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 compound_page_dtor * const compound_page_dtors[] = {
229 #ifdef CONFIG_HUGETLB_PAGE
232 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
237 int min_free_kbytes = 1024;
238 int user_min_free_kbytes = -1;
240 static unsigned long __meminitdata nr_kernel_pages;
241 static unsigned long __meminitdata nr_all_pages;
242 static unsigned long __meminitdata dma_reserve;
244 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
245 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
246 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
247 static unsigned long __initdata required_kernelcore;
248 static unsigned long __initdata required_movablecore;
249 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
250 static bool mirrored_kernelcore;
252 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
254 EXPORT_SYMBOL(movable_zone);
255 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
258 int nr_node_ids __read_mostly = MAX_NUMNODES;
259 int nr_online_nodes __read_mostly = 1;
260 EXPORT_SYMBOL(nr_node_ids);
261 EXPORT_SYMBOL(nr_online_nodes);
264 int page_group_by_mobility_disabled __read_mostly;
266 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
267 static inline void reset_deferred_meminit(pg_data_t *pgdat)
269 pgdat->first_deferred_pfn = ULONG_MAX;
272 /* Returns true if the struct page for the pfn is uninitialised */
273 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
275 if (pfn >= NODE_DATA(early_pfn_to_nid(pfn))->first_deferred_pfn)
281 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
283 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
290 * Returns false when the remaining initialisation should be deferred until
291 * later in the boot cycle when it can be parallelised.
293 static inline bool update_defer_init(pg_data_t *pgdat,
294 unsigned long pfn, unsigned long zone_end,
295 unsigned long *nr_initialised)
297 /* Always populate low zones for address-contrained allocations */
298 if (zone_end < pgdat_end_pfn(pgdat))
301 /* Initialise at least 2G of the highest zone */
303 if (*nr_initialised > (2UL << (30 - PAGE_SHIFT)) &&
304 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
305 pgdat->first_deferred_pfn = pfn;
312 static inline void reset_deferred_meminit(pg_data_t *pgdat)
316 static inline bool early_page_uninitialised(unsigned long pfn)
321 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
326 static inline bool update_defer_init(pg_data_t *pgdat,
327 unsigned long pfn, unsigned long zone_end,
328 unsigned long *nr_initialised)
335 void set_pageblock_migratetype(struct page *page, int migratetype)
337 if (unlikely(page_group_by_mobility_disabled &&
338 migratetype < MIGRATE_PCPTYPES))
339 migratetype = MIGRATE_UNMOVABLE;
341 set_pageblock_flags_group(page, (unsigned long)migratetype,
342 PB_migrate, PB_migrate_end);
345 #ifdef CONFIG_DEBUG_VM
346 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
350 unsigned long pfn = page_to_pfn(page);
351 unsigned long sp, start_pfn;
354 seq = zone_span_seqbegin(zone);
355 start_pfn = zone->zone_start_pfn;
356 sp = zone->spanned_pages;
357 if (!zone_spans_pfn(zone, pfn))
359 } while (zone_span_seqretry(zone, seq));
362 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
363 pfn, zone_to_nid(zone), zone->name,
364 start_pfn, start_pfn + sp);
369 static int page_is_consistent(struct zone *zone, struct page *page)
371 if (!pfn_valid_within(page_to_pfn(page)))
373 if (zone != page_zone(page))
379 * Temporary debugging check for pages not lying within a given zone.
381 static int bad_range(struct zone *zone, struct page *page)
383 if (page_outside_zone_boundaries(zone, page))
385 if (!page_is_consistent(zone, page))
391 static inline int bad_range(struct zone *zone, struct page *page)
397 static void bad_page(struct page *page, const char *reason,
398 unsigned long bad_flags)
400 static unsigned long resume;
401 static unsigned long nr_shown;
402 static unsigned long nr_unshown;
404 /* Don't complain about poisoned pages */
405 if (PageHWPoison(page)) {
406 page_mapcount_reset(page); /* remove PageBuddy */
411 * Allow a burst of 60 reports, then keep quiet for that minute;
412 * or allow a steady drip of one report per second.
414 if (nr_shown == 60) {
415 if (time_before(jiffies, resume)) {
421 "BUG: Bad page state: %lu messages suppressed\n",
428 resume = jiffies + 60 * HZ;
430 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
431 current->comm, page_to_pfn(page));
432 dump_page_badflags(page, reason, bad_flags);
437 /* Leave bad fields for debug, except PageBuddy could make trouble */
438 page_mapcount_reset(page); /* remove PageBuddy */
439 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
443 * Higher-order pages are called "compound pages". They are structured thusly:
445 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
447 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
448 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
450 * The first tail page's ->compound_dtor holds the offset in array of compound
451 * page destructors. See compound_page_dtors.
453 * The first tail page's ->compound_order holds the order of allocation.
454 * This usage means that zero-order pages may not be compound.
457 void free_compound_page(struct page *page)
459 __free_pages_ok(page, compound_order(page));
462 void prep_compound_page(struct page *page, unsigned int order)
465 int nr_pages = 1 << order;
467 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
468 set_compound_order(page, order);
470 for (i = 1; i < nr_pages; i++) {
471 struct page *p = page + i;
472 set_page_count(p, 0);
473 p->mapping = TAIL_MAPPING;
474 set_compound_head(p, page);
476 atomic_set(compound_mapcount_ptr(page), -1);
479 #ifdef CONFIG_DEBUG_PAGEALLOC
480 unsigned int _debug_guardpage_minorder;
481 bool _debug_pagealloc_enabled __read_mostly
482 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
483 bool _debug_guardpage_enabled __read_mostly;
485 static int __init early_debug_pagealloc(char *buf)
490 if (strcmp(buf, "on") == 0)
491 _debug_pagealloc_enabled = true;
493 if (strcmp(buf, "off") == 0)
494 _debug_pagealloc_enabled = false;
498 early_param("debug_pagealloc", early_debug_pagealloc);
500 static bool need_debug_guardpage(void)
502 /* If we don't use debug_pagealloc, we don't need guard page */
503 if (!debug_pagealloc_enabled())
509 static void init_debug_guardpage(void)
511 if (!debug_pagealloc_enabled())
514 _debug_guardpage_enabled = true;
517 struct page_ext_operations debug_guardpage_ops = {
518 .need = need_debug_guardpage,
519 .init = init_debug_guardpage,
522 static int __init debug_guardpage_minorder_setup(char *buf)
526 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
527 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
530 _debug_guardpage_minorder = res;
531 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
534 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
536 static inline void set_page_guard(struct zone *zone, struct page *page,
537 unsigned int order, int migratetype)
539 struct page_ext *page_ext;
541 if (!debug_guardpage_enabled())
544 page_ext = lookup_page_ext(page);
545 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
547 INIT_LIST_HEAD(&page->lru);
548 set_page_private(page, order);
549 /* Guard pages are not available for any usage */
550 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
553 static inline void clear_page_guard(struct zone *zone, struct page *page,
554 unsigned int order, int migratetype)
556 struct page_ext *page_ext;
558 if (!debug_guardpage_enabled())
561 page_ext = lookup_page_ext(page);
562 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
564 set_page_private(page, 0);
565 if (!is_migrate_isolate(migratetype))
566 __mod_zone_freepage_state(zone, (1 << order), migratetype);
569 struct page_ext_operations debug_guardpage_ops = { NULL, };
570 static inline void set_page_guard(struct zone *zone, struct page *page,
571 unsigned int order, int migratetype) {}
572 static inline void clear_page_guard(struct zone *zone, struct page *page,
573 unsigned int order, int migratetype) {}
576 static inline void set_page_order(struct page *page, unsigned int order)
578 set_page_private(page, order);
579 __SetPageBuddy(page);
582 static inline void rmv_page_order(struct page *page)
584 __ClearPageBuddy(page);
585 set_page_private(page, 0);
589 * This function checks whether a page is free && is the buddy
590 * we can do coalesce a page and its buddy if
591 * (a) the buddy is not in a hole &&
592 * (b) the buddy is in the buddy system &&
593 * (c) a page and its buddy have the same order &&
594 * (d) a page and its buddy are in the same zone.
596 * For recording whether a page is in the buddy system, we set ->_mapcount
597 * PAGE_BUDDY_MAPCOUNT_VALUE.
598 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
599 * serialized by zone->lock.
601 * For recording page's order, we use page_private(page).
603 static inline int page_is_buddy(struct page *page, struct page *buddy,
606 if (!pfn_valid_within(page_to_pfn(buddy)))
609 if (page_is_guard(buddy) && page_order(buddy) == order) {
610 if (page_zone_id(page) != page_zone_id(buddy))
613 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
618 if (PageBuddy(buddy) && page_order(buddy) == order) {
620 * zone check is done late to avoid uselessly
621 * calculating zone/node ids for pages that could
624 if (page_zone_id(page) != page_zone_id(buddy))
627 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
635 * Freeing function for a buddy system allocator.
637 * The concept of a buddy system is to maintain direct-mapped table
638 * (containing bit values) for memory blocks of various "orders".
639 * The bottom level table contains the map for the smallest allocatable
640 * units of memory (here, pages), and each level above it describes
641 * pairs of units from the levels below, hence, "buddies".
642 * At a high level, all that happens here is marking the table entry
643 * at the bottom level available, and propagating the changes upward
644 * as necessary, plus some accounting needed to play nicely with other
645 * parts of the VM system.
646 * At each level, we keep a list of pages, which are heads of continuous
647 * free pages of length of (1 << order) and marked with _mapcount
648 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
650 * So when we are allocating or freeing one, we can derive the state of the
651 * other. That is, if we allocate a small block, and both were
652 * free, the remainder of the region must be split into blocks.
653 * If a block is freed, and its buddy is also free, then this
654 * triggers coalescing into a block of larger size.
659 static inline void __free_one_page(struct page *page,
661 struct zone *zone, unsigned int order,
664 unsigned long page_idx;
665 unsigned long combined_idx;
666 unsigned long uninitialized_var(buddy_idx);
668 unsigned int max_order = MAX_ORDER;
670 VM_BUG_ON(!zone_is_initialized(zone));
671 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
673 VM_BUG_ON(migratetype == -1);
674 if (is_migrate_isolate(migratetype)) {
676 * We restrict max order of merging to prevent merge
677 * between freepages on isolate pageblock and normal
678 * pageblock. Without this, pageblock isolation
679 * could cause incorrect freepage accounting.
681 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
683 __mod_zone_freepage_state(zone, 1 << order, migratetype);
686 page_idx = pfn & ((1 << max_order) - 1);
688 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
689 VM_BUG_ON_PAGE(bad_range(zone, page), page);
691 while (order < max_order - 1) {
692 buddy_idx = __find_buddy_index(page_idx, order);
693 buddy = page + (buddy_idx - page_idx);
694 if (!page_is_buddy(page, buddy, order))
697 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
698 * merge with it and move up one order.
700 if (page_is_guard(buddy)) {
701 clear_page_guard(zone, buddy, order, migratetype);
703 list_del(&buddy->lru);
704 zone->free_area[order].nr_free--;
705 rmv_page_order(buddy);
707 combined_idx = buddy_idx & page_idx;
708 page = page + (combined_idx - page_idx);
709 page_idx = combined_idx;
712 set_page_order(page, order);
715 * If this is not the largest possible page, check if the buddy
716 * of the next-highest order is free. If it is, it's possible
717 * that pages are being freed that will coalesce soon. In case,
718 * that is happening, add the free page to the tail of the list
719 * so it's less likely to be used soon and more likely to be merged
720 * as a higher order page
722 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
723 struct page *higher_page, *higher_buddy;
724 combined_idx = buddy_idx & page_idx;
725 higher_page = page + (combined_idx - page_idx);
726 buddy_idx = __find_buddy_index(combined_idx, order + 1);
727 higher_buddy = higher_page + (buddy_idx - combined_idx);
728 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
729 list_add_tail(&page->lru,
730 &zone->free_area[order].free_list[migratetype]);
735 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
737 zone->free_area[order].nr_free++;
740 static inline int free_pages_check(struct page *page)
742 const char *bad_reason = NULL;
743 unsigned long bad_flags = 0;
745 if (unlikely(atomic_read(&page->_mapcount) != -1))
746 bad_reason = "nonzero mapcount";
747 if (unlikely(page->mapping != NULL))
748 bad_reason = "non-NULL mapping";
749 if (unlikely(atomic_read(&page->_count) != 0))
750 bad_reason = "nonzero _count";
751 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
752 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
753 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
756 if (unlikely(page->mem_cgroup))
757 bad_reason = "page still charged to cgroup";
759 if (unlikely(bad_reason)) {
760 bad_page(page, bad_reason, bad_flags);
763 page_cpupid_reset_last(page);
764 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
765 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
770 * Frees a number of pages from the PCP lists
771 * Assumes all pages on list are in same zone, and of same order.
772 * count is the number of pages to free.
774 * If the zone was previously in an "all pages pinned" state then look to
775 * see if this freeing clears that state.
777 * And clear the zone's pages_scanned counter, to hold off the "all pages are
778 * pinned" detection logic.
780 static void free_pcppages_bulk(struct zone *zone, int count,
781 struct per_cpu_pages *pcp)
786 unsigned long nr_scanned;
788 spin_lock(&zone->lock);
789 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
791 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
795 struct list_head *list;
798 * Remove pages from lists in a round-robin fashion. A
799 * batch_free count is maintained that is incremented when an
800 * empty list is encountered. This is so more pages are freed
801 * off fuller lists instead of spinning excessively around empty
806 if (++migratetype == MIGRATE_PCPTYPES)
808 list = &pcp->lists[migratetype];
809 } while (list_empty(list));
811 /* This is the only non-empty list. Free them all. */
812 if (batch_free == MIGRATE_PCPTYPES)
813 batch_free = to_free;
816 int mt; /* migratetype of the to-be-freed page */
818 page = list_last_entry(list, struct page, lru);
819 /* must delete as __free_one_page list manipulates */
820 list_del(&page->lru);
822 mt = get_pcppage_migratetype(page);
823 /* MIGRATE_ISOLATE page should not go to pcplists */
824 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
825 /* Pageblock could have been isolated meanwhile */
826 if (unlikely(has_isolate_pageblock(zone)))
827 mt = get_pageblock_migratetype(page);
829 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
830 trace_mm_page_pcpu_drain(page, 0, mt);
831 } while (--to_free && --batch_free && !list_empty(list));
833 spin_unlock(&zone->lock);
836 static void free_one_page(struct zone *zone,
837 struct page *page, unsigned long pfn,
841 unsigned long nr_scanned;
842 spin_lock(&zone->lock);
843 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
845 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
847 if (unlikely(has_isolate_pageblock(zone) ||
848 is_migrate_isolate(migratetype))) {
849 migratetype = get_pfnblock_migratetype(page, pfn);
851 __free_one_page(page, pfn, zone, order, migratetype);
852 spin_unlock(&zone->lock);
855 static int free_tail_pages_check(struct page *head_page, struct page *page)
860 * We rely page->lru.next never has bit 0 set, unless the page
861 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
863 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
865 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
869 switch (page - head_page) {
871 /* the first tail page: ->mapping is compound_mapcount() */
872 if (unlikely(compound_mapcount(page))) {
873 bad_page(page, "nonzero compound_mapcount", 0);
879 * the second tail page: ->mapping is
880 * page_deferred_list().next -- ignore value.
884 if (page->mapping != TAIL_MAPPING) {
885 bad_page(page, "corrupted mapping in tail page", 0);
890 if (unlikely(!PageTail(page))) {
891 bad_page(page, "PageTail not set", 0);
894 if (unlikely(compound_head(page) != head_page)) {
895 bad_page(page, "compound_head not consistent", 0);
900 page->mapping = NULL;
901 clear_compound_head(page);
905 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
906 unsigned long zone, int nid)
908 set_page_links(page, zone, nid, pfn);
909 init_page_count(page);
910 page_mapcount_reset(page);
911 page_cpupid_reset_last(page);
913 INIT_LIST_HEAD(&page->lru);
914 #ifdef WANT_PAGE_VIRTUAL
915 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
916 if (!is_highmem_idx(zone))
917 set_page_address(page, __va(pfn << PAGE_SHIFT));
921 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
924 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
927 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
928 static void init_reserved_page(unsigned long pfn)
933 if (!early_page_uninitialised(pfn))
936 nid = early_pfn_to_nid(pfn);
937 pgdat = NODE_DATA(nid);
939 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
940 struct zone *zone = &pgdat->node_zones[zid];
942 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
945 __init_single_pfn(pfn, zid, nid);
948 static inline void init_reserved_page(unsigned long pfn)
951 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
954 * Initialised pages do not have PageReserved set. This function is
955 * called for each range allocated by the bootmem allocator and
956 * marks the pages PageReserved. The remaining valid pages are later
957 * sent to the buddy page allocator.
959 void __meminit reserve_bootmem_region(unsigned long start, unsigned long end)
961 unsigned long start_pfn = PFN_DOWN(start);
962 unsigned long end_pfn = PFN_UP(end);
964 for (; start_pfn < end_pfn; start_pfn++) {
965 if (pfn_valid(start_pfn)) {
966 struct page *page = pfn_to_page(start_pfn);
968 init_reserved_page(start_pfn);
970 /* Avoid false-positive PageTail() */
971 INIT_LIST_HEAD(&page->lru);
973 SetPageReserved(page);
978 static bool free_pages_prepare(struct page *page, unsigned int order)
980 bool compound = PageCompound(page);
983 VM_BUG_ON_PAGE(PageTail(page), page);
984 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
986 trace_mm_page_free(page, order);
987 kmemcheck_free_shadow(page, order);
988 kasan_free_pages(page, order);
991 page->mapping = NULL;
992 bad += free_pages_check(page);
993 for (i = 1; i < (1 << order); i++) {
995 bad += free_tail_pages_check(page, page + i);
996 bad += free_pages_check(page + i);
1001 reset_page_owner(page, order);
1003 if (!PageHighMem(page)) {
1004 debug_check_no_locks_freed(page_address(page),
1005 PAGE_SIZE << order);
1006 debug_check_no_obj_freed(page_address(page),
1007 PAGE_SIZE << order);
1009 arch_free_page(page, order);
1010 kernel_map_pages(page, 1 << order, 0);
1015 static void __free_pages_ok(struct page *page, unsigned int order)
1017 unsigned long flags;
1019 unsigned long pfn = page_to_pfn(page);
1021 if (!free_pages_prepare(page, order))
1024 migratetype = get_pfnblock_migratetype(page, pfn);
1025 local_irq_save(flags);
1026 __count_vm_events(PGFREE, 1 << order);
1027 free_one_page(page_zone(page), page, pfn, order, migratetype);
1028 local_irq_restore(flags);
1031 static void __init __free_pages_boot_core(struct page *page,
1032 unsigned long pfn, unsigned int order)
1034 unsigned int nr_pages = 1 << order;
1035 struct page *p = page;
1039 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1041 __ClearPageReserved(p);
1042 set_page_count(p, 0);
1044 __ClearPageReserved(p);
1045 set_page_count(p, 0);
1047 page_zone(page)->managed_pages += nr_pages;
1048 set_page_refcounted(page);
1049 __free_pages(page, order);
1052 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1053 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1055 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1057 int __meminit early_pfn_to_nid(unsigned long pfn)
1059 static DEFINE_SPINLOCK(early_pfn_lock);
1062 spin_lock(&early_pfn_lock);
1063 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1066 spin_unlock(&early_pfn_lock);
1072 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1073 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1074 struct mminit_pfnnid_cache *state)
1078 nid = __early_pfn_to_nid(pfn, state);
1079 if (nid >= 0 && nid != node)
1084 /* Only safe to use early in boot when initialisation is single-threaded */
1085 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1087 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1092 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1096 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1097 struct mminit_pfnnid_cache *state)
1104 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1107 if (early_page_uninitialised(pfn))
1109 return __free_pages_boot_core(page, pfn, order);
1112 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1113 static void __init deferred_free_range(struct page *page,
1114 unsigned long pfn, int nr_pages)
1121 /* Free a large naturally-aligned chunk if possible */
1122 if (nr_pages == MAX_ORDER_NR_PAGES &&
1123 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1124 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1125 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1129 for (i = 0; i < nr_pages; i++, page++, pfn++)
1130 __free_pages_boot_core(page, pfn, 0);
1133 /* Completion tracking for deferred_init_memmap() threads */
1134 static atomic_t pgdat_init_n_undone __initdata;
1135 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1137 static inline void __init pgdat_init_report_one_done(void)
1139 if (atomic_dec_and_test(&pgdat_init_n_undone))
1140 complete(&pgdat_init_all_done_comp);
1143 /* Initialise remaining memory on a node */
1144 static int __init deferred_init_memmap(void *data)
1146 pg_data_t *pgdat = data;
1147 int nid = pgdat->node_id;
1148 struct mminit_pfnnid_cache nid_init_state = { };
1149 unsigned long start = jiffies;
1150 unsigned long nr_pages = 0;
1151 unsigned long walk_start, walk_end;
1154 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1155 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1157 if (first_init_pfn == ULONG_MAX) {
1158 pgdat_init_report_one_done();
1162 /* Bind memory initialisation thread to a local node if possible */
1163 if (!cpumask_empty(cpumask))
1164 set_cpus_allowed_ptr(current, cpumask);
1166 /* Sanity check boundaries */
1167 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1168 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1169 pgdat->first_deferred_pfn = ULONG_MAX;
1171 /* Only the highest zone is deferred so find it */
1172 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1173 zone = pgdat->node_zones + zid;
1174 if (first_init_pfn < zone_end_pfn(zone))
1178 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1179 unsigned long pfn, end_pfn;
1180 struct page *page = NULL;
1181 struct page *free_base_page = NULL;
1182 unsigned long free_base_pfn = 0;
1185 end_pfn = min(walk_end, zone_end_pfn(zone));
1186 pfn = first_init_pfn;
1187 if (pfn < walk_start)
1189 if (pfn < zone->zone_start_pfn)
1190 pfn = zone->zone_start_pfn;
1192 for (; pfn < end_pfn; pfn++) {
1193 if (!pfn_valid_within(pfn))
1197 * Ensure pfn_valid is checked every
1198 * MAX_ORDER_NR_PAGES for memory holes
1200 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1201 if (!pfn_valid(pfn)) {
1207 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1212 /* Minimise pfn page lookups and scheduler checks */
1213 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1216 nr_pages += nr_to_free;
1217 deferred_free_range(free_base_page,
1218 free_base_pfn, nr_to_free);
1219 free_base_page = NULL;
1220 free_base_pfn = nr_to_free = 0;
1222 page = pfn_to_page(pfn);
1227 VM_BUG_ON(page_zone(page) != zone);
1231 __init_single_page(page, pfn, zid, nid);
1232 if (!free_base_page) {
1233 free_base_page = page;
1234 free_base_pfn = pfn;
1239 /* Where possible, batch up pages for a single free */
1242 /* Free the current block of pages to allocator */
1243 nr_pages += nr_to_free;
1244 deferred_free_range(free_base_page, free_base_pfn,
1246 free_base_page = NULL;
1247 free_base_pfn = nr_to_free = 0;
1250 first_init_pfn = max(end_pfn, first_init_pfn);
1253 /* Sanity check that the next zone really is unpopulated */
1254 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1256 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1257 jiffies_to_msecs(jiffies - start));
1259 pgdat_init_report_one_done();
1263 void __init page_alloc_init_late(void)
1267 /* There will be num_node_state(N_MEMORY) threads */
1268 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1269 for_each_node_state(nid, N_MEMORY) {
1270 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1273 /* Block until all are initialised */
1274 wait_for_completion(&pgdat_init_all_done_comp);
1276 /* Reinit limits that are based on free pages after the kernel is up */
1277 files_maxfiles_init();
1279 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1282 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1283 void __init init_cma_reserved_pageblock(struct page *page)
1285 unsigned i = pageblock_nr_pages;
1286 struct page *p = page;
1289 __ClearPageReserved(p);
1290 set_page_count(p, 0);
1293 set_pageblock_migratetype(page, MIGRATE_CMA);
1295 if (pageblock_order >= MAX_ORDER) {
1296 i = pageblock_nr_pages;
1299 set_page_refcounted(p);
1300 __free_pages(p, MAX_ORDER - 1);
1301 p += MAX_ORDER_NR_PAGES;
1302 } while (i -= MAX_ORDER_NR_PAGES);
1304 set_page_refcounted(page);
1305 __free_pages(page, pageblock_order);
1308 adjust_managed_page_count(page, pageblock_nr_pages);
1313 * The order of subdivision here is critical for the IO subsystem.
1314 * Please do not alter this order without good reasons and regression
1315 * testing. Specifically, as large blocks of memory are subdivided,
1316 * the order in which smaller blocks are delivered depends on the order
1317 * they're subdivided in this function. This is the primary factor
1318 * influencing the order in which pages are delivered to the IO
1319 * subsystem according to empirical testing, and this is also justified
1320 * by considering the behavior of a buddy system containing a single
1321 * large block of memory acted on by a series of small allocations.
1322 * This behavior is a critical factor in sglist merging's success.
1326 static inline void expand(struct zone *zone, struct page *page,
1327 int low, int high, struct free_area *area,
1330 unsigned long size = 1 << high;
1332 while (high > low) {
1336 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1338 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1339 debug_guardpage_enabled() &&
1340 high < debug_guardpage_minorder()) {
1342 * Mark as guard pages (or page), that will allow to
1343 * merge back to allocator when buddy will be freed.
1344 * Corresponding page table entries will not be touched,
1345 * pages will stay not present in virtual address space
1347 set_page_guard(zone, &page[size], high, migratetype);
1350 list_add(&page[size].lru, &area->free_list[migratetype]);
1352 set_page_order(&page[size], high);
1357 * This page is about to be returned from the page allocator
1359 static inline int check_new_page(struct page *page)
1361 const char *bad_reason = NULL;
1362 unsigned long bad_flags = 0;
1364 if (unlikely(atomic_read(&page->_mapcount) != -1))
1365 bad_reason = "nonzero mapcount";
1366 if (unlikely(page->mapping != NULL))
1367 bad_reason = "non-NULL mapping";
1368 if (unlikely(atomic_read(&page->_count) != 0))
1369 bad_reason = "nonzero _count";
1370 if (unlikely(page->flags & __PG_HWPOISON)) {
1371 bad_reason = "HWPoisoned (hardware-corrupted)";
1372 bad_flags = __PG_HWPOISON;
1374 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1375 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1376 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1379 if (unlikely(page->mem_cgroup))
1380 bad_reason = "page still charged to cgroup";
1382 if (unlikely(bad_reason)) {
1383 bad_page(page, bad_reason, bad_flags);
1389 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1394 for (i = 0; i < (1 << order); i++) {
1395 struct page *p = page + i;
1396 if (unlikely(check_new_page(p)))
1400 set_page_private(page, 0);
1401 set_page_refcounted(page);
1403 arch_alloc_page(page, order);
1404 kernel_map_pages(page, 1 << order, 1);
1405 kasan_alloc_pages(page, order);
1407 if (gfp_flags & __GFP_ZERO)
1408 for (i = 0; i < (1 << order); i++)
1409 clear_highpage(page + i);
1411 if (order && (gfp_flags & __GFP_COMP))
1412 prep_compound_page(page, order);
1414 set_page_owner(page, order, gfp_flags);
1417 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1418 * allocate the page. The expectation is that the caller is taking
1419 * steps that will free more memory. The caller should avoid the page
1420 * being used for !PFMEMALLOC purposes.
1422 if (alloc_flags & ALLOC_NO_WATERMARKS)
1423 set_page_pfmemalloc(page);
1425 clear_page_pfmemalloc(page);
1431 * Go through the free lists for the given migratetype and remove
1432 * the smallest available page from the freelists
1435 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1438 unsigned int current_order;
1439 struct free_area *area;
1442 /* Find a page of the appropriate size in the preferred list */
1443 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1444 area = &(zone->free_area[current_order]);
1445 page = list_first_entry_or_null(&area->free_list[migratetype],
1449 list_del(&page->lru);
1450 rmv_page_order(page);
1452 expand(zone, page, order, current_order, area, migratetype);
1453 set_pcppage_migratetype(page, migratetype);
1462 * This array describes the order lists are fallen back to when
1463 * the free lists for the desirable migrate type are depleted
1465 static int fallbacks[MIGRATE_TYPES][4] = {
1466 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1467 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1468 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1470 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1472 #ifdef CONFIG_MEMORY_ISOLATION
1473 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1478 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1481 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1484 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1485 unsigned int order) { return NULL; }
1489 * Move the free pages in a range to the free lists of the requested type.
1490 * Note that start_page and end_pages are not aligned on a pageblock
1491 * boundary. If alignment is required, use move_freepages_block()
1493 int move_freepages(struct zone *zone,
1494 struct page *start_page, struct page *end_page,
1499 int pages_moved = 0;
1501 #ifndef CONFIG_HOLES_IN_ZONE
1503 * page_zone is not safe to call in this context when
1504 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1505 * anyway as we check zone boundaries in move_freepages_block().
1506 * Remove at a later date when no bug reports exist related to
1507 * grouping pages by mobility
1509 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1512 for (page = start_page; page <= end_page;) {
1513 /* Make sure we are not inadvertently changing nodes */
1514 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1516 if (!pfn_valid_within(page_to_pfn(page))) {
1521 if (!PageBuddy(page)) {
1526 order = page_order(page);
1527 list_move(&page->lru,
1528 &zone->free_area[order].free_list[migratetype]);
1530 pages_moved += 1 << order;
1536 int move_freepages_block(struct zone *zone, struct page *page,
1539 unsigned long start_pfn, end_pfn;
1540 struct page *start_page, *end_page;
1542 start_pfn = page_to_pfn(page);
1543 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1544 start_page = pfn_to_page(start_pfn);
1545 end_page = start_page + pageblock_nr_pages - 1;
1546 end_pfn = start_pfn + pageblock_nr_pages - 1;
1548 /* Do not cross zone boundaries */
1549 if (!zone_spans_pfn(zone, start_pfn))
1551 if (!zone_spans_pfn(zone, end_pfn))
1554 return move_freepages(zone, start_page, end_page, migratetype);
1557 static void change_pageblock_range(struct page *pageblock_page,
1558 int start_order, int migratetype)
1560 int nr_pageblocks = 1 << (start_order - pageblock_order);
1562 while (nr_pageblocks--) {
1563 set_pageblock_migratetype(pageblock_page, migratetype);
1564 pageblock_page += pageblock_nr_pages;
1569 * When we are falling back to another migratetype during allocation, try to
1570 * steal extra free pages from the same pageblocks to satisfy further
1571 * allocations, instead of polluting multiple pageblocks.
1573 * If we are stealing a relatively large buddy page, it is likely there will
1574 * be more free pages in the pageblock, so try to steal them all. For
1575 * reclaimable and unmovable allocations, we steal regardless of page size,
1576 * as fragmentation caused by those allocations polluting movable pageblocks
1577 * is worse than movable allocations stealing from unmovable and reclaimable
1580 static bool can_steal_fallback(unsigned int order, int start_mt)
1583 * Leaving this order check is intended, although there is
1584 * relaxed order check in next check. The reason is that
1585 * we can actually steal whole pageblock if this condition met,
1586 * but, below check doesn't guarantee it and that is just heuristic
1587 * so could be changed anytime.
1589 if (order >= pageblock_order)
1592 if (order >= pageblock_order / 2 ||
1593 start_mt == MIGRATE_RECLAIMABLE ||
1594 start_mt == MIGRATE_UNMOVABLE ||
1595 page_group_by_mobility_disabled)
1602 * This function implements actual steal behaviour. If order is large enough,
1603 * we can steal whole pageblock. If not, we first move freepages in this
1604 * pageblock and check whether half of pages are moved or not. If half of
1605 * pages are moved, we can change migratetype of pageblock and permanently
1606 * use it's pages as requested migratetype in the future.
1608 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1611 unsigned int current_order = page_order(page);
1614 /* Take ownership for orders >= pageblock_order */
1615 if (current_order >= pageblock_order) {
1616 change_pageblock_range(page, current_order, start_type);
1620 pages = move_freepages_block(zone, page, start_type);
1622 /* Claim the whole block if over half of it is free */
1623 if (pages >= (1 << (pageblock_order-1)) ||
1624 page_group_by_mobility_disabled)
1625 set_pageblock_migratetype(page, start_type);
1629 * Check whether there is a suitable fallback freepage with requested order.
1630 * If only_stealable is true, this function returns fallback_mt only if
1631 * we can steal other freepages all together. This would help to reduce
1632 * fragmentation due to mixed migratetype pages in one pageblock.
1634 int find_suitable_fallback(struct free_area *area, unsigned int order,
1635 int migratetype, bool only_stealable, bool *can_steal)
1640 if (area->nr_free == 0)
1645 fallback_mt = fallbacks[migratetype][i];
1646 if (fallback_mt == MIGRATE_TYPES)
1649 if (list_empty(&area->free_list[fallback_mt]))
1652 if (can_steal_fallback(order, migratetype))
1655 if (!only_stealable)
1666 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1667 * there are no empty page blocks that contain a page with a suitable order
1669 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1670 unsigned int alloc_order)
1673 unsigned long max_managed, flags;
1676 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1677 * Check is race-prone but harmless.
1679 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1680 if (zone->nr_reserved_highatomic >= max_managed)
1683 spin_lock_irqsave(&zone->lock, flags);
1685 /* Recheck the nr_reserved_highatomic limit under the lock */
1686 if (zone->nr_reserved_highatomic >= max_managed)
1690 mt = get_pageblock_migratetype(page);
1691 if (mt != MIGRATE_HIGHATOMIC &&
1692 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1693 zone->nr_reserved_highatomic += pageblock_nr_pages;
1694 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1695 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1699 spin_unlock_irqrestore(&zone->lock, flags);
1703 * Used when an allocation is about to fail under memory pressure. This
1704 * potentially hurts the reliability of high-order allocations when under
1705 * intense memory pressure but failed atomic allocations should be easier
1706 * to recover from than an OOM.
1708 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1710 struct zonelist *zonelist = ac->zonelist;
1711 unsigned long flags;
1717 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1719 /* Preserve at least one pageblock */
1720 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1723 spin_lock_irqsave(&zone->lock, flags);
1724 for (order = 0; order < MAX_ORDER; order++) {
1725 struct free_area *area = &(zone->free_area[order]);
1727 page = list_first_entry_or_null(
1728 &area->free_list[MIGRATE_HIGHATOMIC],
1734 * It should never happen but changes to locking could
1735 * inadvertently allow a per-cpu drain to add pages
1736 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1737 * and watch for underflows.
1739 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
1740 zone->nr_reserved_highatomic);
1743 * Convert to ac->migratetype and avoid the normal
1744 * pageblock stealing heuristics. Minimally, the caller
1745 * is doing the work and needs the pages. More
1746 * importantly, if the block was always converted to
1747 * MIGRATE_UNMOVABLE or another type then the number
1748 * of pageblocks that cannot be completely freed
1751 set_pageblock_migratetype(page, ac->migratetype);
1752 move_freepages_block(zone, page, ac->migratetype);
1753 spin_unlock_irqrestore(&zone->lock, flags);
1756 spin_unlock_irqrestore(&zone->lock, flags);
1760 /* Remove an element from the buddy allocator from the fallback list */
1761 static inline struct page *
1762 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1764 struct free_area *area;
1765 unsigned int current_order;
1770 /* Find the largest possible block of pages in the other list */
1771 for (current_order = MAX_ORDER-1;
1772 current_order >= order && current_order <= MAX_ORDER-1;
1774 area = &(zone->free_area[current_order]);
1775 fallback_mt = find_suitable_fallback(area, current_order,
1776 start_migratetype, false, &can_steal);
1777 if (fallback_mt == -1)
1780 page = list_first_entry(&area->free_list[fallback_mt],
1783 steal_suitable_fallback(zone, page, start_migratetype);
1785 /* Remove the page from the freelists */
1787 list_del(&page->lru);
1788 rmv_page_order(page);
1790 expand(zone, page, order, current_order, area,
1793 * The pcppage_migratetype may differ from pageblock's
1794 * migratetype depending on the decisions in
1795 * find_suitable_fallback(). This is OK as long as it does not
1796 * differ for MIGRATE_CMA pageblocks. Those can be used as
1797 * fallback only via special __rmqueue_cma_fallback() function
1799 set_pcppage_migratetype(page, start_migratetype);
1801 trace_mm_page_alloc_extfrag(page, order, current_order,
1802 start_migratetype, fallback_mt);
1811 * Do the hard work of removing an element from the buddy allocator.
1812 * Call me with the zone->lock already held.
1814 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1819 page = __rmqueue_smallest(zone, order, migratetype);
1820 if (unlikely(!page)) {
1821 if (migratetype == MIGRATE_MOVABLE)
1822 page = __rmqueue_cma_fallback(zone, order);
1825 page = __rmqueue_fallback(zone, order, migratetype);
1828 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1833 * Obtain a specified number of elements from the buddy allocator, all under
1834 * a single hold of the lock, for efficiency. Add them to the supplied list.
1835 * Returns the number of new pages which were placed at *list.
1837 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1838 unsigned long count, struct list_head *list,
1839 int migratetype, bool cold)
1843 spin_lock(&zone->lock);
1844 for (i = 0; i < count; ++i) {
1845 struct page *page = __rmqueue(zone, order, migratetype);
1846 if (unlikely(page == NULL))
1850 * Split buddy pages returned by expand() are received here
1851 * in physical page order. The page is added to the callers and
1852 * list and the list head then moves forward. From the callers
1853 * perspective, the linked list is ordered by page number in
1854 * some conditions. This is useful for IO devices that can
1855 * merge IO requests if the physical pages are ordered
1859 list_add(&page->lru, list);
1861 list_add_tail(&page->lru, list);
1863 if (is_migrate_cma(get_pcppage_migratetype(page)))
1864 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1867 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1868 spin_unlock(&zone->lock);
1874 * Called from the vmstat counter updater to drain pagesets of this
1875 * currently executing processor on remote nodes after they have
1878 * Note that this function must be called with the thread pinned to
1879 * a single processor.
1881 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1883 unsigned long flags;
1884 int to_drain, batch;
1886 local_irq_save(flags);
1887 batch = READ_ONCE(pcp->batch);
1888 to_drain = min(pcp->count, batch);
1890 free_pcppages_bulk(zone, to_drain, pcp);
1891 pcp->count -= to_drain;
1893 local_irq_restore(flags);
1898 * Drain pcplists of the indicated processor and zone.
1900 * The processor must either be the current processor and the
1901 * thread pinned to the current processor or a processor that
1904 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1906 unsigned long flags;
1907 struct per_cpu_pageset *pset;
1908 struct per_cpu_pages *pcp;
1910 local_irq_save(flags);
1911 pset = per_cpu_ptr(zone->pageset, cpu);
1915 free_pcppages_bulk(zone, pcp->count, pcp);
1918 local_irq_restore(flags);
1922 * Drain pcplists of all zones on the indicated processor.
1924 * The processor must either be the current processor and the
1925 * thread pinned to the current processor or a processor that
1928 static void drain_pages(unsigned int cpu)
1932 for_each_populated_zone(zone) {
1933 drain_pages_zone(cpu, zone);
1938 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1940 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1941 * the single zone's pages.
1943 void drain_local_pages(struct zone *zone)
1945 int cpu = smp_processor_id();
1948 drain_pages_zone(cpu, zone);
1954 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1956 * When zone parameter is non-NULL, spill just the single zone's pages.
1958 * Note that this code is protected against sending an IPI to an offline
1959 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1960 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1961 * nothing keeps CPUs from showing up after we populated the cpumask and
1962 * before the call to on_each_cpu_mask().
1964 void drain_all_pages(struct zone *zone)
1969 * Allocate in the BSS so we wont require allocation in
1970 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1972 static cpumask_t cpus_with_pcps;
1975 * We don't care about racing with CPU hotplug event
1976 * as offline notification will cause the notified
1977 * cpu to drain that CPU pcps and on_each_cpu_mask
1978 * disables preemption as part of its processing
1980 for_each_online_cpu(cpu) {
1981 struct per_cpu_pageset *pcp;
1983 bool has_pcps = false;
1986 pcp = per_cpu_ptr(zone->pageset, cpu);
1990 for_each_populated_zone(z) {
1991 pcp = per_cpu_ptr(z->pageset, cpu);
1992 if (pcp->pcp.count) {
2000 cpumask_set_cpu(cpu, &cpus_with_pcps);
2002 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2004 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2008 #ifdef CONFIG_HIBERNATION
2010 void mark_free_pages(struct zone *zone)
2012 unsigned long pfn, max_zone_pfn;
2013 unsigned long flags;
2014 unsigned int order, t;
2017 if (zone_is_empty(zone))
2020 spin_lock_irqsave(&zone->lock, flags);
2022 max_zone_pfn = zone_end_pfn(zone);
2023 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2024 if (pfn_valid(pfn)) {
2025 page = pfn_to_page(pfn);
2026 if (!swsusp_page_is_forbidden(page))
2027 swsusp_unset_page_free(page);
2030 for_each_migratetype_order(order, t) {
2031 list_for_each_entry(page,
2032 &zone->free_area[order].free_list[t], lru) {
2035 pfn = page_to_pfn(page);
2036 for (i = 0; i < (1UL << order); i++)
2037 swsusp_set_page_free(pfn_to_page(pfn + i));
2040 spin_unlock_irqrestore(&zone->lock, flags);
2042 #endif /* CONFIG_PM */
2045 * Free a 0-order page
2046 * cold == true ? free a cold page : free a hot page
2048 void free_hot_cold_page(struct page *page, bool cold)
2050 struct zone *zone = page_zone(page);
2051 struct per_cpu_pages *pcp;
2052 unsigned long flags;
2053 unsigned long pfn = page_to_pfn(page);
2056 if (!free_pages_prepare(page, 0))
2059 migratetype = get_pfnblock_migratetype(page, pfn);
2060 set_pcppage_migratetype(page, migratetype);
2061 local_irq_save(flags);
2062 __count_vm_event(PGFREE);
2065 * We only track unmovable, reclaimable and movable on pcp lists.
2066 * Free ISOLATE pages back to the allocator because they are being
2067 * offlined but treat RESERVE as movable pages so we can get those
2068 * areas back if necessary. Otherwise, we may have to free
2069 * excessively into the page allocator
2071 if (migratetype >= MIGRATE_PCPTYPES) {
2072 if (unlikely(is_migrate_isolate(migratetype))) {
2073 free_one_page(zone, page, pfn, 0, migratetype);
2076 migratetype = MIGRATE_MOVABLE;
2079 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2081 list_add(&page->lru, &pcp->lists[migratetype]);
2083 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2085 if (pcp->count >= pcp->high) {
2086 unsigned long batch = READ_ONCE(pcp->batch);
2087 free_pcppages_bulk(zone, batch, pcp);
2088 pcp->count -= batch;
2092 local_irq_restore(flags);
2096 * Free a list of 0-order pages
2098 void free_hot_cold_page_list(struct list_head *list, bool cold)
2100 struct page *page, *next;
2102 list_for_each_entry_safe(page, next, list, lru) {
2103 trace_mm_page_free_batched(page, cold);
2104 free_hot_cold_page(page, cold);
2109 * split_page takes a non-compound higher-order page, and splits it into
2110 * n (1<<order) sub-pages: page[0..n]
2111 * Each sub-page must be freed individually.
2113 * Note: this is probably too low level an operation for use in drivers.
2114 * Please consult with lkml before using this in your driver.
2116 void split_page(struct page *page, unsigned int order)
2121 VM_BUG_ON_PAGE(PageCompound(page), page);
2122 VM_BUG_ON_PAGE(!page_count(page), page);
2124 #ifdef CONFIG_KMEMCHECK
2126 * Split shadow pages too, because free(page[0]) would
2127 * otherwise free the whole shadow.
2129 if (kmemcheck_page_is_tracked(page))
2130 split_page(virt_to_page(page[0].shadow), order);
2133 gfp_mask = get_page_owner_gfp(page);
2134 set_page_owner(page, 0, gfp_mask);
2135 for (i = 1; i < (1 << order); i++) {
2136 set_page_refcounted(page + i);
2137 set_page_owner(page + i, 0, gfp_mask);
2140 EXPORT_SYMBOL_GPL(split_page);
2142 int __isolate_free_page(struct page *page, unsigned int order)
2144 unsigned long watermark;
2148 BUG_ON(!PageBuddy(page));
2150 zone = page_zone(page);
2151 mt = get_pageblock_migratetype(page);
2153 if (!is_migrate_isolate(mt)) {
2154 /* Obey watermarks as if the page was being allocated */
2155 watermark = low_wmark_pages(zone) + (1 << order);
2156 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2159 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2162 /* Remove page from free list */
2163 list_del(&page->lru);
2164 zone->free_area[order].nr_free--;
2165 rmv_page_order(page);
2167 set_page_owner(page, order, __GFP_MOVABLE);
2169 /* Set the pageblock if the isolated page is at least a pageblock */
2170 if (order >= pageblock_order - 1) {
2171 struct page *endpage = page + (1 << order) - 1;
2172 for (; page < endpage; page += pageblock_nr_pages) {
2173 int mt = get_pageblock_migratetype(page);
2174 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2175 set_pageblock_migratetype(page,
2181 return 1UL << order;
2185 * Similar to split_page except the page is already free. As this is only
2186 * being used for migration, the migratetype of the block also changes.
2187 * As this is called with interrupts disabled, the caller is responsible
2188 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2191 * Note: this is probably too low level an operation for use in drivers.
2192 * Please consult with lkml before using this in your driver.
2194 int split_free_page(struct page *page)
2199 order = page_order(page);
2201 nr_pages = __isolate_free_page(page, order);
2205 /* Split into individual pages */
2206 set_page_refcounted(page);
2207 split_page(page, order);
2212 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2215 struct page *buffered_rmqueue(struct zone *preferred_zone,
2216 struct zone *zone, unsigned int order,
2217 gfp_t gfp_flags, int alloc_flags, int migratetype)
2219 unsigned long flags;
2221 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2223 if (likely(order == 0)) {
2224 struct per_cpu_pages *pcp;
2225 struct list_head *list;
2227 local_irq_save(flags);
2228 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2229 list = &pcp->lists[migratetype];
2230 if (list_empty(list)) {
2231 pcp->count += rmqueue_bulk(zone, 0,
2234 if (unlikely(list_empty(list)))
2239 page = list_last_entry(list, struct page, lru);
2241 page = list_first_entry(list, struct page, lru);
2243 list_del(&page->lru);
2246 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2248 * __GFP_NOFAIL is not to be used in new code.
2250 * All __GFP_NOFAIL callers should be fixed so that they
2251 * properly detect and handle allocation failures.
2253 * We most definitely don't want callers attempting to
2254 * allocate greater than order-1 page units with
2257 WARN_ON_ONCE(order > 1);
2259 spin_lock_irqsave(&zone->lock, flags);
2262 if (alloc_flags & ALLOC_HARDER) {
2263 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2265 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2268 page = __rmqueue(zone, order, migratetype);
2269 spin_unlock(&zone->lock);
2272 __mod_zone_freepage_state(zone, -(1 << order),
2273 get_pcppage_migratetype(page));
2276 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2277 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2278 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2279 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2281 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2282 zone_statistics(preferred_zone, zone, gfp_flags);
2283 local_irq_restore(flags);
2285 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2289 local_irq_restore(flags);
2293 #ifdef CONFIG_FAIL_PAGE_ALLOC
2296 struct fault_attr attr;
2298 bool ignore_gfp_highmem;
2299 bool ignore_gfp_reclaim;
2301 } fail_page_alloc = {
2302 .attr = FAULT_ATTR_INITIALIZER,
2303 .ignore_gfp_reclaim = true,
2304 .ignore_gfp_highmem = true,
2308 static int __init setup_fail_page_alloc(char *str)
2310 return setup_fault_attr(&fail_page_alloc.attr, str);
2312 __setup("fail_page_alloc=", setup_fail_page_alloc);
2314 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2316 if (order < fail_page_alloc.min_order)
2318 if (gfp_mask & __GFP_NOFAIL)
2320 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2322 if (fail_page_alloc.ignore_gfp_reclaim &&
2323 (gfp_mask & __GFP_DIRECT_RECLAIM))
2326 return should_fail(&fail_page_alloc.attr, 1 << order);
2329 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2331 static int __init fail_page_alloc_debugfs(void)
2333 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2336 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2337 &fail_page_alloc.attr);
2339 return PTR_ERR(dir);
2341 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2342 &fail_page_alloc.ignore_gfp_reclaim))
2344 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2345 &fail_page_alloc.ignore_gfp_highmem))
2347 if (!debugfs_create_u32("min-order", mode, dir,
2348 &fail_page_alloc.min_order))
2353 debugfs_remove_recursive(dir);
2358 late_initcall(fail_page_alloc_debugfs);
2360 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2362 #else /* CONFIG_FAIL_PAGE_ALLOC */
2364 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2369 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2372 * Return true if free base pages are above 'mark'. For high-order checks it
2373 * will return true of the order-0 watermark is reached and there is at least
2374 * one free page of a suitable size. Checking now avoids taking the zone lock
2375 * to check in the allocation paths if no pages are free.
2377 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2378 unsigned long mark, int classzone_idx, int alloc_flags,
2383 const int alloc_harder = (alloc_flags & ALLOC_HARDER);
2385 /* free_pages may go negative - that's OK */
2386 free_pages -= (1 << order) - 1;
2388 if (alloc_flags & ALLOC_HIGH)
2392 * If the caller does not have rights to ALLOC_HARDER then subtract
2393 * the high-atomic reserves. This will over-estimate the size of the
2394 * atomic reserve but it avoids a search.
2396 if (likely(!alloc_harder))
2397 free_pages -= z->nr_reserved_highatomic;
2402 /* If allocation can't use CMA areas don't use free CMA pages */
2403 if (!(alloc_flags & ALLOC_CMA))
2404 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2408 * Check watermarks for an order-0 allocation request. If these
2409 * are not met, then a high-order request also cannot go ahead
2410 * even if a suitable page happened to be free.
2412 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2415 /* If this is an order-0 request then the watermark is fine */
2419 /* For a high-order request, check at least one suitable page is free */
2420 for (o = order; o < MAX_ORDER; o++) {
2421 struct free_area *area = &z->free_area[o];
2430 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2431 if (!list_empty(&area->free_list[mt]))
2436 if ((alloc_flags & ALLOC_CMA) &&
2437 !list_empty(&area->free_list[MIGRATE_CMA])) {
2445 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2446 int classzone_idx, int alloc_flags)
2448 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2449 zone_page_state(z, NR_FREE_PAGES));
2452 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2453 unsigned long mark, int classzone_idx)
2455 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2457 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2458 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2460 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2465 static bool zone_local(struct zone *local_zone, struct zone *zone)
2467 return local_zone->node == zone->node;
2470 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2472 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2475 #else /* CONFIG_NUMA */
2476 static bool zone_local(struct zone *local_zone, struct zone *zone)
2481 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2485 #endif /* CONFIG_NUMA */
2487 static void reset_alloc_batches(struct zone *preferred_zone)
2489 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2492 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2493 high_wmark_pages(zone) - low_wmark_pages(zone) -
2494 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2495 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2496 } while (zone++ != preferred_zone);
2500 * get_page_from_freelist goes through the zonelist trying to allocate
2503 static struct page *
2504 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2505 const struct alloc_context *ac)
2507 struct zonelist *zonelist = ac->zonelist;
2509 struct page *page = NULL;
2511 int nr_fair_skipped = 0;
2512 bool zonelist_rescan;
2515 zonelist_rescan = false;
2518 * Scan zonelist, looking for a zone with enough free.
2519 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2521 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2525 if (cpusets_enabled() &&
2526 (alloc_flags & ALLOC_CPUSET) &&
2527 !cpuset_zone_allowed(zone, gfp_mask))
2530 * Distribute pages in proportion to the individual
2531 * zone size to ensure fair page aging. The zone a
2532 * page was allocated in should have no effect on the
2533 * time the page has in memory before being reclaimed.
2535 if (alloc_flags & ALLOC_FAIR) {
2536 if (!zone_local(ac->preferred_zone, zone))
2538 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2544 * When allocating a page cache page for writing, we
2545 * want to get it from a zone that is within its dirty
2546 * limit, such that no single zone holds more than its
2547 * proportional share of globally allowed dirty pages.
2548 * The dirty limits take into account the zone's
2549 * lowmem reserves and high watermark so that kswapd
2550 * should be able to balance it without having to
2551 * write pages from its LRU list.
2553 * This may look like it could increase pressure on
2554 * lower zones by failing allocations in higher zones
2555 * before they are full. But the pages that do spill
2556 * over are limited as the lower zones are protected
2557 * by this very same mechanism. It should not become
2558 * a practical burden to them.
2560 * XXX: For now, allow allocations to potentially
2561 * exceed the per-zone dirty limit in the slowpath
2562 * (spread_dirty_pages unset) before going into reclaim,
2563 * which is important when on a NUMA setup the allowed
2564 * zones are together not big enough to reach the
2565 * global limit. The proper fix for these situations
2566 * will require awareness of zones in the
2567 * dirty-throttling and the flusher threads.
2569 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2572 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2573 if (!zone_watermark_ok(zone, order, mark,
2574 ac->classzone_idx, alloc_flags)) {
2577 /* Checked here to keep the fast path fast */
2578 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2579 if (alloc_flags & ALLOC_NO_WATERMARKS)
2582 if (zone_reclaim_mode == 0 ||
2583 !zone_allows_reclaim(ac->preferred_zone, zone))
2586 ret = zone_reclaim(zone, gfp_mask, order);
2588 case ZONE_RECLAIM_NOSCAN:
2591 case ZONE_RECLAIM_FULL:
2592 /* scanned but unreclaimable */
2595 /* did we reclaim enough */
2596 if (zone_watermark_ok(zone, order, mark,
2597 ac->classzone_idx, alloc_flags))
2605 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2606 gfp_mask, alloc_flags, ac->migratetype);
2608 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2612 * If this is a high-order atomic allocation then check
2613 * if the pageblock should be reserved for the future
2615 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2616 reserve_highatomic_pageblock(page, zone, order);
2623 * The first pass makes sure allocations are spread fairly within the
2624 * local node. However, the local node might have free pages left
2625 * after the fairness batches are exhausted, and remote zones haven't
2626 * even been considered yet. Try once more without fairness, and
2627 * include remote zones now, before entering the slowpath and waking
2628 * kswapd: prefer spilling to a remote zone over swapping locally.
2630 if (alloc_flags & ALLOC_FAIR) {
2631 alloc_flags &= ~ALLOC_FAIR;
2632 if (nr_fair_skipped) {
2633 zonelist_rescan = true;
2634 reset_alloc_batches(ac->preferred_zone);
2636 if (nr_online_nodes > 1)
2637 zonelist_rescan = true;
2640 if (zonelist_rescan)
2647 * Large machines with many possible nodes should not always dump per-node
2648 * meminfo in irq context.
2650 static inline bool should_suppress_show_mem(void)
2655 ret = in_interrupt();
2660 static DEFINE_RATELIMIT_STATE(nopage_rs,
2661 DEFAULT_RATELIMIT_INTERVAL,
2662 DEFAULT_RATELIMIT_BURST);
2664 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
2666 unsigned int filter = SHOW_MEM_FILTER_NODES;
2668 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2669 debug_guardpage_minorder() > 0)
2673 * This documents exceptions given to allocations in certain
2674 * contexts that are allowed to allocate outside current's set
2677 if (!(gfp_mask & __GFP_NOMEMALLOC))
2678 if (test_thread_flag(TIF_MEMDIE) ||
2679 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2680 filter &= ~SHOW_MEM_FILTER_NODES;
2681 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2682 filter &= ~SHOW_MEM_FILTER_NODES;
2685 struct va_format vaf;
2688 va_start(args, fmt);
2693 pr_warn("%pV", &vaf);
2698 pr_warn("%s: page allocation failure: order:%u, mode:%#x(%pGg)\n",
2699 current->comm, order, gfp_mask, &gfp_mask);
2701 if (!should_suppress_show_mem())
2705 static inline struct page *
2706 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2707 const struct alloc_context *ac, unsigned long *did_some_progress)
2709 struct oom_control oc = {
2710 .zonelist = ac->zonelist,
2711 .nodemask = ac->nodemask,
2712 .gfp_mask = gfp_mask,
2717 *did_some_progress = 0;
2720 * Acquire the oom lock. If that fails, somebody else is
2721 * making progress for us.
2723 if (!mutex_trylock(&oom_lock)) {
2724 *did_some_progress = 1;
2725 schedule_timeout_uninterruptible(1);
2730 * Go through the zonelist yet one more time, keep very high watermark
2731 * here, this is only to catch a parallel oom killing, we must fail if
2732 * we're still under heavy pressure.
2734 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2735 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2739 if (!(gfp_mask & __GFP_NOFAIL)) {
2740 /* Coredumps can quickly deplete all memory reserves */
2741 if (current->flags & PF_DUMPCORE)
2743 /* The OOM killer will not help higher order allocs */
2744 if (order > PAGE_ALLOC_COSTLY_ORDER)
2746 /* The OOM killer does not needlessly kill tasks for lowmem */
2747 if (ac->high_zoneidx < ZONE_NORMAL)
2749 /* The OOM killer does not compensate for IO-less reclaim */
2750 if (!(gfp_mask & __GFP_FS)) {
2752 * XXX: Page reclaim didn't yield anything,
2753 * and the OOM killer can't be invoked, but
2754 * keep looping as per tradition.
2756 *did_some_progress = 1;
2759 if (pm_suspended_storage())
2761 /* The OOM killer may not free memory on a specific node */
2762 if (gfp_mask & __GFP_THISNODE)
2765 /* Exhausted what can be done so it's blamo time */
2766 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
2767 *did_some_progress = 1;
2769 if (gfp_mask & __GFP_NOFAIL) {
2770 page = get_page_from_freelist(gfp_mask, order,
2771 ALLOC_NO_WATERMARKS|ALLOC_CPUSET, ac);
2773 * fallback to ignore cpuset restriction if our nodes
2777 page = get_page_from_freelist(gfp_mask, order,
2778 ALLOC_NO_WATERMARKS, ac);
2782 mutex_unlock(&oom_lock);
2786 #ifdef CONFIG_COMPACTION
2787 /* Try memory compaction for high-order allocations before reclaim */
2788 static struct page *
2789 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2790 int alloc_flags, const struct alloc_context *ac,
2791 enum migrate_mode mode, int *contended_compaction,
2792 bool *deferred_compaction)
2794 unsigned long compact_result;
2800 current->flags |= PF_MEMALLOC;
2801 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2802 mode, contended_compaction);
2803 current->flags &= ~PF_MEMALLOC;
2805 switch (compact_result) {
2806 case COMPACT_DEFERRED:
2807 *deferred_compaction = true;
2809 case COMPACT_SKIPPED:
2816 * At least in one zone compaction wasn't deferred or skipped, so let's
2817 * count a compaction stall
2819 count_vm_event(COMPACTSTALL);
2821 page = get_page_from_freelist(gfp_mask, order,
2822 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2825 struct zone *zone = page_zone(page);
2827 zone->compact_blockskip_flush = false;
2828 compaction_defer_reset(zone, order, true);
2829 count_vm_event(COMPACTSUCCESS);
2834 * It's bad if compaction run occurs and fails. The most likely reason
2835 * is that pages exist, but not enough to satisfy watermarks.
2837 count_vm_event(COMPACTFAIL);
2844 static inline struct page *
2845 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2846 int alloc_flags, const struct alloc_context *ac,
2847 enum migrate_mode mode, int *contended_compaction,
2848 bool *deferred_compaction)
2852 #endif /* CONFIG_COMPACTION */
2854 /* Perform direct synchronous page reclaim */
2856 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2857 const struct alloc_context *ac)
2859 struct reclaim_state reclaim_state;
2864 /* We now go into synchronous reclaim */
2865 cpuset_memory_pressure_bump();
2866 current->flags |= PF_MEMALLOC;
2867 lockdep_set_current_reclaim_state(gfp_mask);
2868 reclaim_state.reclaimed_slab = 0;
2869 current->reclaim_state = &reclaim_state;
2871 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2874 current->reclaim_state = NULL;
2875 lockdep_clear_current_reclaim_state();
2876 current->flags &= ~PF_MEMALLOC;
2883 /* The really slow allocator path where we enter direct reclaim */
2884 static inline struct page *
2885 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2886 int alloc_flags, const struct alloc_context *ac,
2887 unsigned long *did_some_progress)
2889 struct page *page = NULL;
2890 bool drained = false;
2892 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2893 if (unlikely(!(*did_some_progress)))
2897 page = get_page_from_freelist(gfp_mask, order,
2898 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2901 * If an allocation failed after direct reclaim, it could be because
2902 * pages are pinned on the per-cpu lists or in high alloc reserves.
2903 * Shrink them them and try again
2905 if (!page && !drained) {
2906 unreserve_highatomic_pageblock(ac);
2907 drain_all_pages(NULL);
2915 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2920 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2921 ac->high_zoneidx, ac->nodemask)
2922 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2926 gfp_to_alloc_flags(gfp_t gfp_mask)
2928 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2930 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2931 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2934 * The caller may dip into page reserves a bit more if the caller
2935 * cannot run direct reclaim, or if the caller has realtime scheduling
2936 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2937 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
2939 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2941 if (gfp_mask & __GFP_ATOMIC) {
2943 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2944 * if it can't schedule.
2946 if (!(gfp_mask & __GFP_NOMEMALLOC))
2947 alloc_flags |= ALLOC_HARDER;
2949 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2950 * comment for __cpuset_node_allowed().
2952 alloc_flags &= ~ALLOC_CPUSET;
2953 } else if (unlikely(rt_task(current)) && !in_interrupt())
2954 alloc_flags |= ALLOC_HARDER;
2956 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2957 if (gfp_mask & __GFP_MEMALLOC)
2958 alloc_flags |= ALLOC_NO_WATERMARKS;
2959 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2960 alloc_flags |= ALLOC_NO_WATERMARKS;
2961 else if (!in_interrupt() &&
2962 ((current->flags & PF_MEMALLOC) ||
2963 unlikely(test_thread_flag(TIF_MEMDIE))))
2964 alloc_flags |= ALLOC_NO_WATERMARKS;
2967 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2968 alloc_flags |= ALLOC_CMA;
2973 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2975 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2978 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
2980 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
2983 static inline struct page *
2984 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2985 struct alloc_context *ac)
2987 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
2988 struct page *page = NULL;
2990 unsigned long pages_reclaimed = 0;
2991 unsigned long did_some_progress;
2992 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2993 bool deferred_compaction = false;
2994 int contended_compaction = COMPACT_CONTENDED_NONE;
2997 * In the slowpath, we sanity check order to avoid ever trying to
2998 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2999 * be using allocators in order of preference for an area that is
3002 if (order >= MAX_ORDER) {
3003 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3008 * We also sanity check to catch abuse of atomic reserves being used by
3009 * callers that are not in atomic context.
3011 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3012 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3013 gfp_mask &= ~__GFP_ATOMIC;
3016 * If this allocation cannot block and it is for a specific node, then
3017 * fail early. There's no need to wakeup kswapd or retry for a
3018 * speculative node-specific allocation.
3020 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !can_direct_reclaim)
3024 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3025 wake_all_kswapds(order, ac);
3028 * OK, we're below the kswapd watermark and have kicked background
3029 * reclaim. Now things get more complex, so set up alloc_flags according
3030 * to how we want to proceed.
3032 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3035 * Find the true preferred zone if the allocation is unconstrained by
3038 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3039 struct zoneref *preferred_zoneref;
3040 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3041 ac->high_zoneidx, NULL, &ac->preferred_zone);
3042 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3045 /* This is the last chance, in general, before the goto nopage. */
3046 page = get_page_from_freelist(gfp_mask, order,
3047 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3051 /* Allocate without watermarks if the context allows */
3052 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3054 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3055 * the allocation is high priority and these type of
3056 * allocations are system rather than user orientated
3058 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3059 page = get_page_from_freelist(gfp_mask, order,
3060 ALLOC_NO_WATERMARKS, ac);
3065 /* Caller is not willing to reclaim, we can't balance anything */
3066 if (!can_direct_reclaim) {
3068 * All existing users of the __GFP_NOFAIL are blockable, so warn
3069 * of any new users that actually allow this type of allocation
3072 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3076 /* Avoid recursion of direct reclaim */
3077 if (current->flags & PF_MEMALLOC) {
3079 * __GFP_NOFAIL request from this context is rather bizarre
3080 * because we cannot reclaim anything and only can loop waiting
3081 * for somebody to do a work for us.
3083 if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3090 /* Avoid allocations with no watermarks from looping endlessly */
3091 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3095 * Try direct compaction. The first pass is asynchronous. Subsequent
3096 * attempts after direct reclaim are synchronous
3098 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3100 &contended_compaction,
3101 &deferred_compaction);
3105 /* Checks for THP-specific high-order allocations */
3106 if (is_thp_gfp_mask(gfp_mask)) {
3108 * If compaction is deferred for high-order allocations, it is
3109 * because sync compaction recently failed. If this is the case
3110 * and the caller requested a THP allocation, we do not want
3111 * to heavily disrupt the system, so we fail the allocation
3112 * instead of entering direct reclaim.
3114 if (deferred_compaction)
3118 * In all zones where compaction was attempted (and not
3119 * deferred or skipped), lock contention has been detected.
3120 * For THP allocation we do not want to disrupt the others
3121 * so we fallback to base pages instead.
3123 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3127 * If compaction was aborted due to need_resched(), we do not
3128 * want to further increase allocation latency, unless it is
3129 * khugepaged trying to collapse.
3131 if (contended_compaction == COMPACT_CONTENDED_SCHED
3132 && !(current->flags & PF_KTHREAD))
3137 * It can become very expensive to allocate transparent hugepages at
3138 * fault, so use asynchronous memory compaction for THP unless it is
3139 * khugepaged trying to collapse.
3141 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3142 migration_mode = MIGRATE_SYNC_LIGHT;
3144 /* Try direct reclaim and then allocating */
3145 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3146 &did_some_progress);
3150 /* Do not loop if specifically requested */
3151 if (gfp_mask & __GFP_NORETRY)
3154 /* Keep reclaiming pages as long as there is reasonable progress */
3155 pages_reclaimed += did_some_progress;
3156 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3157 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3158 /* Wait for some write requests to complete then retry */
3159 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3163 /* Reclaim has failed us, start killing things */
3164 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3168 /* Retry as long as the OOM killer is making progress */
3169 if (did_some_progress)
3174 * High-order allocations do not necessarily loop after
3175 * direct reclaim and reclaim/compaction depends on compaction
3176 * being called after reclaim so call directly if necessary
3178 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3180 &contended_compaction,
3181 &deferred_compaction);
3185 warn_alloc_failed(gfp_mask, order, NULL);
3191 * This is the 'heart' of the zoned buddy allocator.
3194 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3195 struct zonelist *zonelist, nodemask_t *nodemask)
3197 struct zoneref *preferred_zoneref;
3198 struct page *page = NULL;
3199 unsigned int cpuset_mems_cookie;
3200 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3201 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3202 struct alloc_context ac = {
3203 .high_zoneidx = gfp_zone(gfp_mask),
3204 .nodemask = nodemask,
3205 .migratetype = gfpflags_to_migratetype(gfp_mask),
3208 gfp_mask &= gfp_allowed_mask;
3210 lockdep_trace_alloc(gfp_mask);
3212 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3214 if (should_fail_alloc_page(gfp_mask, order))
3218 * Check the zones suitable for the gfp_mask contain at least one
3219 * valid zone. It's possible to have an empty zonelist as a result
3220 * of __GFP_THISNODE and a memoryless node
3222 if (unlikely(!zonelist->_zonerefs->zone))
3225 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3226 alloc_flags |= ALLOC_CMA;
3229 cpuset_mems_cookie = read_mems_allowed_begin();
3231 /* We set it here, as __alloc_pages_slowpath might have changed it */
3232 ac.zonelist = zonelist;
3234 /* Dirty zone balancing only done in the fast path */
3235 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3237 /* The preferred zone is used for statistics later */
3238 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3239 ac.nodemask ? : &cpuset_current_mems_allowed,
3240 &ac.preferred_zone);
3241 if (!ac.preferred_zone)
3243 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3245 /* First allocation attempt */
3246 alloc_mask = gfp_mask|__GFP_HARDWALL;
3247 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3248 if (unlikely(!page)) {
3250 * Runtime PM, block IO and its error handling path
3251 * can deadlock because I/O on the device might not
3254 alloc_mask = memalloc_noio_flags(gfp_mask);
3255 ac.spread_dirty_pages = false;
3257 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3260 if (kmemcheck_enabled && page)
3261 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3263 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3267 * When updating a task's mems_allowed, it is possible to race with
3268 * parallel threads in such a way that an allocation can fail while
3269 * the mask is being updated. If a page allocation is about to fail,
3270 * check if the cpuset changed during allocation and if so, retry.
3272 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3277 EXPORT_SYMBOL(__alloc_pages_nodemask);
3280 * Common helper functions.
3282 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3287 * __get_free_pages() returns a 32-bit address, which cannot represent
3290 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3292 page = alloc_pages(gfp_mask, order);
3295 return (unsigned long) page_address(page);
3297 EXPORT_SYMBOL(__get_free_pages);
3299 unsigned long get_zeroed_page(gfp_t gfp_mask)
3301 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3303 EXPORT_SYMBOL(get_zeroed_page);
3305 void __free_pages(struct page *page, unsigned int order)
3307 if (put_page_testzero(page)) {
3309 free_hot_cold_page(page, false);
3311 __free_pages_ok(page, order);
3315 EXPORT_SYMBOL(__free_pages);
3317 void free_pages(unsigned long addr, unsigned int order)
3320 VM_BUG_ON(!virt_addr_valid((void *)addr));
3321 __free_pages(virt_to_page((void *)addr), order);
3325 EXPORT_SYMBOL(free_pages);
3329 * An arbitrary-length arbitrary-offset area of memory which resides
3330 * within a 0 or higher order page. Multiple fragments within that page
3331 * are individually refcounted, in the page's reference counter.
3333 * The page_frag functions below provide a simple allocation framework for
3334 * page fragments. This is used by the network stack and network device
3335 * drivers to provide a backing region of memory for use as either an
3336 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3338 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3341 struct page *page = NULL;
3342 gfp_t gfp = gfp_mask;
3344 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3345 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3347 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3348 PAGE_FRAG_CACHE_MAX_ORDER);
3349 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3351 if (unlikely(!page))
3352 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3354 nc->va = page ? page_address(page) : NULL;
3359 void *__alloc_page_frag(struct page_frag_cache *nc,
3360 unsigned int fragsz, gfp_t gfp_mask)
3362 unsigned int size = PAGE_SIZE;
3366 if (unlikely(!nc->va)) {
3368 page = __page_frag_refill(nc, gfp_mask);
3372 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3373 /* if size can vary use size else just use PAGE_SIZE */
3376 /* Even if we own the page, we do not use atomic_set().
3377 * This would break get_page_unless_zero() users.
3379 atomic_add(size - 1, &page->_count);
3381 /* reset page count bias and offset to start of new frag */
3382 nc->pfmemalloc = page_is_pfmemalloc(page);
3383 nc->pagecnt_bias = size;
3387 offset = nc->offset - fragsz;
3388 if (unlikely(offset < 0)) {
3389 page = virt_to_page(nc->va);
3391 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3394 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3395 /* if size can vary use size else just use PAGE_SIZE */
3398 /* OK, page count is 0, we can safely set it */
3399 atomic_set(&page->_count, size);
3401 /* reset page count bias and offset to start of new frag */
3402 nc->pagecnt_bias = size;
3403 offset = size - fragsz;
3407 nc->offset = offset;
3409 return nc->va + offset;
3411 EXPORT_SYMBOL(__alloc_page_frag);
3414 * Frees a page fragment allocated out of either a compound or order 0 page.
3416 void __free_page_frag(void *addr)
3418 struct page *page = virt_to_head_page(addr);
3420 if (unlikely(put_page_testzero(page)))
3421 __free_pages_ok(page, compound_order(page));
3423 EXPORT_SYMBOL(__free_page_frag);
3426 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3427 * of the current memory cgroup if __GFP_ACCOUNT is set, other than that it is
3428 * equivalent to alloc_pages.
3430 * It should be used when the caller would like to use kmalloc, but since the
3431 * allocation is large, it has to fall back to the page allocator.
3433 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3437 page = alloc_pages(gfp_mask, order);
3438 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3439 __free_pages(page, order);
3445 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3449 page = alloc_pages_node(nid, gfp_mask, order);
3450 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3451 __free_pages(page, order);
3458 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3461 void __free_kmem_pages(struct page *page, unsigned int order)
3463 memcg_kmem_uncharge(page, order);
3464 __free_pages(page, order);
3467 void free_kmem_pages(unsigned long addr, unsigned int order)
3470 VM_BUG_ON(!virt_addr_valid((void *)addr));
3471 __free_kmem_pages(virt_to_page((void *)addr), order);
3475 static void *make_alloc_exact(unsigned long addr, unsigned int order,
3479 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3480 unsigned long used = addr + PAGE_ALIGN(size);
3482 split_page(virt_to_page((void *)addr), order);
3483 while (used < alloc_end) {
3488 return (void *)addr;
3492 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3493 * @size: the number of bytes to allocate
3494 * @gfp_mask: GFP flags for the allocation
3496 * This function is similar to alloc_pages(), except that it allocates the
3497 * minimum number of pages to satisfy the request. alloc_pages() can only
3498 * allocate memory in power-of-two pages.
3500 * This function is also limited by MAX_ORDER.
3502 * Memory allocated by this function must be released by free_pages_exact().
3504 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3506 unsigned int order = get_order(size);
3509 addr = __get_free_pages(gfp_mask, order);
3510 return make_alloc_exact(addr, order, size);
3512 EXPORT_SYMBOL(alloc_pages_exact);
3515 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3517 * @nid: the preferred node ID where memory should be allocated
3518 * @size: the number of bytes to allocate
3519 * @gfp_mask: GFP flags for the allocation
3521 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3524 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3526 unsigned int order = get_order(size);
3527 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3530 return make_alloc_exact((unsigned long)page_address(p), order, size);
3534 * free_pages_exact - release memory allocated via alloc_pages_exact()
3535 * @virt: the value returned by alloc_pages_exact.
3536 * @size: size of allocation, same value as passed to alloc_pages_exact().
3538 * Release the memory allocated by a previous call to alloc_pages_exact.
3540 void free_pages_exact(void *virt, size_t size)
3542 unsigned long addr = (unsigned long)virt;
3543 unsigned long end = addr + PAGE_ALIGN(size);
3545 while (addr < end) {
3550 EXPORT_SYMBOL(free_pages_exact);
3553 * nr_free_zone_pages - count number of pages beyond high watermark
3554 * @offset: The zone index of the highest zone
3556 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3557 * high watermark within all zones at or below a given zone index. For each
3558 * zone, the number of pages is calculated as:
3559 * managed_pages - high_pages
3561 static unsigned long nr_free_zone_pages(int offset)
3566 /* Just pick one node, since fallback list is circular */
3567 unsigned long sum = 0;
3569 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3571 for_each_zone_zonelist(zone, z, zonelist, offset) {
3572 unsigned long size = zone->managed_pages;
3573 unsigned long high = high_wmark_pages(zone);
3582 * nr_free_buffer_pages - count number of pages beyond high watermark
3584 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3585 * watermark within ZONE_DMA and ZONE_NORMAL.
3587 unsigned long nr_free_buffer_pages(void)
3589 return nr_free_zone_pages(gfp_zone(GFP_USER));
3591 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3594 * nr_free_pagecache_pages - count number of pages beyond high watermark
3596 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3597 * high watermark within all zones.
3599 unsigned long nr_free_pagecache_pages(void)
3601 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3604 static inline void show_node(struct zone *zone)
3606 if (IS_ENABLED(CONFIG_NUMA))
3607 printk("Node %d ", zone_to_nid(zone));
3610 void si_meminfo(struct sysinfo *val)
3612 val->totalram = totalram_pages;
3613 val->sharedram = global_page_state(NR_SHMEM);
3614 val->freeram = global_page_state(NR_FREE_PAGES);
3615 val->bufferram = nr_blockdev_pages();
3616 val->totalhigh = totalhigh_pages;
3617 val->freehigh = nr_free_highpages();
3618 val->mem_unit = PAGE_SIZE;
3621 EXPORT_SYMBOL(si_meminfo);
3624 void si_meminfo_node(struct sysinfo *val, int nid)
3626 int zone_type; /* needs to be signed */
3627 unsigned long managed_pages = 0;
3628 pg_data_t *pgdat = NODE_DATA(nid);
3630 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3631 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3632 val->totalram = managed_pages;
3633 val->sharedram = node_page_state(nid, NR_SHMEM);
3634 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3635 #ifdef CONFIG_HIGHMEM
3636 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3637 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3643 val->mem_unit = PAGE_SIZE;
3648 * Determine whether the node should be displayed or not, depending on whether
3649 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3651 bool skip_free_areas_node(unsigned int flags, int nid)
3654 unsigned int cpuset_mems_cookie;
3656 if (!(flags & SHOW_MEM_FILTER_NODES))
3660 cpuset_mems_cookie = read_mems_allowed_begin();
3661 ret = !node_isset(nid, cpuset_current_mems_allowed);
3662 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3667 #define K(x) ((x) << (PAGE_SHIFT-10))
3669 static void show_migration_types(unsigned char type)
3671 static const char types[MIGRATE_TYPES] = {
3672 [MIGRATE_UNMOVABLE] = 'U',
3673 [MIGRATE_MOVABLE] = 'M',
3674 [MIGRATE_RECLAIMABLE] = 'E',
3675 [MIGRATE_HIGHATOMIC] = 'H',
3677 [MIGRATE_CMA] = 'C',
3679 #ifdef CONFIG_MEMORY_ISOLATION
3680 [MIGRATE_ISOLATE] = 'I',
3683 char tmp[MIGRATE_TYPES + 1];
3687 for (i = 0; i < MIGRATE_TYPES; i++) {
3688 if (type & (1 << i))
3693 printk("(%s) ", tmp);
3697 * Show free area list (used inside shift_scroll-lock stuff)
3698 * We also calculate the percentage fragmentation. We do this by counting the
3699 * memory on each free list with the exception of the first item on the list.
3702 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3705 void show_free_areas(unsigned int filter)
3707 unsigned long free_pcp = 0;
3711 for_each_populated_zone(zone) {
3712 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3715 for_each_online_cpu(cpu)
3716 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3719 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3720 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3721 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3722 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3723 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3724 " free:%lu free_pcp:%lu free_cma:%lu\n",
3725 global_page_state(NR_ACTIVE_ANON),
3726 global_page_state(NR_INACTIVE_ANON),
3727 global_page_state(NR_ISOLATED_ANON),
3728 global_page_state(NR_ACTIVE_FILE),
3729 global_page_state(NR_INACTIVE_FILE),
3730 global_page_state(NR_ISOLATED_FILE),
3731 global_page_state(NR_UNEVICTABLE),
3732 global_page_state(NR_FILE_DIRTY),
3733 global_page_state(NR_WRITEBACK),
3734 global_page_state(NR_UNSTABLE_NFS),
3735 global_page_state(NR_SLAB_RECLAIMABLE),
3736 global_page_state(NR_SLAB_UNRECLAIMABLE),
3737 global_page_state(NR_FILE_MAPPED),
3738 global_page_state(NR_SHMEM),
3739 global_page_state(NR_PAGETABLE),
3740 global_page_state(NR_BOUNCE),
3741 global_page_state(NR_FREE_PAGES),
3743 global_page_state(NR_FREE_CMA_PAGES));
3745 for_each_populated_zone(zone) {
3748 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3752 for_each_online_cpu(cpu)
3753 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3761 " active_anon:%lukB"
3762 " inactive_anon:%lukB"
3763 " active_file:%lukB"
3764 " inactive_file:%lukB"
3765 " unevictable:%lukB"
3766 " isolated(anon):%lukB"
3767 " isolated(file):%lukB"
3775 " slab_reclaimable:%lukB"
3776 " slab_unreclaimable:%lukB"
3777 " kernel_stack:%lukB"
3784 " writeback_tmp:%lukB"
3785 " pages_scanned:%lu"
3786 " all_unreclaimable? %s"
3789 K(zone_page_state(zone, NR_FREE_PAGES)),
3790 K(min_wmark_pages(zone)),
3791 K(low_wmark_pages(zone)),
3792 K(high_wmark_pages(zone)),
3793 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3794 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3795 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3796 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3797 K(zone_page_state(zone, NR_UNEVICTABLE)),
3798 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3799 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3800 K(zone->present_pages),
3801 K(zone->managed_pages),
3802 K(zone_page_state(zone, NR_MLOCK)),
3803 K(zone_page_state(zone, NR_FILE_DIRTY)),
3804 K(zone_page_state(zone, NR_WRITEBACK)),
3805 K(zone_page_state(zone, NR_FILE_MAPPED)),
3806 K(zone_page_state(zone, NR_SHMEM)),
3807 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3808 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3809 zone_page_state(zone, NR_KERNEL_STACK) *
3811 K(zone_page_state(zone, NR_PAGETABLE)),
3812 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3813 K(zone_page_state(zone, NR_BOUNCE)),
3815 K(this_cpu_read(zone->pageset->pcp.count)),
3816 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3817 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3818 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3819 (!zone_reclaimable(zone) ? "yes" : "no")
3821 printk("lowmem_reserve[]:");
3822 for (i = 0; i < MAX_NR_ZONES; i++)
3823 printk(" %ld", zone->lowmem_reserve[i]);
3827 for_each_populated_zone(zone) {
3829 unsigned long nr[MAX_ORDER], flags, total = 0;
3830 unsigned char types[MAX_ORDER];
3832 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3835 printk("%s: ", zone->name);
3837 spin_lock_irqsave(&zone->lock, flags);
3838 for (order = 0; order < MAX_ORDER; order++) {
3839 struct free_area *area = &zone->free_area[order];
3842 nr[order] = area->nr_free;
3843 total += nr[order] << order;
3846 for (type = 0; type < MIGRATE_TYPES; type++) {
3847 if (!list_empty(&area->free_list[type]))
3848 types[order] |= 1 << type;
3851 spin_unlock_irqrestore(&zone->lock, flags);
3852 for (order = 0; order < MAX_ORDER; order++) {
3853 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3855 show_migration_types(types[order]);
3857 printk("= %lukB\n", K(total));
3860 hugetlb_show_meminfo();
3862 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3864 show_swap_cache_info();
3867 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3869 zoneref->zone = zone;
3870 zoneref->zone_idx = zone_idx(zone);
3874 * Builds allocation fallback zone lists.
3876 * Add all populated zones of a node to the zonelist.
3878 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3882 enum zone_type zone_type = MAX_NR_ZONES;
3886 zone = pgdat->node_zones + zone_type;
3887 if (populated_zone(zone)) {
3888 zoneref_set_zone(zone,
3889 &zonelist->_zonerefs[nr_zones++]);
3890 check_highest_zone(zone_type);
3892 } while (zone_type);
3900 * 0 = automatic detection of better ordering.
3901 * 1 = order by ([node] distance, -zonetype)
3902 * 2 = order by (-zonetype, [node] distance)
3904 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3905 * the same zonelist. So only NUMA can configure this param.
3907 #define ZONELIST_ORDER_DEFAULT 0
3908 #define ZONELIST_ORDER_NODE 1
3909 #define ZONELIST_ORDER_ZONE 2
3911 /* zonelist order in the kernel.
3912 * set_zonelist_order() will set this to NODE or ZONE.
3914 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3915 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3919 /* The value user specified ....changed by config */
3920 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3921 /* string for sysctl */
3922 #define NUMA_ZONELIST_ORDER_LEN 16
3923 char numa_zonelist_order[16] = "default";
3926 * interface for configure zonelist ordering.
3927 * command line option "numa_zonelist_order"
3928 * = "[dD]efault - default, automatic configuration.
3929 * = "[nN]ode - order by node locality, then by zone within node
3930 * = "[zZ]one - order by zone, then by locality within zone
3933 static int __parse_numa_zonelist_order(char *s)
3935 if (*s == 'd' || *s == 'D') {
3936 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3937 } else if (*s == 'n' || *s == 'N') {
3938 user_zonelist_order = ZONELIST_ORDER_NODE;
3939 } else if (*s == 'z' || *s == 'Z') {
3940 user_zonelist_order = ZONELIST_ORDER_ZONE;
3943 "Ignoring invalid numa_zonelist_order value: "
3950 static __init int setup_numa_zonelist_order(char *s)
3957 ret = __parse_numa_zonelist_order(s);
3959 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3963 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3966 * sysctl handler for numa_zonelist_order
3968 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3969 void __user *buffer, size_t *length,
3972 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3974 static DEFINE_MUTEX(zl_order_mutex);
3976 mutex_lock(&zl_order_mutex);
3978 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3982 strcpy(saved_string, (char *)table->data);
3984 ret = proc_dostring(table, write, buffer, length, ppos);
3988 int oldval = user_zonelist_order;
3990 ret = __parse_numa_zonelist_order((char *)table->data);
3993 * bogus value. restore saved string
3995 strncpy((char *)table->data, saved_string,
3996 NUMA_ZONELIST_ORDER_LEN);
3997 user_zonelist_order = oldval;
3998 } else if (oldval != user_zonelist_order) {
3999 mutex_lock(&zonelists_mutex);
4000 build_all_zonelists(NULL, NULL);
4001 mutex_unlock(&zonelists_mutex);
4005 mutex_unlock(&zl_order_mutex);
4010 #define MAX_NODE_LOAD (nr_online_nodes)
4011 static int node_load[MAX_NUMNODES];
4014 * find_next_best_node - find the next node that should appear in a given node's fallback list
4015 * @node: node whose fallback list we're appending
4016 * @used_node_mask: nodemask_t of already used nodes
4018 * We use a number of factors to determine which is the next node that should
4019 * appear on a given node's fallback list. The node should not have appeared
4020 * already in @node's fallback list, and it should be the next closest node
4021 * according to the distance array (which contains arbitrary distance values
4022 * from each node to each node in the system), and should also prefer nodes
4023 * with no CPUs, since presumably they'll have very little allocation pressure
4024 * on them otherwise.
4025 * It returns -1 if no node is found.
4027 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4030 int min_val = INT_MAX;
4031 int best_node = NUMA_NO_NODE;
4032 const struct cpumask *tmp = cpumask_of_node(0);
4034 /* Use the local node if we haven't already */
4035 if (!node_isset(node, *used_node_mask)) {
4036 node_set(node, *used_node_mask);
4040 for_each_node_state(n, N_MEMORY) {
4042 /* Don't want a node to appear more than once */
4043 if (node_isset(n, *used_node_mask))
4046 /* Use the distance array to find the distance */
4047 val = node_distance(node, n);
4049 /* Penalize nodes under us ("prefer the next node") */
4052 /* Give preference to headless and unused nodes */
4053 tmp = cpumask_of_node(n);
4054 if (!cpumask_empty(tmp))
4055 val += PENALTY_FOR_NODE_WITH_CPUS;
4057 /* Slight preference for less loaded node */
4058 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4059 val += node_load[n];
4061 if (val < min_val) {
4068 node_set(best_node, *used_node_mask);
4075 * Build zonelists ordered by node and zones within node.
4076 * This results in maximum locality--normal zone overflows into local
4077 * DMA zone, if any--but risks exhausting DMA zone.
4079 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4082 struct zonelist *zonelist;
4084 zonelist = &pgdat->node_zonelists[0];
4085 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4087 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4088 zonelist->_zonerefs[j].zone = NULL;
4089 zonelist->_zonerefs[j].zone_idx = 0;
4093 * Build gfp_thisnode zonelists
4095 static void build_thisnode_zonelists(pg_data_t *pgdat)
4098 struct zonelist *zonelist;
4100 zonelist = &pgdat->node_zonelists[1];
4101 j = build_zonelists_node(pgdat, zonelist, 0);
4102 zonelist->_zonerefs[j].zone = NULL;
4103 zonelist->_zonerefs[j].zone_idx = 0;
4107 * Build zonelists ordered by zone and nodes within zones.
4108 * This results in conserving DMA zone[s] until all Normal memory is
4109 * exhausted, but results in overflowing to remote node while memory
4110 * may still exist in local DMA zone.
4112 static int node_order[MAX_NUMNODES];
4114 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4117 int zone_type; /* needs to be signed */
4119 struct zonelist *zonelist;
4121 zonelist = &pgdat->node_zonelists[0];
4123 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4124 for (j = 0; j < nr_nodes; j++) {
4125 node = node_order[j];
4126 z = &NODE_DATA(node)->node_zones[zone_type];
4127 if (populated_zone(z)) {
4129 &zonelist->_zonerefs[pos++]);
4130 check_highest_zone(zone_type);
4134 zonelist->_zonerefs[pos].zone = NULL;
4135 zonelist->_zonerefs[pos].zone_idx = 0;
4138 #if defined(CONFIG_64BIT)
4140 * Devices that require DMA32/DMA are relatively rare and do not justify a
4141 * penalty to every machine in case the specialised case applies. Default
4142 * to Node-ordering on 64-bit NUMA machines
4144 static int default_zonelist_order(void)
4146 return ZONELIST_ORDER_NODE;
4150 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4151 * by the kernel. If processes running on node 0 deplete the low memory zone
4152 * then reclaim will occur more frequency increasing stalls and potentially
4153 * be easier to OOM if a large percentage of the zone is under writeback or
4154 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4155 * Hence, default to zone ordering on 32-bit.
4157 static int default_zonelist_order(void)
4159 return ZONELIST_ORDER_ZONE;
4161 #endif /* CONFIG_64BIT */
4163 static void set_zonelist_order(void)
4165 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4166 current_zonelist_order = default_zonelist_order();
4168 current_zonelist_order = user_zonelist_order;
4171 static void build_zonelists(pg_data_t *pgdat)
4174 nodemask_t used_mask;
4175 int local_node, prev_node;
4176 struct zonelist *zonelist;
4177 unsigned int order = current_zonelist_order;
4179 /* initialize zonelists */
4180 for (i = 0; i < MAX_ZONELISTS; i++) {
4181 zonelist = pgdat->node_zonelists + i;
4182 zonelist->_zonerefs[0].zone = NULL;
4183 zonelist->_zonerefs[0].zone_idx = 0;
4186 /* NUMA-aware ordering of nodes */
4187 local_node = pgdat->node_id;
4188 load = nr_online_nodes;
4189 prev_node = local_node;
4190 nodes_clear(used_mask);
4192 memset(node_order, 0, sizeof(node_order));
4195 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4197 * We don't want to pressure a particular node.
4198 * So adding penalty to the first node in same
4199 * distance group to make it round-robin.
4201 if (node_distance(local_node, node) !=
4202 node_distance(local_node, prev_node))
4203 node_load[node] = load;
4207 if (order == ZONELIST_ORDER_NODE)
4208 build_zonelists_in_node_order(pgdat, node);
4210 node_order[i++] = node; /* remember order */
4213 if (order == ZONELIST_ORDER_ZONE) {
4214 /* calculate node order -- i.e., DMA last! */
4215 build_zonelists_in_zone_order(pgdat, i);
4218 build_thisnode_zonelists(pgdat);
4221 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4223 * Return node id of node used for "local" allocations.
4224 * I.e., first node id of first zone in arg node's generic zonelist.
4225 * Used for initializing percpu 'numa_mem', which is used primarily
4226 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4228 int local_memory_node(int node)
4232 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4233 gfp_zone(GFP_KERNEL),
4240 #else /* CONFIG_NUMA */
4242 static void set_zonelist_order(void)
4244 current_zonelist_order = ZONELIST_ORDER_ZONE;
4247 static void build_zonelists(pg_data_t *pgdat)
4249 int node, local_node;
4251 struct zonelist *zonelist;
4253 local_node = pgdat->node_id;
4255 zonelist = &pgdat->node_zonelists[0];
4256 j = build_zonelists_node(pgdat, zonelist, 0);
4259 * Now we build the zonelist so that it contains the zones
4260 * of all the other nodes.
4261 * We don't want to pressure a particular node, so when
4262 * building the zones for node N, we make sure that the
4263 * zones coming right after the local ones are those from
4264 * node N+1 (modulo N)
4266 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4267 if (!node_online(node))
4269 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4271 for (node = 0; node < local_node; node++) {
4272 if (!node_online(node))
4274 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4277 zonelist->_zonerefs[j].zone = NULL;
4278 zonelist->_zonerefs[j].zone_idx = 0;
4281 #endif /* CONFIG_NUMA */
4284 * Boot pageset table. One per cpu which is going to be used for all
4285 * zones and all nodes. The parameters will be set in such a way
4286 * that an item put on a list will immediately be handed over to
4287 * the buddy list. This is safe since pageset manipulation is done
4288 * with interrupts disabled.
4290 * The boot_pagesets must be kept even after bootup is complete for
4291 * unused processors and/or zones. They do play a role for bootstrapping
4292 * hotplugged processors.
4294 * zoneinfo_show() and maybe other functions do
4295 * not check if the processor is online before following the pageset pointer.
4296 * Other parts of the kernel may not check if the zone is available.
4298 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4299 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4300 static void setup_zone_pageset(struct zone *zone);
4303 * Global mutex to protect against size modification of zonelists
4304 * as well as to serialize pageset setup for the new populated zone.
4306 DEFINE_MUTEX(zonelists_mutex);
4308 /* return values int ....just for stop_machine() */
4309 static int __build_all_zonelists(void *data)
4313 pg_data_t *self = data;
4316 memset(node_load, 0, sizeof(node_load));
4319 if (self && !node_online(self->node_id)) {
4320 build_zonelists(self);
4323 for_each_online_node(nid) {
4324 pg_data_t *pgdat = NODE_DATA(nid);
4326 build_zonelists(pgdat);
4330 * Initialize the boot_pagesets that are going to be used
4331 * for bootstrapping processors. The real pagesets for
4332 * each zone will be allocated later when the per cpu
4333 * allocator is available.
4335 * boot_pagesets are used also for bootstrapping offline
4336 * cpus if the system is already booted because the pagesets
4337 * are needed to initialize allocators on a specific cpu too.
4338 * F.e. the percpu allocator needs the page allocator which
4339 * needs the percpu allocator in order to allocate its pagesets
4340 * (a chicken-egg dilemma).
4342 for_each_possible_cpu(cpu) {
4343 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4345 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4347 * We now know the "local memory node" for each node--
4348 * i.e., the node of the first zone in the generic zonelist.
4349 * Set up numa_mem percpu variable for on-line cpus. During
4350 * boot, only the boot cpu should be on-line; we'll init the
4351 * secondary cpus' numa_mem as they come on-line. During
4352 * node/memory hotplug, we'll fixup all on-line cpus.
4354 if (cpu_online(cpu))
4355 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4362 static noinline void __init
4363 build_all_zonelists_init(void)
4365 __build_all_zonelists(NULL);
4366 mminit_verify_zonelist();
4367 cpuset_init_current_mems_allowed();
4371 * Called with zonelists_mutex held always
4372 * unless system_state == SYSTEM_BOOTING.
4374 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4375 * [we're only called with non-NULL zone through __meminit paths] and
4376 * (2) call of __init annotated helper build_all_zonelists_init
4377 * [protected by SYSTEM_BOOTING].
4379 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4381 set_zonelist_order();
4383 if (system_state == SYSTEM_BOOTING) {
4384 build_all_zonelists_init();
4386 #ifdef CONFIG_MEMORY_HOTPLUG
4388 setup_zone_pageset(zone);
4390 /* we have to stop all cpus to guarantee there is no user
4392 stop_machine(__build_all_zonelists, pgdat, NULL);
4393 /* cpuset refresh routine should be here */
4395 vm_total_pages = nr_free_pagecache_pages();
4397 * Disable grouping by mobility if the number of pages in the
4398 * system is too low to allow the mechanism to work. It would be
4399 * more accurate, but expensive to check per-zone. This check is
4400 * made on memory-hotadd so a system can start with mobility
4401 * disabled and enable it later
4403 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4404 page_group_by_mobility_disabled = 1;
4406 page_group_by_mobility_disabled = 0;
4408 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4409 "Total pages: %ld\n",
4411 zonelist_order_name[current_zonelist_order],
4412 page_group_by_mobility_disabled ? "off" : "on",
4415 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4420 * Helper functions to size the waitqueue hash table.
4421 * Essentially these want to choose hash table sizes sufficiently
4422 * large so that collisions trying to wait on pages are rare.
4423 * But in fact, the number of active page waitqueues on typical
4424 * systems is ridiculously low, less than 200. So this is even
4425 * conservative, even though it seems large.
4427 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4428 * waitqueues, i.e. the size of the waitq table given the number of pages.
4430 #define PAGES_PER_WAITQUEUE 256
4432 #ifndef CONFIG_MEMORY_HOTPLUG
4433 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4435 unsigned long size = 1;
4437 pages /= PAGES_PER_WAITQUEUE;
4439 while (size < pages)
4443 * Once we have dozens or even hundreds of threads sleeping
4444 * on IO we've got bigger problems than wait queue collision.
4445 * Limit the size of the wait table to a reasonable size.
4447 size = min(size, 4096UL);
4449 return max(size, 4UL);
4453 * A zone's size might be changed by hot-add, so it is not possible to determine
4454 * a suitable size for its wait_table. So we use the maximum size now.
4456 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4458 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4459 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4460 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4462 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4463 * or more by the traditional way. (See above). It equals:
4465 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4466 * ia64(16K page size) : = ( 8G + 4M)byte.
4467 * powerpc (64K page size) : = (32G +16M)byte.
4469 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4476 * This is an integer logarithm so that shifts can be used later
4477 * to extract the more random high bits from the multiplicative
4478 * hash function before the remainder is taken.
4480 static inline unsigned long wait_table_bits(unsigned long size)
4486 * Initially all pages are reserved - free ones are freed
4487 * up by free_all_bootmem() once the early boot process is
4488 * done. Non-atomic initialization, single-pass.
4490 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4491 unsigned long start_pfn, enum memmap_context context)
4493 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
4494 unsigned long end_pfn = start_pfn + size;
4495 pg_data_t *pgdat = NODE_DATA(nid);
4497 unsigned long nr_initialised = 0;
4498 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4499 struct memblock_region *r = NULL, *tmp;
4502 if (highest_memmap_pfn < end_pfn - 1)
4503 highest_memmap_pfn = end_pfn - 1;
4506 * Honor reservation requested by the driver for this ZONE_DEVICE
4509 if (altmap && start_pfn == altmap->base_pfn)
4510 start_pfn += altmap->reserve;
4512 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4514 * There can be holes in boot-time mem_map[]s handed to this
4515 * function. They do not exist on hotplugged memory.
4517 if (context != MEMMAP_EARLY)
4520 if (!early_pfn_valid(pfn))
4522 if (!early_pfn_in_nid(pfn, nid))
4524 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
4527 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4529 * If not mirrored_kernelcore and ZONE_MOVABLE exists, range
4530 * from zone_movable_pfn[nid] to end of each node should be
4531 * ZONE_MOVABLE not ZONE_NORMAL. skip it.
4533 if (!mirrored_kernelcore && zone_movable_pfn[nid])
4534 if (zone == ZONE_NORMAL && pfn >= zone_movable_pfn[nid])
4538 * Check given memblock attribute by firmware which can affect
4539 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
4540 * mirrored, it's an overlapped memmap init. skip it.
4542 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
4543 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
4544 for_each_memblock(memory, tmp)
4545 if (pfn < memblock_region_memory_end_pfn(tmp))
4549 if (pfn >= memblock_region_memory_base_pfn(r) &&
4550 memblock_is_mirror(r)) {
4551 /* already initialized as NORMAL */
4552 pfn = memblock_region_memory_end_pfn(r);
4560 * Mark the block movable so that blocks are reserved for
4561 * movable at startup. This will force kernel allocations
4562 * to reserve their blocks rather than leaking throughout
4563 * the address space during boot when many long-lived
4564 * kernel allocations are made.
4566 * bitmap is created for zone's valid pfn range. but memmap
4567 * can be created for invalid pages (for alignment)
4568 * check here not to call set_pageblock_migratetype() against
4571 if (!(pfn & (pageblock_nr_pages - 1))) {
4572 struct page *page = pfn_to_page(pfn);
4574 __init_single_page(page, pfn, zone, nid);
4575 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4577 __init_single_pfn(pfn, zone, nid);
4582 static void __meminit zone_init_free_lists(struct zone *zone)
4584 unsigned int order, t;
4585 for_each_migratetype_order(order, t) {
4586 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4587 zone->free_area[order].nr_free = 0;
4591 #ifndef __HAVE_ARCH_MEMMAP_INIT
4592 #define memmap_init(size, nid, zone, start_pfn) \
4593 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4596 static int zone_batchsize(struct zone *zone)
4602 * The per-cpu-pages pools are set to around 1000th of the
4603 * size of the zone. But no more than 1/2 of a meg.
4605 * OK, so we don't know how big the cache is. So guess.
4607 batch = zone->managed_pages / 1024;
4608 if (batch * PAGE_SIZE > 512 * 1024)
4609 batch = (512 * 1024) / PAGE_SIZE;
4610 batch /= 4; /* We effectively *= 4 below */
4615 * Clamp the batch to a 2^n - 1 value. Having a power
4616 * of 2 value was found to be more likely to have
4617 * suboptimal cache aliasing properties in some cases.
4619 * For example if 2 tasks are alternately allocating
4620 * batches of pages, one task can end up with a lot
4621 * of pages of one half of the possible page colors
4622 * and the other with pages of the other colors.
4624 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4629 /* The deferral and batching of frees should be suppressed under NOMMU
4632 * The problem is that NOMMU needs to be able to allocate large chunks
4633 * of contiguous memory as there's no hardware page translation to
4634 * assemble apparent contiguous memory from discontiguous pages.
4636 * Queueing large contiguous runs of pages for batching, however,
4637 * causes the pages to actually be freed in smaller chunks. As there
4638 * can be a significant delay between the individual batches being
4639 * recycled, this leads to the once large chunks of space being
4640 * fragmented and becoming unavailable for high-order allocations.
4647 * pcp->high and pcp->batch values are related and dependent on one another:
4648 * ->batch must never be higher then ->high.
4649 * The following function updates them in a safe manner without read side
4652 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4653 * those fields changing asynchronously (acording the the above rule).
4655 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4656 * outside of boot time (or some other assurance that no concurrent updaters
4659 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4660 unsigned long batch)
4662 /* start with a fail safe value for batch */
4666 /* Update high, then batch, in order */
4673 /* a companion to pageset_set_high() */
4674 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4676 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4679 static void pageset_init(struct per_cpu_pageset *p)
4681 struct per_cpu_pages *pcp;
4684 memset(p, 0, sizeof(*p));
4688 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4689 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4692 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4695 pageset_set_batch(p, batch);
4699 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4700 * to the value high for the pageset p.
4702 static void pageset_set_high(struct per_cpu_pageset *p,
4705 unsigned long batch = max(1UL, high / 4);
4706 if ((high / 4) > (PAGE_SHIFT * 8))
4707 batch = PAGE_SHIFT * 8;
4709 pageset_update(&p->pcp, high, batch);
4712 static void pageset_set_high_and_batch(struct zone *zone,
4713 struct per_cpu_pageset *pcp)
4715 if (percpu_pagelist_fraction)
4716 pageset_set_high(pcp,
4717 (zone->managed_pages /
4718 percpu_pagelist_fraction));
4720 pageset_set_batch(pcp, zone_batchsize(zone));
4723 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4725 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4728 pageset_set_high_and_batch(zone, pcp);
4731 static void __meminit setup_zone_pageset(struct zone *zone)
4734 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4735 for_each_possible_cpu(cpu)
4736 zone_pageset_init(zone, cpu);
4740 * Allocate per cpu pagesets and initialize them.
4741 * Before this call only boot pagesets were available.
4743 void __init setup_per_cpu_pageset(void)
4747 for_each_populated_zone(zone)
4748 setup_zone_pageset(zone);
4751 static noinline __init_refok
4752 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4758 * The per-page waitqueue mechanism uses hashed waitqueues
4761 zone->wait_table_hash_nr_entries =
4762 wait_table_hash_nr_entries(zone_size_pages);
4763 zone->wait_table_bits =
4764 wait_table_bits(zone->wait_table_hash_nr_entries);
4765 alloc_size = zone->wait_table_hash_nr_entries
4766 * sizeof(wait_queue_head_t);
4768 if (!slab_is_available()) {
4769 zone->wait_table = (wait_queue_head_t *)
4770 memblock_virt_alloc_node_nopanic(
4771 alloc_size, zone->zone_pgdat->node_id);
4774 * This case means that a zone whose size was 0 gets new memory
4775 * via memory hot-add.
4776 * But it may be the case that a new node was hot-added. In
4777 * this case vmalloc() will not be able to use this new node's
4778 * memory - this wait_table must be initialized to use this new
4779 * node itself as well.
4780 * To use this new node's memory, further consideration will be
4783 zone->wait_table = vmalloc(alloc_size);
4785 if (!zone->wait_table)
4788 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4789 init_waitqueue_head(zone->wait_table + i);
4794 static __meminit void zone_pcp_init(struct zone *zone)
4797 * per cpu subsystem is not up at this point. The following code
4798 * relies on the ability of the linker to provide the
4799 * offset of a (static) per cpu variable into the per cpu area.
4801 zone->pageset = &boot_pageset;
4803 if (populated_zone(zone))
4804 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4805 zone->name, zone->present_pages,
4806 zone_batchsize(zone));
4809 int __meminit init_currently_empty_zone(struct zone *zone,
4810 unsigned long zone_start_pfn,
4813 struct pglist_data *pgdat = zone->zone_pgdat;
4815 ret = zone_wait_table_init(zone, size);
4818 pgdat->nr_zones = zone_idx(zone) + 1;
4820 zone->zone_start_pfn = zone_start_pfn;
4822 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4823 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4825 (unsigned long)zone_idx(zone),
4826 zone_start_pfn, (zone_start_pfn + size));
4828 zone_init_free_lists(zone);
4833 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4834 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4837 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4839 int __meminit __early_pfn_to_nid(unsigned long pfn,
4840 struct mminit_pfnnid_cache *state)
4842 unsigned long start_pfn, end_pfn;
4845 if (state->last_start <= pfn && pfn < state->last_end)
4846 return state->last_nid;
4848 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4850 state->last_start = start_pfn;
4851 state->last_end = end_pfn;
4852 state->last_nid = nid;
4857 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4860 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4861 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4862 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4864 * If an architecture guarantees that all ranges registered contain no holes
4865 * and may be freed, this this function may be used instead of calling
4866 * memblock_free_early_nid() manually.
4868 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4870 unsigned long start_pfn, end_pfn;
4873 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4874 start_pfn = min(start_pfn, max_low_pfn);
4875 end_pfn = min(end_pfn, max_low_pfn);
4877 if (start_pfn < end_pfn)
4878 memblock_free_early_nid(PFN_PHYS(start_pfn),
4879 (end_pfn - start_pfn) << PAGE_SHIFT,
4885 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4886 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4888 * If an architecture guarantees that all ranges registered contain no holes and may
4889 * be freed, this function may be used instead of calling memory_present() manually.
4891 void __init sparse_memory_present_with_active_regions(int nid)
4893 unsigned long start_pfn, end_pfn;
4896 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4897 memory_present(this_nid, start_pfn, end_pfn);
4901 * get_pfn_range_for_nid - Return the start and end page frames for a node
4902 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4903 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4904 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4906 * It returns the start and end page frame of a node based on information
4907 * provided by memblock_set_node(). If called for a node
4908 * with no available memory, a warning is printed and the start and end
4911 void __meminit get_pfn_range_for_nid(unsigned int nid,
4912 unsigned long *start_pfn, unsigned long *end_pfn)
4914 unsigned long this_start_pfn, this_end_pfn;
4920 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4921 *start_pfn = min(*start_pfn, this_start_pfn);
4922 *end_pfn = max(*end_pfn, this_end_pfn);
4925 if (*start_pfn == -1UL)
4930 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4931 * assumption is made that zones within a node are ordered in monotonic
4932 * increasing memory addresses so that the "highest" populated zone is used
4934 static void __init find_usable_zone_for_movable(void)
4937 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4938 if (zone_index == ZONE_MOVABLE)
4941 if (arch_zone_highest_possible_pfn[zone_index] >
4942 arch_zone_lowest_possible_pfn[zone_index])
4946 VM_BUG_ON(zone_index == -1);
4947 movable_zone = zone_index;
4951 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4952 * because it is sized independent of architecture. Unlike the other zones,
4953 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4954 * in each node depending on the size of each node and how evenly kernelcore
4955 * is distributed. This helper function adjusts the zone ranges
4956 * provided by the architecture for a given node by using the end of the
4957 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4958 * zones within a node are in order of monotonic increases memory addresses
4960 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4961 unsigned long zone_type,
4962 unsigned long node_start_pfn,
4963 unsigned long node_end_pfn,
4964 unsigned long *zone_start_pfn,
4965 unsigned long *zone_end_pfn)
4967 /* Only adjust if ZONE_MOVABLE is on this node */
4968 if (zone_movable_pfn[nid]) {
4969 /* Size ZONE_MOVABLE */
4970 if (zone_type == ZONE_MOVABLE) {
4971 *zone_start_pfn = zone_movable_pfn[nid];
4972 *zone_end_pfn = min(node_end_pfn,
4973 arch_zone_highest_possible_pfn[movable_zone]);
4975 /* Check if this whole range is within ZONE_MOVABLE */
4976 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4977 *zone_start_pfn = *zone_end_pfn;
4982 * Return the number of pages a zone spans in a node, including holes
4983 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4985 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4986 unsigned long zone_type,
4987 unsigned long node_start_pfn,
4988 unsigned long node_end_pfn,
4989 unsigned long *zone_start_pfn,
4990 unsigned long *zone_end_pfn,
4991 unsigned long *ignored)
4993 /* When hotadd a new node from cpu_up(), the node should be empty */
4994 if (!node_start_pfn && !node_end_pfn)
4997 /* Get the start and end of the zone */
4998 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4999 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5000 adjust_zone_range_for_zone_movable(nid, zone_type,
5001 node_start_pfn, node_end_pfn,
5002 zone_start_pfn, zone_end_pfn);
5004 /* Check that this node has pages within the zone's required range */
5005 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5008 /* Move the zone boundaries inside the node if necessary */
5009 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5010 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5012 /* Return the spanned pages */
5013 return *zone_end_pfn - *zone_start_pfn;
5017 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5018 * then all holes in the requested range will be accounted for.
5020 unsigned long __meminit __absent_pages_in_range(int nid,
5021 unsigned long range_start_pfn,
5022 unsigned long range_end_pfn)
5024 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5025 unsigned long start_pfn, end_pfn;
5028 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5029 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5030 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5031 nr_absent -= end_pfn - start_pfn;
5037 * absent_pages_in_range - Return number of page frames in holes within a range
5038 * @start_pfn: The start PFN to start searching for holes
5039 * @end_pfn: The end PFN to stop searching for holes
5041 * It returns the number of pages frames in memory holes within a range.
5043 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5044 unsigned long end_pfn)
5046 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5049 /* Return the number of page frames in holes in a zone on a node */
5050 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5051 unsigned long zone_type,
5052 unsigned long node_start_pfn,
5053 unsigned long node_end_pfn,
5054 unsigned long *ignored)
5056 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5057 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5058 unsigned long zone_start_pfn, zone_end_pfn;
5059 unsigned long nr_absent;
5061 /* When hotadd a new node from cpu_up(), the node should be empty */
5062 if (!node_start_pfn && !node_end_pfn)
5065 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5066 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5068 adjust_zone_range_for_zone_movable(nid, zone_type,
5069 node_start_pfn, node_end_pfn,
5070 &zone_start_pfn, &zone_end_pfn);
5071 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5074 * ZONE_MOVABLE handling.
5075 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5078 if (zone_movable_pfn[nid]) {
5079 if (mirrored_kernelcore) {
5080 unsigned long start_pfn, end_pfn;
5081 struct memblock_region *r;
5083 for_each_memblock(memory, r) {
5084 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5085 zone_start_pfn, zone_end_pfn);
5086 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5087 zone_start_pfn, zone_end_pfn);
5089 if (zone_type == ZONE_MOVABLE &&
5090 memblock_is_mirror(r))
5091 nr_absent += end_pfn - start_pfn;
5093 if (zone_type == ZONE_NORMAL &&
5094 !memblock_is_mirror(r))
5095 nr_absent += end_pfn - start_pfn;
5098 if (zone_type == ZONE_NORMAL)
5099 nr_absent += node_end_pfn - zone_movable_pfn[nid];
5106 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5107 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5108 unsigned long zone_type,
5109 unsigned long node_start_pfn,
5110 unsigned long node_end_pfn,
5111 unsigned long *zone_start_pfn,
5112 unsigned long *zone_end_pfn,
5113 unsigned long *zones_size)
5117 *zone_start_pfn = node_start_pfn;
5118 for (zone = 0; zone < zone_type; zone++)
5119 *zone_start_pfn += zones_size[zone];
5121 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5123 return zones_size[zone_type];
5126 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5127 unsigned long zone_type,
5128 unsigned long node_start_pfn,
5129 unsigned long node_end_pfn,
5130 unsigned long *zholes_size)
5135 return zholes_size[zone_type];
5138 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5140 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5141 unsigned long node_start_pfn,
5142 unsigned long node_end_pfn,
5143 unsigned long *zones_size,
5144 unsigned long *zholes_size)
5146 unsigned long realtotalpages = 0, totalpages = 0;
5149 for (i = 0; i < MAX_NR_ZONES; i++) {
5150 struct zone *zone = pgdat->node_zones + i;
5151 unsigned long zone_start_pfn, zone_end_pfn;
5152 unsigned long size, real_size;
5154 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5160 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5161 node_start_pfn, node_end_pfn,
5164 zone->zone_start_pfn = zone_start_pfn;
5166 zone->zone_start_pfn = 0;
5167 zone->spanned_pages = size;
5168 zone->present_pages = real_size;
5171 realtotalpages += real_size;
5174 pgdat->node_spanned_pages = totalpages;
5175 pgdat->node_present_pages = realtotalpages;
5176 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5180 #ifndef CONFIG_SPARSEMEM
5182 * Calculate the size of the zone->blockflags rounded to an unsigned long
5183 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5184 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5185 * round what is now in bits to nearest long in bits, then return it in
5188 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5190 unsigned long usemapsize;
5192 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5193 usemapsize = roundup(zonesize, pageblock_nr_pages);
5194 usemapsize = usemapsize >> pageblock_order;
5195 usemapsize *= NR_PAGEBLOCK_BITS;
5196 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5198 return usemapsize / 8;
5201 static void __init setup_usemap(struct pglist_data *pgdat,
5203 unsigned long zone_start_pfn,
5204 unsigned long zonesize)
5206 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5207 zone->pageblock_flags = NULL;
5209 zone->pageblock_flags =
5210 memblock_virt_alloc_node_nopanic(usemapsize,
5214 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5215 unsigned long zone_start_pfn, unsigned long zonesize) {}
5216 #endif /* CONFIG_SPARSEMEM */
5218 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5220 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5221 void __paginginit set_pageblock_order(void)
5225 /* Check that pageblock_nr_pages has not already been setup */
5226 if (pageblock_order)
5229 if (HPAGE_SHIFT > PAGE_SHIFT)
5230 order = HUGETLB_PAGE_ORDER;
5232 order = MAX_ORDER - 1;
5235 * Assume the largest contiguous order of interest is a huge page.
5236 * This value may be variable depending on boot parameters on IA64 and
5239 pageblock_order = order;
5241 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5244 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5245 * is unused as pageblock_order is set at compile-time. See
5246 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5249 void __paginginit set_pageblock_order(void)
5253 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5255 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5256 unsigned long present_pages)
5258 unsigned long pages = spanned_pages;
5261 * Provide a more accurate estimation if there are holes within
5262 * the zone and SPARSEMEM is in use. If there are holes within the
5263 * zone, each populated memory region may cost us one or two extra
5264 * memmap pages due to alignment because memmap pages for each
5265 * populated regions may not naturally algined on page boundary.
5266 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5268 if (spanned_pages > present_pages + (present_pages >> 4) &&
5269 IS_ENABLED(CONFIG_SPARSEMEM))
5270 pages = present_pages;
5272 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5276 * Set up the zone data structures:
5277 * - mark all pages reserved
5278 * - mark all memory queues empty
5279 * - clear the memory bitmaps
5281 * NOTE: pgdat should get zeroed by caller.
5283 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5286 int nid = pgdat->node_id;
5289 pgdat_resize_init(pgdat);
5290 #ifdef CONFIG_NUMA_BALANCING
5291 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5292 pgdat->numabalancing_migrate_nr_pages = 0;
5293 pgdat->numabalancing_migrate_next_window = jiffies;
5295 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5296 spin_lock_init(&pgdat->split_queue_lock);
5297 INIT_LIST_HEAD(&pgdat->split_queue);
5298 pgdat->split_queue_len = 0;
5300 init_waitqueue_head(&pgdat->kswapd_wait);
5301 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5302 pgdat_page_ext_init(pgdat);
5304 for (j = 0; j < MAX_NR_ZONES; j++) {
5305 struct zone *zone = pgdat->node_zones + j;
5306 unsigned long size, realsize, freesize, memmap_pages;
5307 unsigned long zone_start_pfn = zone->zone_start_pfn;
5309 size = zone->spanned_pages;
5310 realsize = freesize = zone->present_pages;
5313 * Adjust freesize so that it accounts for how much memory
5314 * is used by this zone for memmap. This affects the watermark
5315 * and per-cpu initialisations
5317 memmap_pages = calc_memmap_size(size, realsize);
5318 if (!is_highmem_idx(j)) {
5319 if (freesize >= memmap_pages) {
5320 freesize -= memmap_pages;
5323 " %s zone: %lu pages used for memmap\n",
5324 zone_names[j], memmap_pages);
5327 " %s zone: %lu pages exceeds freesize %lu\n",
5328 zone_names[j], memmap_pages, freesize);
5331 /* Account for reserved pages */
5332 if (j == 0 && freesize > dma_reserve) {
5333 freesize -= dma_reserve;
5334 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5335 zone_names[0], dma_reserve);
5338 if (!is_highmem_idx(j))
5339 nr_kernel_pages += freesize;
5340 /* Charge for highmem memmap if there are enough kernel pages */
5341 else if (nr_kernel_pages > memmap_pages * 2)
5342 nr_kernel_pages -= memmap_pages;
5343 nr_all_pages += freesize;
5346 * Set an approximate value for lowmem here, it will be adjusted
5347 * when the bootmem allocator frees pages into the buddy system.
5348 * And all highmem pages will be managed by the buddy system.
5350 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5353 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5355 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5357 zone->name = zone_names[j];
5358 spin_lock_init(&zone->lock);
5359 spin_lock_init(&zone->lru_lock);
5360 zone_seqlock_init(zone);
5361 zone->zone_pgdat = pgdat;
5362 zone_pcp_init(zone);
5364 /* For bootup, initialized properly in watermark setup */
5365 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5367 lruvec_init(&zone->lruvec);
5371 set_pageblock_order();
5372 setup_usemap(pgdat, zone, zone_start_pfn, size);
5373 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5375 memmap_init(size, nid, j, zone_start_pfn);
5379 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5381 unsigned long __maybe_unused start = 0;
5382 unsigned long __maybe_unused offset = 0;
5384 /* Skip empty nodes */
5385 if (!pgdat->node_spanned_pages)
5388 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5389 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5390 offset = pgdat->node_start_pfn - start;
5391 /* ia64 gets its own node_mem_map, before this, without bootmem */
5392 if (!pgdat->node_mem_map) {
5393 unsigned long size, end;
5397 * The zone's endpoints aren't required to be MAX_ORDER
5398 * aligned but the node_mem_map endpoints must be in order
5399 * for the buddy allocator to function correctly.
5401 end = pgdat_end_pfn(pgdat);
5402 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5403 size = (end - start) * sizeof(struct page);
5404 map = alloc_remap(pgdat->node_id, size);
5406 map = memblock_virt_alloc_node_nopanic(size,
5408 pgdat->node_mem_map = map + offset;
5410 #ifndef CONFIG_NEED_MULTIPLE_NODES
5412 * With no DISCONTIG, the global mem_map is just set as node 0's
5414 if (pgdat == NODE_DATA(0)) {
5415 mem_map = NODE_DATA(0)->node_mem_map;
5416 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5417 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5419 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5422 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5425 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5426 unsigned long node_start_pfn, unsigned long *zholes_size)
5428 pg_data_t *pgdat = NODE_DATA(nid);
5429 unsigned long start_pfn = 0;
5430 unsigned long end_pfn = 0;
5432 /* pg_data_t should be reset to zero when it's allocated */
5433 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5435 reset_deferred_meminit(pgdat);
5436 pgdat->node_id = nid;
5437 pgdat->node_start_pfn = node_start_pfn;
5438 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5439 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5440 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5441 (u64)start_pfn << PAGE_SHIFT,
5442 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5444 start_pfn = node_start_pfn;
5446 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5447 zones_size, zholes_size);
5449 alloc_node_mem_map(pgdat);
5450 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5451 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5452 nid, (unsigned long)pgdat,
5453 (unsigned long)pgdat->node_mem_map);
5456 free_area_init_core(pgdat);
5459 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5461 #if MAX_NUMNODES > 1
5463 * Figure out the number of possible node ids.
5465 void __init setup_nr_node_ids(void)
5467 unsigned int highest;
5469 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5470 nr_node_ids = highest + 1;
5475 * node_map_pfn_alignment - determine the maximum internode alignment
5477 * This function should be called after node map is populated and sorted.
5478 * It calculates the maximum power of two alignment which can distinguish
5481 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5482 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5483 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5484 * shifted, 1GiB is enough and this function will indicate so.
5486 * This is used to test whether pfn -> nid mapping of the chosen memory
5487 * model has fine enough granularity to avoid incorrect mapping for the
5488 * populated node map.
5490 * Returns the determined alignment in pfn's. 0 if there is no alignment
5491 * requirement (single node).
5493 unsigned long __init node_map_pfn_alignment(void)
5495 unsigned long accl_mask = 0, last_end = 0;
5496 unsigned long start, end, mask;
5500 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5501 if (!start || last_nid < 0 || last_nid == nid) {
5508 * Start with a mask granular enough to pin-point to the
5509 * start pfn and tick off bits one-by-one until it becomes
5510 * too coarse to separate the current node from the last.
5512 mask = ~((1 << __ffs(start)) - 1);
5513 while (mask && last_end <= (start & (mask << 1)))
5516 /* accumulate all internode masks */
5520 /* convert mask to number of pages */
5521 return ~accl_mask + 1;
5524 /* Find the lowest pfn for a node */
5525 static unsigned long __init find_min_pfn_for_node(int nid)
5527 unsigned long min_pfn = ULONG_MAX;
5528 unsigned long start_pfn;
5531 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5532 min_pfn = min(min_pfn, start_pfn);
5534 if (min_pfn == ULONG_MAX) {
5536 "Could not find start_pfn for node %d\n", nid);
5544 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5546 * It returns the minimum PFN based on information provided via
5547 * memblock_set_node().
5549 unsigned long __init find_min_pfn_with_active_regions(void)
5551 return find_min_pfn_for_node(MAX_NUMNODES);
5555 * early_calculate_totalpages()
5556 * Sum pages in active regions for movable zone.
5557 * Populate N_MEMORY for calculating usable_nodes.
5559 static unsigned long __init early_calculate_totalpages(void)
5561 unsigned long totalpages = 0;
5562 unsigned long start_pfn, end_pfn;
5565 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5566 unsigned long pages = end_pfn - start_pfn;
5568 totalpages += pages;
5570 node_set_state(nid, N_MEMORY);
5576 * Find the PFN the Movable zone begins in each node. Kernel memory
5577 * is spread evenly between nodes as long as the nodes have enough
5578 * memory. When they don't, some nodes will have more kernelcore than
5581 static void __init find_zone_movable_pfns_for_nodes(void)
5584 unsigned long usable_startpfn;
5585 unsigned long kernelcore_node, kernelcore_remaining;
5586 /* save the state before borrow the nodemask */
5587 nodemask_t saved_node_state = node_states[N_MEMORY];
5588 unsigned long totalpages = early_calculate_totalpages();
5589 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5590 struct memblock_region *r;
5592 /* Need to find movable_zone earlier when movable_node is specified. */
5593 find_usable_zone_for_movable();
5596 * If movable_node is specified, ignore kernelcore and movablecore
5599 if (movable_node_is_enabled()) {
5600 for_each_memblock(memory, r) {
5601 if (!memblock_is_hotpluggable(r))
5606 usable_startpfn = PFN_DOWN(r->base);
5607 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5608 min(usable_startpfn, zone_movable_pfn[nid]) :
5616 * If kernelcore=mirror is specified, ignore movablecore option
5618 if (mirrored_kernelcore) {
5619 bool mem_below_4gb_not_mirrored = false;
5621 for_each_memblock(memory, r) {
5622 if (memblock_is_mirror(r))
5627 usable_startpfn = memblock_region_memory_base_pfn(r);
5629 if (usable_startpfn < 0x100000) {
5630 mem_below_4gb_not_mirrored = true;
5634 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5635 min(usable_startpfn, zone_movable_pfn[nid]) :
5639 if (mem_below_4gb_not_mirrored)
5640 pr_warn("This configuration results in unmirrored kernel memory.");
5646 * If movablecore=nn[KMG] was specified, calculate what size of
5647 * kernelcore that corresponds so that memory usable for
5648 * any allocation type is evenly spread. If both kernelcore
5649 * and movablecore are specified, then the value of kernelcore
5650 * will be used for required_kernelcore if it's greater than
5651 * what movablecore would have allowed.
5653 if (required_movablecore) {
5654 unsigned long corepages;
5657 * Round-up so that ZONE_MOVABLE is at least as large as what
5658 * was requested by the user
5660 required_movablecore =
5661 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5662 required_movablecore = min(totalpages, required_movablecore);
5663 corepages = totalpages - required_movablecore;
5665 required_kernelcore = max(required_kernelcore, corepages);
5669 * If kernelcore was not specified or kernelcore size is larger
5670 * than totalpages, there is no ZONE_MOVABLE.
5672 if (!required_kernelcore || required_kernelcore >= totalpages)
5675 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5676 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5679 /* Spread kernelcore memory as evenly as possible throughout nodes */
5680 kernelcore_node = required_kernelcore / usable_nodes;
5681 for_each_node_state(nid, N_MEMORY) {
5682 unsigned long start_pfn, end_pfn;
5685 * Recalculate kernelcore_node if the division per node
5686 * now exceeds what is necessary to satisfy the requested
5687 * amount of memory for the kernel
5689 if (required_kernelcore < kernelcore_node)
5690 kernelcore_node = required_kernelcore / usable_nodes;
5693 * As the map is walked, we track how much memory is usable
5694 * by the kernel using kernelcore_remaining. When it is
5695 * 0, the rest of the node is usable by ZONE_MOVABLE
5697 kernelcore_remaining = kernelcore_node;
5699 /* Go through each range of PFNs within this node */
5700 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5701 unsigned long size_pages;
5703 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5704 if (start_pfn >= end_pfn)
5707 /* Account for what is only usable for kernelcore */
5708 if (start_pfn < usable_startpfn) {
5709 unsigned long kernel_pages;
5710 kernel_pages = min(end_pfn, usable_startpfn)
5713 kernelcore_remaining -= min(kernel_pages,
5714 kernelcore_remaining);
5715 required_kernelcore -= min(kernel_pages,
5716 required_kernelcore);
5718 /* Continue if range is now fully accounted */
5719 if (end_pfn <= usable_startpfn) {
5722 * Push zone_movable_pfn to the end so
5723 * that if we have to rebalance
5724 * kernelcore across nodes, we will
5725 * not double account here
5727 zone_movable_pfn[nid] = end_pfn;
5730 start_pfn = usable_startpfn;
5734 * The usable PFN range for ZONE_MOVABLE is from
5735 * start_pfn->end_pfn. Calculate size_pages as the
5736 * number of pages used as kernelcore
5738 size_pages = end_pfn - start_pfn;
5739 if (size_pages > kernelcore_remaining)
5740 size_pages = kernelcore_remaining;
5741 zone_movable_pfn[nid] = start_pfn + size_pages;
5744 * Some kernelcore has been met, update counts and
5745 * break if the kernelcore for this node has been
5748 required_kernelcore -= min(required_kernelcore,
5750 kernelcore_remaining -= size_pages;
5751 if (!kernelcore_remaining)
5757 * If there is still required_kernelcore, we do another pass with one
5758 * less node in the count. This will push zone_movable_pfn[nid] further
5759 * along on the nodes that still have memory until kernelcore is
5763 if (usable_nodes && required_kernelcore > usable_nodes)
5767 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5768 for (nid = 0; nid < MAX_NUMNODES; nid++)
5769 zone_movable_pfn[nid] =
5770 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5773 /* restore the node_state */
5774 node_states[N_MEMORY] = saved_node_state;
5777 /* Any regular or high memory on that node ? */
5778 static void check_for_memory(pg_data_t *pgdat, int nid)
5780 enum zone_type zone_type;
5782 if (N_MEMORY == N_NORMAL_MEMORY)
5785 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5786 struct zone *zone = &pgdat->node_zones[zone_type];
5787 if (populated_zone(zone)) {
5788 node_set_state(nid, N_HIGH_MEMORY);
5789 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5790 zone_type <= ZONE_NORMAL)
5791 node_set_state(nid, N_NORMAL_MEMORY);
5798 * free_area_init_nodes - Initialise all pg_data_t and zone data
5799 * @max_zone_pfn: an array of max PFNs for each zone
5801 * This will call free_area_init_node() for each active node in the system.
5802 * Using the page ranges provided by memblock_set_node(), the size of each
5803 * zone in each node and their holes is calculated. If the maximum PFN
5804 * between two adjacent zones match, it is assumed that the zone is empty.
5805 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5806 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5807 * starts where the previous one ended. For example, ZONE_DMA32 starts
5808 * at arch_max_dma_pfn.
5810 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5812 unsigned long start_pfn, end_pfn;
5815 /* Record where the zone boundaries are */
5816 memset(arch_zone_lowest_possible_pfn, 0,
5817 sizeof(arch_zone_lowest_possible_pfn));
5818 memset(arch_zone_highest_possible_pfn, 0,
5819 sizeof(arch_zone_highest_possible_pfn));
5820 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5821 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5822 for (i = 1; i < MAX_NR_ZONES; i++) {
5823 if (i == ZONE_MOVABLE)
5825 arch_zone_lowest_possible_pfn[i] =
5826 arch_zone_highest_possible_pfn[i-1];
5827 arch_zone_highest_possible_pfn[i] =
5828 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5830 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5831 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5833 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5834 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5835 find_zone_movable_pfns_for_nodes();
5837 /* Print out the zone ranges */
5838 pr_info("Zone ranges:\n");
5839 for (i = 0; i < MAX_NR_ZONES; i++) {
5840 if (i == ZONE_MOVABLE)
5842 pr_info(" %-8s ", zone_names[i]);
5843 if (arch_zone_lowest_possible_pfn[i] ==
5844 arch_zone_highest_possible_pfn[i])
5847 pr_cont("[mem %#018Lx-%#018Lx]\n",
5848 (u64)arch_zone_lowest_possible_pfn[i]
5850 ((u64)arch_zone_highest_possible_pfn[i]
5851 << PAGE_SHIFT) - 1);
5854 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5855 pr_info("Movable zone start for each node\n");
5856 for (i = 0; i < MAX_NUMNODES; i++) {
5857 if (zone_movable_pfn[i])
5858 pr_info(" Node %d: %#018Lx\n", i,
5859 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5862 /* Print out the early node map */
5863 pr_info("Early memory node ranges\n");
5864 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5865 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5866 (u64)start_pfn << PAGE_SHIFT,
5867 ((u64)end_pfn << PAGE_SHIFT) - 1);
5869 /* Initialise every node */
5870 mminit_verify_pageflags_layout();
5871 setup_nr_node_ids();
5872 for_each_online_node(nid) {
5873 pg_data_t *pgdat = NODE_DATA(nid);
5874 free_area_init_node(nid, NULL,
5875 find_min_pfn_for_node(nid), NULL);
5877 /* Any memory on that node */
5878 if (pgdat->node_present_pages)
5879 node_set_state(nid, N_MEMORY);
5880 check_for_memory(pgdat, nid);
5884 static int __init cmdline_parse_core(char *p, unsigned long *core)
5886 unsigned long long coremem;
5890 coremem = memparse(p, &p);
5891 *core = coremem >> PAGE_SHIFT;
5893 /* Paranoid check that UL is enough for the coremem value */
5894 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5900 * kernelcore=size sets the amount of memory for use for allocations that
5901 * cannot be reclaimed or migrated.
5903 static int __init cmdline_parse_kernelcore(char *p)
5905 /* parse kernelcore=mirror */
5906 if (parse_option_str(p, "mirror")) {
5907 mirrored_kernelcore = true;
5911 return cmdline_parse_core(p, &required_kernelcore);
5915 * movablecore=size sets the amount of memory for use for allocations that
5916 * can be reclaimed or migrated.
5918 static int __init cmdline_parse_movablecore(char *p)
5920 return cmdline_parse_core(p, &required_movablecore);
5923 early_param("kernelcore", cmdline_parse_kernelcore);
5924 early_param("movablecore", cmdline_parse_movablecore);
5926 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5928 void adjust_managed_page_count(struct page *page, long count)
5930 spin_lock(&managed_page_count_lock);
5931 page_zone(page)->managed_pages += count;
5932 totalram_pages += count;
5933 #ifdef CONFIG_HIGHMEM
5934 if (PageHighMem(page))
5935 totalhigh_pages += count;
5937 spin_unlock(&managed_page_count_lock);
5939 EXPORT_SYMBOL(adjust_managed_page_count);
5941 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5944 unsigned long pages = 0;
5946 start = (void *)PAGE_ALIGN((unsigned long)start);
5947 end = (void *)((unsigned long)end & PAGE_MASK);
5948 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5949 if ((unsigned int)poison <= 0xFF)
5950 memset(pos, poison, PAGE_SIZE);
5951 free_reserved_page(virt_to_page(pos));
5955 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5956 s, pages << (PAGE_SHIFT - 10), start, end);
5960 EXPORT_SYMBOL(free_reserved_area);
5962 #ifdef CONFIG_HIGHMEM
5963 void free_highmem_page(struct page *page)
5965 __free_reserved_page(page);
5967 page_zone(page)->managed_pages++;
5973 void __init mem_init_print_info(const char *str)
5975 unsigned long physpages, codesize, datasize, rosize, bss_size;
5976 unsigned long init_code_size, init_data_size;
5978 physpages = get_num_physpages();
5979 codesize = _etext - _stext;
5980 datasize = _edata - _sdata;
5981 rosize = __end_rodata - __start_rodata;
5982 bss_size = __bss_stop - __bss_start;
5983 init_data_size = __init_end - __init_begin;
5984 init_code_size = _einittext - _sinittext;
5987 * Detect special cases and adjust section sizes accordingly:
5988 * 1) .init.* may be embedded into .data sections
5989 * 2) .init.text.* may be out of [__init_begin, __init_end],
5990 * please refer to arch/tile/kernel/vmlinux.lds.S.
5991 * 3) .rodata.* may be embedded into .text or .data sections.
5993 #define adj_init_size(start, end, size, pos, adj) \
5995 if (start <= pos && pos < end && size > adj) \
5999 adj_init_size(__init_begin, __init_end, init_data_size,
6000 _sinittext, init_code_size);
6001 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6002 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6003 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6004 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6006 #undef adj_init_size
6008 pr_info("Memory: %luK/%luK available "
6009 "(%luK kernel code, %luK rwdata, %luK rodata, "
6010 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
6011 #ifdef CONFIG_HIGHMEM
6015 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
6016 codesize >> 10, datasize >> 10, rosize >> 10,
6017 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6018 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
6019 totalcma_pages << (PAGE_SHIFT-10),
6020 #ifdef CONFIG_HIGHMEM
6021 totalhigh_pages << (PAGE_SHIFT-10),
6023 str ? ", " : "", str ? str : "");
6027 * set_dma_reserve - set the specified number of pages reserved in the first zone
6028 * @new_dma_reserve: The number of pages to mark reserved
6030 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6031 * In the DMA zone, a significant percentage may be consumed by kernel image
6032 * and other unfreeable allocations which can skew the watermarks badly. This
6033 * function may optionally be used to account for unfreeable pages in the
6034 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6035 * smaller per-cpu batchsize.
6037 void __init set_dma_reserve(unsigned long new_dma_reserve)
6039 dma_reserve = new_dma_reserve;
6042 void __init free_area_init(unsigned long *zones_size)
6044 free_area_init_node(0, zones_size,
6045 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6048 static int page_alloc_cpu_notify(struct notifier_block *self,
6049 unsigned long action, void *hcpu)
6051 int cpu = (unsigned long)hcpu;
6053 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6054 lru_add_drain_cpu(cpu);
6058 * Spill the event counters of the dead processor
6059 * into the current processors event counters.
6060 * This artificially elevates the count of the current
6063 vm_events_fold_cpu(cpu);
6066 * Zero the differential counters of the dead processor
6067 * so that the vm statistics are consistent.
6069 * This is only okay since the processor is dead and cannot
6070 * race with what we are doing.
6072 cpu_vm_stats_fold(cpu);
6077 void __init page_alloc_init(void)
6079 hotcpu_notifier(page_alloc_cpu_notify, 0);
6083 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6084 * or min_free_kbytes changes.
6086 static void calculate_totalreserve_pages(void)
6088 struct pglist_data *pgdat;
6089 unsigned long reserve_pages = 0;
6090 enum zone_type i, j;
6092 for_each_online_pgdat(pgdat) {
6093 for (i = 0; i < MAX_NR_ZONES; i++) {
6094 struct zone *zone = pgdat->node_zones + i;
6097 /* Find valid and maximum lowmem_reserve in the zone */
6098 for (j = i; j < MAX_NR_ZONES; j++) {
6099 if (zone->lowmem_reserve[j] > max)
6100 max = zone->lowmem_reserve[j];
6103 /* we treat the high watermark as reserved pages. */
6104 max += high_wmark_pages(zone);
6106 if (max > zone->managed_pages)
6107 max = zone->managed_pages;
6109 zone->totalreserve_pages = max;
6111 reserve_pages += max;
6114 totalreserve_pages = reserve_pages;
6118 * setup_per_zone_lowmem_reserve - called whenever
6119 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6120 * has a correct pages reserved value, so an adequate number of
6121 * pages are left in the zone after a successful __alloc_pages().
6123 static void setup_per_zone_lowmem_reserve(void)
6125 struct pglist_data *pgdat;
6126 enum zone_type j, idx;
6128 for_each_online_pgdat(pgdat) {
6129 for (j = 0; j < MAX_NR_ZONES; j++) {
6130 struct zone *zone = pgdat->node_zones + j;
6131 unsigned long managed_pages = zone->managed_pages;
6133 zone->lowmem_reserve[j] = 0;
6137 struct zone *lower_zone;
6141 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6142 sysctl_lowmem_reserve_ratio[idx] = 1;
6144 lower_zone = pgdat->node_zones + idx;
6145 lower_zone->lowmem_reserve[j] = managed_pages /
6146 sysctl_lowmem_reserve_ratio[idx];
6147 managed_pages += lower_zone->managed_pages;
6152 /* update totalreserve_pages */
6153 calculate_totalreserve_pages();
6156 static void __setup_per_zone_wmarks(void)
6158 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6159 unsigned long lowmem_pages = 0;
6161 unsigned long flags;
6163 /* Calculate total number of !ZONE_HIGHMEM pages */
6164 for_each_zone(zone) {
6165 if (!is_highmem(zone))
6166 lowmem_pages += zone->managed_pages;
6169 for_each_zone(zone) {
6172 spin_lock_irqsave(&zone->lock, flags);
6173 tmp = (u64)pages_min * zone->managed_pages;
6174 do_div(tmp, lowmem_pages);
6175 if (is_highmem(zone)) {
6177 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6178 * need highmem pages, so cap pages_min to a small
6181 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6182 * deltas control asynch page reclaim, and so should
6183 * not be capped for highmem.
6185 unsigned long min_pages;
6187 min_pages = zone->managed_pages / 1024;
6188 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6189 zone->watermark[WMARK_MIN] = min_pages;
6192 * If it's a lowmem zone, reserve a number of pages
6193 * proportionate to the zone's size.
6195 zone->watermark[WMARK_MIN] = tmp;
6198 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
6199 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
6201 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6202 high_wmark_pages(zone) - low_wmark_pages(zone) -
6203 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6205 spin_unlock_irqrestore(&zone->lock, flags);
6208 /* update totalreserve_pages */
6209 calculate_totalreserve_pages();
6213 * setup_per_zone_wmarks - called when min_free_kbytes changes
6214 * or when memory is hot-{added|removed}
6216 * Ensures that the watermark[min,low,high] values for each zone are set
6217 * correctly with respect to min_free_kbytes.
6219 void setup_per_zone_wmarks(void)
6221 mutex_lock(&zonelists_mutex);
6222 __setup_per_zone_wmarks();
6223 mutex_unlock(&zonelists_mutex);
6227 * The inactive anon list should be small enough that the VM never has to
6228 * do too much work, but large enough that each inactive page has a chance
6229 * to be referenced again before it is swapped out.
6231 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6232 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6233 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6234 * the anonymous pages are kept on the inactive list.
6237 * memory ratio inactive anon
6238 * -------------------------------------
6247 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6249 unsigned int gb, ratio;
6251 /* Zone size in gigabytes */
6252 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6254 ratio = int_sqrt(10 * gb);
6258 zone->inactive_ratio = ratio;
6261 static void __meminit setup_per_zone_inactive_ratio(void)
6266 calculate_zone_inactive_ratio(zone);
6270 * Initialise min_free_kbytes.
6272 * For small machines we want it small (128k min). For large machines
6273 * we want it large (64MB max). But it is not linear, because network
6274 * bandwidth does not increase linearly with machine size. We use
6276 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6277 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6293 int __meminit init_per_zone_wmark_min(void)
6295 unsigned long lowmem_kbytes;
6296 int new_min_free_kbytes;
6298 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6299 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6301 if (new_min_free_kbytes > user_min_free_kbytes) {
6302 min_free_kbytes = new_min_free_kbytes;
6303 if (min_free_kbytes < 128)
6304 min_free_kbytes = 128;
6305 if (min_free_kbytes > 65536)
6306 min_free_kbytes = 65536;
6308 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6309 new_min_free_kbytes, user_min_free_kbytes);
6311 setup_per_zone_wmarks();
6312 refresh_zone_stat_thresholds();
6313 setup_per_zone_lowmem_reserve();
6314 setup_per_zone_inactive_ratio();
6317 module_init(init_per_zone_wmark_min)
6320 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6321 * that we can call two helper functions whenever min_free_kbytes
6324 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6325 void __user *buffer, size_t *length, loff_t *ppos)
6329 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6334 user_min_free_kbytes = min_free_kbytes;
6335 setup_per_zone_wmarks();
6341 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6342 void __user *buffer, size_t *length, loff_t *ppos)
6347 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6352 zone->min_unmapped_pages = (zone->managed_pages *
6353 sysctl_min_unmapped_ratio) / 100;
6357 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6358 void __user *buffer, size_t *length, loff_t *ppos)
6363 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6368 zone->min_slab_pages = (zone->managed_pages *
6369 sysctl_min_slab_ratio) / 100;
6375 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6376 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6377 * whenever sysctl_lowmem_reserve_ratio changes.
6379 * The reserve ratio obviously has absolutely no relation with the
6380 * minimum watermarks. The lowmem reserve ratio can only make sense
6381 * if in function of the boot time zone sizes.
6383 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6384 void __user *buffer, size_t *length, loff_t *ppos)
6386 proc_dointvec_minmax(table, write, buffer, length, ppos);
6387 setup_per_zone_lowmem_reserve();
6392 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6393 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6394 * pagelist can have before it gets flushed back to buddy allocator.
6396 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6397 void __user *buffer, size_t *length, loff_t *ppos)
6400 int old_percpu_pagelist_fraction;
6403 mutex_lock(&pcp_batch_high_lock);
6404 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6406 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6407 if (!write || ret < 0)
6410 /* Sanity checking to avoid pcp imbalance */
6411 if (percpu_pagelist_fraction &&
6412 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6413 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6419 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6422 for_each_populated_zone(zone) {
6425 for_each_possible_cpu(cpu)
6426 pageset_set_high_and_batch(zone,
6427 per_cpu_ptr(zone->pageset, cpu));
6430 mutex_unlock(&pcp_batch_high_lock);
6435 int hashdist = HASHDIST_DEFAULT;
6437 static int __init set_hashdist(char *str)
6441 hashdist = simple_strtoul(str, &str, 0);
6444 __setup("hashdist=", set_hashdist);
6448 * allocate a large system hash table from bootmem
6449 * - it is assumed that the hash table must contain an exact power-of-2
6450 * quantity of entries
6451 * - limit is the number of hash buckets, not the total allocation size
6453 void *__init alloc_large_system_hash(const char *tablename,
6454 unsigned long bucketsize,
6455 unsigned long numentries,
6458 unsigned int *_hash_shift,
6459 unsigned int *_hash_mask,
6460 unsigned long low_limit,
6461 unsigned long high_limit)
6463 unsigned long long max = high_limit;
6464 unsigned long log2qty, size;
6467 /* allow the kernel cmdline to have a say */
6469 /* round applicable memory size up to nearest megabyte */
6470 numentries = nr_kernel_pages;
6472 /* It isn't necessary when PAGE_SIZE >= 1MB */
6473 if (PAGE_SHIFT < 20)
6474 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6476 /* limit to 1 bucket per 2^scale bytes of low memory */
6477 if (scale > PAGE_SHIFT)
6478 numentries >>= (scale - PAGE_SHIFT);
6480 numentries <<= (PAGE_SHIFT - scale);
6482 /* Make sure we've got at least a 0-order allocation.. */
6483 if (unlikely(flags & HASH_SMALL)) {
6484 /* Makes no sense without HASH_EARLY */
6485 WARN_ON(!(flags & HASH_EARLY));
6486 if (!(numentries >> *_hash_shift)) {
6487 numentries = 1UL << *_hash_shift;
6488 BUG_ON(!numentries);
6490 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6491 numentries = PAGE_SIZE / bucketsize;
6493 numentries = roundup_pow_of_two(numentries);
6495 /* limit allocation size to 1/16 total memory by default */
6497 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6498 do_div(max, bucketsize);
6500 max = min(max, 0x80000000ULL);
6502 if (numentries < low_limit)
6503 numentries = low_limit;
6504 if (numentries > max)
6507 log2qty = ilog2(numentries);
6510 size = bucketsize << log2qty;
6511 if (flags & HASH_EARLY)
6512 table = memblock_virt_alloc_nopanic(size, 0);
6514 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6517 * If bucketsize is not a power-of-two, we may free
6518 * some pages at the end of hash table which
6519 * alloc_pages_exact() automatically does
6521 if (get_order(size) < MAX_ORDER) {
6522 table = alloc_pages_exact(size, GFP_ATOMIC);
6523 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6526 } while (!table && size > PAGE_SIZE && --log2qty);
6529 panic("Failed to allocate %s hash table\n", tablename);
6531 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6534 ilog2(size) - PAGE_SHIFT,
6538 *_hash_shift = log2qty;
6540 *_hash_mask = (1 << log2qty) - 1;
6545 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6546 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6549 #ifdef CONFIG_SPARSEMEM
6550 return __pfn_to_section(pfn)->pageblock_flags;
6552 return zone->pageblock_flags;
6553 #endif /* CONFIG_SPARSEMEM */
6556 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6558 #ifdef CONFIG_SPARSEMEM
6559 pfn &= (PAGES_PER_SECTION-1);
6560 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6562 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6563 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6564 #endif /* CONFIG_SPARSEMEM */
6568 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6569 * @page: The page within the block of interest
6570 * @pfn: The target page frame number
6571 * @end_bitidx: The last bit of interest to retrieve
6572 * @mask: mask of bits that the caller is interested in
6574 * Return: pageblock_bits flags
6576 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6577 unsigned long end_bitidx,
6581 unsigned long *bitmap;
6582 unsigned long bitidx, word_bitidx;
6585 zone = page_zone(page);
6586 bitmap = get_pageblock_bitmap(zone, pfn);
6587 bitidx = pfn_to_bitidx(zone, pfn);
6588 word_bitidx = bitidx / BITS_PER_LONG;
6589 bitidx &= (BITS_PER_LONG-1);
6591 word = bitmap[word_bitidx];
6592 bitidx += end_bitidx;
6593 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6597 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6598 * @page: The page within the block of interest
6599 * @flags: The flags to set
6600 * @pfn: The target page frame number
6601 * @end_bitidx: The last bit of interest
6602 * @mask: mask of bits that the caller is interested in
6604 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6606 unsigned long end_bitidx,
6610 unsigned long *bitmap;
6611 unsigned long bitidx, word_bitidx;
6612 unsigned long old_word, word;
6614 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6616 zone = page_zone(page);
6617 bitmap = get_pageblock_bitmap(zone, pfn);
6618 bitidx = pfn_to_bitidx(zone, pfn);
6619 word_bitidx = bitidx / BITS_PER_LONG;
6620 bitidx &= (BITS_PER_LONG-1);
6622 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6624 bitidx += end_bitidx;
6625 mask <<= (BITS_PER_LONG - bitidx - 1);
6626 flags <<= (BITS_PER_LONG - bitidx - 1);
6628 word = READ_ONCE(bitmap[word_bitidx]);
6630 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6631 if (word == old_word)
6638 * This function checks whether pageblock includes unmovable pages or not.
6639 * If @count is not zero, it is okay to include less @count unmovable pages
6641 * PageLRU check without isolation or lru_lock could race so that
6642 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6643 * expect this function should be exact.
6645 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6646 bool skip_hwpoisoned_pages)
6648 unsigned long pfn, iter, found;
6652 * For avoiding noise data, lru_add_drain_all() should be called
6653 * If ZONE_MOVABLE, the zone never contains unmovable pages
6655 if (zone_idx(zone) == ZONE_MOVABLE)
6657 mt = get_pageblock_migratetype(page);
6658 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6661 pfn = page_to_pfn(page);
6662 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6663 unsigned long check = pfn + iter;
6665 if (!pfn_valid_within(check))
6668 page = pfn_to_page(check);
6671 * Hugepages are not in LRU lists, but they're movable.
6672 * We need not scan over tail pages bacause we don't
6673 * handle each tail page individually in migration.
6675 if (PageHuge(page)) {
6676 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6681 * We can't use page_count without pin a page
6682 * because another CPU can free compound page.
6683 * This check already skips compound tails of THP
6684 * because their page->_count is zero at all time.
6686 if (!atomic_read(&page->_count)) {
6687 if (PageBuddy(page))
6688 iter += (1 << page_order(page)) - 1;
6693 * The HWPoisoned page may be not in buddy system, and
6694 * page_count() is not 0.
6696 if (skip_hwpoisoned_pages && PageHWPoison(page))
6702 * If there are RECLAIMABLE pages, we need to check
6703 * it. But now, memory offline itself doesn't call
6704 * shrink_node_slabs() and it still to be fixed.
6707 * If the page is not RAM, page_count()should be 0.
6708 * we don't need more check. This is an _used_ not-movable page.
6710 * The problematic thing here is PG_reserved pages. PG_reserved
6711 * is set to both of a memory hole page and a _used_ kernel
6720 bool is_pageblock_removable_nolock(struct page *page)
6726 * We have to be careful here because we are iterating over memory
6727 * sections which are not zone aware so we might end up outside of
6728 * the zone but still within the section.
6729 * We have to take care about the node as well. If the node is offline
6730 * its NODE_DATA will be NULL - see page_zone.
6732 if (!node_online(page_to_nid(page)))
6735 zone = page_zone(page);
6736 pfn = page_to_pfn(page);
6737 if (!zone_spans_pfn(zone, pfn))
6740 return !has_unmovable_pages(zone, page, 0, true);
6743 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
6745 static unsigned long pfn_max_align_down(unsigned long pfn)
6747 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6748 pageblock_nr_pages) - 1);
6751 static unsigned long pfn_max_align_up(unsigned long pfn)
6753 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6754 pageblock_nr_pages));
6757 /* [start, end) must belong to a single zone. */
6758 static int __alloc_contig_migrate_range(struct compact_control *cc,
6759 unsigned long start, unsigned long end)
6761 /* This function is based on compact_zone() from compaction.c. */
6762 unsigned long nr_reclaimed;
6763 unsigned long pfn = start;
6764 unsigned int tries = 0;
6769 while (pfn < end || !list_empty(&cc->migratepages)) {
6770 if (fatal_signal_pending(current)) {
6775 if (list_empty(&cc->migratepages)) {
6776 cc->nr_migratepages = 0;
6777 pfn = isolate_migratepages_range(cc, pfn, end);
6783 } else if (++tries == 5) {
6784 ret = ret < 0 ? ret : -EBUSY;
6788 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6790 cc->nr_migratepages -= nr_reclaimed;
6792 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6793 NULL, 0, cc->mode, MR_CMA);
6796 putback_movable_pages(&cc->migratepages);
6803 * alloc_contig_range() -- tries to allocate given range of pages
6804 * @start: start PFN to allocate
6805 * @end: one-past-the-last PFN to allocate
6806 * @migratetype: migratetype of the underlaying pageblocks (either
6807 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6808 * in range must have the same migratetype and it must
6809 * be either of the two.
6811 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6812 * aligned, however it's the caller's responsibility to guarantee that
6813 * we are the only thread that changes migrate type of pageblocks the
6816 * The PFN range must belong to a single zone.
6818 * Returns zero on success or negative error code. On success all
6819 * pages which PFN is in [start, end) are allocated for the caller and
6820 * need to be freed with free_contig_range().
6822 int alloc_contig_range(unsigned long start, unsigned long end,
6823 unsigned migratetype)
6825 unsigned long outer_start, outer_end;
6829 struct compact_control cc = {
6830 .nr_migratepages = 0,
6832 .zone = page_zone(pfn_to_page(start)),
6833 .mode = MIGRATE_SYNC,
6834 .ignore_skip_hint = true,
6836 INIT_LIST_HEAD(&cc.migratepages);
6839 * What we do here is we mark all pageblocks in range as
6840 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6841 * have different sizes, and due to the way page allocator
6842 * work, we align the range to biggest of the two pages so
6843 * that page allocator won't try to merge buddies from
6844 * different pageblocks and change MIGRATE_ISOLATE to some
6845 * other migration type.
6847 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6848 * migrate the pages from an unaligned range (ie. pages that
6849 * we are interested in). This will put all the pages in
6850 * range back to page allocator as MIGRATE_ISOLATE.
6852 * When this is done, we take the pages in range from page
6853 * allocator removing them from the buddy system. This way
6854 * page allocator will never consider using them.
6856 * This lets us mark the pageblocks back as
6857 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6858 * aligned range but not in the unaligned, original range are
6859 * put back to page allocator so that buddy can use them.
6862 ret = start_isolate_page_range(pfn_max_align_down(start),
6863 pfn_max_align_up(end), migratetype,
6869 * In case of -EBUSY, we'd like to know which page causes problem.
6870 * So, just fall through. We will check it in test_pages_isolated().
6872 ret = __alloc_contig_migrate_range(&cc, start, end);
6873 if (ret && ret != -EBUSY)
6877 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6878 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6879 * more, all pages in [start, end) are free in page allocator.
6880 * What we are going to do is to allocate all pages from
6881 * [start, end) (that is remove them from page allocator).
6883 * The only problem is that pages at the beginning and at the
6884 * end of interesting range may be not aligned with pages that
6885 * page allocator holds, ie. they can be part of higher order
6886 * pages. Because of this, we reserve the bigger range and
6887 * once this is done free the pages we are not interested in.
6889 * We don't have to hold zone->lock here because the pages are
6890 * isolated thus they won't get removed from buddy.
6893 lru_add_drain_all();
6894 drain_all_pages(cc.zone);
6897 outer_start = start;
6898 while (!PageBuddy(pfn_to_page(outer_start))) {
6899 if (++order >= MAX_ORDER) {
6900 outer_start = start;
6903 outer_start &= ~0UL << order;
6906 if (outer_start != start) {
6907 order = page_order(pfn_to_page(outer_start));
6910 * outer_start page could be small order buddy page and
6911 * it doesn't include start page. Adjust outer_start
6912 * in this case to report failed page properly
6913 * on tracepoint in test_pages_isolated()
6915 if (outer_start + (1UL << order) <= start)
6916 outer_start = start;
6919 /* Make sure the range is really isolated. */
6920 if (test_pages_isolated(outer_start, end, false)) {
6921 pr_info("%s: [%lx, %lx) PFNs busy\n",
6922 __func__, outer_start, end);
6927 /* Grab isolated pages from freelists. */
6928 outer_end = isolate_freepages_range(&cc, outer_start, end);
6934 /* Free head and tail (if any) */
6935 if (start != outer_start)
6936 free_contig_range(outer_start, start - outer_start);
6937 if (end != outer_end)
6938 free_contig_range(end, outer_end - end);
6941 undo_isolate_page_range(pfn_max_align_down(start),
6942 pfn_max_align_up(end), migratetype);
6946 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6948 unsigned int count = 0;
6950 for (; nr_pages--; pfn++) {
6951 struct page *page = pfn_to_page(pfn);
6953 count += page_count(page) != 1;
6956 WARN(count != 0, "%d pages are still in use!\n", count);
6960 #ifdef CONFIG_MEMORY_HOTPLUG
6962 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6963 * page high values need to be recalulated.
6965 void __meminit zone_pcp_update(struct zone *zone)
6968 mutex_lock(&pcp_batch_high_lock);
6969 for_each_possible_cpu(cpu)
6970 pageset_set_high_and_batch(zone,
6971 per_cpu_ptr(zone->pageset, cpu));
6972 mutex_unlock(&pcp_batch_high_lock);
6976 void zone_pcp_reset(struct zone *zone)
6978 unsigned long flags;
6980 struct per_cpu_pageset *pset;
6982 /* avoid races with drain_pages() */
6983 local_irq_save(flags);
6984 if (zone->pageset != &boot_pageset) {
6985 for_each_online_cpu(cpu) {
6986 pset = per_cpu_ptr(zone->pageset, cpu);
6987 drain_zonestat(zone, pset);
6989 free_percpu(zone->pageset);
6990 zone->pageset = &boot_pageset;
6992 local_irq_restore(flags);
6995 #ifdef CONFIG_MEMORY_HOTREMOVE
6997 * All pages in the range must be isolated before calling this.
7000 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7004 unsigned int order, i;
7006 unsigned long flags;
7007 /* find the first valid pfn */
7008 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7013 zone = page_zone(pfn_to_page(pfn));
7014 spin_lock_irqsave(&zone->lock, flags);
7016 while (pfn < end_pfn) {
7017 if (!pfn_valid(pfn)) {
7021 page = pfn_to_page(pfn);
7023 * The HWPoisoned page may be not in buddy system, and
7024 * page_count() is not 0.
7026 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7028 SetPageReserved(page);
7032 BUG_ON(page_count(page));
7033 BUG_ON(!PageBuddy(page));
7034 order = page_order(page);
7035 #ifdef CONFIG_DEBUG_VM
7036 printk(KERN_INFO "remove from free list %lx %d %lx\n",
7037 pfn, 1 << order, end_pfn);
7039 list_del(&page->lru);
7040 rmv_page_order(page);
7041 zone->free_area[order].nr_free--;
7042 for (i = 0; i < (1 << order); i++)
7043 SetPageReserved((page+i));
7044 pfn += (1 << order);
7046 spin_unlock_irqrestore(&zone->lock, flags);
7050 #ifdef CONFIG_MEMORY_FAILURE
7051 bool is_free_buddy_page(struct page *page)
7053 struct zone *zone = page_zone(page);
7054 unsigned long pfn = page_to_pfn(page);
7055 unsigned long flags;
7058 spin_lock_irqsave(&zone->lock, flags);
7059 for (order = 0; order < MAX_ORDER; order++) {
7060 struct page *page_head = page - (pfn & ((1 << order) - 1));
7062 if (PageBuddy(page_head) && page_order(page_head) >= order)
7065 spin_unlock_irqrestore(&zone->lock, flags);
7067 return order < MAX_ORDER;