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>
67 #include <asm/sections.h>
68 #include <asm/tlbflush.h>
69 #include <asm/div64.h>
72 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
73 static DEFINE_MUTEX(pcp_batch_high_lock);
74 #define MIN_PERCPU_PAGELIST_FRACTION (8)
76 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
77 DEFINE_PER_CPU(int, numa_node);
78 EXPORT_PER_CPU_SYMBOL(numa_node);
81 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
83 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
84 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
85 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
86 * defined in <linux/topology.h>.
88 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
89 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
90 int _node_numa_mem_[MAX_NUMNODES];
94 * Array of node states.
96 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
97 [N_POSSIBLE] = NODE_MASK_ALL,
98 [N_ONLINE] = { { [0] = 1UL } },
100 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
101 #ifdef CONFIG_HIGHMEM
102 [N_HIGH_MEMORY] = { { [0] = 1UL } },
104 #ifdef CONFIG_MOVABLE_NODE
105 [N_MEMORY] = { { [0] = 1UL } },
107 [N_CPU] = { { [0] = 1UL } },
110 EXPORT_SYMBOL(node_states);
112 /* Protect totalram_pages and zone->managed_pages */
113 static DEFINE_SPINLOCK(managed_page_count_lock);
115 unsigned long totalram_pages __read_mostly;
116 unsigned long totalreserve_pages __read_mostly;
117 unsigned long totalcma_pages __read_mostly;
119 int percpu_pagelist_fraction;
120 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
123 * A cached value of the page's pageblock's migratetype, used when the page is
124 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
125 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
126 * Also the migratetype set in the page does not necessarily match the pcplist
127 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
128 * other index - this ensures that it will be put on the correct CMA freelist.
130 static inline int get_pcppage_migratetype(struct page *page)
135 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
137 page->index = migratetype;
140 #ifdef CONFIG_PM_SLEEP
142 * The following functions are used by the suspend/hibernate code to temporarily
143 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
144 * while devices are suspended. To avoid races with the suspend/hibernate code,
145 * they should always be called with pm_mutex held (gfp_allowed_mask also should
146 * only be modified with pm_mutex held, unless the suspend/hibernate code is
147 * guaranteed not to run in parallel with that modification).
150 static gfp_t saved_gfp_mask;
152 void pm_restore_gfp_mask(void)
154 WARN_ON(!mutex_is_locked(&pm_mutex));
155 if (saved_gfp_mask) {
156 gfp_allowed_mask = saved_gfp_mask;
161 void pm_restrict_gfp_mask(void)
163 WARN_ON(!mutex_is_locked(&pm_mutex));
164 WARN_ON(saved_gfp_mask);
165 saved_gfp_mask = gfp_allowed_mask;
166 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
169 bool pm_suspended_storage(void)
171 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
175 #endif /* CONFIG_PM_SLEEP */
177 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
178 unsigned int pageblock_order __read_mostly;
181 static void __free_pages_ok(struct page *page, unsigned int order);
184 * results with 256, 32 in the lowmem_reserve sysctl:
185 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
186 * 1G machine -> (16M dma, 784M normal, 224M high)
187 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
188 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
189 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
191 * TBD: should special case ZONE_DMA32 machines here - in those we normally
192 * don't need any ZONE_NORMAL reservation
194 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
195 #ifdef CONFIG_ZONE_DMA
198 #ifdef CONFIG_ZONE_DMA32
201 #ifdef CONFIG_HIGHMEM
207 EXPORT_SYMBOL(totalram_pages);
209 static char * const zone_names[MAX_NR_ZONES] = {
210 #ifdef CONFIG_ZONE_DMA
213 #ifdef CONFIG_ZONE_DMA32
217 #ifdef CONFIG_HIGHMEM
221 #ifdef CONFIG_ZONE_DEVICE
226 char * const migratetype_names[MIGRATE_TYPES] = {
234 #ifdef CONFIG_MEMORY_ISOLATION
239 compound_page_dtor * const compound_page_dtors[] = {
242 #ifdef CONFIG_HUGETLB_PAGE
245 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
250 int min_free_kbytes = 1024;
251 int user_min_free_kbytes = -1;
253 static unsigned long __meminitdata nr_kernel_pages;
254 static unsigned long __meminitdata nr_all_pages;
255 static unsigned long __meminitdata dma_reserve;
257 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
258 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
259 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
260 static unsigned long __initdata required_kernelcore;
261 static unsigned long __initdata required_movablecore;
262 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
263 static bool mirrored_kernelcore;
265 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
267 EXPORT_SYMBOL(movable_zone);
268 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
271 int nr_node_ids __read_mostly = MAX_NUMNODES;
272 int nr_online_nodes __read_mostly = 1;
273 EXPORT_SYMBOL(nr_node_ids);
274 EXPORT_SYMBOL(nr_online_nodes);
277 int page_group_by_mobility_disabled __read_mostly;
279 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
280 static inline void reset_deferred_meminit(pg_data_t *pgdat)
282 pgdat->first_deferred_pfn = ULONG_MAX;
285 /* Returns true if the struct page for the pfn is uninitialised */
286 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
288 if (pfn >= NODE_DATA(early_pfn_to_nid(pfn))->first_deferred_pfn)
294 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
296 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
303 * Returns false when the remaining initialisation should be deferred until
304 * later in the boot cycle when it can be parallelised.
306 static inline bool update_defer_init(pg_data_t *pgdat,
307 unsigned long pfn, unsigned long zone_end,
308 unsigned long *nr_initialised)
310 /* Always populate low zones for address-contrained allocations */
311 if (zone_end < pgdat_end_pfn(pgdat))
314 /* Initialise at least 2G of the highest zone */
316 if (*nr_initialised > (2UL << (30 - PAGE_SHIFT)) &&
317 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
318 pgdat->first_deferred_pfn = pfn;
325 static inline void reset_deferred_meminit(pg_data_t *pgdat)
329 static inline bool early_page_uninitialised(unsigned long pfn)
334 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
339 static inline bool update_defer_init(pg_data_t *pgdat,
340 unsigned long pfn, unsigned long zone_end,
341 unsigned long *nr_initialised)
348 void set_pageblock_migratetype(struct page *page, int migratetype)
350 if (unlikely(page_group_by_mobility_disabled &&
351 migratetype < MIGRATE_PCPTYPES))
352 migratetype = MIGRATE_UNMOVABLE;
354 set_pageblock_flags_group(page, (unsigned long)migratetype,
355 PB_migrate, PB_migrate_end);
358 #ifdef CONFIG_DEBUG_VM
359 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
363 unsigned long pfn = page_to_pfn(page);
364 unsigned long sp, start_pfn;
367 seq = zone_span_seqbegin(zone);
368 start_pfn = zone->zone_start_pfn;
369 sp = zone->spanned_pages;
370 if (!zone_spans_pfn(zone, pfn))
372 } while (zone_span_seqretry(zone, seq));
375 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
376 pfn, zone_to_nid(zone), zone->name,
377 start_pfn, start_pfn + sp);
382 static int page_is_consistent(struct zone *zone, struct page *page)
384 if (!pfn_valid_within(page_to_pfn(page)))
386 if (zone != page_zone(page))
392 * Temporary debugging check for pages not lying within a given zone.
394 static int bad_range(struct zone *zone, struct page *page)
396 if (page_outside_zone_boundaries(zone, page))
398 if (!page_is_consistent(zone, page))
404 static inline int bad_range(struct zone *zone, struct page *page)
410 static void bad_page(struct page *page, const char *reason,
411 unsigned long bad_flags)
413 static unsigned long resume;
414 static unsigned long nr_shown;
415 static unsigned long nr_unshown;
417 /* Don't complain about poisoned pages */
418 if (PageHWPoison(page)) {
419 page_mapcount_reset(page); /* remove PageBuddy */
424 * Allow a burst of 60 reports, then keep quiet for that minute;
425 * or allow a steady drip of one report per second.
427 if (nr_shown == 60) {
428 if (time_before(jiffies, resume)) {
434 "BUG: Bad page state: %lu messages suppressed\n",
441 resume = jiffies + 60 * HZ;
443 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
444 current->comm, page_to_pfn(page));
445 dump_page_badflags(page, reason, bad_flags);
450 /* Leave bad fields for debug, except PageBuddy could make trouble */
451 page_mapcount_reset(page); /* remove PageBuddy */
452 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
456 * Higher-order pages are called "compound pages". They are structured thusly:
458 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
460 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
461 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
463 * The first tail page's ->compound_dtor holds the offset in array of compound
464 * page destructors. See compound_page_dtors.
466 * The first tail page's ->compound_order holds the order of allocation.
467 * This usage means that zero-order pages may not be compound.
470 void free_compound_page(struct page *page)
472 __free_pages_ok(page, compound_order(page));
475 void prep_compound_page(struct page *page, unsigned int order)
478 int nr_pages = 1 << order;
480 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
481 set_compound_order(page, order);
483 for (i = 1; i < nr_pages; i++) {
484 struct page *p = page + i;
485 set_page_count(p, 0);
486 p->mapping = TAIL_MAPPING;
487 set_compound_head(p, page);
489 atomic_set(compound_mapcount_ptr(page), -1);
492 #ifdef CONFIG_DEBUG_PAGEALLOC
493 unsigned int _debug_guardpage_minorder;
494 bool _debug_pagealloc_enabled __read_mostly
495 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
496 bool _debug_guardpage_enabled __read_mostly;
498 static int __init early_debug_pagealloc(char *buf)
503 if (strcmp(buf, "on") == 0)
504 _debug_pagealloc_enabled = true;
506 if (strcmp(buf, "off") == 0)
507 _debug_pagealloc_enabled = false;
511 early_param("debug_pagealloc", early_debug_pagealloc);
513 static bool need_debug_guardpage(void)
515 /* If we don't use debug_pagealloc, we don't need guard page */
516 if (!debug_pagealloc_enabled())
522 static void init_debug_guardpage(void)
524 if (!debug_pagealloc_enabled())
527 _debug_guardpage_enabled = true;
530 struct page_ext_operations debug_guardpage_ops = {
531 .need = need_debug_guardpage,
532 .init = init_debug_guardpage,
535 static int __init debug_guardpage_minorder_setup(char *buf)
539 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
540 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
543 _debug_guardpage_minorder = res;
544 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
547 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
549 static inline void set_page_guard(struct zone *zone, struct page *page,
550 unsigned int order, int migratetype)
552 struct page_ext *page_ext;
554 if (!debug_guardpage_enabled())
557 page_ext = lookup_page_ext(page);
558 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
560 INIT_LIST_HEAD(&page->lru);
561 set_page_private(page, order);
562 /* Guard pages are not available for any usage */
563 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
566 static inline void clear_page_guard(struct zone *zone, struct page *page,
567 unsigned int order, int migratetype)
569 struct page_ext *page_ext;
571 if (!debug_guardpage_enabled())
574 page_ext = lookup_page_ext(page);
575 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
577 set_page_private(page, 0);
578 if (!is_migrate_isolate(migratetype))
579 __mod_zone_freepage_state(zone, (1 << order), migratetype);
582 struct page_ext_operations debug_guardpage_ops = { NULL, };
583 static inline void set_page_guard(struct zone *zone, struct page *page,
584 unsigned int order, int migratetype) {}
585 static inline void clear_page_guard(struct zone *zone, struct page *page,
586 unsigned int order, int migratetype) {}
589 static inline void set_page_order(struct page *page, unsigned int order)
591 set_page_private(page, order);
592 __SetPageBuddy(page);
595 static inline void rmv_page_order(struct page *page)
597 __ClearPageBuddy(page);
598 set_page_private(page, 0);
602 * This function checks whether a page is free && is the buddy
603 * we can do coalesce a page and its buddy if
604 * (a) the buddy is not in a hole &&
605 * (b) the buddy is in the buddy system &&
606 * (c) a page and its buddy have the same order &&
607 * (d) a page and its buddy are in the same zone.
609 * For recording whether a page is in the buddy system, we set ->_mapcount
610 * PAGE_BUDDY_MAPCOUNT_VALUE.
611 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
612 * serialized by zone->lock.
614 * For recording page's order, we use page_private(page).
616 static inline int page_is_buddy(struct page *page, struct page *buddy,
619 if (!pfn_valid_within(page_to_pfn(buddy)))
622 if (page_is_guard(buddy) && page_order(buddy) == order) {
623 if (page_zone_id(page) != page_zone_id(buddy))
626 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
631 if (PageBuddy(buddy) && page_order(buddy) == order) {
633 * zone check is done late to avoid uselessly
634 * calculating zone/node ids for pages that could
637 if (page_zone_id(page) != page_zone_id(buddy))
640 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
648 * Freeing function for a buddy system allocator.
650 * The concept of a buddy system is to maintain direct-mapped table
651 * (containing bit values) for memory blocks of various "orders".
652 * The bottom level table contains the map for the smallest allocatable
653 * units of memory (here, pages), and each level above it describes
654 * pairs of units from the levels below, hence, "buddies".
655 * At a high level, all that happens here is marking the table entry
656 * at the bottom level available, and propagating the changes upward
657 * as necessary, plus some accounting needed to play nicely with other
658 * parts of the VM system.
659 * At each level, we keep a list of pages, which are heads of continuous
660 * free pages of length of (1 << order) and marked with _mapcount
661 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
663 * So when we are allocating or freeing one, we can derive the state of the
664 * other. That is, if we allocate a small block, and both were
665 * free, the remainder of the region must be split into blocks.
666 * If a block is freed, and its buddy is also free, then this
667 * triggers coalescing into a block of larger size.
672 static inline void __free_one_page(struct page *page,
674 struct zone *zone, unsigned int order,
677 unsigned long page_idx;
678 unsigned long combined_idx;
679 unsigned long uninitialized_var(buddy_idx);
681 unsigned int max_order = MAX_ORDER;
683 VM_BUG_ON(!zone_is_initialized(zone));
684 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
686 VM_BUG_ON(migratetype == -1);
687 if (is_migrate_isolate(migratetype)) {
689 * We restrict max order of merging to prevent merge
690 * between freepages on isolate pageblock and normal
691 * pageblock. Without this, pageblock isolation
692 * could cause incorrect freepage accounting.
694 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
696 __mod_zone_freepage_state(zone, 1 << order, migratetype);
699 page_idx = pfn & ((1 << max_order) - 1);
701 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
702 VM_BUG_ON_PAGE(bad_range(zone, page), page);
704 while (order < max_order - 1) {
705 buddy_idx = __find_buddy_index(page_idx, order);
706 buddy = page + (buddy_idx - page_idx);
707 if (!page_is_buddy(page, buddy, order))
710 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
711 * merge with it and move up one order.
713 if (page_is_guard(buddy)) {
714 clear_page_guard(zone, buddy, order, migratetype);
716 list_del(&buddy->lru);
717 zone->free_area[order].nr_free--;
718 rmv_page_order(buddy);
720 combined_idx = buddy_idx & page_idx;
721 page = page + (combined_idx - page_idx);
722 page_idx = combined_idx;
725 set_page_order(page, order);
728 * If this is not the largest possible page, check if the buddy
729 * of the next-highest order is free. If it is, it's possible
730 * that pages are being freed that will coalesce soon. In case,
731 * that is happening, add the free page to the tail of the list
732 * so it's less likely to be used soon and more likely to be merged
733 * as a higher order page
735 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
736 struct page *higher_page, *higher_buddy;
737 combined_idx = buddy_idx & page_idx;
738 higher_page = page + (combined_idx - page_idx);
739 buddy_idx = __find_buddy_index(combined_idx, order + 1);
740 higher_buddy = higher_page + (buddy_idx - combined_idx);
741 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
742 list_add_tail(&page->lru,
743 &zone->free_area[order].free_list[migratetype]);
748 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
750 zone->free_area[order].nr_free++;
753 static inline int free_pages_check(struct page *page)
755 const char *bad_reason = NULL;
756 unsigned long bad_flags = 0;
758 if (unlikely(atomic_read(&page->_mapcount) != -1))
759 bad_reason = "nonzero mapcount";
760 if (unlikely(page->mapping != NULL))
761 bad_reason = "non-NULL mapping";
762 if (unlikely(atomic_read(&page->_count) != 0))
763 bad_reason = "nonzero _count";
764 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
765 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
766 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
769 if (unlikely(page->mem_cgroup))
770 bad_reason = "page still charged to cgroup";
772 if (unlikely(bad_reason)) {
773 bad_page(page, bad_reason, bad_flags);
776 page_cpupid_reset_last(page);
777 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
778 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
783 * Frees a number of pages from the PCP lists
784 * Assumes all pages on list are in same zone, and of same order.
785 * count is the number of pages to free.
787 * If the zone was previously in an "all pages pinned" state then look to
788 * see if this freeing clears that state.
790 * And clear the zone's pages_scanned counter, to hold off the "all pages are
791 * pinned" detection logic.
793 static void free_pcppages_bulk(struct zone *zone, int count,
794 struct per_cpu_pages *pcp)
799 unsigned long nr_scanned;
801 spin_lock(&zone->lock);
802 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
804 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
808 struct list_head *list;
811 * Remove pages from lists in a round-robin fashion. A
812 * batch_free count is maintained that is incremented when an
813 * empty list is encountered. This is so more pages are freed
814 * off fuller lists instead of spinning excessively around empty
819 if (++migratetype == MIGRATE_PCPTYPES)
821 list = &pcp->lists[migratetype];
822 } while (list_empty(list));
824 /* This is the only non-empty list. Free them all. */
825 if (batch_free == MIGRATE_PCPTYPES)
826 batch_free = to_free;
829 int mt; /* migratetype of the to-be-freed page */
831 page = list_last_entry(list, struct page, lru);
832 /* must delete as __free_one_page list manipulates */
833 list_del(&page->lru);
835 mt = get_pcppage_migratetype(page);
836 /* MIGRATE_ISOLATE page should not go to pcplists */
837 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
838 /* Pageblock could have been isolated meanwhile */
839 if (unlikely(has_isolate_pageblock(zone)))
840 mt = get_pageblock_migratetype(page);
842 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
843 trace_mm_page_pcpu_drain(page, 0, mt);
844 } while (--to_free && --batch_free && !list_empty(list));
846 spin_unlock(&zone->lock);
849 static void free_one_page(struct zone *zone,
850 struct page *page, unsigned long pfn,
854 unsigned long nr_scanned;
855 spin_lock(&zone->lock);
856 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
858 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
860 if (unlikely(has_isolate_pageblock(zone) ||
861 is_migrate_isolate(migratetype))) {
862 migratetype = get_pfnblock_migratetype(page, pfn);
864 __free_one_page(page, pfn, zone, order, migratetype);
865 spin_unlock(&zone->lock);
868 static int free_tail_pages_check(struct page *head_page, struct page *page)
873 * We rely page->lru.next never has bit 0 set, unless the page
874 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
876 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
878 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
882 switch (page - head_page) {
884 /* the first tail page: ->mapping is compound_mapcount() */
885 if (unlikely(compound_mapcount(page))) {
886 bad_page(page, "nonzero compound_mapcount", 0);
892 * the second tail page: ->mapping is
893 * page_deferred_list().next -- ignore value.
897 if (page->mapping != TAIL_MAPPING) {
898 bad_page(page, "corrupted mapping in tail page", 0);
903 if (unlikely(!PageTail(page))) {
904 bad_page(page, "PageTail not set", 0);
907 if (unlikely(compound_head(page) != head_page)) {
908 bad_page(page, "compound_head not consistent", 0);
913 page->mapping = NULL;
914 clear_compound_head(page);
918 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
919 unsigned long zone, int nid)
921 set_page_links(page, zone, nid, pfn);
922 init_page_count(page);
923 page_mapcount_reset(page);
924 page_cpupid_reset_last(page);
926 INIT_LIST_HEAD(&page->lru);
927 #ifdef WANT_PAGE_VIRTUAL
928 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
929 if (!is_highmem_idx(zone))
930 set_page_address(page, __va(pfn << PAGE_SHIFT));
934 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
937 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
940 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
941 static void init_reserved_page(unsigned long pfn)
946 if (!early_page_uninitialised(pfn))
949 nid = early_pfn_to_nid(pfn);
950 pgdat = NODE_DATA(nid);
952 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
953 struct zone *zone = &pgdat->node_zones[zid];
955 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
958 __init_single_pfn(pfn, zid, nid);
961 static inline void init_reserved_page(unsigned long pfn)
964 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
967 * Initialised pages do not have PageReserved set. This function is
968 * called for each range allocated by the bootmem allocator and
969 * marks the pages PageReserved. The remaining valid pages are later
970 * sent to the buddy page allocator.
972 void __meminit reserve_bootmem_region(unsigned long start, unsigned long end)
974 unsigned long start_pfn = PFN_DOWN(start);
975 unsigned long end_pfn = PFN_UP(end);
977 for (; start_pfn < end_pfn; start_pfn++) {
978 if (pfn_valid(start_pfn)) {
979 struct page *page = pfn_to_page(start_pfn);
981 init_reserved_page(start_pfn);
983 /* Avoid false-positive PageTail() */
984 INIT_LIST_HEAD(&page->lru);
986 SetPageReserved(page);
991 static bool free_pages_prepare(struct page *page, unsigned int order)
993 bool compound = PageCompound(page);
996 VM_BUG_ON_PAGE(PageTail(page), page);
997 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
999 trace_mm_page_free(page, order);
1000 kmemcheck_free_shadow(page, order);
1001 kasan_free_pages(page, order);
1004 page->mapping = NULL;
1005 bad += free_pages_check(page);
1006 for (i = 1; i < (1 << order); i++) {
1008 bad += free_tail_pages_check(page, page + i);
1009 bad += free_pages_check(page + i);
1014 reset_page_owner(page, order);
1016 if (!PageHighMem(page)) {
1017 debug_check_no_locks_freed(page_address(page),
1018 PAGE_SIZE << order);
1019 debug_check_no_obj_freed(page_address(page),
1020 PAGE_SIZE << order);
1022 arch_free_page(page, order);
1023 kernel_map_pages(page, 1 << order, 0);
1028 static void __free_pages_ok(struct page *page, unsigned int order)
1030 unsigned long flags;
1032 unsigned long pfn = page_to_pfn(page);
1034 if (!free_pages_prepare(page, order))
1037 migratetype = get_pfnblock_migratetype(page, pfn);
1038 local_irq_save(flags);
1039 __count_vm_events(PGFREE, 1 << order);
1040 free_one_page(page_zone(page), page, pfn, order, migratetype);
1041 local_irq_restore(flags);
1044 static void __init __free_pages_boot_core(struct page *page,
1045 unsigned long pfn, unsigned int order)
1047 unsigned int nr_pages = 1 << order;
1048 struct page *p = page;
1052 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1054 __ClearPageReserved(p);
1055 set_page_count(p, 0);
1057 __ClearPageReserved(p);
1058 set_page_count(p, 0);
1060 page_zone(page)->managed_pages += nr_pages;
1061 set_page_refcounted(page);
1062 __free_pages(page, order);
1065 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1066 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1068 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1070 int __meminit early_pfn_to_nid(unsigned long pfn)
1072 static DEFINE_SPINLOCK(early_pfn_lock);
1075 spin_lock(&early_pfn_lock);
1076 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1079 spin_unlock(&early_pfn_lock);
1085 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1086 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1087 struct mminit_pfnnid_cache *state)
1091 nid = __early_pfn_to_nid(pfn, state);
1092 if (nid >= 0 && nid != node)
1097 /* Only safe to use early in boot when initialisation is single-threaded */
1098 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1100 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1105 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1109 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1110 struct mminit_pfnnid_cache *state)
1117 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1120 if (early_page_uninitialised(pfn))
1122 return __free_pages_boot_core(page, pfn, order);
1125 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1126 static void __init deferred_free_range(struct page *page,
1127 unsigned long pfn, int nr_pages)
1134 /* Free a large naturally-aligned chunk if possible */
1135 if (nr_pages == MAX_ORDER_NR_PAGES &&
1136 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1137 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1138 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1142 for (i = 0; i < nr_pages; i++, page++, pfn++)
1143 __free_pages_boot_core(page, pfn, 0);
1146 /* Completion tracking for deferred_init_memmap() threads */
1147 static atomic_t pgdat_init_n_undone __initdata;
1148 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1150 static inline void __init pgdat_init_report_one_done(void)
1152 if (atomic_dec_and_test(&pgdat_init_n_undone))
1153 complete(&pgdat_init_all_done_comp);
1156 /* Initialise remaining memory on a node */
1157 static int __init deferred_init_memmap(void *data)
1159 pg_data_t *pgdat = data;
1160 int nid = pgdat->node_id;
1161 struct mminit_pfnnid_cache nid_init_state = { };
1162 unsigned long start = jiffies;
1163 unsigned long nr_pages = 0;
1164 unsigned long walk_start, walk_end;
1167 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1168 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1170 if (first_init_pfn == ULONG_MAX) {
1171 pgdat_init_report_one_done();
1175 /* Bind memory initialisation thread to a local node if possible */
1176 if (!cpumask_empty(cpumask))
1177 set_cpus_allowed_ptr(current, cpumask);
1179 /* Sanity check boundaries */
1180 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1181 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1182 pgdat->first_deferred_pfn = ULONG_MAX;
1184 /* Only the highest zone is deferred so find it */
1185 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1186 zone = pgdat->node_zones + zid;
1187 if (first_init_pfn < zone_end_pfn(zone))
1191 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1192 unsigned long pfn, end_pfn;
1193 struct page *page = NULL;
1194 struct page *free_base_page = NULL;
1195 unsigned long free_base_pfn = 0;
1198 end_pfn = min(walk_end, zone_end_pfn(zone));
1199 pfn = first_init_pfn;
1200 if (pfn < walk_start)
1202 if (pfn < zone->zone_start_pfn)
1203 pfn = zone->zone_start_pfn;
1205 for (; pfn < end_pfn; pfn++) {
1206 if (!pfn_valid_within(pfn))
1210 * Ensure pfn_valid is checked every
1211 * MAX_ORDER_NR_PAGES for memory holes
1213 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1214 if (!pfn_valid(pfn)) {
1220 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1225 /* Minimise pfn page lookups and scheduler checks */
1226 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1229 nr_pages += nr_to_free;
1230 deferred_free_range(free_base_page,
1231 free_base_pfn, nr_to_free);
1232 free_base_page = NULL;
1233 free_base_pfn = nr_to_free = 0;
1235 page = pfn_to_page(pfn);
1240 VM_BUG_ON(page_zone(page) != zone);
1244 __init_single_page(page, pfn, zid, nid);
1245 if (!free_base_page) {
1246 free_base_page = page;
1247 free_base_pfn = pfn;
1252 /* Where possible, batch up pages for a single free */
1255 /* Free the current block of pages to allocator */
1256 nr_pages += nr_to_free;
1257 deferred_free_range(free_base_page, free_base_pfn,
1259 free_base_page = NULL;
1260 free_base_pfn = nr_to_free = 0;
1263 first_init_pfn = max(end_pfn, first_init_pfn);
1266 /* Sanity check that the next zone really is unpopulated */
1267 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1269 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1270 jiffies_to_msecs(jiffies - start));
1272 pgdat_init_report_one_done();
1276 void __init page_alloc_init_late(void)
1280 /* There will be num_node_state(N_MEMORY) threads */
1281 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1282 for_each_node_state(nid, N_MEMORY) {
1283 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1286 /* Block until all are initialised */
1287 wait_for_completion(&pgdat_init_all_done_comp);
1289 /* Reinit limits that are based on free pages after the kernel is up */
1290 files_maxfiles_init();
1292 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1295 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1296 void __init init_cma_reserved_pageblock(struct page *page)
1298 unsigned i = pageblock_nr_pages;
1299 struct page *p = page;
1302 __ClearPageReserved(p);
1303 set_page_count(p, 0);
1306 set_pageblock_migratetype(page, MIGRATE_CMA);
1308 if (pageblock_order >= MAX_ORDER) {
1309 i = pageblock_nr_pages;
1312 set_page_refcounted(p);
1313 __free_pages(p, MAX_ORDER - 1);
1314 p += MAX_ORDER_NR_PAGES;
1315 } while (i -= MAX_ORDER_NR_PAGES);
1317 set_page_refcounted(page);
1318 __free_pages(page, pageblock_order);
1321 adjust_managed_page_count(page, pageblock_nr_pages);
1326 * The order of subdivision here is critical for the IO subsystem.
1327 * Please do not alter this order without good reasons and regression
1328 * testing. Specifically, as large blocks of memory are subdivided,
1329 * the order in which smaller blocks are delivered depends on the order
1330 * they're subdivided in this function. This is the primary factor
1331 * influencing the order in which pages are delivered to the IO
1332 * subsystem according to empirical testing, and this is also justified
1333 * by considering the behavior of a buddy system containing a single
1334 * large block of memory acted on by a series of small allocations.
1335 * This behavior is a critical factor in sglist merging's success.
1339 static inline void expand(struct zone *zone, struct page *page,
1340 int low, int high, struct free_area *area,
1343 unsigned long size = 1 << high;
1345 while (high > low) {
1349 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1351 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1352 debug_guardpage_enabled() &&
1353 high < debug_guardpage_minorder()) {
1355 * Mark as guard pages (or page), that will allow to
1356 * merge back to allocator when buddy will be freed.
1357 * Corresponding page table entries will not be touched,
1358 * pages will stay not present in virtual address space
1360 set_page_guard(zone, &page[size], high, migratetype);
1363 list_add(&page[size].lru, &area->free_list[migratetype]);
1365 set_page_order(&page[size], high);
1370 * This page is about to be returned from the page allocator
1372 static inline int check_new_page(struct page *page)
1374 const char *bad_reason = NULL;
1375 unsigned long bad_flags = 0;
1377 if (unlikely(atomic_read(&page->_mapcount) != -1))
1378 bad_reason = "nonzero mapcount";
1379 if (unlikely(page->mapping != NULL))
1380 bad_reason = "non-NULL mapping";
1381 if (unlikely(atomic_read(&page->_count) != 0))
1382 bad_reason = "nonzero _count";
1383 if (unlikely(page->flags & __PG_HWPOISON)) {
1384 bad_reason = "HWPoisoned (hardware-corrupted)";
1385 bad_flags = __PG_HWPOISON;
1387 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1388 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1389 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1392 if (unlikely(page->mem_cgroup))
1393 bad_reason = "page still charged to cgroup";
1395 if (unlikely(bad_reason)) {
1396 bad_page(page, bad_reason, bad_flags);
1402 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1407 for (i = 0; i < (1 << order); i++) {
1408 struct page *p = page + i;
1409 if (unlikely(check_new_page(p)))
1413 set_page_private(page, 0);
1414 set_page_refcounted(page);
1416 arch_alloc_page(page, order);
1417 kernel_map_pages(page, 1 << order, 1);
1418 kasan_alloc_pages(page, order);
1420 if (gfp_flags & __GFP_ZERO)
1421 for (i = 0; i < (1 << order); i++)
1422 clear_highpage(page + i);
1424 if (order && (gfp_flags & __GFP_COMP))
1425 prep_compound_page(page, order);
1427 set_page_owner(page, order, gfp_flags);
1430 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1431 * allocate the page. The expectation is that the caller is taking
1432 * steps that will free more memory. The caller should avoid the page
1433 * being used for !PFMEMALLOC purposes.
1435 if (alloc_flags & ALLOC_NO_WATERMARKS)
1436 set_page_pfmemalloc(page);
1438 clear_page_pfmemalloc(page);
1444 * Go through the free lists for the given migratetype and remove
1445 * the smallest available page from the freelists
1448 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1451 unsigned int current_order;
1452 struct free_area *area;
1455 /* Find a page of the appropriate size in the preferred list */
1456 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1457 area = &(zone->free_area[current_order]);
1458 page = list_first_entry_or_null(&area->free_list[migratetype],
1462 list_del(&page->lru);
1463 rmv_page_order(page);
1465 expand(zone, page, order, current_order, area, migratetype);
1466 set_pcppage_migratetype(page, migratetype);
1475 * This array describes the order lists are fallen back to when
1476 * the free lists for the desirable migrate type are depleted
1478 static int fallbacks[MIGRATE_TYPES][4] = {
1479 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1480 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1481 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1483 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1485 #ifdef CONFIG_MEMORY_ISOLATION
1486 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1491 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1494 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1497 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1498 unsigned int order) { return NULL; }
1502 * Move the free pages in a range to the free lists of the requested type.
1503 * Note that start_page and end_pages are not aligned on a pageblock
1504 * boundary. If alignment is required, use move_freepages_block()
1506 int move_freepages(struct zone *zone,
1507 struct page *start_page, struct page *end_page,
1512 int pages_moved = 0;
1514 #ifndef CONFIG_HOLES_IN_ZONE
1516 * page_zone is not safe to call in this context when
1517 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1518 * anyway as we check zone boundaries in move_freepages_block().
1519 * Remove at a later date when no bug reports exist related to
1520 * grouping pages by mobility
1522 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1525 for (page = start_page; page <= end_page;) {
1526 /* Make sure we are not inadvertently changing nodes */
1527 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1529 if (!pfn_valid_within(page_to_pfn(page))) {
1534 if (!PageBuddy(page)) {
1539 order = page_order(page);
1540 list_move(&page->lru,
1541 &zone->free_area[order].free_list[migratetype]);
1543 pages_moved += 1 << order;
1549 int move_freepages_block(struct zone *zone, struct page *page,
1552 unsigned long start_pfn, end_pfn;
1553 struct page *start_page, *end_page;
1555 start_pfn = page_to_pfn(page);
1556 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1557 start_page = pfn_to_page(start_pfn);
1558 end_page = start_page + pageblock_nr_pages - 1;
1559 end_pfn = start_pfn + pageblock_nr_pages - 1;
1561 /* Do not cross zone boundaries */
1562 if (!zone_spans_pfn(zone, start_pfn))
1564 if (!zone_spans_pfn(zone, end_pfn))
1567 return move_freepages(zone, start_page, end_page, migratetype);
1570 static void change_pageblock_range(struct page *pageblock_page,
1571 int start_order, int migratetype)
1573 int nr_pageblocks = 1 << (start_order - pageblock_order);
1575 while (nr_pageblocks--) {
1576 set_pageblock_migratetype(pageblock_page, migratetype);
1577 pageblock_page += pageblock_nr_pages;
1582 * When we are falling back to another migratetype during allocation, try to
1583 * steal extra free pages from the same pageblocks to satisfy further
1584 * allocations, instead of polluting multiple pageblocks.
1586 * If we are stealing a relatively large buddy page, it is likely there will
1587 * be more free pages in the pageblock, so try to steal them all. For
1588 * reclaimable and unmovable allocations, we steal regardless of page size,
1589 * as fragmentation caused by those allocations polluting movable pageblocks
1590 * is worse than movable allocations stealing from unmovable and reclaimable
1593 static bool can_steal_fallback(unsigned int order, int start_mt)
1596 * Leaving this order check is intended, although there is
1597 * relaxed order check in next check. The reason is that
1598 * we can actually steal whole pageblock if this condition met,
1599 * but, below check doesn't guarantee it and that is just heuristic
1600 * so could be changed anytime.
1602 if (order >= pageblock_order)
1605 if (order >= pageblock_order / 2 ||
1606 start_mt == MIGRATE_RECLAIMABLE ||
1607 start_mt == MIGRATE_UNMOVABLE ||
1608 page_group_by_mobility_disabled)
1615 * This function implements actual steal behaviour. If order is large enough,
1616 * we can steal whole pageblock. If not, we first move freepages in this
1617 * pageblock and check whether half of pages are moved or not. If half of
1618 * pages are moved, we can change migratetype of pageblock and permanently
1619 * use it's pages as requested migratetype in the future.
1621 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1624 unsigned int current_order = page_order(page);
1627 /* Take ownership for orders >= pageblock_order */
1628 if (current_order >= pageblock_order) {
1629 change_pageblock_range(page, current_order, start_type);
1633 pages = move_freepages_block(zone, page, start_type);
1635 /* Claim the whole block if over half of it is free */
1636 if (pages >= (1 << (pageblock_order-1)) ||
1637 page_group_by_mobility_disabled)
1638 set_pageblock_migratetype(page, start_type);
1642 * Check whether there is a suitable fallback freepage with requested order.
1643 * If only_stealable is true, this function returns fallback_mt only if
1644 * we can steal other freepages all together. This would help to reduce
1645 * fragmentation due to mixed migratetype pages in one pageblock.
1647 int find_suitable_fallback(struct free_area *area, unsigned int order,
1648 int migratetype, bool only_stealable, bool *can_steal)
1653 if (area->nr_free == 0)
1658 fallback_mt = fallbacks[migratetype][i];
1659 if (fallback_mt == MIGRATE_TYPES)
1662 if (list_empty(&area->free_list[fallback_mt]))
1665 if (can_steal_fallback(order, migratetype))
1668 if (!only_stealable)
1679 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1680 * there are no empty page blocks that contain a page with a suitable order
1682 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1683 unsigned int alloc_order)
1686 unsigned long max_managed, flags;
1689 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1690 * Check is race-prone but harmless.
1692 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1693 if (zone->nr_reserved_highatomic >= max_managed)
1696 spin_lock_irqsave(&zone->lock, flags);
1698 /* Recheck the nr_reserved_highatomic limit under the lock */
1699 if (zone->nr_reserved_highatomic >= max_managed)
1703 mt = get_pageblock_migratetype(page);
1704 if (mt != MIGRATE_HIGHATOMIC &&
1705 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1706 zone->nr_reserved_highatomic += pageblock_nr_pages;
1707 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1708 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1712 spin_unlock_irqrestore(&zone->lock, flags);
1716 * Used when an allocation is about to fail under memory pressure. This
1717 * potentially hurts the reliability of high-order allocations when under
1718 * intense memory pressure but failed atomic allocations should be easier
1719 * to recover from than an OOM.
1721 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1723 struct zonelist *zonelist = ac->zonelist;
1724 unsigned long flags;
1730 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1732 /* Preserve at least one pageblock */
1733 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1736 spin_lock_irqsave(&zone->lock, flags);
1737 for (order = 0; order < MAX_ORDER; order++) {
1738 struct free_area *area = &(zone->free_area[order]);
1740 page = list_first_entry_or_null(
1741 &area->free_list[MIGRATE_HIGHATOMIC],
1747 * It should never happen but changes to locking could
1748 * inadvertently allow a per-cpu drain to add pages
1749 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1750 * and watch for underflows.
1752 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
1753 zone->nr_reserved_highatomic);
1756 * Convert to ac->migratetype and avoid the normal
1757 * pageblock stealing heuristics. Minimally, the caller
1758 * is doing the work and needs the pages. More
1759 * importantly, if the block was always converted to
1760 * MIGRATE_UNMOVABLE or another type then the number
1761 * of pageblocks that cannot be completely freed
1764 set_pageblock_migratetype(page, ac->migratetype);
1765 move_freepages_block(zone, page, ac->migratetype);
1766 spin_unlock_irqrestore(&zone->lock, flags);
1769 spin_unlock_irqrestore(&zone->lock, flags);
1773 /* Remove an element from the buddy allocator from the fallback list */
1774 static inline struct page *
1775 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1777 struct free_area *area;
1778 unsigned int current_order;
1783 /* Find the largest possible block of pages in the other list */
1784 for (current_order = MAX_ORDER-1;
1785 current_order >= order && current_order <= MAX_ORDER-1;
1787 area = &(zone->free_area[current_order]);
1788 fallback_mt = find_suitable_fallback(area, current_order,
1789 start_migratetype, false, &can_steal);
1790 if (fallback_mt == -1)
1793 page = list_first_entry(&area->free_list[fallback_mt],
1796 steal_suitable_fallback(zone, page, start_migratetype);
1798 /* Remove the page from the freelists */
1800 list_del(&page->lru);
1801 rmv_page_order(page);
1803 expand(zone, page, order, current_order, area,
1806 * The pcppage_migratetype may differ from pageblock's
1807 * migratetype depending on the decisions in
1808 * find_suitable_fallback(). This is OK as long as it does not
1809 * differ for MIGRATE_CMA pageblocks. Those can be used as
1810 * fallback only via special __rmqueue_cma_fallback() function
1812 set_pcppage_migratetype(page, start_migratetype);
1814 trace_mm_page_alloc_extfrag(page, order, current_order,
1815 start_migratetype, fallback_mt);
1824 * Do the hard work of removing an element from the buddy allocator.
1825 * Call me with the zone->lock already held.
1827 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1832 page = __rmqueue_smallest(zone, order, migratetype);
1833 if (unlikely(!page)) {
1834 if (migratetype == MIGRATE_MOVABLE)
1835 page = __rmqueue_cma_fallback(zone, order);
1838 page = __rmqueue_fallback(zone, order, migratetype);
1841 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1846 * Obtain a specified number of elements from the buddy allocator, all under
1847 * a single hold of the lock, for efficiency. Add them to the supplied list.
1848 * Returns the number of new pages which were placed at *list.
1850 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1851 unsigned long count, struct list_head *list,
1852 int migratetype, bool cold)
1856 spin_lock(&zone->lock);
1857 for (i = 0; i < count; ++i) {
1858 struct page *page = __rmqueue(zone, order, migratetype);
1859 if (unlikely(page == NULL))
1863 * Split buddy pages returned by expand() are received here
1864 * in physical page order. The page is added to the callers and
1865 * list and the list head then moves forward. From the callers
1866 * perspective, the linked list is ordered by page number in
1867 * some conditions. This is useful for IO devices that can
1868 * merge IO requests if the physical pages are ordered
1872 list_add(&page->lru, list);
1874 list_add_tail(&page->lru, list);
1876 if (is_migrate_cma(get_pcppage_migratetype(page)))
1877 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1880 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1881 spin_unlock(&zone->lock);
1887 * Called from the vmstat counter updater to drain pagesets of this
1888 * currently executing processor on remote nodes after they have
1891 * Note that this function must be called with the thread pinned to
1892 * a single processor.
1894 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1896 unsigned long flags;
1897 int to_drain, batch;
1899 local_irq_save(flags);
1900 batch = READ_ONCE(pcp->batch);
1901 to_drain = min(pcp->count, batch);
1903 free_pcppages_bulk(zone, to_drain, pcp);
1904 pcp->count -= to_drain;
1906 local_irq_restore(flags);
1911 * Drain pcplists of the indicated processor and zone.
1913 * The processor must either be the current processor and the
1914 * thread pinned to the current processor or a processor that
1917 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1919 unsigned long flags;
1920 struct per_cpu_pageset *pset;
1921 struct per_cpu_pages *pcp;
1923 local_irq_save(flags);
1924 pset = per_cpu_ptr(zone->pageset, cpu);
1928 free_pcppages_bulk(zone, pcp->count, pcp);
1931 local_irq_restore(flags);
1935 * Drain pcplists of all zones on the indicated processor.
1937 * The processor must either be the current processor and the
1938 * thread pinned to the current processor or a processor that
1941 static void drain_pages(unsigned int cpu)
1945 for_each_populated_zone(zone) {
1946 drain_pages_zone(cpu, zone);
1951 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1953 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1954 * the single zone's pages.
1956 void drain_local_pages(struct zone *zone)
1958 int cpu = smp_processor_id();
1961 drain_pages_zone(cpu, zone);
1967 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1969 * When zone parameter is non-NULL, spill just the single zone's pages.
1971 * Note that this code is protected against sending an IPI to an offline
1972 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1973 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1974 * nothing keeps CPUs from showing up after we populated the cpumask and
1975 * before the call to on_each_cpu_mask().
1977 void drain_all_pages(struct zone *zone)
1982 * Allocate in the BSS so we wont require allocation in
1983 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1985 static cpumask_t cpus_with_pcps;
1988 * We don't care about racing with CPU hotplug event
1989 * as offline notification will cause the notified
1990 * cpu to drain that CPU pcps and on_each_cpu_mask
1991 * disables preemption as part of its processing
1993 for_each_online_cpu(cpu) {
1994 struct per_cpu_pageset *pcp;
1996 bool has_pcps = false;
1999 pcp = per_cpu_ptr(zone->pageset, cpu);
2003 for_each_populated_zone(z) {
2004 pcp = per_cpu_ptr(z->pageset, cpu);
2005 if (pcp->pcp.count) {
2013 cpumask_set_cpu(cpu, &cpus_with_pcps);
2015 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2017 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2021 #ifdef CONFIG_HIBERNATION
2023 void mark_free_pages(struct zone *zone)
2025 unsigned long pfn, max_zone_pfn;
2026 unsigned long flags;
2027 unsigned int order, t;
2030 if (zone_is_empty(zone))
2033 spin_lock_irqsave(&zone->lock, flags);
2035 max_zone_pfn = zone_end_pfn(zone);
2036 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2037 if (pfn_valid(pfn)) {
2038 page = pfn_to_page(pfn);
2039 if (!swsusp_page_is_forbidden(page))
2040 swsusp_unset_page_free(page);
2043 for_each_migratetype_order(order, t) {
2044 list_for_each_entry(page,
2045 &zone->free_area[order].free_list[t], lru) {
2048 pfn = page_to_pfn(page);
2049 for (i = 0; i < (1UL << order); i++)
2050 swsusp_set_page_free(pfn_to_page(pfn + i));
2053 spin_unlock_irqrestore(&zone->lock, flags);
2055 #endif /* CONFIG_PM */
2058 * Free a 0-order page
2059 * cold == true ? free a cold page : free a hot page
2061 void free_hot_cold_page(struct page *page, bool cold)
2063 struct zone *zone = page_zone(page);
2064 struct per_cpu_pages *pcp;
2065 unsigned long flags;
2066 unsigned long pfn = page_to_pfn(page);
2069 if (!free_pages_prepare(page, 0))
2072 migratetype = get_pfnblock_migratetype(page, pfn);
2073 set_pcppage_migratetype(page, migratetype);
2074 local_irq_save(flags);
2075 __count_vm_event(PGFREE);
2078 * We only track unmovable, reclaimable and movable on pcp lists.
2079 * Free ISOLATE pages back to the allocator because they are being
2080 * offlined but treat RESERVE as movable pages so we can get those
2081 * areas back if necessary. Otherwise, we may have to free
2082 * excessively into the page allocator
2084 if (migratetype >= MIGRATE_PCPTYPES) {
2085 if (unlikely(is_migrate_isolate(migratetype))) {
2086 free_one_page(zone, page, pfn, 0, migratetype);
2089 migratetype = MIGRATE_MOVABLE;
2092 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2094 list_add(&page->lru, &pcp->lists[migratetype]);
2096 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2098 if (pcp->count >= pcp->high) {
2099 unsigned long batch = READ_ONCE(pcp->batch);
2100 free_pcppages_bulk(zone, batch, pcp);
2101 pcp->count -= batch;
2105 local_irq_restore(flags);
2109 * Free a list of 0-order pages
2111 void free_hot_cold_page_list(struct list_head *list, bool cold)
2113 struct page *page, *next;
2115 list_for_each_entry_safe(page, next, list, lru) {
2116 trace_mm_page_free_batched(page, cold);
2117 free_hot_cold_page(page, cold);
2122 * split_page takes a non-compound higher-order page, and splits it into
2123 * n (1<<order) sub-pages: page[0..n]
2124 * Each sub-page must be freed individually.
2126 * Note: this is probably too low level an operation for use in drivers.
2127 * Please consult with lkml before using this in your driver.
2129 void split_page(struct page *page, unsigned int order)
2134 VM_BUG_ON_PAGE(PageCompound(page), page);
2135 VM_BUG_ON_PAGE(!page_count(page), page);
2137 #ifdef CONFIG_KMEMCHECK
2139 * Split shadow pages too, because free(page[0]) would
2140 * otherwise free the whole shadow.
2142 if (kmemcheck_page_is_tracked(page))
2143 split_page(virt_to_page(page[0].shadow), order);
2146 gfp_mask = get_page_owner_gfp(page);
2147 set_page_owner(page, 0, gfp_mask);
2148 for (i = 1; i < (1 << order); i++) {
2149 set_page_refcounted(page + i);
2150 set_page_owner(page + i, 0, gfp_mask);
2153 EXPORT_SYMBOL_GPL(split_page);
2155 int __isolate_free_page(struct page *page, unsigned int order)
2157 unsigned long watermark;
2161 BUG_ON(!PageBuddy(page));
2163 zone = page_zone(page);
2164 mt = get_pageblock_migratetype(page);
2166 if (!is_migrate_isolate(mt)) {
2167 /* Obey watermarks as if the page was being allocated */
2168 watermark = low_wmark_pages(zone) + (1 << order);
2169 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2172 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2175 /* Remove page from free list */
2176 list_del(&page->lru);
2177 zone->free_area[order].nr_free--;
2178 rmv_page_order(page);
2180 set_page_owner(page, order, __GFP_MOVABLE);
2182 /* Set the pageblock if the isolated page is at least a pageblock */
2183 if (order >= pageblock_order - 1) {
2184 struct page *endpage = page + (1 << order) - 1;
2185 for (; page < endpage; page += pageblock_nr_pages) {
2186 int mt = get_pageblock_migratetype(page);
2187 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2188 set_pageblock_migratetype(page,
2194 return 1UL << order;
2198 * Similar to split_page except the page is already free. As this is only
2199 * being used for migration, the migratetype of the block also changes.
2200 * As this is called with interrupts disabled, the caller is responsible
2201 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2204 * Note: this is probably too low level an operation for use in drivers.
2205 * Please consult with lkml before using this in your driver.
2207 int split_free_page(struct page *page)
2212 order = page_order(page);
2214 nr_pages = __isolate_free_page(page, order);
2218 /* Split into individual pages */
2219 set_page_refcounted(page);
2220 split_page(page, order);
2225 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2228 struct page *buffered_rmqueue(struct zone *preferred_zone,
2229 struct zone *zone, unsigned int order,
2230 gfp_t gfp_flags, int alloc_flags, int migratetype)
2232 unsigned long flags;
2234 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2236 if (likely(order == 0)) {
2237 struct per_cpu_pages *pcp;
2238 struct list_head *list;
2240 local_irq_save(flags);
2241 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2242 list = &pcp->lists[migratetype];
2243 if (list_empty(list)) {
2244 pcp->count += rmqueue_bulk(zone, 0,
2247 if (unlikely(list_empty(list)))
2252 page = list_last_entry(list, struct page, lru);
2254 page = list_first_entry(list, struct page, lru);
2256 list_del(&page->lru);
2259 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2261 * __GFP_NOFAIL is not to be used in new code.
2263 * All __GFP_NOFAIL callers should be fixed so that they
2264 * properly detect and handle allocation failures.
2266 * We most definitely don't want callers attempting to
2267 * allocate greater than order-1 page units with
2270 WARN_ON_ONCE(order > 1);
2272 spin_lock_irqsave(&zone->lock, flags);
2275 if (alloc_flags & ALLOC_HARDER) {
2276 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2278 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2281 page = __rmqueue(zone, order, migratetype);
2282 spin_unlock(&zone->lock);
2285 __mod_zone_freepage_state(zone, -(1 << order),
2286 get_pcppage_migratetype(page));
2289 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2290 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2291 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2292 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2294 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2295 zone_statistics(preferred_zone, zone, gfp_flags);
2296 local_irq_restore(flags);
2298 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2302 local_irq_restore(flags);
2306 #ifdef CONFIG_FAIL_PAGE_ALLOC
2309 struct fault_attr attr;
2311 bool ignore_gfp_highmem;
2312 bool ignore_gfp_reclaim;
2314 } fail_page_alloc = {
2315 .attr = FAULT_ATTR_INITIALIZER,
2316 .ignore_gfp_reclaim = true,
2317 .ignore_gfp_highmem = true,
2321 static int __init setup_fail_page_alloc(char *str)
2323 return setup_fault_attr(&fail_page_alloc.attr, str);
2325 __setup("fail_page_alloc=", setup_fail_page_alloc);
2327 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2329 if (order < fail_page_alloc.min_order)
2331 if (gfp_mask & __GFP_NOFAIL)
2333 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2335 if (fail_page_alloc.ignore_gfp_reclaim &&
2336 (gfp_mask & __GFP_DIRECT_RECLAIM))
2339 return should_fail(&fail_page_alloc.attr, 1 << order);
2342 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2344 static int __init fail_page_alloc_debugfs(void)
2346 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2349 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2350 &fail_page_alloc.attr);
2352 return PTR_ERR(dir);
2354 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2355 &fail_page_alloc.ignore_gfp_reclaim))
2357 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2358 &fail_page_alloc.ignore_gfp_highmem))
2360 if (!debugfs_create_u32("min-order", mode, dir,
2361 &fail_page_alloc.min_order))
2366 debugfs_remove_recursive(dir);
2371 late_initcall(fail_page_alloc_debugfs);
2373 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2375 #else /* CONFIG_FAIL_PAGE_ALLOC */
2377 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2382 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2385 * Return true if free base pages are above 'mark'. For high-order checks it
2386 * will return true of the order-0 watermark is reached and there is at least
2387 * one free page of a suitable size. Checking now avoids taking the zone lock
2388 * to check in the allocation paths if no pages are free.
2390 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2391 unsigned long mark, int classzone_idx, int alloc_flags,
2396 const int alloc_harder = (alloc_flags & ALLOC_HARDER);
2398 /* free_pages may go negative - that's OK */
2399 free_pages -= (1 << order) - 1;
2401 if (alloc_flags & ALLOC_HIGH)
2405 * If the caller does not have rights to ALLOC_HARDER then subtract
2406 * the high-atomic reserves. This will over-estimate the size of the
2407 * atomic reserve but it avoids a search.
2409 if (likely(!alloc_harder))
2410 free_pages -= z->nr_reserved_highatomic;
2415 /* If allocation can't use CMA areas don't use free CMA pages */
2416 if (!(alloc_flags & ALLOC_CMA))
2417 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2421 * Check watermarks for an order-0 allocation request. If these
2422 * are not met, then a high-order request also cannot go ahead
2423 * even if a suitable page happened to be free.
2425 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2428 /* If this is an order-0 request then the watermark is fine */
2432 /* For a high-order request, check at least one suitable page is free */
2433 for (o = order; o < MAX_ORDER; o++) {
2434 struct free_area *area = &z->free_area[o];
2443 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2444 if (!list_empty(&area->free_list[mt]))
2449 if ((alloc_flags & ALLOC_CMA) &&
2450 !list_empty(&area->free_list[MIGRATE_CMA])) {
2458 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2459 int classzone_idx, int alloc_flags)
2461 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2462 zone_page_state(z, NR_FREE_PAGES));
2465 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2466 unsigned long mark, int classzone_idx)
2468 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2470 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2471 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2473 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2478 static bool zone_local(struct zone *local_zone, struct zone *zone)
2480 return local_zone->node == zone->node;
2483 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2485 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2488 #else /* CONFIG_NUMA */
2489 static bool zone_local(struct zone *local_zone, struct zone *zone)
2494 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2498 #endif /* CONFIG_NUMA */
2500 static void reset_alloc_batches(struct zone *preferred_zone)
2502 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2505 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2506 high_wmark_pages(zone) - low_wmark_pages(zone) -
2507 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2508 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2509 } while (zone++ != preferred_zone);
2513 * get_page_from_freelist goes through the zonelist trying to allocate
2516 static struct page *
2517 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2518 const struct alloc_context *ac)
2520 struct zonelist *zonelist = ac->zonelist;
2522 struct page *page = NULL;
2524 int nr_fair_skipped = 0;
2525 bool zonelist_rescan;
2528 zonelist_rescan = false;
2531 * Scan zonelist, looking for a zone with enough free.
2532 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2534 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2538 if (cpusets_enabled() &&
2539 (alloc_flags & ALLOC_CPUSET) &&
2540 !cpuset_zone_allowed(zone, gfp_mask))
2543 * Distribute pages in proportion to the individual
2544 * zone size to ensure fair page aging. The zone a
2545 * page was allocated in should have no effect on the
2546 * time the page has in memory before being reclaimed.
2548 if (alloc_flags & ALLOC_FAIR) {
2549 if (!zone_local(ac->preferred_zone, zone))
2551 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2557 * When allocating a page cache page for writing, we
2558 * want to get it from a zone that is within its dirty
2559 * limit, such that no single zone holds more than its
2560 * proportional share of globally allowed dirty pages.
2561 * The dirty limits take into account the zone's
2562 * lowmem reserves and high watermark so that kswapd
2563 * should be able to balance it without having to
2564 * write pages from its LRU list.
2566 * This may look like it could increase pressure on
2567 * lower zones by failing allocations in higher zones
2568 * before they are full. But the pages that do spill
2569 * over are limited as the lower zones are protected
2570 * by this very same mechanism. It should not become
2571 * a practical burden to them.
2573 * XXX: For now, allow allocations to potentially
2574 * exceed the per-zone dirty limit in the slowpath
2575 * (spread_dirty_pages unset) before going into reclaim,
2576 * which is important when on a NUMA setup the allowed
2577 * zones are together not big enough to reach the
2578 * global limit. The proper fix for these situations
2579 * will require awareness of zones in the
2580 * dirty-throttling and the flusher threads.
2582 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2585 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2586 if (!zone_watermark_ok(zone, order, mark,
2587 ac->classzone_idx, alloc_flags)) {
2590 /* Checked here to keep the fast path fast */
2591 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2592 if (alloc_flags & ALLOC_NO_WATERMARKS)
2595 if (zone_reclaim_mode == 0 ||
2596 !zone_allows_reclaim(ac->preferred_zone, zone))
2599 ret = zone_reclaim(zone, gfp_mask, order);
2601 case ZONE_RECLAIM_NOSCAN:
2604 case ZONE_RECLAIM_FULL:
2605 /* scanned but unreclaimable */
2608 /* did we reclaim enough */
2609 if (zone_watermark_ok(zone, order, mark,
2610 ac->classzone_idx, alloc_flags))
2618 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2619 gfp_mask, alloc_flags, ac->migratetype);
2621 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2625 * If this is a high-order atomic allocation then check
2626 * if the pageblock should be reserved for the future
2628 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2629 reserve_highatomic_pageblock(page, zone, order);
2636 * The first pass makes sure allocations are spread fairly within the
2637 * local node. However, the local node might have free pages left
2638 * after the fairness batches are exhausted, and remote zones haven't
2639 * even been considered yet. Try once more without fairness, and
2640 * include remote zones now, before entering the slowpath and waking
2641 * kswapd: prefer spilling to a remote zone over swapping locally.
2643 if (alloc_flags & ALLOC_FAIR) {
2644 alloc_flags &= ~ALLOC_FAIR;
2645 if (nr_fair_skipped) {
2646 zonelist_rescan = true;
2647 reset_alloc_batches(ac->preferred_zone);
2649 if (nr_online_nodes > 1)
2650 zonelist_rescan = true;
2653 if (zonelist_rescan)
2660 * Large machines with many possible nodes should not always dump per-node
2661 * meminfo in irq context.
2663 static inline bool should_suppress_show_mem(void)
2668 ret = in_interrupt();
2673 static DEFINE_RATELIMIT_STATE(nopage_rs,
2674 DEFAULT_RATELIMIT_INTERVAL,
2675 DEFAULT_RATELIMIT_BURST);
2677 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
2679 unsigned int filter = SHOW_MEM_FILTER_NODES;
2681 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2682 debug_guardpage_minorder() > 0)
2686 * This documents exceptions given to allocations in certain
2687 * contexts that are allowed to allocate outside current's set
2690 if (!(gfp_mask & __GFP_NOMEMALLOC))
2691 if (test_thread_flag(TIF_MEMDIE) ||
2692 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2693 filter &= ~SHOW_MEM_FILTER_NODES;
2694 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2695 filter &= ~SHOW_MEM_FILTER_NODES;
2698 struct va_format vaf;
2701 va_start(args, fmt);
2706 pr_warn("%pV", &vaf);
2711 pr_warn("%s: page allocation failure: order:%u, mode:%#x(%pGg)\n",
2712 current->comm, order, gfp_mask, &gfp_mask);
2714 if (!should_suppress_show_mem())
2718 static inline struct page *
2719 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2720 const struct alloc_context *ac, unsigned long *did_some_progress)
2722 struct oom_control oc = {
2723 .zonelist = ac->zonelist,
2724 .nodemask = ac->nodemask,
2725 .gfp_mask = gfp_mask,
2730 *did_some_progress = 0;
2733 * Acquire the oom lock. If that fails, somebody else is
2734 * making progress for us.
2736 if (!mutex_trylock(&oom_lock)) {
2737 *did_some_progress = 1;
2738 schedule_timeout_uninterruptible(1);
2743 * Go through the zonelist yet one more time, keep very high watermark
2744 * here, this is only to catch a parallel oom killing, we must fail if
2745 * we're still under heavy pressure.
2747 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2748 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2752 if (!(gfp_mask & __GFP_NOFAIL)) {
2753 /* Coredumps can quickly deplete all memory reserves */
2754 if (current->flags & PF_DUMPCORE)
2756 /* The OOM killer will not help higher order allocs */
2757 if (order > PAGE_ALLOC_COSTLY_ORDER)
2759 /* The OOM killer does not needlessly kill tasks for lowmem */
2760 if (ac->high_zoneidx < ZONE_NORMAL)
2762 /* The OOM killer does not compensate for IO-less reclaim */
2763 if (!(gfp_mask & __GFP_FS)) {
2765 * XXX: Page reclaim didn't yield anything,
2766 * and the OOM killer can't be invoked, but
2767 * keep looping as per tradition.
2769 *did_some_progress = 1;
2772 if (pm_suspended_storage())
2774 /* The OOM killer may not free memory on a specific node */
2775 if (gfp_mask & __GFP_THISNODE)
2778 /* Exhausted what can be done so it's blamo time */
2779 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
2780 *did_some_progress = 1;
2782 if (gfp_mask & __GFP_NOFAIL) {
2783 page = get_page_from_freelist(gfp_mask, order,
2784 ALLOC_NO_WATERMARKS|ALLOC_CPUSET, ac);
2786 * fallback to ignore cpuset restriction if our nodes
2790 page = get_page_from_freelist(gfp_mask, order,
2791 ALLOC_NO_WATERMARKS, ac);
2795 mutex_unlock(&oom_lock);
2799 #ifdef CONFIG_COMPACTION
2800 /* Try memory compaction for high-order allocations before reclaim */
2801 static struct page *
2802 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2803 int alloc_flags, const struct alloc_context *ac,
2804 enum migrate_mode mode, int *contended_compaction,
2805 bool *deferred_compaction)
2807 unsigned long compact_result;
2813 current->flags |= PF_MEMALLOC;
2814 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2815 mode, contended_compaction);
2816 current->flags &= ~PF_MEMALLOC;
2818 switch (compact_result) {
2819 case COMPACT_DEFERRED:
2820 *deferred_compaction = true;
2822 case COMPACT_SKIPPED:
2829 * At least in one zone compaction wasn't deferred or skipped, so let's
2830 * count a compaction stall
2832 count_vm_event(COMPACTSTALL);
2834 page = get_page_from_freelist(gfp_mask, order,
2835 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2838 struct zone *zone = page_zone(page);
2840 zone->compact_blockskip_flush = false;
2841 compaction_defer_reset(zone, order, true);
2842 count_vm_event(COMPACTSUCCESS);
2847 * It's bad if compaction run occurs and fails. The most likely reason
2848 * is that pages exist, but not enough to satisfy watermarks.
2850 count_vm_event(COMPACTFAIL);
2857 static inline struct page *
2858 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2859 int alloc_flags, const struct alloc_context *ac,
2860 enum migrate_mode mode, int *contended_compaction,
2861 bool *deferred_compaction)
2865 #endif /* CONFIG_COMPACTION */
2867 /* Perform direct synchronous page reclaim */
2869 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2870 const struct alloc_context *ac)
2872 struct reclaim_state reclaim_state;
2877 /* We now go into synchronous reclaim */
2878 cpuset_memory_pressure_bump();
2879 current->flags |= PF_MEMALLOC;
2880 lockdep_set_current_reclaim_state(gfp_mask);
2881 reclaim_state.reclaimed_slab = 0;
2882 current->reclaim_state = &reclaim_state;
2884 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2887 current->reclaim_state = NULL;
2888 lockdep_clear_current_reclaim_state();
2889 current->flags &= ~PF_MEMALLOC;
2896 /* The really slow allocator path where we enter direct reclaim */
2897 static inline struct page *
2898 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2899 int alloc_flags, const struct alloc_context *ac,
2900 unsigned long *did_some_progress)
2902 struct page *page = NULL;
2903 bool drained = false;
2905 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2906 if (unlikely(!(*did_some_progress)))
2910 page = get_page_from_freelist(gfp_mask, order,
2911 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2914 * If an allocation failed after direct reclaim, it could be because
2915 * pages are pinned on the per-cpu lists or in high alloc reserves.
2916 * Shrink them them and try again
2918 if (!page && !drained) {
2919 unreserve_highatomic_pageblock(ac);
2920 drain_all_pages(NULL);
2928 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2933 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2934 ac->high_zoneidx, ac->nodemask)
2935 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2939 gfp_to_alloc_flags(gfp_t gfp_mask)
2941 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2943 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2944 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2947 * The caller may dip into page reserves a bit more if the caller
2948 * cannot run direct reclaim, or if the caller has realtime scheduling
2949 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2950 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
2952 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2954 if (gfp_mask & __GFP_ATOMIC) {
2956 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2957 * if it can't schedule.
2959 if (!(gfp_mask & __GFP_NOMEMALLOC))
2960 alloc_flags |= ALLOC_HARDER;
2962 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2963 * comment for __cpuset_node_allowed().
2965 alloc_flags &= ~ALLOC_CPUSET;
2966 } else if (unlikely(rt_task(current)) && !in_interrupt())
2967 alloc_flags |= ALLOC_HARDER;
2969 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2970 if (gfp_mask & __GFP_MEMALLOC)
2971 alloc_flags |= ALLOC_NO_WATERMARKS;
2972 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2973 alloc_flags |= ALLOC_NO_WATERMARKS;
2974 else if (!in_interrupt() &&
2975 ((current->flags & PF_MEMALLOC) ||
2976 unlikely(test_thread_flag(TIF_MEMDIE))))
2977 alloc_flags |= ALLOC_NO_WATERMARKS;
2980 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2981 alloc_flags |= ALLOC_CMA;
2986 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2988 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2991 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
2993 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
2996 static inline struct page *
2997 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2998 struct alloc_context *ac)
3000 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3001 struct page *page = NULL;
3003 unsigned long pages_reclaimed = 0;
3004 unsigned long did_some_progress;
3005 enum migrate_mode migration_mode = MIGRATE_ASYNC;
3006 bool deferred_compaction = false;
3007 int contended_compaction = COMPACT_CONTENDED_NONE;
3010 * In the slowpath, we sanity check order to avoid ever trying to
3011 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3012 * be using allocators in order of preference for an area that is
3015 if (order >= MAX_ORDER) {
3016 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3021 * We also sanity check to catch abuse of atomic reserves being used by
3022 * callers that are not in atomic context.
3024 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3025 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3026 gfp_mask &= ~__GFP_ATOMIC;
3029 * If this allocation cannot block and it is for a specific node, then
3030 * fail early. There's no need to wakeup kswapd or retry for a
3031 * speculative node-specific allocation.
3033 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !can_direct_reclaim)
3037 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3038 wake_all_kswapds(order, ac);
3041 * OK, we're below the kswapd watermark and have kicked background
3042 * reclaim. Now things get more complex, so set up alloc_flags according
3043 * to how we want to proceed.
3045 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3048 * Find the true preferred zone if the allocation is unconstrained by
3051 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3052 struct zoneref *preferred_zoneref;
3053 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3054 ac->high_zoneidx, NULL, &ac->preferred_zone);
3055 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3058 /* This is the last chance, in general, before the goto nopage. */
3059 page = get_page_from_freelist(gfp_mask, order,
3060 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3064 /* Allocate without watermarks if the context allows */
3065 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3067 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3068 * the allocation is high priority and these type of
3069 * allocations are system rather than user orientated
3071 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3072 page = get_page_from_freelist(gfp_mask, order,
3073 ALLOC_NO_WATERMARKS, ac);
3078 /* Caller is not willing to reclaim, we can't balance anything */
3079 if (!can_direct_reclaim) {
3081 * All existing users of the __GFP_NOFAIL are blockable, so warn
3082 * of any new users that actually allow this type of allocation
3085 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3089 /* Avoid recursion of direct reclaim */
3090 if (current->flags & PF_MEMALLOC) {
3092 * __GFP_NOFAIL request from this context is rather bizarre
3093 * because we cannot reclaim anything and only can loop waiting
3094 * for somebody to do a work for us.
3096 if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3103 /* Avoid allocations with no watermarks from looping endlessly */
3104 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3108 * Try direct compaction. The first pass is asynchronous. Subsequent
3109 * attempts after direct reclaim are synchronous
3111 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3113 &contended_compaction,
3114 &deferred_compaction);
3118 /* Checks for THP-specific high-order allocations */
3119 if (is_thp_gfp_mask(gfp_mask)) {
3121 * If compaction is deferred for high-order allocations, it is
3122 * because sync compaction recently failed. If this is the case
3123 * and the caller requested a THP allocation, we do not want
3124 * to heavily disrupt the system, so we fail the allocation
3125 * instead of entering direct reclaim.
3127 if (deferred_compaction)
3131 * In all zones where compaction was attempted (and not
3132 * deferred or skipped), lock contention has been detected.
3133 * For THP allocation we do not want to disrupt the others
3134 * so we fallback to base pages instead.
3136 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3140 * If compaction was aborted due to need_resched(), we do not
3141 * want to further increase allocation latency, unless it is
3142 * khugepaged trying to collapse.
3144 if (contended_compaction == COMPACT_CONTENDED_SCHED
3145 && !(current->flags & PF_KTHREAD))
3150 * It can become very expensive to allocate transparent hugepages at
3151 * fault, so use asynchronous memory compaction for THP unless it is
3152 * khugepaged trying to collapse.
3154 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3155 migration_mode = MIGRATE_SYNC_LIGHT;
3157 /* Try direct reclaim and then allocating */
3158 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3159 &did_some_progress);
3163 /* Do not loop if specifically requested */
3164 if (gfp_mask & __GFP_NORETRY)
3167 /* Keep reclaiming pages as long as there is reasonable progress */
3168 pages_reclaimed += did_some_progress;
3169 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3170 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3171 /* Wait for some write requests to complete then retry */
3172 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3176 /* Reclaim has failed us, start killing things */
3177 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3181 /* Retry as long as the OOM killer is making progress */
3182 if (did_some_progress)
3187 * High-order allocations do not necessarily loop after
3188 * direct reclaim and reclaim/compaction depends on compaction
3189 * being called after reclaim so call directly if necessary
3191 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3193 &contended_compaction,
3194 &deferred_compaction);
3198 warn_alloc_failed(gfp_mask, order, NULL);
3204 * This is the 'heart' of the zoned buddy allocator.
3207 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3208 struct zonelist *zonelist, nodemask_t *nodemask)
3210 struct zoneref *preferred_zoneref;
3211 struct page *page = NULL;
3212 unsigned int cpuset_mems_cookie;
3213 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3214 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3215 struct alloc_context ac = {
3216 .high_zoneidx = gfp_zone(gfp_mask),
3217 .nodemask = nodemask,
3218 .migratetype = gfpflags_to_migratetype(gfp_mask),
3221 gfp_mask &= gfp_allowed_mask;
3223 lockdep_trace_alloc(gfp_mask);
3225 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3227 if (should_fail_alloc_page(gfp_mask, order))
3231 * Check the zones suitable for the gfp_mask contain at least one
3232 * valid zone. It's possible to have an empty zonelist as a result
3233 * of __GFP_THISNODE and a memoryless node
3235 if (unlikely(!zonelist->_zonerefs->zone))
3238 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3239 alloc_flags |= ALLOC_CMA;
3242 cpuset_mems_cookie = read_mems_allowed_begin();
3244 /* We set it here, as __alloc_pages_slowpath might have changed it */
3245 ac.zonelist = zonelist;
3247 /* Dirty zone balancing only done in the fast path */
3248 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3250 /* The preferred zone is used for statistics later */
3251 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3252 ac.nodemask ? : &cpuset_current_mems_allowed,
3253 &ac.preferred_zone);
3254 if (!ac.preferred_zone)
3256 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3258 /* First allocation attempt */
3259 alloc_mask = gfp_mask|__GFP_HARDWALL;
3260 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3261 if (unlikely(!page)) {
3263 * Runtime PM, block IO and its error handling path
3264 * can deadlock because I/O on the device might not
3267 alloc_mask = memalloc_noio_flags(gfp_mask);
3268 ac.spread_dirty_pages = false;
3270 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3273 if (kmemcheck_enabled && page)
3274 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3276 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3280 * When updating a task's mems_allowed, it is possible to race with
3281 * parallel threads in such a way that an allocation can fail while
3282 * the mask is being updated. If a page allocation is about to fail,
3283 * check if the cpuset changed during allocation and if so, retry.
3285 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3290 EXPORT_SYMBOL(__alloc_pages_nodemask);
3293 * Common helper functions.
3295 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3300 * __get_free_pages() returns a 32-bit address, which cannot represent
3303 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3305 page = alloc_pages(gfp_mask, order);
3308 return (unsigned long) page_address(page);
3310 EXPORT_SYMBOL(__get_free_pages);
3312 unsigned long get_zeroed_page(gfp_t gfp_mask)
3314 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3316 EXPORT_SYMBOL(get_zeroed_page);
3318 void __free_pages(struct page *page, unsigned int order)
3320 if (put_page_testzero(page)) {
3322 free_hot_cold_page(page, false);
3324 __free_pages_ok(page, order);
3328 EXPORT_SYMBOL(__free_pages);
3330 void free_pages(unsigned long addr, unsigned int order)
3333 VM_BUG_ON(!virt_addr_valid((void *)addr));
3334 __free_pages(virt_to_page((void *)addr), order);
3338 EXPORT_SYMBOL(free_pages);
3342 * An arbitrary-length arbitrary-offset area of memory which resides
3343 * within a 0 or higher order page. Multiple fragments within that page
3344 * are individually refcounted, in the page's reference counter.
3346 * The page_frag functions below provide a simple allocation framework for
3347 * page fragments. This is used by the network stack and network device
3348 * drivers to provide a backing region of memory for use as either an
3349 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3351 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3354 struct page *page = NULL;
3355 gfp_t gfp = gfp_mask;
3357 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3358 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3360 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3361 PAGE_FRAG_CACHE_MAX_ORDER);
3362 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3364 if (unlikely(!page))
3365 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3367 nc->va = page ? page_address(page) : NULL;
3372 void *__alloc_page_frag(struct page_frag_cache *nc,
3373 unsigned int fragsz, gfp_t gfp_mask)
3375 unsigned int size = PAGE_SIZE;
3379 if (unlikely(!nc->va)) {
3381 page = __page_frag_refill(nc, gfp_mask);
3385 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3386 /* if size can vary use size else just use PAGE_SIZE */
3389 /* Even if we own the page, we do not use atomic_set().
3390 * This would break get_page_unless_zero() users.
3392 atomic_add(size - 1, &page->_count);
3394 /* reset page count bias and offset to start of new frag */
3395 nc->pfmemalloc = page_is_pfmemalloc(page);
3396 nc->pagecnt_bias = size;
3400 offset = nc->offset - fragsz;
3401 if (unlikely(offset < 0)) {
3402 page = virt_to_page(nc->va);
3404 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3407 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3408 /* if size can vary use size else just use PAGE_SIZE */
3411 /* OK, page count is 0, we can safely set it */
3412 atomic_set(&page->_count, size);
3414 /* reset page count bias and offset to start of new frag */
3415 nc->pagecnt_bias = size;
3416 offset = size - fragsz;
3420 nc->offset = offset;
3422 return nc->va + offset;
3424 EXPORT_SYMBOL(__alloc_page_frag);
3427 * Frees a page fragment allocated out of either a compound or order 0 page.
3429 void __free_page_frag(void *addr)
3431 struct page *page = virt_to_head_page(addr);
3433 if (unlikely(put_page_testzero(page)))
3434 __free_pages_ok(page, compound_order(page));
3436 EXPORT_SYMBOL(__free_page_frag);
3439 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3440 * of the current memory cgroup if __GFP_ACCOUNT is set, other than that it is
3441 * equivalent to alloc_pages.
3443 * It should be used when the caller would like to use kmalloc, but since the
3444 * allocation is large, it has to fall back to the page allocator.
3446 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3450 page = alloc_pages(gfp_mask, order);
3451 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3452 __free_pages(page, order);
3458 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3462 page = alloc_pages_node(nid, gfp_mask, order);
3463 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3464 __free_pages(page, order);
3471 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3474 void __free_kmem_pages(struct page *page, unsigned int order)
3476 memcg_kmem_uncharge(page, order);
3477 __free_pages(page, order);
3480 void free_kmem_pages(unsigned long addr, unsigned int order)
3483 VM_BUG_ON(!virt_addr_valid((void *)addr));
3484 __free_kmem_pages(virt_to_page((void *)addr), order);
3488 static void *make_alloc_exact(unsigned long addr, unsigned int order,
3492 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3493 unsigned long used = addr + PAGE_ALIGN(size);
3495 split_page(virt_to_page((void *)addr), order);
3496 while (used < alloc_end) {
3501 return (void *)addr;
3505 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3506 * @size: the number of bytes to allocate
3507 * @gfp_mask: GFP flags for the allocation
3509 * This function is similar to alloc_pages(), except that it allocates the
3510 * minimum number of pages to satisfy the request. alloc_pages() can only
3511 * allocate memory in power-of-two pages.
3513 * This function is also limited by MAX_ORDER.
3515 * Memory allocated by this function must be released by free_pages_exact().
3517 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3519 unsigned int order = get_order(size);
3522 addr = __get_free_pages(gfp_mask, order);
3523 return make_alloc_exact(addr, order, size);
3525 EXPORT_SYMBOL(alloc_pages_exact);
3528 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3530 * @nid: the preferred node ID where memory should be allocated
3531 * @size: the number of bytes to allocate
3532 * @gfp_mask: GFP flags for the allocation
3534 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3537 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3539 unsigned int order = get_order(size);
3540 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3543 return make_alloc_exact((unsigned long)page_address(p), order, size);
3547 * free_pages_exact - release memory allocated via alloc_pages_exact()
3548 * @virt: the value returned by alloc_pages_exact.
3549 * @size: size of allocation, same value as passed to alloc_pages_exact().
3551 * Release the memory allocated by a previous call to alloc_pages_exact.
3553 void free_pages_exact(void *virt, size_t size)
3555 unsigned long addr = (unsigned long)virt;
3556 unsigned long end = addr + PAGE_ALIGN(size);
3558 while (addr < end) {
3563 EXPORT_SYMBOL(free_pages_exact);
3566 * nr_free_zone_pages - count number of pages beyond high watermark
3567 * @offset: The zone index of the highest zone
3569 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3570 * high watermark within all zones at or below a given zone index. For each
3571 * zone, the number of pages is calculated as:
3572 * managed_pages - high_pages
3574 static unsigned long nr_free_zone_pages(int offset)
3579 /* Just pick one node, since fallback list is circular */
3580 unsigned long sum = 0;
3582 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3584 for_each_zone_zonelist(zone, z, zonelist, offset) {
3585 unsigned long size = zone->managed_pages;
3586 unsigned long high = high_wmark_pages(zone);
3595 * nr_free_buffer_pages - count number of pages beyond high watermark
3597 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3598 * watermark within ZONE_DMA and ZONE_NORMAL.
3600 unsigned long nr_free_buffer_pages(void)
3602 return nr_free_zone_pages(gfp_zone(GFP_USER));
3604 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3607 * nr_free_pagecache_pages - count number of pages beyond high watermark
3609 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3610 * high watermark within all zones.
3612 unsigned long nr_free_pagecache_pages(void)
3614 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3617 static inline void show_node(struct zone *zone)
3619 if (IS_ENABLED(CONFIG_NUMA))
3620 printk("Node %d ", zone_to_nid(zone));
3623 void si_meminfo(struct sysinfo *val)
3625 val->totalram = totalram_pages;
3626 val->sharedram = global_page_state(NR_SHMEM);
3627 val->freeram = global_page_state(NR_FREE_PAGES);
3628 val->bufferram = nr_blockdev_pages();
3629 val->totalhigh = totalhigh_pages;
3630 val->freehigh = nr_free_highpages();
3631 val->mem_unit = PAGE_SIZE;
3634 EXPORT_SYMBOL(si_meminfo);
3637 void si_meminfo_node(struct sysinfo *val, int nid)
3639 int zone_type; /* needs to be signed */
3640 unsigned long managed_pages = 0;
3641 pg_data_t *pgdat = NODE_DATA(nid);
3643 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3644 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3645 val->totalram = managed_pages;
3646 val->sharedram = node_page_state(nid, NR_SHMEM);
3647 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3648 #ifdef CONFIG_HIGHMEM
3649 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3650 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3656 val->mem_unit = PAGE_SIZE;
3661 * Determine whether the node should be displayed or not, depending on whether
3662 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3664 bool skip_free_areas_node(unsigned int flags, int nid)
3667 unsigned int cpuset_mems_cookie;
3669 if (!(flags & SHOW_MEM_FILTER_NODES))
3673 cpuset_mems_cookie = read_mems_allowed_begin();
3674 ret = !node_isset(nid, cpuset_current_mems_allowed);
3675 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3680 #define K(x) ((x) << (PAGE_SHIFT-10))
3682 static void show_migration_types(unsigned char type)
3684 static const char types[MIGRATE_TYPES] = {
3685 [MIGRATE_UNMOVABLE] = 'U',
3686 [MIGRATE_MOVABLE] = 'M',
3687 [MIGRATE_RECLAIMABLE] = 'E',
3688 [MIGRATE_HIGHATOMIC] = 'H',
3690 [MIGRATE_CMA] = 'C',
3692 #ifdef CONFIG_MEMORY_ISOLATION
3693 [MIGRATE_ISOLATE] = 'I',
3696 char tmp[MIGRATE_TYPES + 1];
3700 for (i = 0; i < MIGRATE_TYPES; i++) {
3701 if (type & (1 << i))
3706 printk("(%s) ", tmp);
3710 * Show free area list (used inside shift_scroll-lock stuff)
3711 * We also calculate the percentage fragmentation. We do this by counting the
3712 * memory on each free list with the exception of the first item on the list.
3715 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3718 void show_free_areas(unsigned int filter)
3720 unsigned long free_pcp = 0;
3724 for_each_populated_zone(zone) {
3725 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3728 for_each_online_cpu(cpu)
3729 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3732 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3733 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3734 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3735 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3736 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3737 " free:%lu free_pcp:%lu free_cma:%lu\n",
3738 global_page_state(NR_ACTIVE_ANON),
3739 global_page_state(NR_INACTIVE_ANON),
3740 global_page_state(NR_ISOLATED_ANON),
3741 global_page_state(NR_ACTIVE_FILE),
3742 global_page_state(NR_INACTIVE_FILE),
3743 global_page_state(NR_ISOLATED_FILE),
3744 global_page_state(NR_UNEVICTABLE),
3745 global_page_state(NR_FILE_DIRTY),
3746 global_page_state(NR_WRITEBACK),
3747 global_page_state(NR_UNSTABLE_NFS),
3748 global_page_state(NR_SLAB_RECLAIMABLE),
3749 global_page_state(NR_SLAB_UNRECLAIMABLE),
3750 global_page_state(NR_FILE_MAPPED),
3751 global_page_state(NR_SHMEM),
3752 global_page_state(NR_PAGETABLE),
3753 global_page_state(NR_BOUNCE),
3754 global_page_state(NR_FREE_PAGES),
3756 global_page_state(NR_FREE_CMA_PAGES));
3758 for_each_populated_zone(zone) {
3761 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3765 for_each_online_cpu(cpu)
3766 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3774 " active_anon:%lukB"
3775 " inactive_anon:%lukB"
3776 " active_file:%lukB"
3777 " inactive_file:%lukB"
3778 " unevictable:%lukB"
3779 " isolated(anon):%lukB"
3780 " isolated(file):%lukB"
3788 " slab_reclaimable:%lukB"
3789 " slab_unreclaimable:%lukB"
3790 " kernel_stack:%lukB"
3797 " writeback_tmp:%lukB"
3798 " pages_scanned:%lu"
3799 " all_unreclaimable? %s"
3802 K(zone_page_state(zone, NR_FREE_PAGES)),
3803 K(min_wmark_pages(zone)),
3804 K(low_wmark_pages(zone)),
3805 K(high_wmark_pages(zone)),
3806 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3807 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3808 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3809 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3810 K(zone_page_state(zone, NR_UNEVICTABLE)),
3811 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3812 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3813 K(zone->present_pages),
3814 K(zone->managed_pages),
3815 K(zone_page_state(zone, NR_MLOCK)),
3816 K(zone_page_state(zone, NR_FILE_DIRTY)),
3817 K(zone_page_state(zone, NR_WRITEBACK)),
3818 K(zone_page_state(zone, NR_FILE_MAPPED)),
3819 K(zone_page_state(zone, NR_SHMEM)),
3820 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3821 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3822 zone_page_state(zone, NR_KERNEL_STACK) *
3824 K(zone_page_state(zone, NR_PAGETABLE)),
3825 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3826 K(zone_page_state(zone, NR_BOUNCE)),
3828 K(this_cpu_read(zone->pageset->pcp.count)),
3829 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3830 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3831 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3832 (!zone_reclaimable(zone) ? "yes" : "no")
3834 printk("lowmem_reserve[]:");
3835 for (i = 0; i < MAX_NR_ZONES; i++)
3836 printk(" %ld", zone->lowmem_reserve[i]);
3840 for_each_populated_zone(zone) {
3842 unsigned long nr[MAX_ORDER], flags, total = 0;
3843 unsigned char types[MAX_ORDER];
3845 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3848 printk("%s: ", zone->name);
3850 spin_lock_irqsave(&zone->lock, flags);
3851 for (order = 0; order < MAX_ORDER; order++) {
3852 struct free_area *area = &zone->free_area[order];
3855 nr[order] = area->nr_free;
3856 total += nr[order] << order;
3859 for (type = 0; type < MIGRATE_TYPES; type++) {
3860 if (!list_empty(&area->free_list[type]))
3861 types[order] |= 1 << type;
3864 spin_unlock_irqrestore(&zone->lock, flags);
3865 for (order = 0; order < MAX_ORDER; order++) {
3866 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3868 show_migration_types(types[order]);
3870 printk("= %lukB\n", K(total));
3873 hugetlb_show_meminfo();
3875 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3877 show_swap_cache_info();
3880 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3882 zoneref->zone = zone;
3883 zoneref->zone_idx = zone_idx(zone);
3887 * Builds allocation fallback zone lists.
3889 * Add all populated zones of a node to the zonelist.
3891 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3895 enum zone_type zone_type = MAX_NR_ZONES;
3899 zone = pgdat->node_zones + zone_type;
3900 if (populated_zone(zone)) {
3901 zoneref_set_zone(zone,
3902 &zonelist->_zonerefs[nr_zones++]);
3903 check_highest_zone(zone_type);
3905 } while (zone_type);
3913 * 0 = automatic detection of better ordering.
3914 * 1 = order by ([node] distance, -zonetype)
3915 * 2 = order by (-zonetype, [node] distance)
3917 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3918 * the same zonelist. So only NUMA can configure this param.
3920 #define ZONELIST_ORDER_DEFAULT 0
3921 #define ZONELIST_ORDER_NODE 1
3922 #define ZONELIST_ORDER_ZONE 2
3924 /* zonelist order in the kernel.
3925 * set_zonelist_order() will set this to NODE or ZONE.
3927 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3928 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3932 /* The value user specified ....changed by config */
3933 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3934 /* string for sysctl */
3935 #define NUMA_ZONELIST_ORDER_LEN 16
3936 char numa_zonelist_order[16] = "default";
3939 * interface for configure zonelist ordering.
3940 * command line option "numa_zonelist_order"
3941 * = "[dD]efault - default, automatic configuration.
3942 * = "[nN]ode - order by node locality, then by zone within node
3943 * = "[zZ]one - order by zone, then by locality within zone
3946 static int __parse_numa_zonelist_order(char *s)
3948 if (*s == 'd' || *s == 'D') {
3949 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3950 } else if (*s == 'n' || *s == 'N') {
3951 user_zonelist_order = ZONELIST_ORDER_NODE;
3952 } else if (*s == 'z' || *s == 'Z') {
3953 user_zonelist_order = ZONELIST_ORDER_ZONE;
3956 "Ignoring invalid numa_zonelist_order value: "
3963 static __init int setup_numa_zonelist_order(char *s)
3970 ret = __parse_numa_zonelist_order(s);
3972 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3976 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3979 * sysctl handler for numa_zonelist_order
3981 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3982 void __user *buffer, size_t *length,
3985 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3987 static DEFINE_MUTEX(zl_order_mutex);
3989 mutex_lock(&zl_order_mutex);
3991 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3995 strcpy(saved_string, (char *)table->data);
3997 ret = proc_dostring(table, write, buffer, length, ppos);
4001 int oldval = user_zonelist_order;
4003 ret = __parse_numa_zonelist_order((char *)table->data);
4006 * bogus value. restore saved string
4008 strncpy((char *)table->data, saved_string,
4009 NUMA_ZONELIST_ORDER_LEN);
4010 user_zonelist_order = oldval;
4011 } else if (oldval != user_zonelist_order) {
4012 mutex_lock(&zonelists_mutex);
4013 build_all_zonelists(NULL, NULL);
4014 mutex_unlock(&zonelists_mutex);
4018 mutex_unlock(&zl_order_mutex);
4023 #define MAX_NODE_LOAD (nr_online_nodes)
4024 static int node_load[MAX_NUMNODES];
4027 * find_next_best_node - find the next node that should appear in a given node's fallback list
4028 * @node: node whose fallback list we're appending
4029 * @used_node_mask: nodemask_t of already used nodes
4031 * We use a number of factors to determine which is the next node that should
4032 * appear on a given node's fallback list. The node should not have appeared
4033 * already in @node's fallback list, and it should be the next closest node
4034 * according to the distance array (which contains arbitrary distance values
4035 * from each node to each node in the system), and should also prefer nodes
4036 * with no CPUs, since presumably they'll have very little allocation pressure
4037 * on them otherwise.
4038 * It returns -1 if no node is found.
4040 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4043 int min_val = INT_MAX;
4044 int best_node = NUMA_NO_NODE;
4045 const struct cpumask *tmp = cpumask_of_node(0);
4047 /* Use the local node if we haven't already */
4048 if (!node_isset(node, *used_node_mask)) {
4049 node_set(node, *used_node_mask);
4053 for_each_node_state(n, N_MEMORY) {
4055 /* Don't want a node to appear more than once */
4056 if (node_isset(n, *used_node_mask))
4059 /* Use the distance array to find the distance */
4060 val = node_distance(node, n);
4062 /* Penalize nodes under us ("prefer the next node") */
4065 /* Give preference to headless and unused nodes */
4066 tmp = cpumask_of_node(n);
4067 if (!cpumask_empty(tmp))
4068 val += PENALTY_FOR_NODE_WITH_CPUS;
4070 /* Slight preference for less loaded node */
4071 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4072 val += node_load[n];
4074 if (val < min_val) {
4081 node_set(best_node, *used_node_mask);
4088 * Build zonelists ordered by node and zones within node.
4089 * This results in maximum locality--normal zone overflows into local
4090 * DMA zone, if any--but risks exhausting DMA zone.
4092 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4095 struct zonelist *zonelist;
4097 zonelist = &pgdat->node_zonelists[0];
4098 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4100 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4101 zonelist->_zonerefs[j].zone = NULL;
4102 zonelist->_zonerefs[j].zone_idx = 0;
4106 * Build gfp_thisnode zonelists
4108 static void build_thisnode_zonelists(pg_data_t *pgdat)
4111 struct zonelist *zonelist;
4113 zonelist = &pgdat->node_zonelists[1];
4114 j = build_zonelists_node(pgdat, zonelist, 0);
4115 zonelist->_zonerefs[j].zone = NULL;
4116 zonelist->_zonerefs[j].zone_idx = 0;
4120 * Build zonelists ordered by zone and nodes within zones.
4121 * This results in conserving DMA zone[s] until all Normal memory is
4122 * exhausted, but results in overflowing to remote node while memory
4123 * may still exist in local DMA zone.
4125 static int node_order[MAX_NUMNODES];
4127 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4130 int zone_type; /* needs to be signed */
4132 struct zonelist *zonelist;
4134 zonelist = &pgdat->node_zonelists[0];
4136 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4137 for (j = 0; j < nr_nodes; j++) {
4138 node = node_order[j];
4139 z = &NODE_DATA(node)->node_zones[zone_type];
4140 if (populated_zone(z)) {
4142 &zonelist->_zonerefs[pos++]);
4143 check_highest_zone(zone_type);
4147 zonelist->_zonerefs[pos].zone = NULL;
4148 zonelist->_zonerefs[pos].zone_idx = 0;
4151 #if defined(CONFIG_64BIT)
4153 * Devices that require DMA32/DMA are relatively rare and do not justify a
4154 * penalty to every machine in case the specialised case applies. Default
4155 * to Node-ordering on 64-bit NUMA machines
4157 static int default_zonelist_order(void)
4159 return ZONELIST_ORDER_NODE;
4163 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4164 * by the kernel. If processes running on node 0 deplete the low memory zone
4165 * then reclaim will occur more frequency increasing stalls and potentially
4166 * be easier to OOM if a large percentage of the zone is under writeback or
4167 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4168 * Hence, default to zone ordering on 32-bit.
4170 static int default_zonelist_order(void)
4172 return ZONELIST_ORDER_ZONE;
4174 #endif /* CONFIG_64BIT */
4176 static void set_zonelist_order(void)
4178 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4179 current_zonelist_order = default_zonelist_order();
4181 current_zonelist_order = user_zonelist_order;
4184 static void build_zonelists(pg_data_t *pgdat)
4187 nodemask_t used_mask;
4188 int local_node, prev_node;
4189 struct zonelist *zonelist;
4190 unsigned int order = current_zonelist_order;
4192 /* initialize zonelists */
4193 for (i = 0; i < MAX_ZONELISTS; i++) {
4194 zonelist = pgdat->node_zonelists + i;
4195 zonelist->_zonerefs[0].zone = NULL;
4196 zonelist->_zonerefs[0].zone_idx = 0;
4199 /* NUMA-aware ordering of nodes */
4200 local_node = pgdat->node_id;
4201 load = nr_online_nodes;
4202 prev_node = local_node;
4203 nodes_clear(used_mask);
4205 memset(node_order, 0, sizeof(node_order));
4208 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4210 * We don't want to pressure a particular node.
4211 * So adding penalty to the first node in same
4212 * distance group to make it round-robin.
4214 if (node_distance(local_node, node) !=
4215 node_distance(local_node, prev_node))
4216 node_load[node] = load;
4220 if (order == ZONELIST_ORDER_NODE)
4221 build_zonelists_in_node_order(pgdat, node);
4223 node_order[i++] = node; /* remember order */
4226 if (order == ZONELIST_ORDER_ZONE) {
4227 /* calculate node order -- i.e., DMA last! */
4228 build_zonelists_in_zone_order(pgdat, i);
4231 build_thisnode_zonelists(pgdat);
4234 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4236 * Return node id of node used for "local" allocations.
4237 * I.e., first node id of first zone in arg node's generic zonelist.
4238 * Used for initializing percpu 'numa_mem', which is used primarily
4239 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4241 int local_memory_node(int node)
4245 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4246 gfp_zone(GFP_KERNEL),
4253 #else /* CONFIG_NUMA */
4255 static void set_zonelist_order(void)
4257 current_zonelist_order = ZONELIST_ORDER_ZONE;
4260 static void build_zonelists(pg_data_t *pgdat)
4262 int node, local_node;
4264 struct zonelist *zonelist;
4266 local_node = pgdat->node_id;
4268 zonelist = &pgdat->node_zonelists[0];
4269 j = build_zonelists_node(pgdat, zonelist, 0);
4272 * Now we build the zonelist so that it contains the zones
4273 * of all the other nodes.
4274 * We don't want to pressure a particular node, so when
4275 * building the zones for node N, we make sure that the
4276 * zones coming right after the local ones are those from
4277 * node N+1 (modulo N)
4279 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4280 if (!node_online(node))
4282 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4284 for (node = 0; node < local_node; node++) {
4285 if (!node_online(node))
4287 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4290 zonelist->_zonerefs[j].zone = NULL;
4291 zonelist->_zonerefs[j].zone_idx = 0;
4294 #endif /* CONFIG_NUMA */
4297 * Boot pageset table. One per cpu which is going to be used for all
4298 * zones and all nodes. The parameters will be set in such a way
4299 * that an item put on a list will immediately be handed over to
4300 * the buddy list. This is safe since pageset manipulation is done
4301 * with interrupts disabled.
4303 * The boot_pagesets must be kept even after bootup is complete for
4304 * unused processors and/or zones. They do play a role for bootstrapping
4305 * hotplugged processors.
4307 * zoneinfo_show() and maybe other functions do
4308 * not check if the processor is online before following the pageset pointer.
4309 * Other parts of the kernel may not check if the zone is available.
4311 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4312 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4313 static void setup_zone_pageset(struct zone *zone);
4316 * Global mutex to protect against size modification of zonelists
4317 * as well as to serialize pageset setup for the new populated zone.
4319 DEFINE_MUTEX(zonelists_mutex);
4321 /* return values int ....just for stop_machine() */
4322 static int __build_all_zonelists(void *data)
4326 pg_data_t *self = data;
4329 memset(node_load, 0, sizeof(node_load));
4332 if (self && !node_online(self->node_id)) {
4333 build_zonelists(self);
4336 for_each_online_node(nid) {
4337 pg_data_t *pgdat = NODE_DATA(nid);
4339 build_zonelists(pgdat);
4343 * Initialize the boot_pagesets that are going to be used
4344 * for bootstrapping processors. The real pagesets for
4345 * each zone will be allocated later when the per cpu
4346 * allocator is available.
4348 * boot_pagesets are used also for bootstrapping offline
4349 * cpus if the system is already booted because the pagesets
4350 * are needed to initialize allocators on a specific cpu too.
4351 * F.e. the percpu allocator needs the page allocator which
4352 * needs the percpu allocator in order to allocate its pagesets
4353 * (a chicken-egg dilemma).
4355 for_each_possible_cpu(cpu) {
4356 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4358 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4360 * We now know the "local memory node" for each node--
4361 * i.e., the node of the first zone in the generic zonelist.
4362 * Set up numa_mem percpu variable for on-line cpus. During
4363 * boot, only the boot cpu should be on-line; we'll init the
4364 * secondary cpus' numa_mem as they come on-line. During
4365 * node/memory hotplug, we'll fixup all on-line cpus.
4367 if (cpu_online(cpu))
4368 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4375 static noinline void __init
4376 build_all_zonelists_init(void)
4378 __build_all_zonelists(NULL);
4379 mminit_verify_zonelist();
4380 cpuset_init_current_mems_allowed();
4384 * Called with zonelists_mutex held always
4385 * unless system_state == SYSTEM_BOOTING.
4387 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4388 * [we're only called with non-NULL zone through __meminit paths] and
4389 * (2) call of __init annotated helper build_all_zonelists_init
4390 * [protected by SYSTEM_BOOTING].
4392 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4394 set_zonelist_order();
4396 if (system_state == SYSTEM_BOOTING) {
4397 build_all_zonelists_init();
4399 #ifdef CONFIG_MEMORY_HOTPLUG
4401 setup_zone_pageset(zone);
4403 /* we have to stop all cpus to guarantee there is no user
4405 stop_machine(__build_all_zonelists, pgdat, NULL);
4406 /* cpuset refresh routine should be here */
4408 vm_total_pages = nr_free_pagecache_pages();
4410 * Disable grouping by mobility if the number of pages in the
4411 * system is too low to allow the mechanism to work. It would be
4412 * more accurate, but expensive to check per-zone. This check is
4413 * made on memory-hotadd so a system can start with mobility
4414 * disabled and enable it later
4416 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4417 page_group_by_mobility_disabled = 1;
4419 page_group_by_mobility_disabled = 0;
4421 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4422 "Total pages: %ld\n",
4424 zonelist_order_name[current_zonelist_order],
4425 page_group_by_mobility_disabled ? "off" : "on",
4428 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4433 * Helper functions to size the waitqueue hash table.
4434 * Essentially these want to choose hash table sizes sufficiently
4435 * large so that collisions trying to wait on pages are rare.
4436 * But in fact, the number of active page waitqueues on typical
4437 * systems is ridiculously low, less than 200. So this is even
4438 * conservative, even though it seems large.
4440 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4441 * waitqueues, i.e. the size of the waitq table given the number of pages.
4443 #define PAGES_PER_WAITQUEUE 256
4445 #ifndef CONFIG_MEMORY_HOTPLUG
4446 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4448 unsigned long size = 1;
4450 pages /= PAGES_PER_WAITQUEUE;
4452 while (size < pages)
4456 * Once we have dozens or even hundreds of threads sleeping
4457 * on IO we've got bigger problems than wait queue collision.
4458 * Limit the size of the wait table to a reasonable size.
4460 size = min(size, 4096UL);
4462 return max(size, 4UL);
4466 * A zone's size might be changed by hot-add, so it is not possible to determine
4467 * a suitable size for its wait_table. So we use the maximum size now.
4469 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4471 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4472 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4473 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4475 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4476 * or more by the traditional way. (See above). It equals:
4478 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4479 * ia64(16K page size) : = ( 8G + 4M)byte.
4480 * powerpc (64K page size) : = (32G +16M)byte.
4482 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4489 * This is an integer logarithm so that shifts can be used later
4490 * to extract the more random high bits from the multiplicative
4491 * hash function before the remainder is taken.
4493 static inline unsigned long wait_table_bits(unsigned long size)
4499 * Initially all pages are reserved - free ones are freed
4500 * up by free_all_bootmem() once the early boot process is
4501 * done. Non-atomic initialization, single-pass.
4503 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4504 unsigned long start_pfn, enum memmap_context context)
4506 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
4507 unsigned long end_pfn = start_pfn + size;
4508 pg_data_t *pgdat = NODE_DATA(nid);
4510 unsigned long nr_initialised = 0;
4511 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4512 struct memblock_region *r = NULL, *tmp;
4515 if (highest_memmap_pfn < end_pfn - 1)
4516 highest_memmap_pfn = end_pfn - 1;
4519 * Honor reservation requested by the driver for this ZONE_DEVICE
4522 if (altmap && start_pfn == altmap->base_pfn)
4523 start_pfn += altmap->reserve;
4525 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4527 * There can be holes in boot-time mem_map[]s handed to this
4528 * function. They do not exist on hotplugged memory.
4530 if (context != MEMMAP_EARLY)
4533 if (!early_pfn_valid(pfn))
4535 if (!early_pfn_in_nid(pfn, nid))
4537 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
4540 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4542 * If not mirrored_kernelcore and ZONE_MOVABLE exists, range
4543 * from zone_movable_pfn[nid] to end of each node should be
4544 * ZONE_MOVABLE not ZONE_NORMAL. skip it.
4546 if (!mirrored_kernelcore && zone_movable_pfn[nid])
4547 if (zone == ZONE_NORMAL && pfn >= zone_movable_pfn[nid])
4551 * Check given memblock attribute by firmware which can affect
4552 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
4553 * mirrored, it's an overlapped memmap init. skip it.
4555 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
4556 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
4557 for_each_memblock(memory, tmp)
4558 if (pfn < memblock_region_memory_end_pfn(tmp))
4562 if (pfn >= memblock_region_memory_base_pfn(r) &&
4563 memblock_is_mirror(r)) {
4564 /* already initialized as NORMAL */
4565 pfn = memblock_region_memory_end_pfn(r);
4573 * Mark the block movable so that blocks are reserved for
4574 * movable at startup. This will force kernel allocations
4575 * to reserve their blocks rather than leaking throughout
4576 * the address space during boot when many long-lived
4577 * kernel allocations are made.
4579 * bitmap is created for zone's valid pfn range. but memmap
4580 * can be created for invalid pages (for alignment)
4581 * check here not to call set_pageblock_migratetype() against
4584 if (!(pfn & (pageblock_nr_pages - 1))) {
4585 struct page *page = pfn_to_page(pfn);
4587 __init_single_page(page, pfn, zone, nid);
4588 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4590 __init_single_pfn(pfn, zone, nid);
4595 static void __meminit zone_init_free_lists(struct zone *zone)
4597 unsigned int order, t;
4598 for_each_migratetype_order(order, t) {
4599 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4600 zone->free_area[order].nr_free = 0;
4604 #ifndef __HAVE_ARCH_MEMMAP_INIT
4605 #define memmap_init(size, nid, zone, start_pfn) \
4606 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4609 static int zone_batchsize(struct zone *zone)
4615 * The per-cpu-pages pools are set to around 1000th of the
4616 * size of the zone. But no more than 1/2 of a meg.
4618 * OK, so we don't know how big the cache is. So guess.
4620 batch = zone->managed_pages / 1024;
4621 if (batch * PAGE_SIZE > 512 * 1024)
4622 batch = (512 * 1024) / PAGE_SIZE;
4623 batch /= 4; /* We effectively *= 4 below */
4628 * Clamp the batch to a 2^n - 1 value. Having a power
4629 * of 2 value was found to be more likely to have
4630 * suboptimal cache aliasing properties in some cases.
4632 * For example if 2 tasks are alternately allocating
4633 * batches of pages, one task can end up with a lot
4634 * of pages of one half of the possible page colors
4635 * and the other with pages of the other colors.
4637 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4642 /* The deferral and batching of frees should be suppressed under NOMMU
4645 * The problem is that NOMMU needs to be able to allocate large chunks
4646 * of contiguous memory as there's no hardware page translation to
4647 * assemble apparent contiguous memory from discontiguous pages.
4649 * Queueing large contiguous runs of pages for batching, however,
4650 * causes the pages to actually be freed in smaller chunks. As there
4651 * can be a significant delay between the individual batches being
4652 * recycled, this leads to the once large chunks of space being
4653 * fragmented and becoming unavailable for high-order allocations.
4660 * pcp->high and pcp->batch values are related and dependent on one another:
4661 * ->batch must never be higher then ->high.
4662 * The following function updates them in a safe manner without read side
4665 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4666 * those fields changing asynchronously (acording the the above rule).
4668 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4669 * outside of boot time (or some other assurance that no concurrent updaters
4672 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4673 unsigned long batch)
4675 /* start with a fail safe value for batch */
4679 /* Update high, then batch, in order */
4686 /* a companion to pageset_set_high() */
4687 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4689 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4692 static void pageset_init(struct per_cpu_pageset *p)
4694 struct per_cpu_pages *pcp;
4697 memset(p, 0, sizeof(*p));
4701 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4702 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4705 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4708 pageset_set_batch(p, batch);
4712 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4713 * to the value high for the pageset p.
4715 static void pageset_set_high(struct per_cpu_pageset *p,
4718 unsigned long batch = max(1UL, high / 4);
4719 if ((high / 4) > (PAGE_SHIFT * 8))
4720 batch = PAGE_SHIFT * 8;
4722 pageset_update(&p->pcp, high, batch);
4725 static void pageset_set_high_and_batch(struct zone *zone,
4726 struct per_cpu_pageset *pcp)
4728 if (percpu_pagelist_fraction)
4729 pageset_set_high(pcp,
4730 (zone->managed_pages /
4731 percpu_pagelist_fraction));
4733 pageset_set_batch(pcp, zone_batchsize(zone));
4736 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4738 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4741 pageset_set_high_and_batch(zone, pcp);
4744 static void __meminit setup_zone_pageset(struct zone *zone)
4747 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4748 for_each_possible_cpu(cpu)
4749 zone_pageset_init(zone, cpu);
4753 * Allocate per cpu pagesets and initialize them.
4754 * Before this call only boot pagesets were available.
4756 void __init setup_per_cpu_pageset(void)
4760 for_each_populated_zone(zone)
4761 setup_zone_pageset(zone);
4764 static noinline __init_refok
4765 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4771 * The per-page waitqueue mechanism uses hashed waitqueues
4774 zone->wait_table_hash_nr_entries =
4775 wait_table_hash_nr_entries(zone_size_pages);
4776 zone->wait_table_bits =
4777 wait_table_bits(zone->wait_table_hash_nr_entries);
4778 alloc_size = zone->wait_table_hash_nr_entries
4779 * sizeof(wait_queue_head_t);
4781 if (!slab_is_available()) {
4782 zone->wait_table = (wait_queue_head_t *)
4783 memblock_virt_alloc_node_nopanic(
4784 alloc_size, zone->zone_pgdat->node_id);
4787 * This case means that a zone whose size was 0 gets new memory
4788 * via memory hot-add.
4789 * But it may be the case that a new node was hot-added. In
4790 * this case vmalloc() will not be able to use this new node's
4791 * memory - this wait_table must be initialized to use this new
4792 * node itself as well.
4793 * To use this new node's memory, further consideration will be
4796 zone->wait_table = vmalloc(alloc_size);
4798 if (!zone->wait_table)
4801 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4802 init_waitqueue_head(zone->wait_table + i);
4807 static __meminit void zone_pcp_init(struct zone *zone)
4810 * per cpu subsystem is not up at this point. The following code
4811 * relies on the ability of the linker to provide the
4812 * offset of a (static) per cpu variable into the per cpu area.
4814 zone->pageset = &boot_pageset;
4816 if (populated_zone(zone))
4817 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4818 zone->name, zone->present_pages,
4819 zone_batchsize(zone));
4822 int __meminit init_currently_empty_zone(struct zone *zone,
4823 unsigned long zone_start_pfn,
4826 struct pglist_data *pgdat = zone->zone_pgdat;
4828 ret = zone_wait_table_init(zone, size);
4831 pgdat->nr_zones = zone_idx(zone) + 1;
4833 zone->zone_start_pfn = zone_start_pfn;
4835 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4836 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4838 (unsigned long)zone_idx(zone),
4839 zone_start_pfn, (zone_start_pfn + size));
4841 zone_init_free_lists(zone);
4846 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4847 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4850 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4852 int __meminit __early_pfn_to_nid(unsigned long pfn,
4853 struct mminit_pfnnid_cache *state)
4855 unsigned long start_pfn, end_pfn;
4858 if (state->last_start <= pfn && pfn < state->last_end)
4859 return state->last_nid;
4861 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4863 state->last_start = start_pfn;
4864 state->last_end = end_pfn;
4865 state->last_nid = nid;
4870 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4873 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4874 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4875 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4877 * If an architecture guarantees that all ranges registered contain no holes
4878 * and may be freed, this this function may be used instead of calling
4879 * memblock_free_early_nid() manually.
4881 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4883 unsigned long start_pfn, end_pfn;
4886 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4887 start_pfn = min(start_pfn, max_low_pfn);
4888 end_pfn = min(end_pfn, max_low_pfn);
4890 if (start_pfn < end_pfn)
4891 memblock_free_early_nid(PFN_PHYS(start_pfn),
4892 (end_pfn - start_pfn) << PAGE_SHIFT,
4898 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4899 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4901 * If an architecture guarantees that all ranges registered contain no holes and may
4902 * be freed, this function may be used instead of calling memory_present() manually.
4904 void __init sparse_memory_present_with_active_regions(int nid)
4906 unsigned long start_pfn, end_pfn;
4909 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4910 memory_present(this_nid, start_pfn, end_pfn);
4914 * get_pfn_range_for_nid - Return the start and end page frames for a node
4915 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4916 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4917 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4919 * It returns the start and end page frame of a node based on information
4920 * provided by memblock_set_node(). If called for a node
4921 * with no available memory, a warning is printed and the start and end
4924 void __meminit get_pfn_range_for_nid(unsigned int nid,
4925 unsigned long *start_pfn, unsigned long *end_pfn)
4927 unsigned long this_start_pfn, this_end_pfn;
4933 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4934 *start_pfn = min(*start_pfn, this_start_pfn);
4935 *end_pfn = max(*end_pfn, this_end_pfn);
4938 if (*start_pfn == -1UL)
4943 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4944 * assumption is made that zones within a node are ordered in monotonic
4945 * increasing memory addresses so that the "highest" populated zone is used
4947 static void __init find_usable_zone_for_movable(void)
4950 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4951 if (zone_index == ZONE_MOVABLE)
4954 if (arch_zone_highest_possible_pfn[zone_index] >
4955 arch_zone_lowest_possible_pfn[zone_index])
4959 VM_BUG_ON(zone_index == -1);
4960 movable_zone = zone_index;
4964 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4965 * because it is sized independent of architecture. Unlike the other zones,
4966 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4967 * in each node depending on the size of each node and how evenly kernelcore
4968 * is distributed. This helper function adjusts the zone ranges
4969 * provided by the architecture for a given node by using the end of the
4970 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4971 * zones within a node are in order of monotonic increases memory addresses
4973 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4974 unsigned long zone_type,
4975 unsigned long node_start_pfn,
4976 unsigned long node_end_pfn,
4977 unsigned long *zone_start_pfn,
4978 unsigned long *zone_end_pfn)
4980 /* Only adjust if ZONE_MOVABLE is on this node */
4981 if (zone_movable_pfn[nid]) {
4982 /* Size ZONE_MOVABLE */
4983 if (zone_type == ZONE_MOVABLE) {
4984 *zone_start_pfn = zone_movable_pfn[nid];
4985 *zone_end_pfn = min(node_end_pfn,
4986 arch_zone_highest_possible_pfn[movable_zone]);
4988 /* Check if this whole range is within ZONE_MOVABLE */
4989 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4990 *zone_start_pfn = *zone_end_pfn;
4995 * Return the number of pages a zone spans in a node, including holes
4996 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4998 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4999 unsigned long zone_type,
5000 unsigned long node_start_pfn,
5001 unsigned long node_end_pfn,
5002 unsigned long *zone_start_pfn,
5003 unsigned long *zone_end_pfn,
5004 unsigned long *ignored)
5006 /* When hotadd a new node from cpu_up(), the node should be empty */
5007 if (!node_start_pfn && !node_end_pfn)
5010 /* Get the start and end of the zone */
5011 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5012 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5013 adjust_zone_range_for_zone_movable(nid, zone_type,
5014 node_start_pfn, node_end_pfn,
5015 zone_start_pfn, zone_end_pfn);
5017 /* Check that this node has pages within the zone's required range */
5018 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5021 /* Move the zone boundaries inside the node if necessary */
5022 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5023 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5025 /* Return the spanned pages */
5026 return *zone_end_pfn - *zone_start_pfn;
5030 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5031 * then all holes in the requested range will be accounted for.
5033 unsigned long __meminit __absent_pages_in_range(int nid,
5034 unsigned long range_start_pfn,
5035 unsigned long range_end_pfn)
5037 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5038 unsigned long start_pfn, end_pfn;
5041 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5042 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5043 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5044 nr_absent -= end_pfn - start_pfn;
5050 * absent_pages_in_range - Return number of page frames in holes within a range
5051 * @start_pfn: The start PFN to start searching for holes
5052 * @end_pfn: The end PFN to stop searching for holes
5054 * It returns the number of pages frames in memory holes within a range.
5056 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5057 unsigned long end_pfn)
5059 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5062 /* Return the number of page frames in holes in a zone on a node */
5063 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5064 unsigned long zone_type,
5065 unsigned long node_start_pfn,
5066 unsigned long node_end_pfn,
5067 unsigned long *ignored)
5069 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5070 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5071 unsigned long zone_start_pfn, zone_end_pfn;
5072 unsigned long nr_absent;
5074 /* When hotadd a new node from cpu_up(), the node should be empty */
5075 if (!node_start_pfn && !node_end_pfn)
5078 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5079 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5081 adjust_zone_range_for_zone_movable(nid, zone_type,
5082 node_start_pfn, node_end_pfn,
5083 &zone_start_pfn, &zone_end_pfn);
5084 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5087 * ZONE_MOVABLE handling.
5088 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5091 if (zone_movable_pfn[nid]) {
5092 if (mirrored_kernelcore) {
5093 unsigned long start_pfn, end_pfn;
5094 struct memblock_region *r;
5096 for_each_memblock(memory, r) {
5097 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5098 zone_start_pfn, zone_end_pfn);
5099 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5100 zone_start_pfn, zone_end_pfn);
5102 if (zone_type == ZONE_MOVABLE &&
5103 memblock_is_mirror(r))
5104 nr_absent += end_pfn - start_pfn;
5106 if (zone_type == ZONE_NORMAL &&
5107 !memblock_is_mirror(r))
5108 nr_absent += end_pfn - start_pfn;
5111 if (zone_type == ZONE_NORMAL)
5112 nr_absent += node_end_pfn - zone_movable_pfn[nid];
5119 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5120 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5121 unsigned long zone_type,
5122 unsigned long node_start_pfn,
5123 unsigned long node_end_pfn,
5124 unsigned long *zone_start_pfn,
5125 unsigned long *zone_end_pfn,
5126 unsigned long *zones_size)
5130 *zone_start_pfn = node_start_pfn;
5131 for (zone = 0; zone < zone_type; zone++)
5132 *zone_start_pfn += zones_size[zone];
5134 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5136 return zones_size[zone_type];
5139 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5140 unsigned long zone_type,
5141 unsigned long node_start_pfn,
5142 unsigned long node_end_pfn,
5143 unsigned long *zholes_size)
5148 return zholes_size[zone_type];
5151 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5153 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5154 unsigned long node_start_pfn,
5155 unsigned long node_end_pfn,
5156 unsigned long *zones_size,
5157 unsigned long *zholes_size)
5159 unsigned long realtotalpages = 0, totalpages = 0;
5162 for (i = 0; i < MAX_NR_ZONES; i++) {
5163 struct zone *zone = pgdat->node_zones + i;
5164 unsigned long zone_start_pfn, zone_end_pfn;
5165 unsigned long size, real_size;
5167 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5173 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5174 node_start_pfn, node_end_pfn,
5177 zone->zone_start_pfn = zone_start_pfn;
5179 zone->zone_start_pfn = 0;
5180 zone->spanned_pages = size;
5181 zone->present_pages = real_size;
5184 realtotalpages += real_size;
5187 pgdat->node_spanned_pages = totalpages;
5188 pgdat->node_present_pages = realtotalpages;
5189 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5193 #ifndef CONFIG_SPARSEMEM
5195 * Calculate the size of the zone->blockflags rounded to an unsigned long
5196 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5197 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5198 * round what is now in bits to nearest long in bits, then return it in
5201 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5203 unsigned long usemapsize;
5205 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5206 usemapsize = roundup(zonesize, pageblock_nr_pages);
5207 usemapsize = usemapsize >> pageblock_order;
5208 usemapsize *= NR_PAGEBLOCK_BITS;
5209 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5211 return usemapsize / 8;
5214 static void __init setup_usemap(struct pglist_data *pgdat,
5216 unsigned long zone_start_pfn,
5217 unsigned long zonesize)
5219 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5220 zone->pageblock_flags = NULL;
5222 zone->pageblock_flags =
5223 memblock_virt_alloc_node_nopanic(usemapsize,
5227 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5228 unsigned long zone_start_pfn, unsigned long zonesize) {}
5229 #endif /* CONFIG_SPARSEMEM */
5231 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5233 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5234 void __paginginit set_pageblock_order(void)
5238 /* Check that pageblock_nr_pages has not already been setup */
5239 if (pageblock_order)
5242 if (HPAGE_SHIFT > PAGE_SHIFT)
5243 order = HUGETLB_PAGE_ORDER;
5245 order = MAX_ORDER - 1;
5248 * Assume the largest contiguous order of interest is a huge page.
5249 * This value may be variable depending on boot parameters on IA64 and
5252 pageblock_order = order;
5254 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5257 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5258 * is unused as pageblock_order is set at compile-time. See
5259 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5262 void __paginginit set_pageblock_order(void)
5266 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5268 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5269 unsigned long present_pages)
5271 unsigned long pages = spanned_pages;
5274 * Provide a more accurate estimation if there are holes within
5275 * the zone and SPARSEMEM is in use. If there are holes within the
5276 * zone, each populated memory region may cost us one or two extra
5277 * memmap pages due to alignment because memmap pages for each
5278 * populated regions may not naturally algined on page boundary.
5279 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5281 if (spanned_pages > present_pages + (present_pages >> 4) &&
5282 IS_ENABLED(CONFIG_SPARSEMEM))
5283 pages = present_pages;
5285 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5289 * Set up the zone data structures:
5290 * - mark all pages reserved
5291 * - mark all memory queues empty
5292 * - clear the memory bitmaps
5294 * NOTE: pgdat should get zeroed by caller.
5296 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5299 int nid = pgdat->node_id;
5302 pgdat_resize_init(pgdat);
5303 #ifdef CONFIG_NUMA_BALANCING
5304 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5305 pgdat->numabalancing_migrate_nr_pages = 0;
5306 pgdat->numabalancing_migrate_next_window = jiffies;
5308 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5309 spin_lock_init(&pgdat->split_queue_lock);
5310 INIT_LIST_HEAD(&pgdat->split_queue);
5311 pgdat->split_queue_len = 0;
5313 init_waitqueue_head(&pgdat->kswapd_wait);
5314 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5315 pgdat_page_ext_init(pgdat);
5317 for (j = 0; j < MAX_NR_ZONES; j++) {
5318 struct zone *zone = pgdat->node_zones + j;
5319 unsigned long size, realsize, freesize, memmap_pages;
5320 unsigned long zone_start_pfn = zone->zone_start_pfn;
5322 size = zone->spanned_pages;
5323 realsize = freesize = zone->present_pages;
5326 * Adjust freesize so that it accounts for how much memory
5327 * is used by this zone for memmap. This affects the watermark
5328 * and per-cpu initialisations
5330 memmap_pages = calc_memmap_size(size, realsize);
5331 if (!is_highmem_idx(j)) {
5332 if (freesize >= memmap_pages) {
5333 freesize -= memmap_pages;
5336 " %s zone: %lu pages used for memmap\n",
5337 zone_names[j], memmap_pages);
5340 " %s zone: %lu pages exceeds freesize %lu\n",
5341 zone_names[j], memmap_pages, freesize);
5344 /* Account for reserved pages */
5345 if (j == 0 && freesize > dma_reserve) {
5346 freesize -= dma_reserve;
5347 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5348 zone_names[0], dma_reserve);
5351 if (!is_highmem_idx(j))
5352 nr_kernel_pages += freesize;
5353 /* Charge for highmem memmap if there are enough kernel pages */
5354 else if (nr_kernel_pages > memmap_pages * 2)
5355 nr_kernel_pages -= memmap_pages;
5356 nr_all_pages += freesize;
5359 * Set an approximate value for lowmem here, it will be adjusted
5360 * when the bootmem allocator frees pages into the buddy system.
5361 * And all highmem pages will be managed by the buddy system.
5363 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5366 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5368 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5370 zone->name = zone_names[j];
5371 spin_lock_init(&zone->lock);
5372 spin_lock_init(&zone->lru_lock);
5373 zone_seqlock_init(zone);
5374 zone->zone_pgdat = pgdat;
5375 zone_pcp_init(zone);
5377 /* For bootup, initialized properly in watermark setup */
5378 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5380 lruvec_init(&zone->lruvec);
5384 set_pageblock_order();
5385 setup_usemap(pgdat, zone, zone_start_pfn, size);
5386 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5388 memmap_init(size, nid, j, zone_start_pfn);
5392 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5394 unsigned long __maybe_unused start = 0;
5395 unsigned long __maybe_unused offset = 0;
5397 /* Skip empty nodes */
5398 if (!pgdat->node_spanned_pages)
5401 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5402 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5403 offset = pgdat->node_start_pfn - start;
5404 /* ia64 gets its own node_mem_map, before this, without bootmem */
5405 if (!pgdat->node_mem_map) {
5406 unsigned long size, end;
5410 * The zone's endpoints aren't required to be MAX_ORDER
5411 * aligned but the node_mem_map endpoints must be in order
5412 * for the buddy allocator to function correctly.
5414 end = pgdat_end_pfn(pgdat);
5415 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5416 size = (end - start) * sizeof(struct page);
5417 map = alloc_remap(pgdat->node_id, size);
5419 map = memblock_virt_alloc_node_nopanic(size,
5421 pgdat->node_mem_map = map + offset;
5423 #ifndef CONFIG_NEED_MULTIPLE_NODES
5425 * With no DISCONTIG, the global mem_map is just set as node 0's
5427 if (pgdat == NODE_DATA(0)) {
5428 mem_map = NODE_DATA(0)->node_mem_map;
5429 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5430 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5432 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5435 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5438 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5439 unsigned long node_start_pfn, unsigned long *zholes_size)
5441 pg_data_t *pgdat = NODE_DATA(nid);
5442 unsigned long start_pfn = 0;
5443 unsigned long end_pfn = 0;
5445 /* pg_data_t should be reset to zero when it's allocated */
5446 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5448 reset_deferred_meminit(pgdat);
5449 pgdat->node_id = nid;
5450 pgdat->node_start_pfn = node_start_pfn;
5451 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5452 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5453 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5454 (u64)start_pfn << PAGE_SHIFT,
5455 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5457 start_pfn = node_start_pfn;
5459 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5460 zones_size, zholes_size);
5462 alloc_node_mem_map(pgdat);
5463 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5464 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5465 nid, (unsigned long)pgdat,
5466 (unsigned long)pgdat->node_mem_map);
5469 free_area_init_core(pgdat);
5472 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5474 #if MAX_NUMNODES > 1
5476 * Figure out the number of possible node ids.
5478 void __init setup_nr_node_ids(void)
5480 unsigned int highest;
5482 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5483 nr_node_ids = highest + 1;
5488 * node_map_pfn_alignment - determine the maximum internode alignment
5490 * This function should be called after node map is populated and sorted.
5491 * It calculates the maximum power of two alignment which can distinguish
5494 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5495 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5496 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5497 * shifted, 1GiB is enough and this function will indicate so.
5499 * This is used to test whether pfn -> nid mapping of the chosen memory
5500 * model has fine enough granularity to avoid incorrect mapping for the
5501 * populated node map.
5503 * Returns the determined alignment in pfn's. 0 if there is no alignment
5504 * requirement (single node).
5506 unsigned long __init node_map_pfn_alignment(void)
5508 unsigned long accl_mask = 0, last_end = 0;
5509 unsigned long start, end, mask;
5513 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5514 if (!start || last_nid < 0 || last_nid == nid) {
5521 * Start with a mask granular enough to pin-point to the
5522 * start pfn and tick off bits one-by-one until it becomes
5523 * too coarse to separate the current node from the last.
5525 mask = ~((1 << __ffs(start)) - 1);
5526 while (mask && last_end <= (start & (mask << 1)))
5529 /* accumulate all internode masks */
5533 /* convert mask to number of pages */
5534 return ~accl_mask + 1;
5537 /* Find the lowest pfn for a node */
5538 static unsigned long __init find_min_pfn_for_node(int nid)
5540 unsigned long min_pfn = ULONG_MAX;
5541 unsigned long start_pfn;
5544 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5545 min_pfn = min(min_pfn, start_pfn);
5547 if (min_pfn == ULONG_MAX) {
5549 "Could not find start_pfn for node %d\n", nid);
5557 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5559 * It returns the minimum PFN based on information provided via
5560 * memblock_set_node().
5562 unsigned long __init find_min_pfn_with_active_regions(void)
5564 return find_min_pfn_for_node(MAX_NUMNODES);
5568 * early_calculate_totalpages()
5569 * Sum pages in active regions for movable zone.
5570 * Populate N_MEMORY for calculating usable_nodes.
5572 static unsigned long __init early_calculate_totalpages(void)
5574 unsigned long totalpages = 0;
5575 unsigned long start_pfn, end_pfn;
5578 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5579 unsigned long pages = end_pfn - start_pfn;
5581 totalpages += pages;
5583 node_set_state(nid, N_MEMORY);
5589 * Find the PFN the Movable zone begins in each node. Kernel memory
5590 * is spread evenly between nodes as long as the nodes have enough
5591 * memory. When they don't, some nodes will have more kernelcore than
5594 static void __init find_zone_movable_pfns_for_nodes(void)
5597 unsigned long usable_startpfn;
5598 unsigned long kernelcore_node, kernelcore_remaining;
5599 /* save the state before borrow the nodemask */
5600 nodemask_t saved_node_state = node_states[N_MEMORY];
5601 unsigned long totalpages = early_calculate_totalpages();
5602 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5603 struct memblock_region *r;
5605 /* Need to find movable_zone earlier when movable_node is specified. */
5606 find_usable_zone_for_movable();
5609 * If movable_node is specified, ignore kernelcore and movablecore
5612 if (movable_node_is_enabled()) {
5613 for_each_memblock(memory, r) {
5614 if (!memblock_is_hotpluggable(r))
5619 usable_startpfn = PFN_DOWN(r->base);
5620 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5621 min(usable_startpfn, zone_movable_pfn[nid]) :
5629 * If kernelcore=mirror is specified, ignore movablecore option
5631 if (mirrored_kernelcore) {
5632 bool mem_below_4gb_not_mirrored = false;
5634 for_each_memblock(memory, r) {
5635 if (memblock_is_mirror(r))
5640 usable_startpfn = memblock_region_memory_base_pfn(r);
5642 if (usable_startpfn < 0x100000) {
5643 mem_below_4gb_not_mirrored = true;
5647 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5648 min(usable_startpfn, zone_movable_pfn[nid]) :
5652 if (mem_below_4gb_not_mirrored)
5653 pr_warn("This configuration results in unmirrored kernel memory.");
5659 * If movablecore=nn[KMG] was specified, calculate what size of
5660 * kernelcore that corresponds so that memory usable for
5661 * any allocation type is evenly spread. If both kernelcore
5662 * and movablecore are specified, then the value of kernelcore
5663 * will be used for required_kernelcore if it's greater than
5664 * what movablecore would have allowed.
5666 if (required_movablecore) {
5667 unsigned long corepages;
5670 * Round-up so that ZONE_MOVABLE is at least as large as what
5671 * was requested by the user
5673 required_movablecore =
5674 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5675 required_movablecore = min(totalpages, required_movablecore);
5676 corepages = totalpages - required_movablecore;
5678 required_kernelcore = max(required_kernelcore, corepages);
5682 * If kernelcore was not specified or kernelcore size is larger
5683 * than totalpages, there is no ZONE_MOVABLE.
5685 if (!required_kernelcore || required_kernelcore >= totalpages)
5688 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5689 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5692 /* Spread kernelcore memory as evenly as possible throughout nodes */
5693 kernelcore_node = required_kernelcore / usable_nodes;
5694 for_each_node_state(nid, N_MEMORY) {
5695 unsigned long start_pfn, end_pfn;
5698 * Recalculate kernelcore_node if the division per node
5699 * now exceeds what is necessary to satisfy the requested
5700 * amount of memory for the kernel
5702 if (required_kernelcore < kernelcore_node)
5703 kernelcore_node = required_kernelcore / usable_nodes;
5706 * As the map is walked, we track how much memory is usable
5707 * by the kernel using kernelcore_remaining. When it is
5708 * 0, the rest of the node is usable by ZONE_MOVABLE
5710 kernelcore_remaining = kernelcore_node;
5712 /* Go through each range of PFNs within this node */
5713 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5714 unsigned long size_pages;
5716 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5717 if (start_pfn >= end_pfn)
5720 /* Account for what is only usable for kernelcore */
5721 if (start_pfn < usable_startpfn) {
5722 unsigned long kernel_pages;
5723 kernel_pages = min(end_pfn, usable_startpfn)
5726 kernelcore_remaining -= min(kernel_pages,
5727 kernelcore_remaining);
5728 required_kernelcore -= min(kernel_pages,
5729 required_kernelcore);
5731 /* Continue if range is now fully accounted */
5732 if (end_pfn <= usable_startpfn) {
5735 * Push zone_movable_pfn to the end so
5736 * that if we have to rebalance
5737 * kernelcore across nodes, we will
5738 * not double account here
5740 zone_movable_pfn[nid] = end_pfn;
5743 start_pfn = usable_startpfn;
5747 * The usable PFN range for ZONE_MOVABLE is from
5748 * start_pfn->end_pfn. Calculate size_pages as the
5749 * number of pages used as kernelcore
5751 size_pages = end_pfn - start_pfn;
5752 if (size_pages > kernelcore_remaining)
5753 size_pages = kernelcore_remaining;
5754 zone_movable_pfn[nid] = start_pfn + size_pages;
5757 * Some kernelcore has been met, update counts and
5758 * break if the kernelcore for this node has been
5761 required_kernelcore -= min(required_kernelcore,
5763 kernelcore_remaining -= size_pages;
5764 if (!kernelcore_remaining)
5770 * If there is still required_kernelcore, we do another pass with one
5771 * less node in the count. This will push zone_movable_pfn[nid] further
5772 * along on the nodes that still have memory until kernelcore is
5776 if (usable_nodes && required_kernelcore > usable_nodes)
5780 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5781 for (nid = 0; nid < MAX_NUMNODES; nid++)
5782 zone_movable_pfn[nid] =
5783 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5786 /* restore the node_state */
5787 node_states[N_MEMORY] = saved_node_state;
5790 /* Any regular or high memory on that node ? */
5791 static void check_for_memory(pg_data_t *pgdat, int nid)
5793 enum zone_type zone_type;
5795 if (N_MEMORY == N_NORMAL_MEMORY)
5798 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5799 struct zone *zone = &pgdat->node_zones[zone_type];
5800 if (populated_zone(zone)) {
5801 node_set_state(nid, N_HIGH_MEMORY);
5802 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5803 zone_type <= ZONE_NORMAL)
5804 node_set_state(nid, N_NORMAL_MEMORY);
5811 * free_area_init_nodes - Initialise all pg_data_t and zone data
5812 * @max_zone_pfn: an array of max PFNs for each zone
5814 * This will call free_area_init_node() for each active node in the system.
5815 * Using the page ranges provided by memblock_set_node(), the size of each
5816 * zone in each node and their holes is calculated. If the maximum PFN
5817 * between two adjacent zones match, it is assumed that the zone is empty.
5818 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5819 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5820 * starts where the previous one ended. For example, ZONE_DMA32 starts
5821 * at arch_max_dma_pfn.
5823 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5825 unsigned long start_pfn, end_pfn;
5828 /* Record where the zone boundaries are */
5829 memset(arch_zone_lowest_possible_pfn, 0,
5830 sizeof(arch_zone_lowest_possible_pfn));
5831 memset(arch_zone_highest_possible_pfn, 0,
5832 sizeof(arch_zone_highest_possible_pfn));
5833 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5834 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5835 for (i = 1; i < MAX_NR_ZONES; i++) {
5836 if (i == ZONE_MOVABLE)
5838 arch_zone_lowest_possible_pfn[i] =
5839 arch_zone_highest_possible_pfn[i-1];
5840 arch_zone_highest_possible_pfn[i] =
5841 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5843 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5844 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5846 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5847 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5848 find_zone_movable_pfns_for_nodes();
5850 /* Print out the zone ranges */
5851 pr_info("Zone ranges:\n");
5852 for (i = 0; i < MAX_NR_ZONES; i++) {
5853 if (i == ZONE_MOVABLE)
5855 pr_info(" %-8s ", zone_names[i]);
5856 if (arch_zone_lowest_possible_pfn[i] ==
5857 arch_zone_highest_possible_pfn[i])
5860 pr_cont("[mem %#018Lx-%#018Lx]\n",
5861 (u64)arch_zone_lowest_possible_pfn[i]
5863 ((u64)arch_zone_highest_possible_pfn[i]
5864 << PAGE_SHIFT) - 1);
5867 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5868 pr_info("Movable zone start for each node\n");
5869 for (i = 0; i < MAX_NUMNODES; i++) {
5870 if (zone_movable_pfn[i])
5871 pr_info(" Node %d: %#018Lx\n", i,
5872 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5875 /* Print out the early node map */
5876 pr_info("Early memory node ranges\n");
5877 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5878 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5879 (u64)start_pfn << PAGE_SHIFT,
5880 ((u64)end_pfn << PAGE_SHIFT) - 1);
5882 /* Initialise every node */
5883 mminit_verify_pageflags_layout();
5884 setup_nr_node_ids();
5885 for_each_online_node(nid) {
5886 pg_data_t *pgdat = NODE_DATA(nid);
5887 free_area_init_node(nid, NULL,
5888 find_min_pfn_for_node(nid), NULL);
5890 /* Any memory on that node */
5891 if (pgdat->node_present_pages)
5892 node_set_state(nid, N_MEMORY);
5893 check_for_memory(pgdat, nid);
5897 static int __init cmdline_parse_core(char *p, unsigned long *core)
5899 unsigned long long coremem;
5903 coremem = memparse(p, &p);
5904 *core = coremem >> PAGE_SHIFT;
5906 /* Paranoid check that UL is enough for the coremem value */
5907 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5913 * kernelcore=size sets the amount of memory for use for allocations that
5914 * cannot be reclaimed or migrated.
5916 static int __init cmdline_parse_kernelcore(char *p)
5918 /* parse kernelcore=mirror */
5919 if (parse_option_str(p, "mirror")) {
5920 mirrored_kernelcore = true;
5924 return cmdline_parse_core(p, &required_kernelcore);
5928 * movablecore=size sets the amount of memory for use for allocations that
5929 * can be reclaimed or migrated.
5931 static int __init cmdline_parse_movablecore(char *p)
5933 return cmdline_parse_core(p, &required_movablecore);
5936 early_param("kernelcore", cmdline_parse_kernelcore);
5937 early_param("movablecore", cmdline_parse_movablecore);
5939 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5941 void adjust_managed_page_count(struct page *page, long count)
5943 spin_lock(&managed_page_count_lock);
5944 page_zone(page)->managed_pages += count;
5945 totalram_pages += count;
5946 #ifdef CONFIG_HIGHMEM
5947 if (PageHighMem(page))
5948 totalhigh_pages += count;
5950 spin_unlock(&managed_page_count_lock);
5952 EXPORT_SYMBOL(adjust_managed_page_count);
5954 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5957 unsigned long pages = 0;
5959 start = (void *)PAGE_ALIGN((unsigned long)start);
5960 end = (void *)((unsigned long)end & PAGE_MASK);
5961 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5962 if ((unsigned int)poison <= 0xFF)
5963 memset(pos, poison, PAGE_SIZE);
5964 free_reserved_page(virt_to_page(pos));
5968 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5969 s, pages << (PAGE_SHIFT - 10), start, end);
5973 EXPORT_SYMBOL(free_reserved_area);
5975 #ifdef CONFIG_HIGHMEM
5976 void free_highmem_page(struct page *page)
5978 __free_reserved_page(page);
5980 page_zone(page)->managed_pages++;
5986 void __init mem_init_print_info(const char *str)
5988 unsigned long physpages, codesize, datasize, rosize, bss_size;
5989 unsigned long init_code_size, init_data_size;
5991 physpages = get_num_physpages();
5992 codesize = _etext - _stext;
5993 datasize = _edata - _sdata;
5994 rosize = __end_rodata - __start_rodata;
5995 bss_size = __bss_stop - __bss_start;
5996 init_data_size = __init_end - __init_begin;
5997 init_code_size = _einittext - _sinittext;
6000 * Detect special cases and adjust section sizes accordingly:
6001 * 1) .init.* may be embedded into .data sections
6002 * 2) .init.text.* may be out of [__init_begin, __init_end],
6003 * please refer to arch/tile/kernel/vmlinux.lds.S.
6004 * 3) .rodata.* may be embedded into .text or .data sections.
6006 #define adj_init_size(start, end, size, pos, adj) \
6008 if (start <= pos && pos < end && size > adj) \
6012 adj_init_size(__init_begin, __init_end, init_data_size,
6013 _sinittext, init_code_size);
6014 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6015 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6016 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6017 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6019 #undef adj_init_size
6021 pr_info("Memory: %luK/%luK available "
6022 "(%luK kernel code, %luK rwdata, %luK rodata, "
6023 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
6024 #ifdef CONFIG_HIGHMEM
6028 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
6029 codesize >> 10, datasize >> 10, rosize >> 10,
6030 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6031 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
6032 totalcma_pages << (PAGE_SHIFT-10),
6033 #ifdef CONFIG_HIGHMEM
6034 totalhigh_pages << (PAGE_SHIFT-10),
6036 str ? ", " : "", str ? str : "");
6040 * set_dma_reserve - set the specified number of pages reserved in the first zone
6041 * @new_dma_reserve: The number of pages to mark reserved
6043 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6044 * In the DMA zone, a significant percentage may be consumed by kernel image
6045 * and other unfreeable allocations which can skew the watermarks badly. This
6046 * function may optionally be used to account for unfreeable pages in the
6047 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6048 * smaller per-cpu batchsize.
6050 void __init set_dma_reserve(unsigned long new_dma_reserve)
6052 dma_reserve = new_dma_reserve;
6055 void __init free_area_init(unsigned long *zones_size)
6057 free_area_init_node(0, zones_size,
6058 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6061 static int page_alloc_cpu_notify(struct notifier_block *self,
6062 unsigned long action, void *hcpu)
6064 int cpu = (unsigned long)hcpu;
6066 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6067 lru_add_drain_cpu(cpu);
6071 * Spill the event counters of the dead processor
6072 * into the current processors event counters.
6073 * This artificially elevates the count of the current
6076 vm_events_fold_cpu(cpu);
6079 * Zero the differential counters of the dead processor
6080 * so that the vm statistics are consistent.
6082 * This is only okay since the processor is dead and cannot
6083 * race with what we are doing.
6085 cpu_vm_stats_fold(cpu);
6090 void __init page_alloc_init(void)
6092 hotcpu_notifier(page_alloc_cpu_notify, 0);
6096 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6097 * or min_free_kbytes changes.
6099 static void calculate_totalreserve_pages(void)
6101 struct pglist_data *pgdat;
6102 unsigned long reserve_pages = 0;
6103 enum zone_type i, j;
6105 for_each_online_pgdat(pgdat) {
6106 for (i = 0; i < MAX_NR_ZONES; i++) {
6107 struct zone *zone = pgdat->node_zones + i;
6110 /* Find valid and maximum lowmem_reserve in the zone */
6111 for (j = i; j < MAX_NR_ZONES; j++) {
6112 if (zone->lowmem_reserve[j] > max)
6113 max = zone->lowmem_reserve[j];
6116 /* we treat the high watermark as reserved pages. */
6117 max += high_wmark_pages(zone);
6119 if (max > zone->managed_pages)
6120 max = zone->managed_pages;
6122 zone->totalreserve_pages = max;
6124 reserve_pages += max;
6127 totalreserve_pages = reserve_pages;
6131 * setup_per_zone_lowmem_reserve - called whenever
6132 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6133 * has a correct pages reserved value, so an adequate number of
6134 * pages are left in the zone after a successful __alloc_pages().
6136 static void setup_per_zone_lowmem_reserve(void)
6138 struct pglist_data *pgdat;
6139 enum zone_type j, idx;
6141 for_each_online_pgdat(pgdat) {
6142 for (j = 0; j < MAX_NR_ZONES; j++) {
6143 struct zone *zone = pgdat->node_zones + j;
6144 unsigned long managed_pages = zone->managed_pages;
6146 zone->lowmem_reserve[j] = 0;
6150 struct zone *lower_zone;
6154 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6155 sysctl_lowmem_reserve_ratio[idx] = 1;
6157 lower_zone = pgdat->node_zones + idx;
6158 lower_zone->lowmem_reserve[j] = managed_pages /
6159 sysctl_lowmem_reserve_ratio[idx];
6160 managed_pages += lower_zone->managed_pages;
6165 /* update totalreserve_pages */
6166 calculate_totalreserve_pages();
6169 static void __setup_per_zone_wmarks(void)
6171 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6172 unsigned long lowmem_pages = 0;
6174 unsigned long flags;
6176 /* Calculate total number of !ZONE_HIGHMEM pages */
6177 for_each_zone(zone) {
6178 if (!is_highmem(zone))
6179 lowmem_pages += zone->managed_pages;
6182 for_each_zone(zone) {
6185 spin_lock_irqsave(&zone->lock, flags);
6186 tmp = (u64)pages_min * zone->managed_pages;
6187 do_div(tmp, lowmem_pages);
6188 if (is_highmem(zone)) {
6190 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6191 * need highmem pages, so cap pages_min to a small
6194 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6195 * deltas control asynch page reclaim, and so should
6196 * not be capped for highmem.
6198 unsigned long min_pages;
6200 min_pages = zone->managed_pages / 1024;
6201 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6202 zone->watermark[WMARK_MIN] = min_pages;
6205 * If it's a lowmem zone, reserve a number of pages
6206 * proportionate to the zone's size.
6208 zone->watermark[WMARK_MIN] = tmp;
6211 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
6212 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
6214 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6215 high_wmark_pages(zone) - low_wmark_pages(zone) -
6216 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6218 spin_unlock_irqrestore(&zone->lock, flags);
6221 /* update totalreserve_pages */
6222 calculate_totalreserve_pages();
6226 * setup_per_zone_wmarks - called when min_free_kbytes changes
6227 * or when memory is hot-{added|removed}
6229 * Ensures that the watermark[min,low,high] values for each zone are set
6230 * correctly with respect to min_free_kbytes.
6232 void setup_per_zone_wmarks(void)
6234 mutex_lock(&zonelists_mutex);
6235 __setup_per_zone_wmarks();
6236 mutex_unlock(&zonelists_mutex);
6240 * The inactive anon list should be small enough that the VM never has to
6241 * do too much work, but large enough that each inactive page has a chance
6242 * to be referenced again before it is swapped out.
6244 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6245 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6246 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6247 * the anonymous pages are kept on the inactive list.
6250 * memory ratio inactive anon
6251 * -------------------------------------
6260 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6262 unsigned int gb, ratio;
6264 /* Zone size in gigabytes */
6265 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6267 ratio = int_sqrt(10 * gb);
6271 zone->inactive_ratio = ratio;
6274 static void __meminit setup_per_zone_inactive_ratio(void)
6279 calculate_zone_inactive_ratio(zone);
6283 * Initialise min_free_kbytes.
6285 * For small machines we want it small (128k min). For large machines
6286 * we want it large (64MB max). But it is not linear, because network
6287 * bandwidth does not increase linearly with machine size. We use
6289 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6290 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6306 int __meminit init_per_zone_wmark_min(void)
6308 unsigned long lowmem_kbytes;
6309 int new_min_free_kbytes;
6311 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6312 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6314 if (new_min_free_kbytes > user_min_free_kbytes) {
6315 min_free_kbytes = new_min_free_kbytes;
6316 if (min_free_kbytes < 128)
6317 min_free_kbytes = 128;
6318 if (min_free_kbytes > 65536)
6319 min_free_kbytes = 65536;
6321 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6322 new_min_free_kbytes, user_min_free_kbytes);
6324 setup_per_zone_wmarks();
6325 refresh_zone_stat_thresholds();
6326 setup_per_zone_lowmem_reserve();
6327 setup_per_zone_inactive_ratio();
6330 module_init(init_per_zone_wmark_min)
6333 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6334 * that we can call two helper functions whenever min_free_kbytes
6337 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6338 void __user *buffer, size_t *length, loff_t *ppos)
6342 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6347 user_min_free_kbytes = min_free_kbytes;
6348 setup_per_zone_wmarks();
6354 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6355 void __user *buffer, size_t *length, loff_t *ppos)
6360 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6365 zone->min_unmapped_pages = (zone->managed_pages *
6366 sysctl_min_unmapped_ratio) / 100;
6370 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6371 void __user *buffer, size_t *length, loff_t *ppos)
6376 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6381 zone->min_slab_pages = (zone->managed_pages *
6382 sysctl_min_slab_ratio) / 100;
6388 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6389 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6390 * whenever sysctl_lowmem_reserve_ratio changes.
6392 * The reserve ratio obviously has absolutely no relation with the
6393 * minimum watermarks. The lowmem reserve ratio can only make sense
6394 * if in function of the boot time zone sizes.
6396 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6397 void __user *buffer, size_t *length, loff_t *ppos)
6399 proc_dointvec_minmax(table, write, buffer, length, ppos);
6400 setup_per_zone_lowmem_reserve();
6405 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6406 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6407 * pagelist can have before it gets flushed back to buddy allocator.
6409 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6410 void __user *buffer, size_t *length, loff_t *ppos)
6413 int old_percpu_pagelist_fraction;
6416 mutex_lock(&pcp_batch_high_lock);
6417 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6419 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6420 if (!write || ret < 0)
6423 /* Sanity checking to avoid pcp imbalance */
6424 if (percpu_pagelist_fraction &&
6425 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6426 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6432 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6435 for_each_populated_zone(zone) {
6438 for_each_possible_cpu(cpu)
6439 pageset_set_high_and_batch(zone,
6440 per_cpu_ptr(zone->pageset, cpu));
6443 mutex_unlock(&pcp_batch_high_lock);
6448 int hashdist = HASHDIST_DEFAULT;
6450 static int __init set_hashdist(char *str)
6454 hashdist = simple_strtoul(str, &str, 0);
6457 __setup("hashdist=", set_hashdist);
6461 * allocate a large system hash table from bootmem
6462 * - it is assumed that the hash table must contain an exact power-of-2
6463 * quantity of entries
6464 * - limit is the number of hash buckets, not the total allocation size
6466 void *__init alloc_large_system_hash(const char *tablename,
6467 unsigned long bucketsize,
6468 unsigned long numentries,
6471 unsigned int *_hash_shift,
6472 unsigned int *_hash_mask,
6473 unsigned long low_limit,
6474 unsigned long high_limit)
6476 unsigned long long max = high_limit;
6477 unsigned long log2qty, size;
6480 /* allow the kernel cmdline to have a say */
6482 /* round applicable memory size up to nearest megabyte */
6483 numentries = nr_kernel_pages;
6485 /* It isn't necessary when PAGE_SIZE >= 1MB */
6486 if (PAGE_SHIFT < 20)
6487 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6489 /* limit to 1 bucket per 2^scale bytes of low memory */
6490 if (scale > PAGE_SHIFT)
6491 numentries >>= (scale - PAGE_SHIFT);
6493 numentries <<= (PAGE_SHIFT - scale);
6495 /* Make sure we've got at least a 0-order allocation.. */
6496 if (unlikely(flags & HASH_SMALL)) {
6497 /* Makes no sense without HASH_EARLY */
6498 WARN_ON(!(flags & HASH_EARLY));
6499 if (!(numentries >> *_hash_shift)) {
6500 numentries = 1UL << *_hash_shift;
6501 BUG_ON(!numentries);
6503 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6504 numentries = PAGE_SIZE / bucketsize;
6506 numentries = roundup_pow_of_two(numentries);
6508 /* limit allocation size to 1/16 total memory by default */
6510 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6511 do_div(max, bucketsize);
6513 max = min(max, 0x80000000ULL);
6515 if (numentries < low_limit)
6516 numentries = low_limit;
6517 if (numentries > max)
6520 log2qty = ilog2(numentries);
6523 size = bucketsize << log2qty;
6524 if (flags & HASH_EARLY)
6525 table = memblock_virt_alloc_nopanic(size, 0);
6527 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6530 * If bucketsize is not a power-of-two, we may free
6531 * some pages at the end of hash table which
6532 * alloc_pages_exact() automatically does
6534 if (get_order(size) < MAX_ORDER) {
6535 table = alloc_pages_exact(size, GFP_ATOMIC);
6536 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6539 } while (!table && size > PAGE_SIZE && --log2qty);
6542 panic("Failed to allocate %s hash table\n", tablename);
6544 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6547 ilog2(size) - PAGE_SHIFT,
6551 *_hash_shift = log2qty;
6553 *_hash_mask = (1 << log2qty) - 1;
6558 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6559 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6562 #ifdef CONFIG_SPARSEMEM
6563 return __pfn_to_section(pfn)->pageblock_flags;
6565 return zone->pageblock_flags;
6566 #endif /* CONFIG_SPARSEMEM */
6569 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6571 #ifdef CONFIG_SPARSEMEM
6572 pfn &= (PAGES_PER_SECTION-1);
6573 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6575 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6576 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6577 #endif /* CONFIG_SPARSEMEM */
6581 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6582 * @page: The page within the block of interest
6583 * @pfn: The target page frame number
6584 * @end_bitidx: The last bit of interest to retrieve
6585 * @mask: mask of bits that the caller is interested in
6587 * Return: pageblock_bits flags
6589 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6590 unsigned long end_bitidx,
6594 unsigned long *bitmap;
6595 unsigned long bitidx, word_bitidx;
6598 zone = page_zone(page);
6599 bitmap = get_pageblock_bitmap(zone, pfn);
6600 bitidx = pfn_to_bitidx(zone, pfn);
6601 word_bitidx = bitidx / BITS_PER_LONG;
6602 bitidx &= (BITS_PER_LONG-1);
6604 word = bitmap[word_bitidx];
6605 bitidx += end_bitidx;
6606 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6610 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6611 * @page: The page within the block of interest
6612 * @flags: The flags to set
6613 * @pfn: The target page frame number
6614 * @end_bitidx: The last bit of interest
6615 * @mask: mask of bits that the caller is interested in
6617 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6619 unsigned long end_bitidx,
6623 unsigned long *bitmap;
6624 unsigned long bitidx, word_bitidx;
6625 unsigned long old_word, word;
6627 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6629 zone = page_zone(page);
6630 bitmap = get_pageblock_bitmap(zone, pfn);
6631 bitidx = pfn_to_bitidx(zone, pfn);
6632 word_bitidx = bitidx / BITS_PER_LONG;
6633 bitidx &= (BITS_PER_LONG-1);
6635 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6637 bitidx += end_bitidx;
6638 mask <<= (BITS_PER_LONG - bitidx - 1);
6639 flags <<= (BITS_PER_LONG - bitidx - 1);
6641 word = READ_ONCE(bitmap[word_bitidx]);
6643 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6644 if (word == old_word)
6651 * This function checks whether pageblock includes unmovable pages or not.
6652 * If @count is not zero, it is okay to include less @count unmovable pages
6654 * PageLRU check without isolation or lru_lock could race so that
6655 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6656 * expect this function should be exact.
6658 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6659 bool skip_hwpoisoned_pages)
6661 unsigned long pfn, iter, found;
6665 * For avoiding noise data, lru_add_drain_all() should be called
6666 * If ZONE_MOVABLE, the zone never contains unmovable pages
6668 if (zone_idx(zone) == ZONE_MOVABLE)
6670 mt = get_pageblock_migratetype(page);
6671 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6674 pfn = page_to_pfn(page);
6675 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6676 unsigned long check = pfn + iter;
6678 if (!pfn_valid_within(check))
6681 page = pfn_to_page(check);
6684 * Hugepages are not in LRU lists, but they're movable.
6685 * We need not scan over tail pages bacause we don't
6686 * handle each tail page individually in migration.
6688 if (PageHuge(page)) {
6689 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6694 * We can't use page_count without pin a page
6695 * because another CPU can free compound page.
6696 * This check already skips compound tails of THP
6697 * because their page->_count is zero at all time.
6699 if (!atomic_read(&page->_count)) {
6700 if (PageBuddy(page))
6701 iter += (1 << page_order(page)) - 1;
6706 * The HWPoisoned page may be not in buddy system, and
6707 * page_count() is not 0.
6709 if (skip_hwpoisoned_pages && PageHWPoison(page))
6715 * If there are RECLAIMABLE pages, we need to check
6716 * it. But now, memory offline itself doesn't call
6717 * shrink_node_slabs() and it still to be fixed.
6720 * If the page is not RAM, page_count()should be 0.
6721 * we don't need more check. This is an _used_ not-movable page.
6723 * The problematic thing here is PG_reserved pages. PG_reserved
6724 * is set to both of a memory hole page and a _used_ kernel
6733 bool is_pageblock_removable_nolock(struct page *page)
6739 * We have to be careful here because we are iterating over memory
6740 * sections which are not zone aware so we might end up outside of
6741 * the zone but still within the section.
6742 * We have to take care about the node as well. If the node is offline
6743 * its NODE_DATA will be NULL - see page_zone.
6745 if (!node_online(page_to_nid(page)))
6748 zone = page_zone(page);
6749 pfn = page_to_pfn(page);
6750 if (!zone_spans_pfn(zone, pfn))
6753 return !has_unmovable_pages(zone, page, 0, true);
6756 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
6758 static unsigned long pfn_max_align_down(unsigned long pfn)
6760 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6761 pageblock_nr_pages) - 1);
6764 static unsigned long pfn_max_align_up(unsigned long pfn)
6766 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6767 pageblock_nr_pages));
6770 /* [start, end) must belong to a single zone. */
6771 static int __alloc_contig_migrate_range(struct compact_control *cc,
6772 unsigned long start, unsigned long end)
6774 /* This function is based on compact_zone() from compaction.c. */
6775 unsigned long nr_reclaimed;
6776 unsigned long pfn = start;
6777 unsigned int tries = 0;
6782 while (pfn < end || !list_empty(&cc->migratepages)) {
6783 if (fatal_signal_pending(current)) {
6788 if (list_empty(&cc->migratepages)) {
6789 cc->nr_migratepages = 0;
6790 pfn = isolate_migratepages_range(cc, pfn, end);
6796 } else if (++tries == 5) {
6797 ret = ret < 0 ? ret : -EBUSY;
6801 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6803 cc->nr_migratepages -= nr_reclaimed;
6805 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6806 NULL, 0, cc->mode, MR_CMA);
6809 putback_movable_pages(&cc->migratepages);
6816 * alloc_contig_range() -- tries to allocate given range of pages
6817 * @start: start PFN to allocate
6818 * @end: one-past-the-last PFN to allocate
6819 * @migratetype: migratetype of the underlaying pageblocks (either
6820 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6821 * in range must have the same migratetype and it must
6822 * be either of the two.
6824 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6825 * aligned, however it's the caller's responsibility to guarantee that
6826 * we are the only thread that changes migrate type of pageblocks the
6829 * The PFN range must belong to a single zone.
6831 * Returns zero on success or negative error code. On success all
6832 * pages which PFN is in [start, end) are allocated for the caller and
6833 * need to be freed with free_contig_range().
6835 int alloc_contig_range(unsigned long start, unsigned long end,
6836 unsigned migratetype)
6838 unsigned long outer_start, outer_end;
6842 struct compact_control cc = {
6843 .nr_migratepages = 0,
6845 .zone = page_zone(pfn_to_page(start)),
6846 .mode = MIGRATE_SYNC,
6847 .ignore_skip_hint = true,
6849 INIT_LIST_HEAD(&cc.migratepages);
6852 * What we do here is we mark all pageblocks in range as
6853 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6854 * have different sizes, and due to the way page allocator
6855 * work, we align the range to biggest of the two pages so
6856 * that page allocator won't try to merge buddies from
6857 * different pageblocks and change MIGRATE_ISOLATE to some
6858 * other migration type.
6860 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6861 * migrate the pages from an unaligned range (ie. pages that
6862 * we are interested in). This will put all the pages in
6863 * range back to page allocator as MIGRATE_ISOLATE.
6865 * When this is done, we take the pages in range from page
6866 * allocator removing them from the buddy system. This way
6867 * page allocator will never consider using them.
6869 * This lets us mark the pageblocks back as
6870 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6871 * aligned range but not in the unaligned, original range are
6872 * put back to page allocator so that buddy can use them.
6875 ret = start_isolate_page_range(pfn_max_align_down(start),
6876 pfn_max_align_up(end), migratetype,
6882 * In case of -EBUSY, we'd like to know which page causes problem.
6883 * So, just fall through. We will check it in test_pages_isolated().
6885 ret = __alloc_contig_migrate_range(&cc, start, end);
6886 if (ret && ret != -EBUSY)
6890 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6891 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6892 * more, all pages in [start, end) are free in page allocator.
6893 * What we are going to do is to allocate all pages from
6894 * [start, end) (that is remove them from page allocator).
6896 * The only problem is that pages at the beginning and at the
6897 * end of interesting range may be not aligned with pages that
6898 * page allocator holds, ie. they can be part of higher order
6899 * pages. Because of this, we reserve the bigger range and
6900 * once this is done free the pages we are not interested in.
6902 * We don't have to hold zone->lock here because the pages are
6903 * isolated thus they won't get removed from buddy.
6906 lru_add_drain_all();
6907 drain_all_pages(cc.zone);
6910 outer_start = start;
6911 while (!PageBuddy(pfn_to_page(outer_start))) {
6912 if (++order >= MAX_ORDER) {
6913 outer_start = start;
6916 outer_start &= ~0UL << order;
6919 if (outer_start != start) {
6920 order = page_order(pfn_to_page(outer_start));
6923 * outer_start page could be small order buddy page and
6924 * it doesn't include start page. Adjust outer_start
6925 * in this case to report failed page properly
6926 * on tracepoint in test_pages_isolated()
6928 if (outer_start + (1UL << order) <= start)
6929 outer_start = start;
6932 /* Make sure the range is really isolated. */
6933 if (test_pages_isolated(outer_start, end, false)) {
6934 pr_info("%s: [%lx, %lx) PFNs busy\n",
6935 __func__, outer_start, end);
6940 /* Grab isolated pages from freelists. */
6941 outer_end = isolate_freepages_range(&cc, outer_start, end);
6947 /* Free head and tail (if any) */
6948 if (start != outer_start)
6949 free_contig_range(outer_start, start - outer_start);
6950 if (end != outer_end)
6951 free_contig_range(end, outer_end - end);
6954 undo_isolate_page_range(pfn_max_align_down(start),
6955 pfn_max_align_up(end), migratetype);
6959 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6961 unsigned int count = 0;
6963 for (; nr_pages--; pfn++) {
6964 struct page *page = pfn_to_page(pfn);
6966 count += page_count(page) != 1;
6969 WARN(count != 0, "%d pages are still in use!\n", count);
6973 #ifdef CONFIG_MEMORY_HOTPLUG
6975 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6976 * page high values need to be recalulated.
6978 void __meminit zone_pcp_update(struct zone *zone)
6981 mutex_lock(&pcp_batch_high_lock);
6982 for_each_possible_cpu(cpu)
6983 pageset_set_high_and_batch(zone,
6984 per_cpu_ptr(zone->pageset, cpu));
6985 mutex_unlock(&pcp_batch_high_lock);
6989 void zone_pcp_reset(struct zone *zone)
6991 unsigned long flags;
6993 struct per_cpu_pageset *pset;
6995 /* avoid races with drain_pages() */
6996 local_irq_save(flags);
6997 if (zone->pageset != &boot_pageset) {
6998 for_each_online_cpu(cpu) {
6999 pset = per_cpu_ptr(zone->pageset, cpu);
7000 drain_zonestat(zone, pset);
7002 free_percpu(zone->pageset);
7003 zone->pageset = &boot_pageset;
7005 local_irq_restore(flags);
7008 #ifdef CONFIG_MEMORY_HOTREMOVE
7010 * All pages in the range must be isolated before calling this.
7013 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7017 unsigned int order, i;
7019 unsigned long flags;
7020 /* find the first valid pfn */
7021 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7026 zone = page_zone(pfn_to_page(pfn));
7027 spin_lock_irqsave(&zone->lock, flags);
7029 while (pfn < end_pfn) {
7030 if (!pfn_valid(pfn)) {
7034 page = pfn_to_page(pfn);
7036 * The HWPoisoned page may be not in buddy system, and
7037 * page_count() is not 0.
7039 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7041 SetPageReserved(page);
7045 BUG_ON(page_count(page));
7046 BUG_ON(!PageBuddy(page));
7047 order = page_order(page);
7048 #ifdef CONFIG_DEBUG_VM
7049 printk(KERN_INFO "remove from free list %lx %d %lx\n",
7050 pfn, 1 << order, end_pfn);
7052 list_del(&page->lru);
7053 rmv_page_order(page);
7054 zone->free_area[order].nr_free--;
7055 for (i = 0; i < (1 << order); i++)
7056 SetPageReserved((page+i));
7057 pfn += (1 << order);
7059 spin_unlock_irqrestore(&zone->lock, flags);
7063 #ifdef CONFIG_MEMORY_FAILURE
7064 bool is_free_buddy_page(struct page *page)
7066 struct zone *zone = page_zone(page);
7067 unsigned long pfn = page_to_pfn(page);
7068 unsigned long flags;
7071 spin_lock_irqsave(&zone->lock, flags);
7072 for (order = 0; order < MAX_ORDER; order++) {
7073 struct page *page_head = page - (pfn & ((1 << order) - 1));
7075 if (PageBuddy(page_head) && page_order(page_head) >= order)
7078 spin_unlock_irqrestore(&zone->lock, flags);
7080 return order < MAX_ORDER;