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/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/notifier.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/memremap.h>
47 #include <linux/stop_machine.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/page_ext.h>
54 #include <linux/debugobjects.h>
55 #include <linux/kmemleak.h>
56 #include <linux/compaction.h>
57 #include <trace/events/kmem.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/page_ext.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
66 #include <linux/memcontrol.h>
67 #include <linux/khugepaged.h>
69 #include <asm/sections.h>
70 #include <asm/tlbflush.h>
71 #include <asm/div64.h>
74 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
75 static DEFINE_MUTEX(pcp_batch_high_lock);
76 #define MIN_PERCPU_PAGELIST_FRACTION (8)
78 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
79 DEFINE_PER_CPU(int, numa_node);
80 EXPORT_PER_CPU_SYMBOL(numa_node);
83 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
85 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
86 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
87 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
88 * defined in <linux/topology.h>.
90 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
91 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
92 int _node_numa_mem_[MAX_NUMNODES];
95 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
96 volatile unsigned long latent_entropy __latent_entropy;
97 EXPORT_SYMBOL(latent_entropy);
101 * Array of node states.
103 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
104 [N_POSSIBLE] = NODE_MASK_ALL,
105 [N_ONLINE] = { { [0] = 1UL } },
107 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
108 #ifdef CONFIG_HIGHMEM
109 [N_HIGH_MEMORY] = { { [0] = 1UL } },
111 #ifdef CONFIG_MOVABLE_NODE
112 [N_MEMORY] = { { [0] = 1UL } },
114 [N_CPU] = { { [0] = 1UL } },
117 EXPORT_SYMBOL(node_states);
119 /* Protect totalram_pages and zone->managed_pages */
120 static DEFINE_SPINLOCK(managed_page_count_lock);
122 unsigned long totalram_pages __read_mostly;
123 unsigned long totalreserve_pages __read_mostly;
124 unsigned long totalcma_pages __read_mostly;
126 int percpu_pagelist_fraction;
127 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
130 * A cached value of the page's pageblock's migratetype, used when the page is
131 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
132 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
133 * Also the migratetype set in the page does not necessarily match the pcplist
134 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
135 * other index - this ensures that it will be put on the correct CMA freelist.
137 static inline int get_pcppage_migratetype(struct page *page)
142 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
144 page->index = migratetype;
147 #ifdef CONFIG_PM_SLEEP
149 * The following functions are used by the suspend/hibernate code to temporarily
150 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
151 * while devices are suspended. To avoid races with the suspend/hibernate code,
152 * they should always be called with pm_mutex held (gfp_allowed_mask also should
153 * only be modified with pm_mutex held, unless the suspend/hibernate code is
154 * guaranteed not to run in parallel with that modification).
157 static gfp_t saved_gfp_mask;
159 void pm_restore_gfp_mask(void)
161 WARN_ON(!mutex_is_locked(&pm_mutex));
162 if (saved_gfp_mask) {
163 gfp_allowed_mask = saved_gfp_mask;
168 void pm_restrict_gfp_mask(void)
170 WARN_ON(!mutex_is_locked(&pm_mutex));
171 WARN_ON(saved_gfp_mask);
172 saved_gfp_mask = gfp_allowed_mask;
173 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
176 bool pm_suspended_storage(void)
178 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
182 #endif /* CONFIG_PM_SLEEP */
184 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
185 unsigned int pageblock_order __read_mostly;
188 static void __free_pages_ok(struct page *page, unsigned int order);
191 * results with 256, 32 in the lowmem_reserve sysctl:
192 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
193 * 1G machine -> (16M dma, 784M normal, 224M high)
194 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
195 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
196 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
198 * TBD: should special case ZONE_DMA32 machines here - in those we normally
199 * don't need any ZONE_NORMAL reservation
201 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
202 #ifdef CONFIG_ZONE_DMA
205 #ifdef CONFIG_ZONE_DMA32
208 #ifdef CONFIG_HIGHMEM
214 EXPORT_SYMBOL(totalram_pages);
216 static char * const zone_names[MAX_NR_ZONES] = {
217 #ifdef CONFIG_ZONE_DMA
220 #ifdef CONFIG_ZONE_DMA32
224 #ifdef CONFIG_HIGHMEM
228 #ifdef CONFIG_ZONE_DEVICE
233 char * const migratetype_names[MIGRATE_TYPES] = {
241 #ifdef CONFIG_MEMORY_ISOLATION
246 compound_page_dtor * const compound_page_dtors[] = {
249 #ifdef CONFIG_HUGETLB_PAGE
252 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
257 int min_free_kbytes = 1024;
258 int user_min_free_kbytes = -1;
259 int watermark_scale_factor = 10;
261 static unsigned long __meminitdata nr_kernel_pages;
262 static unsigned long __meminitdata nr_all_pages;
263 static unsigned long __meminitdata dma_reserve;
265 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
266 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
267 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
268 static unsigned long __initdata required_kernelcore;
269 static unsigned long __initdata required_movablecore;
270 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
271 static bool mirrored_kernelcore;
273 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
275 EXPORT_SYMBOL(movable_zone);
276 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
279 int nr_node_ids __read_mostly = MAX_NUMNODES;
280 int nr_online_nodes __read_mostly = 1;
281 EXPORT_SYMBOL(nr_node_ids);
282 EXPORT_SYMBOL(nr_online_nodes);
285 int page_group_by_mobility_disabled __read_mostly;
287 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
290 * Determine how many pages need to be initialized durig early boot
291 * (non-deferred initialization).
292 * The value of first_deferred_pfn will be set later, once non-deferred pages
293 * are initialized, but for now set it ULONG_MAX.
295 static inline void reset_deferred_meminit(pg_data_t *pgdat)
297 phys_addr_t start_addr, end_addr;
298 unsigned long max_pgcnt;
299 unsigned long reserved;
302 * Initialise at least 2G of a node but also take into account that
303 * two large system hashes that can take up 1GB for 0.25TB/node.
305 max_pgcnt = max(2UL << (30 - PAGE_SHIFT),
306 (pgdat->node_spanned_pages >> 8));
309 * Compensate the all the memblock reservations (e.g. crash kernel)
310 * from the initial estimation to make sure we will initialize enough
313 start_addr = PFN_PHYS(pgdat->node_start_pfn);
314 end_addr = PFN_PHYS(pgdat->node_start_pfn + max_pgcnt);
315 reserved = memblock_reserved_memory_within(start_addr, end_addr);
316 max_pgcnt += PHYS_PFN(reserved);
318 pgdat->static_init_pgcnt = min(max_pgcnt, pgdat->node_spanned_pages);
319 pgdat->first_deferred_pfn = ULONG_MAX;
322 /* Returns true if the struct page for the pfn is uninitialised */
323 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
325 int nid = early_pfn_to_nid(pfn);
327 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
334 * Returns false when the remaining initialisation should be deferred until
335 * later in the boot cycle when it can be parallelised.
337 static inline bool update_defer_init(pg_data_t *pgdat,
338 unsigned long pfn, unsigned long zone_end,
339 unsigned long *nr_initialised)
341 /* Always populate low zones for address-contrained allocations */
342 if (zone_end < pgdat_end_pfn(pgdat))
345 if ((*nr_initialised > pgdat->static_init_pgcnt) &&
346 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
347 pgdat->first_deferred_pfn = pfn;
354 static inline void reset_deferred_meminit(pg_data_t *pgdat)
358 static inline bool early_page_uninitialised(unsigned long pfn)
363 static inline bool update_defer_init(pg_data_t *pgdat,
364 unsigned long pfn, unsigned long zone_end,
365 unsigned long *nr_initialised)
371 /* Return a pointer to the bitmap storing bits affecting a block of pages */
372 static inline unsigned long *get_pageblock_bitmap(struct page *page,
375 #ifdef CONFIG_SPARSEMEM
376 return __pfn_to_section(pfn)->pageblock_flags;
378 return page_zone(page)->pageblock_flags;
379 #endif /* CONFIG_SPARSEMEM */
382 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
384 #ifdef CONFIG_SPARSEMEM
385 pfn &= (PAGES_PER_SECTION-1);
386 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
388 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
389 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
390 #endif /* CONFIG_SPARSEMEM */
394 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
395 * @page: The page within the block of interest
396 * @pfn: The target page frame number
397 * @end_bitidx: The last bit of interest to retrieve
398 * @mask: mask of bits that the caller is interested in
400 * Return: pageblock_bits flags
402 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
404 unsigned long end_bitidx,
407 unsigned long *bitmap;
408 unsigned long bitidx, word_bitidx;
411 bitmap = get_pageblock_bitmap(page, pfn);
412 bitidx = pfn_to_bitidx(page, pfn);
413 word_bitidx = bitidx / BITS_PER_LONG;
414 bitidx &= (BITS_PER_LONG-1);
416 word = bitmap[word_bitidx];
417 bitidx += end_bitidx;
418 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
421 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
422 unsigned long end_bitidx,
425 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
428 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
430 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
434 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
435 * @page: The page within the block of interest
436 * @flags: The flags to set
437 * @pfn: The target page frame number
438 * @end_bitidx: The last bit of interest
439 * @mask: mask of bits that the caller is interested in
441 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
443 unsigned long end_bitidx,
446 unsigned long *bitmap;
447 unsigned long bitidx, word_bitidx;
448 unsigned long old_word, word;
450 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
452 bitmap = get_pageblock_bitmap(page, pfn);
453 bitidx = pfn_to_bitidx(page, pfn);
454 word_bitidx = bitidx / BITS_PER_LONG;
455 bitidx &= (BITS_PER_LONG-1);
457 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
459 bitidx += end_bitidx;
460 mask <<= (BITS_PER_LONG - bitidx - 1);
461 flags <<= (BITS_PER_LONG - bitidx - 1);
463 word = READ_ONCE(bitmap[word_bitidx]);
465 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
466 if (word == old_word)
472 void set_pageblock_migratetype(struct page *page, int migratetype)
474 if (unlikely(page_group_by_mobility_disabled &&
475 migratetype < MIGRATE_PCPTYPES))
476 migratetype = MIGRATE_UNMOVABLE;
478 set_pageblock_flags_group(page, (unsigned long)migratetype,
479 PB_migrate, PB_migrate_end);
482 #ifdef CONFIG_DEBUG_VM
483 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
487 unsigned long pfn = page_to_pfn(page);
488 unsigned long sp, start_pfn;
491 seq = zone_span_seqbegin(zone);
492 start_pfn = zone->zone_start_pfn;
493 sp = zone->spanned_pages;
494 if (!zone_spans_pfn(zone, pfn))
496 } while (zone_span_seqretry(zone, seq));
499 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
500 pfn, zone_to_nid(zone), zone->name,
501 start_pfn, start_pfn + sp);
506 static int page_is_consistent(struct zone *zone, struct page *page)
508 if (!pfn_valid_within(page_to_pfn(page)))
510 if (zone != page_zone(page))
516 * Temporary debugging check for pages not lying within a given zone.
518 static int bad_range(struct zone *zone, struct page *page)
520 if (page_outside_zone_boundaries(zone, page))
522 if (!page_is_consistent(zone, page))
528 static inline int bad_range(struct zone *zone, struct page *page)
534 static void bad_page(struct page *page, const char *reason,
535 unsigned long bad_flags)
537 static unsigned long resume;
538 static unsigned long nr_shown;
539 static unsigned long nr_unshown;
542 * Allow a burst of 60 reports, then keep quiet for that minute;
543 * or allow a steady drip of one report per second.
545 if (nr_shown == 60) {
546 if (time_before(jiffies, resume)) {
552 "BUG: Bad page state: %lu messages suppressed\n",
559 resume = jiffies + 60 * HZ;
561 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
562 current->comm, page_to_pfn(page));
563 __dump_page(page, reason);
564 bad_flags &= page->flags;
566 pr_alert("bad because of flags: %#lx(%pGp)\n",
567 bad_flags, &bad_flags);
568 dump_page_owner(page);
573 /* Leave bad fields for debug, except PageBuddy could make trouble */
574 page_mapcount_reset(page); /* remove PageBuddy */
575 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
579 * Higher-order pages are called "compound pages". They are structured thusly:
581 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
583 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
584 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
586 * The first tail page's ->compound_dtor holds the offset in array of compound
587 * page destructors. See compound_page_dtors.
589 * The first tail page's ->compound_order holds the order of allocation.
590 * This usage means that zero-order pages may not be compound.
593 void free_compound_page(struct page *page)
595 __free_pages_ok(page, compound_order(page));
598 void prep_compound_page(struct page *page, unsigned int order)
601 int nr_pages = 1 << order;
603 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
604 set_compound_order(page, order);
606 for (i = 1; i < nr_pages; i++) {
607 struct page *p = page + i;
608 set_page_count(p, 0);
609 p->mapping = TAIL_MAPPING;
610 set_compound_head(p, page);
612 atomic_set(compound_mapcount_ptr(page), -1);
615 #ifdef CONFIG_DEBUG_PAGEALLOC
616 unsigned int _debug_guardpage_minorder;
617 bool _debug_pagealloc_enabled __read_mostly
618 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
619 EXPORT_SYMBOL(_debug_pagealloc_enabled);
620 bool _debug_guardpage_enabled __read_mostly;
622 static int __init early_debug_pagealloc(char *buf)
626 return kstrtobool(buf, &_debug_pagealloc_enabled);
628 early_param("debug_pagealloc", early_debug_pagealloc);
630 static bool need_debug_guardpage(void)
632 /* If we don't use debug_pagealloc, we don't need guard page */
633 if (!debug_pagealloc_enabled())
636 if (!debug_guardpage_minorder())
642 static void init_debug_guardpage(void)
644 if (!debug_pagealloc_enabled())
647 if (!debug_guardpage_minorder())
650 _debug_guardpage_enabled = true;
653 struct page_ext_operations debug_guardpage_ops = {
654 .need = need_debug_guardpage,
655 .init = init_debug_guardpage,
658 static int __init debug_guardpage_minorder_setup(char *buf)
662 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
663 pr_err("Bad debug_guardpage_minorder value\n");
666 _debug_guardpage_minorder = res;
667 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
670 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
672 static inline bool set_page_guard(struct zone *zone, struct page *page,
673 unsigned int order, int migratetype)
675 struct page_ext *page_ext;
677 if (!debug_guardpage_enabled())
680 if (order >= debug_guardpage_minorder())
683 page_ext = lookup_page_ext(page);
684 if (unlikely(!page_ext))
687 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
689 INIT_LIST_HEAD(&page->lru);
690 set_page_private(page, order);
691 /* Guard pages are not available for any usage */
692 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
697 static inline void clear_page_guard(struct zone *zone, struct page *page,
698 unsigned int order, int migratetype)
700 struct page_ext *page_ext;
702 if (!debug_guardpage_enabled())
705 page_ext = lookup_page_ext(page);
706 if (unlikely(!page_ext))
709 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
711 set_page_private(page, 0);
712 if (!is_migrate_isolate(migratetype))
713 __mod_zone_freepage_state(zone, (1 << order), migratetype);
716 struct page_ext_operations debug_guardpage_ops;
717 static inline bool set_page_guard(struct zone *zone, struct page *page,
718 unsigned int order, int migratetype) { return false; }
719 static inline void clear_page_guard(struct zone *zone, struct page *page,
720 unsigned int order, int migratetype) {}
723 static inline void set_page_order(struct page *page, unsigned int order)
725 set_page_private(page, order);
726 __SetPageBuddy(page);
729 static inline void rmv_page_order(struct page *page)
731 __ClearPageBuddy(page);
732 set_page_private(page, 0);
736 * This function checks whether a page is free && is the buddy
737 * we can do coalesce a page and its buddy if
738 * (a) the buddy is not in a hole &&
739 * (b) the buddy is in the buddy system &&
740 * (c) a page and its buddy have the same order &&
741 * (d) a page and its buddy are in the same zone.
743 * For recording whether a page is in the buddy system, we set ->_mapcount
744 * PAGE_BUDDY_MAPCOUNT_VALUE.
745 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
746 * serialized by zone->lock.
748 * For recording page's order, we use page_private(page).
750 static inline int page_is_buddy(struct page *page, struct page *buddy,
753 if (!pfn_valid_within(page_to_pfn(buddy)))
756 if (page_is_guard(buddy) && page_order(buddy) == order) {
757 if (page_zone_id(page) != page_zone_id(buddy))
760 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
765 if (PageBuddy(buddy) && page_order(buddy) == order) {
767 * zone check is done late to avoid uselessly
768 * calculating zone/node ids for pages that could
771 if (page_zone_id(page) != page_zone_id(buddy))
774 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
782 * Freeing function for a buddy system allocator.
784 * The concept of a buddy system is to maintain direct-mapped table
785 * (containing bit values) for memory blocks of various "orders".
786 * The bottom level table contains the map for the smallest allocatable
787 * units of memory (here, pages), and each level above it describes
788 * pairs of units from the levels below, hence, "buddies".
789 * At a high level, all that happens here is marking the table entry
790 * at the bottom level available, and propagating the changes upward
791 * as necessary, plus some accounting needed to play nicely with other
792 * parts of the VM system.
793 * At each level, we keep a list of pages, which are heads of continuous
794 * free pages of length of (1 << order) and marked with _mapcount
795 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
797 * So when we are allocating or freeing one, we can derive the state of the
798 * other. That is, if we allocate a small block, and both were
799 * free, the remainder of the region must be split into blocks.
800 * If a block is freed, and its buddy is also free, then this
801 * triggers coalescing into a block of larger size.
806 static inline void __free_one_page(struct page *page,
808 struct zone *zone, unsigned int order,
811 unsigned long page_idx;
812 unsigned long combined_idx;
813 unsigned long uninitialized_var(buddy_idx);
815 unsigned int max_order;
817 max_order = min_t(unsigned int, MAX_ORDER - 1, pageblock_order);
819 VM_BUG_ON(!zone_is_initialized(zone));
820 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
822 VM_BUG_ON(migratetype == -1);
823 if (likely(!is_migrate_isolate(migratetype)))
824 __mod_zone_freepage_state(zone, 1 << order, migratetype);
826 page_idx = pfn & ((1 << MAX_ORDER) - 1);
828 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
829 VM_BUG_ON_PAGE(bad_range(zone, page), page);
832 while (order < max_order) {
833 buddy_idx = __find_buddy_index(page_idx, order);
834 buddy = page + (buddy_idx - page_idx);
835 if (!page_is_buddy(page, buddy, order))
838 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
839 * merge with it and move up one order.
841 if (page_is_guard(buddy)) {
842 clear_page_guard(zone, buddy, order, migratetype);
844 list_del(&buddy->lru);
845 zone->free_area[order].nr_free--;
846 rmv_page_order(buddy);
848 combined_idx = buddy_idx & page_idx;
849 page = page + (combined_idx - page_idx);
850 page_idx = combined_idx;
853 if (order < MAX_ORDER - 1) {
854 /* If we are here, it means order is >= pageblock_order.
855 * We want to prevent merge between freepages on isolate
856 * pageblock and normal pageblock. Without this, pageblock
857 * isolation could cause incorrect freepage or CMA accounting.
859 * We don't want to hit this code for the more frequent
862 if (unlikely(has_isolate_pageblock(zone))) {
865 buddy_idx = __find_buddy_index(page_idx, order);
866 buddy = page + (buddy_idx - page_idx);
867 buddy_mt = get_pageblock_migratetype(buddy);
869 if (migratetype != buddy_mt
870 && (is_migrate_isolate(migratetype) ||
871 is_migrate_isolate(buddy_mt)))
874 max_order = order + 1;
875 goto continue_merging;
879 set_page_order(page, order);
882 * If this is not the largest possible page, check if the buddy
883 * of the next-highest order is free. If it is, it's possible
884 * that pages are being freed that will coalesce soon. In case,
885 * that is happening, add the free page to the tail of the list
886 * so it's less likely to be used soon and more likely to be merged
887 * as a higher order page
889 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
890 struct page *higher_page, *higher_buddy;
891 combined_idx = buddy_idx & page_idx;
892 higher_page = page + (combined_idx - page_idx);
893 buddy_idx = __find_buddy_index(combined_idx, order + 1);
894 higher_buddy = higher_page + (buddy_idx - combined_idx);
895 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
896 list_add_tail(&page->lru,
897 &zone->free_area[order].free_list[migratetype]);
902 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
904 zone->free_area[order].nr_free++;
908 * A bad page could be due to a number of fields. Instead of multiple branches,
909 * try and check multiple fields with one check. The caller must do a detailed
910 * check if necessary.
912 static inline bool page_expected_state(struct page *page,
913 unsigned long check_flags)
915 if (unlikely(atomic_read(&page->_mapcount) != -1))
918 if (unlikely((unsigned long)page->mapping |
919 page_ref_count(page) |
921 (unsigned long)page->mem_cgroup |
923 (page->flags & check_flags)))
929 static void free_pages_check_bad(struct page *page)
931 const char *bad_reason;
932 unsigned long bad_flags;
937 if (unlikely(atomic_read(&page->_mapcount) != -1))
938 bad_reason = "nonzero mapcount";
939 if (unlikely(page->mapping != NULL))
940 bad_reason = "non-NULL mapping";
941 if (unlikely(page_ref_count(page) != 0))
942 bad_reason = "nonzero _refcount";
943 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
944 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
945 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
948 if (unlikely(page->mem_cgroup))
949 bad_reason = "page still charged to cgroup";
951 bad_page(page, bad_reason, bad_flags);
954 static inline int free_pages_check(struct page *page)
956 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
959 /* Something has gone sideways, find it */
960 free_pages_check_bad(page);
964 static int free_tail_pages_check(struct page *head_page, struct page *page)
969 * We rely page->lru.next never has bit 0 set, unless the page
970 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
972 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
974 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
978 switch (page - head_page) {
980 /* the first tail page: ->mapping is compound_mapcount() */
981 if (unlikely(compound_mapcount(page))) {
982 bad_page(page, "nonzero compound_mapcount", 0);
988 * the second tail page: ->mapping is
989 * page_deferred_list().next -- ignore value.
993 if (page->mapping != TAIL_MAPPING) {
994 bad_page(page, "corrupted mapping in tail page", 0);
999 if (unlikely(!PageTail(page))) {
1000 bad_page(page, "PageTail not set", 0);
1003 if (unlikely(compound_head(page) != head_page)) {
1004 bad_page(page, "compound_head not consistent", 0);
1009 page->mapping = NULL;
1010 clear_compound_head(page);
1014 static __always_inline bool free_pages_prepare(struct page *page,
1015 unsigned int order, bool check_free)
1019 VM_BUG_ON_PAGE(PageTail(page), page);
1021 trace_mm_page_free(page, order);
1022 kmemcheck_free_shadow(page, order);
1025 * Check tail pages before head page information is cleared to
1026 * avoid checking PageCompound for order-0 pages.
1028 if (unlikely(order)) {
1029 bool compound = PageCompound(page);
1032 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1035 ClearPageDoubleMap(page);
1036 for (i = 1; i < (1 << order); i++) {
1038 bad += free_tail_pages_check(page, page + i);
1039 if (unlikely(free_pages_check(page + i))) {
1043 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1046 if (PageMappingFlags(page))
1047 page->mapping = NULL;
1048 if (memcg_kmem_enabled() && PageKmemcg(page))
1049 memcg_kmem_uncharge(page, order);
1051 bad += free_pages_check(page);
1055 page_cpupid_reset_last(page);
1056 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1057 reset_page_owner(page, order);
1059 if (!PageHighMem(page)) {
1060 debug_check_no_locks_freed(page_address(page),
1061 PAGE_SIZE << order);
1062 debug_check_no_obj_freed(page_address(page),
1063 PAGE_SIZE << order);
1065 arch_free_page(page, order);
1066 kernel_poison_pages(page, 1 << order, 0);
1067 kernel_map_pages(page, 1 << order, 0);
1068 kasan_free_pages(page, order);
1073 #ifdef CONFIG_DEBUG_VM
1074 static inline bool free_pcp_prepare(struct page *page)
1076 return free_pages_prepare(page, 0, true);
1079 static inline bool bulkfree_pcp_prepare(struct page *page)
1084 static bool free_pcp_prepare(struct page *page)
1086 return free_pages_prepare(page, 0, false);
1089 static bool bulkfree_pcp_prepare(struct page *page)
1091 return free_pages_check(page);
1093 #endif /* CONFIG_DEBUG_VM */
1096 * Frees a number of pages from the PCP lists
1097 * Assumes all pages on list are in same zone, and of same order.
1098 * count is the number of pages to free.
1100 * If the zone was previously in an "all pages pinned" state then look to
1101 * see if this freeing clears that state.
1103 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1104 * pinned" detection logic.
1106 static void free_pcppages_bulk(struct zone *zone, int count,
1107 struct per_cpu_pages *pcp)
1109 int migratetype = 0;
1111 unsigned long nr_scanned;
1112 bool isolated_pageblocks;
1114 spin_lock(&zone->lock);
1115 isolated_pageblocks = has_isolate_pageblock(zone);
1116 nr_scanned = node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED);
1118 __mod_node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED, -nr_scanned);
1121 * Ensure proper count is passed which otherwise would stuck in the
1122 * below while (list_empty(list)) loop.
1124 count = min(pcp->count, count);
1127 struct list_head *list;
1130 * Remove pages from lists in a round-robin fashion. A
1131 * batch_free count is maintained that is incremented when an
1132 * empty list is encountered. This is so more pages are freed
1133 * off fuller lists instead of spinning excessively around empty
1138 if (++migratetype == MIGRATE_PCPTYPES)
1140 list = &pcp->lists[migratetype];
1141 } while (list_empty(list));
1143 /* This is the only non-empty list. Free them all. */
1144 if (batch_free == MIGRATE_PCPTYPES)
1148 int mt; /* migratetype of the to-be-freed page */
1150 page = list_last_entry(list, struct page, lru);
1151 /* must delete as __free_one_page list manipulates */
1152 list_del(&page->lru);
1154 mt = get_pcppage_migratetype(page);
1155 /* MIGRATE_ISOLATE page should not go to pcplists */
1156 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1157 /* Pageblock could have been isolated meanwhile */
1158 if (unlikely(isolated_pageblocks))
1159 mt = get_pageblock_migratetype(page);
1161 if (bulkfree_pcp_prepare(page))
1164 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1165 trace_mm_page_pcpu_drain(page, 0, mt);
1166 } while (--count && --batch_free && !list_empty(list));
1168 spin_unlock(&zone->lock);
1171 static void free_one_page(struct zone *zone,
1172 struct page *page, unsigned long pfn,
1176 unsigned long nr_scanned;
1177 spin_lock(&zone->lock);
1178 nr_scanned = node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED);
1180 __mod_node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED, -nr_scanned);
1182 if (unlikely(has_isolate_pageblock(zone) ||
1183 is_migrate_isolate(migratetype))) {
1184 migratetype = get_pfnblock_migratetype(page, pfn);
1186 __free_one_page(page, pfn, zone, order, migratetype);
1187 spin_unlock(&zone->lock);
1190 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1191 unsigned long zone, int nid)
1193 set_page_links(page, zone, nid, pfn);
1194 init_page_count(page);
1195 page_mapcount_reset(page);
1196 page_cpupid_reset_last(page);
1198 INIT_LIST_HEAD(&page->lru);
1199 #ifdef WANT_PAGE_VIRTUAL
1200 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1201 if (!is_highmem_idx(zone))
1202 set_page_address(page, __va(pfn << PAGE_SHIFT));
1206 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1209 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
1212 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1213 static void init_reserved_page(unsigned long pfn)
1218 if (!early_page_uninitialised(pfn))
1221 nid = early_pfn_to_nid(pfn);
1222 pgdat = NODE_DATA(nid);
1224 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1225 struct zone *zone = &pgdat->node_zones[zid];
1227 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1230 __init_single_pfn(pfn, zid, nid);
1233 static inline void init_reserved_page(unsigned long pfn)
1236 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1239 * Initialised pages do not have PageReserved set. This function is
1240 * called for each range allocated by the bootmem allocator and
1241 * marks the pages PageReserved. The remaining valid pages are later
1242 * sent to the buddy page allocator.
1244 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1246 unsigned long start_pfn = PFN_DOWN(start);
1247 unsigned long end_pfn = PFN_UP(end);
1249 for (; start_pfn < end_pfn; start_pfn++) {
1250 if (pfn_valid(start_pfn)) {
1251 struct page *page = pfn_to_page(start_pfn);
1253 init_reserved_page(start_pfn);
1255 /* Avoid false-positive PageTail() */
1256 INIT_LIST_HEAD(&page->lru);
1258 SetPageReserved(page);
1263 static void __free_pages_ok(struct page *page, unsigned int order)
1265 unsigned long flags;
1267 unsigned long pfn = page_to_pfn(page);
1269 if (!free_pages_prepare(page, order, true))
1272 migratetype = get_pfnblock_migratetype(page, pfn);
1273 local_irq_save(flags);
1274 __count_vm_events(PGFREE, 1 << order);
1275 free_one_page(page_zone(page), page, pfn, order, migratetype);
1276 local_irq_restore(flags);
1279 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1281 unsigned int nr_pages = 1 << order;
1282 struct page *p = page;
1286 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1288 __ClearPageReserved(p);
1289 set_page_count(p, 0);
1291 __ClearPageReserved(p);
1292 set_page_count(p, 0);
1294 page_zone(page)->managed_pages += nr_pages;
1295 set_page_refcounted(page);
1296 __free_pages(page, order);
1299 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1300 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1302 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1304 int __meminit early_pfn_to_nid(unsigned long pfn)
1306 static DEFINE_SPINLOCK(early_pfn_lock);
1309 spin_lock(&early_pfn_lock);
1310 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1312 nid = first_online_node;
1313 spin_unlock(&early_pfn_lock);
1319 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1320 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1321 struct mminit_pfnnid_cache *state)
1325 nid = __early_pfn_to_nid(pfn, state);
1326 if (nid >= 0 && nid != node)
1331 /* Only safe to use early in boot when initialisation is single-threaded */
1332 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1334 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1339 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1343 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1344 struct mminit_pfnnid_cache *state)
1351 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1354 if (early_page_uninitialised(pfn))
1356 return __free_pages_boot_core(page, order);
1360 * Check that the whole (or subset of) a pageblock given by the interval of
1361 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1362 * with the migration of free compaction scanner. The scanners then need to
1363 * use only pfn_valid_within() check for arches that allow holes within
1366 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1368 * It's possible on some configurations to have a setup like node0 node1 node0
1369 * i.e. it's possible that all pages within a zones range of pages do not
1370 * belong to a single zone. We assume that a border between node0 and node1
1371 * can occur within a single pageblock, but not a node0 node1 node0
1372 * interleaving within a single pageblock. It is therefore sufficient to check
1373 * the first and last page of a pageblock and avoid checking each individual
1374 * page in a pageblock.
1376 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1377 unsigned long end_pfn, struct zone *zone)
1379 struct page *start_page;
1380 struct page *end_page;
1382 /* end_pfn is one past the range we are checking */
1385 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1388 start_page = pfn_to_page(start_pfn);
1390 if (page_zone(start_page) != zone)
1393 end_page = pfn_to_page(end_pfn);
1395 /* This gives a shorter code than deriving page_zone(end_page) */
1396 if (page_zone_id(start_page) != page_zone_id(end_page))
1402 void set_zone_contiguous(struct zone *zone)
1404 unsigned long block_start_pfn = zone->zone_start_pfn;
1405 unsigned long block_end_pfn;
1407 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1408 for (; block_start_pfn < zone_end_pfn(zone);
1409 block_start_pfn = block_end_pfn,
1410 block_end_pfn += pageblock_nr_pages) {
1412 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1414 if (!__pageblock_pfn_to_page(block_start_pfn,
1415 block_end_pfn, zone))
1420 /* We confirm that there is no hole */
1421 zone->contiguous = true;
1424 void clear_zone_contiguous(struct zone *zone)
1426 zone->contiguous = false;
1429 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1430 static void __init deferred_free_range(struct page *page,
1431 unsigned long pfn, int nr_pages)
1438 /* Free a large naturally-aligned chunk if possible */
1439 if (nr_pages == pageblock_nr_pages &&
1440 (pfn & (pageblock_nr_pages - 1)) == 0) {
1441 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1442 __free_pages_boot_core(page, pageblock_order);
1446 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1447 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1448 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1449 __free_pages_boot_core(page, 0);
1453 /* Completion tracking for deferred_init_memmap() threads */
1454 static atomic_t pgdat_init_n_undone __initdata;
1455 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1457 static inline void __init pgdat_init_report_one_done(void)
1459 if (atomic_dec_and_test(&pgdat_init_n_undone))
1460 complete(&pgdat_init_all_done_comp);
1463 /* Initialise remaining memory on a node */
1464 static int __init deferred_init_memmap(void *data)
1466 pg_data_t *pgdat = data;
1467 int nid = pgdat->node_id;
1468 struct mminit_pfnnid_cache nid_init_state = { };
1469 unsigned long start = jiffies;
1470 unsigned long nr_pages = 0;
1471 unsigned long walk_start, walk_end;
1474 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1475 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1477 if (first_init_pfn == ULONG_MAX) {
1478 pgdat_init_report_one_done();
1482 /* Bind memory initialisation thread to a local node if possible */
1483 if (!cpumask_empty(cpumask))
1484 set_cpus_allowed_ptr(current, cpumask);
1486 /* Sanity check boundaries */
1487 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1488 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1489 pgdat->first_deferred_pfn = ULONG_MAX;
1491 /* Only the highest zone is deferred so find it */
1492 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1493 zone = pgdat->node_zones + zid;
1494 if (first_init_pfn < zone_end_pfn(zone))
1498 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1499 unsigned long pfn, end_pfn;
1500 struct page *page = NULL;
1501 struct page *free_base_page = NULL;
1502 unsigned long free_base_pfn = 0;
1505 end_pfn = min(walk_end, zone_end_pfn(zone));
1506 pfn = first_init_pfn;
1507 if (pfn < walk_start)
1509 if (pfn < zone->zone_start_pfn)
1510 pfn = zone->zone_start_pfn;
1512 for (; pfn < end_pfn; pfn++) {
1513 if (!pfn_valid_within(pfn))
1517 * Ensure pfn_valid is checked every
1518 * pageblock_nr_pages for memory holes
1520 if ((pfn & (pageblock_nr_pages - 1)) == 0) {
1521 if (!pfn_valid(pfn)) {
1527 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1532 /* Minimise pfn page lookups and scheduler checks */
1533 if (page && (pfn & (pageblock_nr_pages - 1)) != 0) {
1536 nr_pages += nr_to_free;
1537 deferred_free_range(free_base_page,
1538 free_base_pfn, nr_to_free);
1539 free_base_page = NULL;
1540 free_base_pfn = nr_to_free = 0;
1542 page = pfn_to_page(pfn);
1547 VM_BUG_ON(page_zone(page) != zone);
1551 __init_single_page(page, pfn, zid, nid);
1552 if (!free_base_page) {
1553 free_base_page = page;
1554 free_base_pfn = pfn;
1559 /* Where possible, batch up pages for a single free */
1562 /* Free the current block of pages to allocator */
1563 nr_pages += nr_to_free;
1564 deferred_free_range(free_base_page, free_base_pfn,
1566 free_base_page = NULL;
1567 free_base_pfn = nr_to_free = 0;
1569 /* Free the last block of pages to allocator */
1570 nr_pages += nr_to_free;
1571 deferred_free_range(free_base_page, free_base_pfn, nr_to_free);
1573 first_init_pfn = max(end_pfn, first_init_pfn);
1576 /* Sanity check that the next zone really is unpopulated */
1577 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1579 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1580 jiffies_to_msecs(jiffies - start));
1582 pgdat_init_report_one_done();
1585 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1587 void __init page_alloc_init_late(void)
1591 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1594 /* There will be num_node_state(N_MEMORY) threads */
1595 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1596 for_each_node_state(nid, N_MEMORY) {
1597 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1600 /* Block until all are initialised */
1601 wait_for_completion(&pgdat_init_all_done_comp);
1603 /* Reinit limits that are based on free pages after the kernel is up */
1604 files_maxfiles_init();
1606 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1607 /* Discard memblock private memory */
1611 for_each_populated_zone(zone)
1612 set_zone_contiguous(zone);
1616 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1617 void __init init_cma_reserved_pageblock(struct page *page)
1619 unsigned i = pageblock_nr_pages;
1620 struct page *p = page;
1623 __ClearPageReserved(p);
1624 set_page_count(p, 0);
1627 set_pageblock_migratetype(page, MIGRATE_CMA);
1629 if (pageblock_order >= MAX_ORDER) {
1630 i = pageblock_nr_pages;
1633 set_page_refcounted(p);
1634 __free_pages(p, MAX_ORDER - 1);
1635 p += MAX_ORDER_NR_PAGES;
1636 } while (i -= MAX_ORDER_NR_PAGES);
1638 set_page_refcounted(page);
1639 __free_pages(page, pageblock_order);
1642 adjust_managed_page_count(page, pageblock_nr_pages);
1647 * The order of subdivision here is critical for the IO subsystem.
1648 * Please do not alter this order without good reasons and regression
1649 * testing. Specifically, as large blocks of memory are subdivided,
1650 * the order in which smaller blocks are delivered depends on the order
1651 * they're subdivided in this function. This is the primary factor
1652 * influencing the order in which pages are delivered to the IO
1653 * subsystem according to empirical testing, and this is also justified
1654 * by considering the behavior of a buddy system containing a single
1655 * large block of memory acted on by a series of small allocations.
1656 * This behavior is a critical factor in sglist merging's success.
1660 static inline void expand(struct zone *zone, struct page *page,
1661 int low, int high, struct free_area *area,
1664 unsigned long size = 1 << high;
1666 while (high > low) {
1670 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1673 * Mark as guard pages (or page), that will allow to
1674 * merge back to allocator when buddy will be freed.
1675 * Corresponding page table entries will not be touched,
1676 * pages will stay not present in virtual address space
1678 if (set_page_guard(zone, &page[size], high, migratetype))
1681 list_add(&page[size].lru, &area->free_list[migratetype]);
1683 set_page_order(&page[size], high);
1687 static void check_new_page_bad(struct page *page)
1689 const char *bad_reason = NULL;
1690 unsigned long bad_flags = 0;
1692 if (unlikely(atomic_read(&page->_mapcount) != -1))
1693 bad_reason = "nonzero mapcount";
1694 if (unlikely(page->mapping != NULL))
1695 bad_reason = "non-NULL mapping";
1696 if (unlikely(page_ref_count(page) != 0))
1697 bad_reason = "nonzero _count";
1698 if (unlikely(page->flags & __PG_HWPOISON)) {
1699 bad_reason = "HWPoisoned (hardware-corrupted)";
1700 bad_flags = __PG_HWPOISON;
1701 /* Don't complain about hwpoisoned pages */
1702 page_mapcount_reset(page); /* remove PageBuddy */
1705 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1706 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1707 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1710 if (unlikely(page->mem_cgroup))
1711 bad_reason = "page still charged to cgroup";
1713 bad_page(page, bad_reason, bad_flags);
1717 * This page is about to be returned from the page allocator
1719 static inline int check_new_page(struct page *page)
1721 if (likely(page_expected_state(page,
1722 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1725 check_new_page_bad(page);
1729 static inline bool free_pages_prezeroed(bool poisoned)
1731 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1732 page_poisoning_enabled() && poisoned;
1735 #ifdef CONFIG_DEBUG_VM
1736 static bool check_pcp_refill(struct page *page)
1741 static bool check_new_pcp(struct page *page)
1743 return check_new_page(page);
1746 static bool check_pcp_refill(struct page *page)
1748 return check_new_page(page);
1750 static bool check_new_pcp(struct page *page)
1754 #endif /* CONFIG_DEBUG_VM */
1756 static bool check_new_pages(struct page *page, unsigned int order)
1759 for (i = 0; i < (1 << order); i++) {
1760 struct page *p = page + i;
1762 if (unlikely(check_new_page(p)))
1769 inline void post_alloc_hook(struct page *page, unsigned int order,
1772 set_page_private(page, 0);
1773 set_page_refcounted(page);
1775 arch_alloc_page(page, order);
1776 kernel_map_pages(page, 1 << order, 1);
1777 kernel_poison_pages(page, 1 << order, 1);
1778 kasan_alloc_pages(page, order);
1779 set_page_owner(page, order, gfp_flags);
1782 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1783 unsigned int alloc_flags)
1786 bool poisoned = true;
1788 for (i = 0; i < (1 << order); i++) {
1789 struct page *p = page + i;
1791 poisoned &= page_is_poisoned(p);
1794 post_alloc_hook(page, order, gfp_flags);
1796 if (!free_pages_prezeroed(poisoned) && (gfp_flags & __GFP_ZERO))
1797 for (i = 0; i < (1 << order); i++)
1798 clear_highpage(page + i);
1800 if (order && (gfp_flags & __GFP_COMP))
1801 prep_compound_page(page, order);
1804 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1805 * allocate the page. The expectation is that the caller is taking
1806 * steps that will free more memory. The caller should avoid the page
1807 * being used for !PFMEMALLOC purposes.
1809 if (alloc_flags & ALLOC_NO_WATERMARKS)
1810 set_page_pfmemalloc(page);
1812 clear_page_pfmemalloc(page);
1816 * Go through the free lists for the given migratetype and remove
1817 * the smallest available page from the freelists
1820 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1823 unsigned int current_order;
1824 struct free_area *area;
1827 /* Find a page of the appropriate size in the preferred list */
1828 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1829 area = &(zone->free_area[current_order]);
1830 page = list_first_entry_or_null(&area->free_list[migratetype],
1834 list_del(&page->lru);
1835 rmv_page_order(page);
1837 expand(zone, page, order, current_order, area, migratetype);
1838 set_pcppage_migratetype(page, migratetype);
1847 * This array describes the order lists are fallen back to when
1848 * the free lists for the desirable migrate type are depleted
1850 static int fallbacks[MIGRATE_TYPES][4] = {
1851 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1852 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1853 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1855 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1857 #ifdef CONFIG_MEMORY_ISOLATION
1858 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1863 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1866 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1869 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1870 unsigned int order) { return NULL; }
1874 * Move the free pages in a range to the free lists of the requested type.
1875 * Note that start_page and end_pages are not aligned on a pageblock
1876 * boundary. If alignment is required, use move_freepages_block()
1878 int move_freepages(struct zone *zone,
1879 struct page *start_page, struct page *end_page,
1884 int pages_moved = 0;
1886 #ifndef CONFIG_HOLES_IN_ZONE
1888 * page_zone is not safe to call in this context when
1889 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1890 * anyway as we check zone boundaries in move_freepages_block().
1891 * Remove at a later date when no bug reports exist related to
1892 * grouping pages by mobility
1894 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1897 for (page = start_page; page <= end_page;) {
1898 if (!pfn_valid_within(page_to_pfn(page))) {
1903 /* Make sure we are not inadvertently changing nodes */
1904 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1906 if (!PageBuddy(page)) {
1911 order = page_order(page);
1912 list_move(&page->lru,
1913 &zone->free_area[order].free_list[migratetype]);
1915 pages_moved += 1 << order;
1921 int move_freepages_block(struct zone *zone, struct page *page,
1924 unsigned long start_pfn, end_pfn;
1925 struct page *start_page, *end_page;
1927 start_pfn = page_to_pfn(page);
1928 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1929 start_page = pfn_to_page(start_pfn);
1930 end_page = start_page + pageblock_nr_pages - 1;
1931 end_pfn = start_pfn + pageblock_nr_pages - 1;
1933 /* Do not cross zone boundaries */
1934 if (!zone_spans_pfn(zone, start_pfn))
1936 if (!zone_spans_pfn(zone, end_pfn))
1939 return move_freepages(zone, start_page, end_page, migratetype);
1942 static void change_pageblock_range(struct page *pageblock_page,
1943 int start_order, int migratetype)
1945 int nr_pageblocks = 1 << (start_order - pageblock_order);
1947 while (nr_pageblocks--) {
1948 set_pageblock_migratetype(pageblock_page, migratetype);
1949 pageblock_page += pageblock_nr_pages;
1954 * When we are falling back to another migratetype during allocation, try to
1955 * steal extra free pages from the same pageblocks to satisfy further
1956 * allocations, instead of polluting multiple pageblocks.
1958 * If we are stealing a relatively large buddy page, it is likely there will
1959 * be more free pages in the pageblock, so try to steal them all. For
1960 * reclaimable and unmovable allocations, we steal regardless of page size,
1961 * as fragmentation caused by those allocations polluting movable pageblocks
1962 * is worse than movable allocations stealing from unmovable and reclaimable
1965 static bool can_steal_fallback(unsigned int order, int start_mt)
1968 * Leaving this order check is intended, although there is
1969 * relaxed order check in next check. The reason is that
1970 * we can actually steal whole pageblock if this condition met,
1971 * but, below check doesn't guarantee it and that is just heuristic
1972 * so could be changed anytime.
1974 if (order >= pageblock_order)
1977 if (order >= pageblock_order / 2 ||
1978 start_mt == MIGRATE_RECLAIMABLE ||
1979 start_mt == MIGRATE_UNMOVABLE ||
1980 page_group_by_mobility_disabled)
1987 * This function implements actual steal behaviour. If order is large enough,
1988 * we can steal whole pageblock. If not, we first move freepages in this
1989 * pageblock and check whether half of pages are moved or not. If half of
1990 * pages are moved, we can change migratetype of pageblock and permanently
1991 * use it's pages as requested migratetype in the future.
1993 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1996 unsigned int current_order = page_order(page);
1999 /* Take ownership for orders >= pageblock_order */
2000 if (current_order >= pageblock_order) {
2001 change_pageblock_range(page, current_order, start_type);
2005 pages = move_freepages_block(zone, page, start_type);
2007 /* Claim the whole block if over half of it is free */
2008 if (pages >= (1 << (pageblock_order-1)) ||
2009 page_group_by_mobility_disabled)
2010 set_pageblock_migratetype(page, start_type);
2014 * Check whether there is a suitable fallback freepage with requested order.
2015 * If only_stealable is true, this function returns fallback_mt only if
2016 * we can steal other freepages all together. This would help to reduce
2017 * fragmentation due to mixed migratetype pages in one pageblock.
2019 int find_suitable_fallback(struct free_area *area, unsigned int order,
2020 int migratetype, bool only_stealable, bool *can_steal)
2025 if (area->nr_free == 0)
2030 fallback_mt = fallbacks[migratetype][i];
2031 if (fallback_mt == MIGRATE_TYPES)
2034 if (list_empty(&area->free_list[fallback_mt]))
2037 if (can_steal_fallback(order, migratetype))
2040 if (!only_stealable)
2051 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2052 * there are no empty page blocks that contain a page with a suitable order
2054 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2055 unsigned int alloc_order)
2058 unsigned long max_managed, flags;
2061 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2062 * Check is race-prone but harmless.
2064 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2065 if (zone->nr_reserved_highatomic >= max_managed)
2068 spin_lock_irqsave(&zone->lock, flags);
2070 /* Recheck the nr_reserved_highatomic limit under the lock */
2071 if (zone->nr_reserved_highatomic >= max_managed)
2075 mt = get_pageblock_migratetype(page);
2076 if (mt != MIGRATE_HIGHATOMIC &&
2077 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
2078 zone->nr_reserved_highatomic += pageblock_nr_pages;
2079 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2080 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
2084 spin_unlock_irqrestore(&zone->lock, flags);
2088 * Used when an allocation is about to fail under memory pressure. This
2089 * potentially hurts the reliability of high-order allocations when under
2090 * intense memory pressure but failed atomic allocations should be easier
2091 * to recover from than an OOM.
2093 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
2095 struct zonelist *zonelist = ac->zonelist;
2096 unsigned long flags;
2102 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2104 /* Preserve at least one pageblock */
2105 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
2108 spin_lock_irqsave(&zone->lock, flags);
2109 for (order = 0; order < MAX_ORDER; order++) {
2110 struct free_area *area = &(zone->free_area[order]);
2112 page = list_first_entry_or_null(
2113 &area->free_list[MIGRATE_HIGHATOMIC],
2119 * In page freeing path, migratetype change is racy so
2120 * we can counter several free pages in a pageblock
2121 * in this loop althoug we changed the pageblock type
2122 * from highatomic to ac->migratetype. So we should
2123 * adjust the count once.
2125 if (get_pageblock_migratetype(page) ==
2126 MIGRATE_HIGHATOMIC) {
2128 * It should never happen but changes to
2129 * locking could inadvertently allow a per-cpu
2130 * drain to add pages to MIGRATE_HIGHATOMIC
2131 * while unreserving so be safe and watch for
2134 zone->nr_reserved_highatomic -= min(
2136 zone->nr_reserved_highatomic);
2140 * Convert to ac->migratetype and avoid the normal
2141 * pageblock stealing heuristics. Minimally, the caller
2142 * is doing the work and needs the pages. More
2143 * importantly, if the block was always converted to
2144 * MIGRATE_UNMOVABLE or another type then the number
2145 * of pageblocks that cannot be completely freed
2148 set_pageblock_migratetype(page, ac->migratetype);
2149 move_freepages_block(zone, page, ac->migratetype);
2150 spin_unlock_irqrestore(&zone->lock, flags);
2153 spin_unlock_irqrestore(&zone->lock, flags);
2157 /* Remove an element from the buddy allocator from the fallback list */
2158 static inline struct page *
2159 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
2161 struct free_area *area;
2162 unsigned int current_order;
2167 /* Find the largest possible block of pages in the other list */
2168 for (current_order = MAX_ORDER-1;
2169 current_order >= order && current_order <= MAX_ORDER-1;
2171 area = &(zone->free_area[current_order]);
2172 fallback_mt = find_suitable_fallback(area, current_order,
2173 start_migratetype, false, &can_steal);
2174 if (fallback_mt == -1)
2177 page = list_first_entry(&area->free_list[fallback_mt],
2180 steal_suitable_fallback(zone, page, start_migratetype);
2182 /* Remove the page from the freelists */
2184 list_del(&page->lru);
2185 rmv_page_order(page);
2187 expand(zone, page, order, current_order, area,
2190 * The pcppage_migratetype may differ from pageblock's
2191 * migratetype depending on the decisions in
2192 * find_suitable_fallback(). This is OK as long as it does not
2193 * differ for MIGRATE_CMA pageblocks. Those can be used as
2194 * fallback only via special __rmqueue_cma_fallback() function
2196 set_pcppage_migratetype(page, start_migratetype);
2198 trace_mm_page_alloc_extfrag(page, order, current_order,
2199 start_migratetype, fallback_mt);
2208 * Do the hard work of removing an element from the buddy allocator.
2209 * Call me with the zone->lock already held.
2211 static struct page *__rmqueue(struct zone *zone, unsigned int order,
2216 page = __rmqueue_smallest(zone, order, migratetype);
2217 if (unlikely(!page)) {
2218 if (migratetype == MIGRATE_MOVABLE)
2219 page = __rmqueue_cma_fallback(zone, order);
2222 page = __rmqueue_fallback(zone, order, migratetype);
2225 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2230 * Obtain a specified number of elements from the buddy allocator, all under
2231 * a single hold of the lock, for efficiency. Add them to the supplied list.
2232 * Returns the number of new pages which were placed at *list.
2234 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2235 unsigned long count, struct list_head *list,
2236 int migratetype, bool cold)
2240 spin_lock(&zone->lock);
2241 for (i = 0; i < count; ++i) {
2242 struct page *page = __rmqueue(zone, order, migratetype);
2243 if (unlikely(page == NULL))
2246 if (unlikely(check_pcp_refill(page)))
2250 * Split buddy pages returned by expand() are received here
2251 * in physical page order. The page is added to the callers and
2252 * list and the list head then moves forward. From the callers
2253 * perspective, the linked list is ordered by page number in
2254 * some conditions. This is useful for IO devices that can
2255 * merge IO requests if the physical pages are ordered
2259 list_add(&page->lru, list);
2261 list_add_tail(&page->lru, list);
2264 if (is_migrate_cma(get_pcppage_migratetype(page)))
2265 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2270 * i pages were removed from the buddy list even if some leak due
2271 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2272 * on i. Do not confuse with 'alloced' which is the number of
2273 * pages added to the pcp list.
2275 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2276 spin_unlock(&zone->lock);
2282 * Called from the vmstat counter updater to drain pagesets of this
2283 * currently executing processor on remote nodes after they have
2286 * Note that this function must be called with the thread pinned to
2287 * a single processor.
2289 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2291 unsigned long flags;
2292 int to_drain, batch;
2294 local_irq_save(flags);
2295 batch = READ_ONCE(pcp->batch);
2296 to_drain = min(pcp->count, batch);
2298 free_pcppages_bulk(zone, to_drain, pcp);
2299 pcp->count -= to_drain;
2301 local_irq_restore(flags);
2306 * Drain pcplists of the indicated processor and zone.
2308 * The processor must either be the current processor and the
2309 * thread pinned to the current processor or a processor that
2312 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2314 unsigned long flags;
2315 struct per_cpu_pageset *pset;
2316 struct per_cpu_pages *pcp;
2318 local_irq_save(flags);
2319 pset = per_cpu_ptr(zone->pageset, cpu);
2323 free_pcppages_bulk(zone, pcp->count, pcp);
2326 local_irq_restore(flags);
2330 * Drain pcplists of all zones on the indicated processor.
2332 * The processor must either be the current processor and the
2333 * thread pinned to the current processor or a processor that
2336 static void drain_pages(unsigned int cpu)
2340 for_each_populated_zone(zone) {
2341 drain_pages_zone(cpu, zone);
2346 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2348 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2349 * the single zone's pages.
2351 void drain_local_pages(struct zone *zone)
2353 int cpu = smp_processor_id();
2356 drain_pages_zone(cpu, zone);
2362 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2364 * When zone parameter is non-NULL, spill just the single zone's pages.
2366 * Note that this code is protected against sending an IPI to an offline
2367 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2368 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2369 * nothing keeps CPUs from showing up after we populated the cpumask and
2370 * before the call to on_each_cpu_mask().
2372 void drain_all_pages(struct zone *zone)
2377 * Allocate in the BSS so we wont require allocation in
2378 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2380 static cpumask_t cpus_with_pcps;
2383 * We don't care about racing with CPU hotplug event
2384 * as offline notification will cause the notified
2385 * cpu to drain that CPU pcps and on_each_cpu_mask
2386 * disables preemption as part of its processing
2388 for_each_online_cpu(cpu) {
2389 struct per_cpu_pageset *pcp;
2391 bool has_pcps = false;
2394 pcp = per_cpu_ptr(zone->pageset, cpu);
2398 for_each_populated_zone(z) {
2399 pcp = per_cpu_ptr(z->pageset, cpu);
2400 if (pcp->pcp.count) {
2408 cpumask_set_cpu(cpu, &cpus_with_pcps);
2410 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2412 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2416 #ifdef CONFIG_HIBERNATION
2418 void mark_free_pages(struct zone *zone)
2420 unsigned long pfn, max_zone_pfn;
2421 unsigned long flags;
2422 unsigned int order, t;
2425 if (zone_is_empty(zone))
2428 spin_lock_irqsave(&zone->lock, flags);
2430 max_zone_pfn = zone_end_pfn(zone);
2431 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2432 if (pfn_valid(pfn)) {
2433 page = pfn_to_page(pfn);
2435 if (page_zone(page) != zone)
2438 if (!swsusp_page_is_forbidden(page))
2439 swsusp_unset_page_free(page);
2442 for_each_migratetype_order(order, t) {
2443 list_for_each_entry(page,
2444 &zone->free_area[order].free_list[t], lru) {
2447 pfn = page_to_pfn(page);
2448 for (i = 0; i < (1UL << order); i++)
2449 swsusp_set_page_free(pfn_to_page(pfn + i));
2452 spin_unlock_irqrestore(&zone->lock, flags);
2454 #endif /* CONFIG_PM */
2457 * Free a 0-order page
2458 * cold == true ? free a cold page : free a hot page
2460 void free_hot_cold_page(struct page *page, bool cold)
2462 struct zone *zone = page_zone(page);
2463 struct per_cpu_pages *pcp;
2464 unsigned long flags;
2465 unsigned long pfn = page_to_pfn(page);
2468 if (!free_pcp_prepare(page))
2471 migratetype = get_pfnblock_migratetype(page, pfn);
2472 set_pcppage_migratetype(page, migratetype);
2473 local_irq_save(flags);
2474 __count_vm_event(PGFREE);
2477 * We only track unmovable, reclaimable and movable on pcp lists.
2478 * Free ISOLATE pages back to the allocator because they are being
2479 * offlined but treat RESERVE as movable pages so we can get those
2480 * areas back if necessary. Otherwise, we may have to free
2481 * excessively into the page allocator
2483 if (migratetype >= MIGRATE_PCPTYPES) {
2484 if (unlikely(is_migrate_isolate(migratetype))) {
2485 free_one_page(zone, page, pfn, 0, migratetype);
2488 migratetype = MIGRATE_MOVABLE;
2491 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2493 list_add(&page->lru, &pcp->lists[migratetype]);
2495 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2497 if (pcp->count >= pcp->high) {
2498 unsigned long batch = READ_ONCE(pcp->batch);
2499 free_pcppages_bulk(zone, batch, pcp);
2500 pcp->count -= batch;
2504 local_irq_restore(flags);
2508 * Free a list of 0-order pages
2510 void free_hot_cold_page_list(struct list_head *list, bool cold)
2512 struct page *page, *next;
2514 list_for_each_entry_safe(page, next, list, lru) {
2515 trace_mm_page_free_batched(page, cold);
2516 free_hot_cold_page(page, cold);
2521 * split_page takes a non-compound higher-order page, and splits it into
2522 * n (1<<order) sub-pages: page[0..n]
2523 * Each sub-page must be freed individually.
2525 * Note: this is probably too low level an operation for use in drivers.
2526 * Please consult with lkml before using this in your driver.
2528 void split_page(struct page *page, unsigned int order)
2532 VM_BUG_ON_PAGE(PageCompound(page), page);
2533 VM_BUG_ON_PAGE(!page_count(page), page);
2535 #ifdef CONFIG_KMEMCHECK
2537 * Split shadow pages too, because free(page[0]) would
2538 * otherwise free the whole shadow.
2540 if (kmemcheck_page_is_tracked(page))
2541 split_page(virt_to_page(page[0].shadow), order);
2544 for (i = 1; i < (1 << order); i++)
2545 set_page_refcounted(page + i);
2546 split_page_owner(page, order);
2548 EXPORT_SYMBOL_GPL(split_page);
2550 int __isolate_free_page(struct page *page, unsigned int order)
2552 unsigned long watermark;
2556 BUG_ON(!PageBuddy(page));
2558 zone = page_zone(page);
2559 mt = get_pageblock_migratetype(page);
2561 if (!is_migrate_isolate(mt)) {
2563 * Obey watermarks as if the page was being allocated. We can
2564 * emulate a high-order watermark check with a raised order-0
2565 * watermark, because we already know our high-order page
2568 watermark = min_wmark_pages(zone) + (1UL << order);
2569 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2572 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2575 /* Remove page from free list */
2576 list_del(&page->lru);
2577 zone->free_area[order].nr_free--;
2578 rmv_page_order(page);
2581 * Set the pageblock if the isolated page is at least half of a
2584 if (order >= pageblock_order - 1) {
2585 struct page *endpage = page + (1 << order) - 1;
2586 for (; page < endpage; page += pageblock_nr_pages) {
2587 int mt = get_pageblock_migratetype(page);
2588 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2589 set_pageblock_migratetype(page,
2595 return 1UL << order;
2599 * Update NUMA hit/miss statistics
2601 * Must be called with interrupts disabled.
2603 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
2607 enum zone_stat_item local_stat = NUMA_LOCAL;
2609 if (z->node != numa_node_id())
2610 local_stat = NUMA_OTHER;
2612 if (z->node == preferred_zone->node)
2613 __inc_zone_state(z, NUMA_HIT);
2615 __inc_zone_state(z, NUMA_MISS);
2616 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
2618 __inc_zone_state(z, local_stat);
2623 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2626 struct page *buffered_rmqueue(struct zone *preferred_zone,
2627 struct zone *zone, unsigned int order,
2628 gfp_t gfp_flags, unsigned int alloc_flags,
2631 unsigned long flags;
2633 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2635 if (likely(order == 0)) {
2636 struct per_cpu_pages *pcp;
2637 struct list_head *list;
2639 local_irq_save(flags);
2641 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2642 list = &pcp->lists[migratetype];
2643 if (list_empty(list)) {
2644 pcp->count += rmqueue_bulk(zone, 0,
2647 if (unlikely(list_empty(list)))
2652 page = list_last_entry(list, struct page, lru);
2654 page = list_first_entry(list, struct page, lru);
2656 list_del(&page->lru);
2659 } while (check_new_pcp(page));
2662 * We most definitely don't want callers attempting to
2663 * allocate greater than order-1 page units with __GFP_NOFAIL.
2665 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2666 spin_lock_irqsave(&zone->lock, flags);
2670 if (alloc_flags & ALLOC_HARDER) {
2671 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2673 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2676 page = __rmqueue(zone, order, migratetype);
2677 } while (page && check_new_pages(page, order));
2678 spin_unlock(&zone->lock);
2681 __mod_zone_freepage_state(zone, -(1 << order),
2682 get_pcppage_migratetype(page));
2685 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2686 zone_statistics(preferred_zone, zone, gfp_flags);
2687 local_irq_restore(flags);
2689 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2693 local_irq_restore(flags);
2697 #ifdef CONFIG_FAIL_PAGE_ALLOC
2700 struct fault_attr attr;
2702 bool ignore_gfp_highmem;
2703 bool ignore_gfp_reclaim;
2705 } fail_page_alloc = {
2706 .attr = FAULT_ATTR_INITIALIZER,
2707 .ignore_gfp_reclaim = true,
2708 .ignore_gfp_highmem = true,
2712 static int __init setup_fail_page_alloc(char *str)
2714 return setup_fault_attr(&fail_page_alloc.attr, str);
2716 __setup("fail_page_alloc=", setup_fail_page_alloc);
2718 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2720 if (order < fail_page_alloc.min_order)
2722 if (gfp_mask & __GFP_NOFAIL)
2724 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2726 if (fail_page_alloc.ignore_gfp_reclaim &&
2727 (gfp_mask & __GFP_DIRECT_RECLAIM))
2730 return should_fail(&fail_page_alloc.attr, 1 << order);
2733 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2735 static int __init fail_page_alloc_debugfs(void)
2737 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2740 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2741 &fail_page_alloc.attr);
2743 return PTR_ERR(dir);
2745 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2746 &fail_page_alloc.ignore_gfp_reclaim))
2748 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2749 &fail_page_alloc.ignore_gfp_highmem))
2751 if (!debugfs_create_u32("min-order", mode, dir,
2752 &fail_page_alloc.min_order))
2757 debugfs_remove_recursive(dir);
2762 late_initcall(fail_page_alloc_debugfs);
2764 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2766 #else /* CONFIG_FAIL_PAGE_ALLOC */
2768 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2773 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2776 * Return true if free base pages are above 'mark'. For high-order checks it
2777 * will return true of the order-0 watermark is reached and there is at least
2778 * one free page of a suitable size. Checking now avoids taking the zone lock
2779 * to check in the allocation paths if no pages are free.
2781 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2782 int classzone_idx, unsigned int alloc_flags,
2787 const bool alloc_harder = (alloc_flags & ALLOC_HARDER);
2789 /* free_pages may go negative - that's OK */
2790 free_pages -= (1 << order) - 1;
2792 if (alloc_flags & ALLOC_HIGH)
2796 * If the caller does not have rights to ALLOC_HARDER then subtract
2797 * the high-atomic reserves. This will over-estimate the size of the
2798 * atomic reserve but it avoids a search.
2800 if (likely(!alloc_harder))
2801 free_pages -= z->nr_reserved_highatomic;
2806 /* If allocation can't use CMA areas don't use free CMA pages */
2807 if (!(alloc_flags & ALLOC_CMA))
2808 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2812 * Check watermarks for an order-0 allocation request. If these
2813 * are not met, then a high-order request also cannot go ahead
2814 * even if a suitable page happened to be free.
2816 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2819 /* If this is an order-0 request then the watermark is fine */
2823 /* For a high-order request, check at least one suitable page is free */
2824 for (o = order; o < MAX_ORDER; o++) {
2825 struct free_area *area = &z->free_area[o];
2831 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2832 if (!list_empty(&area->free_list[mt]))
2837 if ((alloc_flags & ALLOC_CMA) &&
2838 !list_empty(&area->free_list[MIGRATE_CMA])) {
2843 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
2849 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2850 int classzone_idx, unsigned int alloc_flags)
2852 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2853 zone_page_state(z, NR_FREE_PAGES));
2856 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
2857 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
2859 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2863 /* If allocation can't use CMA areas don't use free CMA pages */
2864 if (!(alloc_flags & ALLOC_CMA))
2865 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
2869 * Fast check for order-0 only. If this fails then the reserves
2870 * need to be calculated. There is a corner case where the check
2871 * passes but only the high-order atomic reserve are free. If
2872 * the caller is !atomic then it'll uselessly search the free
2873 * list. That corner case is then slower but it is harmless.
2875 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
2878 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2882 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2883 unsigned long mark, int classzone_idx)
2885 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2887 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2888 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2890 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2895 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2897 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
2900 #else /* CONFIG_NUMA */
2901 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2905 #endif /* CONFIG_NUMA */
2908 * get_page_from_freelist goes through the zonelist trying to allocate
2911 static struct page *
2912 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2913 const struct alloc_context *ac)
2915 struct zoneref *z = ac->preferred_zoneref;
2917 struct pglist_data *last_pgdat_dirty_limit = NULL;
2920 * Scan zonelist, looking for a zone with enough free.
2921 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2923 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
2928 if (cpusets_enabled() &&
2929 (alloc_flags & ALLOC_CPUSET) &&
2930 !__cpuset_zone_allowed(zone, gfp_mask))
2933 * When allocating a page cache page for writing, we
2934 * want to get it from a node that is within its dirty
2935 * limit, such that no single node holds more than its
2936 * proportional share of globally allowed dirty pages.
2937 * The dirty limits take into account the node's
2938 * lowmem reserves and high watermark so that kswapd
2939 * should be able to balance it without having to
2940 * write pages from its LRU list.
2942 * XXX: For now, allow allocations to potentially
2943 * exceed the per-node dirty limit in the slowpath
2944 * (spread_dirty_pages unset) before going into reclaim,
2945 * which is important when on a NUMA setup the allowed
2946 * nodes are together not big enough to reach the
2947 * global limit. The proper fix for these situations
2948 * will require awareness of nodes in the
2949 * dirty-throttling and the flusher threads.
2951 if (ac->spread_dirty_pages) {
2952 if (last_pgdat_dirty_limit == zone->zone_pgdat)
2955 if (!node_dirty_ok(zone->zone_pgdat)) {
2956 last_pgdat_dirty_limit = zone->zone_pgdat;
2961 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2962 if (!zone_watermark_fast(zone, order, mark,
2963 ac_classzone_idx(ac), alloc_flags)) {
2966 /* Checked here to keep the fast path fast */
2967 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2968 if (alloc_flags & ALLOC_NO_WATERMARKS)
2971 if (node_reclaim_mode == 0 ||
2972 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
2975 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
2977 case NODE_RECLAIM_NOSCAN:
2980 case NODE_RECLAIM_FULL:
2981 /* scanned but unreclaimable */
2984 /* did we reclaim enough */
2985 if (zone_watermark_ok(zone, order, mark,
2986 ac_classzone_idx(ac), alloc_flags))
2994 page = buffered_rmqueue(ac->preferred_zoneref->zone, zone, order,
2995 gfp_mask, alloc_flags, ac->migratetype);
2997 prep_new_page(page, order, gfp_mask, alloc_flags);
3000 * If this is a high-order atomic allocation then check
3001 * if the pageblock should be reserved for the future
3003 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3004 reserve_highatomic_pageblock(page, zone, order);
3014 * Large machines with many possible nodes should not always dump per-node
3015 * meminfo in irq context.
3017 static inline bool should_suppress_show_mem(void)
3022 ret = in_interrupt();
3027 static DEFINE_RATELIMIT_STATE(nopage_rs,
3028 DEFAULT_RATELIMIT_INTERVAL,
3029 DEFAULT_RATELIMIT_BURST);
3031 void warn_alloc(gfp_t gfp_mask, const char *fmt, ...)
3033 unsigned int filter = SHOW_MEM_FILTER_NODES;
3034 struct va_format vaf;
3037 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
3038 debug_guardpage_minorder() > 0)
3042 * This documents exceptions given to allocations in certain
3043 * contexts that are allowed to allocate outside current's set
3046 if (!(gfp_mask & __GFP_NOMEMALLOC))
3047 if (test_thread_flag(TIF_MEMDIE) ||
3048 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3049 filter &= ~SHOW_MEM_FILTER_NODES;
3050 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3051 filter &= ~SHOW_MEM_FILTER_NODES;
3053 pr_warn("%s: ", current->comm);
3055 va_start(args, fmt);
3058 pr_cont("%pV", &vaf);
3061 pr_cont(", mode:%#x(%pGg)\n", gfp_mask, &gfp_mask);
3064 if (!should_suppress_show_mem())
3068 static inline struct page *
3069 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3070 const struct alloc_context *ac, unsigned long *did_some_progress)
3072 struct oom_control oc = {
3073 .zonelist = ac->zonelist,
3074 .nodemask = ac->nodemask,
3076 .gfp_mask = gfp_mask,
3081 *did_some_progress = 0;
3084 * Acquire the oom lock. If that fails, somebody else is
3085 * making progress for us.
3087 if (!mutex_trylock(&oom_lock)) {
3088 *did_some_progress = 1;
3089 schedule_timeout_uninterruptible(1);
3094 * Go through the zonelist yet one more time, keep very high watermark
3095 * here, this is only to catch a parallel oom killing, we must fail if
3096 * we're still under heavy pressure.
3098 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
3099 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3103 if (!(gfp_mask & __GFP_NOFAIL)) {
3104 /* Coredumps can quickly deplete all memory reserves */
3105 if (current->flags & PF_DUMPCORE)
3107 /* The OOM killer will not help higher order allocs */
3108 if (order > PAGE_ALLOC_COSTLY_ORDER)
3110 /* The OOM killer does not needlessly kill tasks for lowmem */
3111 if (ac->high_zoneidx < ZONE_NORMAL)
3113 if (pm_suspended_storage())
3116 * XXX: GFP_NOFS allocations should rather fail than rely on
3117 * other request to make a forward progress.
3118 * We are in an unfortunate situation where out_of_memory cannot
3119 * do much for this context but let's try it to at least get
3120 * access to memory reserved if the current task is killed (see
3121 * out_of_memory). Once filesystems are ready to handle allocation
3122 * failures more gracefully we should just bail out here.
3125 /* The OOM killer may not free memory on a specific node */
3126 if (gfp_mask & __GFP_THISNODE)
3129 /* Exhausted what can be done so it's blamo time */
3130 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3131 *did_some_progress = 1;
3133 if (gfp_mask & __GFP_NOFAIL) {
3134 page = get_page_from_freelist(gfp_mask, order,
3135 ALLOC_NO_WATERMARKS|ALLOC_CPUSET, ac);
3137 * fallback to ignore cpuset restriction if our nodes
3141 page = get_page_from_freelist(gfp_mask, order,
3142 ALLOC_NO_WATERMARKS, ac);
3146 mutex_unlock(&oom_lock);
3151 * Maximum number of compaction retries wit a progress before OOM
3152 * killer is consider as the only way to move forward.
3154 #define MAX_COMPACT_RETRIES 16
3156 #ifdef CONFIG_COMPACTION
3157 /* Try memory compaction for high-order allocations before reclaim */
3158 static struct page *
3159 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3160 unsigned int alloc_flags, const struct alloc_context *ac,
3161 enum compact_priority prio, enum compact_result *compact_result)
3164 unsigned int noreclaim_flag = current->flags & PF_MEMALLOC;
3169 current->flags |= PF_MEMALLOC;
3170 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3172 current->flags = (current->flags & ~PF_MEMALLOC) | noreclaim_flag;
3174 if (*compact_result <= COMPACT_INACTIVE)
3178 * At least in one zone compaction wasn't deferred or skipped, so let's
3179 * count a compaction stall
3181 count_vm_event(COMPACTSTALL);
3183 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3186 struct zone *zone = page_zone(page);
3188 zone->compact_blockskip_flush = false;
3189 compaction_defer_reset(zone, order, true);
3190 count_vm_event(COMPACTSUCCESS);
3195 * It's bad if compaction run occurs and fails. The most likely reason
3196 * is that pages exist, but not enough to satisfy watermarks.
3198 count_vm_event(COMPACTFAIL);
3206 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3207 enum compact_result compact_result,
3208 enum compact_priority *compact_priority,
3209 int *compaction_retries)
3211 int max_retries = MAX_COMPACT_RETRIES;
3217 if (compaction_made_progress(compact_result))
3218 (*compaction_retries)++;
3221 * compaction considers all the zone as desperately out of memory
3222 * so it doesn't really make much sense to retry except when the
3223 * failure could be caused by insufficient priority
3225 if (compaction_failed(compact_result))
3226 goto check_priority;
3229 * make sure the compaction wasn't deferred or didn't bail out early
3230 * due to locks contention before we declare that we should give up.
3231 * But do not retry if the given zonelist is not suitable for
3234 if (compaction_withdrawn(compact_result))
3235 return compaction_zonelist_suitable(ac, order, alloc_flags);
3238 * !costly requests are much more important than __GFP_REPEAT
3239 * costly ones because they are de facto nofail and invoke OOM
3240 * killer to move on while costly can fail and users are ready
3241 * to cope with that. 1/4 retries is rather arbitrary but we
3242 * would need much more detailed feedback from compaction to
3243 * make a better decision.
3245 if (order > PAGE_ALLOC_COSTLY_ORDER)
3247 if (*compaction_retries <= max_retries)
3251 * Make sure there are attempts at the highest priority if we exhausted
3252 * all retries or failed at the lower priorities.
3255 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3256 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3257 if (*compact_priority > min_priority) {
3258 (*compact_priority)--;
3259 *compaction_retries = 0;
3265 static inline struct page *
3266 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3267 unsigned int alloc_flags, const struct alloc_context *ac,
3268 enum compact_priority prio, enum compact_result *compact_result)
3270 *compact_result = COMPACT_SKIPPED;
3275 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3276 enum compact_result compact_result,
3277 enum compact_priority *compact_priority,
3278 int *compaction_retries)
3283 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3287 * There are setups with compaction disabled which would prefer to loop
3288 * inside the allocator rather than hit the oom killer prematurely.
3289 * Let's give them a good hope and keep retrying while the order-0
3290 * watermarks are OK.
3292 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3294 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3295 ac_classzone_idx(ac), alloc_flags))
3300 #endif /* CONFIG_COMPACTION */
3302 /* Perform direct synchronous page reclaim */
3304 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3305 const struct alloc_context *ac)
3307 struct reclaim_state reclaim_state;
3312 /* We now go into synchronous reclaim */
3313 cpuset_memory_pressure_bump();
3314 current->flags |= PF_MEMALLOC;
3315 lockdep_set_current_reclaim_state(gfp_mask);
3316 reclaim_state.reclaimed_slab = 0;
3317 current->reclaim_state = &reclaim_state;
3319 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3322 current->reclaim_state = NULL;
3323 lockdep_clear_current_reclaim_state();
3324 current->flags &= ~PF_MEMALLOC;
3331 /* The really slow allocator path where we enter direct reclaim */
3332 static inline struct page *
3333 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3334 unsigned int alloc_flags, const struct alloc_context *ac,
3335 unsigned long *did_some_progress)
3337 struct page *page = NULL;
3338 bool drained = false;
3340 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3341 if (unlikely(!(*did_some_progress)))
3345 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3348 * If an allocation failed after direct reclaim, it could be because
3349 * pages are pinned on the per-cpu lists or in high alloc reserves.
3350 * Shrink them them and try again
3352 if (!page && !drained) {
3353 unreserve_highatomic_pageblock(ac);
3354 drain_all_pages(NULL);
3362 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3366 pg_data_t *last_pgdat = NULL;
3368 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3369 ac->high_zoneidx, ac->nodemask) {
3370 if (last_pgdat != zone->zone_pgdat)
3371 wakeup_kswapd(zone, order, ac->high_zoneidx);
3372 last_pgdat = zone->zone_pgdat;
3376 static inline unsigned int
3377 gfp_to_alloc_flags(gfp_t gfp_mask)
3379 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3381 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3382 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3385 * The caller may dip into page reserves a bit more if the caller
3386 * cannot run direct reclaim, or if the caller has realtime scheduling
3387 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3388 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3390 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3392 if (gfp_mask & __GFP_ATOMIC) {
3394 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3395 * if it can't schedule.
3397 if (!(gfp_mask & __GFP_NOMEMALLOC))
3398 alloc_flags |= ALLOC_HARDER;
3400 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3401 * comment for __cpuset_node_allowed().
3403 alloc_flags &= ~ALLOC_CPUSET;
3404 } else if (unlikely(rt_task(current)) && !in_interrupt())
3405 alloc_flags |= ALLOC_HARDER;
3408 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3409 alloc_flags |= ALLOC_CMA;
3414 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3416 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3419 if (gfp_mask & __GFP_MEMALLOC)
3421 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3423 if (!in_interrupt() &&
3424 ((current->flags & PF_MEMALLOC) ||
3425 unlikely(test_thread_flag(TIF_MEMDIE))))
3432 * Checks whether it makes sense to retry the reclaim to make a forward progress
3433 * for the given allocation request.
3434 * The reclaim feedback represented by did_some_progress (any progress during
3435 * the last reclaim round) and no_progress_loops (number of reclaim rounds without
3436 * any progress in a row) is considered as well as the reclaimable pages on the
3437 * applicable zone list (with a backoff mechanism which is a function of
3438 * no_progress_loops).
3440 * Returns true if a retry is viable or false to enter the oom path.
3443 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3444 struct alloc_context *ac, int alloc_flags,
3445 bool did_some_progress, int *no_progress_loops)
3451 * Costly allocations might have made a progress but this doesn't mean
3452 * their order will become available due to high fragmentation so
3453 * always increment the no progress counter for them
3455 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3456 *no_progress_loops = 0;
3458 (*no_progress_loops)++;
3461 * Make sure we converge to OOM if we cannot make any progress
3462 * several times in the row.
3464 if (*no_progress_loops > MAX_RECLAIM_RETRIES)
3468 * Keep reclaiming pages while there is a chance this will lead
3469 * somewhere. If none of the target zones can satisfy our allocation
3470 * request even if all reclaimable pages are considered then we are
3471 * screwed and have to go OOM.
3473 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3475 unsigned long available;
3476 unsigned long reclaimable;
3478 available = reclaimable = zone_reclaimable_pages(zone);
3479 available -= DIV_ROUND_UP((*no_progress_loops) * available,
3480 MAX_RECLAIM_RETRIES);
3481 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3484 * Would the allocation succeed if we reclaimed the whole
3487 if (__zone_watermark_ok(zone, order, min_wmark_pages(zone),
3488 ac_classzone_idx(ac), alloc_flags, available)) {
3490 * If we didn't make any progress and have a lot of
3491 * dirty + writeback pages then we should wait for
3492 * an IO to complete to slow down the reclaim and
3493 * prevent from pre mature OOM
3495 if (!did_some_progress) {
3496 unsigned long write_pending;
3498 write_pending = zone_page_state_snapshot(zone,
3499 NR_ZONE_WRITE_PENDING);
3501 if (2 * write_pending > reclaimable) {
3502 congestion_wait(BLK_RW_ASYNC, HZ/10);
3508 * Memory allocation/reclaim might be called from a WQ
3509 * context and the current implementation of the WQ
3510 * concurrency control doesn't recognize that
3511 * a particular WQ is congested if the worker thread is
3512 * looping without ever sleeping. Therefore we have to
3513 * do a short sleep here rather than calling
3516 if (current->flags & PF_WQ_WORKER)
3517 schedule_timeout_uninterruptible(1);
3528 static inline struct page *
3529 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3530 struct alloc_context *ac)
3532 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3533 struct page *page = NULL;
3534 unsigned int alloc_flags;
3535 unsigned long did_some_progress;
3536 enum compact_priority compact_priority;
3537 enum compact_result compact_result;
3538 int compaction_retries;
3539 int no_progress_loops;
3540 unsigned int cpuset_mems_cookie;
3543 * In the slowpath, we sanity check order to avoid ever trying to
3544 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3545 * be using allocators in order of preference for an area that is
3548 if (order >= MAX_ORDER) {
3549 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3554 * We also sanity check to catch abuse of atomic reserves being used by
3555 * callers that are not in atomic context.
3557 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3558 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3559 gfp_mask &= ~__GFP_ATOMIC;
3562 compaction_retries = 0;
3563 no_progress_loops = 0;
3564 compact_priority = DEF_COMPACT_PRIORITY;
3565 cpuset_mems_cookie = read_mems_allowed_begin();
3567 * We need to recalculate the starting point for the zonelist iterator
3568 * because we might have used different nodemask in the fast path, or
3569 * there was a cpuset modification and we are retrying - otherwise we
3570 * could end up iterating over non-eligible zones endlessly.
3572 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3573 ac->high_zoneidx, ac->nodemask);
3574 if (!ac->preferred_zoneref->zone)
3579 * The fast path uses conservative alloc_flags to succeed only until
3580 * kswapd needs to be woken up, and to avoid the cost of setting up
3581 * alloc_flags precisely. So we do that now.
3583 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3585 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3586 wake_all_kswapds(order, ac);
3589 * The adjusted alloc_flags might result in immediate success, so try
3592 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3597 * For costly allocations, try direct compaction first, as it's likely
3598 * that we have enough base pages and don't need to reclaim. Don't try
3599 * that for allocations that are allowed to ignore watermarks, as the
3600 * ALLOC_NO_WATERMARKS attempt didn't yet happen.
3602 if (can_direct_reclaim && order > PAGE_ALLOC_COSTLY_ORDER &&
3603 !gfp_pfmemalloc_allowed(gfp_mask)) {
3604 page = __alloc_pages_direct_compact(gfp_mask, order,
3606 INIT_COMPACT_PRIORITY,
3612 * Checks for costly allocations with __GFP_NORETRY, which
3613 * includes THP page fault allocations
3615 if (gfp_mask & __GFP_NORETRY) {
3617 * If compaction is deferred for high-order allocations,
3618 * it is because sync compaction recently failed. If
3619 * this is the case and the caller requested a THP
3620 * allocation, we do not want to heavily disrupt the
3621 * system, so we fail the allocation instead of entering
3624 if (compact_result == COMPACT_DEFERRED)
3628 * Looks like reclaim/compaction is worth trying, but
3629 * sync compaction could be very expensive, so keep
3630 * using async compaction.
3632 compact_priority = INIT_COMPACT_PRIORITY;
3637 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
3638 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3639 wake_all_kswapds(order, ac);
3641 if (gfp_pfmemalloc_allowed(gfp_mask))
3642 alloc_flags = ALLOC_NO_WATERMARKS;
3645 * Reset the zonelist iterators if memory policies can be ignored.
3646 * These allocations are high priority and system rather than user
3649 if (!(alloc_flags & ALLOC_CPUSET) || (alloc_flags & ALLOC_NO_WATERMARKS)) {
3650 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3651 ac->high_zoneidx, ac->nodemask);
3654 /* Attempt with potentially adjusted zonelist and alloc_flags */
3655 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3659 /* Caller is not willing to reclaim, we can't balance anything */
3660 if (!can_direct_reclaim) {
3662 * All existing users of the __GFP_NOFAIL are blockable, so warn
3663 * of any new users that actually allow this type of allocation
3666 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3670 /* Avoid recursion of direct reclaim */
3671 if (current->flags & PF_MEMALLOC) {
3673 * __GFP_NOFAIL request from this context is rather bizarre
3674 * because we cannot reclaim anything and only can loop waiting
3675 * for somebody to do a work for us.
3677 if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3684 /* Avoid allocations with no watermarks from looping endlessly */
3685 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3689 /* Try direct reclaim and then allocating */
3690 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3691 &did_some_progress);
3695 /* Try direct compaction and then allocating */
3696 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3697 compact_priority, &compact_result);
3701 /* Do not loop if specifically requested */
3702 if (gfp_mask & __GFP_NORETRY)
3706 * Do not retry costly high order allocations unless they are
3709 if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_REPEAT))
3712 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
3713 did_some_progress > 0, &no_progress_loops))
3717 * It doesn't make any sense to retry for the compaction if the order-0
3718 * reclaim is not able to make any progress because the current
3719 * implementation of the compaction depends on the sufficient amount
3720 * of free memory (see __compaction_suitable)
3722 if (did_some_progress > 0 &&
3723 should_compact_retry(ac, order, alloc_flags,
3724 compact_result, &compact_priority,
3725 &compaction_retries))
3729 * It's possible we raced with cpuset update so the OOM would be
3730 * premature (see below the nopage: label for full explanation).
3732 if (read_mems_allowed_retry(cpuset_mems_cookie))
3735 /* Reclaim has failed us, start killing things */
3736 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3740 /* Retry as long as the OOM killer is making progress */
3741 if (did_some_progress) {
3742 no_progress_loops = 0;
3748 * When updating a task's mems_allowed or mempolicy nodemask, it is
3749 * possible to race with parallel threads in such a way that our
3750 * allocation can fail while the mask is being updated. If we are about
3751 * to fail, check if the cpuset changed during allocation and if so,
3754 if (read_mems_allowed_retry(cpuset_mems_cookie))
3757 warn_alloc(gfp_mask,
3758 "page allocation failure: order:%u", order);
3764 * This is the 'heart' of the zoned buddy allocator.
3767 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3768 struct zonelist *zonelist, nodemask_t *nodemask)
3771 unsigned int alloc_flags = ALLOC_WMARK_LOW;
3772 gfp_t alloc_mask = gfp_mask; /* The gfp_t that was actually used for allocation */
3773 struct alloc_context ac = {
3774 .high_zoneidx = gfp_zone(gfp_mask),
3775 .zonelist = zonelist,
3776 .nodemask = nodemask,
3777 .migratetype = gfpflags_to_migratetype(gfp_mask),
3780 if (cpusets_enabled()) {
3781 alloc_mask |= __GFP_HARDWALL;
3782 alloc_flags |= ALLOC_CPUSET;
3784 ac.nodemask = &cpuset_current_mems_allowed;
3787 gfp_mask &= gfp_allowed_mask;
3789 lockdep_trace_alloc(gfp_mask);
3791 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3793 if (should_fail_alloc_page(gfp_mask, order))
3797 * Check the zones suitable for the gfp_mask contain at least one
3798 * valid zone. It's possible to have an empty zonelist as a result
3799 * of __GFP_THISNODE and a memoryless node
3801 if (unlikely(!zonelist->_zonerefs->zone))
3804 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3805 alloc_flags |= ALLOC_CMA;
3807 /* Dirty zone balancing only done in the fast path */
3808 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3811 * The preferred zone is used for statistics but crucially it is
3812 * also used as the starting point for the zonelist iterator. It
3813 * may get reset for allocations that ignore memory policies.
3815 ac.preferred_zoneref = first_zones_zonelist(ac.zonelist,
3816 ac.high_zoneidx, ac.nodemask);
3817 if (!ac.preferred_zoneref->zone) {
3820 * This might be due to race with cpuset_current_mems_allowed
3821 * update, so make sure we retry with original nodemask in the
3827 /* First allocation attempt */
3828 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3834 * Runtime PM, block IO and its error handling path can deadlock
3835 * because I/O on the device might not complete.
3837 alloc_mask = memalloc_noio_flags(gfp_mask);
3838 ac.spread_dirty_pages = false;
3841 * Restore the original nodemask if it was potentially replaced with
3842 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
3844 if (unlikely(ac.nodemask != nodemask))
3845 ac.nodemask = nodemask;
3847 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3850 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
3851 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
3852 __free_pages(page, order);
3856 if (kmemcheck_enabled && page)
3857 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3859 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3863 EXPORT_SYMBOL(__alloc_pages_nodemask);
3866 * Common helper functions.
3868 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3873 * __get_free_pages() returns a 32-bit address, which cannot represent
3876 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3878 page = alloc_pages(gfp_mask, order);
3881 return (unsigned long) page_address(page);
3883 EXPORT_SYMBOL(__get_free_pages);
3885 unsigned long get_zeroed_page(gfp_t gfp_mask)
3887 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3889 EXPORT_SYMBOL(get_zeroed_page);
3891 void __free_pages(struct page *page, unsigned int order)
3893 if (put_page_testzero(page)) {
3895 free_hot_cold_page(page, false);
3897 __free_pages_ok(page, order);
3901 EXPORT_SYMBOL(__free_pages);
3903 void free_pages(unsigned long addr, unsigned int order)
3906 VM_BUG_ON(!virt_addr_valid((void *)addr));
3907 __free_pages(virt_to_page((void *)addr), order);
3911 EXPORT_SYMBOL(free_pages);
3915 * An arbitrary-length arbitrary-offset area of memory which resides
3916 * within a 0 or higher order page. Multiple fragments within that page
3917 * are individually refcounted, in the page's reference counter.
3919 * The page_frag functions below provide a simple allocation framework for
3920 * page fragments. This is used by the network stack and network device
3921 * drivers to provide a backing region of memory for use as either an
3922 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3924 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3927 struct page *page = NULL;
3928 gfp_t gfp = gfp_mask;
3930 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3931 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3933 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3934 PAGE_FRAG_CACHE_MAX_ORDER);
3935 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3937 if (unlikely(!page))
3938 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3940 nc->va = page ? page_address(page) : NULL;
3945 void *__alloc_page_frag(struct page_frag_cache *nc,
3946 unsigned int fragsz, gfp_t gfp_mask)
3948 unsigned int size = PAGE_SIZE;
3952 if (unlikely(!nc->va)) {
3954 page = __page_frag_refill(nc, gfp_mask);
3958 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3959 /* if size can vary use size else just use PAGE_SIZE */
3962 /* Even if we own the page, we do not use atomic_set().
3963 * This would break get_page_unless_zero() users.
3965 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
3967 /* reset page count bias and offset to start of new frag */
3968 nc->pfmemalloc = page_is_pfmemalloc(page);
3969 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
3973 offset = nc->offset - fragsz;
3974 if (unlikely(offset < 0)) {
3975 page = virt_to_page(nc->va);
3977 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
3980 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3981 /* if size can vary use size else just use PAGE_SIZE */
3984 /* OK, page count is 0, we can safely set it */
3985 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
3987 /* reset page count bias and offset to start of new frag */
3988 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
3989 offset = size - fragsz;
3993 nc->offset = offset;
3995 return nc->va + offset;
3997 EXPORT_SYMBOL(__alloc_page_frag);
4000 * Frees a page fragment allocated out of either a compound or order 0 page.
4002 void __free_page_frag(void *addr)
4004 struct page *page = virt_to_head_page(addr);
4006 if (unlikely(put_page_testzero(page)))
4007 __free_pages_ok(page, compound_order(page));
4009 EXPORT_SYMBOL(__free_page_frag);
4011 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4015 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4016 unsigned long used = addr + PAGE_ALIGN(size);
4018 split_page(virt_to_page((void *)addr), order);
4019 while (used < alloc_end) {
4024 return (void *)addr;
4028 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4029 * @size: the number of bytes to allocate
4030 * @gfp_mask: GFP flags for the allocation
4032 * This function is similar to alloc_pages(), except that it allocates the
4033 * minimum number of pages to satisfy the request. alloc_pages() can only
4034 * allocate memory in power-of-two pages.
4036 * This function is also limited by MAX_ORDER.
4038 * Memory allocated by this function must be released by free_pages_exact().
4040 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4042 unsigned int order = get_order(size);
4045 addr = __get_free_pages(gfp_mask, order);
4046 return make_alloc_exact(addr, order, size);
4048 EXPORT_SYMBOL(alloc_pages_exact);
4051 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4053 * @nid: the preferred node ID where memory should be allocated
4054 * @size: the number of bytes to allocate
4055 * @gfp_mask: GFP flags for the allocation
4057 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4060 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4062 unsigned int order = get_order(size);
4063 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4066 return make_alloc_exact((unsigned long)page_address(p), order, size);
4070 * free_pages_exact - release memory allocated via alloc_pages_exact()
4071 * @virt: the value returned by alloc_pages_exact.
4072 * @size: size of allocation, same value as passed to alloc_pages_exact().
4074 * Release the memory allocated by a previous call to alloc_pages_exact.
4076 void free_pages_exact(void *virt, size_t size)
4078 unsigned long addr = (unsigned long)virt;
4079 unsigned long end = addr + PAGE_ALIGN(size);
4081 while (addr < end) {
4086 EXPORT_SYMBOL(free_pages_exact);
4089 * nr_free_zone_pages - count number of pages beyond high watermark
4090 * @offset: The zone index of the highest zone
4092 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4093 * high watermark within all zones at or below a given zone index. For each
4094 * zone, the number of pages is calculated as:
4095 * managed_pages - high_pages
4097 static unsigned long nr_free_zone_pages(int offset)
4102 /* Just pick one node, since fallback list is circular */
4103 unsigned long sum = 0;
4105 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4107 for_each_zone_zonelist(zone, z, zonelist, offset) {
4108 unsigned long size = zone->managed_pages;
4109 unsigned long high = high_wmark_pages(zone);
4118 * nr_free_buffer_pages - count number of pages beyond high watermark
4120 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4121 * watermark within ZONE_DMA and ZONE_NORMAL.
4123 unsigned long nr_free_buffer_pages(void)
4125 return nr_free_zone_pages(gfp_zone(GFP_USER));
4127 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4130 * nr_free_pagecache_pages - count number of pages beyond high watermark
4132 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4133 * high watermark within all zones.
4135 unsigned long nr_free_pagecache_pages(void)
4137 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4140 static inline void show_node(struct zone *zone)
4142 if (IS_ENABLED(CONFIG_NUMA))
4143 printk("Node %d ", zone_to_nid(zone));
4146 long si_mem_available(void)
4149 unsigned long pagecache;
4150 unsigned long wmark_low = 0;
4151 unsigned long pages[NR_LRU_LISTS];
4155 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4156 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4159 wmark_low += zone->watermark[WMARK_LOW];
4162 * Estimate the amount of memory available for userspace allocations,
4163 * without causing swapping.
4165 available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
4168 * Not all the page cache can be freed, otherwise the system will
4169 * start swapping. Assume at least half of the page cache, or the
4170 * low watermark worth of cache, needs to stay.
4172 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4173 pagecache -= min(pagecache / 2, wmark_low);
4174 available += pagecache;
4177 * Part of the reclaimable slab consists of items that are in use,
4178 * and cannot be freed. Cap this estimate at the low watermark.
4180 available += global_page_state(NR_SLAB_RECLAIMABLE) -
4181 min(global_page_state(NR_SLAB_RECLAIMABLE) / 2, wmark_low);
4187 EXPORT_SYMBOL_GPL(si_mem_available);
4189 void si_meminfo(struct sysinfo *val)
4191 val->totalram = totalram_pages;
4192 val->sharedram = global_node_page_state(NR_SHMEM);
4193 val->freeram = global_page_state(NR_FREE_PAGES);
4194 val->bufferram = nr_blockdev_pages();
4195 val->totalhigh = totalhigh_pages;
4196 val->freehigh = nr_free_highpages();
4197 val->mem_unit = PAGE_SIZE;
4200 EXPORT_SYMBOL(si_meminfo);
4203 void si_meminfo_node(struct sysinfo *val, int nid)
4205 int zone_type; /* needs to be signed */
4206 unsigned long managed_pages = 0;
4207 unsigned long managed_highpages = 0;
4208 unsigned long free_highpages = 0;
4209 pg_data_t *pgdat = NODE_DATA(nid);
4211 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4212 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4213 val->totalram = managed_pages;
4214 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4215 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4216 #ifdef CONFIG_HIGHMEM
4217 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4218 struct zone *zone = &pgdat->node_zones[zone_type];
4220 if (is_highmem(zone)) {
4221 managed_highpages += zone->managed_pages;
4222 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4225 val->totalhigh = managed_highpages;
4226 val->freehigh = free_highpages;
4228 val->totalhigh = managed_highpages;
4229 val->freehigh = free_highpages;
4231 val->mem_unit = PAGE_SIZE;
4236 * Determine whether the node should be displayed or not, depending on whether
4237 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4239 bool skip_free_areas_node(unsigned int flags, int nid)
4242 unsigned int cpuset_mems_cookie;
4244 if (!(flags & SHOW_MEM_FILTER_NODES))
4248 cpuset_mems_cookie = read_mems_allowed_begin();
4249 ret = !node_isset(nid, cpuset_current_mems_allowed);
4250 } while (read_mems_allowed_retry(cpuset_mems_cookie));
4255 #define K(x) ((x) << (PAGE_SHIFT-10))
4257 static void show_migration_types(unsigned char type)
4259 static const char types[MIGRATE_TYPES] = {
4260 [MIGRATE_UNMOVABLE] = 'U',
4261 [MIGRATE_MOVABLE] = 'M',
4262 [MIGRATE_RECLAIMABLE] = 'E',
4263 [MIGRATE_HIGHATOMIC] = 'H',
4265 [MIGRATE_CMA] = 'C',
4267 #ifdef CONFIG_MEMORY_ISOLATION
4268 [MIGRATE_ISOLATE] = 'I',
4271 char tmp[MIGRATE_TYPES + 1];
4275 for (i = 0; i < MIGRATE_TYPES; i++) {
4276 if (type & (1 << i))
4281 printk(KERN_CONT "(%s) ", tmp);
4285 * Show free area list (used inside shift_scroll-lock stuff)
4286 * We also calculate the percentage fragmentation. We do this by counting the
4287 * memory on each free list with the exception of the first item on the list.
4290 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4293 void show_free_areas(unsigned int filter)
4295 unsigned long free_pcp = 0;
4300 for_each_populated_zone(zone) {
4301 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4304 for_each_online_cpu(cpu)
4305 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4308 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4309 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4310 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4311 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4312 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4313 " free:%lu free_pcp:%lu free_cma:%lu\n",
4314 global_node_page_state(NR_ACTIVE_ANON),
4315 global_node_page_state(NR_INACTIVE_ANON),
4316 global_node_page_state(NR_ISOLATED_ANON),
4317 global_node_page_state(NR_ACTIVE_FILE),
4318 global_node_page_state(NR_INACTIVE_FILE),
4319 global_node_page_state(NR_ISOLATED_FILE),
4320 global_node_page_state(NR_UNEVICTABLE),
4321 global_node_page_state(NR_FILE_DIRTY),
4322 global_node_page_state(NR_WRITEBACK),
4323 global_node_page_state(NR_UNSTABLE_NFS),
4324 global_page_state(NR_SLAB_RECLAIMABLE),
4325 global_page_state(NR_SLAB_UNRECLAIMABLE),
4326 global_node_page_state(NR_FILE_MAPPED),
4327 global_node_page_state(NR_SHMEM),
4328 global_page_state(NR_PAGETABLE),
4329 global_page_state(NR_BOUNCE),
4330 global_page_state(NR_FREE_PAGES),
4332 global_page_state(NR_FREE_CMA_PAGES));
4334 for_each_online_pgdat(pgdat) {
4336 " active_anon:%lukB"
4337 " inactive_anon:%lukB"
4338 " active_file:%lukB"
4339 " inactive_file:%lukB"
4340 " unevictable:%lukB"
4341 " isolated(anon):%lukB"
4342 " isolated(file):%lukB"
4347 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4349 " shmem_pmdmapped: %lukB"
4352 " writeback_tmp:%lukB"
4354 " pages_scanned:%lu"
4355 " all_unreclaimable? %s"
4358 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4359 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4360 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4361 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4362 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4363 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4364 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4365 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4366 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4367 K(node_page_state(pgdat, NR_WRITEBACK)),
4368 K(node_page_state(pgdat, NR_SHMEM)),
4369 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4370 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4371 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4373 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4375 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4376 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4377 node_page_state(pgdat, NR_PAGES_SCANNED),
4378 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
4382 for_each_populated_zone(zone) {
4385 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4389 for_each_online_cpu(cpu)
4390 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4399 " active_anon:%lukB"
4400 " inactive_anon:%lukB"
4401 " active_file:%lukB"
4402 " inactive_file:%lukB"
4403 " unevictable:%lukB"
4404 " writepending:%lukB"
4408 " slab_reclaimable:%lukB"
4409 " slab_unreclaimable:%lukB"
4410 " kernel_stack:%lukB"
4418 K(zone_page_state(zone, NR_FREE_PAGES)),
4419 K(min_wmark_pages(zone)),
4420 K(low_wmark_pages(zone)),
4421 K(high_wmark_pages(zone)),
4422 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4423 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4424 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4425 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4426 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4427 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4428 K(zone->present_pages),
4429 K(zone->managed_pages),
4430 K(zone_page_state(zone, NR_MLOCK)),
4431 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
4432 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
4433 zone_page_state(zone, NR_KERNEL_STACK_KB),
4434 K(zone_page_state(zone, NR_PAGETABLE)),
4435 K(zone_page_state(zone, NR_BOUNCE)),
4437 K(this_cpu_read(zone->pageset->pcp.count)),
4438 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4439 printk("lowmem_reserve[]:");
4440 for (i = 0; i < MAX_NR_ZONES; i++)
4441 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
4442 printk(KERN_CONT "\n");
4445 for_each_populated_zone(zone) {
4447 unsigned long nr[MAX_ORDER], flags, total = 0;
4448 unsigned char types[MAX_ORDER];
4450 if (skip_free_areas_node(filter, zone_to_nid(zone)))
4453 printk(KERN_CONT "%s: ", zone->name);
4455 spin_lock_irqsave(&zone->lock, flags);
4456 for (order = 0; order < MAX_ORDER; order++) {
4457 struct free_area *area = &zone->free_area[order];
4460 nr[order] = area->nr_free;
4461 total += nr[order] << order;
4464 for (type = 0; type < MIGRATE_TYPES; type++) {
4465 if (!list_empty(&area->free_list[type]))
4466 types[order] |= 1 << type;
4469 spin_unlock_irqrestore(&zone->lock, flags);
4470 for (order = 0; order < MAX_ORDER; order++) {
4471 printk(KERN_CONT "%lu*%lukB ",
4472 nr[order], K(1UL) << order);
4474 show_migration_types(types[order]);
4476 printk(KERN_CONT "= %lukB\n", K(total));
4479 hugetlb_show_meminfo();
4481 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
4483 show_swap_cache_info();
4486 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4488 zoneref->zone = zone;
4489 zoneref->zone_idx = zone_idx(zone);
4493 * Builds allocation fallback zone lists.
4495 * Add all populated zones of a node to the zonelist.
4497 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
4501 enum zone_type zone_type = MAX_NR_ZONES;
4505 zone = pgdat->node_zones + zone_type;
4506 if (populated_zone(zone)) {
4507 zoneref_set_zone(zone,
4508 &zonelist->_zonerefs[nr_zones++]);
4509 check_highest_zone(zone_type);
4511 } while (zone_type);
4519 * 0 = automatic detection of better ordering.
4520 * 1 = order by ([node] distance, -zonetype)
4521 * 2 = order by (-zonetype, [node] distance)
4523 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4524 * the same zonelist. So only NUMA can configure this param.
4526 #define ZONELIST_ORDER_DEFAULT 0
4527 #define ZONELIST_ORDER_NODE 1
4528 #define ZONELIST_ORDER_ZONE 2
4530 /* zonelist order in the kernel.
4531 * set_zonelist_order() will set this to NODE or ZONE.
4533 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4534 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4538 /* The value user specified ....changed by config */
4539 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4540 /* string for sysctl */
4541 #define NUMA_ZONELIST_ORDER_LEN 16
4542 char numa_zonelist_order[16] = "default";
4545 * interface for configure zonelist ordering.
4546 * command line option "numa_zonelist_order"
4547 * = "[dD]efault - default, automatic configuration.
4548 * = "[nN]ode - order by node locality, then by zone within node
4549 * = "[zZ]one - order by zone, then by locality within zone
4552 static int __parse_numa_zonelist_order(char *s)
4554 if (*s == 'd' || *s == 'D') {
4555 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4556 } else if (*s == 'n' || *s == 'N') {
4557 user_zonelist_order = ZONELIST_ORDER_NODE;
4558 } else if (*s == 'z' || *s == 'Z') {
4559 user_zonelist_order = ZONELIST_ORDER_ZONE;
4561 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s);
4567 static __init int setup_numa_zonelist_order(char *s)
4574 ret = __parse_numa_zonelist_order(s);
4576 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4580 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4583 * sysctl handler for numa_zonelist_order
4585 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4586 void __user *buffer, size_t *length,
4589 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4591 static DEFINE_MUTEX(zl_order_mutex);
4593 mutex_lock(&zl_order_mutex);
4595 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4599 strcpy(saved_string, (char *)table->data);
4601 ret = proc_dostring(table, write, buffer, length, ppos);
4605 int oldval = user_zonelist_order;
4607 ret = __parse_numa_zonelist_order((char *)table->data);
4610 * bogus value. restore saved string
4612 strncpy((char *)table->data, saved_string,
4613 NUMA_ZONELIST_ORDER_LEN);
4614 user_zonelist_order = oldval;
4615 } else if (oldval != user_zonelist_order) {
4616 mutex_lock(&zonelists_mutex);
4617 build_all_zonelists(NULL, NULL, false);
4618 mutex_unlock(&zonelists_mutex);
4622 mutex_unlock(&zl_order_mutex);
4627 #define MAX_NODE_LOAD (nr_online_nodes)
4628 static int node_load[MAX_NUMNODES];
4631 * find_next_best_node - find the next node that should appear in a given node's fallback list
4632 * @node: node whose fallback list we're appending
4633 * @used_node_mask: nodemask_t of already used nodes
4635 * We use a number of factors to determine which is the next node that should
4636 * appear on a given node's fallback list. The node should not have appeared
4637 * already in @node's fallback list, and it should be the next closest node
4638 * according to the distance array (which contains arbitrary distance values
4639 * from each node to each node in the system), and should also prefer nodes
4640 * with no CPUs, since presumably they'll have very little allocation pressure
4641 * on them otherwise.
4642 * It returns -1 if no node is found.
4644 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4647 int min_val = INT_MAX;
4648 int best_node = NUMA_NO_NODE;
4649 const struct cpumask *tmp = cpumask_of_node(0);
4651 /* Use the local node if we haven't already */
4652 if (!node_isset(node, *used_node_mask)) {
4653 node_set(node, *used_node_mask);
4657 for_each_node_state(n, N_MEMORY) {
4659 /* Don't want a node to appear more than once */
4660 if (node_isset(n, *used_node_mask))
4663 /* Use the distance array to find the distance */
4664 val = node_distance(node, n);
4666 /* Penalize nodes under us ("prefer the next node") */
4669 /* Give preference to headless and unused nodes */
4670 tmp = cpumask_of_node(n);
4671 if (!cpumask_empty(tmp))
4672 val += PENALTY_FOR_NODE_WITH_CPUS;
4674 /* Slight preference for less loaded node */
4675 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4676 val += node_load[n];
4678 if (val < min_val) {
4685 node_set(best_node, *used_node_mask);
4692 * Build zonelists ordered by node and zones within node.
4693 * This results in maximum locality--normal zone overflows into local
4694 * DMA zone, if any--but risks exhausting DMA zone.
4696 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4699 struct zonelist *zonelist;
4701 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4702 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4704 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4705 zonelist->_zonerefs[j].zone = NULL;
4706 zonelist->_zonerefs[j].zone_idx = 0;
4710 * Build gfp_thisnode zonelists
4712 static void build_thisnode_zonelists(pg_data_t *pgdat)
4715 struct zonelist *zonelist;
4717 zonelist = &pgdat->node_zonelists[ZONELIST_NOFALLBACK];
4718 j = build_zonelists_node(pgdat, zonelist, 0);
4719 zonelist->_zonerefs[j].zone = NULL;
4720 zonelist->_zonerefs[j].zone_idx = 0;
4724 * Build zonelists ordered by zone and nodes within zones.
4725 * This results in conserving DMA zone[s] until all Normal memory is
4726 * exhausted, but results in overflowing to remote node while memory
4727 * may still exist in local DMA zone.
4729 static int node_order[MAX_NUMNODES];
4731 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4734 int zone_type; /* needs to be signed */
4736 struct zonelist *zonelist;
4738 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4740 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4741 for (j = 0; j < nr_nodes; j++) {
4742 node = node_order[j];
4743 z = &NODE_DATA(node)->node_zones[zone_type];
4744 if (managed_zone(z)) {
4746 &zonelist->_zonerefs[pos++]);
4747 check_highest_zone(zone_type);
4751 zonelist->_zonerefs[pos].zone = NULL;
4752 zonelist->_zonerefs[pos].zone_idx = 0;
4755 #if defined(CONFIG_64BIT)
4757 * Devices that require DMA32/DMA are relatively rare and do not justify a
4758 * penalty to every machine in case the specialised case applies. Default
4759 * to Node-ordering on 64-bit NUMA machines
4761 static int default_zonelist_order(void)
4763 return ZONELIST_ORDER_NODE;
4767 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4768 * by the kernel. If processes running on node 0 deplete the low memory zone
4769 * then reclaim will occur more frequency increasing stalls and potentially
4770 * be easier to OOM if a large percentage of the zone is under writeback or
4771 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4772 * Hence, default to zone ordering on 32-bit.
4774 static int default_zonelist_order(void)
4776 return ZONELIST_ORDER_ZONE;
4778 #endif /* CONFIG_64BIT */
4780 static void set_zonelist_order(void)
4782 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4783 current_zonelist_order = default_zonelist_order();
4785 current_zonelist_order = user_zonelist_order;
4788 static void build_zonelists(pg_data_t *pgdat)
4791 nodemask_t used_mask;
4792 int local_node, prev_node;
4793 struct zonelist *zonelist;
4794 unsigned int order = current_zonelist_order;
4796 /* initialize zonelists */
4797 for (i = 0; i < MAX_ZONELISTS; i++) {
4798 zonelist = pgdat->node_zonelists + i;
4799 zonelist->_zonerefs[0].zone = NULL;
4800 zonelist->_zonerefs[0].zone_idx = 0;
4803 /* NUMA-aware ordering of nodes */
4804 local_node = pgdat->node_id;
4805 load = nr_online_nodes;
4806 prev_node = local_node;
4807 nodes_clear(used_mask);
4809 memset(node_order, 0, sizeof(node_order));
4812 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4814 * We don't want to pressure a particular node.
4815 * So adding penalty to the first node in same
4816 * distance group to make it round-robin.
4818 if (node_distance(local_node, node) !=
4819 node_distance(local_node, prev_node))
4820 node_load[node] = load;
4824 if (order == ZONELIST_ORDER_NODE)
4825 build_zonelists_in_node_order(pgdat, node);
4827 node_order[i++] = node; /* remember order */
4830 if (order == ZONELIST_ORDER_ZONE) {
4831 /* calculate node order -- i.e., DMA last! */
4832 build_zonelists_in_zone_order(pgdat, i);
4835 build_thisnode_zonelists(pgdat);
4838 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4840 * Return node id of node used for "local" allocations.
4841 * I.e., first node id of first zone in arg node's generic zonelist.
4842 * Used for initializing percpu 'numa_mem', which is used primarily
4843 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4845 int local_memory_node(int node)
4849 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4850 gfp_zone(GFP_KERNEL),
4852 return z->zone->node;
4856 static void setup_min_unmapped_ratio(void);
4857 static void setup_min_slab_ratio(void);
4858 #else /* CONFIG_NUMA */
4860 static void set_zonelist_order(void)
4862 current_zonelist_order = ZONELIST_ORDER_ZONE;
4865 static void build_zonelists(pg_data_t *pgdat)
4867 int node, local_node;
4869 struct zonelist *zonelist;
4871 local_node = pgdat->node_id;
4873 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4874 j = build_zonelists_node(pgdat, zonelist, 0);
4877 * Now we build the zonelist so that it contains the zones
4878 * of all the other nodes.
4879 * We don't want to pressure a particular node, so when
4880 * building the zones for node N, we make sure that the
4881 * zones coming right after the local ones are those from
4882 * node N+1 (modulo N)
4884 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4885 if (!node_online(node))
4887 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4889 for (node = 0; node < local_node; node++) {
4890 if (!node_online(node))
4892 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4895 zonelist->_zonerefs[j].zone = NULL;
4896 zonelist->_zonerefs[j].zone_idx = 0;
4899 #endif /* CONFIG_NUMA */
4902 * Boot pageset table. One per cpu which is going to be used for all
4903 * zones and all nodes. The parameters will be set in such a way
4904 * that an item put on a list will immediately be handed over to
4905 * the buddy list. This is safe since pageset manipulation is done
4906 * with interrupts disabled.
4908 * The boot_pagesets must be kept even after bootup is complete for
4909 * unused processors and/or zones. They do play a role for bootstrapping
4910 * hotplugged processors.
4912 * zoneinfo_show() and maybe other functions do
4913 * not check if the processor is online before following the pageset pointer.
4914 * Other parts of the kernel may not check if the zone is available.
4916 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4917 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4918 static void setup_zone_pageset(struct zone *zone);
4921 * Global mutex to protect against size modification of zonelists
4922 * as well as to serialize pageset setup for the new populated zone.
4924 DEFINE_MUTEX(zonelists_mutex);
4926 /* return values int ....just for stop_machine() */
4927 static int __build_all_zonelists(void *data)
4931 pg_data_t *self = data;
4934 memset(node_load, 0, sizeof(node_load));
4937 if (self && !node_online(self->node_id)) {
4938 build_zonelists(self);
4941 for_each_online_node(nid) {
4942 pg_data_t *pgdat = NODE_DATA(nid);
4944 build_zonelists(pgdat);
4948 * Initialize the boot_pagesets that are going to be used
4949 * for bootstrapping processors. The real pagesets for
4950 * each zone will be allocated later when the per cpu
4951 * allocator is available.
4953 * boot_pagesets are used also for bootstrapping offline
4954 * cpus if the system is already booted because the pagesets
4955 * are needed to initialize allocators on a specific cpu too.
4956 * F.e. the percpu allocator needs the page allocator which
4957 * needs the percpu allocator in order to allocate its pagesets
4958 * (a chicken-egg dilemma).
4960 for_each_possible_cpu(cpu) {
4961 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4963 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4965 * We now know the "local memory node" for each node--
4966 * i.e., the node of the first zone in the generic zonelist.
4967 * Set up numa_mem percpu variable for on-line cpus. During
4968 * boot, only the boot cpu should be on-line; we'll init the
4969 * secondary cpus' numa_mem as they come on-line. During
4970 * node/memory hotplug, we'll fixup all on-line cpus.
4972 if (cpu_online(cpu))
4973 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4980 static noinline void __init
4981 build_all_zonelists_init(void)
4983 __build_all_zonelists(NULL);
4984 mminit_verify_zonelist();
4985 cpuset_init_current_mems_allowed();
4989 * Called with zonelists_mutex held always
4990 * unless system_state == SYSTEM_BOOTING.
4992 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4993 * [we're only called with non-NULL zone through __meminit paths] and
4994 * (2) call of __init annotated helper build_all_zonelists_init
4995 * [protected by SYSTEM_BOOTING].
4997 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone,
4998 bool hotplug_context)
5000 set_zonelist_order();
5002 if (system_state == SYSTEM_BOOTING && !hotplug_context) {
5003 build_all_zonelists_init();
5005 #ifdef CONFIG_MEMORY_HOTPLUG
5007 setup_zone_pageset(zone);
5009 /* we have to stop all cpus to guarantee there is no user
5011 stop_machine(__build_all_zonelists, pgdat, NULL);
5012 /* cpuset refresh routine should be here */
5014 vm_total_pages = nr_free_pagecache_pages();
5016 * Disable grouping by mobility if the number of pages in the
5017 * system is too low to allow the mechanism to work. It would be
5018 * more accurate, but expensive to check per-zone. This check is
5019 * made on memory-hotadd so a system can start with mobility
5020 * disabled and enable it later
5022 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5023 page_group_by_mobility_disabled = 1;
5025 page_group_by_mobility_disabled = 0;
5027 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
5029 zonelist_order_name[current_zonelist_order],
5030 page_group_by_mobility_disabled ? "off" : "on",
5033 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5038 * Initially all pages are reserved - free ones are freed
5039 * up by free_all_bootmem() once the early boot process is
5040 * done. Non-atomic initialization, single-pass.
5042 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5043 unsigned long start_pfn, enum memmap_context context)
5045 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5046 unsigned long end_pfn = start_pfn + size;
5047 pg_data_t *pgdat = NODE_DATA(nid);
5049 unsigned long nr_initialised = 0;
5050 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5051 struct memblock_region *r = NULL, *tmp;
5054 if (highest_memmap_pfn < end_pfn - 1)
5055 highest_memmap_pfn = end_pfn - 1;
5058 * Honor reservation requested by the driver for this ZONE_DEVICE
5061 if (altmap && start_pfn == altmap->base_pfn)
5062 start_pfn += altmap->reserve;
5064 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5066 * There can be holes in boot-time mem_map[]s handed to this
5067 * function. They do not exist on hotplugged memory.
5069 if (context != MEMMAP_EARLY)
5072 if (!early_pfn_valid(pfn))
5074 if (!early_pfn_in_nid(pfn, nid))
5076 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5079 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5081 * Check given memblock attribute by firmware which can affect
5082 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5083 * mirrored, it's an overlapped memmap init. skip it.
5085 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5086 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5087 for_each_memblock(memory, tmp)
5088 if (pfn < memblock_region_memory_end_pfn(tmp))
5092 if (pfn >= memblock_region_memory_base_pfn(r) &&
5093 memblock_is_mirror(r)) {
5094 /* already initialized as NORMAL */
5095 pfn = memblock_region_memory_end_pfn(r);
5103 * Mark the block movable so that blocks are reserved for
5104 * movable at startup. This will force kernel allocations
5105 * to reserve their blocks rather than leaking throughout
5106 * the address space during boot when many long-lived
5107 * kernel allocations are made.
5109 * bitmap is created for zone's valid pfn range. but memmap
5110 * can be created for invalid pages (for alignment)
5111 * check here not to call set_pageblock_migratetype() against
5114 if (!(pfn & (pageblock_nr_pages - 1))) {
5115 struct page *page = pfn_to_page(pfn);
5117 __init_single_page(page, pfn, zone, nid);
5118 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5120 __init_single_pfn(pfn, zone, nid);
5125 static void __meminit zone_init_free_lists(struct zone *zone)
5127 unsigned int order, t;
5128 for_each_migratetype_order(order, t) {
5129 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5130 zone->free_area[order].nr_free = 0;
5134 #ifndef __HAVE_ARCH_MEMMAP_INIT
5135 #define memmap_init(size, nid, zone, start_pfn) \
5136 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5139 static int zone_batchsize(struct zone *zone)
5145 * The per-cpu-pages pools are set to around 1000th of the
5146 * size of the zone. But no more than 1/2 of a meg.
5148 * OK, so we don't know how big the cache is. So guess.
5150 batch = zone->managed_pages / 1024;
5151 if (batch * PAGE_SIZE > 512 * 1024)
5152 batch = (512 * 1024) / PAGE_SIZE;
5153 batch /= 4; /* We effectively *= 4 below */
5158 * Clamp the batch to a 2^n - 1 value. Having a power
5159 * of 2 value was found to be more likely to have
5160 * suboptimal cache aliasing properties in some cases.
5162 * For example if 2 tasks are alternately allocating
5163 * batches of pages, one task can end up with a lot
5164 * of pages of one half of the possible page colors
5165 * and the other with pages of the other colors.
5167 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5172 /* The deferral and batching of frees should be suppressed under NOMMU
5175 * The problem is that NOMMU needs to be able to allocate large chunks
5176 * of contiguous memory as there's no hardware page translation to
5177 * assemble apparent contiguous memory from discontiguous pages.
5179 * Queueing large contiguous runs of pages for batching, however,
5180 * causes the pages to actually be freed in smaller chunks. As there
5181 * can be a significant delay between the individual batches being
5182 * recycled, this leads to the once large chunks of space being
5183 * fragmented and becoming unavailable for high-order allocations.
5190 * pcp->high and pcp->batch values are related and dependent on one another:
5191 * ->batch must never be higher then ->high.
5192 * The following function updates them in a safe manner without read side
5195 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5196 * those fields changing asynchronously (acording the the above rule).
5198 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5199 * outside of boot time (or some other assurance that no concurrent updaters
5202 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5203 unsigned long batch)
5205 /* start with a fail safe value for batch */
5209 /* Update high, then batch, in order */
5216 /* a companion to pageset_set_high() */
5217 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5219 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5222 static void pageset_init(struct per_cpu_pageset *p)
5224 struct per_cpu_pages *pcp;
5227 memset(p, 0, sizeof(*p));
5231 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5232 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5235 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5238 pageset_set_batch(p, batch);
5242 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5243 * to the value high for the pageset p.
5245 static void pageset_set_high(struct per_cpu_pageset *p,
5248 unsigned long batch = max(1UL, high / 4);
5249 if ((high / 4) > (PAGE_SHIFT * 8))
5250 batch = PAGE_SHIFT * 8;
5252 pageset_update(&p->pcp, high, batch);
5255 static void pageset_set_high_and_batch(struct zone *zone,
5256 struct per_cpu_pageset *pcp)
5258 if (percpu_pagelist_fraction)
5259 pageset_set_high(pcp,
5260 (zone->managed_pages /
5261 percpu_pagelist_fraction));
5263 pageset_set_batch(pcp, zone_batchsize(zone));
5266 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5268 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5271 pageset_set_high_and_batch(zone, pcp);
5274 static void __meminit setup_zone_pageset(struct zone *zone)
5277 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5278 for_each_possible_cpu(cpu)
5279 zone_pageset_init(zone, cpu);
5283 * Allocate per cpu pagesets and initialize them.
5284 * Before this call only boot pagesets were available.
5286 void __init setup_per_cpu_pageset(void)
5288 struct pglist_data *pgdat;
5291 for_each_populated_zone(zone)
5292 setup_zone_pageset(zone);
5294 for_each_online_pgdat(pgdat)
5295 pgdat->per_cpu_nodestats =
5296 alloc_percpu(struct per_cpu_nodestat);
5299 static __meminit void zone_pcp_init(struct zone *zone)
5302 * per cpu subsystem is not up at this point. The following code
5303 * relies on the ability of the linker to provide the
5304 * offset of a (static) per cpu variable into the per cpu area.
5306 zone->pageset = &boot_pageset;
5308 if (populated_zone(zone))
5309 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5310 zone->name, zone->present_pages,
5311 zone_batchsize(zone));
5314 int __meminit init_currently_empty_zone(struct zone *zone,
5315 unsigned long zone_start_pfn,
5318 struct pglist_data *pgdat = zone->zone_pgdat;
5320 pgdat->nr_zones = zone_idx(zone) + 1;
5322 zone->zone_start_pfn = zone_start_pfn;
5324 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5325 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5327 (unsigned long)zone_idx(zone),
5328 zone_start_pfn, (zone_start_pfn + size));
5330 zone_init_free_lists(zone);
5331 zone->initialized = 1;
5336 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5337 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5340 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5342 int __meminit __early_pfn_to_nid(unsigned long pfn,
5343 struct mminit_pfnnid_cache *state)
5345 unsigned long start_pfn, end_pfn;
5348 if (state->last_start <= pfn && pfn < state->last_end)
5349 return state->last_nid;
5351 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5353 state->last_start = start_pfn;
5354 state->last_end = end_pfn;
5355 state->last_nid = nid;
5360 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5363 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5364 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5365 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5367 * If an architecture guarantees that all ranges registered contain no holes
5368 * and may be freed, this this function may be used instead of calling
5369 * memblock_free_early_nid() manually.
5371 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5373 unsigned long start_pfn, end_pfn;
5376 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5377 start_pfn = min(start_pfn, max_low_pfn);
5378 end_pfn = min(end_pfn, max_low_pfn);
5380 if (start_pfn < end_pfn)
5381 memblock_free_early_nid(PFN_PHYS(start_pfn),
5382 (end_pfn - start_pfn) << PAGE_SHIFT,
5388 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5389 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5391 * If an architecture guarantees that all ranges registered contain no holes and may
5392 * be freed, this function may be used instead of calling memory_present() manually.
5394 void __init sparse_memory_present_with_active_regions(int nid)
5396 unsigned long start_pfn, end_pfn;
5399 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5400 memory_present(this_nid, start_pfn, end_pfn);
5404 * get_pfn_range_for_nid - Return the start and end page frames for a node
5405 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5406 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5407 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5409 * It returns the start and end page frame of a node based on information
5410 * provided by memblock_set_node(). If called for a node
5411 * with no available memory, a warning is printed and the start and end
5414 void __meminit get_pfn_range_for_nid(unsigned int nid,
5415 unsigned long *start_pfn, unsigned long *end_pfn)
5417 unsigned long this_start_pfn, this_end_pfn;
5423 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5424 *start_pfn = min(*start_pfn, this_start_pfn);
5425 *end_pfn = max(*end_pfn, this_end_pfn);
5428 if (*start_pfn == -1UL)
5433 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5434 * assumption is made that zones within a node are ordered in monotonic
5435 * increasing memory addresses so that the "highest" populated zone is used
5437 static void __init find_usable_zone_for_movable(void)
5440 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5441 if (zone_index == ZONE_MOVABLE)
5444 if (arch_zone_highest_possible_pfn[zone_index] >
5445 arch_zone_lowest_possible_pfn[zone_index])
5449 VM_BUG_ON(zone_index == -1);
5450 movable_zone = zone_index;
5454 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5455 * because it is sized independent of architecture. Unlike the other zones,
5456 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5457 * in each node depending on the size of each node and how evenly kernelcore
5458 * is distributed. This helper function adjusts the zone ranges
5459 * provided by the architecture for a given node by using the end of the
5460 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5461 * zones within a node are in order of monotonic increases memory addresses
5463 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5464 unsigned long zone_type,
5465 unsigned long node_start_pfn,
5466 unsigned long node_end_pfn,
5467 unsigned long *zone_start_pfn,
5468 unsigned long *zone_end_pfn)
5470 /* Only adjust if ZONE_MOVABLE is on this node */
5471 if (zone_movable_pfn[nid]) {
5472 /* Size ZONE_MOVABLE */
5473 if (zone_type == ZONE_MOVABLE) {
5474 *zone_start_pfn = zone_movable_pfn[nid];
5475 *zone_end_pfn = min(node_end_pfn,
5476 arch_zone_highest_possible_pfn[movable_zone]);
5478 /* Adjust for ZONE_MOVABLE starting within this range */
5479 } else if (!mirrored_kernelcore &&
5480 *zone_start_pfn < zone_movable_pfn[nid] &&
5481 *zone_end_pfn > zone_movable_pfn[nid]) {
5482 *zone_end_pfn = zone_movable_pfn[nid];
5484 /* Check if this whole range is within ZONE_MOVABLE */
5485 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5486 *zone_start_pfn = *zone_end_pfn;
5491 * Return the number of pages a zone spans in a node, including holes
5492 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5494 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5495 unsigned long zone_type,
5496 unsigned long node_start_pfn,
5497 unsigned long node_end_pfn,
5498 unsigned long *zone_start_pfn,
5499 unsigned long *zone_end_pfn,
5500 unsigned long *ignored)
5502 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5503 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5504 /* When hotadd a new node from cpu_up(), the node should be empty */
5505 if (!node_start_pfn && !node_end_pfn)
5508 /* Get the start and end of the zone */
5509 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5510 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5511 adjust_zone_range_for_zone_movable(nid, zone_type,
5512 node_start_pfn, node_end_pfn,
5513 zone_start_pfn, zone_end_pfn);
5515 /* Check that this node has pages within the zone's required range */
5516 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5519 /* Move the zone boundaries inside the node if necessary */
5520 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5521 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5523 /* Return the spanned pages */
5524 return *zone_end_pfn - *zone_start_pfn;
5528 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5529 * then all holes in the requested range will be accounted for.
5531 unsigned long __meminit __absent_pages_in_range(int nid,
5532 unsigned long range_start_pfn,
5533 unsigned long range_end_pfn)
5535 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5536 unsigned long start_pfn, end_pfn;
5539 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5540 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5541 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5542 nr_absent -= end_pfn - start_pfn;
5548 * absent_pages_in_range - Return number of page frames in holes within a range
5549 * @start_pfn: The start PFN to start searching for holes
5550 * @end_pfn: The end PFN to stop searching for holes
5552 * It returns the number of pages frames in memory holes within a range.
5554 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5555 unsigned long end_pfn)
5557 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5560 /* Return the number of page frames in holes in a zone on a node */
5561 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5562 unsigned long zone_type,
5563 unsigned long node_start_pfn,
5564 unsigned long node_end_pfn,
5565 unsigned long *ignored)
5567 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5568 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5569 unsigned long zone_start_pfn, zone_end_pfn;
5570 unsigned long nr_absent;
5572 /* When hotadd a new node from cpu_up(), the node should be empty */
5573 if (!node_start_pfn && !node_end_pfn)
5576 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5577 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5579 adjust_zone_range_for_zone_movable(nid, zone_type,
5580 node_start_pfn, node_end_pfn,
5581 &zone_start_pfn, &zone_end_pfn);
5582 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5585 * ZONE_MOVABLE handling.
5586 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5589 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
5590 unsigned long start_pfn, end_pfn;
5591 struct memblock_region *r;
5593 for_each_memblock(memory, r) {
5594 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5595 zone_start_pfn, zone_end_pfn);
5596 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5597 zone_start_pfn, zone_end_pfn);
5599 if (zone_type == ZONE_MOVABLE &&
5600 memblock_is_mirror(r))
5601 nr_absent += end_pfn - start_pfn;
5603 if (zone_type == ZONE_NORMAL &&
5604 !memblock_is_mirror(r))
5605 nr_absent += end_pfn - start_pfn;
5612 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5613 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5614 unsigned long zone_type,
5615 unsigned long node_start_pfn,
5616 unsigned long node_end_pfn,
5617 unsigned long *zone_start_pfn,
5618 unsigned long *zone_end_pfn,
5619 unsigned long *zones_size)
5623 *zone_start_pfn = node_start_pfn;
5624 for (zone = 0; zone < zone_type; zone++)
5625 *zone_start_pfn += zones_size[zone];
5627 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5629 return zones_size[zone_type];
5632 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5633 unsigned long zone_type,
5634 unsigned long node_start_pfn,
5635 unsigned long node_end_pfn,
5636 unsigned long *zholes_size)
5641 return zholes_size[zone_type];
5644 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5646 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5647 unsigned long node_start_pfn,
5648 unsigned long node_end_pfn,
5649 unsigned long *zones_size,
5650 unsigned long *zholes_size)
5652 unsigned long realtotalpages = 0, totalpages = 0;
5655 for (i = 0; i < MAX_NR_ZONES; i++) {
5656 struct zone *zone = pgdat->node_zones + i;
5657 unsigned long zone_start_pfn, zone_end_pfn;
5658 unsigned long size, real_size;
5660 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5666 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5667 node_start_pfn, node_end_pfn,
5670 zone->zone_start_pfn = zone_start_pfn;
5672 zone->zone_start_pfn = 0;
5673 zone->spanned_pages = size;
5674 zone->present_pages = real_size;
5677 realtotalpages += real_size;
5680 pgdat->node_spanned_pages = totalpages;
5681 pgdat->node_present_pages = realtotalpages;
5682 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5686 #ifndef CONFIG_SPARSEMEM
5688 * Calculate the size of the zone->blockflags rounded to an unsigned long
5689 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5690 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5691 * round what is now in bits to nearest long in bits, then return it in
5694 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5696 unsigned long usemapsize;
5698 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5699 usemapsize = roundup(zonesize, pageblock_nr_pages);
5700 usemapsize = usemapsize >> pageblock_order;
5701 usemapsize *= NR_PAGEBLOCK_BITS;
5702 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5704 return usemapsize / 8;
5707 static void __init setup_usemap(struct pglist_data *pgdat,
5709 unsigned long zone_start_pfn,
5710 unsigned long zonesize)
5712 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5713 zone->pageblock_flags = NULL;
5715 zone->pageblock_flags =
5716 memblock_virt_alloc_node_nopanic(usemapsize,
5720 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5721 unsigned long zone_start_pfn, unsigned long zonesize) {}
5722 #endif /* CONFIG_SPARSEMEM */
5724 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5726 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5727 void __paginginit set_pageblock_order(void)
5731 /* Check that pageblock_nr_pages has not already been setup */
5732 if (pageblock_order)
5735 if (HPAGE_SHIFT > PAGE_SHIFT)
5736 order = HUGETLB_PAGE_ORDER;
5738 order = MAX_ORDER - 1;
5741 * Assume the largest contiguous order of interest is a huge page.
5742 * This value may be variable depending on boot parameters on IA64 and
5745 pageblock_order = order;
5747 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5750 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5751 * is unused as pageblock_order is set at compile-time. See
5752 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5755 void __paginginit set_pageblock_order(void)
5759 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5761 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5762 unsigned long present_pages)
5764 unsigned long pages = spanned_pages;
5767 * Provide a more accurate estimation if there are holes within
5768 * the zone and SPARSEMEM is in use. If there are holes within the
5769 * zone, each populated memory region may cost us one or two extra
5770 * memmap pages due to alignment because memmap pages for each
5771 * populated regions may not naturally algined on page boundary.
5772 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5774 if (spanned_pages > present_pages + (present_pages >> 4) &&
5775 IS_ENABLED(CONFIG_SPARSEMEM))
5776 pages = present_pages;
5778 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5782 * Set up the zone data structures:
5783 * - mark all pages reserved
5784 * - mark all memory queues empty
5785 * - clear the memory bitmaps
5787 * NOTE: pgdat should get zeroed by caller.
5789 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5792 int nid = pgdat->node_id;
5795 pgdat_resize_init(pgdat);
5796 #ifdef CONFIG_NUMA_BALANCING
5797 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5798 pgdat->numabalancing_migrate_nr_pages = 0;
5799 pgdat->numabalancing_migrate_next_window = jiffies;
5801 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5802 spin_lock_init(&pgdat->split_queue_lock);
5803 INIT_LIST_HEAD(&pgdat->split_queue);
5804 pgdat->split_queue_len = 0;
5806 init_waitqueue_head(&pgdat->kswapd_wait);
5807 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5808 #ifdef CONFIG_COMPACTION
5809 init_waitqueue_head(&pgdat->kcompactd_wait);
5811 pgdat_page_ext_init(pgdat);
5812 spin_lock_init(&pgdat->lru_lock);
5813 lruvec_init(node_lruvec(pgdat));
5815 for (j = 0; j < MAX_NR_ZONES; j++) {
5816 struct zone *zone = pgdat->node_zones + j;
5817 unsigned long size, realsize, freesize, memmap_pages;
5818 unsigned long zone_start_pfn = zone->zone_start_pfn;
5820 size = zone->spanned_pages;
5821 realsize = freesize = zone->present_pages;
5824 * Adjust freesize so that it accounts for how much memory
5825 * is used by this zone for memmap. This affects the watermark
5826 * and per-cpu initialisations
5828 memmap_pages = calc_memmap_size(size, realsize);
5829 if (!is_highmem_idx(j)) {
5830 if (freesize >= memmap_pages) {
5831 freesize -= memmap_pages;
5834 " %s zone: %lu pages used for memmap\n",
5835 zone_names[j], memmap_pages);
5837 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
5838 zone_names[j], memmap_pages, freesize);
5841 /* Account for reserved pages */
5842 if (j == 0 && freesize > dma_reserve) {
5843 freesize -= dma_reserve;
5844 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5845 zone_names[0], dma_reserve);
5848 if (!is_highmem_idx(j))
5849 nr_kernel_pages += freesize;
5850 /* Charge for highmem memmap if there are enough kernel pages */
5851 else if (nr_kernel_pages > memmap_pages * 2)
5852 nr_kernel_pages -= memmap_pages;
5853 nr_all_pages += freesize;
5856 * Set an approximate value for lowmem here, it will be adjusted
5857 * when the bootmem allocator frees pages into the buddy system.
5858 * And all highmem pages will be managed by the buddy system.
5860 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5864 zone->name = zone_names[j];
5865 zone->zone_pgdat = pgdat;
5866 spin_lock_init(&zone->lock);
5867 zone_seqlock_init(zone);
5868 zone_pcp_init(zone);
5873 set_pageblock_order();
5874 setup_usemap(pgdat, zone, zone_start_pfn, size);
5875 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5877 memmap_init(size, nid, j, zone_start_pfn);
5881 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
5883 unsigned long __maybe_unused start = 0;
5884 unsigned long __maybe_unused offset = 0;
5886 /* Skip empty nodes */
5887 if (!pgdat->node_spanned_pages)
5890 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5891 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5892 offset = pgdat->node_start_pfn - start;
5893 /* ia64 gets its own node_mem_map, before this, without bootmem */
5894 if (!pgdat->node_mem_map) {
5895 unsigned long size, end;
5899 * The zone's endpoints aren't required to be MAX_ORDER
5900 * aligned but the node_mem_map endpoints must be in order
5901 * for the buddy allocator to function correctly.
5903 end = pgdat_end_pfn(pgdat);
5904 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5905 size = (end - start) * sizeof(struct page);
5906 map = alloc_remap(pgdat->node_id, size);
5908 map = memblock_virt_alloc_node_nopanic(size,
5910 pgdat->node_mem_map = map + offset;
5912 #ifndef CONFIG_NEED_MULTIPLE_NODES
5914 * With no DISCONTIG, the global mem_map is just set as node 0's
5916 if (pgdat == NODE_DATA(0)) {
5917 mem_map = NODE_DATA(0)->node_mem_map;
5918 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5919 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5921 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5924 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5927 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5928 unsigned long node_start_pfn, unsigned long *zholes_size)
5930 pg_data_t *pgdat = NODE_DATA(nid);
5931 unsigned long start_pfn = 0;
5932 unsigned long end_pfn = 0;
5934 /* pg_data_t should be reset to zero when it's allocated */
5935 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
5937 pgdat->node_id = nid;
5938 pgdat->node_start_pfn = node_start_pfn;
5939 pgdat->per_cpu_nodestats = NULL;
5940 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5941 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5942 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5943 (u64)start_pfn << PAGE_SHIFT,
5944 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5946 start_pfn = node_start_pfn;
5948 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5949 zones_size, zholes_size);
5951 alloc_node_mem_map(pgdat);
5952 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5953 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5954 nid, (unsigned long)pgdat,
5955 (unsigned long)pgdat->node_mem_map);
5958 reset_deferred_meminit(pgdat);
5959 free_area_init_core(pgdat);
5962 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5964 #if MAX_NUMNODES > 1
5966 * Figure out the number of possible node ids.
5968 void __init setup_nr_node_ids(void)
5970 unsigned int highest;
5972 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5973 nr_node_ids = highest + 1;
5978 * node_map_pfn_alignment - determine the maximum internode alignment
5980 * This function should be called after node map is populated and sorted.
5981 * It calculates the maximum power of two alignment which can distinguish
5984 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5985 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5986 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5987 * shifted, 1GiB is enough and this function will indicate so.
5989 * This is used to test whether pfn -> nid mapping of the chosen memory
5990 * model has fine enough granularity to avoid incorrect mapping for the
5991 * populated node map.
5993 * Returns the determined alignment in pfn's. 0 if there is no alignment
5994 * requirement (single node).
5996 unsigned long __init node_map_pfn_alignment(void)
5998 unsigned long accl_mask = 0, last_end = 0;
5999 unsigned long start, end, mask;
6003 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6004 if (!start || last_nid < 0 || last_nid == nid) {
6011 * Start with a mask granular enough to pin-point to the
6012 * start pfn and tick off bits one-by-one until it becomes
6013 * too coarse to separate the current node from the last.
6015 mask = ~((1 << __ffs(start)) - 1);
6016 while (mask && last_end <= (start & (mask << 1)))
6019 /* accumulate all internode masks */
6023 /* convert mask to number of pages */
6024 return ~accl_mask + 1;
6027 /* Find the lowest pfn for a node */
6028 static unsigned long __init find_min_pfn_for_node(int nid)
6030 unsigned long min_pfn = ULONG_MAX;
6031 unsigned long start_pfn;
6034 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6035 min_pfn = min(min_pfn, start_pfn);
6037 if (min_pfn == ULONG_MAX) {
6038 pr_warn("Could not find start_pfn for node %d\n", nid);
6046 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6048 * It returns the minimum PFN based on information provided via
6049 * memblock_set_node().
6051 unsigned long __init find_min_pfn_with_active_regions(void)
6053 return find_min_pfn_for_node(MAX_NUMNODES);
6057 * early_calculate_totalpages()
6058 * Sum pages in active regions for movable zone.
6059 * Populate N_MEMORY for calculating usable_nodes.
6061 static unsigned long __init early_calculate_totalpages(void)
6063 unsigned long totalpages = 0;
6064 unsigned long start_pfn, end_pfn;
6067 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6068 unsigned long pages = end_pfn - start_pfn;
6070 totalpages += pages;
6072 node_set_state(nid, N_MEMORY);
6078 * Find the PFN the Movable zone begins in each node. Kernel memory
6079 * is spread evenly between nodes as long as the nodes have enough
6080 * memory. When they don't, some nodes will have more kernelcore than
6083 static void __init find_zone_movable_pfns_for_nodes(void)
6086 unsigned long usable_startpfn;
6087 unsigned long kernelcore_node, kernelcore_remaining;
6088 /* save the state before borrow the nodemask */
6089 nodemask_t saved_node_state = node_states[N_MEMORY];
6090 unsigned long totalpages = early_calculate_totalpages();
6091 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6092 struct memblock_region *r;
6094 /* Need to find movable_zone earlier when movable_node is specified. */
6095 find_usable_zone_for_movable();
6098 * If movable_node is specified, ignore kernelcore and movablecore
6101 if (movable_node_is_enabled()) {
6102 for_each_memblock(memory, r) {
6103 if (!memblock_is_hotpluggable(r))
6108 usable_startpfn = PFN_DOWN(r->base);
6109 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6110 min(usable_startpfn, zone_movable_pfn[nid]) :
6118 * If kernelcore=mirror is specified, ignore movablecore option
6120 if (mirrored_kernelcore) {
6121 bool mem_below_4gb_not_mirrored = false;
6123 for_each_memblock(memory, r) {
6124 if (memblock_is_mirror(r))
6129 usable_startpfn = memblock_region_memory_base_pfn(r);
6131 if (usable_startpfn < 0x100000) {
6132 mem_below_4gb_not_mirrored = true;
6136 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6137 min(usable_startpfn, zone_movable_pfn[nid]) :
6141 if (mem_below_4gb_not_mirrored)
6142 pr_warn("This configuration results in unmirrored kernel memory.");
6148 * If movablecore=nn[KMG] was specified, calculate what size of
6149 * kernelcore that corresponds so that memory usable for
6150 * any allocation type is evenly spread. If both kernelcore
6151 * and movablecore are specified, then the value of kernelcore
6152 * will be used for required_kernelcore if it's greater than
6153 * what movablecore would have allowed.
6155 if (required_movablecore) {
6156 unsigned long corepages;
6159 * Round-up so that ZONE_MOVABLE is at least as large as what
6160 * was requested by the user
6162 required_movablecore =
6163 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6164 required_movablecore = min(totalpages, required_movablecore);
6165 corepages = totalpages - required_movablecore;
6167 required_kernelcore = max(required_kernelcore, corepages);
6171 * If kernelcore was not specified or kernelcore size is larger
6172 * than totalpages, there is no ZONE_MOVABLE.
6174 if (!required_kernelcore || required_kernelcore >= totalpages)
6177 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6178 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6181 /* Spread kernelcore memory as evenly as possible throughout nodes */
6182 kernelcore_node = required_kernelcore / usable_nodes;
6183 for_each_node_state(nid, N_MEMORY) {
6184 unsigned long start_pfn, end_pfn;
6187 * Recalculate kernelcore_node if the division per node
6188 * now exceeds what is necessary to satisfy the requested
6189 * amount of memory for the kernel
6191 if (required_kernelcore < kernelcore_node)
6192 kernelcore_node = required_kernelcore / usable_nodes;
6195 * As the map is walked, we track how much memory is usable
6196 * by the kernel using kernelcore_remaining. When it is
6197 * 0, the rest of the node is usable by ZONE_MOVABLE
6199 kernelcore_remaining = kernelcore_node;
6201 /* Go through each range of PFNs within this node */
6202 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6203 unsigned long size_pages;
6205 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6206 if (start_pfn >= end_pfn)
6209 /* Account for what is only usable for kernelcore */
6210 if (start_pfn < usable_startpfn) {
6211 unsigned long kernel_pages;
6212 kernel_pages = min(end_pfn, usable_startpfn)
6215 kernelcore_remaining -= min(kernel_pages,
6216 kernelcore_remaining);
6217 required_kernelcore -= min(kernel_pages,
6218 required_kernelcore);
6220 /* Continue if range is now fully accounted */
6221 if (end_pfn <= usable_startpfn) {
6224 * Push zone_movable_pfn to the end so
6225 * that if we have to rebalance
6226 * kernelcore across nodes, we will
6227 * not double account here
6229 zone_movable_pfn[nid] = end_pfn;
6232 start_pfn = usable_startpfn;
6236 * The usable PFN range for ZONE_MOVABLE is from
6237 * start_pfn->end_pfn. Calculate size_pages as the
6238 * number of pages used as kernelcore
6240 size_pages = end_pfn - start_pfn;
6241 if (size_pages > kernelcore_remaining)
6242 size_pages = kernelcore_remaining;
6243 zone_movable_pfn[nid] = start_pfn + size_pages;
6246 * Some kernelcore has been met, update counts and
6247 * break if the kernelcore for this node has been
6250 required_kernelcore -= min(required_kernelcore,
6252 kernelcore_remaining -= size_pages;
6253 if (!kernelcore_remaining)
6259 * If there is still required_kernelcore, we do another pass with one
6260 * less node in the count. This will push zone_movable_pfn[nid] further
6261 * along on the nodes that still have memory until kernelcore is
6265 if (usable_nodes && required_kernelcore > usable_nodes)
6269 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6270 for (nid = 0; nid < MAX_NUMNODES; nid++) {
6271 unsigned long start_pfn, end_pfn;
6273 zone_movable_pfn[nid] =
6274 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6276 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6277 if (zone_movable_pfn[nid] >= end_pfn)
6278 zone_movable_pfn[nid] = 0;
6282 /* restore the node_state */
6283 node_states[N_MEMORY] = saved_node_state;
6286 /* Any regular or high memory on that node ? */
6287 static void check_for_memory(pg_data_t *pgdat, int nid)
6289 enum zone_type zone_type;
6291 if (N_MEMORY == N_NORMAL_MEMORY)
6294 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6295 struct zone *zone = &pgdat->node_zones[zone_type];
6296 if (populated_zone(zone)) {
6297 node_set_state(nid, N_HIGH_MEMORY);
6298 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6299 zone_type <= ZONE_NORMAL)
6300 node_set_state(nid, N_NORMAL_MEMORY);
6307 * free_area_init_nodes - Initialise all pg_data_t and zone data
6308 * @max_zone_pfn: an array of max PFNs for each zone
6310 * This will call free_area_init_node() for each active node in the system.
6311 * Using the page ranges provided by memblock_set_node(), the size of each
6312 * zone in each node and their holes is calculated. If the maximum PFN
6313 * between two adjacent zones match, it is assumed that the zone is empty.
6314 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6315 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6316 * starts where the previous one ended. For example, ZONE_DMA32 starts
6317 * at arch_max_dma_pfn.
6319 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6321 unsigned long start_pfn, end_pfn;
6324 /* Record where the zone boundaries are */
6325 memset(arch_zone_lowest_possible_pfn, 0,
6326 sizeof(arch_zone_lowest_possible_pfn));
6327 memset(arch_zone_highest_possible_pfn, 0,
6328 sizeof(arch_zone_highest_possible_pfn));
6330 start_pfn = find_min_pfn_with_active_regions();
6332 for (i = 0; i < MAX_NR_ZONES; i++) {
6333 if (i == ZONE_MOVABLE)
6336 end_pfn = max(max_zone_pfn[i], start_pfn);
6337 arch_zone_lowest_possible_pfn[i] = start_pfn;
6338 arch_zone_highest_possible_pfn[i] = end_pfn;
6340 start_pfn = end_pfn;
6342 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
6343 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
6345 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6346 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6347 find_zone_movable_pfns_for_nodes();
6349 /* Print out the zone ranges */
6350 pr_info("Zone ranges:\n");
6351 for (i = 0; i < MAX_NR_ZONES; i++) {
6352 if (i == ZONE_MOVABLE)
6354 pr_info(" %-8s ", zone_names[i]);
6355 if (arch_zone_lowest_possible_pfn[i] ==
6356 arch_zone_highest_possible_pfn[i])
6359 pr_cont("[mem %#018Lx-%#018Lx]\n",
6360 (u64)arch_zone_lowest_possible_pfn[i]
6362 ((u64)arch_zone_highest_possible_pfn[i]
6363 << PAGE_SHIFT) - 1);
6366 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6367 pr_info("Movable zone start for each node\n");
6368 for (i = 0; i < MAX_NUMNODES; i++) {
6369 if (zone_movable_pfn[i])
6370 pr_info(" Node %d: %#018Lx\n", i,
6371 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6374 /* Print out the early node map */
6375 pr_info("Early memory node ranges\n");
6376 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6377 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6378 (u64)start_pfn << PAGE_SHIFT,
6379 ((u64)end_pfn << PAGE_SHIFT) - 1);
6381 /* Initialise every node */
6382 mminit_verify_pageflags_layout();
6383 setup_nr_node_ids();
6384 for_each_online_node(nid) {
6385 pg_data_t *pgdat = NODE_DATA(nid);
6386 free_area_init_node(nid, NULL,
6387 find_min_pfn_for_node(nid), NULL);
6389 /* Any memory on that node */
6390 if (pgdat->node_present_pages)
6391 node_set_state(nid, N_MEMORY);
6392 check_for_memory(pgdat, nid);
6396 static int __init cmdline_parse_core(char *p, unsigned long *core)
6398 unsigned long long coremem;
6402 coremem = memparse(p, &p);
6403 *core = coremem >> PAGE_SHIFT;
6405 /* Paranoid check that UL is enough for the coremem value */
6406 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6412 * kernelcore=size sets the amount of memory for use for allocations that
6413 * cannot be reclaimed or migrated.
6415 static int __init cmdline_parse_kernelcore(char *p)
6417 /* parse kernelcore=mirror */
6418 if (parse_option_str(p, "mirror")) {
6419 mirrored_kernelcore = true;
6423 return cmdline_parse_core(p, &required_kernelcore);
6427 * movablecore=size sets the amount of memory for use for allocations that
6428 * can be reclaimed or migrated.
6430 static int __init cmdline_parse_movablecore(char *p)
6432 return cmdline_parse_core(p, &required_movablecore);
6435 early_param("kernelcore", cmdline_parse_kernelcore);
6436 early_param("movablecore", cmdline_parse_movablecore);
6438 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6440 void adjust_managed_page_count(struct page *page, long count)
6442 spin_lock(&managed_page_count_lock);
6443 page_zone(page)->managed_pages += count;
6444 totalram_pages += count;
6445 #ifdef CONFIG_HIGHMEM
6446 if (PageHighMem(page))
6447 totalhigh_pages += count;
6449 spin_unlock(&managed_page_count_lock);
6451 EXPORT_SYMBOL(adjust_managed_page_count);
6453 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6456 unsigned long pages = 0;
6458 start = (void *)PAGE_ALIGN((unsigned long)start);
6459 end = (void *)((unsigned long)end & PAGE_MASK);
6460 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6461 if ((unsigned int)poison <= 0xFF)
6462 memset(pos, poison, PAGE_SIZE);
6463 free_reserved_page(virt_to_page(pos));
6467 pr_info("Freeing %s memory: %ldK\n",
6468 s, pages << (PAGE_SHIFT - 10));
6472 EXPORT_SYMBOL(free_reserved_area);
6474 #ifdef CONFIG_HIGHMEM
6475 void free_highmem_page(struct page *page)
6477 __free_reserved_page(page);
6479 page_zone(page)->managed_pages++;
6485 void __init mem_init_print_info(const char *str)
6487 unsigned long physpages, codesize, datasize, rosize, bss_size;
6488 unsigned long init_code_size, init_data_size;
6490 physpages = get_num_physpages();
6491 codesize = _etext - _stext;
6492 datasize = _edata - _sdata;
6493 rosize = __end_rodata - __start_rodata;
6494 bss_size = __bss_stop - __bss_start;
6495 init_data_size = __init_end - __init_begin;
6496 init_code_size = _einittext - _sinittext;
6499 * Detect special cases and adjust section sizes accordingly:
6500 * 1) .init.* may be embedded into .data sections
6501 * 2) .init.text.* may be out of [__init_begin, __init_end],
6502 * please refer to arch/tile/kernel/vmlinux.lds.S.
6503 * 3) .rodata.* may be embedded into .text or .data sections.
6505 #define adj_init_size(start, end, size, pos, adj) \
6507 if (&start[0] <= &pos[0] && &pos[0] < &end[0] && size > adj) \
6511 adj_init_size(__init_begin, __init_end, init_data_size,
6512 _sinittext, init_code_size);
6513 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6514 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6515 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6516 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6518 #undef adj_init_size
6520 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6521 #ifdef CONFIG_HIGHMEM
6525 nr_free_pages() << (PAGE_SHIFT - 10),
6526 physpages << (PAGE_SHIFT - 10),
6527 codesize >> 10, datasize >> 10, rosize >> 10,
6528 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6529 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6530 totalcma_pages << (PAGE_SHIFT - 10),
6531 #ifdef CONFIG_HIGHMEM
6532 totalhigh_pages << (PAGE_SHIFT - 10),
6534 str ? ", " : "", str ? str : "");
6538 * set_dma_reserve - set the specified number of pages reserved in the first zone
6539 * @new_dma_reserve: The number of pages to mark reserved
6541 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6542 * In the DMA zone, a significant percentage may be consumed by kernel image
6543 * and other unfreeable allocations which can skew the watermarks badly. This
6544 * function may optionally be used to account for unfreeable pages in the
6545 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6546 * smaller per-cpu batchsize.
6548 void __init set_dma_reserve(unsigned long new_dma_reserve)
6550 dma_reserve = new_dma_reserve;
6553 void __init free_area_init(unsigned long *zones_size)
6555 free_area_init_node(0, zones_size,
6556 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6559 static int page_alloc_cpu_notify(struct notifier_block *self,
6560 unsigned long action, void *hcpu)
6562 int cpu = (unsigned long)hcpu;
6564 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6565 lru_add_drain_cpu(cpu);
6569 * Spill the event counters of the dead processor
6570 * into the current processors event counters.
6571 * This artificially elevates the count of the current
6574 vm_events_fold_cpu(cpu);
6577 * Zero the differential counters of the dead processor
6578 * so that the vm statistics are consistent.
6580 * This is only okay since the processor is dead and cannot
6581 * race with what we are doing.
6583 cpu_vm_stats_fold(cpu);
6588 void __init page_alloc_init(void)
6590 hotcpu_notifier(page_alloc_cpu_notify, 0);
6594 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6595 * or min_free_kbytes changes.
6597 static void calculate_totalreserve_pages(void)
6599 struct pglist_data *pgdat;
6600 unsigned long reserve_pages = 0;
6601 enum zone_type i, j;
6603 for_each_online_pgdat(pgdat) {
6605 pgdat->totalreserve_pages = 0;
6607 for (i = 0; i < MAX_NR_ZONES; i++) {
6608 struct zone *zone = pgdat->node_zones + i;
6611 /* Find valid and maximum lowmem_reserve in the zone */
6612 for (j = i; j < MAX_NR_ZONES; j++) {
6613 if (zone->lowmem_reserve[j] > max)
6614 max = zone->lowmem_reserve[j];
6617 /* we treat the high watermark as reserved pages. */
6618 max += high_wmark_pages(zone);
6620 if (max > zone->managed_pages)
6621 max = zone->managed_pages;
6623 pgdat->totalreserve_pages += max;
6625 reserve_pages += max;
6628 totalreserve_pages = reserve_pages;
6632 * setup_per_zone_lowmem_reserve - called whenever
6633 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6634 * has a correct pages reserved value, so an adequate number of
6635 * pages are left in the zone after a successful __alloc_pages().
6637 static void setup_per_zone_lowmem_reserve(void)
6639 struct pglist_data *pgdat;
6640 enum zone_type j, idx;
6642 for_each_online_pgdat(pgdat) {
6643 for (j = 0; j < MAX_NR_ZONES; j++) {
6644 struct zone *zone = pgdat->node_zones + j;
6645 unsigned long managed_pages = zone->managed_pages;
6647 zone->lowmem_reserve[j] = 0;
6651 struct zone *lower_zone;
6655 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6656 sysctl_lowmem_reserve_ratio[idx] = 1;
6658 lower_zone = pgdat->node_zones + idx;
6659 lower_zone->lowmem_reserve[j] = managed_pages /
6660 sysctl_lowmem_reserve_ratio[idx];
6661 managed_pages += lower_zone->managed_pages;
6666 /* update totalreserve_pages */
6667 calculate_totalreserve_pages();
6670 static void __setup_per_zone_wmarks(void)
6672 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6673 unsigned long lowmem_pages = 0;
6675 unsigned long flags;
6677 /* Calculate total number of !ZONE_HIGHMEM pages */
6678 for_each_zone(zone) {
6679 if (!is_highmem(zone))
6680 lowmem_pages += zone->managed_pages;
6683 for_each_zone(zone) {
6686 spin_lock_irqsave(&zone->lock, flags);
6687 tmp = (u64)pages_min * zone->managed_pages;
6688 do_div(tmp, lowmem_pages);
6689 if (is_highmem(zone)) {
6691 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6692 * need highmem pages, so cap pages_min to a small
6695 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6696 * deltas control asynch page reclaim, and so should
6697 * not be capped for highmem.
6699 unsigned long min_pages;
6701 min_pages = zone->managed_pages / 1024;
6702 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6703 zone->watermark[WMARK_MIN] = min_pages;
6706 * If it's a lowmem zone, reserve a number of pages
6707 * proportionate to the zone's size.
6709 zone->watermark[WMARK_MIN] = tmp;
6713 * Set the kswapd watermarks distance according to the
6714 * scale factor in proportion to available memory, but
6715 * ensure a minimum size on small systems.
6717 tmp = max_t(u64, tmp >> 2,
6718 mult_frac(zone->managed_pages,
6719 watermark_scale_factor, 10000));
6721 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6722 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
6724 spin_unlock_irqrestore(&zone->lock, flags);
6727 /* update totalreserve_pages */
6728 calculate_totalreserve_pages();
6732 * setup_per_zone_wmarks - called when min_free_kbytes changes
6733 * or when memory is hot-{added|removed}
6735 * Ensures that the watermark[min,low,high] values for each zone are set
6736 * correctly with respect to min_free_kbytes.
6738 void setup_per_zone_wmarks(void)
6740 mutex_lock(&zonelists_mutex);
6741 __setup_per_zone_wmarks();
6742 mutex_unlock(&zonelists_mutex);
6746 * Initialise min_free_kbytes.
6748 * For small machines we want it small (128k min). For large machines
6749 * we want it large (64MB max). But it is not linear, because network
6750 * bandwidth does not increase linearly with machine size. We use
6752 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6753 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6769 int __meminit init_per_zone_wmark_min(void)
6771 unsigned long lowmem_kbytes;
6772 int new_min_free_kbytes;
6774 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6775 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6777 if (new_min_free_kbytes > user_min_free_kbytes) {
6778 min_free_kbytes = new_min_free_kbytes;
6779 if (min_free_kbytes < 128)
6780 min_free_kbytes = 128;
6781 if (min_free_kbytes > 65536)
6782 min_free_kbytes = 65536;
6784 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6785 new_min_free_kbytes, user_min_free_kbytes);
6787 setup_per_zone_wmarks();
6788 refresh_zone_stat_thresholds();
6789 setup_per_zone_lowmem_reserve();
6792 setup_min_unmapped_ratio();
6793 setup_min_slab_ratio();
6796 khugepaged_min_free_kbytes_update();
6800 postcore_initcall(init_per_zone_wmark_min)
6803 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6804 * that we can call two helper functions whenever min_free_kbytes
6807 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6808 void __user *buffer, size_t *length, loff_t *ppos)
6812 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6817 user_min_free_kbytes = min_free_kbytes;
6818 setup_per_zone_wmarks();
6823 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
6824 void __user *buffer, size_t *length, loff_t *ppos)
6828 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6833 setup_per_zone_wmarks();
6839 static void setup_min_unmapped_ratio(void)
6844 for_each_online_pgdat(pgdat)
6845 pgdat->min_unmapped_pages = 0;
6848 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
6849 sysctl_min_unmapped_ratio) / 100;
6853 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6854 void __user *buffer, size_t *length, loff_t *ppos)
6858 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6862 setup_min_unmapped_ratio();
6867 static void setup_min_slab_ratio(void)
6872 for_each_online_pgdat(pgdat)
6873 pgdat->min_slab_pages = 0;
6876 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
6877 sysctl_min_slab_ratio) / 100;
6880 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6881 void __user *buffer, size_t *length, loff_t *ppos)
6885 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6889 setup_min_slab_ratio();
6896 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6897 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6898 * whenever sysctl_lowmem_reserve_ratio changes.
6900 * The reserve ratio obviously has absolutely no relation with the
6901 * minimum watermarks. The lowmem reserve ratio can only make sense
6902 * if in function of the boot time zone sizes.
6904 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6905 void __user *buffer, size_t *length, loff_t *ppos)
6907 proc_dointvec_minmax(table, write, buffer, length, ppos);
6908 setup_per_zone_lowmem_reserve();
6913 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6914 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6915 * pagelist can have before it gets flushed back to buddy allocator.
6917 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6918 void __user *buffer, size_t *length, loff_t *ppos)
6921 int old_percpu_pagelist_fraction;
6924 mutex_lock(&pcp_batch_high_lock);
6925 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6927 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6928 if (!write || ret < 0)
6931 /* Sanity checking to avoid pcp imbalance */
6932 if (percpu_pagelist_fraction &&
6933 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6934 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6940 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6943 for_each_populated_zone(zone) {
6946 for_each_possible_cpu(cpu)
6947 pageset_set_high_and_batch(zone,
6948 per_cpu_ptr(zone->pageset, cpu));
6951 mutex_unlock(&pcp_batch_high_lock);
6956 int hashdist = HASHDIST_DEFAULT;
6958 static int __init set_hashdist(char *str)
6962 hashdist = simple_strtoul(str, &str, 0);
6965 __setup("hashdist=", set_hashdist);
6968 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
6970 * Returns the number of pages that arch has reserved but
6971 * is not known to alloc_large_system_hash().
6973 static unsigned long __init arch_reserved_kernel_pages(void)
6980 * allocate a large system hash table from bootmem
6981 * - it is assumed that the hash table must contain an exact power-of-2
6982 * quantity of entries
6983 * - limit is the number of hash buckets, not the total allocation size
6985 void *__init alloc_large_system_hash(const char *tablename,
6986 unsigned long bucketsize,
6987 unsigned long numentries,
6990 unsigned int *_hash_shift,
6991 unsigned int *_hash_mask,
6992 unsigned long low_limit,
6993 unsigned long high_limit)
6995 unsigned long long max = high_limit;
6996 unsigned long log2qty, size;
6999 /* allow the kernel cmdline to have a say */
7001 /* round applicable memory size up to nearest megabyte */
7002 numentries = nr_kernel_pages;
7003 numentries -= arch_reserved_kernel_pages();
7005 /* It isn't necessary when PAGE_SIZE >= 1MB */
7006 if (PAGE_SHIFT < 20)
7007 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7009 /* limit to 1 bucket per 2^scale bytes of low memory */
7010 if (scale > PAGE_SHIFT)
7011 numentries >>= (scale - PAGE_SHIFT);
7013 numentries <<= (PAGE_SHIFT - scale);
7015 /* Make sure we've got at least a 0-order allocation.. */
7016 if (unlikely(flags & HASH_SMALL)) {
7017 /* Makes no sense without HASH_EARLY */
7018 WARN_ON(!(flags & HASH_EARLY));
7019 if (!(numentries >> *_hash_shift)) {
7020 numentries = 1UL << *_hash_shift;
7021 BUG_ON(!numentries);
7023 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7024 numentries = PAGE_SIZE / bucketsize;
7026 numentries = roundup_pow_of_two(numentries);
7028 /* limit allocation size to 1/16 total memory by default */
7030 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7031 do_div(max, bucketsize);
7033 max = min(max, 0x80000000ULL);
7035 if (numentries < low_limit)
7036 numentries = low_limit;
7037 if (numentries > max)
7040 log2qty = ilog2(numentries);
7043 size = bucketsize << log2qty;
7044 if (flags & HASH_EARLY)
7045 table = memblock_virt_alloc_nopanic(size, 0);
7047 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
7050 * If bucketsize is not a power-of-two, we may free
7051 * some pages at the end of hash table which
7052 * alloc_pages_exact() automatically does
7054 if (get_order(size) < MAX_ORDER) {
7055 table = alloc_pages_exact(size, GFP_ATOMIC);
7056 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
7059 } while (!table && size > PAGE_SIZE && --log2qty);
7062 panic("Failed to allocate %s hash table\n", tablename);
7064 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7065 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7068 *_hash_shift = log2qty;
7070 *_hash_mask = (1 << log2qty) - 1;
7076 * This function checks whether pageblock includes unmovable pages or not.
7077 * If @count is not zero, it is okay to include less @count unmovable pages
7079 * PageLRU check without isolation or lru_lock could race so that
7080 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
7081 * expect this function should be exact.
7083 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7084 bool skip_hwpoisoned_pages)
7086 unsigned long pfn, iter, found;
7090 * For avoiding noise data, lru_add_drain_all() should be called
7091 * If ZONE_MOVABLE, the zone never contains unmovable pages
7093 if (zone_idx(zone) == ZONE_MOVABLE)
7095 mt = get_pageblock_migratetype(page);
7096 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
7099 pfn = page_to_pfn(page);
7100 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7101 unsigned long check = pfn + iter;
7103 if (!pfn_valid_within(check))
7106 page = pfn_to_page(check);
7109 * Hugepages are not in LRU lists, but they're movable.
7110 * We need not scan over tail pages bacause we don't
7111 * handle each tail page individually in migration.
7113 if (PageHuge(page)) {
7114 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7119 * We can't use page_count without pin a page
7120 * because another CPU can free compound page.
7121 * This check already skips compound tails of THP
7122 * because their page->_refcount is zero at all time.
7124 if (!page_ref_count(page)) {
7125 if (PageBuddy(page))
7126 iter += (1 << page_order(page)) - 1;
7131 * The HWPoisoned page may be not in buddy system, and
7132 * page_count() is not 0.
7134 if (skip_hwpoisoned_pages && PageHWPoison(page))
7140 * If there are RECLAIMABLE pages, we need to check
7141 * it. But now, memory offline itself doesn't call
7142 * shrink_node_slabs() and it still to be fixed.
7145 * If the page is not RAM, page_count()should be 0.
7146 * we don't need more check. This is an _used_ not-movable page.
7148 * The problematic thing here is PG_reserved pages. PG_reserved
7149 * is set to both of a memory hole page and a _used_ kernel
7158 bool is_pageblock_removable_nolock(struct page *page)
7164 * We have to be careful here because we are iterating over memory
7165 * sections which are not zone aware so we might end up outside of
7166 * the zone but still within the section.
7167 * We have to take care about the node as well. If the node is offline
7168 * its NODE_DATA will be NULL - see page_zone.
7170 if (!node_online(page_to_nid(page)))
7173 zone = page_zone(page);
7174 pfn = page_to_pfn(page);
7175 if (!zone_spans_pfn(zone, pfn))
7178 return !has_unmovable_pages(zone, page, 0, true);
7181 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7183 static unsigned long pfn_max_align_down(unsigned long pfn)
7185 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7186 pageblock_nr_pages) - 1);
7189 static unsigned long pfn_max_align_up(unsigned long pfn)
7191 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7192 pageblock_nr_pages));
7195 /* [start, end) must belong to a single zone. */
7196 static int __alloc_contig_migrate_range(struct compact_control *cc,
7197 unsigned long start, unsigned long end)
7199 /* This function is based on compact_zone() from compaction.c. */
7200 unsigned long nr_reclaimed;
7201 unsigned long pfn = start;
7202 unsigned int tries = 0;
7207 while (pfn < end || !list_empty(&cc->migratepages)) {
7208 if (fatal_signal_pending(current)) {
7213 if (list_empty(&cc->migratepages)) {
7214 cc->nr_migratepages = 0;
7215 pfn = isolate_migratepages_range(cc, pfn, end);
7221 } else if (++tries == 5) {
7222 ret = ret < 0 ? ret : -EBUSY;
7226 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7228 cc->nr_migratepages -= nr_reclaimed;
7230 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7231 NULL, 0, cc->mode, MR_CMA);
7234 putback_movable_pages(&cc->migratepages);
7241 * alloc_contig_range() -- tries to allocate given range of pages
7242 * @start: start PFN to allocate
7243 * @end: one-past-the-last PFN to allocate
7244 * @migratetype: migratetype of the underlaying pageblocks (either
7245 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7246 * in range must have the same migratetype and it must
7247 * be either of the two.
7249 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7250 * aligned, however it's the caller's responsibility to guarantee that
7251 * we are the only thread that changes migrate type of pageblocks the
7254 * The PFN range must belong to a single zone.
7256 * Returns zero on success or negative error code. On success all
7257 * pages which PFN is in [start, end) are allocated for the caller and
7258 * need to be freed with free_contig_range().
7260 int alloc_contig_range(unsigned long start, unsigned long end,
7261 unsigned migratetype)
7263 unsigned long outer_start, outer_end;
7267 struct compact_control cc = {
7268 .nr_migratepages = 0,
7270 .zone = page_zone(pfn_to_page(start)),
7271 .mode = MIGRATE_SYNC,
7272 .ignore_skip_hint = true,
7274 INIT_LIST_HEAD(&cc.migratepages);
7277 * What we do here is we mark all pageblocks in range as
7278 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7279 * have different sizes, and due to the way page allocator
7280 * work, we align the range to biggest of the two pages so
7281 * that page allocator won't try to merge buddies from
7282 * different pageblocks and change MIGRATE_ISOLATE to some
7283 * other migration type.
7285 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7286 * migrate the pages from an unaligned range (ie. pages that
7287 * we are interested in). This will put all the pages in
7288 * range back to page allocator as MIGRATE_ISOLATE.
7290 * When this is done, we take the pages in range from page
7291 * allocator removing them from the buddy system. This way
7292 * page allocator will never consider using them.
7294 * This lets us mark the pageblocks back as
7295 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7296 * aligned range but not in the unaligned, original range are
7297 * put back to page allocator so that buddy can use them.
7300 ret = start_isolate_page_range(pfn_max_align_down(start),
7301 pfn_max_align_up(end), migratetype,
7307 * In case of -EBUSY, we'd like to know which page causes problem.
7308 * So, just fall through. test_pages_isolated() has a tracepoint
7309 * which will report the busy page.
7311 * It is possible that busy pages could become available before
7312 * the call to test_pages_isolated, and the range will actually be
7313 * allocated. So, if we fall through be sure to clear ret so that
7314 * -EBUSY is not accidentally used or returned to caller.
7316 ret = __alloc_contig_migrate_range(&cc, start, end);
7317 if (ret && ret != -EBUSY)
7322 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7323 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7324 * more, all pages in [start, end) are free in page allocator.
7325 * What we are going to do is to allocate all pages from
7326 * [start, end) (that is remove them from page allocator).
7328 * The only problem is that pages at the beginning and at the
7329 * end of interesting range may be not aligned with pages that
7330 * page allocator holds, ie. they can be part of higher order
7331 * pages. Because of this, we reserve the bigger range and
7332 * once this is done free the pages we are not interested in.
7334 * We don't have to hold zone->lock here because the pages are
7335 * isolated thus they won't get removed from buddy.
7338 lru_add_drain_all();
7339 drain_all_pages(cc.zone);
7342 outer_start = start;
7343 while (!PageBuddy(pfn_to_page(outer_start))) {
7344 if (++order >= MAX_ORDER) {
7345 outer_start = start;
7348 outer_start &= ~0UL << order;
7351 if (outer_start != start) {
7352 order = page_order(pfn_to_page(outer_start));
7355 * outer_start page could be small order buddy page and
7356 * it doesn't include start page. Adjust outer_start
7357 * in this case to report failed page properly
7358 * on tracepoint in test_pages_isolated()
7360 if (outer_start + (1UL << order) <= start)
7361 outer_start = start;
7364 /* Make sure the range is really isolated. */
7365 if (test_pages_isolated(outer_start, end, false)) {
7366 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
7367 __func__, outer_start, end);
7372 /* Grab isolated pages from freelists. */
7373 outer_end = isolate_freepages_range(&cc, outer_start, end);
7379 /* Free head and tail (if any) */
7380 if (start != outer_start)
7381 free_contig_range(outer_start, start - outer_start);
7382 if (end != outer_end)
7383 free_contig_range(end, outer_end - end);
7386 undo_isolate_page_range(pfn_max_align_down(start),
7387 pfn_max_align_up(end), migratetype);
7391 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7393 unsigned int count = 0;
7395 for (; nr_pages--; pfn++) {
7396 struct page *page = pfn_to_page(pfn);
7398 count += page_count(page) != 1;
7401 WARN(count != 0, "%d pages are still in use!\n", count);
7405 #ifdef CONFIG_MEMORY_HOTPLUG
7407 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7408 * page high values need to be recalulated.
7410 void __meminit zone_pcp_update(struct zone *zone)
7413 mutex_lock(&pcp_batch_high_lock);
7414 for_each_possible_cpu(cpu)
7415 pageset_set_high_and_batch(zone,
7416 per_cpu_ptr(zone->pageset, cpu));
7417 mutex_unlock(&pcp_batch_high_lock);
7421 void zone_pcp_reset(struct zone *zone)
7423 unsigned long flags;
7425 struct per_cpu_pageset *pset;
7427 /* avoid races with drain_pages() */
7428 local_irq_save(flags);
7429 if (zone->pageset != &boot_pageset) {
7430 for_each_online_cpu(cpu) {
7431 pset = per_cpu_ptr(zone->pageset, cpu);
7432 drain_zonestat(zone, pset);
7434 free_percpu(zone->pageset);
7435 zone->pageset = &boot_pageset;
7437 local_irq_restore(flags);
7440 #ifdef CONFIG_MEMORY_HOTREMOVE
7442 * All pages in the range must be in a single zone and isolated
7443 * before calling this.
7446 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7450 unsigned int order, i;
7452 unsigned long flags;
7453 /* find the first valid pfn */
7454 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7459 zone = page_zone(pfn_to_page(pfn));
7460 spin_lock_irqsave(&zone->lock, flags);
7462 while (pfn < end_pfn) {
7463 if (!pfn_valid(pfn)) {
7467 page = pfn_to_page(pfn);
7469 * The HWPoisoned page may be not in buddy system, and
7470 * page_count() is not 0.
7472 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7474 SetPageReserved(page);
7478 BUG_ON(page_count(page));
7479 BUG_ON(!PageBuddy(page));
7480 order = page_order(page);
7481 #ifdef CONFIG_DEBUG_VM
7482 pr_info("remove from free list %lx %d %lx\n",
7483 pfn, 1 << order, end_pfn);
7485 list_del(&page->lru);
7486 rmv_page_order(page);
7487 zone->free_area[order].nr_free--;
7488 for (i = 0; i < (1 << order); i++)
7489 SetPageReserved((page+i));
7490 pfn += (1 << order);
7492 spin_unlock_irqrestore(&zone->lock, flags);
7496 bool is_free_buddy_page(struct page *page)
7498 struct zone *zone = page_zone(page);
7499 unsigned long pfn = page_to_pfn(page);
7500 unsigned long flags;
7503 spin_lock_irqsave(&zone->lock, flags);
7504 for (order = 0; order < MAX_ORDER; order++) {
7505 struct page *page_head = page - (pfn & ((1 << order) - 1));
7507 if (PageBuddy(page_head) && page_order(page_head) >= order)
7510 spin_unlock_irqrestore(&zone->lock, flags);
7512 return order < MAX_ORDER;