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 */
100 DEFINE_MUTEX(pcpu_drain_mutex);
101 DEFINE_PER_CPU(struct work_struct, pcpu_drain);
103 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
104 volatile unsigned long latent_entropy __latent_entropy;
105 EXPORT_SYMBOL(latent_entropy);
109 * Array of node states.
111 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
112 [N_POSSIBLE] = NODE_MASK_ALL,
113 [N_ONLINE] = { { [0] = 1UL } },
115 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
116 #ifdef CONFIG_HIGHMEM
117 [N_HIGH_MEMORY] = { { [0] = 1UL } },
119 [N_MEMORY] = { { [0] = 1UL } },
120 [N_CPU] = { { [0] = 1UL } },
123 EXPORT_SYMBOL(node_states);
125 atomic_long_t _totalram_pages __read_mostly;
126 EXPORT_SYMBOL(_totalram_pages);
127 unsigned long totalreserve_pages __read_mostly;
128 unsigned long totalcma_pages __read_mostly;
130 int percpu_pagelist_fraction;
131 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
134 * A cached value of the page's pageblock's migratetype, used when the page is
135 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
136 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
137 * Also the migratetype set in the page does not necessarily match the pcplist
138 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
139 * other index - this ensures that it will be put on the correct CMA freelist.
141 static inline int get_pcppage_migratetype(struct page *page)
146 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
148 page->index = migratetype;
151 #ifdef CONFIG_PM_SLEEP
153 * The following functions are used by the suspend/hibernate code to temporarily
154 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
155 * while devices are suspended. To avoid races with the suspend/hibernate code,
156 * they should always be called with system_transition_mutex held
157 * (gfp_allowed_mask also should only be modified with system_transition_mutex
158 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
159 * with that modification).
162 static gfp_t saved_gfp_mask;
164 void pm_restore_gfp_mask(void)
166 WARN_ON(!mutex_is_locked(&system_transition_mutex));
167 if (saved_gfp_mask) {
168 gfp_allowed_mask = saved_gfp_mask;
173 void pm_restrict_gfp_mask(void)
175 WARN_ON(!mutex_is_locked(&system_transition_mutex));
176 WARN_ON(saved_gfp_mask);
177 saved_gfp_mask = gfp_allowed_mask;
178 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
181 bool pm_suspended_storage(void)
183 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
187 #endif /* CONFIG_PM_SLEEP */
189 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
190 unsigned int pageblock_order __read_mostly;
193 static void __free_pages_ok(struct page *page, unsigned int order);
196 * results with 256, 32 in the lowmem_reserve sysctl:
197 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
198 * 1G machine -> (16M dma, 784M normal, 224M high)
199 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
200 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
201 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
203 * TBD: should special case ZONE_DMA32 machines here - in those we normally
204 * don't need any ZONE_NORMAL reservation
206 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
207 #ifdef CONFIG_ZONE_DMA
210 #ifdef CONFIG_ZONE_DMA32
214 #ifdef CONFIG_HIGHMEM
220 EXPORT_SYMBOL(totalram_pages);
222 static char * const zone_names[MAX_NR_ZONES] = {
223 #ifdef CONFIG_ZONE_DMA
226 #ifdef CONFIG_ZONE_DMA32
230 #ifdef CONFIG_HIGHMEM
234 #ifdef CONFIG_ZONE_DEVICE
239 const char * const migratetype_names[MIGRATE_TYPES] = {
247 #ifdef CONFIG_MEMORY_ISOLATION
252 compound_page_dtor * const compound_page_dtors[] = {
255 #ifdef CONFIG_HUGETLB_PAGE
258 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
263 int min_free_kbytes = 1024;
264 int user_min_free_kbytes = -1;
265 int watermark_boost_factor __read_mostly = 15000;
266 int watermark_scale_factor = 10;
268 static unsigned long nr_kernel_pages __initdata;
269 static unsigned long nr_all_pages __initdata;
270 static unsigned long dma_reserve __initdata;
272 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
273 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
274 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
275 static unsigned long required_kernelcore __initdata;
276 static unsigned long required_kernelcore_percent __initdata;
277 static unsigned long required_movablecore __initdata;
278 static unsigned long required_movablecore_percent __initdata;
279 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
280 static bool mirrored_kernelcore __meminitdata;
282 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
284 EXPORT_SYMBOL(movable_zone);
285 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
288 int nr_node_ids __read_mostly = MAX_NUMNODES;
289 int nr_online_nodes __read_mostly = 1;
290 EXPORT_SYMBOL(nr_node_ids);
291 EXPORT_SYMBOL(nr_online_nodes);
294 int page_group_by_mobility_disabled __read_mostly;
296 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
298 * During boot we initialize deferred pages on-demand, as needed, but once
299 * page_alloc_init_late() has finished, the deferred pages are all initialized,
300 * and we can permanently disable that path.
302 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
305 * Calling kasan_free_pages() only after deferred memory initialization
306 * has completed. Poisoning pages during deferred memory init will greatly
307 * lengthen the process and cause problem in large memory systems as the
308 * deferred pages initialization is done with interrupt disabled.
310 * Assuming that there will be no reference to those newly initialized
311 * pages before they are ever allocated, this should have no effect on
312 * KASAN memory tracking as the poison will be properly inserted at page
313 * allocation time. The only corner case is when pages are allocated by
314 * on-demand allocation and then freed again before the deferred pages
315 * initialization is done, but this is not likely to happen.
317 static inline void kasan_free_nondeferred_pages(struct page *page, int order)
319 if (!static_branch_unlikely(&deferred_pages))
320 kasan_free_pages(page, order);
323 /* Returns true if the struct page for the pfn is uninitialised */
324 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
326 int nid = early_pfn_to_nid(pfn);
328 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
335 * Returns true when the remaining initialisation should be deferred until
336 * later in the boot cycle when it can be parallelised.
338 static bool __meminit
339 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
341 static unsigned long prev_end_pfn, nr_initialised;
344 * prev_end_pfn static that contains the end of previous zone
345 * No need to protect because called very early in boot before smp_init.
347 if (prev_end_pfn != end_pfn) {
348 prev_end_pfn = end_pfn;
352 /* Always populate low zones for address-constrained allocations */
353 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
357 * We start only with one section of pages, more pages are added as
358 * needed until the rest of deferred pages are initialized.
361 if ((nr_initialised > PAGES_PER_SECTION) &&
362 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
363 NODE_DATA(nid)->first_deferred_pfn = pfn;
369 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
371 static inline bool early_page_uninitialised(unsigned long pfn)
376 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
382 /* Return a pointer to the bitmap storing bits affecting a block of pages */
383 static inline unsigned long *get_pageblock_bitmap(struct page *page,
386 #ifdef CONFIG_SPARSEMEM
387 return __pfn_to_section(pfn)->pageblock_flags;
389 return page_zone(page)->pageblock_flags;
390 #endif /* CONFIG_SPARSEMEM */
393 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
395 #ifdef CONFIG_SPARSEMEM
396 pfn &= (PAGES_PER_SECTION-1);
397 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
399 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
400 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
401 #endif /* CONFIG_SPARSEMEM */
405 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
406 * @page: The page within the block of interest
407 * @pfn: The target page frame number
408 * @end_bitidx: The last bit of interest to retrieve
409 * @mask: mask of bits that the caller is interested in
411 * Return: pageblock_bits flags
413 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
415 unsigned long end_bitidx,
418 unsigned long *bitmap;
419 unsigned long bitidx, word_bitidx;
422 bitmap = get_pageblock_bitmap(page, pfn);
423 bitidx = pfn_to_bitidx(page, pfn);
424 word_bitidx = bitidx / BITS_PER_LONG;
425 bitidx &= (BITS_PER_LONG-1);
427 word = bitmap[word_bitidx];
428 bitidx += end_bitidx;
429 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
432 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
433 unsigned long end_bitidx,
436 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
439 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
441 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
445 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
446 * @page: The page within the block of interest
447 * @flags: The flags to set
448 * @pfn: The target page frame number
449 * @end_bitidx: The last bit of interest
450 * @mask: mask of bits that the caller is interested in
452 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
454 unsigned long end_bitidx,
457 unsigned long *bitmap;
458 unsigned long bitidx, word_bitidx;
459 unsigned long old_word, word;
461 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
462 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
464 bitmap = get_pageblock_bitmap(page, pfn);
465 bitidx = pfn_to_bitidx(page, pfn);
466 word_bitidx = bitidx / BITS_PER_LONG;
467 bitidx &= (BITS_PER_LONG-1);
469 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
471 bitidx += end_bitidx;
472 mask <<= (BITS_PER_LONG - bitidx - 1);
473 flags <<= (BITS_PER_LONG - bitidx - 1);
475 word = READ_ONCE(bitmap[word_bitidx]);
477 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
478 if (word == old_word)
484 void set_pageblock_migratetype(struct page *page, int migratetype)
486 if (unlikely(page_group_by_mobility_disabled &&
487 migratetype < MIGRATE_PCPTYPES))
488 migratetype = MIGRATE_UNMOVABLE;
490 set_pageblock_flags_group(page, (unsigned long)migratetype,
491 PB_migrate, PB_migrate_end);
494 #ifdef CONFIG_DEBUG_VM
495 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
499 unsigned long pfn = page_to_pfn(page);
500 unsigned long sp, start_pfn;
503 seq = zone_span_seqbegin(zone);
504 start_pfn = zone->zone_start_pfn;
505 sp = zone->spanned_pages;
506 if (!zone_spans_pfn(zone, pfn))
508 } while (zone_span_seqretry(zone, seq));
511 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
512 pfn, zone_to_nid(zone), zone->name,
513 start_pfn, start_pfn + sp);
518 static int page_is_consistent(struct zone *zone, struct page *page)
520 if (!pfn_valid_within(page_to_pfn(page)))
522 if (zone != page_zone(page))
528 * Temporary debugging check for pages not lying within a given zone.
530 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
532 if (page_outside_zone_boundaries(zone, page))
534 if (!page_is_consistent(zone, page))
540 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
546 static void bad_page(struct page *page, const char *reason,
547 unsigned long bad_flags)
549 static unsigned long resume;
550 static unsigned long nr_shown;
551 static unsigned long nr_unshown;
554 * Allow a burst of 60 reports, then keep quiet for that minute;
555 * or allow a steady drip of one report per second.
557 if (nr_shown == 60) {
558 if (time_before(jiffies, resume)) {
564 "BUG: Bad page state: %lu messages suppressed\n",
571 resume = jiffies + 60 * HZ;
573 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
574 current->comm, page_to_pfn(page));
575 __dump_page(page, reason);
576 bad_flags &= page->flags;
578 pr_alert("bad because of flags: %#lx(%pGp)\n",
579 bad_flags, &bad_flags);
580 dump_page_owner(page);
585 /* Leave bad fields for debug, except PageBuddy could make trouble */
586 page_mapcount_reset(page); /* remove PageBuddy */
587 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
591 * Higher-order pages are called "compound pages". They are structured thusly:
593 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
595 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
596 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
598 * The first tail page's ->compound_dtor holds the offset in array of compound
599 * page destructors. See compound_page_dtors.
601 * The first tail page's ->compound_order holds the order of allocation.
602 * This usage means that zero-order pages may not be compound.
605 void free_compound_page(struct page *page)
607 __free_pages_ok(page, compound_order(page));
610 void prep_compound_page(struct page *page, unsigned int order)
613 int nr_pages = 1 << order;
615 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
616 set_compound_order(page, order);
618 for (i = 1; i < nr_pages; i++) {
619 struct page *p = page + i;
620 set_page_count(p, 0);
621 p->mapping = TAIL_MAPPING;
622 set_compound_head(p, page);
624 atomic_set(compound_mapcount_ptr(page), -1);
627 #ifdef CONFIG_DEBUG_PAGEALLOC
628 unsigned int _debug_guardpage_minorder;
629 bool _debug_pagealloc_enabled __read_mostly
630 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
631 EXPORT_SYMBOL(_debug_pagealloc_enabled);
632 bool _debug_guardpage_enabled __read_mostly;
634 static int __init early_debug_pagealloc(char *buf)
638 return kstrtobool(buf, &_debug_pagealloc_enabled);
640 early_param("debug_pagealloc", early_debug_pagealloc);
642 static bool need_debug_guardpage(void)
644 /* If we don't use debug_pagealloc, we don't need guard page */
645 if (!debug_pagealloc_enabled())
648 if (!debug_guardpage_minorder())
654 static void init_debug_guardpage(void)
656 if (!debug_pagealloc_enabled())
659 if (!debug_guardpage_minorder())
662 _debug_guardpage_enabled = true;
665 struct page_ext_operations debug_guardpage_ops = {
666 .need = need_debug_guardpage,
667 .init = init_debug_guardpage,
670 static int __init debug_guardpage_minorder_setup(char *buf)
674 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
675 pr_err("Bad debug_guardpage_minorder value\n");
678 _debug_guardpage_minorder = res;
679 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
682 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
684 static inline bool set_page_guard(struct zone *zone, struct page *page,
685 unsigned int order, int migratetype)
687 struct page_ext *page_ext;
689 if (!debug_guardpage_enabled())
692 if (order >= debug_guardpage_minorder())
695 page_ext = lookup_page_ext(page);
696 if (unlikely(!page_ext))
699 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
701 INIT_LIST_HEAD(&page->lru);
702 set_page_private(page, order);
703 /* Guard pages are not available for any usage */
704 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
709 static inline void clear_page_guard(struct zone *zone, struct page *page,
710 unsigned int order, int migratetype)
712 struct page_ext *page_ext;
714 if (!debug_guardpage_enabled())
717 page_ext = lookup_page_ext(page);
718 if (unlikely(!page_ext))
721 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
723 set_page_private(page, 0);
724 if (!is_migrate_isolate(migratetype))
725 __mod_zone_freepage_state(zone, (1 << order), migratetype);
728 struct page_ext_operations debug_guardpage_ops;
729 static inline bool set_page_guard(struct zone *zone, struct page *page,
730 unsigned int order, int migratetype) { return false; }
731 static inline void clear_page_guard(struct zone *zone, struct page *page,
732 unsigned int order, int migratetype) {}
735 static inline void set_page_order(struct page *page, unsigned int order)
737 set_page_private(page, order);
738 __SetPageBuddy(page);
741 static inline void rmv_page_order(struct page *page)
743 __ClearPageBuddy(page);
744 set_page_private(page, 0);
748 * This function checks whether a page is free && is the buddy
749 * we can coalesce a page and its buddy if
750 * (a) the buddy is not in a hole (check before calling!) &&
751 * (b) the buddy is in the buddy system &&
752 * (c) a page and its buddy have the same order &&
753 * (d) a page and its buddy are in the same zone.
755 * For recording whether a page is in the buddy system, we set PageBuddy.
756 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
758 * For recording page's order, we use page_private(page).
760 static inline int page_is_buddy(struct page *page, struct page *buddy,
763 if (page_is_guard(buddy) && page_order(buddy) == order) {
764 if (page_zone_id(page) != page_zone_id(buddy))
767 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
772 if (PageBuddy(buddy) && page_order(buddy) == order) {
774 * zone check is done late to avoid uselessly
775 * calculating zone/node ids for pages that could
778 if (page_zone_id(page) != page_zone_id(buddy))
781 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
789 * Freeing function for a buddy system allocator.
791 * The concept of a buddy system is to maintain direct-mapped table
792 * (containing bit values) for memory blocks of various "orders".
793 * The bottom level table contains the map for the smallest allocatable
794 * units of memory (here, pages), and each level above it describes
795 * pairs of units from the levels below, hence, "buddies".
796 * At a high level, all that happens here is marking the table entry
797 * at the bottom level available, and propagating the changes upward
798 * as necessary, plus some accounting needed to play nicely with other
799 * parts of the VM system.
800 * At each level, we keep a list of pages, which are heads of continuous
801 * free pages of length of (1 << order) and marked with PageBuddy.
802 * Page's order is recorded in page_private(page) field.
803 * So when we are allocating or freeing one, we can derive the state of the
804 * other. That is, if we allocate a small block, and both were
805 * free, the remainder of the region must be split into blocks.
806 * If a block is freed, and its buddy is also free, then this
807 * triggers coalescing into a block of larger size.
812 static inline void __free_one_page(struct page *page,
814 struct zone *zone, unsigned int order,
817 unsigned long combined_pfn;
818 unsigned long uninitialized_var(buddy_pfn);
820 unsigned int max_order;
822 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
824 VM_BUG_ON(!zone_is_initialized(zone));
825 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
827 VM_BUG_ON(migratetype == -1);
828 if (likely(!is_migrate_isolate(migratetype)))
829 __mod_zone_freepage_state(zone, 1 << order, migratetype);
831 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
832 VM_BUG_ON_PAGE(bad_range(zone, page), page);
835 while (order < max_order - 1) {
836 buddy_pfn = __find_buddy_pfn(pfn, order);
837 buddy = page + (buddy_pfn - pfn);
839 if (!pfn_valid_within(buddy_pfn))
841 if (!page_is_buddy(page, buddy, order))
844 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
845 * merge with it and move up one order.
847 if (page_is_guard(buddy)) {
848 clear_page_guard(zone, buddy, order, migratetype);
850 list_del(&buddy->lru);
851 zone->free_area[order].nr_free--;
852 rmv_page_order(buddy);
854 combined_pfn = buddy_pfn & pfn;
855 page = page + (combined_pfn - pfn);
859 if (max_order < MAX_ORDER) {
860 /* If we are here, it means order is >= pageblock_order.
861 * We want to prevent merge between freepages on isolate
862 * pageblock and normal pageblock. Without this, pageblock
863 * isolation could cause incorrect freepage or CMA accounting.
865 * We don't want to hit this code for the more frequent
868 if (unlikely(has_isolate_pageblock(zone))) {
871 buddy_pfn = __find_buddy_pfn(pfn, order);
872 buddy = page + (buddy_pfn - pfn);
873 buddy_mt = get_pageblock_migratetype(buddy);
875 if (migratetype != buddy_mt
876 && (is_migrate_isolate(migratetype) ||
877 is_migrate_isolate(buddy_mt)))
881 goto continue_merging;
885 set_page_order(page, order);
888 * If this is not the largest possible page, check if the buddy
889 * of the next-highest order is free. If it is, it's possible
890 * that pages are being freed that will coalesce soon. In case,
891 * that is happening, add the free page to the tail of the list
892 * so it's less likely to be used soon and more likely to be merged
893 * as a higher order page
895 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
896 struct page *higher_page, *higher_buddy;
897 combined_pfn = buddy_pfn & pfn;
898 higher_page = page + (combined_pfn - pfn);
899 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
900 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
901 if (pfn_valid_within(buddy_pfn) &&
902 page_is_buddy(higher_page, higher_buddy, order + 1)) {
903 list_add_tail(&page->lru,
904 &zone->free_area[order].free_list[migratetype]);
909 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
911 zone->free_area[order].nr_free++;
915 * A bad page could be due to a number of fields. Instead of multiple branches,
916 * try and check multiple fields with one check. The caller must do a detailed
917 * check if necessary.
919 static inline bool page_expected_state(struct page *page,
920 unsigned long check_flags)
922 if (unlikely(atomic_read(&page->_mapcount) != -1))
925 if (unlikely((unsigned long)page->mapping |
926 page_ref_count(page) |
928 (unsigned long)page->mem_cgroup |
930 (page->flags & check_flags)))
936 static void free_pages_check_bad(struct page *page)
938 const char *bad_reason;
939 unsigned long bad_flags;
944 if (unlikely(atomic_read(&page->_mapcount) != -1))
945 bad_reason = "nonzero mapcount";
946 if (unlikely(page->mapping != NULL))
947 bad_reason = "non-NULL mapping";
948 if (unlikely(page_ref_count(page) != 0))
949 bad_reason = "nonzero _refcount";
950 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
951 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
952 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
955 if (unlikely(page->mem_cgroup))
956 bad_reason = "page still charged to cgroup";
958 bad_page(page, bad_reason, bad_flags);
961 static inline int free_pages_check(struct page *page)
963 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
966 /* Something has gone sideways, find it */
967 free_pages_check_bad(page);
971 static int free_tail_pages_check(struct page *head_page, struct page *page)
976 * We rely page->lru.next never has bit 0 set, unless the page
977 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
979 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
981 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
985 switch (page - head_page) {
987 /* the first tail page: ->mapping may be compound_mapcount() */
988 if (unlikely(compound_mapcount(page))) {
989 bad_page(page, "nonzero compound_mapcount", 0);
995 * the second tail page: ->mapping is
996 * deferred_list.next -- ignore value.
1000 if (page->mapping != TAIL_MAPPING) {
1001 bad_page(page, "corrupted mapping in tail page", 0);
1006 if (unlikely(!PageTail(page))) {
1007 bad_page(page, "PageTail not set", 0);
1010 if (unlikely(compound_head(page) != head_page)) {
1011 bad_page(page, "compound_head not consistent", 0);
1016 page->mapping = NULL;
1017 clear_compound_head(page);
1021 static __always_inline bool free_pages_prepare(struct page *page,
1022 unsigned int order, bool check_free)
1026 VM_BUG_ON_PAGE(PageTail(page), page);
1028 trace_mm_page_free(page, order);
1031 * Check tail pages before head page information is cleared to
1032 * avoid checking PageCompound for order-0 pages.
1034 if (unlikely(order)) {
1035 bool compound = PageCompound(page);
1038 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1041 ClearPageDoubleMap(page);
1042 for (i = 1; i < (1 << order); i++) {
1044 bad += free_tail_pages_check(page, page + i);
1045 if (unlikely(free_pages_check(page + i))) {
1049 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1052 if (PageMappingFlags(page))
1053 page->mapping = NULL;
1054 if (memcg_kmem_enabled() && PageKmemcg(page))
1055 memcg_kmem_uncharge(page, order);
1057 bad += free_pages_check(page);
1061 page_cpupid_reset_last(page);
1062 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1063 reset_page_owner(page, order);
1065 if (!PageHighMem(page)) {
1066 debug_check_no_locks_freed(page_address(page),
1067 PAGE_SIZE << order);
1068 debug_check_no_obj_freed(page_address(page),
1069 PAGE_SIZE << order);
1071 arch_free_page(page, order);
1072 kernel_poison_pages(page, 1 << order, 0);
1073 kernel_map_pages(page, 1 << order, 0);
1074 kasan_free_nondeferred_pages(page, order);
1079 #ifdef CONFIG_DEBUG_VM
1080 static inline bool free_pcp_prepare(struct page *page)
1082 return free_pages_prepare(page, 0, true);
1085 static inline bool bulkfree_pcp_prepare(struct page *page)
1090 static bool free_pcp_prepare(struct page *page)
1092 return free_pages_prepare(page, 0, false);
1095 static bool bulkfree_pcp_prepare(struct page *page)
1097 return free_pages_check(page);
1099 #endif /* CONFIG_DEBUG_VM */
1101 static inline void prefetch_buddy(struct page *page)
1103 unsigned long pfn = page_to_pfn(page);
1104 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1105 struct page *buddy = page + (buddy_pfn - pfn);
1111 * Frees a number of pages from the PCP lists
1112 * Assumes all pages on list are in same zone, and of same order.
1113 * count is the number of pages to free.
1115 * If the zone was previously in an "all pages pinned" state then look to
1116 * see if this freeing clears that state.
1118 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1119 * pinned" detection logic.
1121 static void free_pcppages_bulk(struct zone *zone, int count,
1122 struct per_cpu_pages *pcp)
1124 int migratetype = 0;
1126 int prefetch_nr = 0;
1127 bool isolated_pageblocks;
1128 struct page *page, *tmp;
1132 struct list_head *list;
1135 * Remove pages from lists in a round-robin fashion. A
1136 * batch_free count is maintained that is incremented when an
1137 * empty list is encountered. This is so more pages are freed
1138 * off fuller lists instead of spinning excessively around empty
1143 if (++migratetype == MIGRATE_PCPTYPES)
1145 list = &pcp->lists[migratetype];
1146 } while (list_empty(list));
1148 /* This is the only non-empty list. Free them all. */
1149 if (batch_free == MIGRATE_PCPTYPES)
1153 page = list_last_entry(list, struct page, lru);
1154 /* must delete to avoid corrupting pcp list */
1155 list_del(&page->lru);
1158 if (bulkfree_pcp_prepare(page))
1161 list_add_tail(&page->lru, &head);
1164 * We are going to put the page back to the global
1165 * pool, prefetch its buddy to speed up later access
1166 * under zone->lock. It is believed the overhead of
1167 * an additional test and calculating buddy_pfn here
1168 * can be offset by reduced memory latency later. To
1169 * avoid excessive prefetching due to large count, only
1170 * prefetch buddy for the first pcp->batch nr of pages.
1172 if (prefetch_nr++ < pcp->batch)
1173 prefetch_buddy(page);
1174 } while (--count && --batch_free && !list_empty(list));
1177 spin_lock(&zone->lock);
1178 isolated_pageblocks = has_isolate_pageblock(zone);
1181 * Use safe version since after __free_one_page(),
1182 * page->lru.next will not point to original list.
1184 list_for_each_entry_safe(page, tmp, &head, lru) {
1185 int mt = get_pcppage_migratetype(page);
1186 /* MIGRATE_ISOLATE page should not go to pcplists */
1187 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1188 /* Pageblock could have been isolated meanwhile */
1189 if (unlikely(isolated_pageblocks))
1190 mt = get_pageblock_migratetype(page);
1192 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1193 trace_mm_page_pcpu_drain(page, 0, mt);
1195 spin_unlock(&zone->lock);
1198 static void free_one_page(struct zone *zone,
1199 struct page *page, unsigned long pfn,
1203 spin_lock(&zone->lock);
1204 if (unlikely(has_isolate_pageblock(zone) ||
1205 is_migrate_isolate(migratetype))) {
1206 migratetype = get_pfnblock_migratetype(page, pfn);
1208 __free_one_page(page, pfn, zone, order, migratetype);
1209 spin_unlock(&zone->lock);
1212 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1213 unsigned long zone, int nid)
1215 mm_zero_struct_page(page);
1216 set_page_links(page, zone, nid, pfn);
1217 init_page_count(page);
1218 page_mapcount_reset(page);
1219 page_cpupid_reset_last(page);
1220 page_kasan_tag_reset(page);
1222 INIT_LIST_HEAD(&page->lru);
1223 #ifdef WANT_PAGE_VIRTUAL
1224 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1225 if (!is_highmem_idx(zone))
1226 set_page_address(page, __va(pfn << PAGE_SHIFT));
1230 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1231 static void __meminit init_reserved_page(unsigned long pfn)
1236 if (!early_page_uninitialised(pfn))
1239 nid = early_pfn_to_nid(pfn);
1240 pgdat = NODE_DATA(nid);
1242 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1243 struct zone *zone = &pgdat->node_zones[zid];
1245 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1248 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1251 static inline void init_reserved_page(unsigned long pfn)
1254 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1257 * Initialised pages do not have PageReserved set. This function is
1258 * called for each range allocated by the bootmem allocator and
1259 * marks the pages PageReserved. The remaining valid pages are later
1260 * sent to the buddy page allocator.
1262 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1264 unsigned long start_pfn = PFN_DOWN(start);
1265 unsigned long end_pfn = PFN_UP(end);
1267 for (; start_pfn < end_pfn; start_pfn++) {
1268 if (pfn_valid(start_pfn)) {
1269 struct page *page = pfn_to_page(start_pfn);
1271 init_reserved_page(start_pfn);
1273 /* Avoid false-positive PageTail() */
1274 INIT_LIST_HEAD(&page->lru);
1277 * no need for atomic set_bit because the struct
1278 * page is not visible yet so nobody should
1281 __SetPageReserved(page);
1286 static void __free_pages_ok(struct page *page, unsigned int order)
1288 unsigned long flags;
1290 unsigned long pfn = page_to_pfn(page);
1292 if (!free_pages_prepare(page, order, true))
1295 migratetype = get_pfnblock_migratetype(page, pfn);
1296 local_irq_save(flags);
1297 __count_vm_events(PGFREE, 1 << order);
1298 free_one_page(page_zone(page), page, pfn, order, migratetype);
1299 local_irq_restore(flags);
1302 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1304 unsigned int nr_pages = 1 << order;
1305 struct page *p = page;
1309 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1311 __ClearPageReserved(p);
1312 set_page_count(p, 0);
1314 __ClearPageReserved(p);
1315 set_page_count(p, 0);
1317 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1318 set_page_refcounted(page);
1319 __free_pages(page, order);
1322 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1323 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1325 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1327 int __meminit early_pfn_to_nid(unsigned long pfn)
1329 static DEFINE_SPINLOCK(early_pfn_lock);
1332 spin_lock(&early_pfn_lock);
1333 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1335 nid = first_online_node;
1336 spin_unlock(&early_pfn_lock);
1342 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1343 static inline bool __meminit __maybe_unused
1344 meminit_pfn_in_nid(unsigned long pfn, int node,
1345 struct mminit_pfnnid_cache *state)
1349 nid = __early_pfn_to_nid(pfn, state);
1350 if (nid >= 0 && nid != node)
1355 /* Only safe to use early in boot when initialisation is single-threaded */
1356 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1358 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1363 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1367 static inline bool __meminit __maybe_unused
1368 meminit_pfn_in_nid(unsigned long pfn, int node,
1369 struct mminit_pfnnid_cache *state)
1376 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1379 if (early_page_uninitialised(pfn))
1381 return __free_pages_boot_core(page, order);
1385 * Check that the whole (or subset of) a pageblock given by the interval of
1386 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1387 * with the migration of free compaction scanner. The scanners then need to
1388 * use only pfn_valid_within() check for arches that allow holes within
1391 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1393 * It's possible on some configurations to have a setup like node0 node1 node0
1394 * i.e. it's possible that all pages within a zones range of pages do not
1395 * belong to a single zone. We assume that a border between node0 and node1
1396 * can occur within a single pageblock, but not a node0 node1 node0
1397 * interleaving within a single pageblock. It is therefore sufficient to check
1398 * the first and last page of a pageblock and avoid checking each individual
1399 * page in a pageblock.
1401 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1402 unsigned long end_pfn, struct zone *zone)
1404 struct page *start_page;
1405 struct page *end_page;
1407 /* end_pfn is one past the range we are checking */
1410 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1413 start_page = pfn_to_online_page(start_pfn);
1417 if (page_zone(start_page) != zone)
1420 end_page = pfn_to_page(end_pfn);
1422 /* This gives a shorter code than deriving page_zone(end_page) */
1423 if (page_zone_id(start_page) != page_zone_id(end_page))
1429 void set_zone_contiguous(struct zone *zone)
1431 unsigned long block_start_pfn = zone->zone_start_pfn;
1432 unsigned long block_end_pfn;
1434 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1435 for (; block_start_pfn < zone_end_pfn(zone);
1436 block_start_pfn = block_end_pfn,
1437 block_end_pfn += pageblock_nr_pages) {
1439 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1441 if (!__pageblock_pfn_to_page(block_start_pfn,
1442 block_end_pfn, zone))
1446 /* We confirm that there is no hole */
1447 zone->contiguous = true;
1450 void clear_zone_contiguous(struct zone *zone)
1452 zone->contiguous = false;
1455 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1456 static void __init deferred_free_range(unsigned long pfn,
1457 unsigned long nr_pages)
1465 page = pfn_to_page(pfn);
1467 /* Free a large naturally-aligned chunk if possible */
1468 if (nr_pages == pageblock_nr_pages &&
1469 (pfn & (pageblock_nr_pages - 1)) == 0) {
1470 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1471 __free_pages_boot_core(page, pageblock_order);
1475 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1476 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1477 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1478 __free_pages_boot_core(page, 0);
1482 /* Completion tracking for deferred_init_memmap() threads */
1483 static atomic_t pgdat_init_n_undone __initdata;
1484 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1486 static inline void __init pgdat_init_report_one_done(void)
1488 if (atomic_dec_and_test(&pgdat_init_n_undone))
1489 complete(&pgdat_init_all_done_comp);
1493 * Returns true if page needs to be initialized or freed to buddy allocator.
1495 * First we check if pfn is valid on architectures where it is possible to have
1496 * holes within pageblock_nr_pages. On systems where it is not possible, this
1497 * function is optimized out.
1499 * Then, we check if a current large page is valid by only checking the validity
1502 * Finally, meminit_pfn_in_nid is checked on systems where pfns can interleave
1503 * within a node: a pfn is between start and end of a node, but does not belong
1504 * to this memory node.
1506 static inline bool __init
1507 deferred_pfn_valid(int nid, unsigned long pfn,
1508 struct mminit_pfnnid_cache *nid_init_state)
1510 if (!pfn_valid_within(pfn))
1512 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1514 if (!meminit_pfn_in_nid(pfn, nid, nid_init_state))
1520 * Free pages to buddy allocator. Try to free aligned pages in
1521 * pageblock_nr_pages sizes.
1523 static void __init deferred_free_pages(int nid, int zid, unsigned long pfn,
1524 unsigned long end_pfn)
1526 struct mminit_pfnnid_cache nid_init_state = { };
1527 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1528 unsigned long nr_free = 0;
1530 for (; pfn < end_pfn; pfn++) {
1531 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1532 deferred_free_range(pfn - nr_free, nr_free);
1534 } else if (!(pfn & nr_pgmask)) {
1535 deferred_free_range(pfn - nr_free, nr_free);
1537 touch_nmi_watchdog();
1542 /* Free the last block of pages to allocator */
1543 deferred_free_range(pfn - nr_free, nr_free);
1547 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1548 * by performing it only once every pageblock_nr_pages.
1549 * Return number of pages initialized.
1551 static unsigned long __init deferred_init_pages(int nid, int zid,
1553 unsigned long end_pfn)
1555 struct mminit_pfnnid_cache nid_init_state = { };
1556 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1557 unsigned long nr_pages = 0;
1558 struct page *page = NULL;
1560 for (; pfn < end_pfn; pfn++) {
1561 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1564 } else if (!page || !(pfn & nr_pgmask)) {
1565 page = pfn_to_page(pfn);
1566 touch_nmi_watchdog();
1570 __init_single_page(page, pfn, zid, nid);
1576 /* Initialise remaining memory on a node */
1577 static int __init deferred_init_memmap(void *data)
1579 pg_data_t *pgdat = data;
1580 int nid = pgdat->node_id;
1581 unsigned long start = jiffies;
1582 unsigned long nr_pages = 0;
1583 unsigned long spfn, epfn, first_init_pfn, flags;
1584 phys_addr_t spa, epa;
1587 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1590 /* Bind memory initialisation thread to a local node if possible */
1591 if (!cpumask_empty(cpumask))
1592 set_cpus_allowed_ptr(current, cpumask);
1594 pgdat_resize_lock(pgdat, &flags);
1595 first_init_pfn = pgdat->first_deferred_pfn;
1596 if (first_init_pfn == ULONG_MAX) {
1597 pgdat_resize_unlock(pgdat, &flags);
1598 pgdat_init_report_one_done();
1602 /* Sanity check boundaries */
1603 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1604 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1605 pgdat->first_deferred_pfn = ULONG_MAX;
1607 /* Only the highest zone is deferred so find it */
1608 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1609 zone = pgdat->node_zones + zid;
1610 if (first_init_pfn < zone_end_pfn(zone))
1613 first_init_pfn = max(zone->zone_start_pfn, first_init_pfn);
1616 * Initialize and free pages. We do it in two loops: first we initialize
1617 * struct page, than free to buddy allocator, because while we are
1618 * freeing pages we can access pages that are ahead (computing buddy
1619 * page in __free_one_page()).
1621 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1622 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1623 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1624 nr_pages += deferred_init_pages(nid, zid, spfn, epfn);
1626 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1627 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1628 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1629 deferred_free_pages(nid, zid, spfn, epfn);
1631 pgdat_resize_unlock(pgdat, &flags);
1633 /* Sanity check that the next zone really is unpopulated */
1634 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1636 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1637 jiffies_to_msecs(jiffies - start));
1639 pgdat_init_report_one_done();
1644 * If this zone has deferred pages, try to grow it by initializing enough
1645 * deferred pages to satisfy the allocation specified by order, rounded up to
1646 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1647 * of SECTION_SIZE bytes by initializing struct pages in increments of
1648 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1650 * Return true when zone was grown, otherwise return false. We return true even
1651 * when we grow less than requested, to let the caller decide if there are
1652 * enough pages to satisfy the allocation.
1654 * Note: We use noinline because this function is needed only during boot, and
1655 * it is called from a __ref function _deferred_grow_zone. This way we are
1656 * making sure that it is not inlined into permanent text section.
1658 static noinline bool __init
1659 deferred_grow_zone(struct zone *zone, unsigned int order)
1661 int zid = zone_idx(zone);
1662 int nid = zone_to_nid(zone);
1663 pg_data_t *pgdat = NODE_DATA(nid);
1664 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1665 unsigned long nr_pages = 0;
1666 unsigned long first_init_pfn, spfn, epfn, t, flags;
1667 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1668 phys_addr_t spa, epa;
1671 /* Only the last zone may have deferred pages */
1672 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1675 pgdat_resize_lock(pgdat, &flags);
1678 * If deferred pages have been initialized while we were waiting for
1679 * the lock, return true, as the zone was grown. The caller will retry
1680 * this zone. We won't return to this function since the caller also
1681 * has this static branch.
1683 if (!static_branch_unlikely(&deferred_pages)) {
1684 pgdat_resize_unlock(pgdat, &flags);
1689 * If someone grew this zone while we were waiting for spinlock, return
1690 * true, as there might be enough pages already.
1692 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1693 pgdat_resize_unlock(pgdat, &flags);
1697 first_init_pfn = max(zone->zone_start_pfn, first_deferred_pfn);
1699 if (first_init_pfn >= pgdat_end_pfn(pgdat)) {
1700 pgdat_resize_unlock(pgdat, &flags);
1704 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1705 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1706 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1708 while (spfn < epfn && nr_pages < nr_pages_needed) {
1709 t = ALIGN(spfn + PAGES_PER_SECTION, PAGES_PER_SECTION);
1710 first_deferred_pfn = min(t, epfn);
1711 nr_pages += deferred_init_pages(nid, zid, spfn,
1712 first_deferred_pfn);
1713 spfn = first_deferred_pfn;
1716 if (nr_pages >= nr_pages_needed)
1720 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1721 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1722 epfn = min_t(unsigned long, first_deferred_pfn, PFN_DOWN(epa));
1723 deferred_free_pages(nid, zid, spfn, epfn);
1725 if (first_deferred_pfn == epfn)
1728 pgdat->first_deferred_pfn = first_deferred_pfn;
1729 pgdat_resize_unlock(pgdat, &flags);
1731 return nr_pages > 0;
1735 * deferred_grow_zone() is __init, but it is called from
1736 * get_page_from_freelist() during early boot until deferred_pages permanently
1737 * disables this call. This is why we have refdata wrapper to avoid warning,
1738 * and to ensure that the function body gets unloaded.
1741 _deferred_grow_zone(struct zone *zone, unsigned int order)
1743 return deferred_grow_zone(zone, order);
1746 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1748 void __init page_alloc_init_late(void)
1752 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1755 /* There will be num_node_state(N_MEMORY) threads */
1756 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1757 for_each_node_state(nid, N_MEMORY) {
1758 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1761 /* Block until all are initialised */
1762 wait_for_completion(&pgdat_init_all_done_comp);
1765 * We initialized the rest of the deferred pages. Permanently disable
1766 * on-demand struct page initialization.
1768 static_branch_disable(&deferred_pages);
1770 /* Reinit limits that are based on free pages after the kernel is up */
1771 files_maxfiles_init();
1773 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1774 /* Discard memblock private memory */
1778 for_each_populated_zone(zone)
1779 set_zone_contiguous(zone);
1783 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1784 void __init init_cma_reserved_pageblock(struct page *page)
1786 unsigned i = pageblock_nr_pages;
1787 struct page *p = page;
1790 __ClearPageReserved(p);
1791 set_page_count(p, 0);
1794 set_pageblock_migratetype(page, MIGRATE_CMA);
1796 if (pageblock_order >= MAX_ORDER) {
1797 i = pageblock_nr_pages;
1800 set_page_refcounted(p);
1801 __free_pages(p, MAX_ORDER - 1);
1802 p += MAX_ORDER_NR_PAGES;
1803 } while (i -= MAX_ORDER_NR_PAGES);
1805 set_page_refcounted(page);
1806 __free_pages(page, pageblock_order);
1809 adjust_managed_page_count(page, pageblock_nr_pages);
1814 * The order of subdivision here is critical for the IO subsystem.
1815 * Please do not alter this order without good reasons and regression
1816 * testing. Specifically, as large blocks of memory are subdivided,
1817 * the order in which smaller blocks are delivered depends on the order
1818 * they're subdivided in this function. This is the primary factor
1819 * influencing the order in which pages are delivered to the IO
1820 * subsystem according to empirical testing, and this is also justified
1821 * by considering the behavior of a buddy system containing a single
1822 * large block of memory acted on by a series of small allocations.
1823 * This behavior is a critical factor in sglist merging's success.
1827 static inline void expand(struct zone *zone, struct page *page,
1828 int low, int high, struct free_area *area,
1831 unsigned long size = 1 << high;
1833 while (high > low) {
1837 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1840 * Mark as guard pages (or page), that will allow to
1841 * merge back to allocator when buddy will be freed.
1842 * Corresponding page table entries will not be touched,
1843 * pages will stay not present in virtual address space
1845 if (set_page_guard(zone, &page[size], high, migratetype))
1848 list_add(&page[size].lru, &area->free_list[migratetype]);
1850 set_page_order(&page[size], high);
1854 static void check_new_page_bad(struct page *page)
1856 const char *bad_reason = NULL;
1857 unsigned long bad_flags = 0;
1859 if (unlikely(atomic_read(&page->_mapcount) != -1))
1860 bad_reason = "nonzero mapcount";
1861 if (unlikely(page->mapping != NULL))
1862 bad_reason = "non-NULL mapping";
1863 if (unlikely(page_ref_count(page) != 0))
1864 bad_reason = "nonzero _count";
1865 if (unlikely(page->flags & __PG_HWPOISON)) {
1866 bad_reason = "HWPoisoned (hardware-corrupted)";
1867 bad_flags = __PG_HWPOISON;
1868 /* Don't complain about hwpoisoned pages */
1869 page_mapcount_reset(page); /* remove PageBuddy */
1872 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1873 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1874 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1877 if (unlikely(page->mem_cgroup))
1878 bad_reason = "page still charged to cgroup";
1880 bad_page(page, bad_reason, bad_flags);
1884 * This page is about to be returned from the page allocator
1886 static inline int check_new_page(struct page *page)
1888 if (likely(page_expected_state(page,
1889 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1892 check_new_page_bad(page);
1896 static inline bool free_pages_prezeroed(void)
1898 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1899 page_poisoning_enabled();
1902 #ifdef CONFIG_DEBUG_VM
1903 static bool check_pcp_refill(struct page *page)
1908 static bool check_new_pcp(struct page *page)
1910 return check_new_page(page);
1913 static bool check_pcp_refill(struct page *page)
1915 return check_new_page(page);
1917 static bool check_new_pcp(struct page *page)
1921 #endif /* CONFIG_DEBUG_VM */
1923 static bool check_new_pages(struct page *page, unsigned int order)
1926 for (i = 0; i < (1 << order); i++) {
1927 struct page *p = page + i;
1929 if (unlikely(check_new_page(p)))
1936 inline void post_alloc_hook(struct page *page, unsigned int order,
1939 set_page_private(page, 0);
1940 set_page_refcounted(page);
1942 arch_alloc_page(page, order);
1943 kernel_map_pages(page, 1 << order, 1);
1944 kernel_poison_pages(page, 1 << order, 1);
1945 kasan_alloc_pages(page, order);
1946 set_page_owner(page, order, gfp_flags);
1949 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1950 unsigned int alloc_flags)
1954 post_alloc_hook(page, order, gfp_flags);
1956 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
1957 for (i = 0; i < (1 << order); i++)
1958 clear_highpage(page + i);
1960 if (order && (gfp_flags & __GFP_COMP))
1961 prep_compound_page(page, order);
1964 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1965 * allocate the page. The expectation is that the caller is taking
1966 * steps that will free more memory. The caller should avoid the page
1967 * being used for !PFMEMALLOC purposes.
1969 if (alloc_flags & ALLOC_NO_WATERMARKS)
1970 set_page_pfmemalloc(page);
1972 clear_page_pfmemalloc(page);
1976 * Go through the free lists for the given migratetype and remove
1977 * the smallest available page from the freelists
1979 static __always_inline
1980 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1983 unsigned int current_order;
1984 struct free_area *area;
1987 /* Find a page of the appropriate size in the preferred list */
1988 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1989 area = &(zone->free_area[current_order]);
1990 page = list_first_entry_or_null(&area->free_list[migratetype],
1994 list_del(&page->lru);
1995 rmv_page_order(page);
1997 expand(zone, page, order, current_order, area, migratetype);
1998 set_pcppage_migratetype(page, migratetype);
2007 * This array describes the order lists are fallen back to when
2008 * the free lists for the desirable migrate type are depleted
2010 static int fallbacks[MIGRATE_TYPES][4] = {
2011 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2012 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2013 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2015 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2017 #ifdef CONFIG_MEMORY_ISOLATION
2018 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2023 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2026 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2029 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2030 unsigned int order) { return NULL; }
2034 * Move the free pages in a range to the free lists of the requested type.
2035 * Note that start_page and end_pages are not aligned on a pageblock
2036 * boundary. If alignment is required, use move_freepages_block()
2038 static int move_freepages(struct zone *zone,
2039 struct page *start_page, struct page *end_page,
2040 int migratetype, int *num_movable)
2044 int pages_moved = 0;
2046 #ifndef CONFIG_HOLES_IN_ZONE
2048 * page_zone is not safe to call in this context when
2049 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
2050 * anyway as we check zone boundaries in move_freepages_block().
2051 * Remove at a later date when no bug reports exist related to
2052 * grouping pages by mobility
2054 VM_BUG_ON(pfn_valid(page_to_pfn(start_page)) &&
2055 pfn_valid(page_to_pfn(end_page)) &&
2056 page_zone(start_page) != page_zone(end_page));
2058 for (page = start_page; page <= end_page;) {
2059 if (!pfn_valid_within(page_to_pfn(page))) {
2064 /* Make sure we are not inadvertently changing nodes */
2065 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2067 if (!PageBuddy(page)) {
2069 * We assume that pages that could be isolated for
2070 * migration are movable. But we don't actually try
2071 * isolating, as that would be expensive.
2074 (PageLRU(page) || __PageMovable(page)))
2081 order = page_order(page);
2082 list_move(&page->lru,
2083 &zone->free_area[order].free_list[migratetype]);
2085 pages_moved += 1 << order;
2091 int move_freepages_block(struct zone *zone, struct page *page,
2092 int migratetype, int *num_movable)
2094 unsigned long start_pfn, end_pfn;
2095 struct page *start_page, *end_page;
2100 start_pfn = page_to_pfn(page);
2101 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2102 start_page = pfn_to_page(start_pfn);
2103 end_page = start_page + pageblock_nr_pages - 1;
2104 end_pfn = start_pfn + pageblock_nr_pages - 1;
2106 /* Do not cross zone boundaries */
2107 if (!zone_spans_pfn(zone, start_pfn))
2109 if (!zone_spans_pfn(zone, end_pfn))
2112 return move_freepages(zone, start_page, end_page, migratetype,
2116 static void change_pageblock_range(struct page *pageblock_page,
2117 int start_order, int migratetype)
2119 int nr_pageblocks = 1 << (start_order - pageblock_order);
2121 while (nr_pageblocks--) {
2122 set_pageblock_migratetype(pageblock_page, migratetype);
2123 pageblock_page += pageblock_nr_pages;
2128 * When we are falling back to another migratetype during allocation, try to
2129 * steal extra free pages from the same pageblocks to satisfy further
2130 * allocations, instead of polluting multiple pageblocks.
2132 * If we are stealing a relatively large buddy page, it is likely there will
2133 * be more free pages in the pageblock, so try to steal them all. For
2134 * reclaimable and unmovable allocations, we steal regardless of page size,
2135 * as fragmentation caused by those allocations polluting movable pageblocks
2136 * is worse than movable allocations stealing from unmovable and reclaimable
2139 static bool can_steal_fallback(unsigned int order, int start_mt)
2142 * Leaving this order check is intended, although there is
2143 * relaxed order check in next check. The reason is that
2144 * we can actually steal whole pageblock if this condition met,
2145 * but, below check doesn't guarantee it and that is just heuristic
2146 * so could be changed anytime.
2148 if (order >= pageblock_order)
2151 if (order >= pageblock_order / 2 ||
2152 start_mt == MIGRATE_RECLAIMABLE ||
2153 start_mt == MIGRATE_UNMOVABLE ||
2154 page_group_by_mobility_disabled)
2160 static inline void boost_watermark(struct zone *zone)
2162 unsigned long max_boost;
2164 if (!watermark_boost_factor)
2167 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2168 watermark_boost_factor, 10000);
2169 max_boost = max(pageblock_nr_pages, max_boost);
2171 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2176 * This function implements actual steal behaviour. If order is large enough,
2177 * we can steal whole pageblock. If not, we first move freepages in this
2178 * pageblock to our migratetype and determine how many already-allocated pages
2179 * are there in the pageblock with a compatible migratetype. If at least half
2180 * of pages are free or compatible, we can change migratetype of the pageblock
2181 * itself, so pages freed in the future will be put on the correct free list.
2183 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2184 unsigned int alloc_flags, int start_type, bool whole_block)
2186 unsigned int current_order = page_order(page);
2187 struct free_area *area;
2188 int free_pages, movable_pages, alike_pages;
2191 old_block_type = get_pageblock_migratetype(page);
2194 * This can happen due to races and we want to prevent broken
2195 * highatomic accounting.
2197 if (is_migrate_highatomic(old_block_type))
2200 /* Take ownership for orders >= pageblock_order */
2201 if (current_order >= pageblock_order) {
2202 change_pageblock_range(page, current_order, start_type);
2207 * Boost watermarks to increase reclaim pressure to reduce the
2208 * likelihood of future fallbacks. Wake kswapd now as the node
2209 * may be balanced overall and kswapd will not wake naturally.
2211 boost_watermark(zone);
2212 if (alloc_flags & ALLOC_KSWAPD)
2213 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
2215 /* We are not allowed to try stealing from the whole block */
2219 free_pages = move_freepages_block(zone, page, start_type,
2222 * Determine how many pages are compatible with our allocation.
2223 * For movable allocation, it's the number of movable pages which
2224 * we just obtained. For other types it's a bit more tricky.
2226 if (start_type == MIGRATE_MOVABLE) {
2227 alike_pages = movable_pages;
2230 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2231 * to MOVABLE pageblock, consider all non-movable pages as
2232 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2233 * vice versa, be conservative since we can't distinguish the
2234 * exact migratetype of non-movable pages.
2236 if (old_block_type == MIGRATE_MOVABLE)
2237 alike_pages = pageblock_nr_pages
2238 - (free_pages + movable_pages);
2243 /* moving whole block can fail due to zone boundary conditions */
2248 * If a sufficient number of pages in the block are either free or of
2249 * comparable migratability as our allocation, claim the whole block.
2251 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2252 page_group_by_mobility_disabled)
2253 set_pageblock_migratetype(page, start_type);
2258 area = &zone->free_area[current_order];
2259 list_move(&page->lru, &area->free_list[start_type]);
2263 * Check whether there is a suitable fallback freepage with requested order.
2264 * If only_stealable is true, this function returns fallback_mt only if
2265 * we can steal other freepages all together. This would help to reduce
2266 * fragmentation due to mixed migratetype pages in one pageblock.
2268 int find_suitable_fallback(struct free_area *area, unsigned int order,
2269 int migratetype, bool only_stealable, bool *can_steal)
2274 if (area->nr_free == 0)
2279 fallback_mt = fallbacks[migratetype][i];
2280 if (fallback_mt == MIGRATE_TYPES)
2283 if (list_empty(&area->free_list[fallback_mt]))
2286 if (can_steal_fallback(order, migratetype))
2289 if (!only_stealable)
2300 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2301 * there are no empty page blocks that contain a page with a suitable order
2303 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2304 unsigned int alloc_order)
2307 unsigned long max_managed, flags;
2310 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2311 * Check is race-prone but harmless.
2313 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2314 if (zone->nr_reserved_highatomic >= max_managed)
2317 spin_lock_irqsave(&zone->lock, flags);
2319 /* Recheck the nr_reserved_highatomic limit under the lock */
2320 if (zone->nr_reserved_highatomic >= max_managed)
2324 mt = get_pageblock_migratetype(page);
2325 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2326 && !is_migrate_cma(mt)) {
2327 zone->nr_reserved_highatomic += pageblock_nr_pages;
2328 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2329 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2333 spin_unlock_irqrestore(&zone->lock, flags);
2337 * Used when an allocation is about to fail under memory pressure. This
2338 * potentially hurts the reliability of high-order allocations when under
2339 * intense memory pressure but failed atomic allocations should be easier
2340 * to recover from than an OOM.
2342 * If @force is true, try to unreserve a pageblock even though highatomic
2343 * pageblock is exhausted.
2345 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2348 struct zonelist *zonelist = ac->zonelist;
2349 unsigned long flags;
2356 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2359 * Preserve at least one pageblock unless memory pressure
2362 if (!force && zone->nr_reserved_highatomic <=
2366 spin_lock_irqsave(&zone->lock, flags);
2367 for (order = 0; order < MAX_ORDER; order++) {
2368 struct free_area *area = &(zone->free_area[order]);
2370 page = list_first_entry_or_null(
2371 &area->free_list[MIGRATE_HIGHATOMIC],
2377 * In page freeing path, migratetype change is racy so
2378 * we can counter several free pages in a pageblock
2379 * in this loop althoug we changed the pageblock type
2380 * from highatomic to ac->migratetype. So we should
2381 * adjust the count once.
2383 if (is_migrate_highatomic_page(page)) {
2385 * It should never happen but changes to
2386 * locking could inadvertently allow a per-cpu
2387 * drain to add pages to MIGRATE_HIGHATOMIC
2388 * while unreserving so be safe and watch for
2391 zone->nr_reserved_highatomic -= min(
2393 zone->nr_reserved_highatomic);
2397 * Convert to ac->migratetype and avoid the normal
2398 * pageblock stealing heuristics. Minimally, the caller
2399 * is doing the work and needs the pages. More
2400 * importantly, if the block was always converted to
2401 * MIGRATE_UNMOVABLE or another type then the number
2402 * of pageblocks that cannot be completely freed
2405 set_pageblock_migratetype(page, ac->migratetype);
2406 ret = move_freepages_block(zone, page, ac->migratetype,
2409 spin_unlock_irqrestore(&zone->lock, flags);
2413 spin_unlock_irqrestore(&zone->lock, flags);
2420 * Try finding a free buddy page on the fallback list and put it on the free
2421 * list of requested migratetype, possibly along with other pages from the same
2422 * block, depending on fragmentation avoidance heuristics. Returns true if
2423 * fallback was found so that __rmqueue_smallest() can grab it.
2425 * The use of signed ints for order and current_order is a deliberate
2426 * deviation from the rest of this file, to make the for loop
2427 * condition simpler.
2429 static __always_inline bool
2430 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2431 unsigned int alloc_flags)
2433 struct free_area *area;
2435 int min_order = order;
2441 * Do not steal pages from freelists belonging to other pageblocks
2442 * i.e. orders < pageblock_order. If there are no local zones free,
2443 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2445 if (alloc_flags & ALLOC_NOFRAGMENT)
2446 min_order = pageblock_order;
2449 * Find the largest available free page in the other list. This roughly
2450 * approximates finding the pageblock with the most free pages, which
2451 * would be too costly to do exactly.
2453 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2455 area = &(zone->free_area[current_order]);
2456 fallback_mt = find_suitable_fallback(area, current_order,
2457 start_migratetype, false, &can_steal);
2458 if (fallback_mt == -1)
2462 * We cannot steal all free pages from the pageblock and the
2463 * requested migratetype is movable. In that case it's better to
2464 * steal and split the smallest available page instead of the
2465 * largest available page, because even if the next movable
2466 * allocation falls back into a different pageblock than this
2467 * one, it won't cause permanent fragmentation.
2469 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2470 && current_order > order)
2479 for (current_order = order; current_order < MAX_ORDER;
2481 area = &(zone->free_area[current_order]);
2482 fallback_mt = find_suitable_fallback(area, current_order,
2483 start_migratetype, false, &can_steal);
2484 if (fallback_mt != -1)
2489 * This should not happen - we already found a suitable fallback
2490 * when looking for the largest page.
2492 VM_BUG_ON(current_order == MAX_ORDER);
2495 page = list_first_entry(&area->free_list[fallback_mt],
2498 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2501 trace_mm_page_alloc_extfrag(page, order, current_order,
2502 start_migratetype, fallback_mt);
2509 * Do the hard work of removing an element from the buddy allocator.
2510 * Call me with the zone->lock already held.
2512 static __always_inline struct page *
2513 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2514 unsigned int alloc_flags)
2519 page = __rmqueue_smallest(zone, order, migratetype);
2520 if (unlikely(!page)) {
2521 if (migratetype == MIGRATE_MOVABLE)
2522 page = __rmqueue_cma_fallback(zone, order);
2524 if (!page && __rmqueue_fallback(zone, order, migratetype,
2529 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2534 * Obtain a specified number of elements from the buddy allocator, all under
2535 * a single hold of the lock, for efficiency. Add them to the supplied list.
2536 * Returns the number of new pages which were placed at *list.
2538 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2539 unsigned long count, struct list_head *list,
2540 int migratetype, unsigned int alloc_flags)
2544 spin_lock(&zone->lock);
2545 for (i = 0; i < count; ++i) {
2546 struct page *page = __rmqueue(zone, order, migratetype,
2548 if (unlikely(page == NULL))
2551 if (unlikely(check_pcp_refill(page)))
2555 * Split buddy pages returned by expand() are received here in
2556 * physical page order. The page is added to the tail of
2557 * caller's list. From the callers perspective, the linked list
2558 * is ordered by page number under some conditions. This is
2559 * useful for IO devices that can forward direction from the
2560 * head, thus also in the physical page order. This is useful
2561 * for IO devices that can merge IO requests if the physical
2562 * pages are ordered properly.
2564 list_add_tail(&page->lru, list);
2566 if (is_migrate_cma(get_pcppage_migratetype(page)))
2567 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2572 * i pages were removed from the buddy list even if some leak due
2573 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2574 * on i. Do not confuse with 'alloced' which is the number of
2575 * pages added to the pcp list.
2577 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2578 spin_unlock(&zone->lock);
2584 * Called from the vmstat counter updater to drain pagesets of this
2585 * currently executing processor on remote nodes after they have
2588 * Note that this function must be called with the thread pinned to
2589 * a single processor.
2591 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2593 unsigned long flags;
2594 int to_drain, batch;
2596 local_irq_save(flags);
2597 batch = READ_ONCE(pcp->batch);
2598 to_drain = min(pcp->count, batch);
2600 free_pcppages_bulk(zone, to_drain, pcp);
2601 local_irq_restore(flags);
2606 * Drain pcplists of the indicated processor and zone.
2608 * The processor must either be the current processor and the
2609 * thread pinned to the current processor or a processor that
2612 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2614 unsigned long flags;
2615 struct per_cpu_pageset *pset;
2616 struct per_cpu_pages *pcp;
2618 local_irq_save(flags);
2619 pset = per_cpu_ptr(zone->pageset, cpu);
2623 free_pcppages_bulk(zone, pcp->count, pcp);
2624 local_irq_restore(flags);
2628 * Drain pcplists of all zones on the indicated processor.
2630 * The processor must either be the current processor and the
2631 * thread pinned to the current processor or a processor that
2634 static void drain_pages(unsigned int cpu)
2638 for_each_populated_zone(zone) {
2639 drain_pages_zone(cpu, zone);
2644 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2646 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2647 * the single zone's pages.
2649 void drain_local_pages(struct zone *zone)
2651 int cpu = smp_processor_id();
2654 drain_pages_zone(cpu, zone);
2659 static void drain_local_pages_wq(struct work_struct *work)
2662 * drain_all_pages doesn't use proper cpu hotplug protection so
2663 * we can race with cpu offline when the WQ can move this from
2664 * a cpu pinned worker to an unbound one. We can operate on a different
2665 * cpu which is allright but we also have to make sure to not move to
2669 drain_local_pages(NULL);
2674 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2676 * When zone parameter is non-NULL, spill just the single zone's pages.
2678 * Note that this can be extremely slow as the draining happens in a workqueue.
2680 void drain_all_pages(struct zone *zone)
2685 * Allocate in the BSS so we wont require allocation in
2686 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2688 static cpumask_t cpus_with_pcps;
2691 * Make sure nobody triggers this path before mm_percpu_wq is fully
2694 if (WARN_ON_ONCE(!mm_percpu_wq))
2698 * Do not drain if one is already in progress unless it's specific to
2699 * a zone. Such callers are primarily CMA and memory hotplug and need
2700 * the drain to be complete when the call returns.
2702 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2705 mutex_lock(&pcpu_drain_mutex);
2709 * We don't care about racing with CPU hotplug event
2710 * as offline notification will cause the notified
2711 * cpu to drain that CPU pcps and on_each_cpu_mask
2712 * disables preemption as part of its processing
2714 for_each_online_cpu(cpu) {
2715 struct per_cpu_pageset *pcp;
2717 bool has_pcps = false;
2720 pcp = per_cpu_ptr(zone->pageset, cpu);
2724 for_each_populated_zone(z) {
2725 pcp = per_cpu_ptr(z->pageset, cpu);
2726 if (pcp->pcp.count) {
2734 cpumask_set_cpu(cpu, &cpus_with_pcps);
2736 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2739 for_each_cpu(cpu, &cpus_with_pcps) {
2740 struct work_struct *work = per_cpu_ptr(&pcpu_drain, cpu);
2741 INIT_WORK(work, drain_local_pages_wq);
2742 queue_work_on(cpu, mm_percpu_wq, work);
2744 for_each_cpu(cpu, &cpus_with_pcps)
2745 flush_work(per_cpu_ptr(&pcpu_drain, cpu));
2747 mutex_unlock(&pcpu_drain_mutex);
2750 #ifdef CONFIG_HIBERNATION
2753 * Touch the watchdog for every WD_PAGE_COUNT pages.
2755 #define WD_PAGE_COUNT (128*1024)
2757 void mark_free_pages(struct zone *zone)
2759 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2760 unsigned long flags;
2761 unsigned int order, t;
2764 if (zone_is_empty(zone))
2767 spin_lock_irqsave(&zone->lock, flags);
2769 max_zone_pfn = zone_end_pfn(zone);
2770 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2771 if (pfn_valid(pfn)) {
2772 page = pfn_to_page(pfn);
2774 if (!--page_count) {
2775 touch_nmi_watchdog();
2776 page_count = WD_PAGE_COUNT;
2779 if (page_zone(page) != zone)
2782 if (!swsusp_page_is_forbidden(page))
2783 swsusp_unset_page_free(page);
2786 for_each_migratetype_order(order, t) {
2787 list_for_each_entry(page,
2788 &zone->free_area[order].free_list[t], lru) {
2791 pfn = page_to_pfn(page);
2792 for (i = 0; i < (1UL << order); i++) {
2793 if (!--page_count) {
2794 touch_nmi_watchdog();
2795 page_count = WD_PAGE_COUNT;
2797 swsusp_set_page_free(pfn_to_page(pfn + i));
2801 spin_unlock_irqrestore(&zone->lock, flags);
2803 #endif /* CONFIG_PM */
2805 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
2809 if (!free_pcp_prepare(page))
2812 migratetype = get_pfnblock_migratetype(page, pfn);
2813 set_pcppage_migratetype(page, migratetype);
2817 static void free_unref_page_commit(struct page *page, unsigned long pfn)
2819 struct zone *zone = page_zone(page);
2820 struct per_cpu_pages *pcp;
2823 migratetype = get_pcppage_migratetype(page);
2824 __count_vm_event(PGFREE);
2827 * We only track unmovable, reclaimable and movable on pcp lists.
2828 * Free ISOLATE pages back to the allocator because they are being
2829 * offlined but treat HIGHATOMIC as movable pages so we can get those
2830 * areas back if necessary. Otherwise, we may have to free
2831 * excessively into the page allocator
2833 if (migratetype >= MIGRATE_PCPTYPES) {
2834 if (unlikely(is_migrate_isolate(migratetype))) {
2835 free_one_page(zone, page, pfn, 0, migratetype);
2838 migratetype = MIGRATE_MOVABLE;
2841 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2842 list_add(&page->lru, &pcp->lists[migratetype]);
2844 if (pcp->count >= pcp->high) {
2845 unsigned long batch = READ_ONCE(pcp->batch);
2846 free_pcppages_bulk(zone, batch, pcp);
2851 * Free a 0-order page
2853 void free_unref_page(struct page *page)
2855 unsigned long flags;
2856 unsigned long pfn = page_to_pfn(page);
2858 if (!free_unref_page_prepare(page, pfn))
2861 local_irq_save(flags);
2862 free_unref_page_commit(page, pfn);
2863 local_irq_restore(flags);
2867 * Free a list of 0-order pages
2869 void free_unref_page_list(struct list_head *list)
2871 struct page *page, *next;
2872 unsigned long flags, pfn;
2873 int batch_count = 0;
2875 /* Prepare pages for freeing */
2876 list_for_each_entry_safe(page, next, list, lru) {
2877 pfn = page_to_pfn(page);
2878 if (!free_unref_page_prepare(page, pfn))
2879 list_del(&page->lru);
2880 set_page_private(page, pfn);
2883 local_irq_save(flags);
2884 list_for_each_entry_safe(page, next, list, lru) {
2885 unsigned long pfn = page_private(page);
2887 set_page_private(page, 0);
2888 trace_mm_page_free_batched(page);
2889 free_unref_page_commit(page, pfn);
2892 * Guard against excessive IRQ disabled times when we get
2893 * a large list of pages to free.
2895 if (++batch_count == SWAP_CLUSTER_MAX) {
2896 local_irq_restore(flags);
2898 local_irq_save(flags);
2901 local_irq_restore(flags);
2905 * split_page takes a non-compound higher-order page, and splits it into
2906 * n (1<<order) sub-pages: page[0..n]
2907 * Each sub-page must be freed individually.
2909 * Note: this is probably too low level an operation for use in drivers.
2910 * Please consult with lkml before using this in your driver.
2912 void split_page(struct page *page, unsigned int order)
2916 VM_BUG_ON_PAGE(PageCompound(page), page);
2917 VM_BUG_ON_PAGE(!page_count(page), page);
2919 for (i = 1; i < (1 << order); i++)
2920 set_page_refcounted(page + i);
2921 split_page_owner(page, order);
2923 EXPORT_SYMBOL_GPL(split_page);
2925 int __isolate_free_page(struct page *page, unsigned int order)
2927 unsigned long watermark;
2931 BUG_ON(!PageBuddy(page));
2933 zone = page_zone(page);
2934 mt = get_pageblock_migratetype(page);
2936 if (!is_migrate_isolate(mt)) {
2938 * Obey watermarks as if the page was being allocated. We can
2939 * emulate a high-order watermark check with a raised order-0
2940 * watermark, because we already know our high-order page
2943 watermark = min_wmark_pages(zone) + (1UL << order);
2944 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2947 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2950 /* Remove page from free list */
2951 list_del(&page->lru);
2952 zone->free_area[order].nr_free--;
2953 rmv_page_order(page);
2956 * Set the pageblock if the isolated page is at least half of a
2959 if (order >= pageblock_order - 1) {
2960 struct page *endpage = page + (1 << order) - 1;
2961 for (; page < endpage; page += pageblock_nr_pages) {
2962 int mt = get_pageblock_migratetype(page);
2963 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2964 && !is_migrate_highatomic(mt))
2965 set_pageblock_migratetype(page,
2971 return 1UL << order;
2975 * Update NUMA hit/miss statistics
2977 * Must be called with interrupts disabled.
2979 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2982 enum numa_stat_item local_stat = NUMA_LOCAL;
2984 /* skip numa counters update if numa stats is disabled */
2985 if (!static_branch_likely(&vm_numa_stat_key))
2988 if (zone_to_nid(z) != numa_node_id())
2989 local_stat = NUMA_OTHER;
2991 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
2992 __inc_numa_state(z, NUMA_HIT);
2994 __inc_numa_state(z, NUMA_MISS);
2995 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
2997 __inc_numa_state(z, local_stat);
3001 /* Remove page from the per-cpu list, caller must protect the list */
3002 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3003 unsigned int alloc_flags,
3004 struct per_cpu_pages *pcp,
3005 struct list_head *list)
3010 if (list_empty(list)) {
3011 pcp->count += rmqueue_bulk(zone, 0,
3013 migratetype, alloc_flags);
3014 if (unlikely(list_empty(list)))
3018 page = list_first_entry(list, struct page, lru);
3019 list_del(&page->lru);
3021 } while (check_new_pcp(page));
3026 /* Lock and remove page from the per-cpu list */
3027 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3028 struct zone *zone, unsigned int order,
3029 gfp_t gfp_flags, int migratetype,
3030 unsigned int alloc_flags)
3032 struct per_cpu_pages *pcp;
3033 struct list_head *list;
3035 unsigned long flags;
3037 local_irq_save(flags);
3038 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3039 list = &pcp->lists[migratetype];
3040 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3042 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3043 zone_statistics(preferred_zone, zone);
3045 local_irq_restore(flags);
3050 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3053 struct page *rmqueue(struct zone *preferred_zone,
3054 struct zone *zone, unsigned int order,
3055 gfp_t gfp_flags, unsigned int alloc_flags,
3058 unsigned long flags;
3061 if (likely(order == 0)) {
3062 page = rmqueue_pcplist(preferred_zone, zone, order,
3063 gfp_flags, migratetype, alloc_flags);
3068 * We most definitely don't want callers attempting to
3069 * allocate greater than order-1 page units with __GFP_NOFAIL.
3071 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3072 spin_lock_irqsave(&zone->lock, flags);
3076 if (alloc_flags & ALLOC_HARDER) {
3077 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3079 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3082 page = __rmqueue(zone, order, migratetype, alloc_flags);
3083 } while (page && check_new_pages(page, order));
3084 spin_unlock(&zone->lock);
3087 __mod_zone_freepage_state(zone, -(1 << order),
3088 get_pcppage_migratetype(page));
3090 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3091 zone_statistics(preferred_zone, zone);
3092 local_irq_restore(flags);
3095 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3099 local_irq_restore(flags);
3103 #ifdef CONFIG_FAIL_PAGE_ALLOC
3106 struct fault_attr attr;
3108 bool ignore_gfp_highmem;
3109 bool ignore_gfp_reclaim;
3111 } fail_page_alloc = {
3112 .attr = FAULT_ATTR_INITIALIZER,
3113 .ignore_gfp_reclaim = true,
3114 .ignore_gfp_highmem = true,
3118 static int __init setup_fail_page_alloc(char *str)
3120 return setup_fault_attr(&fail_page_alloc.attr, str);
3122 __setup("fail_page_alloc=", setup_fail_page_alloc);
3124 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3126 if (order < fail_page_alloc.min_order)
3128 if (gfp_mask & __GFP_NOFAIL)
3130 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3132 if (fail_page_alloc.ignore_gfp_reclaim &&
3133 (gfp_mask & __GFP_DIRECT_RECLAIM))
3136 return should_fail(&fail_page_alloc.attr, 1 << order);
3139 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3141 static int __init fail_page_alloc_debugfs(void)
3143 umode_t mode = S_IFREG | 0600;
3146 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3147 &fail_page_alloc.attr);
3149 return PTR_ERR(dir);
3151 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
3152 &fail_page_alloc.ignore_gfp_reclaim))
3154 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3155 &fail_page_alloc.ignore_gfp_highmem))
3157 if (!debugfs_create_u32("min-order", mode, dir,
3158 &fail_page_alloc.min_order))
3163 debugfs_remove_recursive(dir);
3168 late_initcall(fail_page_alloc_debugfs);
3170 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3172 #else /* CONFIG_FAIL_PAGE_ALLOC */
3174 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3179 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3182 * Return true if free base pages are above 'mark'. For high-order checks it
3183 * will return true of the order-0 watermark is reached and there is at least
3184 * one free page of a suitable size. Checking now avoids taking the zone lock
3185 * to check in the allocation paths if no pages are free.
3187 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3188 int classzone_idx, unsigned int alloc_flags,
3193 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3195 /* free_pages may go negative - that's OK */
3196 free_pages -= (1 << order) - 1;
3198 if (alloc_flags & ALLOC_HIGH)
3202 * If the caller does not have rights to ALLOC_HARDER then subtract
3203 * the high-atomic reserves. This will over-estimate the size of the
3204 * atomic reserve but it avoids a search.
3206 if (likely(!alloc_harder)) {
3207 free_pages -= z->nr_reserved_highatomic;
3210 * OOM victims can try even harder than normal ALLOC_HARDER
3211 * users on the grounds that it's definitely going to be in
3212 * the exit path shortly and free memory. Any allocation it
3213 * makes during the free path will be small and short-lived.
3215 if (alloc_flags & ALLOC_OOM)
3223 /* If allocation can't use CMA areas don't use free CMA pages */
3224 if (!(alloc_flags & ALLOC_CMA))
3225 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3229 * Check watermarks for an order-0 allocation request. If these
3230 * are not met, then a high-order request also cannot go ahead
3231 * even if a suitable page happened to be free.
3233 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3236 /* If this is an order-0 request then the watermark is fine */
3240 /* For a high-order request, check at least one suitable page is free */
3241 for (o = order; o < MAX_ORDER; o++) {
3242 struct free_area *area = &z->free_area[o];
3248 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3249 if (!list_empty(&area->free_list[mt]))
3254 if ((alloc_flags & ALLOC_CMA) &&
3255 !list_empty(&area->free_list[MIGRATE_CMA])) {
3260 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3266 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3267 int classzone_idx, unsigned int alloc_flags)
3269 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3270 zone_page_state(z, NR_FREE_PAGES));
3273 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3274 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3276 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3280 /* If allocation can't use CMA areas don't use free CMA pages */
3281 if (!(alloc_flags & ALLOC_CMA))
3282 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3286 * Fast check for order-0 only. If this fails then the reserves
3287 * need to be calculated. There is a corner case where the check
3288 * passes but only the high-order atomic reserve are free. If
3289 * the caller is !atomic then it'll uselessly search the free
3290 * list. That corner case is then slower but it is harmless.
3292 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3295 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3299 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3300 unsigned long mark, int classzone_idx)
3302 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3304 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3305 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3307 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3312 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3314 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3317 #else /* CONFIG_NUMA */
3318 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3322 #endif /* CONFIG_NUMA */
3325 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3326 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3327 * premature use of a lower zone may cause lowmem pressure problems that
3328 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3329 * probably too small. It only makes sense to spread allocations to avoid
3330 * fragmentation between the Normal and DMA32 zones.
3332 static inline unsigned int
3333 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3335 unsigned int alloc_flags = 0;
3337 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3338 alloc_flags |= ALLOC_KSWAPD;
3340 #ifdef CONFIG_ZONE_DMA32
3341 if (zone_idx(zone) != ZONE_NORMAL)
3345 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3346 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3347 * on UMA that if Normal is populated then so is DMA32.
3349 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3350 if (nr_online_nodes > 1 && !populated_zone(--zone))
3354 #endif /* CONFIG_ZONE_DMA32 */
3359 * get_page_from_freelist goes through the zonelist trying to allocate
3362 static struct page *
3363 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3364 const struct alloc_context *ac)
3368 struct pglist_data *last_pgdat_dirty_limit = NULL;
3373 * Scan zonelist, looking for a zone with enough free.
3374 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3376 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3377 z = ac->preferred_zoneref;
3378 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3383 if (cpusets_enabled() &&
3384 (alloc_flags & ALLOC_CPUSET) &&
3385 !__cpuset_zone_allowed(zone, gfp_mask))
3388 * When allocating a page cache page for writing, we
3389 * want to get it from a node that is within its dirty
3390 * limit, such that no single node holds more than its
3391 * proportional share of globally allowed dirty pages.
3392 * The dirty limits take into account the node's
3393 * lowmem reserves and high watermark so that kswapd
3394 * should be able to balance it without having to
3395 * write pages from its LRU list.
3397 * XXX: For now, allow allocations to potentially
3398 * exceed the per-node dirty limit in the slowpath
3399 * (spread_dirty_pages unset) before going into reclaim,
3400 * which is important when on a NUMA setup the allowed
3401 * nodes are together not big enough to reach the
3402 * global limit. The proper fix for these situations
3403 * will require awareness of nodes in the
3404 * dirty-throttling and the flusher threads.
3406 if (ac->spread_dirty_pages) {
3407 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3410 if (!node_dirty_ok(zone->zone_pgdat)) {
3411 last_pgdat_dirty_limit = zone->zone_pgdat;
3416 if (no_fallback && nr_online_nodes > 1 &&
3417 zone != ac->preferred_zoneref->zone) {
3421 * If moving to a remote node, retry but allow
3422 * fragmenting fallbacks. Locality is more important
3423 * than fragmentation avoidance.
3425 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3426 if (zone_to_nid(zone) != local_nid) {
3427 alloc_flags &= ~ALLOC_NOFRAGMENT;
3432 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3433 if (!zone_watermark_fast(zone, order, mark,
3434 ac_classzone_idx(ac), alloc_flags)) {
3437 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3439 * Watermark failed for this zone, but see if we can
3440 * grow this zone if it contains deferred pages.
3442 if (static_branch_unlikely(&deferred_pages)) {
3443 if (_deferred_grow_zone(zone, order))
3447 /* Checked here to keep the fast path fast */
3448 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3449 if (alloc_flags & ALLOC_NO_WATERMARKS)
3452 if (node_reclaim_mode == 0 ||
3453 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3456 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3458 case NODE_RECLAIM_NOSCAN:
3461 case NODE_RECLAIM_FULL:
3462 /* scanned but unreclaimable */
3465 /* did we reclaim enough */
3466 if (zone_watermark_ok(zone, order, mark,
3467 ac_classzone_idx(ac), alloc_flags))
3475 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3476 gfp_mask, alloc_flags, ac->migratetype);
3478 prep_new_page(page, order, gfp_mask, alloc_flags);
3481 * If this is a high-order atomic allocation then check
3482 * if the pageblock should be reserved for the future
3484 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3485 reserve_highatomic_pageblock(page, zone, order);
3489 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3490 /* Try again if zone has deferred pages */
3491 if (static_branch_unlikely(&deferred_pages)) {
3492 if (_deferred_grow_zone(zone, order))
3500 * It's possible on a UMA machine to get through all zones that are
3501 * fragmented. If avoiding fragmentation, reset and try again.
3504 alloc_flags &= ~ALLOC_NOFRAGMENT;
3511 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3513 unsigned int filter = SHOW_MEM_FILTER_NODES;
3514 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3516 if (!__ratelimit(&show_mem_rs))
3520 * This documents exceptions given to allocations in certain
3521 * contexts that are allowed to allocate outside current's set
3524 if (!(gfp_mask & __GFP_NOMEMALLOC))
3525 if (tsk_is_oom_victim(current) ||
3526 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3527 filter &= ~SHOW_MEM_FILTER_NODES;
3528 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3529 filter &= ~SHOW_MEM_FILTER_NODES;
3531 show_mem(filter, nodemask);
3534 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3536 struct va_format vaf;
3538 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3539 DEFAULT_RATELIMIT_BURST);
3541 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3544 va_start(args, fmt);
3547 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3548 current->comm, &vaf, gfp_mask, &gfp_mask,
3549 nodemask_pr_args(nodemask));
3552 cpuset_print_current_mems_allowed();
3555 warn_alloc_show_mem(gfp_mask, nodemask);
3558 static inline struct page *
3559 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3560 unsigned int alloc_flags,
3561 const struct alloc_context *ac)
3565 page = get_page_from_freelist(gfp_mask, order,
3566 alloc_flags|ALLOC_CPUSET, ac);
3568 * fallback to ignore cpuset restriction if our nodes
3572 page = get_page_from_freelist(gfp_mask, order,
3578 static inline struct page *
3579 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3580 const struct alloc_context *ac, unsigned long *did_some_progress)
3582 struct oom_control oc = {
3583 .zonelist = ac->zonelist,
3584 .nodemask = ac->nodemask,
3586 .gfp_mask = gfp_mask,
3591 *did_some_progress = 0;
3594 * Acquire the oom lock. If that fails, somebody else is
3595 * making progress for us.
3597 if (!mutex_trylock(&oom_lock)) {
3598 *did_some_progress = 1;
3599 schedule_timeout_uninterruptible(1);
3604 * Go through the zonelist yet one more time, keep very high watermark
3605 * here, this is only to catch a parallel oom killing, we must fail if
3606 * we're still under heavy pressure. But make sure that this reclaim
3607 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3608 * allocation which will never fail due to oom_lock already held.
3610 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3611 ~__GFP_DIRECT_RECLAIM, order,
3612 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3616 /* Coredumps can quickly deplete all memory reserves */
3617 if (current->flags & PF_DUMPCORE)
3619 /* The OOM killer will not help higher order allocs */
3620 if (order > PAGE_ALLOC_COSTLY_ORDER)
3623 * We have already exhausted all our reclaim opportunities without any
3624 * success so it is time to admit defeat. We will skip the OOM killer
3625 * because it is very likely that the caller has a more reasonable
3626 * fallback than shooting a random task.
3628 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3630 /* The OOM killer does not needlessly kill tasks for lowmem */
3631 if (ac->high_zoneidx < ZONE_NORMAL)
3633 if (pm_suspended_storage())
3636 * XXX: GFP_NOFS allocations should rather fail than rely on
3637 * other request to make a forward progress.
3638 * We are in an unfortunate situation where out_of_memory cannot
3639 * do much for this context but let's try it to at least get
3640 * access to memory reserved if the current task is killed (see
3641 * out_of_memory). Once filesystems are ready to handle allocation
3642 * failures more gracefully we should just bail out here.
3645 /* The OOM killer may not free memory on a specific node */
3646 if (gfp_mask & __GFP_THISNODE)
3649 /* Exhausted what can be done so it's blame time */
3650 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3651 *did_some_progress = 1;
3654 * Help non-failing allocations by giving them access to memory
3657 if (gfp_mask & __GFP_NOFAIL)
3658 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3659 ALLOC_NO_WATERMARKS, ac);
3662 mutex_unlock(&oom_lock);
3667 * Maximum number of compaction retries wit a progress before OOM
3668 * killer is consider as the only way to move forward.
3670 #define MAX_COMPACT_RETRIES 16
3672 #ifdef CONFIG_COMPACTION
3673 /* Try memory compaction for high-order allocations before reclaim */
3674 static struct page *
3675 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3676 unsigned int alloc_flags, const struct alloc_context *ac,
3677 enum compact_priority prio, enum compact_result *compact_result)
3680 unsigned long pflags;
3681 unsigned int noreclaim_flag;
3686 psi_memstall_enter(&pflags);
3687 noreclaim_flag = memalloc_noreclaim_save();
3689 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3692 memalloc_noreclaim_restore(noreclaim_flag);
3693 psi_memstall_leave(&pflags);
3695 if (*compact_result <= COMPACT_INACTIVE)
3699 * At least in one zone compaction wasn't deferred or skipped, so let's
3700 * count a compaction stall
3702 count_vm_event(COMPACTSTALL);
3704 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3707 struct zone *zone = page_zone(page);
3709 zone->compact_blockskip_flush = false;
3710 compaction_defer_reset(zone, order, true);
3711 count_vm_event(COMPACTSUCCESS);
3716 * It's bad if compaction run occurs and fails. The most likely reason
3717 * is that pages exist, but not enough to satisfy watermarks.
3719 count_vm_event(COMPACTFAIL);
3727 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3728 enum compact_result compact_result,
3729 enum compact_priority *compact_priority,
3730 int *compaction_retries)
3732 int max_retries = MAX_COMPACT_RETRIES;
3735 int retries = *compaction_retries;
3736 enum compact_priority priority = *compact_priority;
3741 if (compaction_made_progress(compact_result))
3742 (*compaction_retries)++;
3745 * compaction considers all the zone as desperately out of memory
3746 * so it doesn't really make much sense to retry except when the
3747 * failure could be caused by insufficient priority
3749 if (compaction_failed(compact_result))
3750 goto check_priority;
3753 * make sure the compaction wasn't deferred or didn't bail out early
3754 * due to locks contention before we declare that we should give up.
3755 * But do not retry if the given zonelist is not suitable for
3758 if (compaction_withdrawn(compact_result)) {
3759 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3764 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3765 * costly ones because they are de facto nofail and invoke OOM
3766 * killer to move on while costly can fail and users are ready
3767 * to cope with that. 1/4 retries is rather arbitrary but we
3768 * would need much more detailed feedback from compaction to
3769 * make a better decision.
3771 if (order > PAGE_ALLOC_COSTLY_ORDER)
3773 if (*compaction_retries <= max_retries) {
3779 * Make sure there are attempts at the highest priority if we exhausted
3780 * all retries or failed at the lower priorities.
3783 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3784 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3786 if (*compact_priority > min_priority) {
3787 (*compact_priority)--;
3788 *compaction_retries = 0;
3792 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3796 static inline struct page *
3797 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3798 unsigned int alloc_flags, const struct alloc_context *ac,
3799 enum compact_priority prio, enum compact_result *compact_result)
3801 *compact_result = COMPACT_SKIPPED;
3806 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3807 enum compact_result compact_result,
3808 enum compact_priority *compact_priority,
3809 int *compaction_retries)
3814 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3818 * There are setups with compaction disabled which would prefer to loop
3819 * inside the allocator rather than hit the oom killer prematurely.
3820 * Let's give them a good hope and keep retrying while the order-0
3821 * watermarks are OK.
3823 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3825 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3826 ac_classzone_idx(ac), alloc_flags))
3831 #endif /* CONFIG_COMPACTION */
3833 #ifdef CONFIG_LOCKDEP
3834 static struct lockdep_map __fs_reclaim_map =
3835 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3837 static bool __need_fs_reclaim(gfp_t gfp_mask)
3839 gfp_mask = current_gfp_context(gfp_mask);
3841 /* no reclaim without waiting on it */
3842 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3845 /* this guy won't enter reclaim */
3846 if (current->flags & PF_MEMALLOC)
3849 /* We're only interested __GFP_FS allocations for now */
3850 if (!(gfp_mask & __GFP_FS))
3853 if (gfp_mask & __GFP_NOLOCKDEP)
3859 void __fs_reclaim_acquire(void)
3861 lock_map_acquire(&__fs_reclaim_map);
3864 void __fs_reclaim_release(void)
3866 lock_map_release(&__fs_reclaim_map);
3869 void fs_reclaim_acquire(gfp_t gfp_mask)
3871 if (__need_fs_reclaim(gfp_mask))
3872 __fs_reclaim_acquire();
3874 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3876 void fs_reclaim_release(gfp_t gfp_mask)
3878 if (__need_fs_reclaim(gfp_mask))
3879 __fs_reclaim_release();
3881 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3884 /* Perform direct synchronous page reclaim */
3886 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3887 const struct alloc_context *ac)
3889 struct reclaim_state reclaim_state;
3891 unsigned int noreclaim_flag;
3892 unsigned long pflags;
3896 /* We now go into synchronous reclaim */
3897 cpuset_memory_pressure_bump();
3898 psi_memstall_enter(&pflags);
3899 fs_reclaim_acquire(gfp_mask);
3900 noreclaim_flag = memalloc_noreclaim_save();
3901 reclaim_state.reclaimed_slab = 0;
3902 current->reclaim_state = &reclaim_state;
3904 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3907 current->reclaim_state = NULL;
3908 memalloc_noreclaim_restore(noreclaim_flag);
3909 fs_reclaim_release(gfp_mask);
3910 psi_memstall_leave(&pflags);
3917 /* The really slow allocator path where we enter direct reclaim */
3918 static inline struct page *
3919 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3920 unsigned int alloc_flags, const struct alloc_context *ac,
3921 unsigned long *did_some_progress)
3923 struct page *page = NULL;
3924 bool drained = false;
3926 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3927 if (unlikely(!(*did_some_progress)))
3931 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3934 * If an allocation failed after direct reclaim, it could be because
3935 * pages are pinned on the per-cpu lists or in high alloc reserves.
3936 * Shrink them them and try again
3938 if (!page && !drained) {
3939 unreserve_highatomic_pageblock(ac, false);
3940 drain_all_pages(NULL);
3948 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
3949 const struct alloc_context *ac)
3953 pg_data_t *last_pgdat = NULL;
3954 enum zone_type high_zoneidx = ac->high_zoneidx;
3956 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
3958 if (last_pgdat != zone->zone_pgdat)
3959 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
3960 last_pgdat = zone->zone_pgdat;
3964 static inline unsigned int
3965 gfp_to_alloc_flags(gfp_t gfp_mask)
3967 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3969 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3970 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3973 * The caller may dip into page reserves a bit more if the caller
3974 * cannot run direct reclaim, or if the caller has realtime scheduling
3975 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3976 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3978 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3980 if (gfp_mask & __GFP_ATOMIC) {
3982 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3983 * if it can't schedule.
3985 if (!(gfp_mask & __GFP_NOMEMALLOC))
3986 alloc_flags |= ALLOC_HARDER;
3988 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3989 * comment for __cpuset_node_allowed().
3991 alloc_flags &= ~ALLOC_CPUSET;
3992 } else if (unlikely(rt_task(current)) && !in_interrupt())
3993 alloc_flags |= ALLOC_HARDER;
3995 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3996 alloc_flags |= ALLOC_KSWAPD;
3999 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4000 alloc_flags |= ALLOC_CMA;
4005 static bool oom_reserves_allowed(struct task_struct *tsk)
4007 if (!tsk_is_oom_victim(tsk))
4011 * !MMU doesn't have oom reaper so give access to memory reserves
4012 * only to the thread with TIF_MEMDIE set
4014 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4021 * Distinguish requests which really need access to full memory
4022 * reserves from oom victims which can live with a portion of it
4024 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4026 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4028 if (gfp_mask & __GFP_MEMALLOC)
4029 return ALLOC_NO_WATERMARKS;
4030 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4031 return ALLOC_NO_WATERMARKS;
4032 if (!in_interrupt()) {
4033 if (current->flags & PF_MEMALLOC)
4034 return ALLOC_NO_WATERMARKS;
4035 else if (oom_reserves_allowed(current))
4042 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4044 return !!__gfp_pfmemalloc_flags(gfp_mask);
4048 * Checks whether it makes sense to retry the reclaim to make a forward progress
4049 * for the given allocation request.
4051 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4052 * without success, or when we couldn't even meet the watermark if we
4053 * reclaimed all remaining pages on the LRU lists.
4055 * Returns true if a retry is viable or false to enter the oom path.
4058 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4059 struct alloc_context *ac, int alloc_flags,
4060 bool did_some_progress, int *no_progress_loops)
4067 * Costly allocations might have made a progress but this doesn't mean
4068 * their order will become available due to high fragmentation so
4069 * always increment the no progress counter for them
4071 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4072 *no_progress_loops = 0;
4074 (*no_progress_loops)++;
4077 * Make sure we converge to OOM if we cannot make any progress
4078 * several times in the row.
4080 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4081 /* Before OOM, exhaust highatomic_reserve */
4082 return unreserve_highatomic_pageblock(ac, true);
4086 * Keep reclaiming pages while there is a chance this will lead
4087 * somewhere. If none of the target zones can satisfy our allocation
4088 * request even if all reclaimable pages are considered then we are
4089 * screwed and have to go OOM.
4091 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4093 unsigned long available;
4094 unsigned long reclaimable;
4095 unsigned long min_wmark = min_wmark_pages(zone);
4098 available = reclaimable = zone_reclaimable_pages(zone);
4099 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4102 * Would the allocation succeed if we reclaimed all
4103 * reclaimable pages?
4105 wmark = __zone_watermark_ok(zone, order, min_wmark,
4106 ac_classzone_idx(ac), alloc_flags, available);
4107 trace_reclaim_retry_zone(z, order, reclaimable,
4108 available, min_wmark, *no_progress_loops, wmark);
4111 * If we didn't make any progress and have a lot of
4112 * dirty + writeback pages then we should wait for
4113 * an IO to complete to slow down the reclaim and
4114 * prevent from pre mature OOM
4116 if (!did_some_progress) {
4117 unsigned long write_pending;
4119 write_pending = zone_page_state_snapshot(zone,
4120 NR_ZONE_WRITE_PENDING);
4122 if (2 * write_pending > reclaimable) {
4123 congestion_wait(BLK_RW_ASYNC, HZ/10);
4135 * Memory allocation/reclaim might be called from a WQ context and the
4136 * current implementation of the WQ concurrency control doesn't
4137 * recognize that a particular WQ is congested if the worker thread is
4138 * looping without ever sleeping. Therefore we have to do a short sleep
4139 * here rather than calling cond_resched().
4141 if (current->flags & PF_WQ_WORKER)
4142 schedule_timeout_uninterruptible(1);
4149 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4152 * It's possible that cpuset's mems_allowed and the nodemask from
4153 * mempolicy don't intersect. This should be normally dealt with by
4154 * policy_nodemask(), but it's possible to race with cpuset update in
4155 * such a way the check therein was true, and then it became false
4156 * before we got our cpuset_mems_cookie here.
4157 * This assumes that for all allocations, ac->nodemask can come only
4158 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4159 * when it does not intersect with the cpuset restrictions) or the
4160 * caller can deal with a violated nodemask.
4162 if (cpusets_enabled() && ac->nodemask &&
4163 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4164 ac->nodemask = NULL;
4169 * When updating a task's mems_allowed or mempolicy nodemask, it is
4170 * possible to race with parallel threads in such a way that our
4171 * allocation can fail while the mask is being updated. If we are about
4172 * to fail, check if the cpuset changed during allocation and if so,
4175 if (read_mems_allowed_retry(cpuset_mems_cookie))
4181 static inline struct page *
4182 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4183 struct alloc_context *ac)
4185 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4186 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4187 struct page *page = NULL;
4188 unsigned int alloc_flags;
4189 unsigned long did_some_progress;
4190 enum compact_priority compact_priority;
4191 enum compact_result compact_result;
4192 int compaction_retries;
4193 int no_progress_loops;
4194 unsigned int cpuset_mems_cookie;
4198 * We also sanity check to catch abuse of atomic reserves being used by
4199 * callers that are not in atomic context.
4201 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4202 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4203 gfp_mask &= ~__GFP_ATOMIC;
4206 compaction_retries = 0;
4207 no_progress_loops = 0;
4208 compact_priority = DEF_COMPACT_PRIORITY;
4209 cpuset_mems_cookie = read_mems_allowed_begin();
4212 * The fast path uses conservative alloc_flags to succeed only until
4213 * kswapd needs to be woken up, and to avoid the cost of setting up
4214 * alloc_flags precisely. So we do that now.
4216 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4219 * We need to recalculate the starting point for the zonelist iterator
4220 * because we might have used different nodemask in the fast path, or
4221 * there was a cpuset modification and we are retrying - otherwise we
4222 * could end up iterating over non-eligible zones endlessly.
4224 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4225 ac->high_zoneidx, ac->nodemask);
4226 if (!ac->preferred_zoneref->zone)
4229 if (alloc_flags & ALLOC_KSWAPD)
4230 wake_all_kswapds(order, gfp_mask, ac);
4233 * The adjusted alloc_flags might result in immediate success, so try
4236 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4241 * For costly allocations, try direct compaction first, as it's likely
4242 * that we have enough base pages and don't need to reclaim. For non-
4243 * movable high-order allocations, do that as well, as compaction will
4244 * try prevent permanent fragmentation by migrating from blocks of the
4246 * Don't try this for allocations that are allowed to ignore
4247 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4249 if (can_direct_reclaim &&
4251 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4252 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4253 page = __alloc_pages_direct_compact(gfp_mask, order,
4255 INIT_COMPACT_PRIORITY,
4261 * Checks for costly allocations with __GFP_NORETRY, which
4262 * includes THP page fault allocations
4264 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4266 * If compaction is deferred for high-order allocations,
4267 * it is because sync compaction recently failed. If
4268 * this is the case and the caller requested a THP
4269 * allocation, we do not want to heavily disrupt the
4270 * system, so we fail the allocation instead of entering
4273 if (compact_result == COMPACT_DEFERRED)
4277 * Looks like reclaim/compaction is worth trying, but
4278 * sync compaction could be very expensive, so keep
4279 * using async compaction.
4281 compact_priority = INIT_COMPACT_PRIORITY;
4286 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4287 if (alloc_flags & ALLOC_KSWAPD)
4288 wake_all_kswapds(order, gfp_mask, ac);
4290 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4292 alloc_flags = reserve_flags;
4295 * Reset the nodemask and zonelist iterators if memory policies can be
4296 * ignored. These allocations are high priority and system rather than
4299 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4300 ac->nodemask = NULL;
4301 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4302 ac->high_zoneidx, ac->nodemask);
4305 /* Attempt with potentially adjusted zonelist and alloc_flags */
4306 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4310 /* Caller is not willing to reclaim, we can't balance anything */
4311 if (!can_direct_reclaim)
4314 /* Avoid recursion of direct reclaim */
4315 if (current->flags & PF_MEMALLOC)
4318 /* Try direct reclaim and then allocating */
4319 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4320 &did_some_progress);
4324 /* Try direct compaction and then allocating */
4325 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4326 compact_priority, &compact_result);
4330 /* Do not loop if specifically requested */
4331 if (gfp_mask & __GFP_NORETRY)
4335 * Do not retry costly high order allocations unless they are
4336 * __GFP_RETRY_MAYFAIL
4338 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4341 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4342 did_some_progress > 0, &no_progress_loops))
4346 * It doesn't make any sense to retry for the compaction if the order-0
4347 * reclaim is not able to make any progress because the current
4348 * implementation of the compaction depends on the sufficient amount
4349 * of free memory (see __compaction_suitable)
4351 if (did_some_progress > 0 &&
4352 should_compact_retry(ac, order, alloc_flags,
4353 compact_result, &compact_priority,
4354 &compaction_retries))
4358 /* Deal with possible cpuset update races before we start OOM killing */
4359 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4362 /* Reclaim has failed us, start killing things */
4363 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4367 /* Avoid allocations with no watermarks from looping endlessly */
4368 if (tsk_is_oom_victim(current) &&
4369 (alloc_flags == ALLOC_OOM ||
4370 (gfp_mask & __GFP_NOMEMALLOC)))
4373 /* Retry as long as the OOM killer is making progress */
4374 if (did_some_progress) {
4375 no_progress_loops = 0;
4380 /* Deal with possible cpuset update races before we fail */
4381 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4385 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4388 if (gfp_mask & __GFP_NOFAIL) {
4390 * All existing users of the __GFP_NOFAIL are blockable, so warn
4391 * of any new users that actually require GFP_NOWAIT
4393 if (WARN_ON_ONCE(!can_direct_reclaim))
4397 * PF_MEMALLOC request from this context is rather bizarre
4398 * because we cannot reclaim anything and only can loop waiting
4399 * for somebody to do a work for us
4401 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4404 * non failing costly orders are a hard requirement which we
4405 * are not prepared for much so let's warn about these users
4406 * so that we can identify them and convert them to something
4409 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4412 * Help non-failing allocations by giving them access to memory
4413 * reserves but do not use ALLOC_NO_WATERMARKS because this
4414 * could deplete whole memory reserves which would just make
4415 * the situation worse
4417 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4425 warn_alloc(gfp_mask, ac->nodemask,
4426 "page allocation failure: order:%u", order);
4431 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4432 int preferred_nid, nodemask_t *nodemask,
4433 struct alloc_context *ac, gfp_t *alloc_mask,
4434 unsigned int *alloc_flags)
4436 ac->high_zoneidx = gfp_zone(gfp_mask);
4437 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4438 ac->nodemask = nodemask;
4439 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4441 if (cpusets_enabled()) {
4442 *alloc_mask |= __GFP_HARDWALL;
4444 ac->nodemask = &cpuset_current_mems_allowed;
4446 *alloc_flags |= ALLOC_CPUSET;
4449 fs_reclaim_acquire(gfp_mask);
4450 fs_reclaim_release(gfp_mask);
4452 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4454 if (should_fail_alloc_page(gfp_mask, order))
4457 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4458 *alloc_flags |= ALLOC_CMA;
4463 /* Determine whether to spread dirty pages and what the first usable zone */
4464 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4466 /* Dirty zone balancing only done in the fast path */
4467 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4470 * The preferred zone is used for statistics but crucially it is
4471 * also used as the starting point for the zonelist iterator. It
4472 * may get reset for allocations that ignore memory policies.
4474 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4475 ac->high_zoneidx, ac->nodemask);
4479 * This is the 'heart' of the zoned buddy allocator.
4482 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4483 nodemask_t *nodemask)
4486 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4487 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4488 struct alloc_context ac = { };
4491 * There are several places where we assume that the order value is sane
4492 * so bail out early if the request is out of bound.
4494 if (unlikely(order >= MAX_ORDER)) {
4495 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4499 gfp_mask &= gfp_allowed_mask;
4500 alloc_mask = gfp_mask;
4501 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4504 finalise_ac(gfp_mask, &ac);
4507 * Forbid the first pass from falling back to types that fragment
4508 * memory until all local zones are considered.
4510 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4512 /* First allocation attempt */
4513 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4518 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4519 * resp. GFP_NOIO which has to be inherited for all allocation requests
4520 * from a particular context which has been marked by
4521 * memalloc_no{fs,io}_{save,restore}.
4523 alloc_mask = current_gfp_context(gfp_mask);
4524 ac.spread_dirty_pages = false;
4527 * Restore the original nodemask if it was potentially replaced with
4528 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4530 if (unlikely(ac.nodemask != nodemask))
4531 ac.nodemask = nodemask;
4533 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4536 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4537 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4538 __free_pages(page, order);
4542 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4546 EXPORT_SYMBOL(__alloc_pages_nodemask);
4549 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4550 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4551 * you need to access high mem.
4553 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4557 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4560 return (unsigned long) page_address(page);
4562 EXPORT_SYMBOL(__get_free_pages);
4564 unsigned long get_zeroed_page(gfp_t gfp_mask)
4566 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4568 EXPORT_SYMBOL(get_zeroed_page);
4570 static inline void free_the_page(struct page *page, unsigned int order)
4572 if (order == 0) /* Via pcp? */
4573 free_unref_page(page);
4575 __free_pages_ok(page, order);
4578 void __free_pages(struct page *page, unsigned int order)
4580 if (put_page_testzero(page))
4581 free_the_page(page, order);
4583 EXPORT_SYMBOL(__free_pages);
4585 void free_pages(unsigned long addr, unsigned int order)
4588 VM_BUG_ON(!virt_addr_valid((void *)addr));
4589 __free_pages(virt_to_page((void *)addr), order);
4593 EXPORT_SYMBOL(free_pages);
4597 * An arbitrary-length arbitrary-offset area of memory which resides
4598 * within a 0 or higher order page. Multiple fragments within that page
4599 * are individually refcounted, in the page's reference counter.
4601 * The page_frag functions below provide a simple allocation framework for
4602 * page fragments. This is used by the network stack and network device
4603 * drivers to provide a backing region of memory for use as either an
4604 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4606 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4609 struct page *page = NULL;
4610 gfp_t gfp = gfp_mask;
4612 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4613 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4615 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4616 PAGE_FRAG_CACHE_MAX_ORDER);
4617 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4619 if (unlikely(!page))
4620 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4622 nc->va = page ? page_address(page) : NULL;
4627 void __page_frag_cache_drain(struct page *page, unsigned int count)
4629 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4631 if (page_ref_sub_and_test(page, count))
4632 free_the_page(page, compound_order(page));
4634 EXPORT_SYMBOL(__page_frag_cache_drain);
4636 void *page_frag_alloc(struct page_frag_cache *nc,
4637 unsigned int fragsz, gfp_t gfp_mask)
4639 unsigned int size = PAGE_SIZE;
4643 if (unlikely(!nc->va)) {
4645 page = __page_frag_cache_refill(nc, gfp_mask);
4649 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4650 /* if size can vary use size else just use PAGE_SIZE */
4653 /* Even if we own the page, we do not use atomic_set().
4654 * This would break get_page_unless_zero() users.
4656 page_ref_add(page, size - 1);
4658 /* reset page count bias and offset to start of new frag */
4659 nc->pfmemalloc = page_is_pfmemalloc(page);
4660 nc->pagecnt_bias = size;
4664 offset = nc->offset - fragsz;
4665 if (unlikely(offset < 0)) {
4666 page = virt_to_page(nc->va);
4668 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4671 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4672 /* if size can vary use size else just use PAGE_SIZE */
4675 /* OK, page count is 0, we can safely set it */
4676 set_page_count(page, size);
4678 /* reset page count bias and offset to start of new frag */
4679 nc->pagecnt_bias = size;
4680 offset = size - fragsz;
4684 nc->offset = offset;
4686 return nc->va + offset;
4688 EXPORT_SYMBOL(page_frag_alloc);
4691 * Frees a page fragment allocated out of either a compound or order 0 page.
4693 void page_frag_free(void *addr)
4695 struct page *page = virt_to_head_page(addr);
4697 if (unlikely(put_page_testzero(page)))
4698 free_the_page(page, compound_order(page));
4700 EXPORT_SYMBOL(page_frag_free);
4702 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4706 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4707 unsigned long used = addr + PAGE_ALIGN(size);
4709 split_page(virt_to_page((void *)addr), order);
4710 while (used < alloc_end) {
4715 return (void *)addr;
4719 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4720 * @size: the number of bytes to allocate
4721 * @gfp_mask: GFP flags for the allocation
4723 * This function is similar to alloc_pages(), except that it allocates the
4724 * minimum number of pages to satisfy the request. alloc_pages() can only
4725 * allocate memory in power-of-two pages.
4727 * This function is also limited by MAX_ORDER.
4729 * Memory allocated by this function must be released by free_pages_exact().
4731 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4733 unsigned int order = get_order(size);
4736 addr = __get_free_pages(gfp_mask, order);
4737 return make_alloc_exact(addr, order, size);
4739 EXPORT_SYMBOL(alloc_pages_exact);
4742 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4744 * @nid: the preferred node ID where memory should be allocated
4745 * @size: the number of bytes to allocate
4746 * @gfp_mask: GFP flags for the allocation
4748 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4751 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4753 unsigned int order = get_order(size);
4754 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4757 return make_alloc_exact((unsigned long)page_address(p), order, size);
4761 * free_pages_exact - release memory allocated via alloc_pages_exact()
4762 * @virt: the value returned by alloc_pages_exact.
4763 * @size: size of allocation, same value as passed to alloc_pages_exact().
4765 * Release the memory allocated by a previous call to alloc_pages_exact.
4767 void free_pages_exact(void *virt, size_t size)
4769 unsigned long addr = (unsigned long)virt;
4770 unsigned long end = addr + PAGE_ALIGN(size);
4772 while (addr < end) {
4777 EXPORT_SYMBOL(free_pages_exact);
4780 * nr_free_zone_pages - count number of pages beyond high watermark
4781 * @offset: The zone index of the highest zone
4783 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4784 * high watermark within all zones at or below a given zone index. For each
4785 * zone, the number of pages is calculated as:
4787 * nr_free_zone_pages = managed_pages - high_pages
4789 static unsigned long nr_free_zone_pages(int offset)
4794 /* Just pick one node, since fallback list is circular */
4795 unsigned long sum = 0;
4797 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4799 for_each_zone_zonelist(zone, z, zonelist, offset) {
4800 unsigned long size = zone_managed_pages(zone);
4801 unsigned long high = high_wmark_pages(zone);
4810 * nr_free_buffer_pages - count number of pages beyond high watermark
4812 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4813 * watermark within ZONE_DMA and ZONE_NORMAL.
4815 unsigned long nr_free_buffer_pages(void)
4817 return nr_free_zone_pages(gfp_zone(GFP_USER));
4819 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4822 * nr_free_pagecache_pages - count number of pages beyond high watermark
4824 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4825 * high watermark within all zones.
4827 unsigned long nr_free_pagecache_pages(void)
4829 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4832 static inline void show_node(struct zone *zone)
4834 if (IS_ENABLED(CONFIG_NUMA))
4835 printk("Node %d ", zone_to_nid(zone));
4838 long si_mem_available(void)
4841 unsigned long pagecache;
4842 unsigned long wmark_low = 0;
4843 unsigned long pages[NR_LRU_LISTS];
4844 unsigned long reclaimable;
4848 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4849 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4852 wmark_low += low_wmark_pages(zone);
4855 * Estimate the amount of memory available for userspace allocations,
4856 * without causing swapping.
4858 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
4861 * Not all the page cache can be freed, otherwise the system will
4862 * start swapping. Assume at least half of the page cache, or the
4863 * low watermark worth of cache, needs to stay.
4865 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4866 pagecache -= min(pagecache / 2, wmark_low);
4867 available += pagecache;
4870 * Part of the reclaimable slab and other kernel memory consists of
4871 * items that are in use, and cannot be freed. Cap this estimate at the
4874 reclaimable = global_node_page_state(NR_SLAB_RECLAIMABLE) +
4875 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
4876 available += reclaimable - min(reclaimable / 2, wmark_low);
4882 EXPORT_SYMBOL_GPL(si_mem_available);
4884 void si_meminfo(struct sysinfo *val)
4886 val->totalram = totalram_pages();
4887 val->sharedram = global_node_page_state(NR_SHMEM);
4888 val->freeram = global_zone_page_state(NR_FREE_PAGES);
4889 val->bufferram = nr_blockdev_pages();
4890 val->totalhigh = totalhigh_pages();
4891 val->freehigh = nr_free_highpages();
4892 val->mem_unit = PAGE_SIZE;
4895 EXPORT_SYMBOL(si_meminfo);
4898 void si_meminfo_node(struct sysinfo *val, int nid)
4900 int zone_type; /* needs to be signed */
4901 unsigned long managed_pages = 0;
4902 unsigned long managed_highpages = 0;
4903 unsigned long free_highpages = 0;
4904 pg_data_t *pgdat = NODE_DATA(nid);
4906 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4907 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
4908 val->totalram = managed_pages;
4909 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4910 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4911 #ifdef CONFIG_HIGHMEM
4912 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4913 struct zone *zone = &pgdat->node_zones[zone_type];
4915 if (is_highmem(zone)) {
4916 managed_highpages += zone_managed_pages(zone);
4917 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4920 val->totalhigh = managed_highpages;
4921 val->freehigh = free_highpages;
4923 val->totalhigh = managed_highpages;
4924 val->freehigh = free_highpages;
4926 val->mem_unit = PAGE_SIZE;
4931 * Determine whether the node should be displayed or not, depending on whether
4932 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4934 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4936 if (!(flags & SHOW_MEM_FILTER_NODES))
4940 * no node mask - aka implicit memory numa policy. Do not bother with
4941 * the synchronization - read_mems_allowed_begin - because we do not
4942 * have to be precise here.
4945 nodemask = &cpuset_current_mems_allowed;
4947 return !node_isset(nid, *nodemask);
4950 #define K(x) ((x) << (PAGE_SHIFT-10))
4952 static void show_migration_types(unsigned char type)
4954 static const char types[MIGRATE_TYPES] = {
4955 [MIGRATE_UNMOVABLE] = 'U',
4956 [MIGRATE_MOVABLE] = 'M',
4957 [MIGRATE_RECLAIMABLE] = 'E',
4958 [MIGRATE_HIGHATOMIC] = 'H',
4960 [MIGRATE_CMA] = 'C',
4962 #ifdef CONFIG_MEMORY_ISOLATION
4963 [MIGRATE_ISOLATE] = 'I',
4966 char tmp[MIGRATE_TYPES + 1];
4970 for (i = 0; i < MIGRATE_TYPES; i++) {
4971 if (type & (1 << i))
4976 printk(KERN_CONT "(%s) ", tmp);
4980 * Show free area list (used inside shift_scroll-lock stuff)
4981 * We also calculate the percentage fragmentation. We do this by counting the
4982 * memory on each free list with the exception of the first item on the list.
4985 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4988 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
4990 unsigned long free_pcp = 0;
4995 for_each_populated_zone(zone) {
4996 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4999 for_each_online_cpu(cpu)
5000 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5003 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5004 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5005 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
5006 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5007 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5008 " free:%lu free_pcp:%lu free_cma:%lu\n",
5009 global_node_page_state(NR_ACTIVE_ANON),
5010 global_node_page_state(NR_INACTIVE_ANON),
5011 global_node_page_state(NR_ISOLATED_ANON),
5012 global_node_page_state(NR_ACTIVE_FILE),
5013 global_node_page_state(NR_INACTIVE_FILE),
5014 global_node_page_state(NR_ISOLATED_FILE),
5015 global_node_page_state(NR_UNEVICTABLE),
5016 global_node_page_state(NR_FILE_DIRTY),
5017 global_node_page_state(NR_WRITEBACK),
5018 global_node_page_state(NR_UNSTABLE_NFS),
5019 global_node_page_state(NR_SLAB_RECLAIMABLE),
5020 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
5021 global_node_page_state(NR_FILE_MAPPED),
5022 global_node_page_state(NR_SHMEM),
5023 global_zone_page_state(NR_PAGETABLE),
5024 global_zone_page_state(NR_BOUNCE),
5025 global_zone_page_state(NR_FREE_PAGES),
5027 global_zone_page_state(NR_FREE_CMA_PAGES));
5029 for_each_online_pgdat(pgdat) {
5030 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5034 " active_anon:%lukB"
5035 " inactive_anon:%lukB"
5036 " active_file:%lukB"
5037 " inactive_file:%lukB"
5038 " unevictable:%lukB"
5039 " isolated(anon):%lukB"
5040 " isolated(file):%lukB"
5045 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5047 " shmem_pmdmapped: %lukB"
5050 " writeback_tmp:%lukB"
5052 " all_unreclaimable? %s"
5055 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5056 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5057 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5058 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5059 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5060 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5061 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5062 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5063 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5064 K(node_page_state(pgdat, NR_WRITEBACK)),
5065 K(node_page_state(pgdat, NR_SHMEM)),
5066 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5067 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5068 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5070 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5072 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5073 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
5074 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5078 for_each_populated_zone(zone) {
5081 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5085 for_each_online_cpu(cpu)
5086 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5095 " active_anon:%lukB"
5096 " inactive_anon:%lukB"
5097 " active_file:%lukB"
5098 " inactive_file:%lukB"
5099 " unevictable:%lukB"
5100 " writepending:%lukB"
5104 " kernel_stack:%lukB"
5112 K(zone_page_state(zone, NR_FREE_PAGES)),
5113 K(min_wmark_pages(zone)),
5114 K(low_wmark_pages(zone)),
5115 K(high_wmark_pages(zone)),
5116 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5117 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5118 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5119 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5120 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5121 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5122 K(zone->present_pages),
5123 K(zone_managed_pages(zone)),
5124 K(zone_page_state(zone, NR_MLOCK)),
5125 zone_page_state(zone, NR_KERNEL_STACK_KB),
5126 K(zone_page_state(zone, NR_PAGETABLE)),
5127 K(zone_page_state(zone, NR_BOUNCE)),
5129 K(this_cpu_read(zone->pageset->pcp.count)),
5130 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5131 printk("lowmem_reserve[]:");
5132 for (i = 0; i < MAX_NR_ZONES; i++)
5133 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5134 printk(KERN_CONT "\n");
5137 for_each_populated_zone(zone) {
5139 unsigned long nr[MAX_ORDER], flags, total = 0;
5140 unsigned char types[MAX_ORDER];
5142 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5145 printk(KERN_CONT "%s: ", zone->name);
5147 spin_lock_irqsave(&zone->lock, flags);
5148 for (order = 0; order < MAX_ORDER; order++) {
5149 struct free_area *area = &zone->free_area[order];
5152 nr[order] = area->nr_free;
5153 total += nr[order] << order;
5156 for (type = 0; type < MIGRATE_TYPES; type++) {
5157 if (!list_empty(&area->free_list[type]))
5158 types[order] |= 1 << type;
5161 spin_unlock_irqrestore(&zone->lock, flags);
5162 for (order = 0; order < MAX_ORDER; order++) {
5163 printk(KERN_CONT "%lu*%lukB ",
5164 nr[order], K(1UL) << order);
5166 show_migration_types(types[order]);
5168 printk(KERN_CONT "= %lukB\n", K(total));
5171 hugetlb_show_meminfo();
5173 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5175 show_swap_cache_info();
5178 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5180 zoneref->zone = zone;
5181 zoneref->zone_idx = zone_idx(zone);
5185 * Builds allocation fallback zone lists.
5187 * Add all populated zones of a node to the zonelist.
5189 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5192 enum zone_type zone_type = MAX_NR_ZONES;
5197 zone = pgdat->node_zones + zone_type;
5198 if (managed_zone(zone)) {
5199 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5200 check_highest_zone(zone_type);
5202 } while (zone_type);
5209 static int __parse_numa_zonelist_order(char *s)
5212 * We used to support different zonlists modes but they turned
5213 * out to be just not useful. Let's keep the warning in place
5214 * if somebody still use the cmd line parameter so that we do
5215 * not fail it silently
5217 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5218 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5224 static __init int setup_numa_zonelist_order(char *s)
5229 return __parse_numa_zonelist_order(s);
5231 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5233 char numa_zonelist_order[] = "Node";
5236 * sysctl handler for numa_zonelist_order
5238 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5239 void __user *buffer, size_t *length,
5246 return proc_dostring(table, write, buffer, length, ppos);
5247 str = memdup_user_nul(buffer, 16);
5249 return PTR_ERR(str);
5251 ret = __parse_numa_zonelist_order(str);
5257 #define MAX_NODE_LOAD (nr_online_nodes)
5258 static int node_load[MAX_NUMNODES];
5261 * find_next_best_node - find the next node that should appear in a given node's fallback list
5262 * @node: node whose fallback list we're appending
5263 * @used_node_mask: nodemask_t of already used nodes
5265 * We use a number of factors to determine which is the next node that should
5266 * appear on a given node's fallback list. The node should not have appeared
5267 * already in @node's fallback list, and it should be the next closest node
5268 * according to the distance array (which contains arbitrary distance values
5269 * from each node to each node in the system), and should also prefer nodes
5270 * with no CPUs, since presumably they'll have very little allocation pressure
5271 * on them otherwise.
5272 * It returns -1 if no node is found.
5274 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5277 int min_val = INT_MAX;
5278 int best_node = NUMA_NO_NODE;
5279 const struct cpumask *tmp = cpumask_of_node(0);
5281 /* Use the local node if we haven't already */
5282 if (!node_isset(node, *used_node_mask)) {
5283 node_set(node, *used_node_mask);
5287 for_each_node_state(n, N_MEMORY) {
5289 /* Don't want a node to appear more than once */
5290 if (node_isset(n, *used_node_mask))
5293 /* Use the distance array to find the distance */
5294 val = node_distance(node, n);
5296 /* Penalize nodes under us ("prefer the next node") */
5299 /* Give preference to headless and unused nodes */
5300 tmp = cpumask_of_node(n);
5301 if (!cpumask_empty(tmp))
5302 val += PENALTY_FOR_NODE_WITH_CPUS;
5304 /* Slight preference for less loaded node */
5305 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5306 val += node_load[n];
5308 if (val < min_val) {
5315 node_set(best_node, *used_node_mask);
5322 * Build zonelists ordered by node and zones within node.
5323 * This results in maximum locality--normal zone overflows into local
5324 * DMA zone, if any--but risks exhausting DMA zone.
5326 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5329 struct zoneref *zonerefs;
5332 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5334 for (i = 0; i < nr_nodes; i++) {
5337 pg_data_t *node = NODE_DATA(node_order[i]);
5339 nr_zones = build_zonerefs_node(node, zonerefs);
5340 zonerefs += nr_zones;
5342 zonerefs->zone = NULL;
5343 zonerefs->zone_idx = 0;
5347 * Build gfp_thisnode zonelists
5349 static void build_thisnode_zonelists(pg_data_t *pgdat)
5351 struct zoneref *zonerefs;
5354 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5355 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5356 zonerefs += nr_zones;
5357 zonerefs->zone = NULL;
5358 zonerefs->zone_idx = 0;
5362 * Build zonelists ordered by zone and nodes within zones.
5363 * This results in conserving DMA zone[s] until all Normal memory is
5364 * exhausted, but results in overflowing to remote node while memory
5365 * may still exist in local DMA zone.
5368 static void build_zonelists(pg_data_t *pgdat)
5370 static int node_order[MAX_NUMNODES];
5371 int node, load, nr_nodes = 0;
5372 nodemask_t used_mask;
5373 int local_node, prev_node;
5375 /* NUMA-aware ordering of nodes */
5376 local_node = pgdat->node_id;
5377 load = nr_online_nodes;
5378 prev_node = local_node;
5379 nodes_clear(used_mask);
5381 memset(node_order, 0, sizeof(node_order));
5382 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5384 * We don't want to pressure a particular node.
5385 * So adding penalty to the first node in same
5386 * distance group to make it round-robin.
5388 if (node_distance(local_node, node) !=
5389 node_distance(local_node, prev_node))
5390 node_load[node] = load;
5392 node_order[nr_nodes++] = node;
5397 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5398 build_thisnode_zonelists(pgdat);
5401 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5403 * Return node id of node used for "local" allocations.
5404 * I.e., first node id of first zone in arg node's generic zonelist.
5405 * Used for initializing percpu 'numa_mem', which is used primarily
5406 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5408 int local_memory_node(int node)
5412 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5413 gfp_zone(GFP_KERNEL),
5415 return zone_to_nid(z->zone);
5419 static void setup_min_unmapped_ratio(void);
5420 static void setup_min_slab_ratio(void);
5421 #else /* CONFIG_NUMA */
5423 static void build_zonelists(pg_data_t *pgdat)
5425 int node, local_node;
5426 struct zoneref *zonerefs;
5429 local_node = pgdat->node_id;
5431 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5432 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5433 zonerefs += nr_zones;
5436 * Now we build the zonelist so that it contains the zones
5437 * of all the other nodes.
5438 * We don't want to pressure a particular node, so when
5439 * building the zones for node N, we make sure that the
5440 * zones coming right after the local ones are those from
5441 * node N+1 (modulo N)
5443 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5444 if (!node_online(node))
5446 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5447 zonerefs += nr_zones;
5449 for (node = 0; node < local_node; node++) {
5450 if (!node_online(node))
5452 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5453 zonerefs += nr_zones;
5456 zonerefs->zone = NULL;
5457 zonerefs->zone_idx = 0;
5460 #endif /* CONFIG_NUMA */
5463 * Boot pageset table. One per cpu which is going to be used for all
5464 * zones and all nodes. The parameters will be set in such a way
5465 * that an item put on a list will immediately be handed over to
5466 * the buddy list. This is safe since pageset manipulation is done
5467 * with interrupts disabled.
5469 * The boot_pagesets must be kept even after bootup is complete for
5470 * unused processors and/or zones. They do play a role for bootstrapping
5471 * hotplugged processors.
5473 * zoneinfo_show() and maybe other functions do
5474 * not check if the processor is online before following the pageset pointer.
5475 * Other parts of the kernel may not check if the zone is available.
5477 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5478 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5479 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5481 static void __build_all_zonelists(void *data)
5484 int __maybe_unused cpu;
5485 pg_data_t *self = data;
5486 static DEFINE_SPINLOCK(lock);
5491 memset(node_load, 0, sizeof(node_load));
5495 * This node is hotadded and no memory is yet present. So just
5496 * building zonelists is fine - no need to touch other nodes.
5498 if (self && !node_online(self->node_id)) {
5499 build_zonelists(self);
5501 for_each_online_node(nid) {
5502 pg_data_t *pgdat = NODE_DATA(nid);
5504 build_zonelists(pgdat);
5507 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5509 * We now know the "local memory node" for each node--
5510 * i.e., the node of the first zone in the generic zonelist.
5511 * Set up numa_mem percpu variable for on-line cpus. During
5512 * boot, only the boot cpu should be on-line; we'll init the
5513 * secondary cpus' numa_mem as they come on-line. During
5514 * node/memory hotplug, we'll fixup all on-line cpus.
5516 for_each_online_cpu(cpu)
5517 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5524 static noinline void __init
5525 build_all_zonelists_init(void)
5529 __build_all_zonelists(NULL);
5532 * Initialize the boot_pagesets that are going to be used
5533 * for bootstrapping processors. The real pagesets for
5534 * each zone will be allocated later when the per cpu
5535 * allocator is available.
5537 * boot_pagesets are used also for bootstrapping offline
5538 * cpus if the system is already booted because the pagesets
5539 * are needed to initialize allocators on a specific cpu too.
5540 * F.e. the percpu allocator needs the page allocator which
5541 * needs the percpu allocator in order to allocate its pagesets
5542 * (a chicken-egg dilemma).
5544 for_each_possible_cpu(cpu)
5545 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5547 mminit_verify_zonelist();
5548 cpuset_init_current_mems_allowed();
5552 * unless system_state == SYSTEM_BOOTING.
5554 * __ref due to call of __init annotated helper build_all_zonelists_init
5555 * [protected by SYSTEM_BOOTING].
5557 void __ref build_all_zonelists(pg_data_t *pgdat)
5559 if (system_state == SYSTEM_BOOTING) {
5560 build_all_zonelists_init();
5562 __build_all_zonelists(pgdat);
5563 /* cpuset refresh routine should be here */
5565 vm_total_pages = nr_free_pagecache_pages();
5567 * Disable grouping by mobility if the number of pages in the
5568 * system is too low to allow the mechanism to work. It would be
5569 * more accurate, but expensive to check per-zone. This check is
5570 * made on memory-hotadd so a system can start with mobility
5571 * disabled and enable it later
5573 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5574 page_group_by_mobility_disabled = 1;
5576 page_group_by_mobility_disabled = 0;
5578 pr_info("Built %i zonelists, mobility grouping %s. Total pages: %ld\n",
5580 page_group_by_mobility_disabled ? "off" : "on",
5583 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5587 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5588 static bool __meminit
5589 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5591 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5592 static struct memblock_region *r;
5594 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5595 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5596 for_each_memblock(memory, r) {
5597 if (*pfn < memblock_region_memory_end_pfn(r))
5601 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5602 memblock_is_mirror(r)) {
5603 *pfn = memblock_region_memory_end_pfn(r);
5612 * Initially all pages are reserved - free ones are freed
5613 * up by memblock_free_all() once the early boot process is
5614 * done. Non-atomic initialization, single-pass.
5616 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5617 unsigned long start_pfn, enum memmap_context context,
5618 struct vmem_altmap *altmap)
5620 unsigned long pfn, end_pfn = start_pfn + size;
5623 if (highest_memmap_pfn < end_pfn - 1)
5624 highest_memmap_pfn = end_pfn - 1;
5626 #ifdef CONFIG_ZONE_DEVICE
5628 * Honor reservation requested by the driver for this ZONE_DEVICE
5629 * memory. We limit the total number of pages to initialize to just
5630 * those that might contain the memory mapping. We will defer the
5631 * ZONE_DEVICE page initialization until after we have released
5634 if (zone == ZONE_DEVICE) {
5638 if (start_pfn == altmap->base_pfn)
5639 start_pfn += altmap->reserve;
5640 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5644 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5646 * There can be holes in boot-time mem_map[]s handed to this
5647 * function. They do not exist on hotplugged memory.
5649 if (context == MEMMAP_EARLY) {
5650 if (!early_pfn_valid(pfn))
5652 if (!early_pfn_in_nid(pfn, nid))
5654 if (overlap_memmap_init(zone, &pfn))
5656 if (defer_init(nid, pfn, end_pfn))
5660 page = pfn_to_page(pfn);
5661 __init_single_page(page, pfn, zone, nid);
5662 if (context == MEMMAP_HOTPLUG)
5663 __SetPageReserved(page);
5666 * Mark the block movable so that blocks are reserved for
5667 * movable at startup. This will force kernel allocations
5668 * to reserve their blocks rather than leaking throughout
5669 * the address space during boot when many long-lived
5670 * kernel allocations are made.
5672 * bitmap is created for zone's valid pfn range. but memmap
5673 * can be created for invalid pages (for alignment)
5674 * check here not to call set_pageblock_migratetype() against
5677 if (!(pfn & (pageblock_nr_pages - 1))) {
5678 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5682 #ifdef CONFIG_SPARSEMEM
5684 * If the zone does not span the rest of the section then
5685 * we should at least initialize those pages. Otherwise we
5686 * could blow up on a poisoned page in some paths which depend
5687 * on full sections being initialized (e.g. memory hotplug).
5689 while (end_pfn % PAGES_PER_SECTION) {
5690 __init_single_page(pfn_to_page(end_pfn), end_pfn, zone, nid);
5696 #ifdef CONFIG_ZONE_DEVICE
5697 void __ref memmap_init_zone_device(struct zone *zone,
5698 unsigned long start_pfn,
5700 struct dev_pagemap *pgmap)
5702 unsigned long pfn, end_pfn = start_pfn + size;
5703 struct pglist_data *pgdat = zone->zone_pgdat;
5704 unsigned long zone_idx = zone_idx(zone);
5705 unsigned long start = jiffies;
5706 int nid = pgdat->node_id;
5708 if (WARN_ON_ONCE(!pgmap || !is_dev_zone(zone)))
5712 * The call to memmap_init_zone should have already taken care
5713 * of the pages reserved for the memmap, so we can just jump to
5714 * the end of that region and start processing the device pages.
5716 if (pgmap->altmap_valid) {
5717 struct vmem_altmap *altmap = &pgmap->altmap;
5719 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5720 size = end_pfn - start_pfn;
5723 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5724 struct page *page = pfn_to_page(pfn);
5726 __init_single_page(page, pfn, zone_idx, nid);
5729 * Mark page reserved as it will need to wait for onlining
5730 * phase for it to be fully associated with a zone.
5732 * We can use the non-atomic __set_bit operation for setting
5733 * the flag as we are still initializing the pages.
5735 __SetPageReserved(page);
5738 * ZONE_DEVICE pages union ->lru with a ->pgmap back
5739 * pointer and hmm_data. It is a bug if a ZONE_DEVICE
5740 * page is ever freed or placed on a driver-private list.
5742 page->pgmap = pgmap;
5746 * Mark the block movable so that blocks are reserved for
5747 * movable at startup. This will force kernel allocations
5748 * to reserve their blocks rather than leaking throughout
5749 * the address space during boot when many long-lived
5750 * kernel allocations are made.
5752 * bitmap is created for zone's valid pfn range. but memmap
5753 * can be created for invalid pages (for alignment)
5754 * check here not to call set_pageblock_migratetype() against
5757 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
5758 * because this is done early in sparse_add_one_section
5760 if (!(pfn & (pageblock_nr_pages - 1))) {
5761 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5766 pr_info("%s initialised, %lu pages in %ums\n", dev_name(pgmap->dev),
5767 size, jiffies_to_msecs(jiffies - start));
5771 static void __meminit zone_init_free_lists(struct zone *zone)
5773 unsigned int order, t;
5774 for_each_migratetype_order(order, t) {
5775 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5776 zone->free_area[order].nr_free = 0;
5780 void __meminit __weak memmap_init(unsigned long size, int nid,
5781 unsigned long zone, unsigned long start_pfn)
5783 memmap_init_zone(size, nid, zone, start_pfn, MEMMAP_EARLY, NULL);
5786 static int zone_batchsize(struct zone *zone)
5792 * The per-cpu-pages pools are set to around 1000th of the
5795 batch = zone_managed_pages(zone) / 1024;
5796 /* But no more than a meg. */
5797 if (batch * PAGE_SIZE > 1024 * 1024)
5798 batch = (1024 * 1024) / PAGE_SIZE;
5799 batch /= 4; /* We effectively *= 4 below */
5804 * Clamp the batch to a 2^n - 1 value. Having a power
5805 * of 2 value was found to be more likely to have
5806 * suboptimal cache aliasing properties in some cases.
5808 * For example if 2 tasks are alternately allocating
5809 * batches of pages, one task can end up with a lot
5810 * of pages of one half of the possible page colors
5811 * and the other with pages of the other colors.
5813 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5818 /* The deferral and batching of frees should be suppressed under NOMMU
5821 * The problem is that NOMMU needs to be able to allocate large chunks
5822 * of contiguous memory as there's no hardware page translation to
5823 * assemble apparent contiguous memory from discontiguous pages.
5825 * Queueing large contiguous runs of pages for batching, however,
5826 * causes the pages to actually be freed in smaller chunks. As there
5827 * can be a significant delay between the individual batches being
5828 * recycled, this leads to the once large chunks of space being
5829 * fragmented and becoming unavailable for high-order allocations.
5836 * pcp->high and pcp->batch values are related and dependent on one another:
5837 * ->batch must never be higher then ->high.
5838 * The following function updates them in a safe manner without read side
5841 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5842 * those fields changing asynchronously (acording the the above rule).
5844 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5845 * outside of boot time (or some other assurance that no concurrent updaters
5848 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5849 unsigned long batch)
5851 /* start with a fail safe value for batch */
5855 /* Update high, then batch, in order */
5862 /* a companion to pageset_set_high() */
5863 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5865 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5868 static void pageset_init(struct per_cpu_pageset *p)
5870 struct per_cpu_pages *pcp;
5873 memset(p, 0, sizeof(*p));
5876 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5877 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5880 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5883 pageset_set_batch(p, batch);
5887 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5888 * to the value high for the pageset p.
5890 static void pageset_set_high(struct per_cpu_pageset *p,
5893 unsigned long batch = max(1UL, high / 4);
5894 if ((high / 4) > (PAGE_SHIFT * 8))
5895 batch = PAGE_SHIFT * 8;
5897 pageset_update(&p->pcp, high, batch);
5900 static void pageset_set_high_and_batch(struct zone *zone,
5901 struct per_cpu_pageset *pcp)
5903 if (percpu_pagelist_fraction)
5904 pageset_set_high(pcp,
5905 (zone_managed_pages(zone) /
5906 percpu_pagelist_fraction));
5908 pageset_set_batch(pcp, zone_batchsize(zone));
5911 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5913 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5916 pageset_set_high_and_batch(zone, pcp);
5919 void __meminit setup_zone_pageset(struct zone *zone)
5922 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5923 for_each_possible_cpu(cpu)
5924 zone_pageset_init(zone, cpu);
5928 * Allocate per cpu pagesets and initialize them.
5929 * Before this call only boot pagesets were available.
5931 void __init setup_per_cpu_pageset(void)
5933 struct pglist_data *pgdat;
5936 for_each_populated_zone(zone)
5937 setup_zone_pageset(zone);
5939 for_each_online_pgdat(pgdat)
5940 pgdat->per_cpu_nodestats =
5941 alloc_percpu(struct per_cpu_nodestat);
5944 static __meminit void zone_pcp_init(struct zone *zone)
5947 * per cpu subsystem is not up at this point. The following code
5948 * relies on the ability of the linker to provide the
5949 * offset of a (static) per cpu variable into the per cpu area.
5951 zone->pageset = &boot_pageset;
5953 if (populated_zone(zone))
5954 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5955 zone->name, zone->present_pages,
5956 zone_batchsize(zone));
5959 void __meminit init_currently_empty_zone(struct zone *zone,
5960 unsigned long zone_start_pfn,
5963 struct pglist_data *pgdat = zone->zone_pgdat;
5964 int zone_idx = zone_idx(zone) + 1;
5966 if (zone_idx > pgdat->nr_zones)
5967 pgdat->nr_zones = zone_idx;
5969 zone->zone_start_pfn = zone_start_pfn;
5971 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5972 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5974 (unsigned long)zone_idx(zone),
5975 zone_start_pfn, (zone_start_pfn + size));
5977 zone_init_free_lists(zone);
5978 zone->initialized = 1;
5981 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5982 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5985 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5987 int __meminit __early_pfn_to_nid(unsigned long pfn,
5988 struct mminit_pfnnid_cache *state)
5990 unsigned long start_pfn, end_pfn;
5993 if (state->last_start <= pfn && pfn < state->last_end)
5994 return state->last_nid;
5996 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5998 state->last_start = start_pfn;
5999 state->last_end = end_pfn;
6000 state->last_nid = nid;
6005 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
6008 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
6009 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
6010 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
6012 * If an architecture guarantees that all ranges registered contain no holes
6013 * and may be freed, this this function may be used instead of calling
6014 * memblock_free_early_nid() manually.
6016 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
6018 unsigned long start_pfn, end_pfn;
6021 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
6022 start_pfn = min(start_pfn, max_low_pfn);
6023 end_pfn = min(end_pfn, max_low_pfn);
6025 if (start_pfn < end_pfn)
6026 memblock_free_early_nid(PFN_PHYS(start_pfn),
6027 (end_pfn - start_pfn) << PAGE_SHIFT,
6033 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6034 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6036 * If an architecture guarantees that all ranges registered contain no holes and may
6037 * be freed, this function may be used instead of calling memory_present() manually.
6039 void __init sparse_memory_present_with_active_regions(int nid)
6041 unsigned long start_pfn, end_pfn;
6044 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
6045 memory_present(this_nid, start_pfn, end_pfn);
6049 * get_pfn_range_for_nid - Return the start and end page frames for a node
6050 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6051 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6052 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6054 * It returns the start and end page frame of a node based on information
6055 * provided by memblock_set_node(). If called for a node
6056 * with no available memory, a warning is printed and the start and end
6059 void __init get_pfn_range_for_nid(unsigned int nid,
6060 unsigned long *start_pfn, unsigned long *end_pfn)
6062 unsigned long this_start_pfn, this_end_pfn;
6068 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6069 *start_pfn = min(*start_pfn, this_start_pfn);
6070 *end_pfn = max(*end_pfn, this_end_pfn);
6073 if (*start_pfn == -1UL)
6078 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6079 * assumption is made that zones within a node are ordered in monotonic
6080 * increasing memory addresses so that the "highest" populated zone is used
6082 static void __init find_usable_zone_for_movable(void)
6085 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6086 if (zone_index == ZONE_MOVABLE)
6089 if (arch_zone_highest_possible_pfn[zone_index] >
6090 arch_zone_lowest_possible_pfn[zone_index])
6094 VM_BUG_ON(zone_index == -1);
6095 movable_zone = zone_index;
6099 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6100 * because it is sized independent of architecture. Unlike the other zones,
6101 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6102 * in each node depending on the size of each node and how evenly kernelcore
6103 * is distributed. This helper function adjusts the zone ranges
6104 * provided by the architecture for a given node by using the end of the
6105 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6106 * zones within a node are in order of monotonic increases memory addresses
6108 static void __init adjust_zone_range_for_zone_movable(int nid,
6109 unsigned long zone_type,
6110 unsigned long node_start_pfn,
6111 unsigned long node_end_pfn,
6112 unsigned long *zone_start_pfn,
6113 unsigned long *zone_end_pfn)
6115 /* Only adjust if ZONE_MOVABLE is on this node */
6116 if (zone_movable_pfn[nid]) {
6117 /* Size ZONE_MOVABLE */
6118 if (zone_type == ZONE_MOVABLE) {
6119 *zone_start_pfn = zone_movable_pfn[nid];
6120 *zone_end_pfn = min(node_end_pfn,
6121 arch_zone_highest_possible_pfn[movable_zone]);
6123 /* Adjust for ZONE_MOVABLE starting within this range */
6124 } else if (!mirrored_kernelcore &&
6125 *zone_start_pfn < zone_movable_pfn[nid] &&
6126 *zone_end_pfn > zone_movable_pfn[nid]) {
6127 *zone_end_pfn = zone_movable_pfn[nid];
6129 /* Check if this whole range is within ZONE_MOVABLE */
6130 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6131 *zone_start_pfn = *zone_end_pfn;
6136 * Return the number of pages a zone spans in a node, including holes
6137 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6139 static unsigned long __init zone_spanned_pages_in_node(int nid,
6140 unsigned long zone_type,
6141 unsigned long node_start_pfn,
6142 unsigned long node_end_pfn,
6143 unsigned long *zone_start_pfn,
6144 unsigned long *zone_end_pfn,
6145 unsigned long *ignored)
6147 /* When hotadd a new node from cpu_up(), the node should be empty */
6148 if (!node_start_pfn && !node_end_pfn)
6151 /* Get the start and end of the zone */
6152 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
6153 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
6154 adjust_zone_range_for_zone_movable(nid, zone_type,
6155 node_start_pfn, node_end_pfn,
6156 zone_start_pfn, zone_end_pfn);
6158 /* Check that this node has pages within the zone's required range */
6159 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6162 /* Move the zone boundaries inside the node if necessary */
6163 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6164 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6166 /* Return the spanned pages */
6167 return *zone_end_pfn - *zone_start_pfn;
6171 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6172 * then all holes in the requested range will be accounted for.
6174 unsigned long __init __absent_pages_in_range(int nid,
6175 unsigned long range_start_pfn,
6176 unsigned long range_end_pfn)
6178 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6179 unsigned long start_pfn, end_pfn;
6182 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6183 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6184 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6185 nr_absent -= end_pfn - start_pfn;
6191 * absent_pages_in_range - Return number of page frames in holes within a range
6192 * @start_pfn: The start PFN to start searching for holes
6193 * @end_pfn: The end PFN to stop searching for holes
6195 * It returns the number of pages frames in memory holes within a range.
6197 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6198 unsigned long end_pfn)
6200 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6203 /* Return the number of page frames in holes in a zone on a node */
6204 static unsigned long __init zone_absent_pages_in_node(int nid,
6205 unsigned long zone_type,
6206 unsigned long node_start_pfn,
6207 unsigned long node_end_pfn,
6208 unsigned long *ignored)
6210 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6211 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6212 unsigned long zone_start_pfn, zone_end_pfn;
6213 unsigned long nr_absent;
6215 /* When hotadd a new node from cpu_up(), the node should be empty */
6216 if (!node_start_pfn && !node_end_pfn)
6219 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6220 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6222 adjust_zone_range_for_zone_movable(nid, zone_type,
6223 node_start_pfn, node_end_pfn,
6224 &zone_start_pfn, &zone_end_pfn);
6225 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6228 * ZONE_MOVABLE handling.
6229 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6232 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6233 unsigned long start_pfn, end_pfn;
6234 struct memblock_region *r;
6236 for_each_memblock(memory, r) {
6237 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6238 zone_start_pfn, zone_end_pfn);
6239 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6240 zone_start_pfn, zone_end_pfn);
6242 if (zone_type == ZONE_MOVABLE &&
6243 memblock_is_mirror(r))
6244 nr_absent += end_pfn - start_pfn;
6246 if (zone_type == ZONE_NORMAL &&
6247 !memblock_is_mirror(r))
6248 nr_absent += end_pfn - start_pfn;
6255 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6256 static inline unsigned long __init zone_spanned_pages_in_node(int nid,
6257 unsigned long zone_type,
6258 unsigned long node_start_pfn,
6259 unsigned long node_end_pfn,
6260 unsigned long *zone_start_pfn,
6261 unsigned long *zone_end_pfn,
6262 unsigned long *zones_size)
6266 *zone_start_pfn = node_start_pfn;
6267 for (zone = 0; zone < zone_type; zone++)
6268 *zone_start_pfn += zones_size[zone];
6270 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6272 return zones_size[zone_type];
6275 static inline unsigned long __init zone_absent_pages_in_node(int nid,
6276 unsigned long zone_type,
6277 unsigned long node_start_pfn,
6278 unsigned long node_end_pfn,
6279 unsigned long *zholes_size)
6284 return zholes_size[zone_type];
6287 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6289 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6290 unsigned long node_start_pfn,
6291 unsigned long node_end_pfn,
6292 unsigned long *zones_size,
6293 unsigned long *zholes_size)
6295 unsigned long realtotalpages = 0, totalpages = 0;
6298 for (i = 0; i < MAX_NR_ZONES; i++) {
6299 struct zone *zone = pgdat->node_zones + i;
6300 unsigned long zone_start_pfn, zone_end_pfn;
6301 unsigned long size, real_size;
6303 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6309 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6310 node_start_pfn, node_end_pfn,
6313 zone->zone_start_pfn = zone_start_pfn;
6315 zone->zone_start_pfn = 0;
6316 zone->spanned_pages = size;
6317 zone->present_pages = real_size;
6320 realtotalpages += real_size;
6323 pgdat->node_spanned_pages = totalpages;
6324 pgdat->node_present_pages = realtotalpages;
6325 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6329 #ifndef CONFIG_SPARSEMEM
6331 * Calculate the size of the zone->blockflags rounded to an unsigned long
6332 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6333 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6334 * round what is now in bits to nearest long in bits, then return it in
6337 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6339 unsigned long usemapsize;
6341 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6342 usemapsize = roundup(zonesize, pageblock_nr_pages);
6343 usemapsize = usemapsize >> pageblock_order;
6344 usemapsize *= NR_PAGEBLOCK_BITS;
6345 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6347 return usemapsize / 8;
6350 static void __ref setup_usemap(struct pglist_data *pgdat,
6352 unsigned long zone_start_pfn,
6353 unsigned long zonesize)
6355 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6356 zone->pageblock_flags = NULL;
6358 zone->pageblock_flags =
6359 memblock_alloc_node_nopanic(usemapsize,
6363 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6364 unsigned long zone_start_pfn, unsigned long zonesize) {}
6365 #endif /* CONFIG_SPARSEMEM */
6367 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6369 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6370 void __init set_pageblock_order(void)
6374 /* Check that pageblock_nr_pages has not already been setup */
6375 if (pageblock_order)
6378 if (HPAGE_SHIFT > PAGE_SHIFT)
6379 order = HUGETLB_PAGE_ORDER;
6381 order = MAX_ORDER - 1;
6384 * Assume the largest contiguous order of interest is a huge page.
6385 * This value may be variable depending on boot parameters on IA64 and
6388 pageblock_order = order;
6390 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6393 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6394 * is unused as pageblock_order is set at compile-time. See
6395 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6398 void __init set_pageblock_order(void)
6402 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6404 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6405 unsigned long present_pages)
6407 unsigned long pages = spanned_pages;
6410 * Provide a more accurate estimation if there are holes within
6411 * the zone and SPARSEMEM is in use. If there are holes within the
6412 * zone, each populated memory region may cost us one or two extra
6413 * memmap pages due to alignment because memmap pages for each
6414 * populated regions may not be naturally aligned on page boundary.
6415 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6417 if (spanned_pages > present_pages + (present_pages >> 4) &&
6418 IS_ENABLED(CONFIG_SPARSEMEM))
6419 pages = present_pages;
6421 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6424 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6425 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6427 spin_lock_init(&pgdat->split_queue_lock);
6428 INIT_LIST_HEAD(&pgdat->split_queue);
6429 pgdat->split_queue_len = 0;
6432 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6435 #ifdef CONFIG_COMPACTION
6436 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6438 init_waitqueue_head(&pgdat->kcompactd_wait);
6441 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6444 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6446 pgdat_resize_init(pgdat);
6448 pgdat_init_split_queue(pgdat);
6449 pgdat_init_kcompactd(pgdat);
6451 init_waitqueue_head(&pgdat->kswapd_wait);
6452 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6454 pgdat_page_ext_init(pgdat);
6455 spin_lock_init(&pgdat->lru_lock);
6456 lruvec_init(node_lruvec(pgdat));
6459 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6460 unsigned long remaining_pages)
6462 atomic_long_set(&zone->managed_pages, remaining_pages);
6463 zone_set_nid(zone, nid);
6464 zone->name = zone_names[idx];
6465 zone->zone_pgdat = NODE_DATA(nid);
6466 spin_lock_init(&zone->lock);
6467 zone_seqlock_init(zone);
6468 zone_pcp_init(zone);
6472 * Set up the zone data structures
6473 * - init pgdat internals
6474 * - init all zones belonging to this node
6476 * NOTE: this function is only called during memory hotplug
6478 #ifdef CONFIG_MEMORY_HOTPLUG
6479 void __ref free_area_init_core_hotplug(int nid)
6482 pg_data_t *pgdat = NODE_DATA(nid);
6484 pgdat_init_internals(pgdat);
6485 for (z = 0; z < MAX_NR_ZONES; z++)
6486 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6491 * Set up the zone data structures:
6492 * - mark all pages reserved
6493 * - mark all memory queues empty
6494 * - clear the memory bitmaps
6496 * NOTE: pgdat should get zeroed by caller.
6497 * NOTE: this function is only called during early init.
6499 static void __init free_area_init_core(struct pglist_data *pgdat)
6502 int nid = pgdat->node_id;
6504 pgdat_init_internals(pgdat);
6505 pgdat->per_cpu_nodestats = &boot_nodestats;
6507 for (j = 0; j < MAX_NR_ZONES; j++) {
6508 struct zone *zone = pgdat->node_zones + j;
6509 unsigned long size, freesize, memmap_pages;
6510 unsigned long zone_start_pfn = zone->zone_start_pfn;
6512 size = zone->spanned_pages;
6513 freesize = zone->present_pages;
6516 * Adjust freesize so that it accounts for how much memory
6517 * is used by this zone for memmap. This affects the watermark
6518 * and per-cpu initialisations
6520 memmap_pages = calc_memmap_size(size, freesize);
6521 if (!is_highmem_idx(j)) {
6522 if (freesize >= memmap_pages) {
6523 freesize -= memmap_pages;
6526 " %s zone: %lu pages used for memmap\n",
6527 zone_names[j], memmap_pages);
6529 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6530 zone_names[j], memmap_pages, freesize);
6533 /* Account for reserved pages */
6534 if (j == 0 && freesize > dma_reserve) {
6535 freesize -= dma_reserve;
6536 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6537 zone_names[0], dma_reserve);
6540 if (!is_highmem_idx(j))
6541 nr_kernel_pages += freesize;
6542 /* Charge for highmem memmap if there are enough kernel pages */
6543 else if (nr_kernel_pages > memmap_pages * 2)
6544 nr_kernel_pages -= memmap_pages;
6545 nr_all_pages += freesize;
6548 * Set an approximate value for lowmem here, it will be adjusted
6549 * when the bootmem allocator frees pages into the buddy system.
6550 * And all highmem pages will be managed by the buddy system.
6552 zone_init_internals(zone, j, nid, freesize);
6557 set_pageblock_order();
6558 setup_usemap(pgdat, zone, zone_start_pfn, size);
6559 init_currently_empty_zone(zone, zone_start_pfn, size);
6560 memmap_init(size, nid, j, zone_start_pfn);
6564 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6565 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6567 unsigned long __maybe_unused start = 0;
6568 unsigned long __maybe_unused offset = 0;
6570 /* Skip empty nodes */
6571 if (!pgdat->node_spanned_pages)
6574 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6575 offset = pgdat->node_start_pfn - start;
6576 /* ia64 gets its own node_mem_map, before this, without bootmem */
6577 if (!pgdat->node_mem_map) {
6578 unsigned long size, end;
6582 * The zone's endpoints aren't required to be MAX_ORDER
6583 * aligned but the node_mem_map endpoints must be in order
6584 * for the buddy allocator to function correctly.
6586 end = pgdat_end_pfn(pgdat);
6587 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6588 size = (end - start) * sizeof(struct page);
6589 map = memblock_alloc_node_nopanic(size, pgdat->node_id);
6590 pgdat->node_mem_map = map + offset;
6592 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6593 __func__, pgdat->node_id, (unsigned long)pgdat,
6594 (unsigned long)pgdat->node_mem_map);
6595 #ifndef CONFIG_NEED_MULTIPLE_NODES
6597 * With no DISCONTIG, the global mem_map is just set as node 0's
6599 if (pgdat == NODE_DATA(0)) {
6600 mem_map = NODE_DATA(0)->node_mem_map;
6601 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6602 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6604 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6609 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6610 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6612 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6613 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6615 pgdat->first_deferred_pfn = ULONG_MAX;
6618 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6621 void __init free_area_init_node(int nid, unsigned long *zones_size,
6622 unsigned long node_start_pfn,
6623 unsigned long *zholes_size)
6625 pg_data_t *pgdat = NODE_DATA(nid);
6626 unsigned long start_pfn = 0;
6627 unsigned long end_pfn = 0;
6629 /* pg_data_t should be reset to zero when it's allocated */
6630 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6632 pgdat->node_id = nid;
6633 pgdat->node_start_pfn = node_start_pfn;
6634 pgdat->per_cpu_nodestats = NULL;
6635 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6636 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6637 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6638 (u64)start_pfn << PAGE_SHIFT,
6639 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6641 start_pfn = node_start_pfn;
6643 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6644 zones_size, zholes_size);
6646 alloc_node_mem_map(pgdat);
6647 pgdat_set_deferred_range(pgdat);
6649 free_area_init_core(pgdat);
6652 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6654 * Zero all valid struct pages in range [spfn, epfn), return number of struct
6657 static u64 zero_pfn_range(unsigned long spfn, unsigned long epfn)
6662 for (pfn = spfn; pfn < epfn; pfn++) {
6663 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6664 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6665 + pageblock_nr_pages - 1;
6668 mm_zero_struct_page(pfn_to_page(pfn));
6676 * Only struct pages that are backed by physical memory are zeroed and
6677 * initialized by going through __init_single_page(). But, there are some
6678 * struct pages which are reserved in memblock allocator and their fields
6679 * may be accessed (for example page_to_pfn() on some configuration accesses
6680 * flags). We must explicitly zero those struct pages.
6682 * This function also addresses a similar issue where struct pages are left
6683 * uninitialized because the physical address range is not covered by
6684 * memblock.memory or memblock.reserved. That could happen when memblock
6685 * layout is manually configured via memmap=.
6687 void __init zero_resv_unavail(void)
6689 phys_addr_t start, end;
6691 phys_addr_t next = 0;
6694 * Loop through unavailable ranges not covered by memblock.memory.
6697 for_each_mem_range(i, &memblock.memory, NULL,
6698 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
6700 pgcnt += zero_pfn_range(PFN_DOWN(next), PFN_UP(start));
6703 pgcnt += zero_pfn_range(PFN_DOWN(next), max_pfn);
6706 * Struct pages that do not have backing memory. This could be because
6707 * firmware is using some of this memory, or for some other reasons.
6710 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
6712 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
6714 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6716 #if MAX_NUMNODES > 1
6718 * Figure out the number of possible node ids.
6720 void __init setup_nr_node_ids(void)
6722 unsigned int highest;
6724 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6725 nr_node_ids = highest + 1;
6730 * node_map_pfn_alignment - determine the maximum internode alignment
6732 * This function should be called after node map is populated and sorted.
6733 * It calculates the maximum power of two alignment which can distinguish
6736 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6737 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6738 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6739 * shifted, 1GiB is enough and this function will indicate so.
6741 * This is used to test whether pfn -> nid mapping of the chosen memory
6742 * model has fine enough granularity to avoid incorrect mapping for the
6743 * populated node map.
6745 * Returns the determined alignment in pfn's. 0 if there is no alignment
6746 * requirement (single node).
6748 unsigned long __init node_map_pfn_alignment(void)
6750 unsigned long accl_mask = 0, last_end = 0;
6751 unsigned long start, end, mask;
6755 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6756 if (!start || last_nid < 0 || last_nid == nid) {
6763 * Start with a mask granular enough to pin-point to the
6764 * start pfn and tick off bits one-by-one until it becomes
6765 * too coarse to separate the current node from the last.
6767 mask = ~((1 << __ffs(start)) - 1);
6768 while (mask && last_end <= (start & (mask << 1)))
6771 /* accumulate all internode masks */
6775 /* convert mask to number of pages */
6776 return ~accl_mask + 1;
6779 /* Find the lowest pfn for a node */
6780 static unsigned long __init find_min_pfn_for_node(int nid)
6782 unsigned long min_pfn = ULONG_MAX;
6783 unsigned long start_pfn;
6786 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6787 min_pfn = min(min_pfn, start_pfn);
6789 if (min_pfn == ULONG_MAX) {
6790 pr_warn("Could not find start_pfn for node %d\n", nid);
6798 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6800 * It returns the minimum PFN based on information provided via
6801 * memblock_set_node().
6803 unsigned long __init find_min_pfn_with_active_regions(void)
6805 return find_min_pfn_for_node(MAX_NUMNODES);
6809 * early_calculate_totalpages()
6810 * Sum pages in active regions for movable zone.
6811 * Populate N_MEMORY for calculating usable_nodes.
6813 static unsigned long __init early_calculate_totalpages(void)
6815 unsigned long totalpages = 0;
6816 unsigned long start_pfn, end_pfn;
6819 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6820 unsigned long pages = end_pfn - start_pfn;
6822 totalpages += pages;
6824 node_set_state(nid, N_MEMORY);
6830 * Find the PFN the Movable zone begins in each node. Kernel memory
6831 * is spread evenly between nodes as long as the nodes have enough
6832 * memory. When they don't, some nodes will have more kernelcore than
6835 static void __init find_zone_movable_pfns_for_nodes(void)
6838 unsigned long usable_startpfn;
6839 unsigned long kernelcore_node, kernelcore_remaining;
6840 /* save the state before borrow the nodemask */
6841 nodemask_t saved_node_state = node_states[N_MEMORY];
6842 unsigned long totalpages = early_calculate_totalpages();
6843 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6844 struct memblock_region *r;
6846 /* Need to find movable_zone earlier when movable_node is specified. */
6847 find_usable_zone_for_movable();
6850 * If movable_node is specified, ignore kernelcore and movablecore
6853 if (movable_node_is_enabled()) {
6854 for_each_memblock(memory, r) {
6855 if (!memblock_is_hotpluggable(r))
6860 usable_startpfn = PFN_DOWN(r->base);
6861 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6862 min(usable_startpfn, zone_movable_pfn[nid]) :
6870 * If kernelcore=mirror is specified, ignore movablecore option
6872 if (mirrored_kernelcore) {
6873 bool mem_below_4gb_not_mirrored = false;
6875 for_each_memblock(memory, r) {
6876 if (memblock_is_mirror(r))
6881 usable_startpfn = memblock_region_memory_base_pfn(r);
6883 if (usable_startpfn < 0x100000) {
6884 mem_below_4gb_not_mirrored = true;
6888 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6889 min(usable_startpfn, zone_movable_pfn[nid]) :
6893 if (mem_below_4gb_not_mirrored)
6894 pr_warn("This configuration results in unmirrored kernel memory.");
6900 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
6901 * amount of necessary memory.
6903 if (required_kernelcore_percent)
6904 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
6906 if (required_movablecore_percent)
6907 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
6911 * If movablecore= was specified, calculate what size of
6912 * kernelcore that corresponds so that memory usable for
6913 * any allocation type is evenly spread. If both kernelcore
6914 * and movablecore are specified, then the value of kernelcore
6915 * will be used for required_kernelcore if it's greater than
6916 * what movablecore would have allowed.
6918 if (required_movablecore) {
6919 unsigned long corepages;
6922 * Round-up so that ZONE_MOVABLE is at least as large as what
6923 * was requested by the user
6925 required_movablecore =
6926 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6927 required_movablecore = min(totalpages, required_movablecore);
6928 corepages = totalpages - required_movablecore;
6930 required_kernelcore = max(required_kernelcore, corepages);
6934 * If kernelcore was not specified or kernelcore size is larger
6935 * than totalpages, there is no ZONE_MOVABLE.
6937 if (!required_kernelcore || required_kernelcore >= totalpages)
6940 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6941 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6944 /* Spread kernelcore memory as evenly as possible throughout nodes */
6945 kernelcore_node = required_kernelcore / usable_nodes;
6946 for_each_node_state(nid, N_MEMORY) {
6947 unsigned long start_pfn, end_pfn;
6950 * Recalculate kernelcore_node if the division per node
6951 * now exceeds what is necessary to satisfy the requested
6952 * amount of memory for the kernel
6954 if (required_kernelcore < kernelcore_node)
6955 kernelcore_node = required_kernelcore / usable_nodes;
6958 * As the map is walked, we track how much memory is usable
6959 * by the kernel using kernelcore_remaining. When it is
6960 * 0, the rest of the node is usable by ZONE_MOVABLE
6962 kernelcore_remaining = kernelcore_node;
6964 /* Go through each range of PFNs within this node */
6965 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6966 unsigned long size_pages;
6968 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6969 if (start_pfn >= end_pfn)
6972 /* Account for what is only usable for kernelcore */
6973 if (start_pfn < usable_startpfn) {
6974 unsigned long kernel_pages;
6975 kernel_pages = min(end_pfn, usable_startpfn)
6978 kernelcore_remaining -= min(kernel_pages,
6979 kernelcore_remaining);
6980 required_kernelcore -= min(kernel_pages,
6981 required_kernelcore);
6983 /* Continue if range is now fully accounted */
6984 if (end_pfn <= usable_startpfn) {
6987 * Push zone_movable_pfn to the end so
6988 * that if we have to rebalance
6989 * kernelcore across nodes, we will
6990 * not double account here
6992 zone_movable_pfn[nid] = end_pfn;
6995 start_pfn = usable_startpfn;
6999 * The usable PFN range for ZONE_MOVABLE is from
7000 * start_pfn->end_pfn. Calculate size_pages as the
7001 * number of pages used as kernelcore
7003 size_pages = end_pfn - start_pfn;
7004 if (size_pages > kernelcore_remaining)
7005 size_pages = kernelcore_remaining;
7006 zone_movable_pfn[nid] = start_pfn + size_pages;
7009 * Some kernelcore has been met, update counts and
7010 * break if the kernelcore for this node has been
7013 required_kernelcore -= min(required_kernelcore,
7015 kernelcore_remaining -= size_pages;
7016 if (!kernelcore_remaining)
7022 * If there is still required_kernelcore, we do another pass with one
7023 * less node in the count. This will push zone_movable_pfn[nid] further
7024 * along on the nodes that still have memory until kernelcore is
7028 if (usable_nodes && required_kernelcore > usable_nodes)
7032 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7033 for (nid = 0; nid < MAX_NUMNODES; nid++)
7034 zone_movable_pfn[nid] =
7035 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7038 /* restore the node_state */
7039 node_states[N_MEMORY] = saved_node_state;
7042 /* Any regular or high memory on that node ? */
7043 static void check_for_memory(pg_data_t *pgdat, int nid)
7045 enum zone_type zone_type;
7047 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7048 struct zone *zone = &pgdat->node_zones[zone_type];
7049 if (populated_zone(zone)) {
7050 if (IS_ENABLED(CONFIG_HIGHMEM))
7051 node_set_state(nid, N_HIGH_MEMORY);
7052 if (zone_type <= ZONE_NORMAL)
7053 node_set_state(nid, N_NORMAL_MEMORY);
7060 * free_area_init_nodes - Initialise all pg_data_t and zone data
7061 * @max_zone_pfn: an array of max PFNs for each zone
7063 * This will call free_area_init_node() for each active node in the system.
7064 * Using the page ranges provided by memblock_set_node(), the size of each
7065 * zone in each node and their holes is calculated. If the maximum PFN
7066 * between two adjacent zones match, it is assumed that the zone is empty.
7067 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7068 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7069 * starts where the previous one ended. For example, ZONE_DMA32 starts
7070 * at arch_max_dma_pfn.
7072 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
7074 unsigned long start_pfn, end_pfn;
7077 /* Record where the zone boundaries are */
7078 memset(arch_zone_lowest_possible_pfn, 0,
7079 sizeof(arch_zone_lowest_possible_pfn));
7080 memset(arch_zone_highest_possible_pfn, 0,
7081 sizeof(arch_zone_highest_possible_pfn));
7083 start_pfn = find_min_pfn_with_active_regions();
7085 for (i = 0; i < MAX_NR_ZONES; i++) {
7086 if (i == ZONE_MOVABLE)
7089 end_pfn = max(max_zone_pfn[i], start_pfn);
7090 arch_zone_lowest_possible_pfn[i] = start_pfn;
7091 arch_zone_highest_possible_pfn[i] = end_pfn;
7093 start_pfn = end_pfn;
7096 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7097 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7098 find_zone_movable_pfns_for_nodes();
7100 /* Print out the zone ranges */
7101 pr_info("Zone ranges:\n");
7102 for (i = 0; i < MAX_NR_ZONES; i++) {
7103 if (i == ZONE_MOVABLE)
7105 pr_info(" %-8s ", zone_names[i]);
7106 if (arch_zone_lowest_possible_pfn[i] ==
7107 arch_zone_highest_possible_pfn[i])
7110 pr_cont("[mem %#018Lx-%#018Lx]\n",
7111 (u64)arch_zone_lowest_possible_pfn[i]
7113 ((u64)arch_zone_highest_possible_pfn[i]
7114 << PAGE_SHIFT) - 1);
7117 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7118 pr_info("Movable zone start for each node\n");
7119 for (i = 0; i < MAX_NUMNODES; i++) {
7120 if (zone_movable_pfn[i])
7121 pr_info(" Node %d: %#018Lx\n", i,
7122 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7125 /* Print out the early node map */
7126 pr_info("Early memory node ranges\n");
7127 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
7128 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7129 (u64)start_pfn << PAGE_SHIFT,
7130 ((u64)end_pfn << PAGE_SHIFT) - 1);
7132 /* Initialise every node */
7133 mminit_verify_pageflags_layout();
7134 setup_nr_node_ids();
7135 zero_resv_unavail();
7136 for_each_online_node(nid) {
7137 pg_data_t *pgdat = NODE_DATA(nid);
7138 free_area_init_node(nid, NULL,
7139 find_min_pfn_for_node(nid), NULL);
7141 /* Any memory on that node */
7142 if (pgdat->node_present_pages)
7143 node_set_state(nid, N_MEMORY);
7144 check_for_memory(pgdat, nid);
7148 static int __init cmdline_parse_core(char *p, unsigned long *core,
7149 unsigned long *percent)
7151 unsigned long long coremem;
7157 /* Value may be a percentage of total memory, otherwise bytes */
7158 coremem = simple_strtoull(p, &endptr, 0);
7159 if (*endptr == '%') {
7160 /* Paranoid check for percent values greater than 100 */
7161 WARN_ON(coremem > 100);
7165 coremem = memparse(p, &p);
7166 /* Paranoid check that UL is enough for the coremem value */
7167 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7169 *core = coremem >> PAGE_SHIFT;
7176 * kernelcore=size sets the amount of memory for use for allocations that
7177 * cannot be reclaimed or migrated.
7179 static int __init cmdline_parse_kernelcore(char *p)
7181 /* parse kernelcore=mirror */
7182 if (parse_option_str(p, "mirror")) {
7183 mirrored_kernelcore = true;
7187 return cmdline_parse_core(p, &required_kernelcore,
7188 &required_kernelcore_percent);
7192 * movablecore=size sets the amount of memory for use for allocations that
7193 * can be reclaimed or migrated.
7195 static int __init cmdline_parse_movablecore(char *p)
7197 return cmdline_parse_core(p, &required_movablecore,
7198 &required_movablecore_percent);
7201 early_param("kernelcore", cmdline_parse_kernelcore);
7202 early_param("movablecore", cmdline_parse_movablecore);
7204 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
7206 void adjust_managed_page_count(struct page *page, long count)
7208 atomic_long_add(count, &page_zone(page)->managed_pages);
7209 totalram_pages_add(count);
7210 #ifdef CONFIG_HIGHMEM
7211 if (PageHighMem(page))
7212 totalhigh_pages_add(count);
7215 EXPORT_SYMBOL(adjust_managed_page_count);
7217 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7220 unsigned long pages = 0;
7222 start = (void *)PAGE_ALIGN((unsigned long)start);
7223 end = (void *)((unsigned long)end & PAGE_MASK);
7224 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7225 struct page *page = virt_to_page(pos);
7226 void *direct_map_addr;
7229 * 'direct_map_addr' might be different from 'pos'
7230 * because some architectures' virt_to_page()
7231 * work with aliases. Getting the direct map
7232 * address ensures that we get a _writeable_
7233 * alias for the memset().
7235 direct_map_addr = page_address(page);
7236 if ((unsigned int)poison <= 0xFF)
7237 memset(direct_map_addr, poison, PAGE_SIZE);
7239 free_reserved_page(page);
7243 pr_info("Freeing %s memory: %ldK\n",
7244 s, pages << (PAGE_SHIFT - 10));
7248 EXPORT_SYMBOL(free_reserved_area);
7250 #ifdef CONFIG_HIGHMEM
7251 void free_highmem_page(struct page *page)
7253 __free_reserved_page(page);
7254 totalram_pages_inc();
7255 atomic_long_inc(&page_zone(page)->managed_pages);
7256 totalhigh_pages_inc();
7261 void __init mem_init_print_info(const char *str)
7263 unsigned long physpages, codesize, datasize, rosize, bss_size;
7264 unsigned long init_code_size, init_data_size;
7266 physpages = get_num_physpages();
7267 codesize = _etext - _stext;
7268 datasize = _edata - _sdata;
7269 rosize = __end_rodata - __start_rodata;
7270 bss_size = __bss_stop - __bss_start;
7271 init_data_size = __init_end - __init_begin;
7272 init_code_size = _einittext - _sinittext;
7275 * Detect special cases and adjust section sizes accordingly:
7276 * 1) .init.* may be embedded into .data sections
7277 * 2) .init.text.* may be out of [__init_begin, __init_end],
7278 * please refer to arch/tile/kernel/vmlinux.lds.S.
7279 * 3) .rodata.* may be embedded into .text or .data sections.
7281 #define adj_init_size(start, end, size, pos, adj) \
7283 if (start <= pos && pos < end && size > adj) \
7287 adj_init_size(__init_begin, __init_end, init_data_size,
7288 _sinittext, init_code_size);
7289 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7290 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7291 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7292 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7294 #undef adj_init_size
7296 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7297 #ifdef CONFIG_HIGHMEM
7301 nr_free_pages() << (PAGE_SHIFT - 10),
7302 physpages << (PAGE_SHIFT - 10),
7303 codesize >> 10, datasize >> 10, rosize >> 10,
7304 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7305 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7306 totalcma_pages << (PAGE_SHIFT - 10),
7307 #ifdef CONFIG_HIGHMEM
7308 totalhigh_pages() << (PAGE_SHIFT - 10),
7310 str ? ", " : "", str ? str : "");
7314 * set_dma_reserve - set the specified number of pages reserved in the first zone
7315 * @new_dma_reserve: The number of pages to mark reserved
7317 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7318 * In the DMA zone, a significant percentage may be consumed by kernel image
7319 * and other unfreeable allocations which can skew the watermarks badly. This
7320 * function may optionally be used to account for unfreeable pages in the
7321 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7322 * smaller per-cpu batchsize.
7324 void __init set_dma_reserve(unsigned long new_dma_reserve)
7326 dma_reserve = new_dma_reserve;
7329 void __init free_area_init(unsigned long *zones_size)
7331 zero_resv_unavail();
7332 free_area_init_node(0, zones_size,
7333 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
7336 static int page_alloc_cpu_dead(unsigned int cpu)
7339 lru_add_drain_cpu(cpu);
7343 * Spill the event counters of the dead processor
7344 * into the current processors event counters.
7345 * This artificially elevates the count of the current
7348 vm_events_fold_cpu(cpu);
7351 * Zero the differential counters of the dead processor
7352 * so that the vm statistics are consistent.
7354 * This is only okay since the processor is dead and cannot
7355 * race with what we are doing.
7357 cpu_vm_stats_fold(cpu);
7361 void __init page_alloc_init(void)
7365 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7366 "mm/page_alloc:dead", NULL,
7367 page_alloc_cpu_dead);
7372 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7373 * or min_free_kbytes changes.
7375 static void calculate_totalreserve_pages(void)
7377 struct pglist_data *pgdat;
7378 unsigned long reserve_pages = 0;
7379 enum zone_type i, j;
7381 for_each_online_pgdat(pgdat) {
7383 pgdat->totalreserve_pages = 0;
7385 for (i = 0; i < MAX_NR_ZONES; i++) {
7386 struct zone *zone = pgdat->node_zones + i;
7388 unsigned long managed_pages = zone_managed_pages(zone);
7390 /* Find valid and maximum lowmem_reserve in the zone */
7391 for (j = i; j < MAX_NR_ZONES; j++) {
7392 if (zone->lowmem_reserve[j] > max)
7393 max = zone->lowmem_reserve[j];
7396 /* we treat the high watermark as reserved pages. */
7397 max += high_wmark_pages(zone);
7399 if (max > managed_pages)
7400 max = managed_pages;
7402 pgdat->totalreserve_pages += max;
7404 reserve_pages += max;
7407 totalreserve_pages = reserve_pages;
7411 * setup_per_zone_lowmem_reserve - called whenever
7412 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7413 * has a correct pages reserved value, so an adequate number of
7414 * pages are left in the zone after a successful __alloc_pages().
7416 static void setup_per_zone_lowmem_reserve(void)
7418 struct pglist_data *pgdat;
7419 enum zone_type j, idx;
7421 for_each_online_pgdat(pgdat) {
7422 for (j = 0; j < MAX_NR_ZONES; j++) {
7423 struct zone *zone = pgdat->node_zones + j;
7424 unsigned long managed_pages = zone_managed_pages(zone);
7426 zone->lowmem_reserve[j] = 0;
7430 struct zone *lower_zone;
7433 lower_zone = pgdat->node_zones + idx;
7435 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7436 sysctl_lowmem_reserve_ratio[idx] = 0;
7437 lower_zone->lowmem_reserve[j] = 0;
7439 lower_zone->lowmem_reserve[j] =
7440 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7442 managed_pages += zone_managed_pages(lower_zone);
7447 /* update totalreserve_pages */
7448 calculate_totalreserve_pages();
7451 static void __setup_per_zone_wmarks(void)
7453 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7454 unsigned long lowmem_pages = 0;
7456 unsigned long flags;
7458 /* Calculate total number of !ZONE_HIGHMEM pages */
7459 for_each_zone(zone) {
7460 if (!is_highmem(zone))
7461 lowmem_pages += zone_managed_pages(zone);
7464 for_each_zone(zone) {
7467 spin_lock_irqsave(&zone->lock, flags);
7468 tmp = (u64)pages_min * zone_managed_pages(zone);
7469 do_div(tmp, lowmem_pages);
7470 if (is_highmem(zone)) {
7472 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7473 * need highmem pages, so cap pages_min to a small
7476 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7477 * deltas control asynch page reclaim, and so should
7478 * not be capped for highmem.
7480 unsigned long min_pages;
7482 min_pages = zone_managed_pages(zone) / 1024;
7483 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7484 zone->_watermark[WMARK_MIN] = min_pages;
7487 * If it's a lowmem zone, reserve a number of pages
7488 * proportionate to the zone's size.
7490 zone->_watermark[WMARK_MIN] = tmp;
7494 * Set the kswapd watermarks distance according to the
7495 * scale factor in proportion to available memory, but
7496 * ensure a minimum size on small systems.
7498 tmp = max_t(u64, tmp >> 2,
7499 mult_frac(zone_managed_pages(zone),
7500 watermark_scale_factor, 10000));
7502 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7503 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7504 zone->watermark_boost = 0;
7506 spin_unlock_irqrestore(&zone->lock, flags);
7509 /* update totalreserve_pages */
7510 calculate_totalreserve_pages();
7514 * setup_per_zone_wmarks - called when min_free_kbytes changes
7515 * or when memory is hot-{added|removed}
7517 * Ensures that the watermark[min,low,high] values for each zone are set
7518 * correctly with respect to min_free_kbytes.
7520 void setup_per_zone_wmarks(void)
7522 static DEFINE_SPINLOCK(lock);
7525 __setup_per_zone_wmarks();
7530 * Initialise min_free_kbytes.
7532 * For small machines we want it small (128k min). For large machines
7533 * we want it large (64MB max). But it is not linear, because network
7534 * bandwidth does not increase linearly with machine size. We use
7536 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7537 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7553 int __meminit init_per_zone_wmark_min(void)
7555 unsigned long lowmem_kbytes;
7556 int new_min_free_kbytes;
7558 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7559 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7561 if (new_min_free_kbytes > user_min_free_kbytes) {
7562 min_free_kbytes = new_min_free_kbytes;
7563 if (min_free_kbytes < 128)
7564 min_free_kbytes = 128;
7565 if (min_free_kbytes > 65536)
7566 min_free_kbytes = 65536;
7568 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7569 new_min_free_kbytes, user_min_free_kbytes);
7571 setup_per_zone_wmarks();
7572 refresh_zone_stat_thresholds();
7573 setup_per_zone_lowmem_reserve();
7576 setup_min_unmapped_ratio();
7577 setup_min_slab_ratio();
7582 core_initcall(init_per_zone_wmark_min)
7585 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7586 * that we can call two helper functions whenever min_free_kbytes
7589 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7590 void __user *buffer, size_t *length, loff_t *ppos)
7594 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7599 user_min_free_kbytes = min_free_kbytes;
7600 setup_per_zone_wmarks();
7605 int watermark_boost_factor_sysctl_handler(struct ctl_table *table, int write,
7606 void __user *buffer, size_t *length, loff_t *ppos)
7610 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7617 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7618 void __user *buffer, size_t *length, loff_t *ppos)
7622 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7627 setup_per_zone_wmarks();
7633 static void setup_min_unmapped_ratio(void)
7638 for_each_online_pgdat(pgdat)
7639 pgdat->min_unmapped_pages = 0;
7642 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7643 sysctl_min_unmapped_ratio) / 100;
7647 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7648 void __user *buffer, size_t *length, loff_t *ppos)
7652 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7656 setup_min_unmapped_ratio();
7661 static void setup_min_slab_ratio(void)
7666 for_each_online_pgdat(pgdat)
7667 pgdat->min_slab_pages = 0;
7670 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7671 sysctl_min_slab_ratio) / 100;
7674 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7675 void __user *buffer, size_t *length, loff_t *ppos)
7679 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7683 setup_min_slab_ratio();
7690 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7691 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7692 * whenever sysctl_lowmem_reserve_ratio changes.
7694 * The reserve ratio obviously has absolutely no relation with the
7695 * minimum watermarks. The lowmem reserve ratio can only make sense
7696 * if in function of the boot time zone sizes.
7698 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7699 void __user *buffer, size_t *length, loff_t *ppos)
7701 proc_dointvec_minmax(table, write, buffer, length, ppos);
7702 setup_per_zone_lowmem_reserve();
7707 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7708 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7709 * pagelist can have before it gets flushed back to buddy allocator.
7711 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7712 void __user *buffer, size_t *length, loff_t *ppos)
7715 int old_percpu_pagelist_fraction;
7718 mutex_lock(&pcp_batch_high_lock);
7719 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7721 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7722 if (!write || ret < 0)
7725 /* Sanity checking to avoid pcp imbalance */
7726 if (percpu_pagelist_fraction &&
7727 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7728 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7734 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7737 for_each_populated_zone(zone) {
7740 for_each_possible_cpu(cpu)
7741 pageset_set_high_and_batch(zone,
7742 per_cpu_ptr(zone->pageset, cpu));
7745 mutex_unlock(&pcp_batch_high_lock);
7750 int hashdist = HASHDIST_DEFAULT;
7752 static int __init set_hashdist(char *str)
7756 hashdist = simple_strtoul(str, &str, 0);
7759 __setup("hashdist=", set_hashdist);
7762 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7764 * Returns the number of pages that arch has reserved but
7765 * is not known to alloc_large_system_hash().
7767 static unsigned long __init arch_reserved_kernel_pages(void)
7774 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7775 * machines. As memory size is increased the scale is also increased but at
7776 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7777 * quadruples the scale is increased by one, which means the size of hash table
7778 * only doubles, instead of quadrupling as well.
7779 * Because 32-bit systems cannot have large physical memory, where this scaling
7780 * makes sense, it is disabled on such platforms.
7782 #if __BITS_PER_LONG > 32
7783 #define ADAPT_SCALE_BASE (64ul << 30)
7784 #define ADAPT_SCALE_SHIFT 2
7785 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7789 * allocate a large system hash table from bootmem
7790 * - it is assumed that the hash table must contain an exact power-of-2
7791 * quantity of entries
7792 * - limit is the number of hash buckets, not the total allocation size
7794 void *__init alloc_large_system_hash(const char *tablename,
7795 unsigned long bucketsize,
7796 unsigned long numentries,
7799 unsigned int *_hash_shift,
7800 unsigned int *_hash_mask,
7801 unsigned long low_limit,
7802 unsigned long high_limit)
7804 unsigned long long max = high_limit;
7805 unsigned long log2qty, size;
7809 /* allow the kernel cmdline to have a say */
7811 /* round applicable memory size up to nearest megabyte */
7812 numentries = nr_kernel_pages;
7813 numentries -= arch_reserved_kernel_pages();
7815 /* It isn't necessary when PAGE_SIZE >= 1MB */
7816 if (PAGE_SHIFT < 20)
7817 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7819 #if __BITS_PER_LONG > 32
7821 unsigned long adapt;
7823 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7824 adapt <<= ADAPT_SCALE_SHIFT)
7829 /* limit to 1 bucket per 2^scale bytes of low memory */
7830 if (scale > PAGE_SHIFT)
7831 numentries >>= (scale - PAGE_SHIFT);
7833 numentries <<= (PAGE_SHIFT - scale);
7835 /* Make sure we've got at least a 0-order allocation.. */
7836 if (unlikely(flags & HASH_SMALL)) {
7837 /* Makes no sense without HASH_EARLY */
7838 WARN_ON(!(flags & HASH_EARLY));
7839 if (!(numentries >> *_hash_shift)) {
7840 numentries = 1UL << *_hash_shift;
7841 BUG_ON(!numentries);
7843 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7844 numentries = PAGE_SIZE / bucketsize;
7846 numentries = roundup_pow_of_two(numentries);
7848 /* limit allocation size to 1/16 total memory by default */
7850 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7851 do_div(max, bucketsize);
7853 max = min(max, 0x80000000ULL);
7855 if (numentries < low_limit)
7856 numentries = low_limit;
7857 if (numentries > max)
7860 log2qty = ilog2(numentries);
7862 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
7864 size = bucketsize << log2qty;
7865 if (flags & HASH_EARLY) {
7866 if (flags & HASH_ZERO)
7867 table = memblock_alloc_nopanic(size,
7870 table = memblock_alloc_raw(size,
7872 } else if (hashdist) {
7873 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
7876 * If bucketsize is not a power-of-two, we may free
7877 * some pages at the end of hash table which
7878 * alloc_pages_exact() automatically does
7880 if (get_order(size) < MAX_ORDER) {
7881 table = alloc_pages_exact(size, gfp_flags);
7882 kmemleak_alloc(table, size, 1, gfp_flags);
7885 } while (!table && size > PAGE_SIZE && --log2qty);
7888 panic("Failed to allocate %s hash table\n", tablename);
7890 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7891 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7894 *_hash_shift = log2qty;
7896 *_hash_mask = (1 << log2qty) - 1;
7902 * This function checks whether pageblock includes unmovable pages or not.
7903 * If @count is not zero, it is okay to include less @count unmovable pages
7905 * PageLRU check without isolation or lru_lock could race so that
7906 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7907 * check without lock_page also may miss some movable non-lru pages at
7908 * race condition. So you can't expect this function should be exact.
7910 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7911 int migratetype, int flags)
7913 unsigned long pfn, iter, found;
7916 * TODO we could make this much more efficient by not checking every
7917 * page in the range if we know all of them are in MOVABLE_ZONE and
7918 * that the movable zone guarantees that pages are migratable but
7919 * the later is not the case right now unfortunatelly. E.g. movablecore
7920 * can still lead to having bootmem allocations in zone_movable.
7924 * CMA allocations (alloc_contig_range) really need to mark isolate
7925 * CMA pageblocks even when they are not movable in fact so consider
7926 * them movable here.
7928 if (is_migrate_cma(migratetype) &&
7929 is_migrate_cma(get_pageblock_migratetype(page)))
7932 pfn = page_to_pfn(page);
7933 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7934 unsigned long check = pfn + iter;
7936 if (!pfn_valid_within(check))
7939 page = pfn_to_page(check);
7941 if (PageReserved(page))
7945 * If the zone is movable and we have ruled out all reserved
7946 * pages then it should be reasonably safe to assume the rest
7949 if (zone_idx(zone) == ZONE_MOVABLE)
7953 * Hugepages are not in LRU lists, but they're movable.
7954 * We need not scan over tail pages bacause we don't
7955 * handle each tail page individually in migration.
7957 if (PageHuge(page)) {
7958 struct page *head = compound_head(page);
7959 unsigned int skip_pages;
7961 if (!hugepage_migration_supported(page_hstate(head)))
7964 skip_pages = (1 << compound_order(head)) - (page - head);
7965 iter += skip_pages - 1;
7970 * We can't use page_count without pin a page
7971 * because another CPU can free compound page.
7972 * This check already skips compound tails of THP
7973 * because their page->_refcount is zero at all time.
7975 if (!page_ref_count(page)) {
7976 if (PageBuddy(page))
7977 iter += (1 << page_order(page)) - 1;
7982 * The HWPoisoned page may be not in buddy system, and
7983 * page_count() is not 0.
7985 if ((flags & SKIP_HWPOISON) && PageHWPoison(page))
7988 if (__PageMovable(page))
7994 * If there are RECLAIMABLE pages, we need to check
7995 * it. But now, memory offline itself doesn't call
7996 * shrink_node_slabs() and it still to be fixed.
7999 * If the page is not RAM, page_count()should be 0.
8000 * we don't need more check. This is an _used_ not-movable page.
8002 * The problematic thing here is PG_reserved pages. PG_reserved
8003 * is set to both of a memory hole page and a _used_ kernel
8011 WARN_ON_ONCE(zone_idx(zone) == ZONE_MOVABLE);
8012 if (flags & REPORT_FAILURE)
8013 dump_page(pfn_to_page(pfn+iter), "unmovable page");
8017 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
8019 static unsigned long pfn_max_align_down(unsigned long pfn)
8021 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8022 pageblock_nr_pages) - 1);
8025 static unsigned long pfn_max_align_up(unsigned long pfn)
8027 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8028 pageblock_nr_pages));
8031 /* [start, end) must belong to a single zone. */
8032 static int __alloc_contig_migrate_range(struct compact_control *cc,
8033 unsigned long start, unsigned long end)
8035 /* This function is based on compact_zone() from compaction.c. */
8036 unsigned long nr_reclaimed;
8037 unsigned long pfn = start;
8038 unsigned int tries = 0;
8043 while (pfn < end || !list_empty(&cc->migratepages)) {
8044 if (fatal_signal_pending(current)) {
8049 if (list_empty(&cc->migratepages)) {
8050 cc->nr_migratepages = 0;
8051 pfn = isolate_migratepages_range(cc, pfn, end);
8057 } else if (++tries == 5) {
8058 ret = ret < 0 ? ret : -EBUSY;
8062 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8064 cc->nr_migratepages -= nr_reclaimed;
8066 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
8067 NULL, 0, cc->mode, MR_CONTIG_RANGE);
8070 putback_movable_pages(&cc->migratepages);
8077 * alloc_contig_range() -- tries to allocate given range of pages
8078 * @start: start PFN to allocate
8079 * @end: one-past-the-last PFN to allocate
8080 * @migratetype: migratetype of the underlaying pageblocks (either
8081 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8082 * in range must have the same migratetype and it must
8083 * be either of the two.
8084 * @gfp_mask: GFP mask to use during compaction
8086 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8087 * aligned. The PFN range must belong to a single zone.
8089 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8090 * pageblocks in the range. Once isolated, the pageblocks should not
8091 * be modified by others.
8093 * Returns zero on success or negative error code. On success all
8094 * pages which PFN is in [start, end) are allocated for the caller and
8095 * need to be freed with free_contig_range().
8097 int alloc_contig_range(unsigned long start, unsigned long end,
8098 unsigned migratetype, gfp_t gfp_mask)
8100 unsigned long outer_start, outer_end;
8104 struct compact_control cc = {
8105 .nr_migratepages = 0,
8107 .zone = page_zone(pfn_to_page(start)),
8108 .mode = MIGRATE_SYNC,
8109 .ignore_skip_hint = true,
8110 .no_set_skip_hint = true,
8111 .gfp_mask = current_gfp_context(gfp_mask),
8113 INIT_LIST_HEAD(&cc.migratepages);
8116 * What we do here is we mark all pageblocks in range as
8117 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8118 * have different sizes, and due to the way page allocator
8119 * work, we align the range to biggest of the two pages so
8120 * that page allocator won't try to merge buddies from
8121 * different pageblocks and change MIGRATE_ISOLATE to some
8122 * other migration type.
8124 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8125 * migrate the pages from an unaligned range (ie. pages that
8126 * we are interested in). This will put all the pages in
8127 * range back to page allocator as MIGRATE_ISOLATE.
8129 * When this is done, we take the pages in range from page
8130 * allocator removing them from the buddy system. This way
8131 * page allocator will never consider using them.
8133 * This lets us mark the pageblocks back as
8134 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8135 * aligned range but not in the unaligned, original range are
8136 * put back to page allocator so that buddy can use them.
8139 ret = start_isolate_page_range(pfn_max_align_down(start),
8140 pfn_max_align_up(end), migratetype, 0);
8145 * In case of -EBUSY, we'd like to know which page causes problem.
8146 * So, just fall through. test_pages_isolated() has a tracepoint
8147 * which will report the busy page.
8149 * It is possible that busy pages could become available before
8150 * the call to test_pages_isolated, and the range will actually be
8151 * allocated. So, if we fall through be sure to clear ret so that
8152 * -EBUSY is not accidentally used or returned to caller.
8154 ret = __alloc_contig_migrate_range(&cc, start, end);
8155 if (ret && ret != -EBUSY)
8160 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8161 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8162 * more, all pages in [start, end) are free in page allocator.
8163 * What we are going to do is to allocate all pages from
8164 * [start, end) (that is remove them from page allocator).
8166 * The only problem is that pages at the beginning and at the
8167 * end of interesting range may be not aligned with pages that
8168 * page allocator holds, ie. they can be part of higher order
8169 * pages. Because of this, we reserve the bigger range and
8170 * once this is done free the pages we are not interested in.
8172 * We don't have to hold zone->lock here because the pages are
8173 * isolated thus they won't get removed from buddy.
8176 lru_add_drain_all();
8177 drain_all_pages(cc.zone);
8180 outer_start = start;
8181 while (!PageBuddy(pfn_to_page(outer_start))) {
8182 if (++order >= MAX_ORDER) {
8183 outer_start = start;
8186 outer_start &= ~0UL << order;
8189 if (outer_start != start) {
8190 order = page_order(pfn_to_page(outer_start));
8193 * outer_start page could be small order buddy page and
8194 * it doesn't include start page. Adjust outer_start
8195 * in this case to report failed page properly
8196 * on tracepoint in test_pages_isolated()
8198 if (outer_start + (1UL << order) <= start)
8199 outer_start = start;
8202 /* Make sure the range is really isolated. */
8203 if (test_pages_isolated(outer_start, end, false)) {
8204 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8205 __func__, outer_start, end);
8210 /* Grab isolated pages from freelists. */
8211 outer_end = isolate_freepages_range(&cc, outer_start, end);
8217 /* Free head and tail (if any) */
8218 if (start != outer_start)
8219 free_contig_range(outer_start, start - outer_start);
8220 if (end != outer_end)
8221 free_contig_range(end, outer_end - end);
8224 undo_isolate_page_range(pfn_max_align_down(start),
8225 pfn_max_align_up(end), migratetype);
8229 void free_contig_range(unsigned long pfn, unsigned nr_pages)
8231 unsigned int count = 0;
8233 for (; nr_pages--; pfn++) {
8234 struct page *page = pfn_to_page(pfn);
8236 count += page_count(page) != 1;
8239 WARN(count != 0, "%d pages are still in use!\n", count);
8243 #ifdef CONFIG_MEMORY_HOTPLUG
8245 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8246 * page high values need to be recalulated.
8248 void __meminit zone_pcp_update(struct zone *zone)
8251 mutex_lock(&pcp_batch_high_lock);
8252 for_each_possible_cpu(cpu)
8253 pageset_set_high_and_batch(zone,
8254 per_cpu_ptr(zone->pageset, cpu));
8255 mutex_unlock(&pcp_batch_high_lock);
8259 void zone_pcp_reset(struct zone *zone)
8261 unsigned long flags;
8263 struct per_cpu_pageset *pset;
8265 /* avoid races with drain_pages() */
8266 local_irq_save(flags);
8267 if (zone->pageset != &boot_pageset) {
8268 for_each_online_cpu(cpu) {
8269 pset = per_cpu_ptr(zone->pageset, cpu);
8270 drain_zonestat(zone, pset);
8272 free_percpu(zone->pageset);
8273 zone->pageset = &boot_pageset;
8275 local_irq_restore(flags);
8278 #ifdef CONFIG_MEMORY_HOTREMOVE
8280 * All pages in the range must be in a single zone and isolated
8281 * before calling this.
8284 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8288 unsigned int order, i;
8290 unsigned long flags;
8291 /* find the first valid pfn */
8292 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8297 offline_mem_sections(pfn, end_pfn);
8298 zone = page_zone(pfn_to_page(pfn));
8299 spin_lock_irqsave(&zone->lock, flags);
8301 while (pfn < end_pfn) {
8302 if (!pfn_valid(pfn)) {
8306 page = pfn_to_page(pfn);
8308 * The HWPoisoned page may be not in buddy system, and
8309 * page_count() is not 0.
8311 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8313 SetPageReserved(page);
8317 BUG_ON(page_count(page));
8318 BUG_ON(!PageBuddy(page));
8319 order = page_order(page);
8320 #ifdef CONFIG_DEBUG_VM
8321 pr_info("remove from free list %lx %d %lx\n",
8322 pfn, 1 << order, end_pfn);
8324 list_del(&page->lru);
8325 rmv_page_order(page);
8326 zone->free_area[order].nr_free--;
8327 for (i = 0; i < (1 << order); i++)
8328 SetPageReserved((page+i));
8329 pfn += (1 << order);
8331 spin_unlock_irqrestore(&zone->lock, flags);
8335 bool is_free_buddy_page(struct page *page)
8337 struct zone *zone = page_zone(page);
8338 unsigned long pfn = page_to_pfn(page);
8339 unsigned long flags;
8342 spin_lock_irqsave(&zone->lock, flags);
8343 for (order = 0; order < MAX_ORDER; order++) {
8344 struct page *page_head = page - (pfn & ((1 << order) - 1));
8346 if (PageBuddy(page_head) && page_order(page_head) >= order)
8349 spin_unlock_irqrestore(&zone->lock, flags);
8351 return order < MAX_ORDER;
8354 #ifdef CONFIG_MEMORY_FAILURE
8356 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8357 * test is performed under the zone lock to prevent a race against page
8360 bool set_hwpoison_free_buddy_page(struct page *page)
8362 struct zone *zone = page_zone(page);
8363 unsigned long pfn = page_to_pfn(page);
8364 unsigned long flags;
8366 bool hwpoisoned = false;
8368 spin_lock_irqsave(&zone->lock, flags);
8369 for (order = 0; order < MAX_ORDER; order++) {
8370 struct page *page_head = page - (pfn & ((1 << order) - 1));
8372 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8373 if (!TestSetPageHWPoison(page))
8378 spin_unlock_irqrestore(&zone->lock, flags);