offset to start of new frag */
4693 nc->pfmemalloc = page_is_pfmemalloc(page);
4694 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4698 offset = nc->offset - fragsz;
4699 if (unlikely(offset < 0)) {
4700 page = virt_to_page(nc->va);
4702 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4705 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4706 /* if size can vary use size else just use PAGE_SIZE */
4709 /* OK, page count is 0, we can safely set it */
4710 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4712 /* reset page count bias and offset to start of new frag */
4713 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4714 offset = size - fragsz;
4718 nc->offset = offset;
4720 return nc->va + offset;
4722 EXPORT_SYMBOL(page_frag_alloc);
4725 * Frees a page fragment allocated out of either a compound or order 0 page.
4727 void page_frag_free(void *addr)
4729 struct page *page = virt_to_head_page(addr);
4731 if (unlikely(put_page_testzero(page)))
4732 free_the_page(page, compound_order(page));
4734 EXPORT_SYMBOL(page_frag_free);
4736 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4740 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4741 unsigned long used = addr + PAGE_ALIGN(size);
4743 split_page(virt_to_page((void *)addr), order);
4744 while (used < alloc_end) {
4749 return (void *)addr;
4753 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4754 * @size: the number of bytes to allocate
4755 * @gfp_mask: GFP flags for the allocation
4757 * This function is similar to alloc_pages(), except that it allocates the
4758 * minimum number of pages to satisfy the request. alloc_pages() can only
4759 * allocate memory in power-of-two pages.
4761 * This function is also limited by MAX_ORDER.
4763 * Memory allocated by this function must be released by free_pages_exact().
4765 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4767 unsigned int order = get_order(size);
4770 addr = __get_free_pages(gfp_mask, order);
4771 return make_alloc_exact(addr, order, size);
4773 EXPORT_SYMBOL(alloc_pages_exact);
4776 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4778 * @nid: the preferred node ID where memory should be allocated
4779 * @size: the number of bytes to allocate
4780 * @gfp_mask: GFP flags for the allocation
4782 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4785 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4787 unsigned int order = get_order(size);
4788 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4791 return make_alloc_exact((unsigned long)page_address(p), order, size);
4795 * free_pages_exact - release memory allocated via alloc_pages_exact()
4796 * @virt: the value returned by alloc_pages_exact.
4797 * @size: size of allocation, same value as passed to alloc_pages_exact().
4799 * Release the memory allocated by a previous call to alloc_pages_exact.
4801 void free_pages_exact(void *virt, size_t size)
4803 unsigned long addr = (unsigned long)virt;
4804 unsigned long end = addr + PAGE_ALIGN(size);
4806 while (addr < end) {
4811 EXPORT_SYMBOL(free_pages_exact);
4814 * nr_free_zone_pages - count number of pages beyond high watermark
4815 * @offset: The zone index of the highest zone
4817 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4818 * high watermark within all zones at or below a given zone index. For each
4819 * zone, the number of pages is calculated as:
4821 * nr_free_zone_pages = managed_pages - high_pages
4823 static unsigned long nr_free_zone_pages(int offset)
4828 /* Just pick one node, since fallback list is circular */
4829 unsigned long sum = 0;
4831 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4833 for_each_zone_zonelist(zone, z, zonelist, offset) {
4834 unsigned long size = zone_managed_pages(zone);
4835 unsigned long high = high_wmark_pages(zone);
4844 * nr_free_buffer_pages - count number of pages beyond high watermark
4846 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4847 * watermark within ZONE_DMA and ZONE_NORMAL.
4849 unsigned long nr_free_buffer_pages(void)
4851 return nr_free_zone_pages(gfp_zone(GFP_USER));
4853 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4856 * nr_free_pagecache_pages - count number of pages beyond high watermark
4858 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4859 * high watermark within all zones.
4861 unsigned long nr_free_pagecache_pages(void)
4863 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4866 static inline void show_node(struct zone *zone)
4868 if (IS_ENABLED(CONFIG_NUMA))
4869 printk("Node %d ", zone_to_nid(zone));
4872 long si_mem_available(void)
4875 unsigned long pagecache;
4876 unsigned long wmark_low = 0;
4877 unsigned long pages[NR_LRU_LISTS];
4878 unsigned long reclaimable;
4882 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4883 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4886 wmark_low += low_wmark_pages(zone);
4889 * Estimate the amount of memory available for userspace allocations,
4890 * without causing swapping.
4892 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
4895 * Not all the page cache can be freed, otherwise the system will
4896 * start swapping. Assume at least half of the page cache, or the
4897 * low watermark worth of cache, needs to stay.
4899 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4900 pagecache -= min(pagecache / 2, wmark_low);
4901 available += pagecache;
4904 * Part of the reclaimable slab and other kernel memory consists of
4905 * items that are in use, and cannot be freed. Cap this estimate at the
4908 reclaimable = global_node_page_state(NR_SLAB_RECLAIMABLE) +
4909 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
4910 available += reclaimable - min(reclaimable / 2, wmark_low);
4916 EXPORT_SYMBOL_GPL(si_mem_available);
4918 void si_meminfo(struct sysinfo *val)
4920 val->totalram = totalram_pages();
4921 val->sharedram = global_node_page_state(NR_SHMEM);
4922 val->freeram = global_zone_page_state(NR_FREE_PAGES);
4923 val->bufferram = nr_blockdev_pages();
4924 val->totalhigh = totalhigh_pages();
4925 val->freehigh = nr_free_highpages();
4926 val->mem_unit = PAGE_SIZE;
4929 EXPORT_SYMBOL(si_meminfo);
4932 void si_meminfo_node(struct sysinfo *val, int nid)
4934 int zone_type; /* needs to be signed */
4935 unsigned long managed_pages = 0;
4936 unsigned long managed_highpages = 0;
4937 unsigned long free_highpages = 0;
4938 pg_data_t *pgdat = NODE_DATA(nid);
4940 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4941 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
4942 val->totalram = managed_pages;
4943 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4944 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4945 #ifdef CONFIG_HIGHMEM
4946 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4947 struct zone *zone = &pgdat->node_zones[zone_type];
4949 if (is_highmem(zone)) {
4950 managed_highpages += zone_managed_pages(zone);
4951 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4954 val->totalhigh = managed_highpages;
4955 val->freehigh = free_highpages;
4957 val->totalhigh = managed_highpages;
4958 val->freehigh = free_highpages;
4960 val->mem_unit = PAGE_SIZE;
4965 * Determine whether the node should be displayed or not, depending on whether
4966 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4968 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4970 if (!(flags & SHOW_MEM_FILTER_NODES))
4974 * no node mask - aka implicit memory numa policy. Do not bother with
4975 * the synchronization - read_mems_allowed_begin - because we do not
4976 * have to be precise here.
4979 nodemask = &cpuset_current_mems_allowed;
4981 return !node_isset(nid, *nodemask);
4984 #define K(x) ((x) << (PAGE_SHIFT-10))
4986 static void show_migration_types(unsigned char type)
4988 static const char types[MIGRATE_TYPES] = {
4989 [MIGRATE_UNMOVABLE] = 'U',
4990 [MIGRATE_MOVABLE] = 'M',
4991 [MIGRATE_RECLAIMABLE] = 'E',
4992 [MIGRATE_HIGHATOMIC] = 'H',
4994 [MIGRATE_CMA] = 'C',
4996 #ifdef CONFIG_MEMORY_ISOLATION
4997 [MIGRATE_ISOLATE] = 'I',
5000 char tmp[MIGRATE_TYPES + 1];
5004 for (i = 0; i < MIGRATE_TYPES; i++) {
5005 if (type & (1 << i))
5010 printk(KERN_CONT "(%s) ", tmp);
5014 * Show free area list (used inside shift_scroll-lock stuff)
5015 * We also calculate the percentage fragmentation. We do this by counting the
5016 * memory on each free list with the exception of the first item on the list.
5019 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5022 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5024 unsigned long free_pcp = 0;
5029 for_each_populated_zone(zone) {
5030 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5033 for_each_online_cpu(cpu)
5034 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5037 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5038 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5039 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
5040 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5041 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5042 " free:%lu free_pcp:%lu free_cma:%lu\n",
5043 global_node_page_state(NR_ACTIVE_ANON),
5044 global_node_page_state(NR_INACTIVE_ANON),
5045 global_node_page_state(NR_ISOLATED_ANON),
5046 global_node_page_state(NR_ACTIVE_FILE),
5047 global_node_page_state(NR_INACTIVE_FILE),
5048 global_node_page_state(NR_ISOLATED_FILE),
5049 global_node_page_state(NR_UNEVICTABLE),
5050 global_node_page_state(NR_FILE_DIRTY),
5051 global_node_page_state(NR_WRITEBACK),
5052 global_node_page_state(NR_UNSTABLE_NFS),
5053 global_node_page_state(NR_SLAB_RECLAIMABLE),
5054 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
5055 global_node_page_state(NR_FILE_MAPPED),
5056 global_node_page_state(NR_SHMEM),
5057 global_zone_page_state(NR_PAGETABLE),
5058 global_zone_page_state(NR_BOUNCE),
5059 global_zone_page_state(NR_FREE_PAGES),
5061 global_zone_page_state(NR_FREE_CMA_PAGES));
5063 for_each_online_pgdat(pgdat) {
5064 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5068 " active_anon:%lukB"
5069 " inactive_anon:%lukB"
5070 " active_file:%lukB"
5071 " inactive_file:%lukB"
5072 " unevictable:%lukB"
5073 " isolated(anon):%lukB"
5074 " isolated(file):%lukB"
5079 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5081 " shmem_pmdmapped: %lukB"
5084 " writeback_tmp:%lukB"
5086 " all_unreclaimable? %s"
5089 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5090 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5091 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5092 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5093 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5094 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5095 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5096 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5097 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5098 K(node_page_state(pgdat, NR_WRITEBACK)),
5099 K(node_page_state(pgdat, NR_SHMEM)),
5100 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5101 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5102 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5104 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5106 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5107 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
5108 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5112 for_each_populated_zone(zone) {
5115 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5119 for_each_online_cpu(cpu)
5120 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5129 " active_anon:%lukB"
5130 " inactive_anon:%lukB"
5131 " active_file:%lukB"
5132 " inactive_file:%lukB"
5133 " unevictable:%lukB"
5134 " writepending:%lukB"
5138 " kernel_stack:%lukB"
5146 K(zone_page_state(zone, NR_FREE_PAGES)),
5147 K(min_wmark_pages(zone)),
5148 K(low_wmark_pages(zone)),
5149 K(high_wmark_pages(zone)),
5150 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5151 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5152 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5153 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5154 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5155 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5156 K(zone->present_pages),
5157 K(zone_managed_pages(zone)),
5158 K(zone_page_state(zone, NR_MLOCK)),
5159 zone_page_state(zone, NR_KERNEL_STACK_KB),
5160 K(zone_page_state(zone, NR_PAGETABLE)),
5161 K(zone_page_state(zone, NR_BOUNCE)),
5163 K(this_cpu_read(zone->pageset->pcp.count)),
5164 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5165 printk("lowmem_reserve[]:");
5166 for (i = 0; i < MAX_NR_ZONES; i++)
5167 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5168 printk(KERN_CONT "\n");
5171 for_each_populated_zone(zone) {
5173 unsigned long nr[MAX_ORDER], flags, total = 0;
5174 unsigned char types[MAX_ORDER];
5176 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5179 printk(KERN_CONT "%s: ", zone->name);
5181 spin_lock_irqsave(&zone->lock, flags);
5182 for (order = 0; order < MAX_ORDER; order++) {
5183 struct free_area *area = &zone->free_area[order];
5186 nr[order] = area->nr_free;
5187 total += nr[order] << order;
5190 for (type = 0; type < MIGRATE_TYPES; type++) {
5191 if (!list_empty(&area->free_list[type]))
5192 types[order] |= 1 << type;
5195 spin_unlock_irqrestore(&zone->lock, flags);
5196 for (order = 0; order < MAX_ORDER; order++) {
5197 printk(KERN_CONT "%lu*%lukB ",
5198 nr[order], K(1UL) << order);
5200 show_migration_types(types[order]);
5202 printk(KERN_CONT "= %lukB\n", K(total));
5205 hugetlb_show_meminfo();
5207 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5209 show_swap_cache_info();
5212 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5214 zoneref->zone = zone;
5215 zoneref->zone_idx = zone_idx(zone);
5219 * Builds allocation fallback zone lists.
5221 * Add all populated zones of a node to the zonelist.
5223 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5226 enum zone_type zone_type = MAX_NR_ZONES;
5231 zone = pgdat->node_zones + zone_type;
5232 if (managed_zone(zone)) {
5233 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5234 check_highest_zone(zone_type);
5236 } while (zone_type);
5243 static int __parse_numa_zonelist_order(char *s)
5246 * We used to support different zonlists modes but they turned
5247 * out to be just not useful. Let's keep the warning in place
5248 * if somebody still use the cmd line parameter so that we do
5249 * not fail it silently
5251 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5252 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5258 static __init int setup_numa_zonelist_order(char *s)
5263 return __parse_numa_zonelist_order(s);
5265 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5267 char numa_zonelist_order[] = "Node";
5270 * sysctl handler for numa_zonelist_order
5272 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5273 void __user *buffer, size_t *length,
5280 return proc_dostring(table, write, buffer, length, ppos);
5281 str = memdup_user_nul(buffer, 16);
5283 return PTR_ERR(str);
5285 ret = __parse_numa_zonelist_order(str);
5291 #define MAX_NODE_LOAD (nr_online_nodes)
5292 static int node_load[MAX_NUMNODES];
5295 * find_next_best_node - find the next node that should appear in a given node's fallback list
5296 * @node: node whose fallback list we're appending
5297 * @used_node_mask: nodemask_t of already used nodes
5299 * We use a number of factors to determine which is the next node that should
5300 * appear on a given node's fallback list. The node should not have appeared
5301 * already in @node's fallback list, and it should be the next closest node
5302 * according to the distance array (which contains arbitrary distance values
5303 * from each node to each node in the system), and should also prefer nodes
5304 * with no CPUs, since presumably they'll have very little allocation pressure
5305 * on them otherwise.
5306 * It returns -1 if no node is found.
5308 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5311 int min_val = INT_MAX;
5312 int best_node = NUMA_NO_NODE;
5313 const struct cpumask *tmp = cpumask_of_node(0);
5315 /* Use the local node if we haven't already */
5316 if (!node_isset(node, *used_node_mask)) {
5317 node_set(node, *used_node_mask);
5321 for_each_node_state(n, N_MEMORY) {
5323 /* Don't want a node to appear more than once */
5324 if (node_isset(n, *used_node_mask))
5327 /* Use the distance array to find the distance */
5328 val = node_distance(node, n);
5330 /* Penalize nodes under us ("prefer the next node") */
5333 /* Give preference to headless and unused nodes */
5334 tmp = cpumask_of_node(n);
5335 if (!cpumask_empty(tmp))
5336 val += PENALTY_FOR_NODE_WITH_CPUS;
5338 /* Slight preference for less loaded node */
5339 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5340 val += node_load[n];
5342 if (val < min_val) {
5349 node_set(best_node, *used_node_mask);
5356 * Build zonelists ordered by node and zones within node.
5357 * This results in maximum locality--normal zone overflows into local
5358 * DMA zone, if any--but risks exhausting DMA zone.
5360 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5363 struct zoneref *zonerefs;
5366 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5368 for (i = 0; i < nr_nodes; i++) {
5371 pg_data_t *node = NODE_DATA(node_order[i]);
5373 nr_zones = build_zonerefs_node(node, zonerefs);
5374 zonerefs += nr_zones;
5376 zonerefs->zone = NULL;
5377 zonerefs->zone_idx = 0;
5381 * Build gfp_thisnode zonelists
5383 static void build_thisnode_zonelists(pg_data_t *pgdat)
5385 struct zoneref *zonerefs;
5388 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5389 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5390 zonerefs += nr_zones;
5391 zonerefs->zone = NULL;
5392 zonerefs->zone_idx = 0;
5396 * Build zonelists ordered by zone and nodes within zones.
5397 * This results in conserving DMA zone[s] until all Normal memory is
5398 * exhausted, but results in overflowing to remote node while memory
5399 * may still exist in local DMA zone.
5402 static void build_zonelists(pg_data_t *pgdat)
5404 static int node_order[MAX_NUMNODES];
5405 int node, load, nr_nodes = 0;
5406 nodemask_t used_mask;
5407 int local_node, prev_node;
5409 /* NUMA-aware ordering of nodes */
5410 local_node = pgdat->node_id;
5411 load = nr_online_nodes;
5412 prev_node = local_node;
5413 nodes_clear(used_mask);
5415 memset(node_order, 0, sizeof(node_order));
5416 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5418 * We don't want to pressure a particular node.
5419 * So adding penalty to the first node in same
5420 * distance group to make it round-robin.
5422 if (node_distance(local_node, node) !=
5423 node_distance(local_node, prev_node))
5424 node_load[node] = load;
5426 node_order[nr_nodes++] = node;
5431 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5432 build_thisnode_zonelists(pgdat);
5435 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5437 * Return node id of node used for "local" allocations.
5438 * I.e., first node id of first zone in arg node's generic zonelist.
5439 * Used for initializing percpu 'numa_mem', which is used primarily
5440 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5442 int local_memory_node(int node)
5446 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5447 gfp_zone(GFP_KERNEL),
5449 return zone_to_nid(z->zone);
5453 static void setup_min_unmapped_ratio(void);
5454 static void setup_min_slab_ratio(void);
5455 #else /* CONFIG_NUMA */
5457 static void build_zonelists(pg_data_t *pgdat)
5459 int node, local_node;
5460 struct zoneref *zonerefs;
5463 local_node = pgdat->node_id;
5465 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5466 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5467 zonerefs += nr_zones;
5470 * Now we build the zonelist so that it contains the zones
5471 * of all the other nodes.
5472 * We don't want to pressure a particular node, so when
5473 * building the zones for node N, we make sure that the
5474 * zones coming right after the local ones are those from
5475 * node N+1 (modulo N)
5477 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5478 if (!node_online(node))
5480 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5481 zonerefs += nr_zones;
5483 for (node = 0; node < local_node; node++) {
5484 if (!node_online(node))
5486 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5487 zonerefs += nr_zones;
5490 zonerefs->zone = NULL;
5491 zonerefs->zone_idx = 0;
5494 #endif /* CONFIG_NUMA */
5497 * Boot pageset table. One per cpu which is going to be used for all
5498 * zones and all nodes. The parameters will be set in such a way
5499 * that an item put on a list will immediately be handed over to
5500 * the buddy list. This is safe since pageset manipulation is done
5501 * with interrupts disabled.
5503 * The boot_pagesets must be kept even after bootup is complete for
5504 * unused processors and/or zones. They do play a role for bootstrapping
5505 * hotplugged processors.
5507 * zoneinfo_show() and maybe other functions do
5508 * not check if the processor is online before following the pageset pointer.
5509 * Other parts of the kernel may not check if the zone is available.
5511 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5512 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5513 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5515 static void __build_all_zonelists(void *data)
5518 int __maybe_unused cpu;
5519 pg_data_t *self = data;
5520 static DEFINE_SPINLOCK(lock);
5525 memset(node_load, 0, sizeof(node_load));
5529 * This node is hotadded and no memory is yet present. So just
5530 * building zonelists is fine - no need to touch other nodes.
5532 if (self && !node_online(self->node_id)) {
5533 build_zonelists(self);
5535 for_each_online_node(nid) {
5536 pg_data_t *pgdat = NODE_DATA(nid);
5538 build_zonelists(pgdat);
5541 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5543 * We now know the "local memory node" for each node--
5544 * i.e., the node of the first zone in the generic zonelist.
5545 * Set up numa_mem percpu variable for on-line cpus. During
5546 * boot, only the boot cpu should be on-line; we'll init the
5547 * secondary cpus' numa_mem as they come on-line. During
5548 * node/memory hotplug, we'll fixup all on-line cpus.
5550 for_each_online_cpu(cpu)
5551 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5558 static noinline void __init
5559 build_all_zonelists_init(void)
5563 __build_all_zonelists(NULL);
5566 * Initialize the boot_pagesets that are going to be used
5567 * for bootstrapping processors. The real pagesets for
5568 * each zone will be allocated later when the per cpu
5569 * allocator is available.
5571 * boot_pagesets are used also for bootstrapping offline
5572 * cpus if the system is already booted because the pagesets
5573 * are needed to initialize allocators on a specific cpu too.
5574 * F.e. the percpu allocator needs the page allocator which
5575 * needs the percpu allocator in order to allocate its pagesets
5576 * (a chicken-egg dilemma).
5578 for_each_possible_cpu(cpu)
5579 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5581 mminit_verify_zonelist();
5582 cpuset_init_current_mems_allowed();
5586 * unless system_state == SYSTEM_BOOTING.
5588 * __ref due to call of __init annotated helper build_all_zonelists_init
5589 * [protected by SYSTEM_BOOTING].
5591 void __ref build_all_zonelists(pg_data_t *pgdat)
5593 if (system_state == SYSTEM_BOOTING) {
5594 build_all_zonelists_init();
5596 __build_all_zonelists(pgdat);
5597 /* cpuset refresh routine should be here */
5599 vm_total_pages = nr_free_pagecache_pages();
5601 * Disable grouping by mobility if the number of pages in the
5602 * system is too low to allow the mechanism to work. It would be
5603 * more accurate, but expensive to check per-zone. This check is
5604 * made on memory-hotadd so a system can start with mobility
5605 * disabled and enable it later
5607 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5608 page_group_by_mobility_disabled = 1;
5610 page_group_by_mobility_disabled = 0;
5612 pr_info("Built %i zonelists, mobility grouping %s. Total pages: %ld\n",
5614 page_group_by_mobility_disabled ? "off" : "on",
5617 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5621 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5622 static bool __meminit
5623 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5625 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5626 static struct memblock_region *r;
5628 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5629 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5630 for_each_memblock(memory, r) {
5631 if (*pfn < memblock_region_memory_end_pfn(r))
5635 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5636 memblock_is_mirror(r)) {
5637 *pfn = memblock_region_memory_end_pfn(r);
5646 * Initially all pages are reserved - free ones are freed
5647 * up by memblock_free_all() once the early boot process is
5648 * done. Non-atomic initialization, single-pass.
5650 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5651 unsigned long start_pfn, enum memmap_context context,
5652 struct vmem_altmap *altmap)
5654 unsigned long pfn, end_pfn = start_pfn + size;
5657 if (highest_memmap_pfn < end_pfn - 1)
5658 highest_memmap_pfn = end_pfn - 1;
5660 #ifdef CONFIG_ZONE_DEVICE
5662 * Honor reservation requested by the driver for this ZONE_DEVICE
5663 * memory. We limit the total number of pages to initialize to just
5664 * those that might contain the memory mapping. We will defer the
5665 * ZONE_DEVICE page initialization until after we have released
5668 if (zone == ZONE_DEVICE) {
5672 if (start_pfn == altmap->base_pfn)
5673 start_pfn += altmap->reserve;
5674 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5678 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5680 * There can be holes in boot-time mem_map[]s handed to this
5681 * function. They do not exist on hotplugged memory.
5683 if (context == MEMMAP_EARLY) {
5684 if (!early_pfn_valid(pfn))
5686 if (!early_pfn_in_nid(pfn, nid))
5688 if (overlap_memmap_init(zone, &pfn))
5690 if (defer_init(nid, pfn, end_pfn))
5694 page = pfn_to_page(pfn);
5695 __init_single_page(page, pfn, zone, nid);
5696 if (context == MEMMAP_HOTPLUG)
5697 __SetPageReserved(page);
5700 * Mark the block movable so that blocks are reserved for
5701 * movable at startup. This will force kernel allocations
5702 * to reserve their blocks rather than leaking throughout
5703 * the address space during boot when many long-lived
5704 * kernel allocations are made.
5706 * bitmap is created for zone's valid pfn range. but memmap
5707 * can be created for invalid pages (for alignment)
5708 * check here not to call set_pageblock_migratetype() against
5711 if (!(pfn & (pageblock_nr_pages - 1))) {
5712 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5718 #ifdef CONFIG_ZONE_DEVICE
5719 void __ref memmap_init_zone_device(struct zone *zone,
5720 unsigned long start_pfn,
5722 struct dev_pagemap *pgmap)
5724 unsigned long pfn, end_pfn = start_pfn + size;
5725 struct pglist_data *pgdat = zone->zone_pgdat;
5726 unsigned long zone_idx = zone_idx(zone);
5727 unsigned long start = jiffies;
5728 int nid = pgdat->node_id;
5730 if (WARN_ON_ONCE(!pgmap || !is_dev_zone(zone)))
5734 * The call to memmap_init_zone should have already taken care
5735 * of the pages reserved for the memmap, so we can just jump to
5736 * the end of that region and start processing the device pages.
5738 if (pgmap->altmap_valid) {
5739 struct vmem_altmap *altmap = &pgmap->altmap;
5741 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5742 size = end_pfn - start_pfn;
5745 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5746 struct page *page = pfn_to_page(pfn);
5748 __init_single_page(page, pfn, zone_idx, nid);
5751 * Mark page reserved as it will need to wait for onlining
5752 * phase for it to be fully associated with a zone.
5754 * We can use the non-atomic __set_bit operation for setting
5755 * the flag as we are still initializing the pages.
5757 __SetPageReserved(page);
5760 * ZONE_DEVICE pages union ->lru with a ->pgmap back
5761 * pointer and hmm_data. It is a bug if a ZONE_DEVICE
5762 * page is ever freed or placed on a driver-private list.
5764 page->pgmap = pgmap;
5768 * Mark the block movable so that blocks are reserved for
5769 * movable at startup. This will force kernel allocations
5770 * to reserve their blocks rather than leaking throughout
5771 * the address space during boot when many long-lived
5772 * kernel allocations are made.
5774 * bitmap is created for zone's valid pfn range. but memmap
5775 * can be created for invalid pages (for alignment)
5776 * check here not to call set_pageblock_migratetype() against
5779 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
5780 * because this is done early in sparse_add_one_section
5782 if (!(pfn & (pageblock_nr_pages - 1))) {
5783 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5788 pr_info("%s initialised, %lu pages in %ums\n", dev_name(pgmap->dev),
5789 size, jiffies_to_msecs(jiffies - start));
5793 static void __meminit zone_init_free_lists(struct zone *zone)
5795 unsigned int order, t;
5796 for_each_migratetype_order(order, t) {
5797 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5798 zone->free_area[order].nr_free = 0;
5802 void __meminit __weak memmap_init(unsigned long size, int nid,
5803 unsigned long zone, unsigned long start_pfn)
5805 memmap_init_zone(size, nid, zone, start_pfn, MEMMAP_EARLY, NULL);
5808 static int zone_batchsize(struct zone *zone)
5814 * The per-cpu-pages pools are set to around 1000th of the
5817 batch = zone_managed_pages(zone) / 1024;
5818 /* But no more than a meg. */
5819 if (batch * PAGE_SIZE > 1024 * 1024)
5820 batch = (1024 * 1024) / PAGE_SIZE;
5821 batch /= 4; /* We effectively *= 4 below */
5826 * Clamp the batch to a 2^n - 1 value. Having a power
5827 * of 2 value was found to be more likely to have
5828 * suboptimal cache aliasing properties in some cases.
5830 * For example if 2 tasks are alternately allocating
5831 * batches of pages, one task can end up with a lot
5832 * of pages of one half of the possible page colors
5833 * and the other with pages of the other colors.
5835 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5840 /* The deferral and batching of frees should be suppressed under NOMMU
5843 * The problem is that NOMMU needs to be able to allocate large chunks
5844 * of contiguous memory as there's no hardware page translation to
5845 * assemble apparent contiguous memory from discontiguous pages.
5847 * Queueing large contiguous runs of pages for batching, however,
5848 * causes the pages to actually be freed in smaller chunks. As there
5849 * can be a significant delay between the individual batches being
5850 * recycled, this leads to the once large chunks of space being
5851 * fragmented and becoming unavailable for high-order allocations.
5858 * pcp->high and pcp->batch values are related and dependent on one another:
5859 * ->batch must never be higher then ->high.
5860 * The following function updates them in a safe manner without read side
5863 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5864 * those fields changing asynchronously (acording the the above rule).
5866 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5867 * outside of boot time (or some other assurance that no concurrent updaters
5870 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5871 unsigned long batch)
5873 /* start with a fail safe value for batch */
5877 /* Update high, then batch, in order */
5884 /* a companion to pageset_set_high() */
5885 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5887 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5890 static void pageset_init(struct per_cpu_pageset *p)
5892 struct per_cpu_pages *pcp;
5895 memset(p, 0, sizeof(*p));
5898 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5899 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5902 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5905 pageset_set_batch(p, batch);
5909 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5910 * to the value high for the pageset p.
5912 static void pageset_set_high(struct per_cpu_pageset *p,
5915 unsigned long batch = max(1UL, high / 4);
5916 if ((high / 4) > (PAGE_SHIFT * 8))
5917 batch = PAGE_SHIFT * 8;
5919 pageset_update(&p->pcp, high, batch);
5922 static void pageset_set_high_and_batch(struct zone *zone,
5923 struct per_cpu_pageset *pcp)
5925 if (percpu_pagelist_fraction)
5926 pageset_set_high(pcp,
5927 (zone_managed_pages(zone) /
5928 percpu_pagelist_fraction));
5930 pageset_set_batch(pcp, zone_batchsize(zone));
5933 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5935 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5938 pageset_set_high_and_batch(zone, pcp);
5941 void __meminit setup_zone_pageset(struct zone *zone)
5944 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5945 for_each_possible_cpu(cpu)
5946 zone_pageset_init(zone, cpu);
5950 * Allocate per cpu pagesets and initialize them.
5951 * Before this call only boot pagesets were available.
5953 void __init setup_per_cpu_pageset(void)
5955 struct pglist_data *pgdat;
5958 for_each_populated_zone(zone)
5959 setup_zone_pageset(zone);
5961 for_each_online_pgdat(pgdat)
5962 pgdat->per_cpu_nodestats =
5963 alloc_percpu(struct per_cpu_nodestat);
5966 static __meminit void zone_pcp_init(struct zone *zone)
5969 * per cpu subsystem is not up at this point. The following code
5970 * relies on the ability of the linker to provide the
5971 * offset of a (static) per cpu variable into the per cpu area.
5973 zone->pageset = &boot_pageset;
5975 if (populated_zone(zone))
5976 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5977 zone->name, zone->present_pages,
5978 zone_batchsize(zone));
5981 void __meminit init_currently_empty_zone(struct zone *zone,
5982 unsigned long zone_start_pfn,
5985 struct pglist_data *pgdat = zone->zone_pgdat;
5986 int zone_idx = zone_idx(zone) + 1;
5988 if (zone_idx > pgdat->nr_zones)
5989 pgdat->nr_zones = zone_idx;
5991 zone->zone_start_pfn = zone_start_pfn;
5993 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5994 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5996 (unsigned long)zone_idx(zone),
5997 zone_start_pfn, (zone_start_pfn + size));
5999 zone_init_free_lists(zone);
6000 zone->initialized = 1;
6003 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6004 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
6007 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
6009 int __meminit __early_pfn_to_nid(unsigned long pfn,
6010 struct mminit_pfnnid_cache *state)
6012 unsigned long start_pfn, end_pfn;
6015 if (state->last_start <= pfn && pfn < state->last_end)
6016 return state->last_nid;
6018 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
6019 if (nid != NUMA_NO_NODE) {
6020 state->last_start = start_pfn;
6021 state->last_end = end_pfn;
6022 state->last_nid = nid;
6027 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
6030 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
6031 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
6032 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
6034 * If an architecture guarantees that all ranges registered contain no holes
6035 * and may be freed, this this function may be used instead of calling
6036 * memblock_free_early_nid() manually.
6038 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
6040 unsigned long start_pfn, end_pfn;
6043 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
6044 start_pfn = min(start_pfn, max_low_pfn);
6045 end_pfn = min(end_pfn, max_low_pfn);
6047 if (start_pfn < end_pfn)
6048 memblock_free_early_nid(PFN_PHYS(start_pfn),
6049 (end_pfn - start_pfn) << PAGE_SHIFT,
6055 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6056 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6058 * If an architecture guarantees that all ranges registered contain no holes and may
6059 * be freed, this function may be used instead of calling memory_present() manually.
6061 void __init sparse_memory_present_with_active_regions(int nid)
6063 unsigned long start_pfn, end_pfn;
6066 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
6067 memory_present(this_nid, start_pfn, end_pfn);
6071 * get_pfn_range_for_nid - Return the start and end page frames for a node
6072 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6073 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6074 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6076 * It returns the start and end page frame of a node based on information
6077 * provided by memblock_set_node(). If called for a node
6078 * with no available memory, a warning is printed and the start and end
6081 void __init get_pfn_range_for_nid(unsigned int nid,
6082 unsigned long *start_pfn, unsigned long *end_pfn)
6084 unsigned long this_start_pfn, this_end_pfn;
6090 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6091 *start_pfn = min(*start_pfn, this_start_pfn);
6092 *end_pfn = max(*end_pfn, this_end_pfn);
6095 if (*start_pfn == -1UL)
6100 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6101 * assumption is made that zones within a node are ordered in monotonic
6102 * increasing memory addresses so that the "highest" populated zone is used
6104 static void __init find_usable_zone_for_movable(void)
6107 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6108 if (zone_index == ZONE_MOVABLE)
6111 if (arch_zone_highest_possible_pfn[zone_index] >
6112 arch_zone_lowest_possible_pfn[zone_index])
6116 VM_BUG_ON(zone_index == -1);
6117 movable_zone = zone_index;
6121 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6122 * because it is sized independent of architecture. Unlike the other zones,
6123 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6124 * in each node depending on the size of each node and how evenly kernelcore
6125 * is distributed. This helper function adjusts the zone ranges
6126 * provided by the architecture for a given node by using the end of the
6127 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6128 * zones within a node are in order of monotonic increases memory addresses
6130 static void __init adjust_zone_range_for_zone_movable(int nid,
6131 unsigned long zone_type,
6132 unsigned long node_start_pfn,
6133 unsigned long node_end_pfn,
6134 unsigned long *zone_start_pfn,
6135 unsigned long *zone_end_pfn)
6137 /* Only adjust if ZONE_MOVABLE is on this node */
6138 if (zone_movable_pfn[nid]) {
6139 /* Size ZONE_MOVABLE */
6140 if (zone_type == ZONE_MOVABLE) {
6141 *zone_start_pfn = zone_movable_pfn[nid];
6142 *zone_end_pfn = min(node_end_pfn,
6143 arch_zone_highest_possible_pfn[movable_zone]);
6145 /* Adjust for ZONE_MOVABLE starting within this range */
6146 } else if (!mirrored_kernelcore &&
6147 *zone_start_pfn < zone_movable_pfn[nid] &&
6148 *zone_end_pfn > zone_movable_pfn[nid]) {
6149 *zone_end_pfn = zone_movable_pfn[nid];
6151 /* Check if this whole range is within ZONE_MOVABLE */
6152 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6153 *zone_start_pfn = *zone_end_pfn;
6158 * Return the number of pages a zone spans in a node, including holes
6159 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6161 static unsigned long __init zone_spanned_pages_in_node(int nid,
6162 unsigned long zone_type,
6163 unsigned long node_start_pfn,
6164 unsigned long node_end_pfn,
6165 unsigned long *zone_start_pfn,
6166 unsigned long *zone_end_pfn,
6167 unsigned long *ignored)
6169 /* When hotadd a new node from cpu_up(), the node should be empty */
6170 if (!node_start_pfn && !node_end_pfn)
6173 /* Get the start and end of the zone */
6174 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
6175 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
6176 adjust_zone_range_for_zone_movable(nid, zone_type,
6177 node_start_pfn, node_end_pfn,
6178 zone_start_pfn, zone_end_pfn);
6180 /* Check that this node has pages within the zone's required range */
6181 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6184 /* Move the zone boundaries inside the node if necessary */
6185 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6186 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6188 /* Return the spanned pages */
6189 return *zone_end_pfn - *zone_start_pfn;
6193 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6194 * then all holes in the requested range will be accounted for.
6196 unsigned long __init __absent_pages_in_range(int nid,
6197 unsigned long range_start_pfn,
6198 unsigned long range_end_pfn)
6200 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6201 unsigned long start_pfn, end_pfn;
6204 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6205 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6206 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6207 nr_absent -= end_pfn - start_pfn;
6213 * absent_pages_in_range - Return number of page frames in holes within a range
6214 * @start_pfn: The start PFN to start searching for holes
6215 * @end_pfn: The end PFN to stop searching for holes
6217 * It returns the number of pages frames in memory holes within a range.
6219 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6220 unsigned long end_pfn)
6222 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6225 /* Return the number of page frames in holes in a zone on a node */
6226 static unsigned long __init zone_absent_pages_in_node(int nid,
6227 unsigned long zone_type,
6228 unsigned long node_start_pfn,
6229 unsigned long node_end_pfn,
6230 unsigned long *ignored)
6232 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6233 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6234 unsigned long zone_start_pfn, zone_end_pfn;
6235 unsigned long nr_absent;
6237 /* When hotadd a new node from cpu_up(), the node should be empty */
6238 if (!node_start_pfn && !node_end_pfn)
6241 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6242 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6244 adjust_zone_range_for_zone_movable(nid, zone_type,
6245 node_start_pfn, node_end_pfn,
6246 &zone_start_pfn, &zone_end_pfn);
6247 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6250 * ZONE_MOVABLE handling.
6251 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6254 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6255 unsigned long start_pfn, end_pfn;
6256 struct memblock_region *r;
6258 for_each_memblock(memory, r) {
6259 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6260 zone_start_pfn, zone_end_pfn);
6261 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6262 zone_start_pfn, zone_end_pfn);
6264 if (zone_type == ZONE_MOVABLE &&
6265 memblock_is_mirror(r))
6266 nr_absent += end_pfn - start_pfn;
6268 if (zone_type == ZONE_NORMAL &&
6269 !memblock_is_mirror(r))
6270 nr_absent += end_pfn - start_pfn;
6277 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6278 static inline unsigned long __init zone_spanned_pages_in_node(int nid,
6279 unsigned long zone_type,
6280 unsigned long node_start_pfn,
6281 unsigned long node_end_pfn,
6282 unsigned long *zone_start_pfn,
6283 unsigned long *zone_end_pfn,
6284 unsigned long *zones_size)
6288 *zone_start_pfn = node_start_pfn;
6289 for (zone = 0; zone < zone_type; zone++)
6290 *zone_start_pfn += zones_size[zone];
6292 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6294 return zones_size[zone_type];
6297 static inline unsigned long __init zone_absent_pages_in_node(int nid,
6298 unsigned long zone_type,
6299 unsigned long node_start_pfn,
6300 unsigned long node_end_pfn,
6301 unsigned long *zholes_size)
6306 return zholes_size[zone_type];
6309 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6311 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6312 unsigned long node_start_pfn,
6313 unsigned long node_end_pfn,
6314 unsigned long *zones_size,
6315 unsigned long *zholes_size)
6317 unsigned long realtotalpages = 0, totalpages = 0;
6320 for (i = 0; i < MAX_NR_ZONES; i++) {
6321 struct zone *zone = pgdat->node_zones + i;
6322 unsigned long zone_start_pfn, zone_end_pfn;
6323 unsigned long size, real_size;
6325 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6331 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6332 node_start_pfn, node_end_pfn,
6335 zone->zone_start_pfn = zone_start_pfn;
6337 zone->zone_start_pfn = 0;
6338 zone->spanned_pages = size;
6339 zone->present_pages = real_size;
6342 realtotalpages += real_size;
6345 pgdat->node_spanned_pages = totalpages;
6346 pgdat->node_present_pages = realtotalpages;
6347 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6351 #ifndef CONFIG_SPARSEMEM
6353 * Calculate the size of the zone->blockflags rounded to an unsigned long
6354 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6355 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6356 * round what is now in bits to nearest long in bits, then return it in
6359 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6361 unsigned long usemapsize;
6363 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6364 usemapsize = roundup(zonesize, pageblock_nr_pages);
6365 usemapsize = usemapsize >> pageblock_order;
6366 usemapsize *= NR_PAGEBLOCK_BITS;
6367 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6369 return usemapsize / 8;
6372 static void __ref setup_usemap(struct pglist_data *pgdat,
6374 unsigned long zone_start_pfn,
6375 unsigned long zonesize)
6377 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6378 zone->pageblock_flags = NULL;
6380 zone->pageblock_flags =
6381 memblock_alloc_node_nopanic(usemapsize,
6385 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6386 unsigned long zone_start_pfn, unsigned long zonesize) {}
6387 #endif /* CONFIG_SPARSEMEM */
6389 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6391 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6392 void __init set_pageblock_order(void)
6396 /* Check that pageblock_nr_pages has not already been setup */
6397 if (pageblock_order)
6400 if (HPAGE_SHIFT > PAGE_SHIFT)
6401 order = HUGETLB_PAGE_ORDER;
6403 order = MAX_ORDER - 1;
6406 * Assume the largest contiguous order of interest is a huge page.
6407 * This value may be variable depending on boot parameters on IA64 and
6410 pageblock_order = order;
6412 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6415 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6416 * is unused as pageblock_order is set at compile-time. See
6417 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6420 void __init set_pageblock_order(void)
6424 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6426 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6427 unsigned long present_pages)
6429 unsigned long pages = spanned_pages;
6432 * Provide a more accurate estimation if there are holes within
6433 * the zone and SPARSEMEM is in use. If there are holes within the
6434 * zone, each populated memory region may cost us one or two extra
6435 * memmap pages due to alignment because memmap pages for each
6436 * populated regions may not be naturally aligned on page boundary.
6437 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6439 if (spanned_pages > present_pages + (present_pages >> 4) &&
6440 IS_ENABLED(CONFIG_SPARSEMEM))
6441 pages = present_pages;
6443 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6446 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6447 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6449 spin_lock_init(&pgdat->split_queue_lock);
6450 INIT_LIST_HEAD(&pgdat->split_queue);
6451 pgdat->split_queue_len = 0;
6454 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6457 #ifdef CONFIG_COMPACTION
6458 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6460 init_waitqueue_head(&pgdat->kcompactd_wait);
6463 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6466 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6468 pgdat_resize_init(pgdat);
6470 pgdat_init_split_queue(pgdat);
6471 pgdat_init_kcompactd(pgdat);
6473 init_waitqueue_head(&pgdat->kswapd_wait);
6474 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6476 pgdat_page_ext_init(pgdat);
6477 spin_lock_init(&pgdat->lru_lock);
6478 lruvec_init(node_lruvec(pgdat));
6481 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6482 unsigned long remaining_pages)
6484 atomic_long_set(&zone->managed_pages, remaining_pages);
6485 zone_set_nid(zone, nid);
6486 zone->name = zone_names[idx];
6487 zone->zone_pgdat = NODE_DATA(nid);
6488 spin_lock_init(&zone->lock);
6489 zone_seqlock_init(zone);
6490 zone_pcp_init(zone);
6494 * Set up the zone data structures
6495 * - init pgdat internals
6496 * - init all zones belonging to this node
6498 * NOTE: this function is only called during memory hotplug
6500 #ifdef CONFIG_MEMORY_HOTPLUG
6501 void __ref free_area_init_core_hotplug(int nid)
6504 pg_data_t *pgdat = NODE_DATA(nid);
6506 pgdat_init_internals(pgdat);
6507 for (z = 0; z < MAX_NR_ZONES; z++)
6508 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6513 * Set up the zone data structures:
6514 * - mark all pages reserved
6515 * - mark all memory queues empty
6516 * - clear the memory bitmaps
6518 * NOTE: pgdat should get zeroed by caller.
6519 * NOTE: this function is only called during early init.
6521 static void __init free_area_init_core(struct pglist_data *pgdat)
6524 int nid = pgdat->node_id;
6526 pgdat_init_internals(pgdat);
6527 pgdat->per_cpu_nodestats = &boot_nodestats;
6529 for (j = 0; j < MAX_NR_ZONES; j++) {
6530 struct zone *zone = pgdat->node_zones + j;
6531 unsigned long size, freesize, memmap_pages;
6532 unsigned long zone_start_pfn = zone->zone_start_pfn;
6534 size = zone->spanned_pages;
6535 freesize = zone->present_pages;
6538 * Adjust freesize so that it accounts for how much memory
6539 * is used by this zone for memmap. This affects the watermark
6540 * and per-cpu initialisations
6542 memmap_pages = calc_memmap_size(size, freesize);
6543 if (!is_highmem_idx(j)) {
6544 if (freesize >= memmap_pages) {
6545 freesize -= memmap_pages;
6548 " %s zone: %lu pages used for memmap\n",
6549 zone_names[j], memmap_pages);
6551 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6552 zone_names[j], memmap_pages, freesize);
6555 /* Account for reserved pages */
6556 if (j == 0 && freesize > dma_reserve) {
6557 freesize -= dma_reserve;
6558 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6559 zone_names[0], dma_reserve);
6562 if (!is_highmem_idx(j))
6563 nr_kernel_pages += freesize;
6564 /* Charge for highmem memmap if there are enough kernel pages */
6565 else if (nr_kernel_pages > memmap_pages * 2)
6566 nr_kernel_pages -= memmap_pages;
6567 nr_all_pages += freesize;
6570 * Set an approximate value for lowmem here, it will be adjusted
6571 * when the bootmem allocator frees pages into the buddy system.
6572 * And all highmem pages will be managed by the buddy system.
6574 zone_init_internals(zone, j, nid, freesize);
6579 set_pageblock_order();
6580 setup_usemap(pgdat, zone, zone_start_pfn, size);
6581 init_currently_empty_zone(zone, zone_start_pfn, size);
6582 memmap_init(size, nid, j, zone_start_pfn);
6586 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6587 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6589 unsigned long __maybe_unused start = 0;
6590 unsigned long __maybe_unused offset = 0;
6592 /* Skip empty nodes */
6593 if (!pgdat->node_spanned_pages)
6596 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6597 offset = pgdat->node_start_pfn - start;
6598 /* ia64 gets its own node_mem_map, before this, without bootmem */
6599 if (!pgdat->node_mem_map) {
6600 unsigned long size, end;
6604 * The zone's endpoints aren't required to be MAX_ORDER
6605 * aligned but the node_mem_map endpoints must be in order
6606 * for the buddy allocator to function correctly.
6608 end = pgdat_end_pfn(pgdat);
6609 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6610 size = (end - start) * sizeof(struct page);
6611 map = memblock_alloc_node_nopanic(size, pgdat->node_id);
6612 pgdat->node_mem_map = map + offset;
6614 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6615 __func__, pgdat->node_id, (unsigned long)pgdat,
6616 (unsigned long)pgdat->node_mem_map);
6617 #ifndef CONFIG_NEED_MULTIPLE_NODES
6619 * With no DISCONTIG, the global mem_map is just set as node 0's
6621 if (pgdat == NODE_DATA(0)) {
6622 mem_map = NODE_DATA(0)->node_mem_map;
6623 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6624 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6626 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6631 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6632 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6634 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6635 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6637 pgdat->first_deferred_pfn = ULONG_MAX;
6640 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6643 void __init free_area_init_node(int nid, unsigned long *zones_size,
6644 unsigned long node_start_pfn,
6645 unsigned long *zholes_size)
6647 pg_data_t *pgdat = NODE_DATA(nid);
6648 unsigned long start_pfn = 0;
6649 unsigned long end_pfn = 0;
6651 /* pg_data_t should be reset to zero when it's allocated */
6652 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6654 pgdat->node_id = nid;
6655 pgdat->node_start_pfn = node_start_pfn;
6656 pgdat->per_cpu_nodestats = NULL;
6657 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6658 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6659 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6660 (u64)start_pfn << PAGE_SHIFT,
6661 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6663 start_pfn = node_start_pfn;
6665 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6666 zones_size, zholes_size);
6668 alloc_node_mem_map(pgdat);
6669 pgdat_set_deferred_range(pgdat);
6671 free_area_init_core(pgdat);
6674 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6676 * Zero all valid struct pages in range [spfn, epfn), return number of struct
6679 static u64 zero_pfn_range(unsigned long spfn, unsigned long epfn)
6684 for (pfn = spfn; pfn < epfn; pfn++) {
6685 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6686 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6687 + pageblock_nr_pages - 1;
6690 mm_zero_struct_page(pfn_to_page(pfn));
6698 * Only struct pages that are backed by physical memory are zeroed and
6699 * initialized by going through __init_single_page(). But, there are some
6700 * struct pages which are reserved in memblock allocator and their fields
6701 * may be accessed (for example page_to_pfn() on some configuration accesses
6702 * flags). We must explicitly zero those struct pages.
6704 * This function also addresses a similar issue where struct pages are left
6705 * uninitialized because the physical address range is not covered by
6706 * memblock.memory or memblock.reserved. That could happen when memblock
6707 * layout is manually configured via memmap=.
6709 void __init zero_resv_unavail(void)
6711 phys_addr_t start, end;
6713 phys_addr_t next = 0;
6716 * Loop through unavailable ranges not covered by memblock.memory.
6719 for_each_mem_range(i, &memblock.memory, NULL,
6720 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
6722 pgcnt += zero_pfn_range(PFN_DOWN(next), PFN_UP(start));
6725 pgcnt += zero_pfn_range(PFN_DOWN(next), max_pfn);
6728 * Struct pages that do not have backing memory. This could be because
6729 * firmware is using some of this memory, or for some other reasons.
6732 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
6734 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
6736 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6738 #if MAX_NUMNODES > 1
6740 * Figure out the number of possible node ids.
6742 void __init setup_nr_node_ids(void)
6744 unsigned int highest;
6746 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6747 nr_node_ids = highest + 1;
6752 * node_map_pfn_alignment - determine the maximum internode alignment
6754 * This function should be called after node map is populated and sorted.
6755 * It calculates the maximum power of two alignment which can distinguish
6758 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6759 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6760 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6761 * shifted, 1GiB is enough and this function will indicate so.
6763 * This is used to test whether pfn -> nid mapping of the chosen memory
6764 * model has fine enough granularity to avoid incorrect mapping for the
6765 * populated node map.
6767 * Returns the determined alignment in pfn's. 0 if there is no alignment
6768 * requirement (single node).
6770 unsigned long __init node_map_pfn_alignment(void)
6772 unsigned long accl_mask = 0, last_end = 0;
6773 unsigned long start, end, mask;
6774 int last_nid = NUMA_NO_NODE;
6777 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6778 if (!start || last_nid < 0 || last_nid == nid) {
6785 * Start with a mask granular enough to pin-point to the
6786 * start pfn and tick off bits one-by-one until it becomes
6787 * too coarse to separate the current node from the last.
6789 mask = ~((1 << __ffs(start)) - 1);
6790 while (mask && last_end <= (start & (mask << 1)))
6793 /* accumulate all internode masks */
6797 /* convert mask to number of pages */
6798 return ~accl_mask + 1;
6801 /* Find the lowest pfn for a node */
6802 static unsigned long __init find_min_pfn_for_node(int nid)
6804 unsigned long min_pfn = ULONG_MAX;
6805 unsigned long start_pfn;
6808 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6809 min_pfn = min(min_pfn, start_pfn);
6811 if (min_pfn == ULONG_MAX) {
6812 pr_warn("Could not find start_pfn for node %d\n", nid);
6820 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6822 * It returns the minimum PFN based on information provided via
6823 * memblock_set_node().
6825 unsigned long __init find_min_pfn_with_active_regions(void)
6827 return find_min_pfn_for_node(MAX_NUMNODES);
6831 * early_calculate_totalpages()
6832 * Sum pages in active regions for movable zone.
6833 * Populate N_MEMORY for calculating usable_nodes.
6835 static unsigned long __init early_calculate_totalpages(void)
6837 unsigned long totalpages = 0;
6838 unsigned long start_pfn, end_pfn;
6841 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6842 unsigned long pages = end_pfn - start_pfn;
6844 totalpages += pages;
6846 node_set_state(nid, N_MEMORY);
6852 * Find the PFN the Movable zone begins in each node. Kernel memory
6853 * is spread evenly between nodes as long as the nodes have enough
6854 * memory. When they don't, some nodes will have more kernelcore than
6857 static void __init find_zone_movable_pfns_for_nodes(void)
6860 unsigned long usable_startpfn;
6861 unsigned long kernelcore_node, kernelcore_remaining;
6862 /* save the state before borrow the nodemask */
6863 nodemask_t saved_node_state = node_states[N_MEMORY];
6864 unsigned long totalpages = early_calculate_totalpages();
6865 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6866 struct memblock_region *r;
6868 /* Need to find movable_zone earlier when movable_node is specified. */
6869 find_usable_zone_for_movable();
6872 * If movable_node is specified, ignore kernelcore and movablecore
6875 if (movable_node_is_enabled()) {
6876 for_each_memblock(memory, r) {
6877 if (!memblock_is_hotpluggable(r))
6882 usable_startpfn = PFN_DOWN(r->base);
6883 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6884 min(usable_startpfn, zone_movable_pfn[nid]) :
6892 * If kernelcore=mirror is specified, ignore movablecore option
6894 if (mirrored_kernelcore) {
6895 bool mem_below_4gb_not_mirrored = false;
6897 for_each_memblock(memory, r) {
6898 if (memblock_is_mirror(r))
6903 usable_startpfn = memblock_region_memory_base_pfn(r);
6905 if (usable_startpfn < 0x100000) {
6906 mem_below_4gb_not_mirrored = true;
6910 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6911 min(usable_startpfn, zone_movable_pfn[nid]) :
6915 if (mem_below_4gb_not_mirrored)
6916 pr_warn("This configuration results in unmirrored kernel memory.");
6922 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
6923 * amount of necessary memory.
6925 if (required_kernelcore_percent)
6926 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
6928 if (required_movablecore_percent)
6929 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
6933 * If movablecore= was specified, calculate what size of
6934 * kernelcore that corresponds so that memory usable for
6935 * any allocation type is evenly spread. If both kernelcore
6936 * and movablecore are specified, then the value of kernelcore
6937 * will be used for required_kernelcore if it's greater than
6938 * what movablecore would have allowed.
6940 if (required_movablecore) {
6941 unsigned long corepages;
6944 * Round-up so that ZONE_MOVABLE is at least as large as what
6945 * was requested by the user
6947 required_movablecore =
6948 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6949 required_movablecore = min(totalpages, required_movablecore);
6950 corepages = totalpages - required_movablecore;
6952 required_kernelcore = max(required_kernelcore, corepages);
6956 * If kernelcore was not specified or kernelcore size is larger
6957 * than totalpages, there is no ZONE_MOVABLE.
6959 if (!required_kernelcore || required_kernelcore >= totalpages)
6962 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6963 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6966 /* Spread kernelcore memory as evenly as possible throughout nodes */
6967 kernelcore_node = required_kernelcore / usable_nodes;
6968 for_each_node_state(nid, N_MEMORY) {
6969 unsigned long start_pfn, end_pfn;
6972 * Recalculate kernelcore_node if the division per node
6973 * now exceeds what is necessary to satisfy the requested
6974 * amount of memory for the kernel
6976 if (required_kernelcore < kernelcore_node)
6977 kernelcore_node = required_kernelcore / usable_nodes;
6980 * As the map is walked, we track how much memory is usable
6981 * by the kernel using kernelcore_remaining. When it is
6982 * 0, the rest of the node is usable by ZONE_MOVABLE
6984 kernelcore_remaining = kernelcore_node;
6986 /* Go through each range of PFNs within this node */
6987 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6988 unsigned long size_pages;
6990 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6991 if (start_pfn >= end_pfn)
6994 /* Account for what is only usable for kernelcore */
6995 if (start_pfn < usable_startpfn) {
6996 unsigned long kernel_pages;
6997 kernel_pages = min(end_pfn, usable_startpfn)
7000 kernelcore_remaining -= min(kernel_pages,
7001 kernelcore_remaining);
7002 required_kernelcore -= min(kernel_pages,
7003 required_kernelcore);
7005 /* Continue if range is now fully accounted */
7006 if (end_pfn <= usable_startpfn) {
7009 * Push zone_movable_pfn to the end so
7010 * that if we have to rebalance
7011 * kernelcore across nodes, we will
7012 * not double account here
7014 zone_movable_pfn[nid] = end_pfn;
7017 start_pfn = usable_startpfn;
7021 * The usable PFN range for ZONE_MOVABLE is from
7022 * start_pfn->end_pfn. Calculate size_pages as the
7023 * number of pages used as kernelcore
7025 size_pages = end_pfn - start_pfn;
7026 if (size_pages > kernelcore_remaining)
7027 size_pages = kernelcore_remaining;
7028 zone_movable_pfn[nid] = start_pfn + size_pages;
7031 * Some kernelcore has been met, update counts and
7032 * break if the kernelcore for this node has been
7035 required_kernelcore -= min(required_kernelcore,
7037 kernelcore_remaining -= size_pages;
7038 if (!kernelcore_remaining)
7044 * If there is still required_kernelcore, we do another pass with one
7045 * less node in the count. This will push zone_movable_pfn[nid] further
7046 * along on the nodes that still have memory until kernelcore is
7050 if (usable_nodes && required_kernelcore > usable_nodes)
7054 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7055 for (nid = 0; nid < MAX_NUMNODES; nid++)
7056 zone_movable_pfn[nid] =
7057 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7060 /* restore the node_state */
7061 node_states[N_MEMORY] = saved_node_state;
7064 /* Any regular or high memory on that node ? */
7065 static void check_for_memory(pg_data_t *pgdat, int nid)
7067 enum zone_type zone_type;
7069 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7070 struct zone *zone = &pgdat->node_zones[zone_type];
7071 if (populated_zone(zone)) {
7072 if (IS_ENABLED(CONFIG_HIGHMEM))
7073 node_set_state(nid, N_HIGH_MEMORY);
7074 if (zone_type <= ZONE_NORMAL)
7075 node_set_state(nid, N_NORMAL_MEMORY);
7082 * free_area_init_nodes - Initialise all pg_data_t and zone data
7083 * @max_zone_pfn: an array of max PFNs for each zone
7085 * This will call free_area_init_node() for each active node in the system.
7086 * Using the page ranges provided by memblock_set_node(), the size of each
7087 * zone in each node and their holes is calculated. If the maximum PFN
7088 * between two adjacent zones match, it is assumed that the zone is empty.
7089 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7090 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7091 * starts where the previous one ended. For example, ZONE_DMA32 starts
7092 * at arch_max_dma_pfn.
7094 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
7096 unsigned long start_pfn, end_pfn;
7099 /* Record where the zone boundaries are */
7100 memset(arch_zone_lowest_possible_pfn, 0,
7101 sizeof(arch_zone_lowest_possible_pfn));
7102 memset(arch_zone_highest_possible_pfn, 0,
7103 sizeof(arch_zone_highest_possible_pfn));
7105 start_pfn = find_min_pfn_with_active_regions();
7107 for (i = 0; i < MAX_NR_ZONES; i++) {
7108 if (i == ZONE_MOVABLE)
7111 end_pfn = max(max_zone_pfn[i], start_pfn);
7112 arch_zone_lowest_possible_pfn[i] = start_pfn;
7113 arch_zone_highest_possible_pfn[i] = end_pfn;
7115 start_pfn = end_pfn;
7118 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7119 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7120 find_zone_movable_pfns_for_nodes();
7122 /* Print out the zone ranges */
7123 pr_info("Zone ranges:\n");
7124 for (i = 0; i < MAX_NR_ZONES; i++) {
7125 if (i == ZONE_MOVABLE)
7127 pr_info(" %-8s ", zone_names[i]);
7128 if (arch_zone_lowest_possible_pfn[i] ==
7129 arch_zone_highest_possible_pfn[i])
7132 pr_cont("[mem %#018Lx-%#018Lx]\n",
7133 (u64)arch_zone_lowest_possible_pfn[i]
7135 ((u64)arch_zone_highest_possible_pfn[i]
7136 << PAGE_SHIFT) - 1);
7139 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7140 pr_info("Movable zone start for each node\n");
7141 for (i = 0; i < MAX_NUMNODES; i++) {
7142 if (zone_movable_pfn[i])
7143 pr_info(" Node %d: %#018Lx\n", i,
7144 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7147 /* Print out the early node map */
7148 pr_info("Early memory node ranges\n");
7149 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
7150 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7151 (u64)start_pfn << PAGE_SHIFT,
7152 ((u64)end_pfn << PAGE_SHIFT) - 1);
7154 /* Initialise every node */
7155 mminit_verify_pageflags_layout();
7156 setup_nr_node_ids();
7157 zero_resv_unavail();
7158 for_each_online_node(nid) {
7159 pg_data_t *pgdat = NODE_DATA(nid);
7160 free_area_init_node(nid, NULL,
7161 find_min_pfn_for_node(nid), NULL);
7163 /* Any memory on that node */
7164 if (pgdat->node_present_pages)
7165 node_set_state(nid, N_MEMORY);
7166 check_for_memory(pgdat, nid);
7170 static int __init cmdline_parse_core(char *p, unsigned long *core,
7171 unsigned long *percent)
7173 unsigned long long coremem;
7179 /* Value may be a percentage of total memory, otherwise bytes */
7180 coremem = simple_strtoull(p, &endptr, 0);
7181 if (*endptr == '%') {
7182 /* Paranoid check for percent values greater than 100 */
7183 WARN_ON(coremem > 100);
7187 coremem = memparse(p, &p);
7188 /* Paranoid check that UL is enough for the coremem value */
7189 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7191 *core = coremem >> PAGE_SHIFT;
7198 * kernelcore=size sets the amount of memory for use for allocations that
7199 * cannot be reclaimed or migrated.
7201 static int __init cmdline_parse_kernelcore(char *p)
7203 /* parse kernelcore=mirror */
7204 if (parse_option_str(p, "mirror")) {
7205 mirrored_kernelcore = true;
7209 return cmdline_parse_core(p, &required_kernelcore,
7210 &required_kernelcore_percent);
7214 * movablecore=size sets the amount of memory for use for allocations that
7215 * can be reclaimed or migrated.
7217 static int __init cmdline_parse_movablecore(char *p)
7219 return cmdline_parse_core(p, &required_movablecore,
7220 &required_movablecore_percent);
7223 early_param("kernelcore", cmdline_parse_kernelcore);
7224 early_param("movablecore", cmdline_parse_movablecore);
7226 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
7228 void adjust_managed_page_count(struct page *page, long count)
7230 atomic_long_add(count, &page_zone(page)->managed_pages);
7231 totalram_pages_add(count);
7232 #ifdef CONFIG_HIGHMEM
7233 if (PageHighMem(page))
7234 totalhigh_pages_add(count);
7237 EXPORT_SYMBOL(adjust_managed_page_count);
7239 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7242 unsigned long pages = 0;
7244 start = (void *)PAGE_ALIGN((unsigned long)start);
7245 end = (void *)((unsigned long)end & PAGE_MASK);
7246 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7247 struct page *page = virt_to_page(pos);
7248 void *direct_map_addr;
7251 * 'direct_map_addr' might be different from 'pos'
7252 * because some architectures' virt_to_page()
7253 * work with aliases. Getting the direct map
7254 * address ensures that we get a _writeable_
7255 * alias for the memset().
7257 direct_map_addr = page_address(page);
7258 if ((unsigned int)poison <= 0xFF)
7259 memset(direct_map_addr, poison, PAGE_SIZE);
7261 free_reserved_page(page);
7265 pr_info("Freeing %s memory: %ldK\n",
7266 s, pages << (PAGE_SHIFT - 10));
7270 EXPORT_SYMBOL(free_reserved_area);
7272 #ifdef CONFIG_HIGHMEM
7273 void free_highmem_page(struct page *page)
7275 __free_reserved_page(page);
7276 totalram_pages_inc();
7277 atomic_long_inc(&page_zone(page)->managed_pages);
7278 totalhigh_pages_inc();
7283 void __init mem_init_print_info(const char *str)
7285 unsigned long physpages, codesize, datasize, rosize, bss_size;
7286 unsigned long init_code_size, init_data_size;
7288 physpages = get_num_physpages();
7289 codesize = _etext - _stext;
7290 datasize = _edata - _sdata;
7291 rosize = __end_rodata - __start_rodata;
7292 bss_size = __bss_stop - __bss_start;
7293 init_data_size = __init_end - __init_begin;
7294 init_code_size = _einittext - _sinittext;
7297 * Detect special cases and adjust section sizes accordingly:
7298 * 1) .init.* may be embedded into .data sections
7299 * 2) .init.text.* may be out of [__init_begin, __init_end],
7300 * please refer to arch/tile/kernel/vmlinux.lds.S.
7301 * 3) .rodata.* may be embedded into .text or .data sections.
7303 #define adj_init_size(start, end, size, pos, adj) \
7305 if (start <= pos && pos < end && size > adj) \
7309 adj_init_size(__init_begin, __init_end, init_data_size,
7310 _sinittext, init_code_size);
7311 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7312 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7313 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7314 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7316 #undef adj_init_size
7318 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7319 #ifdef CONFIG_HIGHMEM
7323 nr_free_pages() << (PAGE_SHIFT - 10),
7324 physpages << (PAGE_SHIFT - 10),
7325 codesize >> 10, datasize >> 10, rosize >> 10,
7326 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7327 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7328 totalcma_pages << (PAGE_SHIFT - 10),
7329 #ifdef CONFIG_HIGHMEM
7330 totalhigh_pages() << (PAGE_SHIFT - 10),
7332 str ? ", " : "", str ? str : "");
7336 * set_dma_reserve - set the specified number of pages reserved in the first zone
7337 * @new_dma_reserve: The number of pages to mark reserved
7339 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7340 * In the DMA zone, a significant percentage may be consumed by kernel image
7341 * and other unfreeable allocations which can skew the watermarks badly. This
7342 * function may optionally be used to account for unfreeable pages in the
7343 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7344 * smaller per-cpu batchsize.
7346 void __init set_dma_reserve(unsigned long new_dma_reserve)
7348 dma_reserve = new_dma_reserve;
7351 void __init free_area_init(unsigned long *zones_size)
7353 zero_resv_unavail();
7354 free_area_init_node(0, zones_size,
7355 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
7358 static int page_alloc_cpu_dead(unsigned int cpu)
7361 lru_add_drain_cpu(cpu);
7365 * Spill the event counters of the dead processor
7366 * into the current processors event counters.
7367 * This artificially elevates the count of the current
7370 vm_events_fold_cpu(cpu);
7373 * Zero the differential counters of the dead processor
7374 * so that the vm statistics are consistent.
7376 * This is only okay since the processor is dead and cannot
7377 * race with what we are doing.
7379 cpu_vm_stats_fold(cpu);
7383 void __init page_alloc_init(void)
7387 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7388 "mm/page_alloc:dead", NULL,
7389 page_alloc_cpu_dead);
7394 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7395 * or min_free_kbytes changes.
7397 static void calculate_totalreserve_pages(void)
7399 struct pglist_data *pgdat;
7400 unsigned long reserve_pages = 0;
7401 enum zone_type i, j;
7403 for_each_online_pgdat(pgdat) {
7405 pgdat->totalreserve_pages = 0;
7407 for (i = 0; i < MAX_NR_ZONES; i++) {
7408 struct zone *zone = pgdat->node_zones + i;
7410 unsigned long managed_pages = zone_managed_pages(zone);
7412 /* Find valid and maximum lowmem_reserve in the zone */
7413 for (j = i; j < MAX_NR_ZONES; j++) {
7414 if (zone->lowmem_reserve[j] > max)
7415 max = zone->lowmem_reserve[j];
7418 /* we treat the high watermark as reserved pages. */
7419 max += high_wmark_pages(zone);
7421 if (max > managed_pages)
7422 max = managed_pages;
7424 pgdat->totalreserve_pages += max;
7426 reserve_pages += max;
7429 totalreserve_pages = reserve_pages;
7433 * setup_per_zone_lowmem_reserve - called whenever
7434 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7435 * has a correct pages reserved value, so an adequate number of
7436 * pages are left in the zone after a successful __alloc_pages().
7438 static void setup_per_zone_lowmem_reserve(void)
7440 struct pglist_data *pgdat;
7441 enum zone_type j, idx;
7443 for_each_online_pgdat(pgdat) {
7444 for (j = 0; j < MAX_NR_ZONES; j++) {
7445 struct zone *zone = pgdat->node_zones + j;
7446 unsigned long managed_pages = zone_managed_pages(zone);
7448 zone->lowmem_reserve[j] = 0;
7452 struct zone *lower_zone;
7455 lower_zone = pgdat->node_zones + idx;
7457 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7458 sysctl_lowmem_reserve_ratio[idx] = 0;
7459 lower_zone->lowmem_reserve[j] = 0;
7461 lower_zone->lowmem_reserve[j] =
7462 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7464 managed_pages += zone_managed_pages(lower_zone);
7469 /* update totalreserve_pages */
7470 calculate_totalreserve_pages();
7473 static void __setup_per_zone_wmarks(void)
7475 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7476 unsigned long lowmem_pages = 0;
7478 unsigned long flags;
7480 /* Calculate total number of !ZONE_HIGHMEM pages */
7481 for_each_zone(zone) {
7482 if (!is_highmem(zone))
7483 lowmem_pages += zone_managed_pages(zone);
7486 for_each_zone(zone) {
7489 spin_lock_irqsave(&zone->lock, flags);
7490 tmp = (u64)pages_min * zone_managed_pages(zone);
7491 do_div(tmp, lowmem_pages);
7492 if (is_highmem(zone)) {
7494 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7495 * need highmem pages, so cap pages_min to a small
7498 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7499 * deltas control asynch page reclaim, and so should
7500 * not be capped for highmem.
7502 unsigned long min_pages;
7504 min_pages = zone_managed_pages(zone) / 1024;
7505 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7506 zone->_watermark[WMARK_MIN] = min_pages;
7509 * If it's a lowmem zone, reserve a number of pages
7510 * proportionate to the zone's size.
7512 zone->_watermark[WMARK_MIN] = tmp;
7516 * Set the kswapd watermarks distance according to the
7517 * scale factor in proportion to available memory, but
7518 * ensure a minimum size on small systems.
7520 tmp = max_t(u64, tmp >> 2,
7521 mult_frac(zone_managed_pages(zone),
7522 watermark_scale_factor, 10000));
7524 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7525 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7526 zone->watermark_boost = 0;
7528 spin_unlock_irqrestore(&zone->lock, flags);
7531 /* update totalreserve_pages */
7532 calculate_totalreserve_pages();
7536 * setup_per_zone_wmarks - called when min_free_kbytes changes
7537 * or when memory is hot-{added|removed}
7539 * Ensures that the watermark[min,low,high] values for each zone are set
7540 * correctly with respect to min_free_kbytes.
7542 void setup_per_zone_wmarks(void)
7544 static DEFINE_SPINLOCK(lock);
7547 __setup_per_zone_wmarks();
7552 * Initialise min_free_kbytes.
7554 * For small machines we want it small (128k min). For large machines
7555 * we want it large (64MB max). But it is not linear, because network
7556 * bandwidth does not increase linearly with machine size. We use
7558 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7559 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7575 int __meminit init_per_zone_wmark_min(void)
7577 unsigned long lowmem_kbytes;
7578 int new_min_free_kbytes;
7580 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7581 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7583 if (new_min_free_kbytes > user_min_free_kbytes) {
7584 min_free_kbytes = new_min_free_kbytes;
7585 if (min_free_kbytes < 128)
7586 min_free_kbytes = 128;
7587 if (min_free_kbytes > 65536)
7588 min_free_kbytes = 65536;
7590 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7591 new_min_free_kbytes, user_min_free_kbytes);
7593 setup_per_zone_wmarks();
7594 refresh_zone_stat_thresholds();
7595 setup_per_zone_lowmem_reserve();
7598 setup_min_unmapped_ratio();
7599 setup_min_slab_ratio();
7604 core_initcall(init_per_zone_wmark_min)
7607 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7608 * that we can call two helper functions whenever min_free_kbytes
7611 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7612 void __user *buffer, size_t *length, loff_t *ppos)
7616 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7621 user_min_free_kbytes = min_free_kbytes;
7622 setup_per_zone_wmarks();
7627 int watermark_boost_factor_sysctl_handler(struct ctl_table *table, int write,
7628 void __user *buffer, size_t *length, loff_t *ppos)
7632 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7639 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7640 void __user *buffer, size_t *length, loff_t *ppos)
7644 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7649 setup_per_zone_wmarks();
7655 static void setup_min_unmapped_ratio(void)
7660 for_each_online_pgdat(pgdat)
7661 pgdat->min_unmapped_pages = 0;
7664 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7665 sysctl_min_unmapped_ratio) / 100;
7669 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7670 void __user *buffer, size_t *length, loff_t *ppos)
7674 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7678 setup_min_unmapped_ratio();
7683 static void setup_min_slab_ratio(void)
7688 for_each_online_pgdat(pgdat)
7689 pgdat->min_slab_pages = 0;
7692 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7693 sysctl_min_slab_ratio) / 100;
7696 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7697 void __user *buffer, size_t *length, loff_t *ppos)
7701 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7705 setup_min_slab_ratio();
7712 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7713 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7714 * whenever sysctl_lowmem_reserve_ratio changes.
7716 * The reserve ratio obviously has absolutely no relation with the
7717 * minimum watermarks. The lowmem reserve ratio can only make sense
7718 * if in function of the boot time zone sizes.
7720 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7721 void __user *buffer, size_t *length, loff_t *ppos)
7723 proc_dointvec_minmax(table, write, buffer, length, ppos);
7724 setup_per_zone_lowmem_reserve();
7729 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7730 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7731 * pagelist can have before it gets flushed back to buddy allocator.
7733 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7734 void __user *buffer, size_t *length, loff_t *ppos)
7737 int old_percpu_pagelist_fraction;
7740 mutex_lock(&pcp_batch_high_lock);
7741 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7743 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7744 if (!write || ret < 0)
7747 /* Sanity checking to avoid pcp imbalance */
7748 if (percpu_pagelist_fraction &&
7749 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7750 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7756 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7759 for_each_populated_zone(zone) {
7762 for_each_possible_cpu(cpu)
7763 pageset_set_high_and_batch(zone,
7764 per_cpu_ptr(zone->pageset, cpu));
7767 mutex_unlock(&pcp_batch_high_lock);
7772 int hashdist = HASHDIST_DEFAULT;
7774 static int __init set_hashdist(char *str)
7778 hashdist = simple_strtoul(str, &str, 0);
7781 __setup("hashdist=", set_hashdist);
7784 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7786 * Returns the number of pages that arch has reserved but
7787 * is not known to alloc_large_system_hash().
7789 static unsigned long __init arch_reserved_kernel_pages(void)
7796 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7797 * machines. As memory size is increased the scale is also increased but at
7798 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7799 * quadruples the scale is increased by one, which means the size of hash table
7800 * only doubles, instead of quadrupling as well.
7801 * Because 32-bit systems cannot have large physical memory, where this scaling
7802 * makes sense, it is disabled on such platforms.
7804 #if __BITS_PER_LONG > 32
7805 #define ADAPT_SCALE_BASE (64ul << 30)
7806 #define ADAPT_SCALE_SHIFT 2
7807 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7811 * allocate a large system hash table from bootmem
7812 * - it is assumed that the hash table must contain an exact power-of-2
7813 * quantity of entries
7814 * - limit is the number of hash buckets, not the total allocation size
7816 void *__init alloc_large_system_hash(const char *tablename,
7817 unsigned long bucketsize,
7818 unsigned long numentries,
7821 unsigned int *_hash_shift,
7822 unsigned int *_hash_mask,
7823 unsigned long low_limit,
7824 unsigned long high_limit)
7826 unsigned long long max = high_limit;
7827 unsigned long log2qty, size;
7831 /* allow the kernel cmdline to have a say */
7833 /* round applicable memory size up to nearest megabyte */
7834 numentries = nr_kernel_pages;
7835 numentries -= arch_reserved_kernel_pages();
7837 /* It isn't necessary when PAGE_SIZE >= 1MB */
7838 if (PAGE_SHIFT < 20)
7839 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7841 #if __BITS_PER_LONG > 32
7843 unsigned long adapt;
7845 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7846 adapt <<= ADAPT_SCALE_SHIFT)
7851 /* limit to 1 bucket per 2^scale bytes of low memory */
7852 if (scale > PAGE_SHIFT)
7853 numentries >>= (scale - PAGE_SHIFT);
7855 numentries <<= (PAGE_SHIFT - scale);
7857 /* Make sure we've got at least a 0-order allocation.. */
7858 if (unlikely(flags & HASH_SMALL)) {
7859 /* Makes no sense without HASH_EARLY */
7860 WARN_ON(!(flags & HASH_EARLY));
7861 if (!(numentries >> *_hash_shift)) {
7862 numentries = 1UL << *_hash_shift;
7863 BUG_ON(!numentries);
7865 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7866 numentries = PAGE_SIZE / bucketsize;
7868 numentries = roundup_pow_of_two(numentries);
7870 /* limit allocation size to 1/16 total memory by default */
7872 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7873 do_div(max, bucketsize);
7875 max = min(max, 0x80000000ULL);
7877 if (numentries < low_limit)
7878 numentries = low_limit;
7879 if (numentries > max)
7882 log2qty = ilog2(numentries);
7884 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
7886 size = bucketsize << log2qty;
7887 if (flags & HASH_EARLY) {
7888 if (flags & HASH_ZERO)
7889 table = memblock_alloc_nopanic(size,
7892 table = memblock_alloc_raw(size,
7894 } else if (hashdist) {
7895 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
7898 * If bucketsize is not a power-of-two, we may free
7899 * some pages at the end of hash table which
7900 * alloc_pages_exact() automatically does
7902 if (get_order(size) < MAX_ORDER) {
7903 table = alloc_pages_exact(size, gfp_flags);
7904 kmemleak_alloc(table, size, 1, gfp_flags);
7907 } while (!table && size > PAGE_SIZE && --log2qty);
7910 panic("Failed to allocate %s hash table\n", tablename);
7912 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7913 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7916 *_hash_shift = log2qty;
7918 *_hash_mask = (1 << log2qty) - 1;
7924 * This function checks whether pageblock includes unmovable pages or not.
7925 * If @count is not zero, it is okay to include less @count unmovable pages
7927 * PageLRU check without isolation or lru_lock could race so that
7928 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7929 * check without lock_page also may miss some movable non-lru pages at
7930 * race condition. So you can't expect this function should be exact.
7932 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7933 int migratetype, int flags)
7935 unsigned long pfn, iter, found;
7938 * TODO we could make this much more efficient by not checking every
7939 * page in the range if we know all of them are in MOVABLE_ZONE and
7940 * that the movable zone guarantees that pages are migratable but
7941 * the later is not the case right now unfortunatelly. E.g. movablecore
7942 * can still lead to having bootmem allocations in zone_movable.
7946 * CMA allocations (alloc_contig_range) really need to mark isolate
7947 * CMA pageblocks even when they are not movable in fact so consider
7948 * them movable here.
7950 if (is_migrate_cma(migratetype) &&
7951 is_migrate_cma(get_pageblock_migratetype(page)))
7954 pfn = page_to_pfn(page);
7955 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7956 unsigned long check = pfn + iter;
7958 if (!pfn_valid_within(check))
7961 page = pfn_to_page(check);
7963 if (PageReserved(page))
7967 * If the zone is movable and we have ruled out all reserved
7968 * pages then it should be reasonably safe to assume the rest
7971 if (zone_idx(zone) == ZONE_MOVABLE)
7975 * Hugepages are not in LRU lists, but they're movable.
7976 * We need not scan over tail pages bacause we don't
7977 * handle each tail page individually in migration.
7979 if (PageHuge(page)) {
7980 struct page *head = compound_head(page);
7981 unsigned int skip_pages;
7983 if (!hugepage_migration_supported(page_hstate(head)))
7986 skip_pages = (1 << compound_order(head)) - (page - head);
7987 iter += skip_pages - 1;
7992 * We can't use page_count without pin a page
7993 * because another CPU can free compound page.
7994 * This check already skips compound tails of THP
7995 * because their page->_refcount is zero at all time.
7997 if (!page_ref_count(page)) {
7998 if (PageBuddy(page))
7999 iter += (1 << page_order(page)) - 1;
8004 * The HWPoisoned page may be not in buddy system, and
8005 * page_count() is not 0.
8007 if ((flags & SKIP_HWPOISON) && PageHWPoison(page))
8010 if (__PageMovable(page))
8016 * If there are RECLAIMABLE pages, we need to check
8017 * it. But now, memory offline itself doesn't call
8018 * shrink_node_slabs() and it still to be fixed.
8021 * If the page is not RAM, page_count()should be 0.
8022 * we don't need more check. This is an _used_ not-movable page.
8024 * The problematic thing here is PG_reserved pages. PG_reserved
8025 * is set to both of a memory hole page and a _used_ kernel
8033 WARN_ON_ONCE(zone_idx(zone) == ZONE_MOVABLE);
8034 if (flags & REPORT_FAILURE)
8035 dump_page(pfn_to_page(pfn+iter), "unmovable page");
8039 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
8041 static unsigned long pfn_max_align_down(unsigned long pfn)
8043 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8044 pageblock_nr_pages) - 1);
8047 static unsigned long pfn_max_align_up(unsigned long pfn)
8049 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8050 pageblock_nr_pages));
8053 /* [start, end) must belong to a single zone. */
8054 static int __alloc_contig_migrate_range(struct compact_control *cc,
8055 unsigned long start, unsigned long end)
8057 /* This function is based on compact_zone() from compaction.c. */
8058 unsigned long nr_reclaimed;
8059 unsigned long pfn = start;
8060 unsigned int tries = 0;
8065 while (pfn < end || !list_empty(&cc->migratepages)) {
8066 if (fatal_signal_pending(current)) {
8071 if (list_empty(&cc->migratepages)) {
8072 cc->nr_migratepages = 0;
8073 pfn = isolate_migratepages_range(cc, pfn, end);
8079 } else if (++tries == 5) {
8080 ret = ret < 0 ? ret : -EBUSY;
8084 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8086 cc->nr_migratepages -= nr_reclaimed;
8088 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
8089 NULL, 0, cc->mode, MR_CONTIG_RANGE);
8092 putback_movable_pages(&cc->migratepages);
8099 * alloc_contig_range() -- tries to allocate given range of pages
8100 * @start: start PFN to allocate
8101 * @end: one-past-the-last PFN to allocate
8102 * @migratetype: migratetype of the underlaying pageblocks (either
8103 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8104 * in range must have the same migratetype and it must
8105 * be either of the two.
8106 * @gfp_mask: GFP mask to use during compaction
8108 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8109 * aligned. The PFN range must belong to a single zone.
8111 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8112 * pageblocks in the range. Once isolated, the pageblocks should not
8113 * be modified by others.
8115 * Returns zero on success or negative error code. On success all
8116 * pages which PFN is in [start, end) are allocated for the caller and
8117 * need to be freed with free_contig_range().
8119 int alloc_contig_range(unsigned long start, unsigned long end,
8120 unsigned migratetype, gfp_t gfp_mask)
8122 unsigned long outer_start, outer_end;
8126 struct compact_control cc = {
8127 .nr_migratepages = 0,
8129 .zone = page_zone(pfn_to_page(start)),
8130 .mode = MIGRATE_SYNC,
8131 .ignore_skip_hint = true,
8132 .no_set_skip_hint = true,
8133 .gfp_mask = current_gfp_context(gfp_mask),
8135 INIT_LIST_HEAD(&cc.migratepages);
8138 * What we do here is we mark all pageblocks in range as
8139 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8140 * have different sizes, and due to the way page allocator
8141 * work, we align the range to biggest of the two pages so
8142 * that page allocator won't try to merge buddies from
8143 * different pageblocks and change MIGRATE_ISOLATE to some
8144 * other migration type.
8146 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8147 * migrate the pages from an unaligned range (ie. pages that
8148 * we are interested in). This will put all the pages in
8149 * range back to page allocator as MIGRATE_ISOLATE.
8151 * When this is done, we take the pages in range from page
8152 * allocator removing them from the buddy system. This way
8153 * page allocator will never consider using them.
8155 * This lets us mark the pageblocks back as
8156 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8157 * aligned range but not in the unaligned, original range are
8158 * put back to page allocator so that buddy can use them.
8161 ret = start_isolate_page_range(pfn_max_align_down(start),
8162 pfn_max_align_up(end), migratetype, 0);
8167 * In case of -EBUSY, we'd like to know which page causes problem.
8168 * So, just fall through. test_pages_isolated() has a tracepoint
8169 * which will report the busy page.
8171 * It is possible that busy pages could become available before
8172 * the call to test_pages_isolated, and the range will actually be
8173 * allocated. So, if we fall through be sure to clear ret so that
8174 * -EBUSY is not accidentally used or returned to caller.
8176 ret = __alloc_contig_migrate_range(&cc, start, end);
8177 if (ret && ret != -EBUSY)
8182 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8183 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8184 * more, all pages in [start, end) are free in page allocator.
8185 * What we are going to do is to allocate all pages from
8186 * [start, end) (that is remove them from page allocator).
8188 * The only problem is that pages at the beginning and at the
8189 * end of interesting range may be not aligned with pages that
8190 * page allocator holds, ie. they can be part of higher order
8191 * pages. Because of this, we reserve the bigger range and
8192 * once this is done free the pages we are not interested in.
8194 * We don't have to hold zone->lock here because the pages are
8195 * isolated thus they won't get removed from buddy.
8198 lru_add_drain_all();
8199 drain_all_pages(cc.zone);
8202 outer_start = start;
8203 while (!PageBuddy(pfn_to_page(outer_start))) {
8204 if (++order >= MAX_ORDER) {
8205 outer_start = start;
8208 outer_start &= ~0UL << order;
8211 if (outer_start != start) {
8212 order = page_order(pfn_to_page(outer_start));
8215 * outer_start page could be small order buddy page and
8216 * it doesn't include start page. Adjust outer_start
8217 * in this case to report failed page properly
8218 * on tracepoint in test_pages_isolated()
8220 if (outer_start + (1UL << order) <= start)
8221 outer_start = start;
8224 /* Make sure the range is really isolated. */
8225 if (test_pages_isolated(outer_start, end, false)) {
8226 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8227 __func__, outer_start, end);
8232 /* Grab isolated pages from freelists. */
8233 outer_end = isolate_freepages_range(&cc, outer_start, end);
8239 /* Free head and tail (if any) */
8240 if (start != outer_start)
8241 free_contig_range(outer_start, start - outer_start);
8242 if (end != outer_end)
8243 free_contig_range(end, outer_end - end);
8246 undo_isolate_page_range(pfn_max_align_down(start),
8247 pfn_max_align_up(end), migratetype);
8251 void free_contig_range(unsigned long pfn, unsigned nr_pages)
8253 unsigned int count = 0;
8255 for (; nr_pages--; pfn++) {
8256 struct page *page = pfn_to_page(pfn);
8258 count += page_count(page) != 1;
8261 WARN(count != 0, "%d pages are still in use!\n", count);
8265 #ifdef CONFIG_MEMORY_HOTPLUG
8267 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8268 * page high values need to be recalulated.
8270 void __meminit zone_pcp_update(struct zone *zone)
8273 mutex_lock(&pcp_batch_high_lock);
8274 for_each_possible_cpu(cpu)
8275 pageset_set_high_and_batch(zone,
8276 per_cpu_ptr(zone->pageset, cpu));
8277 mutex_unlock(&pcp_batch_high_lock);
8281 void zone_pcp_reset(struct zone *zone)
8283 unsigned long flags;
8285 struct per_cpu_pageset *pset;
8287 /* avoid races with drain_pages() */
8288 local_irq_save(flags);
8289 if (zone->pageset != &boot_pageset) {
8290 for_each_online_cpu(cpu) {
8291 pset = per_cpu_ptr(zone->pageset, cpu);
8292 drain_zonestat(zone, pset);
8294 free_percpu(zone->pageset);
8295 zone->pageset = &boot_pageset;
8297 local_irq_restore(flags);
8300 #ifdef CONFIG_MEMORY_HOTREMOVE
8302 * All pages in the range must be in a single zone and isolated
8303 * before calling this.
8306 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8310 unsigned int order, i;
8312 unsigned long flags;
8313 /* find the first valid pfn */
8314 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8319 offline_mem_sections(pfn, end_pfn);
8320 zone = page_zone(pfn_to_page(pfn));
8321 spin_lock_irqsave(&zone->lock, flags);
8323 while (pfn < end_pfn) {
8324 if (!pfn_valid(pfn)) {
8328 page = pfn_to_page(pfn);
8330 * The HWPoisoned page may be not in buddy system, and
8331 * page_count() is not 0.
8333 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8335 SetPageReserved(page);
8339 BUG_ON(page_count(page));
8340 BUG_ON(!PageBuddy(page));
8341 order = page_order(page);
8342 #ifdef CONFIG_DEBUG_VM
8343 pr_info("remove from free list %lx %d %lx\n",
8344 pfn, 1 << order, end_pfn);
8346 list_del(&page->lru);
8347 rmv_page_order(page);
8348 zone->free_area[order].nr_free--;
8349 for (i = 0; i < (1 << order); i++)
8350 SetPageReserved((page+i));
8351 pfn += (1 << order);
8353 spin_unlock_irqrestore(&zone->lock, flags);
8357 bool is_free_buddy_page(struct page *page)
8359 struct zone *zone = page_zone(page);
8360 unsigned long pfn = page_to_pfn(page);
8361 unsigned long flags;
8364 spin_lock_irqsave(&zone->lock, flags);
8365 for (order = 0; order < MAX_ORDER; order++) {
8366 struct page *page_head = page - (pfn & ((1 << order) - 1));
8368 if (PageBuddy(page_head) && page_order(page_head) >= order)
8371 spin_unlock_irqrestore(&zone->lock, flags);
8373 return order < MAX_ORDER;
8376 #ifdef CONFIG_MEMORY_FAILURE
8378 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8379 * test is performed under the zone lock to prevent a race against page
8382 bool set_hwpoison_free_buddy_page(struct page *page)
8384 struct zone *zone = page_zone(page);
8385 unsigned long pfn = page_to_pfn(page);
8386 unsigned long flags;
8388 bool hwpoisoned = false;
8390 spin_lock_irqsave(&zone->lock, flags);
8391 for (order = 0; order < MAX_ORDER; order++) {
8392 struct page *page_head = page - (pfn & ((1 << order) - 1));
8394 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8395 if (!TestSetPageHWPoison(page))
8400 spin_unlock_irqrestore(&zone->lock, flags);