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/stop_machine.h>
47 #include <linux/sort.h>
48 #include <linux/pfn.h>
49 #include <linux/backing-dev.h>
50 #include <linux/fault-inject.h>
51 #include <linux/page-isolation.h>
52 #include <linux/page_ext.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <linux/prefetch.h>
58 #include <linux/mm_inline.h>
59 #include <linux/migrate.h>
60 #include <linux/page_ext.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/page_owner.h>
64 #include <linux/kthread.h>
66 #include <asm/sections.h>
67 #include <asm/tlbflush.h>
68 #include <asm/div64.h>
71 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
72 static DEFINE_MUTEX(pcp_batch_high_lock);
73 #define MIN_PERCPU_PAGELIST_FRACTION (8)
75 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
76 DEFINE_PER_CPU(int, numa_node);
77 EXPORT_PER_CPU_SYMBOL(numa_node);
80 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
82 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
83 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
84 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
85 * defined in <linux/topology.h>.
87 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
88 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
89 int _node_numa_mem_[MAX_NUMNODES];
93 * Array of node states.
95 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
96 [N_POSSIBLE] = NODE_MASK_ALL,
97 [N_ONLINE] = { { [0] = 1UL } },
99 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
100 #ifdef CONFIG_HIGHMEM
101 [N_HIGH_MEMORY] = { { [0] = 1UL } },
103 #ifdef CONFIG_MOVABLE_NODE
104 [N_MEMORY] = { { [0] = 1UL } },
106 [N_CPU] = { { [0] = 1UL } },
109 EXPORT_SYMBOL(node_states);
111 /* Protect totalram_pages and zone->managed_pages */
112 static DEFINE_SPINLOCK(managed_page_count_lock);
114 unsigned long totalram_pages __read_mostly;
115 unsigned long totalreserve_pages __read_mostly;
116 unsigned long totalcma_pages __read_mostly;
118 int percpu_pagelist_fraction;
119 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
122 * A cached value of the page's pageblock's migratetype, used when the page is
123 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
124 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
125 * Also the migratetype set in the page does not necessarily match the pcplist
126 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
127 * other index - this ensures that it will be put on the correct CMA freelist.
129 static inline int get_pcppage_migratetype(struct page *page)
134 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
136 page->index = migratetype;
139 #ifdef CONFIG_PM_SLEEP
141 * The following functions are used by the suspend/hibernate code to temporarily
142 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
143 * while devices are suspended. To avoid races with the suspend/hibernate code,
144 * they should always be called with pm_mutex held (gfp_allowed_mask also should
145 * only be modified with pm_mutex held, unless the suspend/hibernate code is
146 * guaranteed not to run in parallel with that modification).
149 static gfp_t saved_gfp_mask;
151 void pm_restore_gfp_mask(void)
153 WARN_ON(!mutex_is_locked(&pm_mutex));
154 if (saved_gfp_mask) {
155 gfp_allowed_mask = saved_gfp_mask;
160 void pm_restrict_gfp_mask(void)
162 WARN_ON(!mutex_is_locked(&pm_mutex));
163 WARN_ON(saved_gfp_mask);
164 saved_gfp_mask = gfp_allowed_mask;
165 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
168 bool pm_suspended_storage(void)
170 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
174 #endif /* CONFIG_PM_SLEEP */
176 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
177 unsigned int pageblock_order __read_mostly;
180 static void __free_pages_ok(struct page *page, unsigned int order);
183 * results with 256, 32 in the lowmem_reserve sysctl:
184 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
185 * 1G machine -> (16M dma, 784M normal, 224M high)
186 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
187 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
188 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
190 * TBD: should special case ZONE_DMA32 machines here - in those we normally
191 * don't need any ZONE_NORMAL reservation
193 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
194 #ifdef CONFIG_ZONE_DMA
197 #ifdef CONFIG_ZONE_DMA32
200 #ifdef CONFIG_HIGHMEM
206 EXPORT_SYMBOL(totalram_pages);
208 static char * const zone_names[MAX_NR_ZONES] = {
209 #ifdef CONFIG_ZONE_DMA
212 #ifdef CONFIG_ZONE_DMA32
216 #ifdef CONFIG_HIGHMEM
220 #ifdef CONFIG_ZONE_DEVICE
225 static void free_compound_page(struct page *page);
226 compound_page_dtor * const compound_page_dtors[] = {
229 #ifdef CONFIG_HUGETLB_PAGE
234 int min_free_kbytes = 1024;
235 int user_min_free_kbytes = -1;
237 static unsigned long __meminitdata nr_kernel_pages;
238 static unsigned long __meminitdata nr_all_pages;
239 static unsigned long __meminitdata dma_reserve;
241 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
242 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
243 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
244 static unsigned long __initdata required_kernelcore;
245 static unsigned long __initdata required_movablecore;
246 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
248 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
250 EXPORT_SYMBOL(movable_zone);
251 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
254 int nr_node_ids __read_mostly = MAX_NUMNODES;
255 int nr_online_nodes __read_mostly = 1;
256 EXPORT_SYMBOL(nr_node_ids);
257 EXPORT_SYMBOL(nr_online_nodes);
260 int page_group_by_mobility_disabled __read_mostly;
262 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
263 static inline void reset_deferred_meminit(pg_data_t *pgdat)
265 pgdat->first_deferred_pfn = ULONG_MAX;
268 /* Returns true if the struct page for the pfn is uninitialised */
269 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
271 if (pfn >= NODE_DATA(early_pfn_to_nid(pfn))->first_deferred_pfn)
277 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
279 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
286 * Returns false when the remaining initialisation should be deferred until
287 * later in the boot cycle when it can be parallelised.
289 static inline bool update_defer_init(pg_data_t *pgdat,
290 unsigned long pfn, unsigned long zone_end,
291 unsigned long *nr_initialised)
293 /* Always populate low zones for address-contrained allocations */
294 if (zone_end < pgdat_end_pfn(pgdat))
297 /* Initialise at least 2G of the highest zone */
299 if (*nr_initialised > (2UL << (30 - PAGE_SHIFT)) &&
300 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
301 pgdat->first_deferred_pfn = pfn;
308 static inline void reset_deferred_meminit(pg_data_t *pgdat)
312 static inline bool early_page_uninitialised(unsigned long pfn)
317 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
322 static inline bool update_defer_init(pg_data_t *pgdat,
323 unsigned long pfn, unsigned long zone_end,
324 unsigned long *nr_initialised)
331 void set_pageblock_migratetype(struct page *page, int migratetype)
333 if (unlikely(page_group_by_mobility_disabled &&
334 migratetype < MIGRATE_PCPTYPES))
335 migratetype = MIGRATE_UNMOVABLE;
337 set_pageblock_flags_group(page, (unsigned long)migratetype,
338 PB_migrate, PB_migrate_end);
341 #ifdef CONFIG_DEBUG_VM
342 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
346 unsigned long pfn = page_to_pfn(page);
347 unsigned long sp, start_pfn;
350 seq = zone_span_seqbegin(zone);
351 start_pfn = zone->zone_start_pfn;
352 sp = zone->spanned_pages;
353 if (!zone_spans_pfn(zone, pfn))
355 } while (zone_span_seqretry(zone, seq));
358 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
359 pfn, zone_to_nid(zone), zone->name,
360 start_pfn, start_pfn + sp);
365 static int page_is_consistent(struct zone *zone, struct page *page)
367 if (!pfn_valid_within(page_to_pfn(page)))
369 if (zone != page_zone(page))
375 * Temporary debugging check for pages not lying within a given zone.
377 static int bad_range(struct zone *zone, struct page *page)
379 if (page_outside_zone_boundaries(zone, page))
381 if (!page_is_consistent(zone, page))
387 static inline int bad_range(struct zone *zone, struct page *page)
393 static void bad_page(struct page *page, const char *reason,
394 unsigned long bad_flags)
396 static unsigned long resume;
397 static unsigned long nr_shown;
398 static unsigned long nr_unshown;
400 /* Don't complain about poisoned pages */
401 if (PageHWPoison(page)) {
402 page_mapcount_reset(page); /* remove PageBuddy */
407 * Allow a burst of 60 reports, then keep quiet for that minute;
408 * or allow a steady drip of one report per second.
410 if (nr_shown == 60) {
411 if (time_before(jiffies, resume)) {
417 "BUG: Bad page state: %lu messages suppressed\n",
424 resume = jiffies + 60 * HZ;
426 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
427 current->comm, page_to_pfn(page));
428 dump_page_badflags(page, reason, bad_flags);
433 /* Leave bad fields for debug, except PageBuddy could make trouble */
434 page_mapcount_reset(page); /* remove PageBuddy */
435 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
439 * Higher-order pages are called "compound pages". They are structured thusly:
441 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
443 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
444 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
446 * The first tail page's ->compound_dtor holds the offset in array of compound
447 * page destructors. See compound_page_dtors.
449 * The first tail page's ->compound_order holds the order of allocation.
450 * This usage means that zero-order pages may not be compound.
453 static void free_compound_page(struct page *page)
455 __free_pages_ok(page, compound_order(page));
458 void prep_compound_page(struct page *page, unsigned int order)
461 int nr_pages = 1 << order;
463 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
464 set_compound_order(page, order);
466 for (i = 1; i < nr_pages; i++) {
467 struct page *p = page + i;
468 set_page_count(p, 0);
469 set_compound_head(p, page);
473 #ifdef CONFIG_DEBUG_PAGEALLOC
474 unsigned int _debug_guardpage_minorder;
475 bool _debug_pagealloc_enabled __read_mostly;
476 bool _debug_guardpage_enabled __read_mostly;
478 static int __init early_debug_pagealloc(char *buf)
483 if (strcmp(buf, "on") == 0)
484 _debug_pagealloc_enabled = true;
488 early_param("debug_pagealloc", early_debug_pagealloc);
490 static bool need_debug_guardpage(void)
492 /* If we don't use debug_pagealloc, we don't need guard page */
493 if (!debug_pagealloc_enabled())
499 static void init_debug_guardpage(void)
501 if (!debug_pagealloc_enabled())
504 _debug_guardpage_enabled = true;
507 struct page_ext_operations debug_guardpage_ops = {
508 .need = need_debug_guardpage,
509 .init = init_debug_guardpage,
512 static int __init debug_guardpage_minorder_setup(char *buf)
516 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
517 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
520 _debug_guardpage_minorder = res;
521 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
524 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
526 static inline void set_page_guard(struct zone *zone, struct page *page,
527 unsigned int order, int migratetype)
529 struct page_ext *page_ext;
531 if (!debug_guardpage_enabled())
534 page_ext = lookup_page_ext(page);
535 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
537 INIT_LIST_HEAD(&page->lru);
538 set_page_private(page, order);
539 /* Guard pages are not available for any usage */
540 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
543 static inline void clear_page_guard(struct zone *zone, struct page *page,
544 unsigned int order, int migratetype)
546 struct page_ext *page_ext;
548 if (!debug_guardpage_enabled())
551 page_ext = lookup_page_ext(page);
552 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
554 set_page_private(page, 0);
555 if (!is_migrate_isolate(migratetype))
556 __mod_zone_freepage_state(zone, (1 << order), migratetype);
559 struct page_ext_operations debug_guardpage_ops = { NULL, };
560 static inline void set_page_guard(struct zone *zone, struct page *page,
561 unsigned int order, int migratetype) {}
562 static inline void clear_page_guard(struct zone *zone, struct page *page,
563 unsigned int order, int migratetype) {}
566 static inline void set_page_order(struct page *page, unsigned int order)
568 set_page_private(page, order);
569 __SetPageBuddy(page);
572 static inline void rmv_page_order(struct page *page)
574 __ClearPageBuddy(page);
575 set_page_private(page, 0);
579 * This function checks whether a page is free && is the buddy
580 * we can do coalesce a page and its buddy if
581 * (a) the buddy is not in a hole &&
582 * (b) the buddy is in the buddy system &&
583 * (c) a page and its buddy have the same order &&
584 * (d) a page and its buddy are in the same zone.
586 * For recording whether a page is in the buddy system, we set ->_mapcount
587 * PAGE_BUDDY_MAPCOUNT_VALUE.
588 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
589 * serialized by zone->lock.
591 * For recording page's order, we use page_private(page).
593 static inline int page_is_buddy(struct page *page, struct page *buddy,
596 if (!pfn_valid_within(page_to_pfn(buddy)))
599 if (page_is_guard(buddy) && page_order(buddy) == order) {
600 if (page_zone_id(page) != page_zone_id(buddy))
603 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
608 if (PageBuddy(buddy) && page_order(buddy) == order) {
610 * zone check is done late to avoid uselessly
611 * calculating zone/node ids for pages that could
614 if (page_zone_id(page) != page_zone_id(buddy))
617 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
625 * Freeing function for a buddy system allocator.
627 * The concept of a buddy system is to maintain direct-mapped table
628 * (containing bit values) for memory blocks of various "orders".
629 * The bottom level table contains the map for the smallest allocatable
630 * units of memory (here, pages), and each level above it describes
631 * pairs of units from the levels below, hence, "buddies".
632 * At a high level, all that happens here is marking the table entry
633 * at the bottom level available, and propagating the changes upward
634 * as necessary, plus some accounting needed to play nicely with other
635 * parts of the VM system.
636 * At each level, we keep a list of pages, which are heads of continuous
637 * free pages of length of (1 << order) and marked with _mapcount
638 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
640 * So when we are allocating or freeing one, we can derive the state of the
641 * other. That is, if we allocate a small block, and both were
642 * free, the remainder of the region must be split into blocks.
643 * If a block is freed, and its buddy is also free, then this
644 * triggers coalescing into a block of larger size.
649 static inline void __free_one_page(struct page *page,
651 struct zone *zone, unsigned int order,
654 unsigned long page_idx;
655 unsigned long combined_idx;
656 unsigned long uninitialized_var(buddy_idx);
658 unsigned int max_order = MAX_ORDER;
660 VM_BUG_ON(!zone_is_initialized(zone));
661 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
663 VM_BUG_ON(migratetype == -1);
664 if (is_migrate_isolate(migratetype)) {
666 * We restrict max order of merging to prevent merge
667 * between freepages on isolate pageblock and normal
668 * pageblock. Without this, pageblock isolation
669 * could cause incorrect freepage accounting.
671 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
673 __mod_zone_freepage_state(zone, 1 << order, migratetype);
676 page_idx = pfn & ((1 << max_order) - 1);
678 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
679 VM_BUG_ON_PAGE(bad_range(zone, page), page);
681 while (order < max_order - 1) {
682 buddy_idx = __find_buddy_index(page_idx, order);
683 buddy = page + (buddy_idx - page_idx);
684 if (!page_is_buddy(page, buddy, order))
687 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
688 * merge with it and move up one order.
690 if (page_is_guard(buddy)) {
691 clear_page_guard(zone, buddy, order, migratetype);
693 list_del(&buddy->lru);
694 zone->free_area[order].nr_free--;
695 rmv_page_order(buddy);
697 combined_idx = buddy_idx & page_idx;
698 page = page + (combined_idx - page_idx);
699 page_idx = combined_idx;
702 set_page_order(page, order);
705 * If this is not the largest possible page, check if the buddy
706 * of the next-highest order is free. If it is, it's possible
707 * that pages are being freed that will coalesce soon. In case,
708 * that is happening, add the free page to the tail of the list
709 * so it's less likely to be used soon and more likely to be merged
710 * as a higher order page
712 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
713 struct page *higher_page, *higher_buddy;
714 combined_idx = buddy_idx & page_idx;
715 higher_page = page + (combined_idx - page_idx);
716 buddy_idx = __find_buddy_index(combined_idx, order + 1);
717 higher_buddy = higher_page + (buddy_idx - combined_idx);
718 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
719 list_add_tail(&page->lru,
720 &zone->free_area[order].free_list[migratetype]);
725 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
727 zone->free_area[order].nr_free++;
730 static inline int free_pages_check(struct page *page)
732 const char *bad_reason = NULL;
733 unsigned long bad_flags = 0;
735 if (unlikely(page_mapcount(page)))
736 bad_reason = "nonzero mapcount";
737 if (unlikely(page->mapping != NULL))
738 bad_reason = "non-NULL mapping";
739 if (unlikely(atomic_read(&page->_count) != 0))
740 bad_reason = "nonzero _count";
741 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
742 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
743 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
746 if (unlikely(page->mem_cgroup))
747 bad_reason = "page still charged to cgroup";
749 if (unlikely(bad_reason)) {
750 bad_page(page, bad_reason, bad_flags);
753 page_cpupid_reset_last(page);
754 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
755 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
760 * Frees a number of pages from the PCP lists
761 * Assumes all pages on list are in same zone, and of same order.
762 * count is the number of pages to free.
764 * If the zone was previously in an "all pages pinned" state then look to
765 * see if this freeing clears that state.
767 * And clear the zone's pages_scanned counter, to hold off the "all pages are
768 * pinned" detection logic.
770 static void free_pcppages_bulk(struct zone *zone, int count,
771 struct per_cpu_pages *pcp)
776 unsigned long nr_scanned;
778 spin_lock(&zone->lock);
779 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
781 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
785 struct list_head *list;
788 * Remove pages from lists in a round-robin fashion. A
789 * batch_free count is maintained that is incremented when an
790 * empty list is encountered. This is so more pages are freed
791 * off fuller lists instead of spinning excessively around empty
796 if (++migratetype == MIGRATE_PCPTYPES)
798 list = &pcp->lists[migratetype];
799 } while (list_empty(list));
801 /* This is the only non-empty list. Free them all. */
802 if (batch_free == MIGRATE_PCPTYPES)
803 batch_free = to_free;
806 int mt; /* migratetype of the to-be-freed page */
808 page = list_last_entry(list, struct page, lru);
809 /* must delete as __free_one_page list manipulates */
810 list_del(&page->lru);
812 mt = get_pcppage_migratetype(page);
813 /* MIGRATE_ISOLATE page should not go to pcplists */
814 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
815 /* Pageblock could have been isolated meanwhile */
816 if (unlikely(has_isolate_pageblock(zone)))
817 mt = get_pageblock_migratetype(page);
819 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
820 trace_mm_page_pcpu_drain(page, 0, mt);
821 } while (--to_free && --batch_free && !list_empty(list));
823 spin_unlock(&zone->lock);
826 static void free_one_page(struct zone *zone,
827 struct page *page, unsigned long pfn,
831 unsigned long nr_scanned;
832 spin_lock(&zone->lock);
833 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
835 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
837 if (unlikely(has_isolate_pageblock(zone) ||
838 is_migrate_isolate(migratetype))) {
839 migratetype = get_pfnblock_migratetype(page, pfn);
841 __free_one_page(page, pfn, zone, order, migratetype);
842 spin_unlock(&zone->lock);
845 static int free_tail_pages_check(struct page *head_page, struct page *page)
850 * We rely page->lru.next never has bit 0 set, unless the page
851 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
853 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
855 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
859 if (unlikely(!PageTail(page))) {
860 bad_page(page, "PageTail not set", 0);
863 if (unlikely(compound_head(page) != head_page)) {
864 bad_page(page, "compound_head not consistent", 0);
869 clear_compound_head(page);
873 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
874 unsigned long zone, int nid)
876 set_page_links(page, zone, nid, pfn);
877 init_page_count(page);
878 page_mapcount_reset(page);
879 page_cpupid_reset_last(page);
881 INIT_LIST_HEAD(&page->lru);
882 #ifdef WANT_PAGE_VIRTUAL
883 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
884 if (!is_highmem_idx(zone))
885 set_page_address(page, __va(pfn << PAGE_SHIFT));
889 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
892 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
895 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
896 static void init_reserved_page(unsigned long pfn)
901 if (!early_page_uninitialised(pfn))
904 nid = early_pfn_to_nid(pfn);
905 pgdat = NODE_DATA(nid);
907 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
908 struct zone *zone = &pgdat->node_zones[zid];
910 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
913 __init_single_pfn(pfn, zid, nid);
916 static inline void init_reserved_page(unsigned long pfn)
919 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
922 * Initialised pages do not have PageReserved set. This function is
923 * called for each range allocated by the bootmem allocator and
924 * marks the pages PageReserved. The remaining valid pages are later
925 * sent to the buddy page allocator.
927 void __meminit reserve_bootmem_region(unsigned long start, unsigned long end)
929 unsigned long start_pfn = PFN_DOWN(start);
930 unsigned long end_pfn = PFN_UP(end);
932 for (; start_pfn < end_pfn; start_pfn++) {
933 if (pfn_valid(start_pfn)) {
934 struct page *page = pfn_to_page(start_pfn);
936 init_reserved_page(start_pfn);
938 /* Avoid false-positive PageTail() */
939 INIT_LIST_HEAD(&page->lru);
941 SetPageReserved(page);
946 static bool free_pages_prepare(struct page *page, unsigned int order)
948 bool compound = PageCompound(page);
951 VM_BUG_ON_PAGE(PageTail(page), page);
952 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
954 trace_mm_page_free(page, order);
955 kmemcheck_free_shadow(page, order);
956 kasan_free_pages(page, order);
959 page->mapping = NULL;
960 bad += free_pages_check(page);
961 for (i = 1; i < (1 << order); i++) {
963 bad += free_tail_pages_check(page, page + i);
964 bad += free_pages_check(page + i);
969 reset_page_owner(page, order);
971 if (!PageHighMem(page)) {
972 debug_check_no_locks_freed(page_address(page),
974 debug_check_no_obj_freed(page_address(page),
977 arch_free_page(page, order);
978 kernel_map_pages(page, 1 << order, 0);
983 static void __free_pages_ok(struct page *page, unsigned int order)
987 unsigned long pfn = page_to_pfn(page);
989 if (!free_pages_prepare(page, order))
992 migratetype = get_pfnblock_migratetype(page, pfn);
993 local_irq_save(flags);
994 __count_vm_events(PGFREE, 1 << order);
995 free_one_page(page_zone(page), page, pfn, order, migratetype);
996 local_irq_restore(flags);
999 static void __init __free_pages_boot_core(struct page *page,
1000 unsigned long pfn, unsigned int order)
1002 unsigned int nr_pages = 1 << order;
1003 struct page *p = page;
1007 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1009 __ClearPageReserved(p);
1010 set_page_count(p, 0);
1012 __ClearPageReserved(p);
1013 set_page_count(p, 0);
1015 page_zone(page)->managed_pages += nr_pages;
1016 set_page_refcounted(page);
1017 __free_pages(page, order);
1020 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1021 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1023 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1025 int __meminit early_pfn_to_nid(unsigned long pfn)
1027 static DEFINE_SPINLOCK(early_pfn_lock);
1030 spin_lock(&early_pfn_lock);
1031 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1034 spin_unlock(&early_pfn_lock);
1040 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1041 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1042 struct mminit_pfnnid_cache *state)
1046 nid = __early_pfn_to_nid(pfn, state);
1047 if (nid >= 0 && nid != node)
1052 /* Only safe to use early in boot when initialisation is single-threaded */
1053 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1055 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1060 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1064 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1065 struct mminit_pfnnid_cache *state)
1072 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1075 if (early_page_uninitialised(pfn))
1077 return __free_pages_boot_core(page, pfn, order);
1080 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1081 static void __init deferred_free_range(struct page *page,
1082 unsigned long pfn, int nr_pages)
1089 /* Free a large naturally-aligned chunk if possible */
1090 if (nr_pages == MAX_ORDER_NR_PAGES &&
1091 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1092 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1093 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1097 for (i = 0; i < nr_pages; i++, page++, pfn++)
1098 __free_pages_boot_core(page, pfn, 0);
1101 /* Completion tracking for deferred_init_memmap() threads */
1102 static atomic_t pgdat_init_n_undone __initdata;
1103 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1105 static inline void __init pgdat_init_report_one_done(void)
1107 if (atomic_dec_and_test(&pgdat_init_n_undone))
1108 complete(&pgdat_init_all_done_comp);
1111 /* Initialise remaining memory on a node */
1112 static int __init deferred_init_memmap(void *data)
1114 pg_data_t *pgdat = data;
1115 int nid = pgdat->node_id;
1116 struct mminit_pfnnid_cache nid_init_state = { };
1117 unsigned long start = jiffies;
1118 unsigned long nr_pages = 0;
1119 unsigned long walk_start, walk_end;
1122 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1123 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1125 if (first_init_pfn == ULONG_MAX) {
1126 pgdat_init_report_one_done();
1130 /* Bind memory initialisation thread to a local node if possible */
1131 if (!cpumask_empty(cpumask))
1132 set_cpus_allowed_ptr(current, cpumask);
1134 /* Sanity check boundaries */
1135 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1136 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1137 pgdat->first_deferred_pfn = ULONG_MAX;
1139 /* Only the highest zone is deferred so find it */
1140 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1141 zone = pgdat->node_zones + zid;
1142 if (first_init_pfn < zone_end_pfn(zone))
1146 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1147 unsigned long pfn, end_pfn;
1148 struct page *page = NULL;
1149 struct page *free_base_page = NULL;
1150 unsigned long free_base_pfn = 0;
1153 end_pfn = min(walk_end, zone_end_pfn(zone));
1154 pfn = first_init_pfn;
1155 if (pfn < walk_start)
1157 if (pfn < zone->zone_start_pfn)
1158 pfn = zone->zone_start_pfn;
1160 for (; pfn < end_pfn; pfn++) {
1161 if (!pfn_valid_within(pfn))
1165 * Ensure pfn_valid is checked every
1166 * MAX_ORDER_NR_PAGES for memory holes
1168 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1169 if (!pfn_valid(pfn)) {
1175 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1180 /* Minimise pfn page lookups and scheduler checks */
1181 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1184 nr_pages += nr_to_free;
1185 deferred_free_range(free_base_page,
1186 free_base_pfn, nr_to_free);
1187 free_base_page = NULL;
1188 free_base_pfn = nr_to_free = 0;
1190 page = pfn_to_page(pfn);
1195 VM_BUG_ON(page_zone(page) != zone);
1199 __init_single_page(page, pfn, zid, nid);
1200 if (!free_base_page) {
1201 free_base_page = page;
1202 free_base_pfn = pfn;
1207 /* Where possible, batch up pages for a single free */
1210 /* Free the current block of pages to allocator */
1211 nr_pages += nr_to_free;
1212 deferred_free_range(free_base_page, free_base_pfn,
1214 free_base_page = NULL;
1215 free_base_pfn = nr_to_free = 0;
1218 first_init_pfn = max(end_pfn, first_init_pfn);
1221 /* Sanity check that the next zone really is unpopulated */
1222 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1224 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1225 jiffies_to_msecs(jiffies - start));
1227 pgdat_init_report_one_done();
1231 void __init page_alloc_init_late(void)
1235 /* There will be num_node_state(N_MEMORY) threads */
1236 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1237 for_each_node_state(nid, N_MEMORY) {
1238 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1241 /* Block until all are initialised */
1242 wait_for_completion(&pgdat_init_all_done_comp);
1244 /* Reinit limits that are based on free pages after the kernel is up */
1245 files_maxfiles_init();
1247 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1250 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1251 void __init init_cma_reserved_pageblock(struct page *page)
1253 unsigned i = pageblock_nr_pages;
1254 struct page *p = page;
1257 __ClearPageReserved(p);
1258 set_page_count(p, 0);
1261 set_pageblock_migratetype(page, MIGRATE_CMA);
1263 if (pageblock_order >= MAX_ORDER) {
1264 i = pageblock_nr_pages;
1267 set_page_refcounted(p);
1268 __free_pages(p, MAX_ORDER - 1);
1269 p += MAX_ORDER_NR_PAGES;
1270 } while (i -= MAX_ORDER_NR_PAGES);
1272 set_page_refcounted(page);
1273 __free_pages(page, pageblock_order);
1276 adjust_managed_page_count(page, pageblock_nr_pages);
1281 * The order of subdivision here is critical for the IO subsystem.
1282 * Please do not alter this order without good reasons and regression
1283 * testing. Specifically, as large blocks of memory are subdivided,
1284 * the order in which smaller blocks are delivered depends on the order
1285 * they're subdivided in this function. This is the primary factor
1286 * influencing the order in which pages are delivered to the IO
1287 * subsystem according to empirical testing, and this is also justified
1288 * by considering the behavior of a buddy system containing a single
1289 * large block of memory acted on by a series of small allocations.
1290 * This behavior is a critical factor in sglist merging's success.
1294 static inline void expand(struct zone *zone, struct page *page,
1295 int low, int high, struct free_area *area,
1298 unsigned long size = 1 << high;
1300 while (high > low) {
1304 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1306 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1307 debug_guardpage_enabled() &&
1308 high < debug_guardpage_minorder()) {
1310 * Mark as guard pages (or page), that will allow to
1311 * merge back to allocator when buddy will be freed.
1312 * Corresponding page table entries will not be touched,
1313 * pages will stay not present in virtual address space
1315 set_page_guard(zone, &page[size], high, migratetype);
1318 list_add(&page[size].lru, &area->free_list[migratetype]);
1320 set_page_order(&page[size], high);
1325 * This page is about to be returned from the page allocator
1327 static inline int check_new_page(struct page *page)
1329 const char *bad_reason = NULL;
1330 unsigned long bad_flags = 0;
1332 if (unlikely(page_mapcount(page)))
1333 bad_reason = "nonzero mapcount";
1334 if (unlikely(page->mapping != NULL))
1335 bad_reason = "non-NULL mapping";
1336 if (unlikely(atomic_read(&page->_count) != 0))
1337 bad_reason = "nonzero _count";
1338 if (unlikely(page->flags & __PG_HWPOISON)) {
1339 bad_reason = "HWPoisoned (hardware-corrupted)";
1340 bad_flags = __PG_HWPOISON;
1342 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1343 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1344 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1347 if (unlikely(page->mem_cgroup))
1348 bad_reason = "page still charged to cgroup";
1350 if (unlikely(bad_reason)) {
1351 bad_page(page, bad_reason, bad_flags);
1357 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1362 for (i = 0; i < (1 << order); i++) {
1363 struct page *p = page + i;
1364 if (unlikely(check_new_page(p)))
1368 set_page_private(page, 0);
1369 set_page_refcounted(page);
1371 arch_alloc_page(page, order);
1372 kernel_map_pages(page, 1 << order, 1);
1373 kasan_alloc_pages(page, order);
1375 if (gfp_flags & __GFP_ZERO)
1376 for (i = 0; i < (1 << order); i++)
1377 clear_highpage(page + i);
1379 if (order && (gfp_flags & __GFP_COMP))
1380 prep_compound_page(page, order);
1382 set_page_owner(page, order, gfp_flags);
1385 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1386 * allocate the page. The expectation is that the caller is taking
1387 * steps that will free more memory. The caller should avoid the page
1388 * being used for !PFMEMALLOC purposes.
1390 if (alloc_flags & ALLOC_NO_WATERMARKS)
1391 set_page_pfmemalloc(page);
1393 clear_page_pfmemalloc(page);
1399 * Go through the free lists for the given migratetype and remove
1400 * the smallest available page from the freelists
1403 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1406 unsigned int current_order;
1407 struct free_area *area;
1410 /* Find a page of the appropriate size in the preferred list */
1411 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1412 area = &(zone->free_area[current_order]);
1413 page = list_first_entry_or_null(&area->free_list[migratetype],
1417 list_del(&page->lru);
1418 rmv_page_order(page);
1420 expand(zone, page, order, current_order, area, migratetype);
1421 set_pcppage_migratetype(page, migratetype);
1430 * This array describes the order lists are fallen back to when
1431 * the free lists for the desirable migrate type are depleted
1433 static int fallbacks[MIGRATE_TYPES][4] = {
1434 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1435 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1436 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1438 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1440 #ifdef CONFIG_MEMORY_ISOLATION
1441 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1446 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1449 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1452 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1453 unsigned int order) { return NULL; }
1457 * Move the free pages in a range to the free lists of the requested type.
1458 * Note that start_page and end_pages are not aligned on a pageblock
1459 * boundary. If alignment is required, use move_freepages_block()
1461 int move_freepages(struct zone *zone,
1462 struct page *start_page, struct page *end_page,
1467 int pages_moved = 0;
1469 #ifndef CONFIG_HOLES_IN_ZONE
1471 * page_zone is not safe to call in this context when
1472 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1473 * anyway as we check zone boundaries in move_freepages_block().
1474 * Remove at a later date when no bug reports exist related to
1475 * grouping pages by mobility
1477 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1480 for (page = start_page; page <= end_page;) {
1481 /* Make sure we are not inadvertently changing nodes */
1482 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1484 if (!pfn_valid_within(page_to_pfn(page))) {
1489 if (!PageBuddy(page)) {
1494 order = page_order(page);
1495 list_move(&page->lru,
1496 &zone->free_area[order].free_list[migratetype]);
1498 pages_moved += 1 << order;
1504 int move_freepages_block(struct zone *zone, struct page *page,
1507 unsigned long start_pfn, end_pfn;
1508 struct page *start_page, *end_page;
1510 start_pfn = page_to_pfn(page);
1511 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1512 start_page = pfn_to_page(start_pfn);
1513 end_page = start_page + pageblock_nr_pages - 1;
1514 end_pfn = start_pfn + pageblock_nr_pages - 1;
1516 /* Do not cross zone boundaries */
1517 if (!zone_spans_pfn(zone, start_pfn))
1519 if (!zone_spans_pfn(zone, end_pfn))
1522 return move_freepages(zone, start_page, end_page, migratetype);
1525 static void change_pageblock_range(struct page *pageblock_page,
1526 int start_order, int migratetype)
1528 int nr_pageblocks = 1 << (start_order - pageblock_order);
1530 while (nr_pageblocks--) {
1531 set_pageblock_migratetype(pageblock_page, migratetype);
1532 pageblock_page += pageblock_nr_pages;
1537 * When we are falling back to another migratetype during allocation, try to
1538 * steal extra free pages from the same pageblocks to satisfy further
1539 * allocations, instead of polluting multiple pageblocks.
1541 * If we are stealing a relatively large buddy page, it is likely there will
1542 * be more free pages in the pageblock, so try to steal them all. For
1543 * reclaimable and unmovable allocations, we steal regardless of page size,
1544 * as fragmentation caused by those allocations polluting movable pageblocks
1545 * is worse than movable allocations stealing from unmovable and reclaimable
1548 static bool can_steal_fallback(unsigned int order, int start_mt)
1551 * Leaving this order check is intended, although there is
1552 * relaxed order check in next check. The reason is that
1553 * we can actually steal whole pageblock if this condition met,
1554 * but, below check doesn't guarantee it and that is just heuristic
1555 * so could be changed anytime.
1557 if (order >= pageblock_order)
1560 if (order >= pageblock_order / 2 ||
1561 start_mt == MIGRATE_RECLAIMABLE ||
1562 start_mt == MIGRATE_UNMOVABLE ||
1563 page_group_by_mobility_disabled)
1570 * This function implements actual steal behaviour. If order is large enough,
1571 * we can steal whole pageblock. If not, we first move freepages in this
1572 * pageblock and check whether half of pages are moved or not. If half of
1573 * pages are moved, we can change migratetype of pageblock and permanently
1574 * use it's pages as requested migratetype in the future.
1576 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1579 unsigned int current_order = page_order(page);
1582 /* Take ownership for orders >= pageblock_order */
1583 if (current_order >= pageblock_order) {
1584 change_pageblock_range(page, current_order, start_type);
1588 pages = move_freepages_block(zone, page, start_type);
1590 /* Claim the whole block if over half of it is free */
1591 if (pages >= (1 << (pageblock_order-1)) ||
1592 page_group_by_mobility_disabled)
1593 set_pageblock_migratetype(page, start_type);
1597 * Check whether there is a suitable fallback freepage with requested order.
1598 * If only_stealable is true, this function returns fallback_mt only if
1599 * we can steal other freepages all together. This would help to reduce
1600 * fragmentation due to mixed migratetype pages in one pageblock.
1602 int find_suitable_fallback(struct free_area *area, unsigned int order,
1603 int migratetype, bool only_stealable, bool *can_steal)
1608 if (area->nr_free == 0)
1613 fallback_mt = fallbacks[migratetype][i];
1614 if (fallback_mt == MIGRATE_TYPES)
1617 if (list_empty(&area->free_list[fallback_mt]))
1620 if (can_steal_fallback(order, migratetype))
1623 if (!only_stealable)
1634 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1635 * there are no empty page blocks that contain a page with a suitable order
1637 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1638 unsigned int alloc_order)
1641 unsigned long max_managed, flags;
1644 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1645 * Check is race-prone but harmless.
1647 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1648 if (zone->nr_reserved_highatomic >= max_managed)
1651 spin_lock_irqsave(&zone->lock, flags);
1653 /* Recheck the nr_reserved_highatomic limit under the lock */
1654 if (zone->nr_reserved_highatomic >= max_managed)
1658 mt = get_pageblock_migratetype(page);
1659 if (mt != MIGRATE_HIGHATOMIC &&
1660 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1661 zone->nr_reserved_highatomic += pageblock_nr_pages;
1662 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1663 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1667 spin_unlock_irqrestore(&zone->lock, flags);
1671 * Used when an allocation is about to fail under memory pressure. This
1672 * potentially hurts the reliability of high-order allocations when under
1673 * intense memory pressure but failed atomic allocations should be easier
1674 * to recover from than an OOM.
1676 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1678 struct zonelist *zonelist = ac->zonelist;
1679 unsigned long flags;
1685 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1687 /* Preserve at least one pageblock */
1688 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1691 spin_lock_irqsave(&zone->lock, flags);
1692 for (order = 0; order < MAX_ORDER; order++) {
1693 struct free_area *area = &(zone->free_area[order]);
1695 page = list_first_entry_or_null(
1696 &area->free_list[MIGRATE_HIGHATOMIC],
1702 * It should never happen but changes to locking could
1703 * inadvertently allow a per-cpu drain to add pages
1704 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1705 * and watch for underflows.
1707 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
1708 zone->nr_reserved_highatomic);
1711 * Convert to ac->migratetype and avoid the normal
1712 * pageblock stealing heuristics. Minimally, the caller
1713 * is doing the work and needs the pages. More
1714 * importantly, if the block was always converted to
1715 * MIGRATE_UNMOVABLE or another type then the number
1716 * of pageblocks that cannot be completely freed
1719 set_pageblock_migratetype(page, ac->migratetype);
1720 move_freepages_block(zone, page, ac->migratetype);
1721 spin_unlock_irqrestore(&zone->lock, flags);
1724 spin_unlock_irqrestore(&zone->lock, flags);
1728 /* Remove an element from the buddy allocator from the fallback list */
1729 static inline struct page *
1730 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1732 struct free_area *area;
1733 unsigned int current_order;
1738 /* Find the largest possible block of pages in the other list */
1739 for (current_order = MAX_ORDER-1;
1740 current_order >= order && current_order <= MAX_ORDER-1;
1742 area = &(zone->free_area[current_order]);
1743 fallback_mt = find_suitable_fallback(area, current_order,
1744 start_migratetype, false, &can_steal);
1745 if (fallback_mt == -1)
1748 page = list_first_entry(&area->free_list[fallback_mt],
1751 steal_suitable_fallback(zone, page, start_migratetype);
1753 /* Remove the page from the freelists */
1755 list_del(&page->lru);
1756 rmv_page_order(page);
1758 expand(zone, page, order, current_order, area,
1761 * The pcppage_migratetype may differ from pageblock's
1762 * migratetype depending on the decisions in
1763 * find_suitable_fallback(). This is OK as long as it does not
1764 * differ for MIGRATE_CMA pageblocks. Those can be used as
1765 * fallback only via special __rmqueue_cma_fallback() function
1767 set_pcppage_migratetype(page, start_migratetype);
1769 trace_mm_page_alloc_extfrag(page, order, current_order,
1770 start_migratetype, fallback_mt);
1779 * Do the hard work of removing an element from the buddy allocator.
1780 * Call me with the zone->lock already held.
1782 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1787 page = __rmqueue_smallest(zone, order, migratetype);
1788 if (unlikely(!page)) {
1789 if (migratetype == MIGRATE_MOVABLE)
1790 page = __rmqueue_cma_fallback(zone, order);
1793 page = __rmqueue_fallback(zone, order, migratetype);
1796 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1801 * Obtain a specified number of elements from the buddy allocator, all under
1802 * a single hold of the lock, for efficiency. Add them to the supplied list.
1803 * Returns the number of new pages which were placed at *list.
1805 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1806 unsigned long count, struct list_head *list,
1807 int migratetype, bool cold)
1811 spin_lock(&zone->lock);
1812 for (i = 0; i < count; ++i) {
1813 struct page *page = __rmqueue(zone, order, migratetype);
1814 if (unlikely(page == NULL))
1818 * Split buddy pages returned by expand() are received here
1819 * in physical page order. The page is added to the callers and
1820 * list and the list head then moves forward. From the callers
1821 * perspective, the linked list is ordered by page number in
1822 * some conditions. This is useful for IO devices that can
1823 * merge IO requests if the physical pages are ordered
1827 list_add(&page->lru, list);
1829 list_add_tail(&page->lru, list);
1831 if (is_migrate_cma(get_pcppage_migratetype(page)))
1832 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1835 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1836 spin_unlock(&zone->lock);
1842 * Called from the vmstat counter updater to drain pagesets of this
1843 * currently executing processor on remote nodes after they have
1846 * Note that this function must be called with the thread pinned to
1847 * a single processor.
1849 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1851 unsigned long flags;
1852 int to_drain, batch;
1854 local_irq_save(flags);
1855 batch = READ_ONCE(pcp->batch);
1856 to_drain = min(pcp->count, batch);
1858 free_pcppages_bulk(zone, to_drain, pcp);
1859 pcp->count -= to_drain;
1861 local_irq_restore(flags);
1866 * Drain pcplists of the indicated processor and zone.
1868 * The processor must either be the current processor and the
1869 * thread pinned to the current processor or a processor that
1872 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1874 unsigned long flags;
1875 struct per_cpu_pageset *pset;
1876 struct per_cpu_pages *pcp;
1878 local_irq_save(flags);
1879 pset = per_cpu_ptr(zone->pageset, cpu);
1883 free_pcppages_bulk(zone, pcp->count, pcp);
1886 local_irq_restore(flags);
1890 * Drain pcplists of all zones on the indicated processor.
1892 * The processor must either be the current processor and the
1893 * thread pinned to the current processor or a processor that
1896 static void drain_pages(unsigned int cpu)
1900 for_each_populated_zone(zone) {
1901 drain_pages_zone(cpu, zone);
1906 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1908 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1909 * the single zone's pages.
1911 void drain_local_pages(struct zone *zone)
1913 int cpu = smp_processor_id();
1916 drain_pages_zone(cpu, zone);
1922 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1924 * When zone parameter is non-NULL, spill just the single zone's pages.
1926 * Note that this code is protected against sending an IPI to an offline
1927 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1928 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1929 * nothing keeps CPUs from showing up after we populated the cpumask and
1930 * before the call to on_each_cpu_mask().
1932 void drain_all_pages(struct zone *zone)
1937 * Allocate in the BSS so we wont require allocation in
1938 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1940 static cpumask_t cpus_with_pcps;
1943 * We don't care about racing with CPU hotplug event
1944 * as offline notification will cause the notified
1945 * cpu to drain that CPU pcps and on_each_cpu_mask
1946 * disables preemption as part of its processing
1948 for_each_online_cpu(cpu) {
1949 struct per_cpu_pageset *pcp;
1951 bool has_pcps = false;
1954 pcp = per_cpu_ptr(zone->pageset, cpu);
1958 for_each_populated_zone(z) {
1959 pcp = per_cpu_ptr(z->pageset, cpu);
1960 if (pcp->pcp.count) {
1968 cpumask_set_cpu(cpu, &cpus_with_pcps);
1970 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1972 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
1976 #ifdef CONFIG_HIBERNATION
1978 void mark_free_pages(struct zone *zone)
1980 unsigned long pfn, max_zone_pfn;
1981 unsigned long flags;
1982 unsigned int order, t;
1985 if (zone_is_empty(zone))
1988 spin_lock_irqsave(&zone->lock, flags);
1990 max_zone_pfn = zone_end_pfn(zone);
1991 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1992 if (pfn_valid(pfn)) {
1993 page = pfn_to_page(pfn);
1994 if (!swsusp_page_is_forbidden(page))
1995 swsusp_unset_page_free(page);
1998 for_each_migratetype_order(order, t) {
1999 list_for_each_entry(page,
2000 &zone->free_area[order].free_list[t], lru) {
2003 pfn = page_to_pfn(page);
2004 for (i = 0; i < (1UL << order); i++)
2005 swsusp_set_page_free(pfn_to_page(pfn + i));
2008 spin_unlock_irqrestore(&zone->lock, flags);
2010 #endif /* CONFIG_PM */
2013 * Free a 0-order page
2014 * cold == true ? free a cold page : free a hot page
2016 void free_hot_cold_page(struct page *page, bool cold)
2018 struct zone *zone = page_zone(page);
2019 struct per_cpu_pages *pcp;
2020 unsigned long flags;
2021 unsigned long pfn = page_to_pfn(page);
2024 if (!free_pages_prepare(page, 0))
2027 migratetype = get_pfnblock_migratetype(page, pfn);
2028 set_pcppage_migratetype(page, migratetype);
2029 local_irq_save(flags);
2030 __count_vm_event(PGFREE);
2033 * We only track unmovable, reclaimable and movable on pcp lists.
2034 * Free ISOLATE pages back to the allocator because they are being
2035 * offlined but treat RESERVE as movable pages so we can get those
2036 * areas back if necessary. Otherwise, we may have to free
2037 * excessively into the page allocator
2039 if (migratetype >= MIGRATE_PCPTYPES) {
2040 if (unlikely(is_migrate_isolate(migratetype))) {
2041 free_one_page(zone, page, pfn, 0, migratetype);
2044 migratetype = MIGRATE_MOVABLE;
2047 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2049 list_add(&page->lru, &pcp->lists[migratetype]);
2051 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2053 if (pcp->count >= pcp->high) {
2054 unsigned long batch = READ_ONCE(pcp->batch);
2055 free_pcppages_bulk(zone, batch, pcp);
2056 pcp->count -= batch;
2060 local_irq_restore(flags);
2064 * Free a list of 0-order pages
2066 void free_hot_cold_page_list(struct list_head *list, bool cold)
2068 struct page *page, *next;
2070 list_for_each_entry_safe(page, next, list, lru) {
2071 trace_mm_page_free_batched(page, cold);
2072 free_hot_cold_page(page, cold);
2077 * split_page takes a non-compound higher-order page, and splits it into
2078 * n (1<<order) sub-pages: page[0..n]
2079 * Each sub-page must be freed individually.
2081 * Note: this is probably too low level an operation for use in drivers.
2082 * Please consult with lkml before using this in your driver.
2084 void split_page(struct page *page, unsigned int order)
2089 VM_BUG_ON_PAGE(PageCompound(page), page);
2090 VM_BUG_ON_PAGE(!page_count(page), page);
2092 #ifdef CONFIG_KMEMCHECK
2094 * Split shadow pages too, because free(page[0]) would
2095 * otherwise free the whole shadow.
2097 if (kmemcheck_page_is_tracked(page))
2098 split_page(virt_to_page(page[0].shadow), order);
2101 gfp_mask = get_page_owner_gfp(page);
2102 set_page_owner(page, 0, gfp_mask);
2103 for (i = 1; i < (1 << order); i++) {
2104 set_page_refcounted(page + i);
2105 set_page_owner(page + i, 0, gfp_mask);
2108 EXPORT_SYMBOL_GPL(split_page);
2110 int __isolate_free_page(struct page *page, unsigned int order)
2112 unsigned long watermark;
2116 BUG_ON(!PageBuddy(page));
2118 zone = page_zone(page);
2119 mt = get_pageblock_migratetype(page);
2121 if (!is_migrate_isolate(mt)) {
2122 /* Obey watermarks as if the page was being allocated */
2123 watermark = low_wmark_pages(zone) + (1 << order);
2124 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2127 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2130 /* Remove page from free list */
2131 list_del(&page->lru);
2132 zone->free_area[order].nr_free--;
2133 rmv_page_order(page);
2135 set_page_owner(page, order, __GFP_MOVABLE);
2137 /* Set the pageblock if the isolated page is at least a pageblock */
2138 if (order >= pageblock_order - 1) {
2139 struct page *endpage = page + (1 << order) - 1;
2140 for (; page < endpage; page += pageblock_nr_pages) {
2141 int mt = get_pageblock_migratetype(page);
2142 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2143 set_pageblock_migratetype(page,
2149 return 1UL << order;
2153 * Similar to split_page except the page is already free. As this is only
2154 * being used for migration, the migratetype of the block also changes.
2155 * As this is called with interrupts disabled, the caller is responsible
2156 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2159 * Note: this is probably too low level an operation for use in drivers.
2160 * Please consult with lkml before using this in your driver.
2162 int split_free_page(struct page *page)
2167 order = page_order(page);
2169 nr_pages = __isolate_free_page(page, order);
2173 /* Split into individual pages */
2174 set_page_refcounted(page);
2175 split_page(page, order);
2180 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2183 struct page *buffered_rmqueue(struct zone *preferred_zone,
2184 struct zone *zone, unsigned int order,
2185 gfp_t gfp_flags, int alloc_flags, int migratetype)
2187 unsigned long flags;
2189 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2191 if (likely(order == 0)) {
2192 struct per_cpu_pages *pcp;
2193 struct list_head *list;
2195 local_irq_save(flags);
2196 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2197 list = &pcp->lists[migratetype];
2198 if (list_empty(list)) {
2199 pcp->count += rmqueue_bulk(zone, 0,
2202 if (unlikely(list_empty(list)))
2207 page = list_last_entry(list, struct page, lru);
2209 page = list_first_entry(list, struct page, lru);
2211 list_del(&page->lru);
2214 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2216 * __GFP_NOFAIL is not to be used in new code.
2218 * All __GFP_NOFAIL callers should be fixed so that they
2219 * properly detect and handle allocation failures.
2221 * We most definitely don't want callers attempting to
2222 * allocate greater than order-1 page units with
2225 WARN_ON_ONCE(order > 1);
2227 spin_lock_irqsave(&zone->lock, flags);
2230 if (alloc_flags & ALLOC_HARDER) {
2231 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2233 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2236 page = __rmqueue(zone, order, migratetype);
2237 spin_unlock(&zone->lock);
2240 __mod_zone_freepage_state(zone, -(1 << order),
2241 get_pcppage_migratetype(page));
2244 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2245 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2246 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2247 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2249 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2250 zone_statistics(preferred_zone, zone, gfp_flags);
2251 local_irq_restore(flags);
2253 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2257 local_irq_restore(flags);
2261 #ifdef CONFIG_FAIL_PAGE_ALLOC
2264 struct fault_attr attr;
2266 bool ignore_gfp_highmem;
2267 bool ignore_gfp_reclaim;
2269 } fail_page_alloc = {
2270 .attr = FAULT_ATTR_INITIALIZER,
2271 .ignore_gfp_reclaim = true,
2272 .ignore_gfp_highmem = true,
2276 static int __init setup_fail_page_alloc(char *str)
2278 return setup_fault_attr(&fail_page_alloc.attr, str);
2280 __setup("fail_page_alloc=", setup_fail_page_alloc);
2282 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2284 if (order < fail_page_alloc.min_order)
2286 if (gfp_mask & __GFP_NOFAIL)
2288 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2290 if (fail_page_alloc.ignore_gfp_reclaim &&
2291 (gfp_mask & __GFP_DIRECT_RECLAIM))
2294 return should_fail(&fail_page_alloc.attr, 1 << order);
2297 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2299 static int __init fail_page_alloc_debugfs(void)
2301 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2304 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2305 &fail_page_alloc.attr);
2307 return PTR_ERR(dir);
2309 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2310 &fail_page_alloc.ignore_gfp_reclaim))
2312 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2313 &fail_page_alloc.ignore_gfp_highmem))
2315 if (!debugfs_create_u32("min-order", mode, dir,
2316 &fail_page_alloc.min_order))
2321 debugfs_remove_recursive(dir);
2326 late_initcall(fail_page_alloc_debugfs);
2328 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2330 #else /* CONFIG_FAIL_PAGE_ALLOC */
2332 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2337 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2340 * Return true if free base pages are above 'mark'. For high-order checks it
2341 * will return true of the order-0 watermark is reached and there is at least
2342 * one free page of a suitable size. Checking now avoids taking the zone lock
2343 * to check in the allocation paths if no pages are free.
2345 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2346 unsigned long mark, int classzone_idx, int alloc_flags,
2351 const int alloc_harder = (alloc_flags & ALLOC_HARDER);
2353 /* free_pages may go negative - that's OK */
2354 free_pages -= (1 << order) - 1;
2356 if (alloc_flags & ALLOC_HIGH)
2360 * If the caller does not have rights to ALLOC_HARDER then subtract
2361 * the high-atomic reserves. This will over-estimate the size of the
2362 * atomic reserve but it avoids a search.
2364 if (likely(!alloc_harder))
2365 free_pages -= z->nr_reserved_highatomic;
2370 /* If allocation can't use CMA areas don't use free CMA pages */
2371 if (!(alloc_flags & ALLOC_CMA))
2372 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2376 * Check watermarks for an order-0 allocation request. If these
2377 * are not met, then a high-order request also cannot go ahead
2378 * even if a suitable page happened to be free.
2380 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2383 /* If this is an order-0 request then the watermark is fine */
2387 /* For a high-order request, check at least one suitable page is free */
2388 for (o = order; o < MAX_ORDER; o++) {
2389 struct free_area *area = &z->free_area[o];
2398 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2399 if (!list_empty(&area->free_list[mt]))
2404 if ((alloc_flags & ALLOC_CMA) &&
2405 !list_empty(&area->free_list[MIGRATE_CMA])) {
2413 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2414 int classzone_idx, int alloc_flags)
2416 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2417 zone_page_state(z, NR_FREE_PAGES));
2420 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2421 unsigned long mark, int classzone_idx)
2423 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2425 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2426 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2428 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2433 static bool zone_local(struct zone *local_zone, struct zone *zone)
2435 return local_zone->node == zone->node;
2438 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2440 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2443 #else /* CONFIG_NUMA */
2444 static bool zone_local(struct zone *local_zone, struct zone *zone)
2449 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2453 #endif /* CONFIG_NUMA */
2455 static void reset_alloc_batches(struct zone *preferred_zone)
2457 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2460 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2461 high_wmark_pages(zone) - low_wmark_pages(zone) -
2462 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2463 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2464 } while (zone++ != preferred_zone);
2468 * get_page_from_freelist goes through the zonelist trying to allocate
2471 static struct page *
2472 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2473 const struct alloc_context *ac)
2475 struct zonelist *zonelist = ac->zonelist;
2477 struct page *page = NULL;
2479 int nr_fair_skipped = 0;
2480 bool zonelist_rescan;
2483 zonelist_rescan = false;
2486 * Scan zonelist, looking for a zone with enough free.
2487 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2489 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2493 if (cpusets_enabled() &&
2494 (alloc_flags & ALLOC_CPUSET) &&
2495 !cpuset_zone_allowed(zone, gfp_mask))
2498 * Distribute pages in proportion to the individual
2499 * zone size to ensure fair page aging. The zone a
2500 * page was allocated in should have no effect on the
2501 * time the page has in memory before being reclaimed.
2503 if (alloc_flags & ALLOC_FAIR) {
2504 if (!zone_local(ac->preferred_zone, zone))
2506 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2512 * When allocating a page cache page for writing, we
2513 * want to get it from a zone that is within its dirty
2514 * limit, such that no single zone holds more than its
2515 * proportional share of globally allowed dirty pages.
2516 * The dirty limits take into account the zone's
2517 * lowmem reserves and high watermark so that kswapd
2518 * should be able to balance it without having to
2519 * write pages from its LRU list.
2521 * This may look like it could increase pressure on
2522 * lower zones by failing allocations in higher zones
2523 * before they are full. But the pages that do spill
2524 * over are limited as the lower zones are protected
2525 * by this very same mechanism. It should not become
2526 * a practical burden to them.
2528 * XXX: For now, allow allocations to potentially
2529 * exceed the per-zone dirty limit in the slowpath
2530 * (spread_dirty_pages unset) before going into reclaim,
2531 * which is important when on a NUMA setup the allowed
2532 * zones are together not big enough to reach the
2533 * global limit. The proper fix for these situations
2534 * will require awareness of zones in the
2535 * dirty-throttling and the flusher threads.
2537 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2540 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2541 if (!zone_watermark_ok(zone, order, mark,
2542 ac->classzone_idx, alloc_flags)) {
2545 /* Checked here to keep the fast path fast */
2546 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2547 if (alloc_flags & ALLOC_NO_WATERMARKS)
2550 if (zone_reclaim_mode == 0 ||
2551 !zone_allows_reclaim(ac->preferred_zone, zone))
2554 ret = zone_reclaim(zone, gfp_mask, order);
2556 case ZONE_RECLAIM_NOSCAN:
2559 case ZONE_RECLAIM_FULL:
2560 /* scanned but unreclaimable */
2563 /* did we reclaim enough */
2564 if (zone_watermark_ok(zone, order, mark,
2565 ac->classzone_idx, alloc_flags))
2573 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2574 gfp_mask, alloc_flags, ac->migratetype);
2576 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2580 * If this is a high-order atomic allocation then check
2581 * if the pageblock should be reserved for the future
2583 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2584 reserve_highatomic_pageblock(page, zone, order);
2591 * The first pass makes sure allocations are spread fairly within the
2592 * local node. However, the local node might have free pages left
2593 * after the fairness batches are exhausted, and remote zones haven't
2594 * even been considered yet. Try once more without fairness, and
2595 * include remote zones now, before entering the slowpath and waking
2596 * kswapd: prefer spilling to a remote zone over swapping locally.
2598 if (alloc_flags & ALLOC_FAIR) {
2599 alloc_flags &= ~ALLOC_FAIR;
2600 if (nr_fair_skipped) {
2601 zonelist_rescan = true;
2602 reset_alloc_batches(ac->preferred_zone);
2604 if (nr_online_nodes > 1)
2605 zonelist_rescan = true;
2608 if (zonelist_rescan)
2615 * Large machines with many possible nodes should not always dump per-node
2616 * meminfo in irq context.
2618 static inline bool should_suppress_show_mem(void)
2623 ret = in_interrupt();
2628 static DEFINE_RATELIMIT_STATE(nopage_rs,
2629 DEFAULT_RATELIMIT_INTERVAL,
2630 DEFAULT_RATELIMIT_BURST);
2632 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
2634 unsigned int filter = SHOW_MEM_FILTER_NODES;
2636 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2637 debug_guardpage_minorder() > 0)
2641 * This documents exceptions given to allocations in certain
2642 * contexts that are allowed to allocate outside current's set
2645 if (!(gfp_mask & __GFP_NOMEMALLOC))
2646 if (test_thread_flag(TIF_MEMDIE) ||
2647 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2648 filter &= ~SHOW_MEM_FILTER_NODES;
2649 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2650 filter &= ~SHOW_MEM_FILTER_NODES;
2653 struct va_format vaf;
2656 va_start(args, fmt);
2661 pr_warn("%pV", &vaf);
2666 pr_warn("%s: page allocation failure: order:%u, mode:0x%x\n",
2667 current->comm, order, gfp_mask);
2670 if (!should_suppress_show_mem())
2674 static inline struct page *
2675 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2676 const struct alloc_context *ac, unsigned long *did_some_progress)
2678 struct oom_control oc = {
2679 .zonelist = ac->zonelist,
2680 .nodemask = ac->nodemask,
2681 .gfp_mask = gfp_mask,
2686 *did_some_progress = 0;
2689 * Acquire the oom lock. If that fails, somebody else is
2690 * making progress for us.
2692 if (!mutex_trylock(&oom_lock)) {
2693 *did_some_progress = 1;
2694 schedule_timeout_uninterruptible(1);
2699 * Go through the zonelist yet one more time, keep very high watermark
2700 * here, this is only to catch a parallel oom killing, we must fail if
2701 * we're still under heavy pressure.
2703 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2704 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2708 if (!(gfp_mask & __GFP_NOFAIL)) {
2709 /* Coredumps can quickly deplete all memory reserves */
2710 if (current->flags & PF_DUMPCORE)
2712 /* The OOM killer will not help higher order allocs */
2713 if (order > PAGE_ALLOC_COSTLY_ORDER)
2715 /* The OOM killer does not needlessly kill tasks for lowmem */
2716 if (ac->high_zoneidx < ZONE_NORMAL)
2718 /* The OOM killer does not compensate for IO-less reclaim */
2719 if (!(gfp_mask & __GFP_FS)) {
2721 * XXX: Page reclaim didn't yield anything,
2722 * and the OOM killer can't be invoked, but
2723 * keep looping as per tradition.
2725 *did_some_progress = 1;
2728 if (pm_suspended_storage())
2730 /* The OOM killer may not free memory on a specific node */
2731 if (gfp_mask & __GFP_THISNODE)
2734 /* Exhausted what can be done so it's blamo time */
2735 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
2736 *did_some_progress = 1;
2738 if (gfp_mask & __GFP_NOFAIL) {
2739 page = get_page_from_freelist(gfp_mask, order,
2740 ALLOC_NO_WATERMARKS|ALLOC_CPUSET, ac);
2742 * fallback to ignore cpuset restriction if our nodes
2746 page = get_page_from_freelist(gfp_mask, order,
2747 ALLOC_NO_WATERMARKS, ac);
2751 mutex_unlock(&oom_lock);
2755 #ifdef CONFIG_COMPACTION
2756 /* Try memory compaction for high-order allocations before reclaim */
2757 static struct page *
2758 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2759 int alloc_flags, const struct alloc_context *ac,
2760 enum migrate_mode mode, int *contended_compaction,
2761 bool *deferred_compaction)
2763 unsigned long compact_result;
2769 current->flags |= PF_MEMALLOC;
2770 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2771 mode, contended_compaction);
2772 current->flags &= ~PF_MEMALLOC;
2774 switch (compact_result) {
2775 case COMPACT_DEFERRED:
2776 *deferred_compaction = true;
2778 case COMPACT_SKIPPED:
2785 * At least in one zone compaction wasn't deferred or skipped, so let's
2786 * count a compaction stall
2788 count_vm_event(COMPACTSTALL);
2790 page = get_page_from_freelist(gfp_mask, order,
2791 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2794 struct zone *zone = page_zone(page);
2796 zone->compact_blockskip_flush = false;
2797 compaction_defer_reset(zone, order, true);
2798 count_vm_event(COMPACTSUCCESS);
2803 * It's bad if compaction run occurs and fails. The most likely reason
2804 * is that pages exist, but not enough to satisfy watermarks.
2806 count_vm_event(COMPACTFAIL);
2813 static inline struct page *
2814 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2815 int alloc_flags, const struct alloc_context *ac,
2816 enum migrate_mode mode, int *contended_compaction,
2817 bool *deferred_compaction)
2821 #endif /* CONFIG_COMPACTION */
2823 /* Perform direct synchronous page reclaim */
2825 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2826 const struct alloc_context *ac)
2828 struct reclaim_state reclaim_state;
2833 /* We now go into synchronous reclaim */
2834 cpuset_memory_pressure_bump();
2835 current->flags |= PF_MEMALLOC;
2836 lockdep_set_current_reclaim_state(gfp_mask);
2837 reclaim_state.reclaimed_slab = 0;
2838 current->reclaim_state = &reclaim_state;
2840 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2843 current->reclaim_state = NULL;
2844 lockdep_clear_current_reclaim_state();
2845 current->flags &= ~PF_MEMALLOC;
2852 /* The really slow allocator path where we enter direct reclaim */
2853 static inline struct page *
2854 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2855 int alloc_flags, const struct alloc_context *ac,
2856 unsigned long *did_some_progress)
2858 struct page *page = NULL;
2859 bool drained = false;
2861 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2862 if (unlikely(!(*did_some_progress)))
2866 page = get_page_from_freelist(gfp_mask, order,
2867 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2870 * If an allocation failed after direct reclaim, it could be because
2871 * pages are pinned on the per-cpu lists or in high alloc reserves.
2872 * Shrink them them and try again
2874 if (!page && !drained) {
2875 unreserve_highatomic_pageblock(ac);
2876 drain_all_pages(NULL);
2884 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2889 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2890 ac->high_zoneidx, ac->nodemask)
2891 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2895 gfp_to_alloc_flags(gfp_t gfp_mask)
2897 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2899 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2900 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2903 * The caller may dip into page reserves a bit more if the caller
2904 * cannot run direct reclaim, or if the caller has realtime scheduling
2905 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2906 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
2908 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2910 if (gfp_mask & __GFP_ATOMIC) {
2912 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2913 * if it can't schedule.
2915 if (!(gfp_mask & __GFP_NOMEMALLOC))
2916 alloc_flags |= ALLOC_HARDER;
2918 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2919 * comment for __cpuset_node_allowed().
2921 alloc_flags &= ~ALLOC_CPUSET;
2922 } else if (unlikely(rt_task(current)) && !in_interrupt())
2923 alloc_flags |= ALLOC_HARDER;
2925 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2926 if (gfp_mask & __GFP_MEMALLOC)
2927 alloc_flags |= ALLOC_NO_WATERMARKS;
2928 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2929 alloc_flags |= ALLOC_NO_WATERMARKS;
2930 else if (!in_interrupt() &&
2931 ((current->flags & PF_MEMALLOC) ||
2932 unlikely(test_thread_flag(TIF_MEMDIE))))
2933 alloc_flags |= ALLOC_NO_WATERMARKS;
2936 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2937 alloc_flags |= ALLOC_CMA;
2942 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2944 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2947 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
2949 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
2952 static inline struct page *
2953 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2954 struct alloc_context *ac)
2956 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
2957 struct page *page = NULL;
2959 unsigned long pages_reclaimed = 0;
2960 unsigned long did_some_progress;
2961 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2962 bool deferred_compaction = false;
2963 int contended_compaction = COMPACT_CONTENDED_NONE;
2966 * In the slowpath, we sanity check order to avoid ever trying to
2967 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2968 * be using allocators in order of preference for an area that is
2971 if (order >= MAX_ORDER) {
2972 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2977 * We also sanity check to catch abuse of atomic reserves being used by
2978 * callers that are not in atomic context.
2980 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
2981 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
2982 gfp_mask &= ~__GFP_ATOMIC;
2985 * If this allocation cannot block and it is for a specific node, then
2986 * fail early. There's no need to wakeup kswapd or retry for a
2987 * speculative node-specific allocation.
2989 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !can_direct_reclaim)
2993 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
2994 wake_all_kswapds(order, ac);
2997 * OK, we're below the kswapd watermark and have kicked background
2998 * reclaim. Now things get more complex, so set up alloc_flags according
2999 * to how we want to proceed.
3001 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3004 * Find the true preferred zone if the allocation is unconstrained by
3007 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3008 struct zoneref *preferred_zoneref;
3009 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3010 ac->high_zoneidx, NULL, &ac->preferred_zone);
3011 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3014 /* This is the last chance, in general, before the goto nopage. */
3015 page = get_page_from_freelist(gfp_mask, order,
3016 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3020 /* Allocate without watermarks if the context allows */
3021 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3023 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3024 * the allocation is high priority and these type of
3025 * allocations are system rather than user orientated
3027 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3028 page = get_page_from_freelist(gfp_mask, order,
3029 ALLOC_NO_WATERMARKS, ac);
3034 /* Caller is not willing to reclaim, we can't balance anything */
3035 if (!can_direct_reclaim) {
3037 * All existing users of the __GFP_NOFAIL are blockable, so warn
3038 * of any new users that actually allow this type of allocation
3041 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3045 /* Avoid recursion of direct reclaim */
3046 if (current->flags & PF_MEMALLOC) {
3048 * __GFP_NOFAIL request from this context is rather bizarre
3049 * because we cannot reclaim anything and only can loop waiting
3050 * for somebody to do a work for us.
3052 if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3059 /* Avoid allocations with no watermarks from looping endlessly */
3060 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3064 * Try direct compaction. The first pass is asynchronous. Subsequent
3065 * attempts after direct reclaim are synchronous
3067 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3069 &contended_compaction,
3070 &deferred_compaction);
3074 /* Checks for THP-specific high-order allocations */
3075 if (is_thp_gfp_mask(gfp_mask)) {
3077 * If compaction is deferred for high-order allocations, it is
3078 * because sync compaction recently failed. If this is the case
3079 * and the caller requested a THP allocation, we do not want
3080 * to heavily disrupt the system, so we fail the allocation
3081 * instead of entering direct reclaim.
3083 if (deferred_compaction)
3087 * In all zones where compaction was attempted (and not
3088 * deferred or skipped), lock contention has been detected.
3089 * For THP allocation we do not want to disrupt the others
3090 * so we fallback to base pages instead.
3092 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3096 * If compaction was aborted due to need_resched(), we do not
3097 * want to further increase allocation latency, unless it is
3098 * khugepaged trying to collapse.
3100 if (contended_compaction == COMPACT_CONTENDED_SCHED
3101 && !(current->flags & PF_KTHREAD))
3106 * It can become very expensive to allocate transparent hugepages at
3107 * fault, so use asynchronous memory compaction for THP unless it is
3108 * khugepaged trying to collapse.
3110 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3111 migration_mode = MIGRATE_SYNC_LIGHT;
3113 /* Try direct reclaim and then allocating */
3114 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3115 &did_some_progress);
3119 /* Do not loop if specifically requested */
3120 if (gfp_mask & __GFP_NORETRY)
3123 /* Keep reclaiming pages as long as there is reasonable progress */
3124 pages_reclaimed += did_some_progress;
3125 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3126 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3127 /* Wait for some write requests to complete then retry */
3128 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3132 /* Reclaim has failed us, start killing things */
3133 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3137 /* Retry as long as the OOM killer is making progress */
3138 if (did_some_progress)
3143 * High-order allocations do not necessarily loop after
3144 * direct reclaim and reclaim/compaction depends on compaction
3145 * being called after reclaim so call directly if necessary
3147 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3149 &contended_compaction,
3150 &deferred_compaction);
3154 warn_alloc_failed(gfp_mask, order, NULL);
3160 * This is the 'heart' of the zoned buddy allocator.
3163 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3164 struct zonelist *zonelist, nodemask_t *nodemask)
3166 struct zoneref *preferred_zoneref;
3167 struct page *page = NULL;
3168 unsigned int cpuset_mems_cookie;
3169 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3170 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3171 struct alloc_context ac = {
3172 .high_zoneidx = gfp_zone(gfp_mask),
3173 .nodemask = nodemask,
3174 .migratetype = gfpflags_to_migratetype(gfp_mask),
3177 gfp_mask &= gfp_allowed_mask;
3179 lockdep_trace_alloc(gfp_mask);
3181 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3183 if (should_fail_alloc_page(gfp_mask, order))
3187 * Check the zones suitable for the gfp_mask contain at least one
3188 * valid zone. It's possible to have an empty zonelist as a result
3189 * of __GFP_THISNODE and a memoryless node
3191 if (unlikely(!zonelist->_zonerefs->zone))
3194 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3195 alloc_flags |= ALLOC_CMA;
3198 cpuset_mems_cookie = read_mems_allowed_begin();
3200 /* We set it here, as __alloc_pages_slowpath might have changed it */
3201 ac.zonelist = zonelist;
3203 /* Dirty zone balancing only done in the fast path */
3204 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3206 /* The preferred zone is used for statistics later */
3207 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3208 ac.nodemask ? : &cpuset_current_mems_allowed,
3209 &ac.preferred_zone);
3210 if (!ac.preferred_zone)
3212 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3214 /* First allocation attempt */
3215 alloc_mask = gfp_mask|__GFP_HARDWALL;
3216 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3217 if (unlikely(!page)) {
3219 * Runtime PM, block IO and its error handling path
3220 * can deadlock because I/O on the device might not
3223 alloc_mask = memalloc_noio_flags(gfp_mask);
3224 ac.spread_dirty_pages = false;
3226 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3229 if (kmemcheck_enabled && page)
3230 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3232 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3236 * When updating a task's mems_allowed, it is possible to race with
3237 * parallel threads in such a way that an allocation can fail while
3238 * the mask is being updated. If a page allocation is about to fail,
3239 * check if the cpuset changed during allocation and if so, retry.
3241 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3246 EXPORT_SYMBOL(__alloc_pages_nodemask);
3249 * Common helper functions.
3251 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3256 * __get_free_pages() returns a 32-bit address, which cannot represent
3259 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3261 page = alloc_pages(gfp_mask, order);
3264 return (unsigned long) page_address(page);
3266 EXPORT_SYMBOL(__get_free_pages);
3268 unsigned long get_zeroed_page(gfp_t gfp_mask)
3270 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3272 EXPORT_SYMBOL(get_zeroed_page);
3274 void __free_pages(struct page *page, unsigned int order)
3276 if (put_page_testzero(page)) {
3278 free_hot_cold_page(page, false);
3280 __free_pages_ok(page, order);
3284 EXPORT_SYMBOL(__free_pages);
3286 void free_pages(unsigned long addr, unsigned int order)
3289 VM_BUG_ON(!virt_addr_valid((void *)addr));
3290 __free_pages(virt_to_page((void *)addr), order);
3294 EXPORT_SYMBOL(free_pages);
3298 * An arbitrary-length arbitrary-offset area of memory which resides
3299 * within a 0 or higher order page. Multiple fragments within that page
3300 * are individually refcounted, in the page's reference counter.
3302 * The page_frag functions below provide a simple allocation framework for
3303 * page fragments. This is used by the network stack and network device
3304 * drivers to provide a backing region of memory for use as either an
3305 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3307 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3310 struct page *page = NULL;
3311 gfp_t gfp = gfp_mask;
3313 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3314 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3316 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3317 PAGE_FRAG_CACHE_MAX_ORDER);
3318 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3320 if (unlikely(!page))
3321 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3323 nc->va = page ? page_address(page) : NULL;
3328 void *__alloc_page_frag(struct page_frag_cache *nc,
3329 unsigned int fragsz, gfp_t gfp_mask)
3331 unsigned int size = PAGE_SIZE;
3335 if (unlikely(!nc->va)) {
3337 page = __page_frag_refill(nc, gfp_mask);
3341 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3342 /* if size can vary use size else just use PAGE_SIZE */
3345 /* Even if we own the page, we do not use atomic_set().
3346 * This would break get_page_unless_zero() users.
3348 atomic_add(size - 1, &page->_count);
3350 /* reset page count bias and offset to start of new frag */
3351 nc->pfmemalloc = page_is_pfmemalloc(page);
3352 nc->pagecnt_bias = size;
3356 offset = nc->offset - fragsz;
3357 if (unlikely(offset < 0)) {
3358 page = virt_to_page(nc->va);
3360 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3363 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3364 /* if size can vary use size else just use PAGE_SIZE */
3367 /* OK, page count is 0, we can safely set it */
3368 atomic_set(&page->_count, size);
3370 /* reset page count bias and offset to start of new frag */
3371 nc->pagecnt_bias = size;
3372 offset = size - fragsz;
3376 nc->offset = offset;
3378 return nc->va + offset;
3380 EXPORT_SYMBOL(__alloc_page_frag);
3383 * Frees a page fragment allocated out of either a compound or order 0 page.
3385 void __free_page_frag(void *addr)
3387 struct page *page = virt_to_head_page(addr);
3389 if (unlikely(put_page_testzero(page)))
3390 __free_pages_ok(page, compound_order(page));
3392 EXPORT_SYMBOL(__free_page_frag);
3395 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3396 * of the current memory cgroup if __GFP_ACCOUNT is set, other than that it is
3397 * equivalent to alloc_pages.
3399 * It should be used when the caller would like to use kmalloc, but since the
3400 * allocation is large, it has to fall back to the page allocator.
3402 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3406 page = alloc_pages(gfp_mask, order);
3407 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3408 __free_pages(page, order);
3414 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3418 page = alloc_pages_node(nid, gfp_mask, order);
3419 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3420 __free_pages(page, order);
3427 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3430 void __free_kmem_pages(struct page *page, unsigned int order)
3432 memcg_kmem_uncharge(page, order);
3433 __free_pages(page, order);
3436 void free_kmem_pages(unsigned long addr, unsigned int order)
3439 VM_BUG_ON(!virt_addr_valid((void *)addr));
3440 __free_kmem_pages(virt_to_page((void *)addr), order);
3444 static void *make_alloc_exact(unsigned long addr, unsigned int order,
3448 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3449 unsigned long used = addr + PAGE_ALIGN(size);
3451 split_page(virt_to_page((void *)addr), order);
3452 while (used < alloc_end) {
3457 return (void *)addr;
3461 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3462 * @size: the number of bytes to allocate
3463 * @gfp_mask: GFP flags for the allocation
3465 * This function is similar to alloc_pages(), except that it allocates the
3466 * minimum number of pages to satisfy the request. alloc_pages() can only
3467 * allocate memory in power-of-two pages.
3469 * This function is also limited by MAX_ORDER.
3471 * Memory allocated by this function must be released by free_pages_exact().
3473 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3475 unsigned int order = get_order(size);
3478 addr = __get_free_pages(gfp_mask, order);
3479 return make_alloc_exact(addr, order, size);
3481 EXPORT_SYMBOL(alloc_pages_exact);
3484 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3486 * @nid: the preferred node ID where memory should be allocated
3487 * @size: the number of bytes to allocate
3488 * @gfp_mask: GFP flags for the allocation
3490 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3493 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3495 unsigned int order = get_order(size);
3496 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3499 return make_alloc_exact((unsigned long)page_address(p), order, size);
3503 * free_pages_exact - release memory allocated via alloc_pages_exact()
3504 * @virt: the value returned by alloc_pages_exact.
3505 * @size: size of allocation, same value as passed to alloc_pages_exact().
3507 * Release the memory allocated by a previous call to alloc_pages_exact.
3509 void free_pages_exact(void *virt, size_t size)
3511 unsigned long addr = (unsigned long)virt;
3512 unsigned long end = addr + PAGE_ALIGN(size);
3514 while (addr < end) {
3519 EXPORT_SYMBOL(free_pages_exact);
3522 * nr_free_zone_pages - count number of pages beyond high watermark
3523 * @offset: The zone index of the highest zone
3525 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3526 * high watermark within all zones at or below a given zone index. For each
3527 * zone, the number of pages is calculated as:
3528 * managed_pages - high_pages
3530 static unsigned long nr_free_zone_pages(int offset)
3535 /* Just pick one node, since fallback list is circular */
3536 unsigned long sum = 0;
3538 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3540 for_each_zone_zonelist(zone, z, zonelist, offset) {
3541 unsigned long size = zone->managed_pages;
3542 unsigned long high = high_wmark_pages(zone);
3551 * nr_free_buffer_pages - count number of pages beyond high watermark
3553 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3554 * watermark within ZONE_DMA and ZONE_NORMAL.
3556 unsigned long nr_free_buffer_pages(void)
3558 return nr_free_zone_pages(gfp_zone(GFP_USER));
3560 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3563 * nr_free_pagecache_pages - count number of pages beyond high watermark
3565 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3566 * high watermark within all zones.
3568 unsigned long nr_free_pagecache_pages(void)
3570 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3573 static inline void show_node(struct zone *zone)
3575 if (IS_ENABLED(CONFIG_NUMA))
3576 printk("Node %d ", zone_to_nid(zone));
3579 void si_meminfo(struct sysinfo *val)
3581 val->totalram = totalram_pages;
3582 val->sharedram = global_page_state(NR_SHMEM);
3583 val->freeram = global_page_state(NR_FREE_PAGES);
3584 val->bufferram = nr_blockdev_pages();
3585 val->totalhigh = totalhigh_pages;
3586 val->freehigh = nr_free_highpages();
3587 val->mem_unit = PAGE_SIZE;
3590 EXPORT_SYMBOL(si_meminfo);
3593 void si_meminfo_node(struct sysinfo *val, int nid)
3595 int zone_type; /* needs to be signed */
3596 unsigned long managed_pages = 0;
3597 pg_data_t *pgdat = NODE_DATA(nid);
3599 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3600 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3601 val->totalram = managed_pages;
3602 val->sharedram = node_page_state(nid, NR_SHMEM);
3603 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3604 #ifdef CONFIG_HIGHMEM
3605 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3606 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3612 val->mem_unit = PAGE_SIZE;
3617 * Determine whether the node should be displayed or not, depending on whether
3618 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3620 bool skip_free_areas_node(unsigned int flags, int nid)
3623 unsigned int cpuset_mems_cookie;
3625 if (!(flags & SHOW_MEM_FILTER_NODES))
3629 cpuset_mems_cookie = read_mems_allowed_begin();
3630 ret = !node_isset(nid, cpuset_current_mems_allowed);
3631 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3636 #define K(x) ((x) << (PAGE_SHIFT-10))
3638 static void show_migration_types(unsigned char type)
3640 static const char types[MIGRATE_TYPES] = {
3641 [MIGRATE_UNMOVABLE] = 'U',
3642 [MIGRATE_MOVABLE] = 'M',
3643 [MIGRATE_RECLAIMABLE] = 'E',
3644 [MIGRATE_HIGHATOMIC] = 'H',
3646 [MIGRATE_CMA] = 'C',
3648 #ifdef CONFIG_MEMORY_ISOLATION
3649 [MIGRATE_ISOLATE] = 'I',
3652 char tmp[MIGRATE_TYPES + 1];
3656 for (i = 0; i < MIGRATE_TYPES; i++) {
3657 if (type & (1 << i))
3662 printk("(%s) ", tmp);
3666 * Show free area list (used inside shift_scroll-lock stuff)
3667 * We also calculate the percentage fragmentation. We do this by counting the
3668 * memory on each free list with the exception of the first item on the list.
3671 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3674 void show_free_areas(unsigned int filter)
3676 unsigned long free_pcp = 0;
3680 for_each_populated_zone(zone) {
3681 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3684 for_each_online_cpu(cpu)
3685 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3688 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3689 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3690 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3691 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3692 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3693 " free:%lu free_pcp:%lu free_cma:%lu\n",
3694 global_page_state(NR_ACTIVE_ANON),
3695 global_page_state(NR_INACTIVE_ANON),
3696 global_page_state(NR_ISOLATED_ANON),
3697 global_page_state(NR_ACTIVE_FILE),
3698 global_page_state(NR_INACTIVE_FILE),
3699 global_page_state(NR_ISOLATED_FILE),
3700 global_page_state(NR_UNEVICTABLE),
3701 global_page_state(NR_FILE_DIRTY),
3702 global_page_state(NR_WRITEBACK),
3703 global_page_state(NR_UNSTABLE_NFS),
3704 global_page_state(NR_SLAB_RECLAIMABLE),
3705 global_page_state(NR_SLAB_UNRECLAIMABLE),
3706 global_page_state(NR_FILE_MAPPED),
3707 global_page_state(NR_SHMEM),
3708 global_page_state(NR_PAGETABLE),
3709 global_page_state(NR_BOUNCE),
3710 global_page_state(NR_FREE_PAGES),
3712 global_page_state(NR_FREE_CMA_PAGES));
3714 for_each_populated_zone(zone) {
3717 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3721 for_each_online_cpu(cpu)
3722 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3730 " active_anon:%lukB"
3731 " inactive_anon:%lukB"
3732 " active_file:%lukB"
3733 " inactive_file:%lukB"
3734 " unevictable:%lukB"
3735 " isolated(anon):%lukB"
3736 " isolated(file):%lukB"
3744 " slab_reclaimable:%lukB"
3745 " slab_unreclaimable:%lukB"
3746 " kernel_stack:%lukB"
3753 " writeback_tmp:%lukB"
3754 " pages_scanned:%lu"
3755 " all_unreclaimable? %s"
3758 K(zone_page_state(zone, NR_FREE_PAGES)),
3759 K(min_wmark_pages(zone)),
3760 K(low_wmark_pages(zone)),
3761 K(high_wmark_pages(zone)),
3762 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3763 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3764 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3765 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3766 K(zone_page_state(zone, NR_UNEVICTABLE)),
3767 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3768 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3769 K(zone->present_pages),
3770 K(zone->managed_pages),
3771 K(zone_page_state(zone, NR_MLOCK)),
3772 K(zone_page_state(zone, NR_FILE_DIRTY)),
3773 K(zone_page_state(zone, NR_WRITEBACK)),
3774 K(zone_page_state(zone, NR_FILE_MAPPED)),
3775 K(zone_page_state(zone, NR_SHMEM)),
3776 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3777 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3778 zone_page_state(zone, NR_KERNEL_STACK) *
3780 K(zone_page_state(zone, NR_PAGETABLE)),
3781 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3782 K(zone_page_state(zone, NR_BOUNCE)),
3784 K(this_cpu_read(zone->pageset->pcp.count)),
3785 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3786 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3787 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3788 (!zone_reclaimable(zone) ? "yes" : "no")
3790 printk("lowmem_reserve[]:");
3791 for (i = 0; i < MAX_NR_ZONES; i++)
3792 printk(" %ld", zone->lowmem_reserve[i]);
3796 for_each_populated_zone(zone) {
3798 unsigned long nr[MAX_ORDER], flags, total = 0;
3799 unsigned char types[MAX_ORDER];
3801 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3804 printk("%s: ", zone->name);
3806 spin_lock_irqsave(&zone->lock, flags);
3807 for (order = 0; order < MAX_ORDER; order++) {
3808 struct free_area *area = &zone->free_area[order];
3811 nr[order] = area->nr_free;
3812 total += nr[order] << order;
3815 for (type = 0; type < MIGRATE_TYPES; type++) {
3816 if (!list_empty(&area->free_list[type]))
3817 types[order] |= 1 << type;
3820 spin_unlock_irqrestore(&zone->lock, flags);
3821 for (order = 0; order < MAX_ORDER; order++) {
3822 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3824 show_migration_types(types[order]);
3826 printk("= %lukB\n", K(total));
3829 hugetlb_show_meminfo();
3831 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3833 show_swap_cache_info();
3836 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3838 zoneref->zone = zone;
3839 zoneref->zone_idx = zone_idx(zone);
3843 * Builds allocation fallback zone lists.
3845 * Add all populated zones of a node to the zonelist.
3847 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3851 enum zone_type zone_type = MAX_NR_ZONES;
3855 zone = pgdat->node_zones + zone_type;
3856 if (populated_zone(zone)) {
3857 zoneref_set_zone(zone,
3858 &zonelist->_zonerefs[nr_zones++]);
3859 check_highest_zone(zone_type);
3861 } while (zone_type);
3869 * 0 = automatic detection of better ordering.
3870 * 1 = order by ([node] distance, -zonetype)
3871 * 2 = order by (-zonetype, [node] distance)
3873 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3874 * the same zonelist. So only NUMA can configure this param.
3876 #define ZONELIST_ORDER_DEFAULT 0
3877 #define ZONELIST_ORDER_NODE 1
3878 #define ZONELIST_ORDER_ZONE 2
3880 /* zonelist order in the kernel.
3881 * set_zonelist_order() will set this to NODE or ZONE.
3883 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3884 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3888 /* The value user specified ....changed by config */
3889 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3890 /* string for sysctl */
3891 #define NUMA_ZONELIST_ORDER_LEN 16
3892 char numa_zonelist_order[16] = "default";
3895 * interface for configure zonelist ordering.
3896 * command line option "numa_zonelist_order"
3897 * = "[dD]efault - default, automatic configuration.
3898 * = "[nN]ode - order by node locality, then by zone within node
3899 * = "[zZ]one - order by zone, then by locality within zone
3902 static int __parse_numa_zonelist_order(char *s)
3904 if (*s == 'd' || *s == 'D') {
3905 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3906 } else if (*s == 'n' || *s == 'N') {
3907 user_zonelist_order = ZONELIST_ORDER_NODE;
3908 } else if (*s == 'z' || *s == 'Z') {
3909 user_zonelist_order = ZONELIST_ORDER_ZONE;
3912 "Ignoring invalid numa_zonelist_order value: "
3919 static __init int setup_numa_zonelist_order(char *s)
3926 ret = __parse_numa_zonelist_order(s);
3928 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3932 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3935 * sysctl handler for numa_zonelist_order
3937 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3938 void __user *buffer, size_t *length,
3941 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3943 static DEFINE_MUTEX(zl_order_mutex);
3945 mutex_lock(&zl_order_mutex);
3947 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3951 strcpy(saved_string, (char *)table->data);
3953 ret = proc_dostring(table, write, buffer, length, ppos);
3957 int oldval = user_zonelist_order;
3959 ret = __parse_numa_zonelist_order((char *)table->data);
3962 * bogus value. restore saved string
3964 strncpy((char *)table->data, saved_string,
3965 NUMA_ZONELIST_ORDER_LEN);
3966 user_zonelist_order = oldval;
3967 } else if (oldval != user_zonelist_order) {
3968 mutex_lock(&zonelists_mutex);
3969 build_all_zonelists(NULL, NULL);
3970 mutex_unlock(&zonelists_mutex);
3974 mutex_unlock(&zl_order_mutex);
3979 #define MAX_NODE_LOAD (nr_online_nodes)
3980 static int node_load[MAX_NUMNODES];
3983 * find_next_best_node - find the next node that should appear in a given node's fallback list
3984 * @node: node whose fallback list we're appending
3985 * @used_node_mask: nodemask_t of already used nodes
3987 * We use a number of factors to determine which is the next node that should
3988 * appear on a given node's fallback list. The node should not have appeared
3989 * already in @node's fallback list, and it should be the next closest node
3990 * according to the distance array (which contains arbitrary distance values
3991 * from each node to each node in the system), and should also prefer nodes
3992 * with no CPUs, since presumably they'll have very little allocation pressure
3993 * on them otherwise.
3994 * It returns -1 if no node is found.
3996 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3999 int min_val = INT_MAX;
4000 int best_node = NUMA_NO_NODE;
4001 const struct cpumask *tmp = cpumask_of_node(0);
4003 /* Use the local node if we haven't already */
4004 if (!node_isset(node, *used_node_mask)) {
4005 node_set(node, *used_node_mask);
4009 for_each_node_state(n, N_MEMORY) {
4011 /* Don't want a node to appear more than once */
4012 if (node_isset(n, *used_node_mask))
4015 /* Use the distance array to find the distance */
4016 val = node_distance(node, n);
4018 /* Penalize nodes under us ("prefer the next node") */
4021 /* Give preference to headless and unused nodes */
4022 tmp = cpumask_of_node(n);
4023 if (!cpumask_empty(tmp))
4024 val += PENALTY_FOR_NODE_WITH_CPUS;
4026 /* Slight preference for less loaded node */
4027 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4028 val += node_load[n];
4030 if (val < min_val) {
4037 node_set(best_node, *used_node_mask);
4044 * Build zonelists ordered by node and zones within node.
4045 * This results in maximum locality--normal zone overflows into local
4046 * DMA zone, if any--but risks exhausting DMA zone.
4048 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4051 struct zonelist *zonelist;
4053 zonelist = &pgdat->node_zonelists[0];
4054 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4056 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4057 zonelist->_zonerefs[j].zone = NULL;
4058 zonelist->_zonerefs[j].zone_idx = 0;
4062 * Build gfp_thisnode zonelists
4064 static void build_thisnode_zonelists(pg_data_t *pgdat)
4067 struct zonelist *zonelist;
4069 zonelist = &pgdat->node_zonelists[1];
4070 j = build_zonelists_node(pgdat, zonelist, 0);
4071 zonelist->_zonerefs[j].zone = NULL;
4072 zonelist->_zonerefs[j].zone_idx = 0;
4076 * Build zonelists ordered by zone and nodes within zones.
4077 * This results in conserving DMA zone[s] until all Normal memory is
4078 * exhausted, but results in overflowing to remote node while memory
4079 * may still exist in local DMA zone.
4081 static int node_order[MAX_NUMNODES];
4083 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4086 int zone_type; /* needs to be signed */
4088 struct zonelist *zonelist;
4090 zonelist = &pgdat->node_zonelists[0];
4092 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4093 for (j = 0; j < nr_nodes; j++) {
4094 node = node_order[j];
4095 z = &NODE_DATA(node)->node_zones[zone_type];
4096 if (populated_zone(z)) {
4098 &zonelist->_zonerefs[pos++]);
4099 check_highest_zone(zone_type);
4103 zonelist->_zonerefs[pos].zone = NULL;
4104 zonelist->_zonerefs[pos].zone_idx = 0;
4107 #if defined(CONFIG_64BIT)
4109 * Devices that require DMA32/DMA are relatively rare and do not justify a
4110 * penalty to every machine in case the specialised case applies. Default
4111 * to Node-ordering on 64-bit NUMA machines
4113 static int default_zonelist_order(void)
4115 return ZONELIST_ORDER_NODE;
4119 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4120 * by the kernel. If processes running on node 0 deplete the low memory zone
4121 * then reclaim will occur more frequency increasing stalls and potentially
4122 * be easier to OOM if a large percentage of the zone is under writeback or
4123 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4124 * Hence, default to zone ordering on 32-bit.
4126 static int default_zonelist_order(void)
4128 return ZONELIST_ORDER_ZONE;
4130 #endif /* CONFIG_64BIT */
4132 static void set_zonelist_order(void)
4134 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4135 current_zonelist_order = default_zonelist_order();
4137 current_zonelist_order = user_zonelist_order;
4140 static void build_zonelists(pg_data_t *pgdat)
4143 nodemask_t used_mask;
4144 int local_node, prev_node;
4145 struct zonelist *zonelist;
4146 unsigned int order = current_zonelist_order;
4148 /* initialize zonelists */
4149 for (i = 0; i < MAX_ZONELISTS; i++) {
4150 zonelist = pgdat->node_zonelists + i;
4151 zonelist->_zonerefs[0].zone = NULL;
4152 zonelist->_zonerefs[0].zone_idx = 0;
4155 /* NUMA-aware ordering of nodes */
4156 local_node = pgdat->node_id;
4157 load = nr_online_nodes;
4158 prev_node = local_node;
4159 nodes_clear(used_mask);
4161 memset(node_order, 0, sizeof(node_order));
4164 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4166 * We don't want to pressure a particular node.
4167 * So adding penalty to the first node in same
4168 * distance group to make it round-robin.
4170 if (node_distance(local_node, node) !=
4171 node_distance(local_node, prev_node))
4172 node_load[node] = load;
4176 if (order == ZONELIST_ORDER_NODE)
4177 build_zonelists_in_node_order(pgdat, node);
4179 node_order[i++] = node; /* remember order */
4182 if (order == ZONELIST_ORDER_ZONE) {
4183 /* calculate node order -- i.e., DMA last! */
4184 build_zonelists_in_zone_order(pgdat, i);
4187 build_thisnode_zonelists(pgdat);
4190 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4192 * Return node id of node used for "local" allocations.
4193 * I.e., first node id of first zone in arg node's generic zonelist.
4194 * Used for initializing percpu 'numa_mem', which is used primarily
4195 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4197 int local_memory_node(int node)
4201 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4202 gfp_zone(GFP_KERNEL),
4209 #else /* CONFIG_NUMA */
4211 static void set_zonelist_order(void)
4213 current_zonelist_order = ZONELIST_ORDER_ZONE;
4216 static void build_zonelists(pg_data_t *pgdat)
4218 int node, local_node;
4220 struct zonelist *zonelist;
4222 local_node = pgdat->node_id;
4224 zonelist = &pgdat->node_zonelists[0];
4225 j = build_zonelists_node(pgdat, zonelist, 0);
4228 * Now we build the zonelist so that it contains the zones
4229 * of all the other nodes.
4230 * We don't want to pressure a particular node, so when
4231 * building the zones for node N, we make sure that the
4232 * zones coming right after the local ones are those from
4233 * node N+1 (modulo N)
4235 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4236 if (!node_online(node))
4238 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4240 for (node = 0; node < local_node; node++) {
4241 if (!node_online(node))
4243 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4246 zonelist->_zonerefs[j].zone = NULL;
4247 zonelist->_zonerefs[j].zone_idx = 0;
4250 #endif /* CONFIG_NUMA */
4253 * Boot pageset table. One per cpu which is going to be used for all
4254 * zones and all nodes. The parameters will be set in such a way
4255 * that an item put on a list will immediately be handed over to
4256 * the buddy list. This is safe since pageset manipulation is done
4257 * with interrupts disabled.
4259 * The boot_pagesets must be kept even after bootup is complete for
4260 * unused processors and/or zones. They do play a role for bootstrapping
4261 * hotplugged processors.
4263 * zoneinfo_show() and maybe other functions do
4264 * not check if the processor is online before following the pageset pointer.
4265 * Other parts of the kernel may not check if the zone is available.
4267 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4268 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4269 static void setup_zone_pageset(struct zone *zone);
4272 * Global mutex to protect against size modification of zonelists
4273 * as well as to serialize pageset setup for the new populated zone.
4275 DEFINE_MUTEX(zonelists_mutex);
4277 /* return values int ....just for stop_machine() */
4278 static int __build_all_zonelists(void *data)
4282 pg_data_t *self = data;
4285 memset(node_load, 0, sizeof(node_load));
4288 if (self && !node_online(self->node_id)) {
4289 build_zonelists(self);
4292 for_each_online_node(nid) {
4293 pg_data_t *pgdat = NODE_DATA(nid);
4295 build_zonelists(pgdat);
4299 * Initialize the boot_pagesets that are going to be used
4300 * for bootstrapping processors. The real pagesets for
4301 * each zone will be allocated later when the per cpu
4302 * allocator is available.
4304 * boot_pagesets are used also for bootstrapping offline
4305 * cpus if the system is already booted because the pagesets
4306 * are needed to initialize allocators on a specific cpu too.
4307 * F.e. the percpu allocator needs the page allocator which
4308 * needs the percpu allocator in order to allocate its pagesets
4309 * (a chicken-egg dilemma).
4311 for_each_possible_cpu(cpu) {
4312 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4314 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4316 * We now know the "local memory node" for each node--
4317 * i.e., the node of the first zone in the generic zonelist.
4318 * Set up numa_mem percpu variable for on-line cpus. During
4319 * boot, only the boot cpu should be on-line; we'll init the
4320 * secondary cpus' numa_mem as they come on-line. During
4321 * node/memory hotplug, we'll fixup all on-line cpus.
4323 if (cpu_online(cpu))
4324 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4331 static noinline void __init
4332 build_all_zonelists_init(void)
4334 __build_all_zonelists(NULL);
4335 mminit_verify_zonelist();
4336 cpuset_init_current_mems_allowed();
4340 * Called with zonelists_mutex held always
4341 * unless system_state == SYSTEM_BOOTING.
4343 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4344 * [we're only called with non-NULL zone through __meminit paths] and
4345 * (2) call of __init annotated helper build_all_zonelists_init
4346 * [protected by SYSTEM_BOOTING].
4348 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4350 set_zonelist_order();
4352 if (system_state == SYSTEM_BOOTING) {
4353 build_all_zonelists_init();
4355 #ifdef CONFIG_MEMORY_HOTPLUG
4357 setup_zone_pageset(zone);
4359 /* we have to stop all cpus to guarantee there is no user
4361 stop_machine(__build_all_zonelists, pgdat, NULL);
4362 /* cpuset refresh routine should be here */
4364 vm_total_pages = nr_free_pagecache_pages();
4366 * Disable grouping by mobility if the number of pages in the
4367 * system is too low to allow the mechanism to work. It would be
4368 * more accurate, but expensive to check per-zone. This check is
4369 * made on memory-hotadd so a system can start with mobility
4370 * disabled and enable it later
4372 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4373 page_group_by_mobility_disabled = 1;
4375 page_group_by_mobility_disabled = 0;
4377 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4378 "Total pages: %ld\n",
4380 zonelist_order_name[current_zonelist_order],
4381 page_group_by_mobility_disabled ? "off" : "on",
4384 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4389 * Helper functions to size the waitqueue hash table.
4390 * Essentially these want to choose hash table sizes sufficiently
4391 * large so that collisions trying to wait on pages are rare.
4392 * But in fact, the number of active page waitqueues on typical
4393 * systems is ridiculously low, less than 200. So this is even
4394 * conservative, even though it seems large.
4396 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4397 * waitqueues, i.e. the size of the waitq table given the number of pages.
4399 #define PAGES_PER_WAITQUEUE 256
4401 #ifndef CONFIG_MEMORY_HOTPLUG
4402 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4404 unsigned long size = 1;
4406 pages /= PAGES_PER_WAITQUEUE;
4408 while (size < pages)
4412 * Once we have dozens or even hundreds of threads sleeping
4413 * on IO we've got bigger problems than wait queue collision.
4414 * Limit the size of the wait table to a reasonable size.
4416 size = min(size, 4096UL);
4418 return max(size, 4UL);
4422 * A zone's size might be changed by hot-add, so it is not possible to determine
4423 * a suitable size for its wait_table. So we use the maximum size now.
4425 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4427 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4428 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4429 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4431 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4432 * or more by the traditional way. (See above). It equals:
4434 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4435 * ia64(16K page size) : = ( 8G + 4M)byte.
4436 * powerpc (64K page size) : = (32G +16M)byte.
4438 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4445 * This is an integer logarithm so that shifts can be used later
4446 * to extract the more random high bits from the multiplicative
4447 * hash function before the remainder is taken.
4449 static inline unsigned long wait_table_bits(unsigned long size)
4455 * Initially all pages are reserved - free ones are freed
4456 * up by free_all_bootmem() once the early boot process is
4457 * done. Non-atomic initialization, single-pass.
4459 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4460 unsigned long start_pfn, enum memmap_context context)
4462 pg_data_t *pgdat = NODE_DATA(nid);
4463 unsigned long end_pfn = start_pfn + size;
4466 unsigned long nr_initialised = 0;
4468 if (highest_memmap_pfn < end_pfn - 1)
4469 highest_memmap_pfn = end_pfn - 1;
4471 z = &pgdat->node_zones[zone];
4472 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4474 * There can be holes in boot-time mem_map[]s
4475 * handed to this function. They do not
4476 * exist on hotplugged memory.
4478 if (context == MEMMAP_EARLY) {
4479 if (!early_pfn_valid(pfn))
4481 if (!early_pfn_in_nid(pfn, nid))
4483 if (!update_defer_init(pgdat, pfn, end_pfn,
4489 * Mark the block movable so that blocks are reserved for
4490 * movable at startup. This will force kernel allocations
4491 * to reserve their blocks rather than leaking throughout
4492 * the address space during boot when many long-lived
4493 * kernel allocations are made.
4495 * bitmap is created for zone's valid pfn range. but memmap
4496 * can be created for invalid pages (for alignment)
4497 * check here not to call set_pageblock_migratetype() against
4500 if (!(pfn & (pageblock_nr_pages - 1))) {
4501 struct page *page = pfn_to_page(pfn);
4503 __init_single_page(page, pfn, zone, nid);
4504 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4506 __init_single_pfn(pfn, zone, nid);
4511 static void __meminit zone_init_free_lists(struct zone *zone)
4513 unsigned int order, t;
4514 for_each_migratetype_order(order, t) {
4515 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4516 zone->free_area[order].nr_free = 0;
4520 #ifndef __HAVE_ARCH_MEMMAP_INIT
4521 #define memmap_init(size, nid, zone, start_pfn) \
4522 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4525 static int zone_batchsize(struct zone *zone)
4531 * The per-cpu-pages pools are set to around 1000th of the
4532 * size of the zone. But no more than 1/2 of a meg.
4534 * OK, so we don't know how big the cache is. So guess.
4536 batch = zone->managed_pages / 1024;
4537 if (batch * PAGE_SIZE > 512 * 1024)
4538 batch = (512 * 1024) / PAGE_SIZE;
4539 batch /= 4; /* We effectively *= 4 below */
4544 * Clamp the batch to a 2^n - 1 value. Having a power
4545 * of 2 value was found to be more likely to have
4546 * suboptimal cache aliasing properties in some cases.
4548 * For example if 2 tasks are alternately allocating
4549 * batches of pages, one task can end up with a lot
4550 * of pages of one half of the possible page colors
4551 * and the other with pages of the other colors.
4553 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4558 /* The deferral and batching of frees should be suppressed under NOMMU
4561 * The problem is that NOMMU needs to be able to allocate large chunks
4562 * of contiguous memory as there's no hardware page translation to
4563 * assemble apparent contiguous memory from discontiguous pages.
4565 * Queueing large contiguous runs of pages for batching, however,
4566 * causes the pages to actually be freed in smaller chunks. As there
4567 * can be a significant delay between the individual batches being
4568 * recycled, this leads to the once large chunks of space being
4569 * fragmented and becoming unavailable for high-order allocations.
4576 * pcp->high and pcp->batch values are related and dependent on one another:
4577 * ->batch must never be higher then ->high.
4578 * The following function updates them in a safe manner without read side
4581 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4582 * those fields changing asynchronously (acording the the above rule).
4584 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4585 * outside of boot time (or some other assurance that no concurrent updaters
4588 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4589 unsigned long batch)
4591 /* start with a fail safe value for batch */
4595 /* Update high, then batch, in order */
4602 /* a companion to pageset_set_high() */
4603 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4605 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4608 static void pageset_init(struct per_cpu_pageset *p)
4610 struct per_cpu_pages *pcp;
4613 memset(p, 0, sizeof(*p));
4617 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4618 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4621 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4624 pageset_set_batch(p, batch);
4628 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4629 * to the value high for the pageset p.
4631 static void pageset_set_high(struct per_cpu_pageset *p,
4634 unsigned long batch = max(1UL, high / 4);
4635 if ((high / 4) > (PAGE_SHIFT * 8))
4636 batch = PAGE_SHIFT * 8;
4638 pageset_update(&p->pcp, high, batch);
4641 static void pageset_set_high_and_batch(struct zone *zone,
4642 struct per_cpu_pageset *pcp)
4644 if (percpu_pagelist_fraction)
4645 pageset_set_high(pcp,
4646 (zone->managed_pages /
4647 percpu_pagelist_fraction));
4649 pageset_set_batch(pcp, zone_batchsize(zone));
4652 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4654 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4657 pageset_set_high_and_batch(zone, pcp);
4660 static void __meminit setup_zone_pageset(struct zone *zone)
4663 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4664 for_each_possible_cpu(cpu)
4665 zone_pageset_init(zone, cpu);
4669 * Allocate per cpu pagesets and initialize them.
4670 * Before this call only boot pagesets were available.
4672 void __init setup_per_cpu_pageset(void)
4676 for_each_populated_zone(zone)
4677 setup_zone_pageset(zone);
4680 static noinline __init_refok
4681 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4687 * The per-page waitqueue mechanism uses hashed waitqueues
4690 zone->wait_table_hash_nr_entries =
4691 wait_table_hash_nr_entries(zone_size_pages);
4692 zone->wait_table_bits =
4693 wait_table_bits(zone->wait_table_hash_nr_entries);
4694 alloc_size = zone->wait_table_hash_nr_entries
4695 * sizeof(wait_queue_head_t);
4697 if (!slab_is_available()) {
4698 zone->wait_table = (wait_queue_head_t *)
4699 memblock_virt_alloc_node_nopanic(
4700 alloc_size, zone->zone_pgdat->node_id);
4703 * This case means that a zone whose size was 0 gets new memory
4704 * via memory hot-add.
4705 * But it may be the case that a new node was hot-added. In
4706 * this case vmalloc() will not be able to use this new node's
4707 * memory - this wait_table must be initialized to use this new
4708 * node itself as well.
4709 * To use this new node's memory, further consideration will be
4712 zone->wait_table = vmalloc(alloc_size);
4714 if (!zone->wait_table)
4717 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4718 init_waitqueue_head(zone->wait_table + i);
4723 static __meminit void zone_pcp_init(struct zone *zone)
4726 * per cpu subsystem is not up at this point. The following code
4727 * relies on the ability of the linker to provide the
4728 * offset of a (static) per cpu variable into the per cpu area.
4730 zone->pageset = &boot_pageset;
4732 if (populated_zone(zone))
4733 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4734 zone->name, zone->present_pages,
4735 zone_batchsize(zone));
4738 int __meminit init_currently_empty_zone(struct zone *zone,
4739 unsigned long zone_start_pfn,
4742 struct pglist_data *pgdat = zone->zone_pgdat;
4744 ret = zone_wait_table_init(zone, size);
4747 pgdat->nr_zones = zone_idx(zone) + 1;
4749 zone->zone_start_pfn = zone_start_pfn;
4751 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4752 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4754 (unsigned long)zone_idx(zone),
4755 zone_start_pfn, (zone_start_pfn + size));
4757 zone_init_free_lists(zone);
4762 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4763 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4766 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4768 int __meminit __early_pfn_to_nid(unsigned long pfn,
4769 struct mminit_pfnnid_cache *state)
4771 unsigned long start_pfn, end_pfn;
4774 if (state->last_start <= pfn && pfn < state->last_end)
4775 return state->last_nid;
4777 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4779 state->last_start = start_pfn;
4780 state->last_end = end_pfn;
4781 state->last_nid = nid;
4786 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4789 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4790 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4791 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4793 * If an architecture guarantees that all ranges registered contain no holes
4794 * and may be freed, this this function may be used instead of calling
4795 * memblock_free_early_nid() manually.
4797 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4799 unsigned long start_pfn, end_pfn;
4802 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4803 start_pfn = min(start_pfn, max_low_pfn);
4804 end_pfn = min(end_pfn, max_low_pfn);
4806 if (start_pfn < end_pfn)
4807 memblock_free_early_nid(PFN_PHYS(start_pfn),
4808 (end_pfn - start_pfn) << PAGE_SHIFT,
4814 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4815 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4817 * If an architecture guarantees that all ranges registered contain no holes and may
4818 * be freed, this function may be used instead of calling memory_present() manually.
4820 void __init sparse_memory_present_with_active_regions(int nid)
4822 unsigned long start_pfn, end_pfn;
4825 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4826 memory_present(this_nid, start_pfn, end_pfn);
4830 * get_pfn_range_for_nid - Return the start and end page frames for a node
4831 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4832 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4833 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4835 * It returns the start and end page frame of a node based on information
4836 * provided by memblock_set_node(). If called for a node
4837 * with no available memory, a warning is printed and the start and end
4840 void __meminit get_pfn_range_for_nid(unsigned int nid,
4841 unsigned long *start_pfn, unsigned long *end_pfn)
4843 unsigned long this_start_pfn, this_end_pfn;
4849 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4850 *start_pfn = min(*start_pfn, this_start_pfn);
4851 *end_pfn = max(*end_pfn, this_end_pfn);
4854 if (*start_pfn == -1UL)
4859 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4860 * assumption is made that zones within a node are ordered in monotonic
4861 * increasing memory addresses so that the "highest" populated zone is used
4863 static void __init find_usable_zone_for_movable(void)
4866 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4867 if (zone_index == ZONE_MOVABLE)
4870 if (arch_zone_highest_possible_pfn[zone_index] >
4871 arch_zone_lowest_possible_pfn[zone_index])
4875 VM_BUG_ON(zone_index == -1);
4876 movable_zone = zone_index;
4880 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4881 * because it is sized independent of architecture. Unlike the other zones,
4882 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4883 * in each node depending on the size of each node and how evenly kernelcore
4884 * is distributed. This helper function adjusts the zone ranges
4885 * provided by the architecture for a given node by using the end of the
4886 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4887 * zones within a node are in order of monotonic increases memory addresses
4889 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4890 unsigned long zone_type,
4891 unsigned long node_start_pfn,
4892 unsigned long node_end_pfn,
4893 unsigned long *zone_start_pfn,
4894 unsigned long *zone_end_pfn)
4896 /* Only adjust if ZONE_MOVABLE is on this node */
4897 if (zone_movable_pfn[nid]) {
4898 /* Size ZONE_MOVABLE */
4899 if (zone_type == ZONE_MOVABLE) {
4900 *zone_start_pfn = zone_movable_pfn[nid];
4901 *zone_end_pfn = min(node_end_pfn,
4902 arch_zone_highest_possible_pfn[movable_zone]);
4904 /* Adjust for ZONE_MOVABLE starting within this range */
4905 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4906 *zone_end_pfn > zone_movable_pfn[nid]) {
4907 *zone_end_pfn = zone_movable_pfn[nid];
4909 /* Check if this whole range is within ZONE_MOVABLE */
4910 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4911 *zone_start_pfn = *zone_end_pfn;
4916 * Return the number of pages a zone spans in a node, including holes
4917 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4919 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4920 unsigned long zone_type,
4921 unsigned long node_start_pfn,
4922 unsigned long node_end_pfn,
4923 unsigned long *ignored)
4925 unsigned long zone_start_pfn, zone_end_pfn;
4927 /* When hotadd a new node from cpu_up(), the node should be empty */
4928 if (!node_start_pfn && !node_end_pfn)
4931 /* Get the start and end of the zone */
4932 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4933 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4934 adjust_zone_range_for_zone_movable(nid, zone_type,
4935 node_start_pfn, node_end_pfn,
4936 &zone_start_pfn, &zone_end_pfn);
4938 /* Check that this node has pages within the zone's required range */
4939 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4942 /* Move the zone boundaries inside the node if necessary */
4943 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4944 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4946 /* Return the spanned pages */
4947 return zone_end_pfn - zone_start_pfn;
4951 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4952 * then all holes in the requested range will be accounted for.
4954 unsigned long __meminit __absent_pages_in_range(int nid,
4955 unsigned long range_start_pfn,
4956 unsigned long range_end_pfn)
4958 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4959 unsigned long start_pfn, end_pfn;
4962 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4963 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4964 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4965 nr_absent -= end_pfn - start_pfn;
4971 * absent_pages_in_range - Return number of page frames in holes within a range
4972 * @start_pfn: The start PFN to start searching for holes
4973 * @end_pfn: The end PFN to stop searching for holes
4975 * It returns the number of pages frames in memory holes within a range.
4977 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4978 unsigned long end_pfn)
4980 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4983 /* Return the number of page frames in holes in a zone on a node */
4984 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4985 unsigned long zone_type,
4986 unsigned long node_start_pfn,
4987 unsigned long node_end_pfn,
4988 unsigned long *ignored)
4990 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4991 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4992 unsigned long zone_start_pfn, zone_end_pfn;
4994 /* When hotadd a new node from cpu_up(), the node should be empty */
4995 if (!node_start_pfn && !node_end_pfn)
4998 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4999 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5001 adjust_zone_range_for_zone_movable(nid, zone_type,
5002 node_start_pfn, node_end_pfn,
5003 &zone_start_pfn, &zone_end_pfn);
5004 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5007 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5008 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5009 unsigned long zone_type,
5010 unsigned long node_start_pfn,
5011 unsigned long node_end_pfn,
5012 unsigned long *zones_size)
5014 return zones_size[zone_type];
5017 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5018 unsigned long zone_type,
5019 unsigned long node_start_pfn,
5020 unsigned long node_end_pfn,
5021 unsigned long *zholes_size)
5026 return zholes_size[zone_type];
5029 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5031 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5032 unsigned long node_start_pfn,
5033 unsigned long node_end_pfn,
5034 unsigned long *zones_size,
5035 unsigned long *zholes_size)
5037 unsigned long realtotalpages = 0, totalpages = 0;
5040 for (i = 0; i < MAX_NR_ZONES; i++) {
5041 struct zone *zone = pgdat->node_zones + i;
5042 unsigned long size, real_size;
5044 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5048 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5049 node_start_pfn, node_end_pfn,
5051 zone->spanned_pages = size;
5052 zone->present_pages = real_size;
5055 realtotalpages += real_size;
5058 pgdat->node_spanned_pages = totalpages;
5059 pgdat->node_present_pages = realtotalpages;
5060 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5064 #ifndef CONFIG_SPARSEMEM
5066 * Calculate the size of the zone->blockflags rounded to an unsigned long
5067 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5068 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5069 * round what is now in bits to nearest long in bits, then return it in
5072 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5074 unsigned long usemapsize;
5076 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5077 usemapsize = roundup(zonesize, pageblock_nr_pages);
5078 usemapsize = usemapsize >> pageblock_order;
5079 usemapsize *= NR_PAGEBLOCK_BITS;
5080 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5082 return usemapsize / 8;
5085 static void __init setup_usemap(struct pglist_data *pgdat,
5087 unsigned long zone_start_pfn,
5088 unsigned long zonesize)
5090 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5091 zone->pageblock_flags = NULL;
5093 zone->pageblock_flags =
5094 memblock_virt_alloc_node_nopanic(usemapsize,
5098 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5099 unsigned long zone_start_pfn, unsigned long zonesize) {}
5100 #endif /* CONFIG_SPARSEMEM */
5102 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5104 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5105 void __paginginit set_pageblock_order(void)
5109 /* Check that pageblock_nr_pages has not already been setup */
5110 if (pageblock_order)
5113 if (HPAGE_SHIFT > PAGE_SHIFT)
5114 order = HUGETLB_PAGE_ORDER;
5116 order = MAX_ORDER - 1;
5119 * Assume the largest contiguous order of interest is a huge page.
5120 * This value may be variable depending on boot parameters on IA64 and
5123 pageblock_order = order;
5125 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5128 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5129 * is unused as pageblock_order is set at compile-time. See
5130 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5133 void __paginginit set_pageblock_order(void)
5137 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5139 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5140 unsigned long present_pages)
5142 unsigned long pages = spanned_pages;
5145 * Provide a more accurate estimation if there are holes within
5146 * the zone and SPARSEMEM is in use. If there are holes within the
5147 * zone, each populated memory region may cost us one or two extra
5148 * memmap pages due to alignment because memmap pages for each
5149 * populated regions may not naturally algined on page boundary.
5150 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5152 if (spanned_pages > present_pages + (present_pages >> 4) &&
5153 IS_ENABLED(CONFIG_SPARSEMEM))
5154 pages = present_pages;
5156 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5160 * Set up the zone data structures:
5161 * - mark all pages reserved
5162 * - mark all memory queues empty
5163 * - clear the memory bitmaps
5165 * NOTE: pgdat should get zeroed by caller.
5167 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5170 int nid = pgdat->node_id;
5171 unsigned long zone_start_pfn = pgdat->node_start_pfn;
5174 pgdat_resize_init(pgdat);
5175 #ifdef CONFIG_NUMA_BALANCING
5176 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5177 pgdat->numabalancing_migrate_nr_pages = 0;
5178 pgdat->numabalancing_migrate_next_window = jiffies;
5180 init_waitqueue_head(&pgdat->kswapd_wait);
5181 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5182 pgdat_page_ext_init(pgdat);
5184 for (j = 0; j < MAX_NR_ZONES; j++) {
5185 struct zone *zone = pgdat->node_zones + j;
5186 unsigned long size, realsize, freesize, memmap_pages;
5188 size = zone->spanned_pages;
5189 realsize = freesize = zone->present_pages;
5192 * Adjust freesize so that it accounts for how much memory
5193 * is used by this zone for memmap. This affects the watermark
5194 * and per-cpu initialisations
5196 memmap_pages = calc_memmap_size(size, realsize);
5197 if (!is_highmem_idx(j)) {
5198 if (freesize >= memmap_pages) {
5199 freesize -= memmap_pages;
5202 " %s zone: %lu pages used for memmap\n",
5203 zone_names[j], memmap_pages);
5206 " %s zone: %lu pages exceeds freesize %lu\n",
5207 zone_names[j], memmap_pages, freesize);
5210 /* Account for reserved pages */
5211 if (j == 0 && freesize > dma_reserve) {
5212 freesize -= dma_reserve;
5213 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5214 zone_names[0], dma_reserve);
5217 if (!is_highmem_idx(j))
5218 nr_kernel_pages += freesize;
5219 /* Charge for highmem memmap if there are enough kernel pages */
5220 else if (nr_kernel_pages > memmap_pages * 2)
5221 nr_kernel_pages -= memmap_pages;
5222 nr_all_pages += freesize;
5225 * Set an approximate value for lowmem here, it will be adjusted
5226 * when the bootmem allocator frees pages into the buddy system.
5227 * And all highmem pages will be managed by the buddy system.
5229 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5232 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5234 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5236 zone->name = zone_names[j];
5237 spin_lock_init(&zone->lock);
5238 spin_lock_init(&zone->lru_lock);
5239 zone_seqlock_init(zone);
5240 zone->zone_pgdat = pgdat;
5241 zone_pcp_init(zone);
5243 /* For bootup, initialized properly in watermark setup */
5244 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5246 lruvec_init(&zone->lruvec);
5250 set_pageblock_order();
5251 setup_usemap(pgdat, zone, zone_start_pfn, size);
5252 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5254 memmap_init(size, nid, j, zone_start_pfn);
5255 zone_start_pfn += size;
5259 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5261 unsigned long __maybe_unused start = 0;
5262 unsigned long __maybe_unused offset = 0;
5264 /* Skip empty nodes */
5265 if (!pgdat->node_spanned_pages)
5268 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5269 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5270 offset = pgdat->node_start_pfn - start;
5271 /* ia64 gets its own node_mem_map, before this, without bootmem */
5272 if (!pgdat->node_mem_map) {
5273 unsigned long size, end;
5277 * The zone's endpoints aren't required to be MAX_ORDER
5278 * aligned but the node_mem_map endpoints must be in order
5279 * for the buddy allocator to function correctly.
5281 end = pgdat_end_pfn(pgdat);
5282 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5283 size = (end - start) * sizeof(struct page);
5284 map = alloc_remap(pgdat->node_id, size);
5286 map = memblock_virt_alloc_node_nopanic(size,
5288 pgdat->node_mem_map = map + offset;
5290 #ifndef CONFIG_NEED_MULTIPLE_NODES
5292 * With no DISCONTIG, the global mem_map is just set as node 0's
5294 if (pgdat == NODE_DATA(0)) {
5295 mem_map = NODE_DATA(0)->node_mem_map;
5296 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5297 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5299 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5302 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5305 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5306 unsigned long node_start_pfn, unsigned long *zholes_size)
5308 pg_data_t *pgdat = NODE_DATA(nid);
5309 unsigned long start_pfn = 0;
5310 unsigned long end_pfn = 0;
5312 /* pg_data_t should be reset to zero when it's allocated */
5313 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5315 reset_deferred_meminit(pgdat);
5316 pgdat->node_id = nid;
5317 pgdat->node_start_pfn = node_start_pfn;
5318 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5319 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5320 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5321 (u64)start_pfn << PAGE_SHIFT,
5322 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5324 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5325 zones_size, zholes_size);
5327 alloc_node_mem_map(pgdat);
5328 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5329 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5330 nid, (unsigned long)pgdat,
5331 (unsigned long)pgdat->node_mem_map);
5334 free_area_init_core(pgdat);
5337 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5339 #if MAX_NUMNODES > 1
5341 * Figure out the number of possible node ids.
5343 void __init setup_nr_node_ids(void)
5345 unsigned int highest;
5347 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5348 nr_node_ids = highest + 1;
5353 * node_map_pfn_alignment - determine the maximum internode alignment
5355 * This function should be called after node map is populated and sorted.
5356 * It calculates the maximum power of two alignment which can distinguish
5359 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5360 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5361 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5362 * shifted, 1GiB is enough and this function will indicate so.
5364 * This is used to test whether pfn -> nid mapping of the chosen memory
5365 * model has fine enough granularity to avoid incorrect mapping for the
5366 * populated node map.
5368 * Returns the determined alignment in pfn's. 0 if there is no alignment
5369 * requirement (single node).
5371 unsigned long __init node_map_pfn_alignment(void)
5373 unsigned long accl_mask = 0, last_end = 0;
5374 unsigned long start, end, mask;
5378 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5379 if (!start || last_nid < 0 || last_nid == nid) {
5386 * Start with a mask granular enough to pin-point to the
5387 * start pfn and tick off bits one-by-one until it becomes
5388 * too coarse to separate the current node from the last.
5390 mask = ~((1 << __ffs(start)) - 1);
5391 while (mask && last_end <= (start & (mask << 1)))
5394 /* accumulate all internode masks */
5398 /* convert mask to number of pages */
5399 return ~accl_mask + 1;
5402 /* Find the lowest pfn for a node */
5403 static unsigned long __init find_min_pfn_for_node(int nid)
5405 unsigned long min_pfn = ULONG_MAX;
5406 unsigned long start_pfn;
5409 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5410 min_pfn = min(min_pfn, start_pfn);
5412 if (min_pfn == ULONG_MAX) {
5414 "Could not find start_pfn for node %d\n", nid);
5422 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5424 * It returns the minimum PFN based on information provided via
5425 * memblock_set_node().
5427 unsigned long __init find_min_pfn_with_active_regions(void)
5429 return find_min_pfn_for_node(MAX_NUMNODES);
5433 * early_calculate_totalpages()
5434 * Sum pages in active regions for movable zone.
5435 * Populate N_MEMORY for calculating usable_nodes.
5437 static unsigned long __init early_calculate_totalpages(void)
5439 unsigned long totalpages = 0;
5440 unsigned long start_pfn, end_pfn;
5443 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5444 unsigned long pages = end_pfn - start_pfn;
5446 totalpages += pages;
5448 node_set_state(nid, N_MEMORY);
5454 * Find the PFN the Movable zone begins in each node. Kernel memory
5455 * is spread evenly between nodes as long as the nodes have enough
5456 * memory. When they don't, some nodes will have more kernelcore than
5459 static void __init find_zone_movable_pfns_for_nodes(void)
5462 unsigned long usable_startpfn;
5463 unsigned long kernelcore_node, kernelcore_remaining;
5464 /* save the state before borrow the nodemask */
5465 nodemask_t saved_node_state = node_states[N_MEMORY];
5466 unsigned long totalpages = early_calculate_totalpages();
5467 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5468 struct memblock_region *r;
5470 /* Need to find movable_zone earlier when movable_node is specified. */
5471 find_usable_zone_for_movable();
5474 * If movable_node is specified, ignore kernelcore and movablecore
5477 if (movable_node_is_enabled()) {
5478 for_each_memblock(memory, r) {
5479 if (!memblock_is_hotpluggable(r))
5484 usable_startpfn = PFN_DOWN(r->base);
5485 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5486 min(usable_startpfn, zone_movable_pfn[nid]) :
5494 * If movablecore=nn[KMG] was specified, calculate what size of
5495 * kernelcore that corresponds so that memory usable for
5496 * any allocation type is evenly spread. If both kernelcore
5497 * and movablecore are specified, then the value of kernelcore
5498 * will be used for required_kernelcore if it's greater than
5499 * what movablecore would have allowed.
5501 if (required_movablecore) {
5502 unsigned long corepages;
5505 * Round-up so that ZONE_MOVABLE is at least as large as what
5506 * was requested by the user
5508 required_movablecore =
5509 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5510 required_movablecore = min(totalpages, required_movablecore);
5511 corepages = totalpages - required_movablecore;
5513 required_kernelcore = max(required_kernelcore, corepages);
5517 * If kernelcore was not specified or kernelcore size is larger
5518 * than totalpages, there is no ZONE_MOVABLE.
5520 if (!required_kernelcore || required_kernelcore >= totalpages)
5523 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5524 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5527 /* Spread kernelcore memory as evenly as possible throughout nodes */
5528 kernelcore_node = required_kernelcore / usable_nodes;
5529 for_each_node_state(nid, N_MEMORY) {
5530 unsigned long start_pfn, end_pfn;
5533 * Recalculate kernelcore_node if the division per node
5534 * now exceeds what is necessary to satisfy the requested
5535 * amount of memory for the kernel
5537 if (required_kernelcore < kernelcore_node)
5538 kernelcore_node = required_kernelcore / usable_nodes;
5541 * As the map is walked, we track how much memory is usable
5542 * by the kernel using kernelcore_remaining. When it is
5543 * 0, the rest of the node is usable by ZONE_MOVABLE
5545 kernelcore_remaining = kernelcore_node;
5547 /* Go through each range of PFNs within this node */
5548 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5549 unsigned long size_pages;
5551 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5552 if (start_pfn >= end_pfn)
5555 /* Account for what is only usable for kernelcore */
5556 if (start_pfn < usable_startpfn) {
5557 unsigned long kernel_pages;
5558 kernel_pages = min(end_pfn, usable_startpfn)
5561 kernelcore_remaining -= min(kernel_pages,
5562 kernelcore_remaining);
5563 required_kernelcore -= min(kernel_pages,
5564 required_kernelcore);
5566 /* Continue if range is now fully accounted */
5567 if (end_pfn <= usable_startpfn) {
5570 * Push zone_movable_pfn to the end so
5571 * that if we have to rebalance
5572 * kernelcore across nodes, we will
5573 * not double account here
5575 zone_movable_pfn[nid] = end_pfn;
5578 start_pfn = usable_startpfn;
5582 * The usable PFN range for ZONE_MOVABLE is from
5583 * start_pfn->end_pfn. Calculate size_pages as the
5584 * number of pages used as kernelcore
5586 size_pages = end_pfn - start_pfn;
5587 if (size_pages > kernelcore_remaining)
5588 size_pages = kernelcore_remaining;
5589 zone_movable_pfn[nid] = start_pfn + size_pages;
5592 * Some kernelcore has been met, update counts and
5593 * break if the kernelcore for this node has been
5596 required_kernelcore -= min(required_kernelcore,
5598 kernelcore_remaining -= size_pages;
5599 if (!kernelcore_remaining)
5605 * If there is still required_kernelcore, we do another pass with one
5606 * less node in the count. This will push zone_movable_pfn[nid] further
5607 * along on the nodes that still have memory until kernelcore is
5611 if (usable_nodes && required_kernelcore > usable_nodes)
5615 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5616 for (nid = 0; nid < MAX_NUMNODES; nid++)
5617 zone_movable_pfn[nid] =
5618 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5621 /* restore the node_state */
5622 node_states[N_MEMORY] = saved_node_state;
5625 /* Any regular or high memory on that node ? */
5626 static void check_for_memory(pg_data_t *pgdat, int nid)
5628 enum zone_type zone_type;
5630 if (N_MEMORY == N_NORMAL_MEMORY)
5633 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5634 struct zone *zone = &pgdat->node_zones[zone_type];
5635 if (populated_zone(zone)) {
5636 node_set_state(nid, N_HIGH_MEMORY);
5637 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5638 zone_type <= ZONE_NORMAL)
5639 node_set_state(nid, N_NORMAL_MEMORY);
5646 * free_area_init_nodes - Initialise all pg_data_t and zone data
5647 * @max_zone_pfn: an array of max PFNs for each zone
5649 * This will call free_area_init_node() for each active node in the system.
5650 * Using the page ranges provided by memblock_set_node(), the size of each
5651 * zone in each node and their holes is calculated. If the maximum PFN
5652 * between two adjacent zones match, it is assumed that the zone is empty.
5653 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5654 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5655 * starts where the previous one ended. For example, ZONE_DMA32 starts
5656 * at arch_max_dma_pfn.
5658 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5660 unsigned long start_pfn, end_pfn;
5663 /* Record where the zone boundaries are */
5664 memset(arch_zone_lowest_possible_pfn, 0,
5665 sizeof(arch_zone_lowest_possible_pfn));
5666 memset(arch_zone_highest_possible_pfn, 0,
5667 sizeof(arch_zone_highest_possible_pfn));
5668 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5669 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5670 for (i = 1; i < MAX_NR_ZONES; i++) {
5671 if (i == ZONE_MOVABLE)
5673 arch_zone_lowest_possible_pfn[i] =
5674 arch_zone_highest_possible_pfn[i-1];
5675 arch_zone_highest_possible_pfn[i] =
5676 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5678 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5679 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5681 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5682 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5683 find_zone_movable_pfns_for_nodes();
5685 /* Print out the zone ranges */
5686 pr_info("Zone ranges:\n");
5687 for (i = 0; i < MAX_NR_ZONES; i++) {
5688 if (i == ZONE_MOVABLE)
5690 pr_info(" %-8s ", zone_names[i]);
5691 if (arch_zone_lowest_possible_pfn[i] ==
5692 arch_zone_highest_possible_pfn[i])
5695 pr_cont("[mem %#018Lx-%#018Lx]\n",
5696 (u64)arch_zone_lowest_possible_pfn[i]
5698 ((u64)arch_zone_highest_possible_pfn[i]
5699 << PAGE_SHIFT) - 1);
5702 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5703 pr_info("Movable zone start for each node\n");
5704 for (i = 0; i < MAX_NUMNODES; i++) {
5705 if (zone_movable_pfn[i])
5706 pr_info(" Node %d: %#018Lx\n", i,
5707 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5710 /* Print out the early node map */
5711 pr_info("Early memory node ranges\n");
5712 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5713 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5714 (u64)start_pfn << PAGE_SHIFT,
5715 ((u64)end_pfn << PAGE_SHIFT) - 1);
5717 /* Initialise every node */
5718 mminit_verify_pageflags_layout();
5719 setup_nr_node_ids();
5720 for_each_online_node(nid) {
5721 pg_data_t *pgdat = NODE_DATA(nid);
5722 free_area_init_node(nid, NULL,
5723 find_min_pfn_for_node(nid), NULL);
5725 /* Any memory on that node */
5726 if (pgdat->node_present_pages)
5727 node_set_state(nid, N_MEMORY);
5728 check_for_memory(pgdat, nid);
5732 static int __init cmdline_parse_core(char *p, unsigned long *core)
5734 unsigned long long coremem;
5738 coremem = memparse(p, &p);
5739 *core = coremem >> PAGE_SHIFT;
5741 /* Paranoid check that UL is enough for the coremem value */
5742 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5748 * kernelcore=size sets the amount of memory for use for allocations that
5749 * cannot be reclaimed or migrated.
5751 static int __init cmdline_parse_kernelcore(char *p)
5753 return cmdline_parse_core(p, &required_kernelcore);
5757 * movablecore=size sets the amount of memory for use for allocations that
5758 * can be reclaimed or migrated.
5760 static int __init cmdline_parse_movablecore(char *p)
5762 return cmdline_parse_core(p, &required_movablecore);
5765 early_param("kernelcore", cmdline_parse_kernelcore);
5766 early_param("movablecore", cmdline_parse_movablecore);
5768 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5770 void adjust_managed_page_count(struct page *page, long count)
5772 spin_lock(&managed_page_count_lock);
5773 page_zone(page)->managed_pages += count;
5774 totalram_pages += count;
5775 #ifdef CONFIG_HIGHMEM
5776 if (PageHighMem(page))
5777 totalhigh_pages += count;
5779 spin_unlock(&managed_page_count_lock);
5781 EXPORT_SYMBOL(adjust_managed_page_count);
5783 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5786 unsigned long pages = 0;
5788 start = (void *)PAGE_ALIGN((unsigned long)start);
5789 end = (void *)((unsigned long)end & PAGE_MASK);
5790 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5791 if ((unsigned int)poison <= 0xFF)
5792 memset(pos, poison, PAGE_SIZE);
5793 free_reserved_page(virt_to_page(pos));
5797 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5798 s, pages << (PAGE_SHIFT - 10), start, end);
5802 EXPORT_SYMBOL(free_reserved_area);
5804 #ifdef CONFIG_HIGHMEM
5805 void free_highmem_page(struct page *page)
5807 __free_reserved_page(page);
5809 page_zone(page)->managed_pages++;
5815 void __init mem_init_print_info(const char *str)
5817 unsigned long physpages, codesize, datasize, rosize, bss_size;
5818 unsigned long init_code_size, init_data_size;
5820 physpages = get_num_physpages();
5821 codesize = _etext - _stext;
5822 datasize = _edata - _sdata;
5823 rosize = __end_rodata - __start_rodata;
5824 bss_size = __bss_stop - __bss_start;
5825 init_data_size = __init_end - __init_begin;
5826 init_code_size = _einittext - _sinittext;
5829 * Detect special cases and adjust section sizes accordingly:
5830 * 1) .init.* may be embedded into .data sections
5831 * 2) .init.text.* may be out of [__init_begin, __init_end],
5832 * please refer to arch/tile/kernel/vmlinux.lds.S.
5833 * 3) .rodata.* may be embedded into .text or .data sections.
5835 #define adj_init_size(start, end, size, pos, adj) \
5837 if (start <= pos && pos < end && size > adj) \
5841 adj_init_size(__init_begin, __init_end, init_data_size,
5842 _sinittext, init_code_size);
5843 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5844 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5845 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5846 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5848 #undef adj_init_size
5850 pr_info("Memory: %luK/%luK available "
5851 "(%luK kernel code, %luK rwdata, %luK rodata, "
5852 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5853 #ifdef CONFIG_HIGHMEM
5857 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5858 codesize >> 10, datasize >> 10, rosize >> 10,
5859 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5860 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5861 totalcma_pages << (PAGE_SHIFT-10),
5862 #ifdef CONFIG_HIGHMEM
5863 totalhigh_pages << (PAGE_SHIFT-10),
5865 str ? ", " : "", str ? str : "");
5869 * set_dma_reserve - set the specified number of pages reserved in the first zone
5870 * @new_dma_reserve: The number of pages to mark reserved
5872 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
5873 * In the DMA zone, a significant percentage may be consumed by kernel image
5874 * and other unfreeable allocations which can skew the watermarks badly. This
5875 * function may optionally be used to account for unfreeable pages in the
5876 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5877 * smaller per-cpu batchsize.
5879 void __init set_dma_reserve(unsigned long new_dma_reserve)
5881 dma_reserve = new_dma_reserve;
5884 void __init free_area_init(unsigned long *zones_size)
5886 free_area_init_node(0, zones_size,
5887 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5890 static int page_alloc_cpu_notify(struct notifier_block *self,
5891 unsigned long action, void *hcpu)
5893 int cpu = (unsigned long)hcpu;
5895 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5896 lru_add_drain_cpu(cpu);
5900 * Spill the event counters of the dead processor
5901 * into the current processors event counters.
5902 * This artificially elevates the count of the current
5905 vm_events_fold_cpu(cpu);
5908 * Zero the differential counters of the dead processor
5909 * so that the vm statistics are consistent.
5911 * This is only okay since the processor is dead and cannot
5912 * race with what we are doing.
5914 cpu_vm_stats_fold(cpu);
5919 void __init page_alloc_init(void)
5921 hotcpu_notifier(page_alloc_cpu_notify, 0);
5925 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5926 * or min_free_kbytes changes.
5928 static void calculate_totalreserve_pages(void)
5930 struct pglist_data *pgdat;
5931 unsigned long reserve_pages = 0;
5932 enum zone_type i, j;
5934 for_each_online_pgdat(pgdat) {
5935 for (i = 0; i < MAX_NR_ZONES; i++) {
5936 struct zone *zone = pgdat->node_zones + i;
5939 /* Find valid and maximum lowmem_reserve in the zone */
5940 for (j = i; j < MAX_NR_ZONES; j++) {
5941 if (zone->lowmem_reserve[j] > max)
5942 max = zone->lowmem_reserve[j];
5945 /* we treat the high watermark as reserved pages. */
5946 max += high_wmark_pages(zone);
5948 if (max > zone->managed_pages)
5949 max = zone->managed_pages;
5951 zone->totalreserve_pages = max;
5953 reserve_pages += max;
5956 totalreserve_pages = reserve_pages;
5960 * setup_per_zone_lowmem_reserve - called whenever
5961 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
5962 * has a correct pages reserved value, so an adequate number of
5963 * pages are left in the zone after a successful __alloc_pages().
5965 static void setup_per_zone_lowmem_reserve(void)
5967 struct pglist_data *pgdat;
5968 enum zone_type j, idx;
5970 for_each_online_pgdat(pgdat) {
5971 for (j = 0; j < MAX_NR_ZONES; j++) {
5972 struct zone *zone = pgdat->node_zones + j;
5973 unsigned long managed_pages = zone->managed_pages;
5975 zone->lowmem_reserve[j] = 0;
5979 struct zone *lower_zone;
5983 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5984 sysctl_lowmem_reserve_ratio[idx] = 1;
5986 lower_zone = pgdat->node_zones + idx;
5987 lower_zone->lowmem_reserve[j] = managed_pages /
5988 sysctl_lowmem_reserve_ratio[idx];
5989 managed_pages += lower_zone->managed_pages;
5994 /* update totalreserve_pages */
5995 calculate_totalreserve_pages();
5998 static void __setup_per_zone_wmarks(void)
6000 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6001 unsigned long lowmem_pages = 0;
6003 unsigned long flags;
6005 /* Calculate total number of !ZONE_HIGHMEM pages */
6006 for_each_zone(zone) {
6007 if (!is_highmem(zone))
6008 lowmem_pages += zone->managed_pages;
6011 for_each_zone(zone) {
6014 spin_lock_irqsave(&zone->lock, flags);
6015 tmp = (u64)pages_min * zone->managed_pages;
6016 do_div(tmp, lowmem_pages);
6017 if (is_highmem(zone)) {
6019 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6020 * need highmem pages, so cap pages_min to a small
6023 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6024 * deltas control asynch page reclaim, and so should
6025 * not be capped for highmem.
6027 unsigned long min_pages;
6029 min_pages = zone->managed_pages / 1024;
6030 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6031 zone->watermark[WMARK_MIN] = min_pages;
6034 * If it's a lowmem zone, reserve a number of pages
6035 * proportionate to the zone's size.
6037 zone->watermark[WMARK_MIN] = tmp;
6040 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
6041 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
6043 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6044 high_wmark_pages(zone) - low_wmark_pages(zone) -
6045 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6047 spin_unlock_irqrestore(&zone->lock, flags);
6050 /* update totalreserve_pages */
6051 calculate_totalreserve_pages();
6055 * setup_per_zone_wmarks - called when min_free_kbytes changes
6056 * or when memory is hot-{added|removed}
6058 * Ensures that the watermark[min,low,high] values for each zone are set
6059 * correctly with respect to min_free_kbytes.
6061 void setup_per_zone_wmarks(void)
6063 mutex_lock(&zonelists_mutex);
6064 __setup_per_zone_wmarks();
6065 mutex_unlock(&zonelists_mutex);
6069 * The inactive anon list should be small enough that the VM never has to
6070 * do too much work, but large enough that each inactive page has a chance
6071 * to be referenced again before it is swapped out.
6073 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6074 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6075 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6076 * the anonymous pages are kept on the inactive list.
6079 * memory ratio inactive anon
6080 * -------------------------------------
6089 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6091 unsigned int gb, ratio;
6093 /* Zone size in gigabytes */
6094 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6096 ratio = int_sqrt(10 * gb);
6100 zone->inactive_ratio = ratio;
6103 static void __meminit setup_per_zone_inactive_ratio(void)
6108 calculate_zone_inactive_ratio(zone);
6112 * Initialise min_free_kbytes.
6114 * For small machines we want it small (128k min). For large machines
6115 * we want it large (64MB max). But it is not linear, because network
6116 * bandwidth does not increase linearly with machine size. We use
6118 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6119 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6135 int __meminit init_per_zone_wmark_min(void)
6137 unsigned long lowmem_kbytes;
6138 int new_min_free_kbytes;
6140 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6141 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6143 if (new_min_free_kbytes > user_min_free_kbytes) {
6144 min_free_kbytes = new_min_free_kbytes;
6145 if (min_free_kbytes < 128)
6146 min_free_kbytes = 128;
6147 if (min_free_kbytes > 65536)
6148 min_free_kbytes = 65536;
6150 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6151 new_min_free_kbytes, user_min_free_kbytes);
6153 setup_per_zone_wmarks();
6154 refresh_zone_stat_thresholds();
6155 setup_per_zone_lowmem_reserve();
6156 setup_per_zone_inactive_ratio();
6159 module_init(init_per_zone_wmark_min)
6162 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6163 * that we can call two helper functions whenever min_free_kbytes
6166 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6167 void __user *buffer, size_t *length, loff_t *ppos)
6171 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6176 user_min_free_kbytes = min_free_kbytes;
6177 setup_per_zone_wmarks();
6183 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6184 void __user *buffer, size_t *length, loff_t *ppos)
6189 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6194 zone->min_unmapped_pages = (zone->managed_pages *
6195 sysctl_min_unmapped_ratio) / 100;
6199 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6200 void __user *buffer, size_t *length, loff_t *ppos)
6205 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6210 zone->min_slab_pages = (zone->managed_pages *
6211 sysctl_min_slab_ratio) / 100;
6217 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6218 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6219 * whenever sysctl_lowmem_reserve_ratio changes.
6221 * The reserve ratio obviously has absolutely no relation with the
6222 * minimum watermarks. The lowmem reserve ratio can only make sense
6223 * if in function of the boot time zone sizes.
6225 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6226 void __user *buffer, size_t *length, loff_t *ppos)
6228 proc_dointvec_minmax(table, write, buffer, length, ppos);
6229 setup_per_zone_lowmem_reserve();
6234 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6235 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6236 * pagelist can have before it gets flushed back to buddy allocator.
6238 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6239 void __user *buffer, size_t *length, loff_t *ppos)
6242 int old_percpu_pagelist_fraction;
6245 mutex_lock(&pcp_batch_high_lock);
6246 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6248 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6249 if (!write || ret < 0)
6252 /* Sanity checking to avoid pcp imbalance */
6253 if (percpu_pagelist_fraction &&
6254 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6255 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6261 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6264 for_each_populated_zone(zone) {
6267 for_each_possible_cpu(cpu)
6268 pageset_set_high_and_batch(zone,
6269 per_cpu_ptr(zone->pageset, cpu));
6272 mutex_unlock(&pcp_batch_high_lock);
6277 int hashdist = HASHDIST_DEFAULT;
6279 static int __init set_hashdist(char *str)
6283 hashdist = simple_strtoul(str, &str, 0);
6286 __setup("hashdist=", set_hashdist);
6290 * allocate a large system hash table from bootmem
6291 * - it is assumed that the hash table must contain an exact power-of-2
6292 * quantity of entries
6293 * - limit is the number of hash buckets, not the total allocation size
6295 void *__init alloc_large_system_hash(const char *tablename,
6296 unsigned long bucketsize,
6297 unsigned long numentries,
6300 unsigned int *_hash_shift,
6301 unsigned int *_hash_mask,
6302 unsigned long low_limit,
6303 unsigned long high_limit)
6305 unsigned long long max = high_limit;
6306 unsigned long log2qty, size;
6309 /* allow the kernel cmdline to have a say */
6311 /* round applicable memory size up to nearest megabyte */
6312 numentries = nr_kernel_pages;
6314 /* It isn't necessary when PAGE_SIZE >= 1MB */
6315 if (PAGE_SHIFT < 20)
6316 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6318 /* limit to 1 bucket per 2^scale bytes of low memory */
6319 if (scale > PAGE_SHIFT)
6320 numentries >>= (scale - PAGE_SHIFT);
6322 numentries <<= (PAGE_SHIFT - scale);
6324 /* Make sure we've got at least a 0-order allocation.. */
6325 if (unlikely(flags & HASH_SMALL)) {
6326 /* Makes no sense without HASH_EARLY */
6327 WARN_ON(!(flags & HASH_EARLY));
6328 if (!(numentries >> *_hash_shift)) {
6329 numentries = 1UL << *_hash_shift;
6330 BUG_ON(!numentries);
6332 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6333 numentries = PAGE_SIZE / bucketsize;
6335 numentries = roundup_pow_of_two(numentries);
6337 /* limit allocation size to 1/16 total memory by default */
6339 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6340 do_div(max, bucketsize);
6342 max = min(max, 0x80000000ULL);
6344 if (numentries < low_limit)
6345 numentries = low_limit;
6346 if (numentries > max)
6349 log2qty = ilog2(numentries);
6352 size = bucketsize << log2qty;
6353 if (flags & HASH_EARLY)
6354 table = memblock_virt_alloc_nopanic(size, 0);
6356 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6359 * If bucketsize is not a power-of-two, we may free
6360 * some pages at the end of hash table which
6361 * alloc_pages_exact() automatically does
6363 if (get_order(size) < MAX_ORDER) {
6364 table = alloc_pages_exact(size, GFP_ATOMIC);
6365 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6368 } while (!table && size > PAGE_SIZE && --log2qty);
6371 panic("Failed to allocate %s hash table\n", tablename);
6373 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6376 ilog2(size) - PAGE_SHIFT,
6380 *_hash_shift = log2qty;
6382 *_hash_mask = (1 << log2qty) - 1;
6387 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6388 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6391 #ifdef CONFIG_SPARSEMEM
6392 return __pfn_to_section(pfn)->pageblock_flags;
6394 return zone->pageblock_flags;
6395 #endif /* CONFIG_SPARSEMEM */
6398 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6400 #ifdef CONFIG_SPARSEMEM
6401 pfn &= (PAGES_PER_SECTION-1);
6402 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6404 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6405 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6406 #endif /* CONFIG_SPARSEMEM */
6410 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6411 * @page: The page within the block of interest
6412 * @pfn: The target page frame number
6413 * @end_bitidx: The last bit of interest to retrieve
6414 * @mask: mask of bits that the caller is interested in
6416 * Return: pageblock_bits flags
6418 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6419 unsigned long end_bitidx,
6423 unsigned long *bitmap;
6424 unsigned long bitidx, word_bitidx;
6427 zone = page_zone(page);
6428 bitmap = get_pageblock_bitmap(zone, pfn);
6429 bitidx = pfn_to_bitidx(zone, pfn);
6430 word_bitidx = bitidx / BITS_PER_LONG;
6431 bitidx &= (BITS_PER_LONG-1);
6433 word = bitmap[word_bitidx];
6434 bitidx += end_bitidx;
6435 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6439 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6440 * @page: The page within the block of interest
6441 * @flags: The flags to set
6442 * @pfn: The target page frame number
6443 * @end_bitidx: The last bit of interest
6444 * @mask: mask of bits that the caller is interested in
6446 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6448 unsigned long end_bitidx,
6452 unsigned long *bitmap;
6453 unsigned long bitidx, word_bitidx;
6454 unsigned long old_word, word;
6456 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6458 zone = page_zone(page);
6459 bitmap = get_pageblock_bitmap(zone, pfn);
6460 bitidx = pfn_to_bitidx(zone, pfn);
6461 word_bitidx = bitidx / BITS_PER_LONG;
6462 bitidx &= (BITS_PER_LONG-1);
6464 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6466 bitidx += end_bitidx;
6467 mask <<= (BITS_PER_LONG - bitidx - 1);
6468 flags <<= (BITS_PER_LONG - bitidx - 1);
6470 word = READ_ONCE(bitmap[word_bitidx]);
6472 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6473 if (word == old_word)
6480 * This function checks whether pageblock includes unmovable pages or not.
6481 * If @count is not zero, it is okay to include less @count unmovable pages
6483 * PageLRU check without isolation or lru_lock could race so that
6484 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6485 * expect this function should be exact.
6487 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6488 bool skip_hwpoisoned_pages)
6490 unsigned long pfn, iter, found;
6494 * For avoiding noise data, lru_add_drain_all() should be called
6495 * If ZONE_MOVABLE, the zone never contains unmovable pages
6497 if (zone_idx(zone) == ZONE_MOVABLE)
6499 mt = get_pageblock_migratetype(page);
6500 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6503 pfn = page_to_pfn(page);
6504 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6505 unsigned long check = pfn + iter;
6507 if (!pfn_valid_within(check))
6510 page = pfn_to_page(check);
6513 * Hugepages are not in LRU lists, but they're movable.
6514 * We need not scan over tail pages bacause we don't
6515 * handle each tail page individually in migration.
6517 if (PageHuge(page)) {
6518 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6523 * We can't use page_count without pin a page
6524 * because another CPU can free compound page.
6525 * This check already skips compound tails of THP
6526 * because their page->_count is zero at all time.
6528 if (!atomic_read(&page->_count)) {
6529 if (PageBuddy(page))
6530 iter += (1 << page_order(page)) - 1;
6535 * The HWPoisoned page may be not in buddy system, and
6536 * page_count() is not 0.
6538 if (skip_hwpoisoned_pages && PageHWPoison(page))
6544 * If there are RECLAIMABLE pages, we need to check
6545 * it. But now, memory offline itself doesn't call
6546 * shrink_node_slabs() and it still to be fixed.
6549 * If the page is not RAM, page_count()should be 0.
6550 * we don't need more check. This is an _used_ not-movable page.
6552 * The problematic thing here is PG_reserved pages. PG_reserved
6553 * is set to both of a memory hole page and a _used_ kernel
6562 bool is_pageblock_removable_nolock(struct page *page)
6568 * We have to be careful here because we are iterating over memory
6569 * sections which are not zone aware so we might end up outside of
6570 * the zone but still within the section.
6571 * We have to take care about the node as well. If the node is offline
6572 * its NODE_DATA will be NULL - see page_zone.
6574 if (!node_online(page_to_nid(page)))
6577 zone = page_zone(page);
6578 pfn = page_to_pfn(page);
6579 if (!zone_spans_pfn(zone, pfn))
6582 return !has_unmovable_pages(zone, page, 0, true);
6587 static unsigned long pfn_max_align_down(unsigned long pfn)
6589 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6590 pageblock_nr_pages) - 1);
6593 static unsigned long pfn_max_align_up(unsigned long pfn)
6595 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6596 pageblock_nr_pages));
6599 /* [start, end) must belong to a single zone. */
6600 static int __alloc_contig_migrate_range(struct compact_control *cc,
6601 unsigned long start, unsigned long end)
6603 /* This function is based on compact_zone() from compaction.c. */
6604 unsigned long nr_reclaimed;
6605 unsigned long pfn = start;
6606 unsigned int tries = 0;
6611 while (pfn < end || !list_empty(&cc->migratepages)) {
6612 if (fatal_signal_pending(current)) {
6617 if (list_empty(&cc->migratepages)) {
6618 cc->nr_migratepages = 0;
6619 pfn = isolate_migratepages_range(cc, pfn, end);
6625 } else if (++tries == 5) {
6626 ret = ret < 0 ? ret : -EBUSY;
6630 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6632 cc->nr_migratepages -= nr_reclaimed;
6634 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6635 NULL, 0, cc->mode, MR_CMA);
6638 putback_movable_pages(&cc->migratepages);
6645 * alloc_contig_range() -- tries to allocate given range of pages
6646 * @start: start PFN to allocate
6647 * @end: one-past-the-last PFN to allocate
6648 * @migratetype: migratetype of the underlaying pageblocks (either
6649 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6650 * in range must have the same migratetype and it must
6651 * be either of the two.
6653 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6654 * aligned, however it's the caller's responsibility to guarantee that
6655 * we are the only thread that changes migrate type of pageblocks the
6658 * The PFN range must belong to a single zone.
6660 * Returns zero on success or negative error code. On success all
6661 * pages which PFN is in [start, end) are allocated for the caller and
6662 * need to be freed with free_contig_range().
6664 int alloc_contig_range(unsigned long start, unsigned long end,
6665 unsigned migratetype)
6667 unsigned long outer_start, outer_end;
6671 struct compact_control cc = {
6672 .nr_migratepages = 0,
6674 .zone = page_zone(pfn_to_page(start)),
6675 .mode = MIGRATE_SYNC,
6676 .ignore_skip_hint = true,
6678 INIT_LIST_HEAD(&cc.migratepages);
6681 * What we do here is we mark all pageblocks in range as
6682 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6683 * have different sizes, and due to the way page allocator
6684 * work, we align the range to biggest of the two pages so
6685 * that page allocator won't try to merge buddies from
6686 * different pageblocks and change MIGRATE_ISOLATE to some
6687 * other migration type.
6689 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6690 * migrate the pages from an unaligned range (ie. pages that
6691 * we are interested in). This will put all the pages in
6692 * range back to page allocator as MIGRATE_ISOLATE.
6694 * When this is done, we take the pages in range from page
6695 * allocator removing them from the buddy system. This way
6696 * page allocator will never consider using them.
6698 * This lets us mark the pageblocks back as
6699 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6700 * aligned range but not in the unaligned, original range are
6701 * put back to page allocator so that buddy can use them.
6704 ret = start_isolate_page_range(pfn_max_align_down(start),
6705 pfn_max_align_up(end), migratetype,
6711 * In case of -EBUSY, we'd like to know which page causes problem.
6712 * So, just fall through. We will check it in test_pages_isolated().
6714 ret = __alloc_contig_migrate_range(&cc, start, end);
6715 if (ret && ret != -EBUSY)
6719 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6720 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6721 * more, all pages in [start, end) are free in page allocator.
6722 * What we are going to do is to allocate all pages from
6723 * [start, end) (that is remove them from page allocator).
6725 * The only problem is that pages at the beginning and at the
6726 * end of interesting range may be not aligned with pages that
6727 * page allocator holds, ie. they can be part of higher order
6728 * pages. Because of this, we reserve the bigger range and
6729 * once this is done free the pages we are not interested in.
6731 * We don't have to hold zone->lock here because the pages are
6732 * isolated thus they won't get removed from buddy.
6735 lru_add_drain_all();
6736 drain_all_pages(cc.zone);
6739 outer_start = start;
6740 while (!PageBuddy(pfn_to_page(outer_start))) {
6741 if (++order >= MAX_ORDER) {
6742 outer_start = start;
6745 outer_start &= ~0UL << order;
6748 if (outer_start != start) {
6749 order = page_order(pfn_to_page(outer_start));
6752 * outer_start page could be small order buddy page and
6753 * it doesn't include start page. Adjust outer_start
6754 * in this case to report failed page properly
6755 * on tracepoint in test_pages_isolated()
6757 if (outer_start + (1UL << order) <= start)
6758 outer_start = start;
6761 /* Make sure the range is really isolated. */
6762 if (test_pages_isolated(outer_start, end, false)) {
6763 pr_info("%s: [%lx, %lx) PFNs busy\n",
6764 __func__, outer_start, end);
6769 /* Grab isolated pages from freelists. */
6770 outer_end = isolate_freepages_range(&cc, outer_start, end);
6776 /* Free head and tail (if any) */
6777 if (start != outer_start)
6778 free_contig_range(outer_start, start - outer_start);
6779 if (end != outer_end)
6780 free_contig_range(end, outer_end - end);
6783 undo_isolate_page_range(pfn_max_align_down(start),
6784 pfn_max_align_up(end), migratetype);
6788 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6790 unsigned int count = 0;
6792 for (; nr_pages--; pfn++) {
6793 struct page *page = pfn_to_page(pfn);
6795 count += page_count(page) != 1;
6798 WARN(count != 0, "%d pages are still in use!\n", count);
6802 #ifdef CONFIG_MEMORY_HOTPLUG
6804 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6805 * page high values need to be recalulated.
6807 void __meminit zone_pcp_update(struct zone *zone)
6810 mutex_lock(&pcp_batch_high_lock);
6811 for_each_possible_cpu(cpu)
6812 pageset_set_high_and_batch(zone,
6813 per_cpu_ptr(zone->pageset, cpu));
6814 mutex_unlock(&pcp_batch_high_lock);
6818 void zone_pcp_reset(struct zone *zone)
6820 unsigned long flags;
6822 struct per_cpu_pageset *pset;
6824 /* avoid races with drain_pages() */
6825 local_irq_save(flags);
6826 if (zone->pageset != &boot_pageset) {
6827 for_each_online_cpu(cpu) {
6828 pset = per_cpu_ptr(zone->pageset, cpu);
6829 drain_zonestat(zone, pset);
6831 free_percpu(zone->pageset);
6832 zone->pageset = &boot_pageset;
6834 local_irq_restore(flags);
6837 #ifdef CONFIG_MEMORY_HOTREMOVE
6839 * All pages in the range must be isolated before calling this.
6842 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6846 unsigned int order, i;
6848 unsigned long flags;
6849 /* find the first valid pfn */
6850 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6855 zone = page_zone(pfn_to_page(pfn));
6856 spin_lock_irqsave(&zone->lock, flags);
6858 while (pfn < end_pfn) {
6859 if (!pfn_valid(pfn)) {
6863 page = pfn_to_page(pfn);
6865 * The HWPoisoned page may be not in buddy system, and
6866 * page_count() is not 0.
6868 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6870 SetPageReserved(page);
6874 BUG_ON(page_count(page));
6875 BUG_ON(!PageBuddy(page));
6876 order = page_order(page);
6877 #ifdef CONFIG_DEBUG_VM
6878 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6879 pfn, 1 << order, end_pfn);
6881 list_del(&page->lru);
6882 rmv_page_order(page);
6883 zone->free_area[order].nr_free--;
6884 for (i = 0; i < (1 << order); i++)
6885 SetPageReserved((page+i));
6886 pfn += (1 << order);
6888 spin_unlock_irqrestore(&zone->lock, flags);
6892 #ifdef CONFIG_MEMORY_FAILURE
6893 bool is_free_buddy_page(struct page *page)
6895 struct zone *zone = page_zone(page);
6896 unsigned long pfn = page_to_pfn(page);
6897 unsigned long flags;
6900 spin_lock_irqsave(&zone->lock, flags);
6901 for (order = 0; order < MAX_ORDER; order++) {
6902 struct page *page_head = page - (pfn & ((1 << order) - 1));
6904 if (PageBuddy(page_head) && page_order(page_head) >= order)
6907 spin_unlock_irqrestore(&zone->lock, flags);
6909 return order < MAX_ORDER;