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 * When calculating the number of globally allowed dirty pages, there
119 * is a certain number of per-zone reserves that should not be
120 * considered dirtyable memory. This is the sum of those reserves
121 * over all existing zones that contribute dirtyable memory.
123 unsigned long dirty_balance_reserve __read_mostly;
125 int percpu_pagelist_fraction;
126 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
129 * A cached value of the page's pageblock's migratetype, used when the page is
130 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
131 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
132 * Also the migratetype set in the page does not necessarily match the pcplist
133 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
134 * other index - this ensures that it will be put on the correct CMA freelist.
136 static inline int get_pcppage_migratetype(struct page *page)
141 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
143 page->index = migratetype;
146 #ifdef CONFIG_PM_SLEEP
148 * The following functions are used by the suspend/hibernate code to temporarily
149 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
150 * while devices are suspended. To avoid races with the suspend/hibernate code,
151 * they should always be called with pm_mutex held (gfp_allowed_mask also should
152 * only be modified with pm_mutex held, unless the suspend/hibernate code is
153 * guaranteed not to run in parallel with that modification).
156 static gfp_t saved_gfp_mask;
158 void pm_restore_gfp_mask(void)
160 WARN_ON(!mutex_is_locked(&pm_mutex));
161 if (saved_gfp_mask) {
162 gfp_allowed_mask = saved_gfp_mask;
167 void pm_restrict_gfp_mask(void)
169 WARN_ON(!mutex_is_locked(&pm_mutex));
170 WARN_ON(saved_gfp_mask);
171 saved_gfp_mask = gfp_allowed_mask;
172 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
175 bool pm_suspended_storage(void)
177 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
181 #endif /* CONFIG_PM_SLEEP */
183 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
184 unsigned int pageblock_order __read_mostly;
187 static void __free_pages_ok(struct page *page, unsigned int order);
190 * results with 256, 32 in the lowmem_reserve sysctl:
191 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
192 * 1G machine -> (16M dma, 784M normal, 224M high)
193 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
194 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
195 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
197 * TBD: should special case ZONE_DMA32 machines here - in those we normally
198 * don't need any ZONE_NORMAL reservation
200 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
201 #ifdef CONFIG_ZONE_DMA
204 #ifdef CONFIG_ZONE_DMA32
207 #ifdef CONFIG_HIGHMEM
213 EXPORT_SYMBOL(totalram_pages);
215 static char * const zone_names[MAX_NR_ZONES] = {
216 #ifdef CONFIG_ZONE_DMA
219 #ifdef CONFIG_ZONE_DMA32
223 #ifdef CONFIG_HIGHMEM
227 #ifdef CONFIG_ZONE_DEVICE
232 static void free_compound_page(struct page *page);
233 compound_page_dtor * const compound_page_dtors[] = {
236 #ifdef CONFIG_HUGETLB_PAGE
241 int min_free_kbytes = 1024;
242 int user_min_free_kbytes = -1;
244 static unsigned long __meminitdata nr_kernel_pages;
245 static unsigned long __meminitdata nr_all_pages;
246 static unsigned long __meminitdata dma_reserve;
248 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
249 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
250 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
251 static unsigned long __initdata required_kernelcore;
252 static unsigned long __initdata required_movablecore;
253 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
255 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
257 EXPORT_SYMBOL(movable_zone);
258 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
261 int nr_node_ids __read_mostly = MAX_NUMNODES;
262 int nr_online_nodes __read_mostly = 1;
263 EXPORT_SYMBOL(nr_node_ids);
264 EXPORT_SYMBOL(nr_online_nodes);
267 int page_group_by_mobility_disabled __read_mostly;
269 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
270 static inline void reset_deferred_meminit(pg_data_t *pgdat)
272 pgdat->first_deferred_pfn = ULONG_MAX;
275 /* Returns true if the struct page for the pfn is uninitialised */
276 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
278 if (pfn >= NODE_DATA(early_pfn_to_nid(pfn))->first_deferred_pfn)
284 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
286 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
293 * Returns false when the remaining initialisation should be deferred until
294 * later in the boot cycle when it can be parallelised.
296 static inline bool update_defer_init(pg_data_t *pgdat,
297 unsigned long pfn, unsigned long zone_end,
298 unsigned long *nr_initialised)
300 /* Always populate low zones for address-contrained allocations */
301 if (zone_end < pgdat_end_pfn(pgdat))
304 /* Initialise at least 2G of the highest zone */
306 if (*nr_initialised > (2UL << (30 - PAGE_SHIFT)) &&
307 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
308 pgdat->first_deferred_pfn = pfn;
315 static inline void reset_deferred_meminit(pg_data_t *pgdat)
319 static inline bool early_page_uninitialised(unsigned long pfn)
324 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
329 static inline bool update_defer_init(pg_data_t *pgdat,
330 unsigned long pfn, unsigned long zone_end,
331 unsigned long *nr_initialised)
338 void set_pageblock_migratetype(struct page *page, int migratetype)
340 if (unlikely(page_group_by_mobility_disabled &&
341 migratetype < MIGRATE_PCPTYPES))
342 migratetype = MIGRATE_UNMOVABLE;
344 set_pageblock_flags_group(page, (unsigned long)migratetype,
345 PB_migrate, PB_migrate_end);
348 #ifdef CONFIG_DEBUG_VM
349 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
353 unsigned long pfn = page_to_pfn(page);
354 unsigned long sp, start_pfn;
357 seq = zone_span_seqbegin(zone);
358 start_pfn = zone->zone_start_pfn;
359 sp = zone->spanned_pages;
360 if (!zone_spans_pfn(zone, pfn))
362 } while (zone_span_seqretry(zone, seq));
365 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
366 pfn, zone_to_nid(zone), zone->name,
367 start_pfn, start_pfn + sp);
372 static int page_is_consistent(struct zone *zone, struct page *page)
374 if (!pfn_valid_within(page_to_pfn(page)))
376 if (zone != page_zone(page))
382 * Temporary debugging check for pages not lying within a given zone.
384 static int bad_range(struct zone *zone, struct page *page)
386 if (page_outside_zone_boundaries(zone, page))
388 if (!page_is_consistent(zone, page))
394 static inline int bad_range(struct zone *zone, struct page *page)
400 static void bad_page(struct page *page, const char *reason,
401 unsigned long bad_flags)
403 static unsigned long resume;
404 static unsigned long nr_shown;
405 static unsigned long nr_unshown;
407 /* Don't complain about poisoned pages */
408 if (PageHWPoison(page)) {
409 page_mapcount_reset(page); /* remove PageBuddy */
414 * Allow a burst of 60 reports, then keep quiet for that minute;
415 * or allow a steady drip of one report per second.
417 if (nr_shown == 60) {
418 if (time_before(jiffies, resume)) {
424 "BUG: Bad page state: %lu messages suppressed\n",
431 resume = jiffies + 60 * HZ;
433 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
434 current->comm, page_to_pfn(page));
435 dump_page_badflags(page, reason, bad_flags);
440 /* Leave bad fields for debug, except PageBuddy could make trouble */
441 page_mapcount_reset(page); /* remove PageBuddy */
442 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
446 * Higher-order pages are called "compound pages". They are structured thusly:
448 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
450 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
451 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
453 * The first tail page's ->compound_dtor holds the offset in array of compound
454 * page destructors. See compound_page_dtors.
456 * The first tail page's ->compound_order holds the order of allocation.
457 * This usage means that zero-order pages may not be compound.
460 static void free_compound_page(struct page *page)
462 __free_pages_ok(page, compound_order(page));
465 void prep_compound_page(struct page *page, unsigned int order)
468 int nr_pages = 1 << order;
470 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
471 set_compound_order(page, order);
473 for (i = 1; i < nr_pages; i++) {
474 struct page *p = page + i;
475 set_page_count(p, 0);
476 set_compound_head(p, page);
480 #ifdef CONFIG_DEBUG_PAGEALLOC
481 unsigned int _debug_guardpage_minorder;
482 bool _debug_pagealloc_enabled __read_mostly;
483 bool _debug_guardpage_enabled __read_mostly;
485 static int __init early_debug_pagealloc(char *buf)
490 if (strcmp(buf, "on") == 0)
491 _debug_pagealloc_enabled = true;
495 early_param("debug_pagealloc", early_debug_pagealloc);
497 static bool need_debug_guardpage(void)
499 /* If we don't use debug_pagealloc, we don't need guard page */
500 if (!debug_pagealloc_enabled())
506 static void init_debug_guardpage(void)
508 if (!debug_pagealloc_enabled())
511 _debug_guardpage_enabled = true;
514 struct page_ext_operations debug_guardpage_ops = {
515 .need = need_debug_guardpage,
516 .init = init_debug_guardpage,
519 static int __init debug_guardpage_minorder_setup(char *buf)
523 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
524 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
527 _debug_guardpage_minorder = res;
528 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
531 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
533 static inline void set_page_guard(struct zone *zone, struct page *page,
534 unsigned int order, int migratetype)
536 struct page_ext *page_ext;
538 if (!debug_guardpage_enabled())
541 page_ext = lookup_page_ext(page);
542 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
544 INIT_LIST_HEAD(&page->lru);
545 set_page_private(page, order);
546 /* Guard pages are not available for any usage */
547 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
550 static inline void clear_page_guard(struct zone *zone, struct page *page,
551 unsigned int order, int migratetype)
553 struct page_ext *page_ext;
555 if (!debug_guardpage_enabled())
558 page_ext = lookup_page_ext(page);
559 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
561 set_page_private(page, 0);
562 if (!is_migrate_isolate(migratetype))
563 __mod_zone_freepage_state(zone, (1 << order), migratetype);
566 struct page_ext_operations debug_guardpage_ops = { NULL, };
567 static inline void set_page_guard(struct zone *zone, struct page *page,
568 unsigned int order, int migratetype) {}
569 static inline void clear_page_guard(struct zone *zone, struct page *page,
570 unsigned int order, int migratetype) {}
573 static inline void set_page_order(struct page *page, unsigned int order)
575 set_page_private(page, order);
576 __SetPageBuddy(page);
579 static inline void rmv_page_order(struct page *page)
581 __ClearPageBuddy(page);
582 set_page_private(page, 0);
586 * This function checks whether a page is free && is the buddy
587 * we can do coalesce a page and its buddy if
588 * (a) the buddy is not in a hole &&
589 * (b) the buddy is in the buddy system &&
590 * (c) a page and its buddy have the same order &&
591 * (d) a page and its buddy are in the same zone.
593 * For recording whether a page is in the buddy system, we set ->_mapcount
594 * PAGE_BUDDY_MAPCOUNT_VALUE.
595 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
596 * serialized by zone->lock.
598 * For recording page's order, we use page_private(page).
600 static inline int page_is_buddy(struct page *page, struct page *buddy,
603 if (!pfn_valid_within(page_to_pfn(buddy)))
606 if (page_is_guard(buddy) && page_order(buddy) == order) {
607 if (page_zone_id(page) != page_zone_id(buddy))
610 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
615 if (PageBuddy(buddy) && page_order(buddy) == order) {
617 * zone check is done late to avoid uselessly
618 * calculating zone/node ids for pages that could
621 if (page_zone_id(page) != page_zone_id(buddy))
624 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
632 * Freeing function for a buddy system allocator.
634 * The concept of a buddy system is to maintain direct-mapped table
635 * (containing bit values) for memory blocks of various "orders".
636 * The bottom level table contains the map for the smallest allocatable
637 * units of memory (here, pages), and each level above it describes
638 * pairs of units from the levels below, hence, "buddies".
639 * At a high level, all that happens here is marking the table entry
640 * at the bottom level available, and propagating the changes upward
641 * as necessary, plus some accounting needed to play nicely with other
642 * parts of the VM system.
643 * At each level, we keep a list of pages, which are heads of continuous
644 * free pages of length of (1 << order) and marked with _mapcount
645 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
647 * So when we are allocating or freeing one, we can derive the state of the
648 * other. That is, if we allocate a small block, and both were
649 * free, the remainder of the region must be split into blocks.
650 * If a block is freed, and its buddy is also free, then this
651 * triggers coalescing into a block of larger size.
656 static inline void __free_one_page(struct page *page,
658 struct zone *zone, unsigned int order,
661 unsigned long page_idx;
662 unsigned long combined_idx;
663 unsigned long uninitialized_var(buddy_idx);
665 unsigned int max_order;
667 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
669 VM_BUG_ON(!zone_is_initialized(zone));
670 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
672 VM_BUG_ON(migratetype == -1);
673 if (likely(!is_migrate_isolate(migratetype)))
674 __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);
682 while (order < max_order - 1) {
683 buddy_idx = __find_buddy_index(page_idx, order);
684 buddy = page + (buddy_idx - page_idx);
685 if (!page_is_buddy(page, buddy, order))
688 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
689 * merge with it and move up one order.
691 if (page_is_guard(buddy)) {
692 clear_page_guard(zone, buddy, order, migratetype);
694 list_del(&buddy->lru);
695 zone->free_area[order].nr_free--;
696 rmv_page_order(buddy);
698 combined_idx = buddy_idx & page_idx;
699 page = page + (combined_idx - page_idx);
700 page_idx = combined_idx;
703 if (max_order < MAX_ORDER) {
704 /* If we are here, it means order is >= pageblock_order.
705 * We want to prevent merge between freepages on isolate
706 * pageblock and normal pageblock. Without this, pageblock
707 * isolation could cause incorrect freepage or CMA accounting.
709 * We don't want to hit this code for the more frequent
712 if (unlikely(has_isolate_pageblock(zone))) {
715 buddy_idx = __find_buddy_index(page_idx, order);
716 buddy = page + (buddy_idx - page_idx);
717 buddy_mt = get_pageblock_migratetype(buddy);
719 if (migratetype != buddy_mt
720 && (is_migrate_isolate(migratetype) ||
721 is_migrate_isolate(buddy_mt)))
725 goto continue_merging;
729 set_page_order(page, order);
732 * If this is not the largest possible page, check if the buddy
733 * of the next-highest order is free. If it is, it's possible
734 * that pages are being freed that will coalesce soon. In case,
735 * that is happening, add the free page to the tail of the list
736 * so it's less likely to be used soon and more likely to be merged
737 * as a higher order page
739 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
740 struct page *higher_page, *higher_buddy;
741 combined_idx = buddy_idx & page_idx;
742 higher_page = page + (combined_idx - page_idx);
743 buddy_idx = __find_buddy_index(combined_idx, order + 1);
744 higher_buddy = higher_page + (buddy_idx - combined_idx);
745 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
746 list_add_tail(&page->lru,
747 &zone->free_area[order].free_list[migratetype]);
752 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
754 zone->free_area[order].nr_free++;
757 static inline int free_pages_check(struct page *page)
759 const char *bad_reason = NULL;
760 unsigned long bad_flags = 0;
762 if (unlikely(page_mapcount(page)))
763 bad_reason = "nonzero mapcount";
764 if (unlikely(page->mapping != NULL))
765 bad_reason = "non-NULL mapping";
766 if (unlikely(atomic_read(&page->_count) != 0))
767 bad_reason = "nonzero _count";
768 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
769 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
770 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
773 if (unlikely(page->mem_cgroup))
774 bad_reason = "page still charged to cgroup";
776 if (unlikely(bad_reason)) {
777 bad_page(page, bad_reason, bad_flags);
780 page_cpupid_reset_last(page);
781 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
782 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
787 * Frees a number of pages from the PCP lists
788 * Assumes all pages on list are in same zone, and of same order.
789 * count is the number of pages to free.
791 * If the zone was previously in an "all pages pinned" state then look to
792 * see if this freeing clears that state.
794 * And clear the zone's pages_scanned counter, to hold off the "all pages are
795 * pinned" detection logic.
797 static void free_pcppages_bulk(struct zone *zone, int count,
798 struct per_cpu_pages *pcp)
803 unsigned long nr_scanned;
805 spin_lock(&zone->lock);
806 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
808 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
812 struct list_head *list;
815 * Remove pages from lists in a round-robin fashion. A
816 * batch_free count is maintained that is incremented when an
817 * empty list is encountered. This is so more pages are freed
818 * off fuller lists instead of spinning excessively around empty
823 if (++migratetype == MIGRATE_PCPTYPES)
825 list = &pcp->lists[migratetype];
826 } while (list_empty(list));
828 /* This is the only non-empty list. Free them all. */
829 if (batch_free == MIGRATE_PCPTYPES)
830 batch_free = to_free;
833 int mt; /* migratetype of the to-be-freed page */
835 page = list_entry(list->prev, struct page, lru);
836 /* must delete as __free_one_page list manipulates */
837 list_del(&page->lru);
839 mt = get_pcppage_migratetype(page);
840 /* MIGRATE_ISOLATE page should not go to pcplists */
841 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
842 /* Pageblock could have been isolated meanwhile */
843 if (unlikely(has_isolate_pageblock(zone)))
844 mt = get_pageblock_migratetype(page);
846 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
847 trace_mm_page_pcpu_drain(page, 0, mt);
848 } while (--to_free && --batch_free && !list_empty(list));
850 spin_unlock(&zone->lock);
853 static void free_one_page(struct zone *zone,
854 struct page *page, unsigned long pfn,
858 unsigned long nr_scanned;
859 spin_lock(&zone->lock);
860 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
862 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
864 if (unlikely(has_isolate_pageblock(zone) ||
865 is_migrate_isolate(migratetype))) {
866 migratetype = get_pfnblock_migratetype(page, pfn);
868 __free_one_page(page, pfn, zone, order, migratetype);
869 spin_unlock(&zone->lock);
872 static int free_tail_pages_check(struct page *head_page, struct page *page)
877 * We rely page->lru.next never has bit 0 set, unless the page
878 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
880 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
882 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
886 if (unlikely(!PageTail(page))) {
887 bad_page(page, "PageTail not set", 0);
890 if (unlikely(compound_head(page) != head_page)) {
891 bad_page(page, "compound_head not consistent", 0);
896 clear_compound_head(page);
900 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
901 unsigned long zone, int nid)
903 set_page_links(page, zone, nid, pfn);
904 init_page_count(page);
905 page_mapcount_reset(page);
906 page_cpupid_reset_last(page);
908 INIT_LIST_HEAD(&page->lru);
909 #ifdef WANT_PAGE_VIRTUAL
910 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
911 if (!is_highmem_idx(zone))
912 set_page_address(page, __va(pfn << PAGE_SHIFT));
916 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
919 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
922 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
923 static void init_reserved_page(unsigned long pfn)
928 if (!early_page_uninitialised(pfn))
931 nid = early_pfn_to_nid(pfn);
932 pgdat = NODE_DATA(nid);
934 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
935 struct zone *zone = &pgdat->node_zones[zid];
937 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
940 __init_single_pfn(pfn, zid, nid);
943 static inline void init_reserved_page(unsigned long pfn)
946 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
949 * Initialised pages do not have PageReserved set. This function is
950 * called for each range allocated by the bootmem allocator and
951 * marks the pages PageReserved. The remaining valid pages are later
952 * sent to the buddy page allocator.
954 void __meminit reserve_bootmem_region(unsigned long start, unsigned long end)
956 unsigned long start_pfn = PFN_DOWN(start);
957 unsigned long end_pfn = PFN_UP(end);
959 for (; start_pfn < end_pfn; start_pfn++) {
960 if (pfn_valid(start_pfn)) {
961 struct page *page = pfn_to_page(start_pfn);
963 init_reserved_page(start_pfn);
965 /* Avoid false-positive PageTail() */
966 INIT_LIST_HEAD(&page->lru);
968 SetPageReserved(page);
973 static bool free_pages_prepare(struct page *page, unsigned int order)
975 bool compound = PageCompound(page);
978 VM_BUG_ON_PAGE(PageTail(page), page);
979 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
981 trace_mm_page_free(page, order);
982 kmemcheck_free_shadow(page, order);
983 kasan_free_pages(page, order);
986 page->mapping = NULL;
987 bad += free_pages_check(page);
988 for (i = 1; i < (1 << order); i++) {
990 bad += free_tail_pages_check(page, page + i);
991 bad += free_pages_check(page + i);
996 reset_page_owner(page, order);
998 if (!PageHighMem(page)) {
999 debug_check_no_locks_freed(page_address(page),
1000 PAGE_SIZE << order);
1001 debug_check_no_obj_freed(page_address(page),
1002 PAGE_SIZE << order);
1004 arch_free_page(page, order);
1005 kernel_map_pages(page, 1 << order, 0);
1010 static void __free_pages_ok(struct page *page, unsigned int order)
1012 unsigned long flags;
1014 unsigned long pfn = page_to_pfn(page);
1016 if (!free_pages_prepare(page, order))
1019 migratetype = get_pfnblock_migratetype(page, pfn);
1020 local_irq_save(flags);
1021 __count_vm_events(PGFREE, 1 << order);
1022 free_one_page(page_zone(page), page, pfn, order, migratetype);
1023 local_irq_restore(flags);
1026 static void __init __free_pages_boot_core(struct page *page,
1027 unsigned long pfn, unsigned int order)
1029 unsigned int nr_pages = 1 << order;
1030 struct page *p = page;
1034 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1036 __ClearPageReserved(p);
1037 set_page_count(p, 0);
1039 __ClearPageReserved(p);
1040 set_page_count(p, 0);
1042 page_zone(page)->managed_pages += nr_pages;
1043 set_page_refcounted(page);
1044 __free_pages(page, order);
1047 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1048 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1050 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1052 int __meminit early_pfn_to_nid(unsigned long pfn)
1054 static DEFINE_SPINLOCK(early_pfn_lock);
1057 spin_lock(&early_pfn_lock);
1058 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1061 spin_unlock(&early_pfn_lock);
1067 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1068 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1069 struct mminit_pfnnid_cache *state)
1073 nid = __early_pfn_to_nid(pfn, state);
1074 if (nid >= 0 && nid != node)
1079 /* Only safe to use early in boot when initialisation is single-threaded */
1080 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1082 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1087 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1091 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1092 struct mminit_pfnnid_cache *state)
1099 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1102 if (early_page_uninitialised(pfn))
1104 return __free_pages_boot_core(page, pfn, order);
1107 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1108 static void __init deferred_free_range(struct page *page,
1109 unsigned long pfn, int nr_pages)
1116 /* Free a large naturally-aligned chunk if possible */
1117 if (nr_pages == MAX_ORDER_NR_PAGES &&
1118 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1119 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1120 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1124 for (i = 0; i < nr_pages; i++, page++, pfn++)
1125 __free_pages_boot_core(page, pfn, 0);
1128 /* Completion tracking for deferred_init_memmap() threads */
1129 static atomic_t pgdat_init_n_undone __initdata;
1130 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1132 static inline void __init pgdat_init_report_one_done(void)
1134 if (atomic_dec_and_test(&pgdat_init_n_undone))
1135 complete(&pgdat_init_all_done_comp);
1138 /* Initialise remaining memory on a node */
1139 static int __init deferred_init_memmap(void *data)
1141 pg_data_t *pgdat = data;
1142 int nid = pgdat->node_id;
1143 struct mminit_pfnnid_cache nid_init_state = { };
1144 unsigned long start = jiffies;
1145 unsigned long nr_pages = 0;
1146 unsigned long walk_start, walk_end;
1149 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1150 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1152 if (first_init_pfn == ULONG_MAX) {
1153 pgdat_init_report_one_done();
1157 /* Bind memory initialisation thread to a local node if possible */
1158 if (!cpumask_empty(cpumask))
1159 set_cpus_allowed_ptr(current, cpumask);
1161 /* Sanity check boundaries */
1162 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1163 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1164 pgdat->first_deferred_pfn = ULONG_MAX;
1166 /* Only the highest zone is deferred so find it */
1167 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1168 zone = pgdat->node_zones + zid;
1169 if (first_init_pfn < zone_end_pfn(zone))
1173 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1174 unsigned long pfn, end_pfn;
1175 struct page *page = NULL;
1176 struct page *free_base_page = NULL;
1177 unsigned long free_base_pfn = 0;
1180 end_pfn = min(walk_end, zone_end_pfn(zone));
1181 pfn = first_init_pfn;
1182 if (pfn < walk_start)
1184 if (pfn < zone->zone_start_pfn)
1185 pfn = zone->zone_start_pfn;
1187 for (; pfn < end_pfn; pfn++) {
1188 if (!pfn_valid_within(pfn))
1192 * Ensure pfn_valid is checked every
1193 * MAX_ORDER_NR_PAGES for memory holes
1195 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1196 if (!pfn_valid(pfn)) {
1202 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1207 /* Minimise pfn page lookups and scheduler checks */
1208 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1211 nr_pages += nr_to_free;
1212 deferred_free_range(free_base_page,
1213 free_base_pfn, nr_to_free);
1214 free_base_page = NULL;
1215 free_base_pfn = nr_to_free = 0;
1217 page = pfn_to_page(pfn);
1222 VM_BUG_ON(page_zone(page) != zone);
1226 __init_single_page(page, pfn, zid, nid);
1227 if (!free_base_page) {
1228 free_base_page = page;
1229 free_base_pfn = pfn;
1234 /* Where possible, batch up pages for a single free */
1237 /* Free the current block of pages to allocator */
1238 nr_pages += nr_to_free;
1239 deferred_free_range(free_base_page, free_base_pfn,
1241 free_base_page = NULL;
1242 free_base_pfn = nr_to_free = 0;
1245 first_init_pfn = max(end_pfn, first_init_pfn);
1248 /* Sanity check that the next zone really is unpopulated */
1249 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1251 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1252 jiffies_to_msecs(jiffies - start));
1254 pgdat_init_report_one_done();
1258 void __init page_alloc_init_late(void)
1262 /* There will be num_node_state(N_MEMORY) threads */
1263 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1264 for_each_node_state(nid, N_MEMORY) {
1265 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1268 /* Block until all are initialised */
1269 wait_for_completion(&pgdat_init_all_done_comp);
1271 /* Reinit limits that are based on free pages after the kernel is up */
1272 files_maxfiles_init();
1274 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1277 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1278 void __init init_cma_reserved_pageblock(struct page *page)
1280 unsigned i = pageblock_nr_pages;
1281 struct page *p = page;
1284 __ClearPageReserved(p);
1285 set_page_count(p, 0);
1288 set_pageblock_migratetype(page, MIGRATE_CMA);
1290 if (pageblock_order >= MAX_ORDER) {
1291 i = pageblock_nr_pages;
1294 set_page_refcounted(p);
1295 __free_pages(p, MAX_ORDER - 1);
1296 p += MAX_ORDER_NR_PAGES;
1297 } while (i -= MAX_ORDER_NR_PAGES);
1299 set_page_refcounted(page);
1300 __free_pages(page, pageblock_order);
1303 adjust_managed_page_count(page, pageblock_nr_pages);
1308 * The order of subdivision here is critical for the IO subsystem.
1309 * Please do not alter this order without good reasons and regression
1310 * testing. Specifically, as large blocks of memory are subdivided,
1311 * the order in which smaller blocks are delivered depends on the order
1312 * they're subdivided in this function. This is the primary factor
1313 * influencing the order in which pages are delivered to the IO
1314 * subsystem according to empirical testing, and this is also justified
1315 * by considering the behavior of a buddy system containing a single
1316 * large block of memory acted on by a series of small allocations.
1317 * This behavior is a critical factor in sglist merging's success.
1321 static inline void expand(struct zone *zone, struct page *page,
1322 int low, int high, struct free_area *area,
1325 unsigned long size = 1 << high;
1327 while (high > low) {
1331 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1333 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1334 debug_guardpage_enabled() &&
1335 high < debug_guardpage_minorder()) {
1337 * Mark as guard pages (or page), that will allow to
1338 * merge back to allocator when buddy will be freed.
1339 * Corresponding page table entries will not be touched,
1340 * pages will stay not present in virtual address space
1342 set_page_guard(zone, &page[size], high, migratetype);
1345 list_add(&page[size].lru, &area->free_list[migratetype]);
1347 set_page_order(&page[size], high);
1352 * This page is about to be returned from the page allocator
1354 static inline int check_new_page(struct page *page)
1356 const char *bad_reason = NULL;
1357 unsigned long bad_flags = 0;
1359 if (unlikely(page_mapcount(page)))
1360 bad_reason = "nonzero mapcount";
1361 if (unlikely(page->mapping != NULL))
1362 bad_reason = "non-NULL mapping";
1363 if (unlikely(atomic_read(&page->_count) != 0))
1364 bad_reason = "nonzero _count";
1365 if (unlikely(page->flags & __PG_HWPOISON)) {
1366 bad_reason = "HWPoisoned (hardware-corrupted)";
1367 bad_flags = __PG_HWPOISON;
1369 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1370 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1371 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1374 if (unlikely(page->mem_cgroup))
1375 bad_reason = "page still charged to cgroup";
1377 if (unlikely(bad_reason)) {
1378 bad_page(page, bad_reason, bad_flags);
1384 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1389 for (i = 0; i < (1 << order); i++) {
1390 struct page *p = page + i;
1391 if (unlikely(check_new_page(p)))
1395 set_page_private(page, 0);
1396 set_page_refcounted(page);
1398 arch_alloc_page(page, order);
1399 kernel_map_pages(page, 1 << order, 1);
1400 kasan_alloc_pages(page, order);
1402 if (gfp_flags & __GFP_ZERO)
1403 for (i = 0; i < (1 << order); i++)
1404 clear_highpage(page + i);
1406 if (order && (gfp_flags & __GFP_COMP))
1407 prep_compound_page(page, order);
1409 set_page_owner(page, order, gfp_flags);
1412 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1413 * allocate the page. The expectation is that the caller is taking
1414 * steps that will free more memory. The caller should avoid the page
1415 * being used for !PFMEMALLOC purposes.
1417 if (alloc_flags & ALLOC_NO_WATERMARKS)
1418 set_page_pfmemalloc(page);
1420 clear_page_pfmemalloc(page);
1426 * Go through the free lists for the given migratetype and remove
1427 * the smallest available page from the freelists
1430 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1433 unsigned int current_order;
1434 struct free_area *area;
1437 /* Find a page of the appropriate size in the preferred list */
1438 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1439 area = &(zone->free_area[current_order]);
1440 if (list_empty(&area->free_list[migratetype]))
1443 page = list_entry(area->free_list[migratetype].next,
1445 list_del(&page->lru);
1446 rmv_page_order(page);
1448 expand(zone, page, order, current_order, area, migratetype);
1449 set_pcppage_migratetype(page, migratetype);
1458 * This array describes the order lists are fallen back to when
1459 * the free lists for the desirable migrate type are depleted
1461 static int fallbacks[MIGRATE_TYPES][4] = {
1462 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1463 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1464 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1466 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1468 #ifdef CONFIG_MEMORY_ISOLATION
1469 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1474 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1477 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1480 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1481 unsigned int order) { return NULL; }
1485 * Move the free pages in a range to the free lists of the requested type.
1486 * Note that start_page and end_pages are not aligned on a pageblock
1487 * boundary. If alignment is required, use move_freepages_block()
1489 int move_freepages(struct zone *zone,
1490 struct page *start_page, struct page *end_page,
1495 int pages_moved = 0;
1497 #ifndef CONFIG_HOLES_IN_ZONE
1499 * page_zone is not safe to call in this context when
1500 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1501 * anyway as we check zone boundaries in move_freepages_block().
1502 * Remove at a later date when no bug reports exist related to
1503 * grouping pages by mobility
1505 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1508 for (page = start_page; page <= end_page;) {
1509 /* Make sure we are not inadvertently changing nodes */
1510 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1512 if (!pfn_valid_within(page_to_pfn(page))) {
1517 if (!PageBuddy(page)) {
1522 order = page_order(page);
1523 list_move(&page->lru,
1524 &zone->free_area[order].free_list[migratetype]);
1526 pages_moved += 1 << order;
1532 int move_freepages_block(struct zone *zone, struct page *page,
1535 unsigned long start_pfn, end_pfn;
1536 struct page *start_page, *end_page;
1538 start_pfn = page_to_pfn(page);
1539 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1540 start_page = pfn_to_page(start_pfn);
1541 end_page = start_page + pageblock_nr_pages - 1;
1542 end_pfn = start_pfn + pageblock_nr_pages - 1;
1544 /* Do not cross zone boundaries */
1545 if (!zone_spans_pfn(zone, start_pfn))
1547 if (!zone_spans_pfn(zone, end_pfn))
1550 return move_freepages(zone, start_page, end_page, migratetype);
1553 static void change_pageblock_range(struct page *pageblock_page,
1554 int start_order, int migratetype)
1556 int nr_pageblocks = 1 << (start_order - pageblock_order);
1558 while (nr_pageblocks--) {
1559 set_pageblock_migratetype(pageblock_page, migratetype);
1560 pageblock_page += pageblock_nr_pages;
1565 * When we are falling back to another migratetype during allocation, try to
1566 * steal extra free pages from the same pageblocks to satisfy further
1567 * allocations, instead of polluting multiple pageblocks.
1569 * If we are stealing a relatively large buddy page, it is likely there will
1570 * be more free pages in the pageblock, so try to steal them all. For
1571 * reclaimable and unmovable allocations, we steal regardless of page size,
1572 * as fragmentation caused by those allocations polluting movable pageblocks
1573 * is worse than movable allocations stealing from unmovable and reclaimable
1576 static bool can_steal_fallback(unsigned int order, int start_mt)
1579 * Leaving this order check is intended, although there is
1580 * relaxed order check in next check. The reason is that
1581 * we can actually steal whole pageblock if this condition met,
1582 * but, below check doesn't guarantee it and that is just heuristic
1583 * so could be changed anytime.
1585 if (order >= pageblock_order)
1588 if (order >= pageblock_order / 2 ||
1589 start_mt == MIGRATE_RECLAIMABLE ||
1590 start_mt == MIGRATE_UNMOVABLE ||
1591 page_group_by_mobility_disabled)
1598 * This function implements actual steal behaviour. If order is large enough,
1599 * we can steal whole pageblock. If not, we first move freepages in this
1600 * pageblock and check whether half of pages are moved or not. If half of
1601 * pages are moved, we can change migratetype of pageblock and permanently
1602 * use it's pages as requested migratetype in the future.
1604 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1607 unsigned int current_order = page_order(page);
1610 /* Take ownership for orders >= pageblock_order */
1611 if (current_order >= pageblock_order) {
1612 change_pageblock_range(page, current_order, start_type);
1616 pages = move_freepages_block(zone, page, start_type);
1618 /* Claim the whole block if over half of it is free */
1619 if (pages >= (1 << (pageblock_order-1)) ||
1620 page_group_by_mobility_disabled)
1621 set_pageblock_migratetype(page, start_type);
1625 * Check whether there is a suitable fallback freepage with requested order.
1626 * If only_stealable is true, this function returns fallback_mt only if
1627 * we can steal other freepages all together. This would help to reduce
1628 * fragmentation due to mixed migratetype pages in one pageblock.
1630 int find_suitable_fallback(struct free_area *area, unsigned int order,
1631 int migratetype, bool only_stealable, bool *can_steal)
1636 if (area->nr_free == 0)
1641 fallback_mt = fallbacks[migratetype][i];
1642 if (fallback_mt == MIGRATE_TYPES)
1645 if (list_empty(&area->free_list[fallback_mt]))
1648 if (can_steal_fallback(order, migratetype))
1651 if (!only_stealable)
1662 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1663 * there are no empty page blocks that contain a page with a suitable order
1665 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1666 unsigned int alloc_order)
1669 unsigned long max_managed, flags;
1672 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1673 * Check is race-prone but harmless.
1675 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1676 if (zone->nr_reserved_highatomic >= max_managed)
1679 spin_lock_irqsave(&zone->lock, flags);
1681 /* Recheck the nr_reserved_highatomic limit under the lock */
1682 if (zone->nr_reserved_highatomic >= max_managed)
1686 mt = get_pageblock_migratetype(page);
1687 if (mt != MIGRATE_HIGHATOMIC &&
1688 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1689 zone->nr_reserved_highatomic += pageblock_nr_pages;
1690 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1691 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1695 spin_unlock_irqrestore(&zone->lock, flags);
1699 * Used when an allocation is about to fail under memory pressure. This
1700 * potentially hurts the reliability of high-order allocations when under
1701 * intense memory pressure but failed atomic allocations should be easier
1702 * to recover from than an OOM.
1704 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1706 struct zonelist *zonelist = ac->zonelist;
1707 unsigned long flags;
1713 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1715 /* Preserve at least one pageblock */
1716 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1719 spin_lock_irqsave(&zone->lock, flags);
1720 for (order = 0; order < MAX_ORDER; order++) {
1721 struct free_area *area = &(zone->free_area[order]);
1723 if (list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
1726 page = list_entry(area->free_list[MIGRATE_HIGHATOMIC].next,
1730 * It should never happen but changes to locking could
1731 * inadvertently allow a per-cpu drain to add pages
1732 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1733 * and watch for underflows.
1735 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
1736 zone->nr_reserved_highatomic);
1739 * Convert to ac->migratetype and avoid the normal
1740 * pageblock stealing heuristics. Minimally, the caller
1741 * is doing the work and needs the pages. More
1742 * importantly, if the block was always converted to
1743 * MIGRATE_UNMOVABLE or another type then the number
1744 * of pageblocks that cannot be completely freed
1747 set_pageblock_migratetype(page, ac->migratetype);
1748 move_freepages_block(zone, page, ac->migratetype);
1749 spin_unlock_irqrestore(&zone->lock, flags);
1752 spin_unlock_irqrestore(&zone->lock, flags);
1756 /* Remove an element from the buddy allocator from the fallback list */
1757 static inline struct page *
1758 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1760 struct free_area *area;
1761 unsigned int current_order;
1766 /* Find the largest possible block of pages in the other list */
1767 for (current_order = MAX_ORDER-1;
1768 current_order >= order && current_order <= MAX_ORDER-1;
1770 area = &(zone->free_area[current_order]);
1771 fallback_mt = find_suitable_fallback(area, current_order,
1772 start_migratetype, false, &can_steal);
1773 if (fallback_mt == -1)
1776 page = list_entry(area->free_list[fallback_mt].next,
1779 steal_suitable_fallback(zone, page, start_migratetype);
1781 /* Remove the page from the freelists */
1783 list_del(&page->lru);
1784 rmv_page_order(page);
1786 expand(zone, page, order, current_order, area,
1789 * The pcppage_migratetype may differ from pageblock's
1790 * migratetype depending on the decisions in
1791 * find_suitable_fallback(). This is OK as long as it does not
1792 * differ for MIGRATE_CMA pageblocks. Those can be used as
1793 * fallback only via special __rmqueue_cma_fallback() function
1795 set_pcppage_migratetype(page, start_migratetype);
1797 trace_mm_page_alloc_extfrag(page, order, current_order,
1798 start_migratetype, fallback_mt);
1807 * Do the hard work of removing an element from the buddy allocator.
1808 * Call me with the zone->lock already held.
1810 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1811 int migratetype, gfp_t gfp_flags)
1815 page = __rmqueue_smallest(zone, order, migratetype);
1816 if (unlikely(!page)) {
1817 if (migratetype == MIGRATE_MOVABLE)
1818 page = __rmqueue_cma_fallback(zone, order);
1821 page = __rmqueue_fallback(zone, order, migratetype);
1824 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1829 * Obtain a specified number of elements from the buddy allocator, all under
1830 * a single hold of the lock, for efficiency. Add them to the supplied list.
1831 * Returns the number of new pages which were placed at *list.
1833 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1834 unsigned long count, struct list_head *list,
1835 int migratetype, bool cold)
1839 spin_lock(&zone->lock);
1840 for (i = 0; i < count; ++i) {
1841 struct page *page = __rmqueue(zone, order, migratetype, 0);
1842 if (unlikely(page == NULL))
1846 * Split buddy pages returned by expand() are received here
1847 * in physical page order. The page is added to the callers and
1848 * list and the list head then moves forward. From the callers
1849 * perspective, the linked list is ordered by page number in
1850 * some conditions. This is useful for IO devices that can
1851 * merge IO requests if the physical pages are ordered
1855 list_add(&page->lru, list);
1857 list_add_tail(&page->lru, list);
1859 if (is_migrate_cma(get_pcppage_migratetype(page)))
1860 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1863 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1864 spin_unlock(&zone->lock);
1870 * Called from the vmstat counter updater to drain pagesets of this
1871 * currently executing processor on remote nodes after they have
1874 * Note that this function must be called with the thread pinned to
1875 * a single processor.
1877 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1879 unsigned long flags;
1880 int to_drain, batch;
1882 local_irq_save(flags);
1883 batch = READ_ONCE(pcp->batch);
1884 to_drain = min(pcp->count, batch);
1886 free_pcppages_bulk(zone, to_drain, pcp);
1887 pcp->count -= to_drain;
1889 local_irq_restore(flags);
1894 * Drain pcplists of the indicated processor and zone.
1896 * The processor must either be the current processor and the
1897 * thread pinned to the current processor or a processor that
1900 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1902 unsigned long flags;
1903 struct per_cpu_pageset *pset;
1904 struct per_cpu_pages *pcp;
1906 local_irq_save(flags);
1907 pset = per_cpu_ptr(zone->pageset, cpu);
1911 free_pcppages_bulk(zone, pcp->count, pcp);
1914 local_irq_restore(flags);
1918 * Drain pcplists of all zones on the indicated processor.
1920 * The processor must either be the current processor and the
1921 * thread pinned to the current processor or a processor that
1924 static void drain_pages(unsigned int cpu)
1928 for_each_populated_zone(zone) {
1929 drain_pages_zone(cpu, zone);
1934 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1936 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1937 * the single zone's pages.
1939 void drain_local_pages(struct zone *zone)
1941 int cpu = smp_processor_id();
1944 drain_pages_zone(cpu, zone);
1950 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1952 * When zone parameter is non-NULL, spill just the single zone's pages.
1954 * Note that this code is protected against sending an IPI to an offline
1955 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1956 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1957 * nothing keeps CPUs from showing up after we populated the cpumask and
1958 * before the call to on_each_cpu_mask().
1960 void drain_all_pages(struct zone *zone)
1965 * Allocate in the BSS so we wont require allocation in
1966 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1968 static cpumask_t cpus_with_pcps;
1971 * We don't care about racing with CPU hotplug event
1972 * as offline notification will cause the notified
1973 * cpu to drain that CPU pcps and on_each_cpu_mask
1974 * disables preemption as part of its processing
1976 for_each_online_cpu(cpu) {
1977 struct per_cpu_pageset *pcp;
1979 bool has_pcps = false;
1982 pcp = per_cpu_ptr(zone->pageset, cpu);
1986 for_each_populated_zone(z) {
1987 pcp = per_cpu_ptr(z->pageset, cpu);
1988 if (pcp->pcp.count) {
1996 cpumask_set_cpu(cpu, &cpus_with_pcps);
1998 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2000 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2004 #ifdef CONFIG_HIBERNATION
2006 void mark_free_pages(struct zone *zone)
2008 unsigned long pfn, max_zone_pfn;
2009 unsigned long flags;
2010 unsigned int order, t;
2011 struct list_head *curr;
2013 if (zone_is_empty(zone))
2016 spin_lock_irqsave(&zone->lock, flags);
2018 max_zone_pfn = zone_end_pfn(zone);
2019 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2020 if (pfn_valid(pfn)) {
2021 struct page *page = pfn_to_page(pfn);
2023 if (!swsusp_page_is_forbidden(page))
2024 swsusp_unset_page_free(page);
2027 for_each_migratetype_order(order, t) {
2028 list_for_each(curr, &zone->free_area[order].free_list[t]) {
2031 pfn = page_to_pfn(list_entry(curr, struct page, lru));
2032 for (i = 0; i < (1UL << order); i++)
2033 swsusp_set_page_free(pfn_to_page(pfn + i));
2036 spin_unlock_irqrestore(&zone->lock, flags);
2038 #endif /* CONFIG_PM */
2041 * Free a 0-order page
2042 * cold == true ? free a cold page : free a hot page
2044 void free_hot_cold_page(struct page *page, bool cold)
2046 struct zone *zone = page_zone(page);
2047 struct per_cpu_pages *pcp;
2048 unsigned long flags;
2049 unsigned long pfn = page_to_pfn(page);
2052 if (!free_pages_prepare(page, 0))
2055 migratetype = get_pfnblock_migratetype(page, pfn);
2056 set_pcppage_migratetype(page, migratetype);
2057 local_irq_save(flags);
2058 __count_vm_event(PGFREE);
2061 * We only track unmovable, reclaimable and movable on pcp lists.
2062 * Free ISOLATE pages back to the allocator because they are being
2063 * offlined but treat RESERVE as movable pages so we can get those
2064 * areas back if necessary. Otherwise, we may have to free
2065 * excessively into the page allocator
2067 if (migratetype >= MIGRATE_PCPTYPES) {
2068 if (unlikely(is_migrate_isolate(migratetype))) {
2069 free_one_page(zone, page, pfn, 0, migratetype);
2072 migratetype = MIGRATE_MOVABLE;
2075 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2077 list_add(&page->lru, &pcp->lists[migratetype]);
2079 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2081 if (pcp->count >= pcp->high) {
2082 unsigned long batch = READ_ONCE(pcp->batch);
2083 free_pcppages_bulk(zone, batch, pcp);
2084 pcp->count -= batch;
2088 local_irq_restore(flags);
2092 * Free a list of 0-order pages
2094 void free_hot_cold_page_list(struct list_head *list, bool cold)
2096 struct page *page, *next;
2098 list_for_each_entry_safe(page, next, list, lru) {
2099 trace_mm_page_free_batched(page, cold);
2100 free_hot_cold_page(page, cold);
2105 * split_page takes a non-compound higher-order page, and splits it into
2106 * n (1<<order) sub-pages: page[0..n]
2107 * Each sub-page must be freed individually.
2109 * Note: this is probably too low level an operation for use in drivers.
2110 * Please consult with lkml before using this in your driver.
2112 void split_page(struct page *page, unsigned int order)
2117 VM_BUG_ON_PAGE(PageCompound(page), page);
2118 VM_BUG_ON_PAGE(!page_count(page), page);
2120 #ifdef CONFIG_KMEMCHECK
2122 * Split shadow pages too, because free(page[0]) would
2123 * otherwise free the whole shadow.
2125 if (kmemcheck_page_is_tracked(page))
2126 split_page(virt_to_page(page[0].shadow), order);
2129 gfp_mask = get_page_owner_gfp(page);
2130 set_page_owner(page, 0, gfp_mask);
2131 for (i = 1; i < (1 << order); i++) {
2132 set_page_refcounted(page + i);
2133 set_page_owner(page + i, 0, gfp_mask);
2136 EXPORT_SYMBOL_GPL(split_page);
2138 int __isolate_free_page(struct page *page, unsigned int order)
2140 unsigned long watermark;
2144 BUG_ON(!PageBuddy(page));
2146 zone = page_zone(page);
2147 mt = get_pageblock_migratetype(page);
2149 if (!is_migrate_isolate(mt)) {
2150 /* Obey watermarks as if the page was being allocated */
2151 watermark = low_wmark_pages(zone) + (1 << order);
2152 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2155 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2158 /* Remove page from free list */
2159 list_del(&page->lru);
2160 zone->free_area[order].nr_free--;
2161 rmv_page_order(page);
2163 set_page_owner(page, order, __GFP_MOVABLE);
2165 /* Set the pageblock if the isolated page is at least a pageblock */
2166 if (order >= pageblock_order - 1) {
2167 struct page *endpage = page + (1 << order) - 1;
2168 for (; page < endpage; page += pageblock_nr_pages) {
2169 int mt = get_pageblock_migratetype(page);
2170 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2171 set_pageblock_migratetype(page,
2177 return 1UL << order;
2181 * Similar to split_page except the page is already free. As this is only
2182 * being used for migration, the migratetype of the block also changes.
2183 * As this is called with interrupts disabled, the caller is responsible
2184 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2187 * Note: this is probably too low level an operation for use in drivers.
2188 * Please consult with lkml before using this in your driver.
2190 int split_free_page(struct page *page)
2195 order = page_order(page);
2197 nr_pages = __isolate_free_page(page, order);
2201 /* Split into individual pages */
2202 set_page_refcounted(page);
2203 split_page(page, order);
2208 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2211 struct page *buffered_rmqueue(struct zone *preferred_zone,
2212 struct zone *zone, unsigned int order,
2213 gfp_t gfp_flags, int alloc_flags, int migratetype)
2215 unsigned long flags;
2217 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2219 if (likely(order == 0)) {
2220 struct per_cpu_pages *pcp;
2221 struct list_head *list;
2223 local_irq_save(flags);
2224 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2225 list = &pcp->lists[migratetype];
2226 if (list_empty(list)) {
2227 pcp->count += rmqueue_bulk(zone, 0,
2230 if (unlikely(list_empty(list)))
2235 page = list_entry(list->prev, struct page, lru);
2237 page = list_entry(list->next, struct page, lru);
2239 list_del(&page->lru);
2242 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2244 * __GFP_NOFAIL is not to be used in new code.
2246 * All __GFP_NOFAIL callers should be fixed so that they
2247 * properly detect and handle allocation failures.
2249 * We most definitely don't want callers attempting to
2250 * allocate greater than order-1 page units with
2253 WARN_ON_ONCE(order > 1);
2255 spin_lock_irqsave(&zone->lock, flags);
2258 if (alloc_flags & ALLOC_HARDER) {
2259 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2261 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2264 page = __rmqueue(zone, order, migratetype, gfp_flags);
2265 spin_unlock(&zone->lock);
2268 __mod_zone_freepage_state(zone, -(1 << order),
2269 get_pcppage_migratetype(page));
2272 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2273 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2274 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2275 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2277 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2278 zone_statistics(preferred_zone, zone, gfp_flags);
2279 local_irq_restore(flags);
2281 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2285 local_irq_restore(flags);
2289 #ifdef CONFIG_FAIL_PAGE_ALLOC
2292 struct fault_attr attr;
2294 bool ignore_gfp_highmem;
2295 bool ignore_gfp_reclaim;
2297 } fail_page_alloc = {
2298 .attr = FAULT_ATTR_INITIALIZER,
2299 .ignore_gfp_reclaim = true,
2300 .ignore_gfp_highmem = true,
2304 static int __init setup_fail_page_alloc(char *str)
2306 return setup_fault_attr(&fail_page_alloc.attr, str);
2308 __setup("fail_page_alloc=", setup_fail_page_alloc);
2310 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2312 if (order < fail_page_alloc.min_order)
2314 if (gfp_mask & __GFP_NOFAIL)
2316 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2318 if (fail_page_alloc.ignore_gfp_reclaim &&
2319 (gfp_mask & __GFP_DIRECT_RECLAIM))
2322 return should_fail(&fail_page_alloc.attr, 1 << order);
2325 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2327 static int __init fail_page_alloc_debugfs(void)
2329 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2332 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2333 &fail_page_alloc.attr);
2335 return PTR_ERR(dir);
2337 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2338 &fail_page_alloc.ignore_gfp_reclaim))
2340 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2341 &fail_page_alloc.ignore_gfp_highmem))
2343 if (!debugfs_create_u32("min-order", mode, dir,
2344 &fail_page_alloc.min_order))
2349 debugfs_remove_recursive(dir);
2354 late_initcall(fail_page_alloc_debugfs);
2356 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2358 #else /* CONFIG_FAIL_PAGE_ALLOC */
2360 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2365 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2368 * Return true if free base pages are above 'mark'. For high-order checks it
2369 * will return true of the order-0 watermark is reached and there is at least
2370 * one free page of a suitable size. Checking now avoids taking the zone lock
2371 * to check in the allocation paths if no pages are free.
2373 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2374 unsigned long mark, int classzone_idx, int alloc_flags,
2379 const int alloc_harder = (alloc_flags & ALLOC_HARDER);
2381 /* free_pages may go negative - that's OK */
2382 free_pages -= (1 << order) - 1;
2384 if (alloc_flags & ALLOC_HIGH)
2388 * If the caller does not have rights to ALLOC_HARDER then subtract
2389 * the high-atomic reserves. This will over-estimate the size of the
2390 * atomic reserve but it avoids a search.
2392 if (likely(!alloc_harder))
2393 free_pages -= z->nr_reserved_highatomic;
2398 /* If allocation can't use CMA areas don't use free CMA pages */
2399 if (!(alloc_flags & ALLOC_CMA))
2400 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2404 * Check watermarks for an order-0 allocation request. If these
2405 * are not met, then a high-order request also cannot go ahead
2406 * even if a suitable page happened to be free.
2408 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2411 /* If this is an order-0 request then the watermark is fine */
2415 /* For a high-order request, check at least one suitable page is free */
2416 for (o = order; o < MAX_ORDER; o++) {
2417 struct free_area *area = &z->free_area[o];
2426 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2427 if (!list_empty(&area->free_list[mt]))
2432 if ((alloc_flags & ALLOC_CMA) &&
2433 !list_empty(&area->free_list[MIGRATE_CMA])) {
2441 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2442 int classzone_idx, int alloc_flags)
2444 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2445 zone_page_state(z, NR_FREE_PAGES));
2448 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2449 unsigned long mark, int classzone_idx)
2451 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2453 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2454 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2456 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2461 static bool zone_local(struct zone *local_zone, struct zone *zone)
2463 return local_zone->node == zone->node;
2466 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2468 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2471 #else /* CONFIG_NUMA */
2472 static bool zone_local(struct zone *local_zone, struct zone *zone)
2477 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2481 #endif /* CONFIG_NUMA */
2483 static void reset_alloc_batches(struct zone *preferred_zone)
2485 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2488 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2489 high_wmark_pages(zone) - low_wmark_pages(zone) -
2490 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2491 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2492 } while (zone++ != preferred_zone);
2496 * get_page_from_freelist goes through the zonelist trying to allocate
2499 static struct page *
2500 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2501 const struct alloc_context *ac)
2503 struct zonelist *zonelist = ac->zonelist;
2505 struct page *page = NULL;
2507 int nr_fair_skipped = 0;
2508 bool zonelist_rescan;
2511 zonelist_rescan = false;
2514 * Scan zonelist, looking for a zone with enough free.
2515 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2517 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2521 if (cpusets_enabled() &&
2522 (alloc_flags & ALLOC_CPUSET) &&
2523 !cpuset_zone_allowed(zone, gfp_mask))
2526 * Distribute pages in proportion to the individual
2527 * zone size to ensure fair page aging. The zone a
2528 * page was allocated in should have no effect on the
2529 * time the page has in memory before being reclaimed.
2531 if (alloc_flags & ALLOC_FAIR) {
2532 if (!zone_local(ac->preferred_zone, zone))
2534 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2540 * When allocating a page cache page for writing, we
2541 * want to get it from a zone that is within its dirty
2542 * limit, such that no single zone holds more than its
2543 * proportional share of globally allowed dirty pages.
2544 * The dirty limits take into account the zone's
2545 * lowmem reserves and high watermark so that kswapd
2546 * should be able to balance it without having to
2547 * write pages from its LRU list.
2549 * This may look like it could increase pressure on
2550 * lower zones by failing allocations in higher zones
2551 * before they are full. But the pages that do spill
2552 * over are limited as the lower zones are protected
2553 * by this very same mechanism. It should not become
2554 * a practical burden to them.
2556 * XXX: For now, allow allocations to potentially
2557 * exceed the per-zone dirty limit in the slowpath
2558 * (spread_dirty_pages unset) before going into reclaim,
2559 * which is important when on a NUMA setup the allowed
2560 * zones are together not big enough to reach the
2561 * global limit. The proper fix for these situations
2562 * will require awareness of zones in the
2563 * dirty-throttling and the flusher threads.
2565 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2568 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2569 if (!zone_watermark_ok(zone, order, mark,
2570 ac->classzone_idx, alloc_flags)) {
2573 /* Checked here to keep the fast path fast */
2574 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2575 if (alloc_flags & ALLOC_NO_WATERMARKS)
2578 if (zone_reclaim_mode == 0 ||
2579 !zone_allows_reclaim(ac->preferred_zone, zone))
2582 ret = zone_reclaim(zone, gfp_mask, order);
2584 case ZONE_RECLAIM_NOSCAN:
2587 case ZONE_RECLAIM_FULL:
2588 /* scanned but unreclaimable */
2591 /* did we reclaim enough */
2592 if (zone_watermark_ok(zone, order, mark,
2593 ac->classzone_idx, alloc_flags))
2601 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2602 gfp_mask, alloc_flags, ac->migratetype);
2604 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2608 * If this is a high-order atomic allocation then check
2609 * if the pageblock should be reserved for the future
2611 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2612 reserve_highatomic_pageblock(page, zone, order);
2619 * The first pass makes sure allocations are spread fairly within the
2620 * local node. However, the local node might have free pages left
2621 * after the fairness batches are exhausted, and remote zones haven't
2622 * even been considered yet. Try once more without fairness, and
2623 * include remote zones now, before entering the slowpath and waking
2624 * kswapd: prefer spilling to a remote zone over swapping locally.
2626 if (alloc_flags & ALLOC_FAIR) {
2627 alloc_flags &= ~ALLOC_FAIR;
2628 if (nr_fair_skipped) {
2629 zonelist_rescan = true;
2630 reset_alloc_batches(ac->preferred_zone);
2632 if (nr_online_nodes > 1)
2633 zonelist_rescan = true;
2636 if (zonelist_rescan)
2643 * Large machines with many possible nodes should not always dump per-node
2644 * meminfo in irq context.
2646 static inline bool should_suppress_show_mem(void)
2651 ret = in_interrupt();
2656 static DEFINE_RATELIMIT_STATE(nopage_rs,
2657 DEFAULT_RATELIMIT_INTERVAL,
2658 DEFAULT_RATELIMIT_BURST);
2660 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
2662 unsigned int filter = SHOW_MEM_FILTER_NODES;
2664 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2665 debug_guardpage_minorder() > 0)
2669 * This documents exceptions given to allocations in certain
2670 * contexts that are allowed to allocate outside current's set
2673 if (!(gfp_mask & __GFP_NOMEMALLOC))
2674 if (test_thread_flag(TIF_MEMDIE) ||
2675 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2676 filter &= ~SHOW_MEM_FILTER_NODES;
2677 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2678 filter &= ~SHOW_MEM_FILTER_NODES;
2681 struct va_format vaf;
2684 va_start(args, fmt);
2689 pr_warn("%pV", &vaf);
2694 pr_warn("%s: page allocation failure: order:%u, mode:0x%x\n",
2695 current->comm, order, gfp_mask);
2698 if (!should_suppress_show_mem())
2702 static inline struct page *
2703 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2704 const struct alloc_context *ac, unsigned long *did_some_progress)
2706 struct oom_control oc = {
2707 .zonelist = ac->zonelist,
2708 .nodemask = ac->nodemask,
2709 .gfp_mask = gfp_mask,
2714 *did_some_progress = 0;
2717 * Acquire the oom lock. If that fails, somebody else is
2718 * making progress for us.
2720 if (!mutex_trylock(&oom_lock)) {
2721 *did_some_progress = 1;
2722 schedule_timeout_uninterruptible(1);
2727 * Go through the zonelist yet one more time, keep very high watermark
2728 * here, this is only to catch a parallel oom killing, we must fail if
2729 * we're still under heavy pressure.
2731 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2732 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2736 if (!(gfp_mask & __GFP_NOFAIL)) {
2737 /* Coredumps can quickly deplete all memory reserves */
2738 if (current->flags & PF_DUMPCORE)
2740 /* The OOM killer will not help higher order allocs */
2741 if (order > PAGE_ALLOC_COSTLY_ORDER)
2743 /* The OOM killer does not needlessly kill tasks for lowmem */
2744 if (ac->high_zoneidx < ZONE_NORMAL)
2746 /* The OOM killer does not compensate for IO-less reclaim */
2747 if (!(gfp_mask & __GFP_FS)) {
2749 * XXX: Page reclaim didn't yield anything,
2750 * and the OOM killer can't be invoked, but
2751 * keep looping as per tradition.
2753 *did_some_progress = 1;
2756 if (pm_suspended_storage())
2758 /* The OOM killer may not free memory on a specific node */
2759 if (gfp_mask & __GFP_THISNODE)
2762 /* Exhausted what can be done so it's blamo time */
2763 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL))
2764 *did_some_progress = 1;
2766 mutex_unlock(&oom_lock);
2770 #ifdef CONFIG_COMPACTION
2771 /* Try memory compaction for high-order allocations before reclaim */
2772 static struct page *
2773 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2774 int alloc_flags, const struct alloc_context *ac,
2775 enum migrate_mode mode, int *contended_compaction,
2776 bool *deferred_compaction)
2778 unsigned long compact_result;
2784 current->flags |= PF_MEMALLOC;
2785 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2786 mode, contended_compaction);
2787 current->flags &= ~PF_MEMALLOC;
2789 switch (compact_result) {
2790 case COMPACT_DEFERRED:
2791 *deferred_compaction = true;
2793 case COMPACT_SKIPPED:
2800 * At least in one zone compaction wasn't deferred or skipped, so let's
2801 * count a compaction stall
2803 count_vm_event(COMPACTSTALL);
2805 page = get_page_from_freelist(gfp_mask, order,
2806 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2809 struct zone *zone = page_zone(page);
2811 zone->compact_blockskip_flush = false;
2812 compaction_defer_reset(zone, order, true);
2813 count_vm_event(COMPACTSUCCESS);
2818 * It's bad if compaction run occurs and fails. The most likely reason
2819 * is that pages exist, but not enough to satisfy watermarks.
2821 count_vm_event(COMPACTFAIL);
2828 static inline struct page *
2829 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2830 int alloc_flags, const struct alloc_context *ac,
2831 enum migrate_mode mode, int *contended_compaction,
2832 bool *deferred_compaction)
2836 #endif /* CONFIG_COMPACTION */
2838 /* Perform direct synchronous page reclaim */
2840 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2841 const struct alloc_context *ac)
2843 struct reclaim_state reclaim_state;
2848 /* We now go into synchronous reclaim */
2849 cpuset_memory_pressure_bump();
2850 current->flags |= PF_MEMALLOC;
2851 lockdep_set_current_reclaim_state(gfp_mask);
2852 reclaim_state.reclaimed_slab = 0;
2853 current->reclaim_state = &reclaim_state;
2855 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2858 current->reclaim_state = NULL;
2859 lockdep_clear_current_reclaim_state();
2860 current->flags &= ~PF_MEMALLOC;
2867 /* The really slow allocator path where we enter direct reclaim */
2868 static inline struct page *
2869 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2870 int alloc_flags, const struct alloc_context *ac,
2871 unsigned long *did_some_progress)
2873 struct page *page = NULL;
2874 bool drained = false;
2876 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2877 if (unlikely(!(*did_some_progress)))
2881 page = get_page_from_freelist(gfp_mask, order,
2882 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2885 * If an allocation failed after direct reclaim, it could be because
2886 * pages are pinned on the per-cpu lists or in high alloc reserves.
2887 * Shrink them them and try again
2889 if (!page && !drained) {
2890 unreserve_highatomic_pageblock(ac);
2891 drain_all_pages(NULL);
2900 * This is called in the allocator slow-path if the allocation request is of
2901 * sufficient urgency to ignore watermarks and take other desperate measures
2903 static inline struct page *
2904 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2905 const struct alloc_context *ac)
2910 page = get_page_from_freelist(gfp_mask, order,
2911 ALLOC_NO_WATERMARKS, ac);
2913 if (!page && gfp_mask & __GFP_NOFAIL)
2914 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC,
2916 } while (!page && (gfp_mask & __GFP_NOFAIL));
2921 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2926 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2927 ac->high_zoneidx, ac->nodemask)
2928 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2932 gfp_to_alloc_flags(gfp_t gfp_mask)
2934 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2936 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2937 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2940 * The caller may dip into page reserves a bit more if the caller
2941 * cannot run direct reclaim, or if the caller has realtime scheduling
2942 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2943 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
2945 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2947 if (gfp_mask & __GFP_ATOMIC) {
2949 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2950 * if it can't schedule.
2952 if (!(gfp_mask & __GFP_NOMEMALLOC))
2953 alloc_flags |= ALLOC_HARDER;
2955 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2956 * comment for __cpuset_node_allowed().
2958 alloc_flags &= ~ALLOC_CPUSET;
2959 } else if (unlikely(rt_task(current)) && !in_interrupt())
2960 alloc_flags |= ALLOC_HARDER;
2962 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2963 if (gfp_mask & __GFP_MEMALLOC)
2964 alloc_flags |= ALLOC_NO_WATERMARKS;
2965 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2966 alloc_flags |= ALLOC_NO_WATERMARKS;
2967 else if (!in_interrupt() &&
2968 ((current->flags & PF_MEMALLOC) ||
2969 unlikely(test_thread_flag(TIF_MEMDIE))))
2970 alloc_flags |= ALLOC_NO_WATERMARKS;
2973 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2974 alloc_flags |= ALLOC_CMA;
2979 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2981 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2984 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
2986 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
2989 static inline struct page *
2990 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2991 struct alloc_context *ac)
2993 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
2994 struct page *page = NULL;
2996 unsigned long pages_reclaimed = 0;
2997 unsigned long did_some_progress;
2998 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2999 bool deferred_compaction = false;
3000 int contended_compaction = COMPACT_CONTENDED_NONE;
3003 * In the slowpath, we sanity check order to avoid ever trying to
3004 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3005 * be using allocators in order of preference for an area that is
3008 if (order >= MAX_ORDER) {
3009 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3014 * We also sanity check to catch abuse of atomic reserves being used by
3015 * callers that are not in atomic context.
3017 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3018 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3019 gfp_mask &= ~__GFP_ATOMIC;
3022 * If this allocation cannot block and it is for a specific node, then
3023 * fail early. There's no need to wakeup kswapd or retry for a
3024 * speculative node-specific allocation.
3026 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !can_direct_reclaim)
3030 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3031 wake_all_kswapds(order, ac);
3034 * OK, we're below the kswapd watermark and have kicked background
3035 * reclaim. Now things get more complex, so set up alloc_flags according
3036 * to how we want to proceed.
3038 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3041 * Find the true preferred zone if the allocation is unconstrained by
3044 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3045 struct zoneref *preferred_zoneref;
3046 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3047 ac->high_zoneidx, NULL, &ac->preferred_zone);
3048 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3051 /* This is the last chance, in general, before the goto nopage. */
3052 page = get_page_from_freelist(gfp_mask, order,
3053 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3057 /* Allocate without watermarks if the context allows */
3058 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3060 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3061 * the allocation is high priority and these type of
3062 * allocations are system rather than user orientated
3064 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3066 page = __alloc_pages_high_priority(gfp_mask, order, ac);
3073 /* Caller is not willing to reclaim, we can't balance anything */
3074 if (!can_direct_reclaim) {
3076 * All existing users of the deprecated __GFP_NOFAIL are
3077 * blockable, so warn of any new users that actually allow this
3078 * type of allocation to fail.
3080 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3084 /* Avoid recursion of direct reclaim */
3085 if (current->flags & PF_MEMALLOC)
3088 /* Avoid allocations with no watermarks from looping endlessly */
3089 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3093 * Try direct compaction. The first pass is asynchronous. Subsequent
3094 * attempts after direct reclaim are synchronous
3096 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3098 &contended_compaction,
3099 &deferred_compaction);
3103 /* Checks for THP-specific high-order allocations */
3104 if (is_thp_gfp_mask(gfp_mask)) {
3106 * If compaction is deferred for high-order allocations, it is
3107 * because sync compaction recently failed. If this is the case
3108 * and the caller requested a THP allocation, we do not want
3109 * to heavily disrupt the system, so we fail the allocation
3110 * instead of entering direct reclaim.
3112 if (deferred_compaction)
3116 * In all zones where compaction was attempted (and not
3117 * deferred or skipped), lock contention has been detected.
3118 * For THP allocation we do not want to disrupt the others
3119 * so we fallback to base pages instead.
3121 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3125 * If compaction was aborted due to need_resched(), we do not
3126 * want to further increase allocation latency, unless it is
3127 * khugepaged trying to collapse.
3129 if (contended_compaction == COMPACT_CONTENDED_SCHED
3130 && !(current->flags & PF_KTHREAD))
3135 * It can become very expensive to allocate transparent hugepages at
3136 * fault, so use asynchronous memory compaction for THP unless it is
3137 * khugepaged trying to collapse.
3139 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3140 migration_mode = MIGRATE_SYNC_LIGHT;
3142 /* Try direct reclaim and then allocating */
3143 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3144 &did_some_progress);
3148 /* Do not loop if specifically requested */
3149 if (gfp_mask & __GFP_NORETRY)
3152 /* Keep reclaiming pages as long as there is reasonable progress */
3153 pages_reclaimed += did_some_progress;
3154 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3155 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3156 /* Wait for some write requests to complete then retry */
3157 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3161 /* Reclaim has failed us, start killing things */
3162 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3166 /* Retry as long as the OOM killer is making progress */
3167 if (did_some_progress)
3172 * High-order allocations do not necessarily loop after
3173 * direct reclaim and reclaim/compaction depends on compaction
3174 * being called after reclaim so call directly if necessary
3176 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3178 &contended_compaction,
3179 &deferred_compaction);
3183 warn_alloc_failed(gfp_mask, order, NULL);
3189 * This is the 'heart' of the zoned buddy allocator.
3192 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3193 struct zonelist *zonelist, nodemask_t *nodemask)
3195 struct zoneref *preferred_zoneref;
3196 struct page *page = NULL;
3197 unsigned int cpuset_mems_cookie;
3198 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3199 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3200 struct alloc_context ac = {
3201 .high_zoneidx = gfp_zone(gfp_mask),
3202 .nodemask = nodemask,
3203 .migratetype = gfpflags_to_migratetype(gfp_mask),
3206 gfp_mask &= gfp_allowed_mask;
3208 lockdep_trace_alloc(gfp_mask);
3210 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3212 if (should_fail_alloc_page(gfp_mask, order))
3216 * Check the zones suitable for the gfp_mask contain at least one
3217 * valid zone. It's possible to have an empty zonelist as a result
3218 * of __GFP_THISNODE and a memoryless node
3220 if (unlikely(!zonelist->_zonerefs->zone))
3223 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3224 alloc_flags |= ALLOC_CMA;
3227 cpuset_mems_cookie = read_mems_allowed_begin();
3229 /* We set it here, as __alloc_pages_slowpath might have changed it */
3230 ac.zonelist = zonelist;
3232 /* Dirty zone balancing only done in the fast path */
3233 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3235 /* The preferred zone is used for statistics later */
3236 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3237 ac.nodemask ? : &cpuset_current_mems_allowed,
3238 &ac.preferred_zone);
3239 if (!ac.preferred_zone)
3241 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3243 /* First allocation attempt */
3244 alloc_mask = gfp_mask|__GFP_HARDWALL;
3245 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3246 if (unlikely(!page)) {
3248 * Runtime PM, block IO and its error handling path
3249 * can deadlock because I/O on the device might not
3252 alloc_mask = memalloc_noio_flags(gfp_mask);
3253 ac.spread_dirty_pages = false;
3255 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3258 if (kmemcheck_enabled && page)
3259 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3261 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3265 * When updating a task's mems_allowed, it is possible to race with
3266 * parallel threads in such a way that an allocation can fail while
3267 * the mask is being updated. If a page allocation is about to fail,
3268 * check if the cpuset changed during allocation and if so, retry.
3270 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3275 EXPORT_SYMBOL(__alloc_pages_nodemask);
3278 * Common helper functions.
3280 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3285 * __get_free_pages() returns a 32-bit address, which cannot represent
3288 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3290 page = alloc_pages(gfp_mask, order);
3293 return (unsigned long) page_address(page);
3295 EXPORT_SYMBOL(__get_free_pages);
3297 unsigned long get_zeroed_page(gfp_t gfp_mask)
3299 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3301 EXPORT_SYMBOL(get_zeroed_page);
3303 void __free_pages(struct page *page, unsigned int order)
3305 if (put_page_testzero(page)) {
3307 free_hot_cold_page(page, false);
3309 __free_pages_ok(page, order);
3313 EXPORT_SYMBOL(__free_pages);
3315 void free_pages(unsigned long addr, unsigned int order)
3318 VM_BUG_ON(!virt_addr_valid((void *)addr));
3319 __free_pages(virt_to_page((void *)addr), order);
3323 EXPORT_SYMBOL(free_pages);
3327 * An arbitrary-length arbitrary-offset area of memory which resides
3328 * within a 0 or higher order page. Multiple fragments within that page
3329 * are individually refcounted, in the page's reference counter.
3331 * The page_frag functions below provide a simple allocation framework for
3332 * page fragments. This is used by the network stack and network device
3333 * drivers to provide a backing region of memory for use as either an
3334 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3336 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3339 struct page *page = NULL;
3340 gfp_t gfp = gfp_mask;
3342 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3343 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3345 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3346 PAGE_FRAG_CACHE_MAX_ORDER);
3347 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3349 if (unlikely(!page))
3350 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3352 nc->va = page ? page_address(page) : NULL;
3357 void *__alloc_page_frag(struct page_frag_cache *nc,
3358 unsigned int fragsz, gfp_t gfp_mask)
3360 unsigned int size = PAGE_SIZE;
3364 if (unlikely(!nc->va)) {
3366 page = __page_frag_refill(nc, gfp_mask);
3370 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3371 /* if size can vary use size else just use PAGE_SIZE */
3374 /* Even if we own the page, we do not use atomic_set().
3375 * This would break get_page_unless_zero() users.
3377 atomic_add(size - 1, &page->_count);
3379 /* reset page count bias and offset to start of new frag */
3380 nc->pfmemalloc = page_is_pfmemalloc(page);
3381 nc->pagecnt_bias = size;
3385 offset = nc->offset - fragsz;
3386 if (unlikely(offset < 0)) {
3387 page = virt_to_page(nc->va);
3389 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3392 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3393 /* if size can vary use size else just use PAGE_SIZE */
3396 /* OK, page count is 0, we can safely set it */
3397 atomic_set(&page->_count, size);
3399 /* reset page count bias and offset to start of new frag */
3400 nc->pagecnt_bias = size;
3401 offset = size - fragsz;
3405 nc->offset = offset;
3407 return nc->va + offset;
3409 EXPORT_SYMBOL(__alloc_page_frag);
3412 * Frees a page fragment allocated out of either a compound or order 0 page.
3414 void __free_page_frag(void *addr)
3416 struct page *page = virt_to_head_page(addr);
3418 if (unlikely(put_page_testzero(page)))
3419 __free_pages_ok(page, compound_order(page));
3421 EXPORT_SYMBOL(__free_page_frag);
3424 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3425 * of the current memory cgroup.
3427 * It should be used when the caller would like to use kmalloc, but since the
3428 * allocation is large, it has to fall back to the page allocator.
3430 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3434 page = alloc_pages(gfp_mask, order);
3435 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3436 __free_pages(page, order);
3442 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3446 page = alloc_pages_node(nid, gfp_mask, order);
3447 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3448 __free_pages(page, order);
3455 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3458 void __free_kmem_pages(struct page *page, unsigned int order)
3460 memcg_kmem_uncharge(page, order);
3461 __free_pages(page, order);
3464 void free_kmem_pages(unsigned long addr, unsigned int order)
3467 VM_BUG_ON(!virt_addr_valid((void *)addr));
3468 __free_kmem_pages(virt_to_page((void *)addr), order);
3472 static void *make_alloc_exact(unsigned long addr, unsigned int order,
3476 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3477 unsigned long used = addr + PAGE_ALIGN(size);
3479 split_page(virt_to_page((void *)addr), order);
3480 while (used < alloc_end) {
3485 return (void *)addr;
3489 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3490 * @size: the number of bytes to allocate
3491 * @gfp_mask: GFP flags for the allocation
3493 * This function is similar to alloc_pages(), except that it allocates the
3494 * minimum number of pages to satisfy the request. alloc_pages() can only
3495 * allocate memory in power-of-two pages.
3497 * This function is also limited by MAX_ORDER.
3499 * Memory allocated by this function must be released by free_pages_exact().
3501 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3503 unsigned int order = get_order(size);
3506 addr = __get_free_pages(gfp_mask, order);
3507 return make_alloc_exact(addr, order, size);
3509 EXPORT_SYMBOL(alloc_pages_exact);
3512 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3514 * @nid: the preferred node ID where memory should be allocated
3515 * @size: the number of bytes to allocate
3516 * @gfp_mask: GFP flags for the allocation
3518 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3521 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3523 unsigned int order = get_order(size);
3524 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3527 return make_alloc_exact((unsigned long)page_address(p), order, size);
3531 * free_pages_exact - release memory allocated via alloc_pages_exact()
3532 * @virt: the value returned by alloc_pages_exact.
3533 * @size: size of allocation, same value as passed to alloc_pages_exact().
3535 * Release the memory allocated by a previous call to alloc_pages_exact.
3537 void free_pages_exact(void *virt, size_t size)
3539 unsigned long addr = (unsigned long)virt;
3540 unsigned long end = addr + PAGE_ALIGN(size);
3542 while (addr < end) {
3547 EXPORT_SYMBOL(free_pages_exact);
3550 * nr_free_zone_pages - count number of pages beyond high watermark
3551 * @offset: The zone index of the highest zone
3553 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3554 * high watermark within all zones at or below a given zone index. For each
3555 * zone, the number of pages is calculated as:
3556 * managed_pages - high_pages
3558 static unsigned long nr_free_zone_pages(int offset)
3563 /* Just pick one node, since fallback list is circular */
3564 unsigned long sum = 0;
3566 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3568 for_each_zone_zonelist(zone, z, zonelist, offset) {
3569 unsigned long size = zone->managed_pages;
3570 unsigned long high = high_wmark_pages(zone);
3579 * nr_free_buffer_pages - count number of pages beyond high watermark
3581 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3582 * watermark within ZONE_DMA and ZONE_NORMAL.
3584 unsigned long nr_free_buffer_pages(void)
3586 return nr_free_zone_pages(gfp_zone(GFP_USER));
3588 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3591 * nr_free_pagecache_pages - count number of pages beyond high watermark
3593 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3594 * high watermark within all zones.
3596 unsigned long nr_free_pagecache_pages(void)
3598 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3601 static inline void show_node(struct zone *zone)
3603 if (IS_ENABLED(CONFIG_NUMA))
3604 printk("Node %d ", zone_to_nid(zone));
3607 void si_meminfo(struct sysinfo *val)
3609 val->totalram = totalram_pages;
3610 val->sharedram = global_page_state(NR_SHMEM);
3611 val->freeram = global_page_state(NR_FREE_PAGES);
3612 val->bufferram = nr_blockdev_pages();
3613 val->totalhigh = totalhigh_pages;
3614 val->freehigh = nr_free_highpages();
3615 val->mem_unit = PAGE_SIZE;
3618 EXPORT_SYMBOL(si_meminfo);
3621 void si_meminfo_node(struct sysinfo *val, int nid)
3623 int zone_type; /* needs to be signed */
3624 unsigned long managed_pages = 0;
3625 pg_data_t *pgdat = NODE_DATA(nid);
3627 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3628 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3629 val->totalram = managed_pages;
3630 val->sharedram = node_page_state(nid, NR_SHMEM);
3631 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3632 #ifdef CONFIG_HIGHMEM
3633 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3634 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3640 val->mem_unit = PAGE_SIZE;
3645 * Determine whether the node should be displayed or not, depending on whether
3646 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3648 bool skip_free_areas_node(unsigned int flags, int nid)
3651 unsigned int cpuset_mems_cookie;
3653 if (!(flags & SHOW_MEM_FILTER_NODES))
3657 cpuset_mems_cookie = read_mems_allowed_begin();
3658 ret = !node_isset(nid, cpuset_current_mems_allowed);
3659 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3664 #define K(x) ((x) << (PAGE_SHIFT-10))
3666 static void show_migration_types(unsigned char type)
3668 static const char types[MIGRATE_TYPES] = {
3669 [MIGRATE_UNMOVABLE] = 'U',
3670 [MIGRATE_MOVABLE] = 'M',
3671 [MIGRATE_RECLAIMABLE] = 'E',
3672 [MIGRATE_HIGHATOMIC] = 'H',
3674 [MIGRATE_CMA] = 'C',
3676 #ifdef CONFIG_MEMORY_ISOLATION
3677 [MIGRATE_ISOLATE] = 'I',
3680 char tmp[MIGRATE_TYPES + 1];
3684 for (i = 0; i < MIGRATE_TYPES; i++) {
3685 if (type & (1 << i))
3690 printk("(%s) ", tmp);
3694 * Show free area list (used inside shift_scroll-lock stuff)
3695 * We also calculate the percentage fragmentation. We do this by counting the
3696 * memory on each free list with the exception of the first item on the list.
3699 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3702 void show_free_areas(unsigned int filter)
3704 unsigned long free_pcp = 0;
3708 for_each_populated_zone(zone) {
3709 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3712 for_each_online_cpu(cpu)
3713 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3716 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3717 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3718 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3719 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3720 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3721 " free:%lu free_pcp:%lu free_cma:%lu\n",
3722 global_page_state(NR_ACTIVE_ANON),
3723 global_page_state(NR_INACTIVE_ANON),
3724 global_page_state(NR_ISOLATED_ANON),
3725 global_page_state(NR_ACTIVE_FILE),
3726 global_page_state(NR_INACTIVE_FILE),
3727 global_page_state(NR_ISOLATED_FILE),
3728 global_page_state(NR_UNEVICTABLE),
3729 global_page_state(NR_FILE_DIRTY),
3730 global_page_state(NR_WRITEBACK),
3731 global_page_state(NR_UNSTABLE_NFS),
3732 global_page_state(NR_SLAB_RECLAIMABLE),
3733 global_page_state(NR_SLAB_UNRECLAIMABLE),
3734 global_page_state(NR_FILE_MAPPED),
3735 global_page_state(NR_SHMEM),
3736 global_page_state(NR_PAGETABLE),
3737 global_page_state(NR_BOUNCE),
3738 global_page_state(NR_FREE_PAGES),
3740 global_page_state(NR_FREE_CMA_PAGES));
3742 for_each_populated_zone(zone) {
3745 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3749 for_each_online_cpu(cpu)
3750 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3758 " active_anon:%lukB"
3759 " inactive_anon:%lukB"
3760 " active_file:%lukB"
3761 " inactive_file:%lukB"
3762 " unevictable:%lukB"
3763 " isolated(anon):%lukB"
3764 " isolated(file):%lukB"
3772 " slab_reclaimable:%lukB"
3773 " slab_unreclaimable:%lukB"
3774 " kernel_stack:%lukB"
3781 " writeback_tmp:%lukB"
3782 " pages_scanned:%lu"
3783 " all_unreclaimable? %s"
3786 K(zone_page_state(zone, NR_FREE_PAGES)),
3787 K(min_wmark_pages(zone)),
3788 K(low_wmark_pages(zone)),
3789 K(high_wmark_pages(zone)),
3790 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3791 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3792 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3793 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3794 K(zone_page_state(zone, NR_UNEVICTABLE)),
3795 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3796 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3797 K(zone->present_pages),
3798 K(zone->managed_pages),
3799 K(zone_page_state(zone, NR_MLOCK)),
3800 K(zone_page_state(zone, NR_FILE_DIRTY)),
3801 K(zone_page_state(zone, NR_WRITEBACK)),
3802 K(zone_page_state(zone, NR_FILE_MAPPED)),
3803 K(zone_page_state(zone, NR_SHMEM)),
3804 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3805 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3806 zone_page_state(zone, NR_KERNEL_STACK) *
3808 K(zone_page_state(zone, NR_PAGETABLE)),
3809 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3810 K(zone_page_state(zone, NR_BOUNCE)),
3812 K(this_cpu_read(zone->pageset->pcp.count)),
3813 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3814 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3815 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3816 (!zone_reclaimable(zone) ? "yes" : "no")
3818 printk("lowmem_reserve[]:");
3819 for (i = 0; i < MAX_NR_ZONES; i++)
3820 printk(" %ld", zone->lowmem_reserve[i]);
3824 for_each_populated_zone(zone) {
3826 unsigned long nr[MAX_ORDER], flags, total = 0;
3827 unsigned char types[MAX_ORDER];
3829 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3832 printk("%s: ", zone->name);
3834 spin_lock_irqsave(&zone->lock, flags);
3835 for (order = 0; order < MAX_ORDER; order++) {
3836 struct free_area *area = &zone->free_area[order];
3839 nr[order] = area->nr_free;
3840 total += nr[order] << order;
3843 for (type = 0; type < MIGRATE_TYPES; type++) {
3844 if (!list_empty(&area->free_list[type]))
3845 types[order] |= 1 << type;
3848 spin_unlock_irqrestore(&zone->lock, flags);
3849 for (order = 0; order < MAX_ORDER; order++) {
3850 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3852 show_migration_types(types[order]);
3854 printk("= %lukB\n", K(total));
3857 hugetlb_show_meminfo();
3859 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3861 show_swap_cache_info();
3864 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3866 zoneref->zone = zone;
3867 zoneref->zone_idx = zone_idx(zone);
3871 * Builds allocation fallback zone lists.
3873 * Add all populated zones of a node to the zonelist.
3875 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3879 enum zone_type zone_type = MAX_NR_ZONES;
3883 zone = pgdat->node_zones + zone_type;
3884 if (populated_zone(zone)) {
3885 zoneref_set_zone(zone,
3886 &zonelist->_zonerefs[nr_zones++]);
3887 check_highest_zone(zone_type);
3889 } while (zone_type);
3897 * 0 = automatic detection of better ordering.
3898 * 1 = order by ([node] distance, -zonetype)
3899 * 2 = order by (-zonetype, [node] distance)
3901 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3902 * the same zonelist. So only NUMA can configure this param.
3904 #define ZONELIST_ORDER_DEFAULT 0
3905 #define ZONELIST_ORDER_NODE 1
3906 #define ZONELIST_ORDER_ZONE 2
3908 /* zonelist order in the kernel.
3909 * set_zonelist_order() will set this to NODE or ZONE.
3911 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3912 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3916 /* The value user specified ....changed by config */
3917 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3918 /* string for sysctl */
3919 #define NUMA_ZONELIST_ORDER_LEN 16
3920 char numa_zonelist_order[16] = "default";
3923 * interface for configure zonelist ordering.
3924 * command line option "numa_zonelist_order"
3925 * = "[dD]efault - default, automatic configuration.
3926 * = "[nN]ode - order by node locality, then by zone within node
3927 * = "[zZ]one - order by zone, then by locality within zone
3930 static int __parse_numa_zonelist_order(char *s)
3932 if (*s == 'd' || *s == 'D') {
3933 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3934 } else if (*s == 'n' || *s == 'N') {
3935 user_zonelist_order = ZONELIST_ORDER_NODE;
3936 } else if (*s == 'z' || *s == 'Z') {
3937 user_zonelist_order = ZONELIST_ORDER_ZONE;
3940 "Ignoring invalid numa_zonelist_order value: "
3947 static __init int setup_numa_zonelist_order(char *s)
3954 ret = __parse_numa_zonelist_order(s);
3956 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3960 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3963 * sysctl handler for numa_zonelist_order
3965 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3966 void __user *buffer, size_t *length,
3969 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3971 static DEFINE_MUTEX(zl_order_mutex);
3973 mutex_lock(&zl_order_mutex);
3975 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3979 strcpy(saved_string, (char *)table->data);
3981 ret = proc_dostring(table, write, buffer, length, ppos);
3985 int oldval = user_zonelist_order;
3987 ret = __parse_numa_zonelist_order((char *)table->data);
3990 * bogus value. restore saved string
3992 strncpy((char *)table->data, saved_string,
3993 NUMA_ZONELIST_ORDER_LEN);
3994 user_zonelist_order = oldval;
3995 } else if (oldval != user_zonelist_order) {
3996 mutex_lock(&zonelists_mutex);
3997 build_all_zonelists(NULL, NULL);
3998 mutex_unlock(&zonelists_mutex);
4002 mutex_unlock(&zl_order_mutex);
4007 #define MAX_NODE_LOAD (nr_online_nodes)
4008 static int node_load[MAX_NUMNODES];
4011 * find_next_best_node - find the next node that should appear in a given node's fallback list
4012 * @node: node whose fallback list we're appending
4013 * @used_node_mask: nodemask_t of already used nodes
4015 * We use a number of factors to determine which is the next node that should
4016 * appear on a given node's fallback list. The node should not have appeared
4017 * already in @node's fallback list, and it should be the next closest node
4018 * according to the distance array (which contains arbitrary distance values
4019 * from each node to each node in the system), and should also prefer nodes
4020 * with no CPUs, since presumably they'll have very little allocation pressure
4021 * on them otherwise.
4022 * It returns -1 if no node is found.
4024 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4027 int min_val = INT_MAX;
4028 int best_node = NUMA_NO_NODE;
4029 const struct cpumask *tmp = cpumask_of_node(0);
4031 /* Use the local node if we haven't already */
4032 if (!node_isset(node, *used_node_mask)) {
4033 node_set(node, *used_node_mask);
4037 for_each_node_state(n, N_MEMORY) {
4039 /* Don't want a node to appear more than once */
4040 if (node_isset(n, *used_node_mask))
4043 /* Use the distance array to find the distance */
4044 val = node_distance(node, n);
4046 /* Penalize nodes under us ("prefer the next node") */
4049 /* Give preference to headless and unused nodes */
4050 tmp = cpumask_of_node(n);
4051 if (!cpumask_empty(tmp))
4052 val += PENALTY_FOR_NODE_WITH_CPUS;
4054 /* Slight preference for less loaded node */
4055 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4056 val += node_load[n];
4058 if (val < min_val) {
4065 node_set(best_node, *used_node_mask);
4072 * Build zonelists ordered by node and zones within node.
4073 * This results in maximum locality--normal zone overflows into local
4074 * DMA zone, if any--but risks exhausting DMA zone.
4076 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4079 struct zonelist *zonelist;
4081 zonelist = &pgdat->node_zonelists[0];
4082 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4084 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4085 zonelist->_zonerefs[j].zone = NULL;
4086 zonelist->_zonerefs[j].zone_idx = 0;
4090 * Build gfp_thisnode zonelists
4092 static void build_thisnode_zonelists(pg_data_t *pgdat)
4095 struct zonelist *zonelist;
4097 zonelist = &pgdat->node_zonelists[1];
4098 j = build_zonelists_node(pgdat, zonelist, 0);
4099 zonelist->_zonerefs[j].zone = NULL;
4100 zonelist->_zonerefs[j].zone_idx = 0;
4104 * Build zonelists ordered by zone and nodes within zones.
4105 * This results in conserving DMA zone[s] until all Normal memory is
4106 * exhausted, but results in overflowing to remote node while memory
4107 * may still exist in local DMA zone.
4109 static int node_order[MAX_NUMNODES];
4111 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4114 int zone_type; /* needs to be signed */
4116 struct zonelist *zonelist;
4118 zonelist = &pgdat->node_zonelists[0];
4120 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4121 for (j = 0; j < nr_nodes; j++) {
4122 node = node_order[j];
4123 z = &NODE_DATA(node)->node_zones[zone_type];
4124 if (populated_zone(z)) {
4126 &zonelist->_zonerefs[pos++]);
4127 check_highest_zone(zone_type);
4131 zonelist->_zonerefs[pos].zone = NULL;
4132 zonelist->_zonerefs[pos].zone_idx = 0;
4135 #if defined(CONFIG_64BIT)
4137 * Devices that require DMA32/DMA are relatively rare and do not justify a
4138 * penalty to every machine in case the specialised case applies. Default
4139 * to Node-ordering on 64-bit NUMA machines
4141 static int default_zonelist_order(void)
4143 return ZONELIST_ORDER_NODE;
4147 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4148 * by the kernel. If processes running on node 0 deplete the low memory zone
4149 * then reclaim will occur more frequency increasing stalls and potentially
4150 * be easier to OOM if a large percentage of the zone is under writeback or
4151 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4152 * Hence, default to zone ordering on 32-bit.
4154 static int default_zonelist_order(void)
4156 return ZONELIST_ORDER_ZONE;
4158 #endif /* CONFIG_64BIT */
4160 static void set_zonelist_order(void)
4162 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4163 current_zonelist_order = default_zonelist_order();
4165 current_zonelist_order = user_zonelist_order;
4168 static void build_zonelists(pg_data_t *pgdat)
4172 nodemask_t used_mask;
4173 int local_node, prev_node;
4174 struct zonelist *zonelist;
4175 unsigned int order = current_zonelist_order;
4177 /* initialize zonelists */
4178 for (i = 0; i < MAX_ZONELISTS; i++) {
4179 zonelist = pgdat->node_zonelists + i;
4180 zonelist->_zonerefs[0].zone = NULL;
4181 zonelist->_zonerefs[0].zone_idx = 0;
4184 /* NUMA-aware ordering of nodes */
4185 local_node = pgdat->node_id;
4186 load = nr_online_nodes;
4187 prev_node = local_node;
4188 nodes_clear(used_mask);
4190 memset(node_order, 0, sizeof(node_order));
4193 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4195 * We don't want to pressure a particular node.
4196 * So adding penalty to the first node in same
4197 * distance group to make it round-robin.
4199 if (node_distance(local_node, node) !=
4200 node_distance(local_node, prev_node))
4201 node_load[node] = load;
4205 if (order == ZONELIST_ORDER_NODE)
4206 build_zonelists_in_node_order(pgdat, node);
4208 node_order[j++] = node; /* remember order */
4211 if (order == ZONELIST_ORDER_ZONE) {
4212 /* calculate node order -- i.e., DMA last! */
4213 build_zonelists_in_zone_order(pgdat, j);
4216 build_thisnode_zonelists(pgdat);
4219 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4221 * Return node id of node used for "local" allocations.
4222 * I.e., first node id of first zone in arg node's generic zonelist.
4223 * Used for initializing percpu 'numa_mem', which is used primarily
4224 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4226 int local_memory_node(int node)
4230 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4231 gfp_zone(GFP_KERNEL),
4238 #else /* CONFIG_NUMA */
4240 static void set_zonelist_order(void)
4242 current_zonelist_order = ZONELIST_ORDER_ZONE;
4245 static void build_zonelists(pg_data_t *pgdat)
4247 int node, local_node;
4249 struct zonelist *zonelist;
4251 local_node = pgdat->node_id;
4253 zonelist = &pgdat->node_zonelists[0];
4254 j = build_zonelists_node(pgdat, zonelist, 0);
4257 * Now we build the zonelist so that it contains the zones
4258 * of all the other nodes.
4259 * We don't want to pressure a particular node, so when
4260 * building the zones for node N, we make sure that the
4261 * zones coming right after the local ones are those from
4262 * node N+1 (modulo N)
4264 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4265 if (!node_online(node))
4267 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4269 for (node = 0; node < local_node; node++) {
4270 if (!node_online(node))
4272 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4275 zonelist->_zonerefs[j].zone = NULL;
4276 zonelist->_zonerefs[j].zone_idx = 0;
4279 #endif /* CONFIG_NUMA */
4282 * Boot pageset table. One per cpu which is going to be used for all
4283 * zones and all nodes. The parameters will be set in such a way
4284 * that an item put on a list will immediately be handed over to
4285 * the buddy list. This is safe since pageset manipulation is done
4286 * with interrupts disabled.
4288 * The boot_pagesets must be kept even after bootup is complete for
4289 * unused processors and/or zones. They do play a role for bootstrapping
4290 * hotplugged processors.
4292 * zoneinfo_show() and maybe other functions do
4293 * not check if the processor is online before following the pageset pointer.
4294 * Other parts of the kernel may not check if the zone is available.
4296 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4297 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4298 static void setup_zone_pageset(struct zone *zone);
4301 * Global mutex to protect against size modification of zonelists
4302 * as well as to serialize pageset setup for the new populated zone.
4304 DEFINE_MUTEX(zonelists_mutex);
4306 /* return values int ....just for stop_machine() */
4307 static int __build_all_zonelists(void *data)
4311 pg_data_t *self = data;
4314 memset(node_load, 0, sizeof(node_load));
4317 if (self && !node_online(self->node_id)) {
4318 build_zonelists(self);
4321 for_each_online_node(nid) {
4322 pg_data_t *pgdat = NODE_DATA(nid);
4324 build_zonelists(pgdat);
4328 * Initialize the boot_pagesets that are going to be used
4329 * for bootstrapping processors. The real pagesets for
4330 * each zone will be allocated later when the per cpu
4331 * allocator is available.
4333 * boot_pagesets are used also for bootstrapping offline
4334 * cpus if the system is already booted because the pagesets
4335 * are needed to initialize allocators on a specific cpu too.
4336 * F.e. the percpu allocator needs the page allocator which
4337 * needs the percpu allocator in order to allocate its pagesets
4338 * (a chicken-egg dilemma).
4340 for_each_possible_cpu(cpu) {
4341 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4343 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4345 * We now know the "local memory node" for each node--
4346 * i.e., the node of the first zone in the generic zonelist.
4347 * Set up numa_mem percpu variable for on-line cpus. During
4348 * boot, only the boot cpu should be on-line; we'll init the
4349 * secondary cpus' numa_mem as they come on-line. During
4350 * node/memory hotplug, we'll fixup all on-line cpus.
4352 if (cpu_online(cpu))
4353 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4360 static noinline void __init
4361 build_all_zonelists_init(void)
4363 __build_all_zonelists(NULL);
4364 mminit_verify_zonelist();
4365 cpuset_init_current_mems_allowed();
4369 * Called with zonelists_mutex held always
4370 * unless system_state == SYSTEM_BOOTING.
4372 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4373 * [we're only called with non-NULL zone through __meminit paths] and
4374 * (2) call of __init annotated helper build_all_zonelists_init
4375 * [protected by SYSTEM_BOOTING].
4377 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4379 set_zonelist_order();
4381 if (system_state == SYSTEM_BOOTING) {
4382 build_all_zonelists_init();
4384 #ifdef CONFIG_MEMORY_HOTPLUG
4386 setup_zone_pageset(zone);
4388 /* we have to stop all cpus to guarantee there is no user
4390 stop_machine(__build_all_zonelists, pgdat, NULL);
4391 /* cpuset refresh routine should be here */
4393 vm_total_pages = nr_free_pagecache_pages();
4395 * Disable grouping by mobility if the number of pages in the
4396 * system is too low to allow the mechanism to work. It would be
4397 * more accurate, but expensive to check per-zone. This check is
4398 * made on memory-hotadd so a system can start with mobility
4399 * disabled and enable it later
4401 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4402 page_group_by_mobility_disabled = 1;
4404 page_group_by_mobility_disabled = 0;
4406 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4407 "Total pages: %ld\n",
4409 zonelist_order_name[current_zonelist_order],
4410 page_group_by_mobility_disabled ? "off" : "on",
4413 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4418 * Helper functions to size the waitqueue hash table.
4419 * Essentially these want to choose hash table sizes sufficiently
4420 * large so that collisions trying to wait on pages are rare.
4421 * But in fact, the number of active page waitqueues on typical
4422 * systems is ridiculously low, less than 200. So this is even
4423 * conservative, even though it seems large.
4425 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4426 * waitqueues, i.e. the size of the waitq table given the number of pages.
4428 #define PAGES_PER_WAITQUEUE 256
4430 #ifndef CONFIG_MEMORY_HOTPLUG
4431 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4433 unsigned long size = 1;
4435 pages /= PAGES_PER_WAITQUEUE;
4437 while (size < pages)
4441 * Once we have dozens or even hundreds of threads sleeping
4442 * on IO we've got bigger problems than wait queue collision.
4443 * Limit the size of the wait table to a reasonable size.
4445 size = min(size, 4096UL);
4447 return max(size, 4UL);
4451 * A zone's size might be changed by hot-add, so it is not possible to determine
4452 * a suitable size for its wait_table. So we use the maximum size now.
4454 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4456 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4457 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4458 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4460 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4461 * or more by the traditional way. (See above). It equals:
4463 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4464 * ia64(16K page size) : = ( 8G + 4M)byte.
4465 * powerpc (64K page size) : = (32G +16M)byte.
4467 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4474 * This is an integer logarithm so that shifts can be used later
4475 * to extract the more random high bits from the multiplicative
4476 * hash function before the remainder is taken.
4478 static inline unsigned long wait_table_bits(unsigned long size)
4484 * Initially all pages are reserved - free ones are freed
4485 * up by free_all_bootmem() once the early boot process is
4486 * done. Non-atomic initialization, single-pass.
4488 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4489 unsigned long start_pfn, enum memmap_context context)
4491 pg_data_t *pgdat = NODE_DATA(nid);
4492 unsigned long end_pfn = start_pfn + size;
4495 unsigned long nr_initialised = 0;
4497 if (highest_memmap_pfn < end_pfn - 1)
4498 highest_memmap_pfn = end_pfn - 1;
4500 z = &pgdat->node_zones[zone];
4501 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4503 * There can be holes in boot-time mem_map[]s
4504 * handed to this function. They do not
4505 * exist on hotplugged memory.
4507 if (context == MEMMAP_EARLY) {
4508 if (!early_pfn_valid(pfn))
4510 if (!early_pfn_in_nid(pfn, nid))
4512 if (!update_defer_init(pgdat, pfn, end_pfn,
4518 * Mark the block movable so that blocks are reserved for
4519 * movable at startup. This will force kernel allocations
4520 * to reserve their blocks rather than leaking throughout
4521 * the address space during boot when many long-lived
4522 * kernel allocations are made.
4524 * bitmap is created for zone's valid pfn range. but memmap
4525 * can be created for invalid pages (for alignment)
4526 * check here not to call set_pageblock_migratetype() against
4529 if (!(pfn & (pageblock_nr_pages - 1))) {
4530 struct page *page = pfn_to_page(pfn);
4532 __init_single_page(page, pfn, zone, nid);
4533 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4535 __init_single_pfn(pfn, zone, nid);
4540 static void __meminit zone_init_free_lists(struct zone *zone)
4542 unsigned int order, t;
4543 for_each_migratetype_order(order, t) {
4544 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4545 zone->free_area[order].nr_free = 0;
4549 #ifndef __HAVE_ARCH_MEMMAP_INIT
4550 #define memmap_init(size, nid, zone, start_pfn) \
4551 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4554 static int zone_batchsize(struct zone *zone)
4560 * The per-cpu-pages pools are set to around 1000th of the
4561 * size of the zone. But no more than 1/2 of a meg.
4563 * OK, so we don't know how big the cache is. So guess.
4565 batch = zone->managed_pages / 1024;
4566 if (batch * PAGE_SIZE > 512 * 1024)
4567 batch = (512 * 1024) / PAGE_SIZE;
4568 batch /= 4; /* We effectively *= 4 below */
4573 * Clamp the batch to a 2^n - 1 value. Having a power
4574 * of 2 value was found to be more likely to have
4575 * suboptimal cache aliasing properties in some cases.
4577 * For example if 2 tasks are alternately allocating
4578 * batches of pages, one task can end up with a lot
4579 * of pages of one half of the possible page colors
4580 * and the other with pages of the other colors.
4582 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4587 /* The deferral and batching of frees should be suppressed under NOMMU
4590 * The problem is that NOMMU needs to be able to allocate large chunks
4591 * of contiguous memory as there's no hardware page translation to
4592 * assemble apparent contiguous memory from discontiguous pages.
4594 * Queueing large contiguous runs of pages for batching, however,
4595 * causes the pages to actually be freed in smaller chunks. As there
4596 * can be a significant delay between the individual batches being
4597 * recycled, this leads to the once large chunks of space being
4598 * fragmented and becoming unavailable for high-order allocations.
4605 * pcp->high and pcp->batch values are related and dependent on one another:
4606 * ->batch must never be higher then ->high.
4607 * The following function updates them in a safe manner without read side
4610 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4611 * those fields changing asynchronously (acording the the above rule).
4613 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4614 * outside of boot time (or some other assurance that no concurrent updaters
4617 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4618 unsigned long batch)
4620 /* start with a fail safe value for batch */
4624 /* Update high, then batch, in order */
4631 /* a companion to pageset_set_high() */
4632 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4634 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4637 static void pageset_init(struct per_cpu_pageset *p)
4639 struct per_cpu_pages *pcp;
4642 memset(p, 0, sizeof(*p));
4646 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4647 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4650 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4653 pageset_set_batch(p, batch);
4657 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4658 * to the value high for the pageset p.
4660 static void pageset_set_high(struct per_cpu_pageset *p,
4663 unsigned long batch = max(1UL, high / 4);
4664 if ((high / 4) > (PAGE_SHIFT * 8))
4665 batch = PAGE_SHIFT * 8;
4667 pageset_update(&p->pcp, high, batch);
4670 static void pageset_set_high_and_batch(struct zone *zone,
4671 struct per_cpu_pageset *pcp)
4673 if (percpu_pagelist_fraction)
4674 pageset_set_high(pcp,
4675 (zone->managed_pages /
4676 percpu_pagelist_fraction));
4678 pageset_set_batch(pcp, zone_batchsize(zone));
4681 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4683 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4686 pageset_set_high_and_batch(zone, pcp);
4689 static void __meminit setup_zone_pageset(struct zone *zone)
4692 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4693 for_each_possible_cpu(cpu)
4694 zone_pageset_init(zone, cpu);
4698 * Allocate per cpu pagesets and initialize them.
4699 * Before this call only boot pagesets were available.
4701 void __init setup_per_cpu_pageset(void)
4705 for_each_populated_zone(zone)
4706 setup_zone_pageset(zone);
4709 static noinline __init_refok
4710 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4716 * The per-page waitqueue mechanism uses hashed waitqueues
4719 zone->wait_table_hash_nr_entries =
4720 wait_table_hash_nr_entries(zone_size_pages);
4721 zone->wait_table_bits =
4722 wait_table_bits(zone->wait_table_hash_nr_entries);
4723 alloc_size = zone->wait_table_hash_nr_entries
4724 * sizeof(wait_queue_head_t);
4726 if (!slab_is_available()) {
4727 zone->wait_table = (wait_queue_head_t *)
4728 memblock_virt_alloc_node_nopanic(
4729 alloc_size, zone->zone_pgdat->node_id);
4732 * This case means that a zone whose size was 0 gets new memory
4733 * via memory hot-add.
4734 * But it may be the case that a new node was hot-added. In
4735 * this case vmalloc() will not be able to use this new node's
4736 * memory - this wait_table must be initialized to use this new
4737 * node itself as well.
4738 * To use this new node's memory, further consideration will be
4741 zone->wait_table = vmalloc(alloc_size);
4743 if (!zone->wait_table)
4746 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4747 init_waitqueue_head(zone->wait_table + i);
4752 static __meminit void zone_pcp_init(struct zone *zone)
4755 * per cpu subsystem is not up at this point. The following code
4756 * relies on the ability of the linker to provide the
4757 * offset of a (static) per cpu variable into the per cpu area.
4759 zone->pageset = &boot_pageset;
4761 if (populated_zone(zone))
4762 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4763 zone->name, zone->present_pages,
4764 zone_batchsize(zone));
4767 int __meminit init_currently_empty_zone(struct zone *zone,
4768 unsigned long zone_start_pfn,
4771 struct pglist_data *pgdat = zone->zone_pgdat;
4773 ret = zone_wait_table_init(zone, size);
4776 pgdat->nr_zones = zone_idx(zone) + 1;
4778 zone->zone_start_pfn = zone_start_pfn;
4780 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4781 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4783 (unsigned long)zone_idx(zone),
4784 zone_start_pfn, (zone_start_pfn + size));
4786 zone_init_free_lists(zone);
4791 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4792 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4795 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4797 int __meminit __early_pfn_to_nid(unsigned long pfn,
4798 struct mminit_pfnnid_cache *state)
4800 unsigned long start_pfn, end_pfn;
4803 if (state->last_start <= pfn && pfn < state->last_end)
4804 return state->last_nid;
4806 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4808 state->last_start = start_pfn;
4809 state->last_end = end_pfn;
4810 state->last_nid = nid;
4815 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4818 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4819 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4820 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4822 * If an architecture guarantees that all ranges registered contain no holes
4823 * and may be freed, this this function may be used instead of calling
4824 * memblock_free_early_nid() manually.
4826 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4828 unsigned long start_pfn, end_pfn;
4831 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4832 start_pfn = min(start_pfn, max_low_pfn);
4833 end_pfn = min(end_pfn, max_low_pfn);
4835 if (start_pfn < end_pfn)
4836 memblock_free_early_nid(PFN_PHYS(start_pfn),
4837 (end_pfn - start_pfn) << PAGE_SHIFT,
4843 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4844 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4846 * If an architecture guarantees that all ranges registered contain no holes and may
4847 * be freed, this function may be used instead of calling memory_present() manually.
4849 void __init sparse_memory_present_with_active_regions(int nid)
4851 unsigned long start_pfn, end_pfn;
4854 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4855 memory_present(this_nid, start_pfn, end_pfn);
4859 * get_pfn_range_for_nid - Return the start and end page frames for a node
4860 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4861 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4862 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4864 * It returns the start and end page frame of a node based on information
4865 * provided by memblock_set_node(). If called for a node
4866 * with no available memory, a warning is printed and the start and end
4869 void __meminit get_pfn_range_for_nid(unsigned int nid,
4870 unsigned long *start_pfn, unsigned long *end_pfn)
4872 unsigned long this_start_pfn, this_end_pfn;
4878 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4879 *start_pfn = min(*start_pfn, this_start_pfn);
4880 *end_pfn = max(*end_pfn, this_end_pfn);
4883 if (*start_pfn == -1UL)
4888 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4889 * assumption is made that zones within a node are ordered in monotonic
4890 * increasing memory addresses so that the "highest" populated zone is used
4892 static void __init find_usable_zone_for_movable(void)
4895 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4896 if (zone_index == ZONE_MOVABLE)
4899 if (arch_zone_highest_possible_pfn[zone_index] >
4900 arch_zone_lowest_possible_pfn[zone_index])
4904 VM_BUG_ON(zone_index == -1);
4905 movable_zone = zone_index;
4909 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4910 * because it is sized independent of architecture. Unlike the other zones,
4911 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4912 * in each node depending on the size of each node and how evenly kernelcore
4913 * is distributed. This helper function adjusts the zone ranges
4914 * provided by the architecture for a given node by using the end of the
4915 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4916 * zones within a node are in order of monotonic increases memory addresses
4918 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4919 unsigned long zone_type,
4920 unsigned long node_start_pfn,
4921 unsigned long node_end_pfn,
4922 unsigned long *zone_start_pfn,
4923 unsigned long *zone_end_pfn)
4925 /* Only adjust if ZONE_MOVABLE is on this node */
4926 if (zone_movable_pfn[nid]) {
4927 /* Size ZONE_MOVABLE */
4928 if (zone_type == ZONE_MOVABLE) {
4929 *zone_start_pfn = zone_movable_pfn[nid];
4930 *zone_end_pfn = min(node_end_pfn,
4931 arch_zone_highest_possible_pfn[movable_zone]);
4933 /* Adjust for ZONE_MOVABLE starting within this range */
4934 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4935 *zone_end_pfn > zone_movable_pfn[nid]) {
4936 *zone_end_pfn = zone_movable_pfn[nid];
4938 /* Check if this whole range is within ZONE_MOVABLE */
4939 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4940 *zone_start_pfn = *zone_end_pfn;
4945 * Return the number of pages a zone spans in a node, including holes
4946 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4948 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4949 unsigned long zone_type,
4950 unsigned long node_start_pfn,
4951 unsigned long node_end_pfn,
4952 unsigned long *ignored)
4954 unsigned long zone_start_pfn, zone_end_pfn;
4956 /* When hotadd a new node from cpu_up(), the node should be empty */
4957 if (!node_start_pfn && !node_end_pfn)
4960 /* Get the start and end of the zone */
4961 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4962 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4963 adjust_zone_range_for_zone_movable(nid, zone_type,
4964 node_start_pfn, node_end_pfn,
4965 &zone_start_pfn, &zone_end_pfn);
4967 /* Check that this node has pages within the zone's required range */
4968 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4971 /* Move the zone boundaries inside the node if necessary */
4972 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4973 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4975 /* Return the spanned pages */
4976 return zone_end_pfn - zone_start_pfn;
4980 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4981 * then all holes in the requested range will be accounted for.
4983 unsigned long __meminit __absent_pages_in_range(int nid,
4984 unsigned long range_start_pfn,
4985 unsigned long range_end_pfn)
4987 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4988 unsigned long start_pfn, end_pfn;
4991 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4992 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4993 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4994 nr_absent -= end_pfn - start_pfn;
5000 * absent_pages_in_range - Return number of page frames in holes within a range
5001 * @start_pfn: The start PFN to start searching for holes
5002 * @end_pfn: The end PFN to stop searching for holes
5004 * It returns the number of pages frames in memory holes within a range.
5006 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5007 unsigned long end_pfn)
5009 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5012 /* Return the number of page frames in holes in a zone on a node */
5013 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5014 unsigned long zone_type,
5015 unsigned long node_start_pfn,
5016 unsigned long node_end_pfn,
5017 unsigned long *ignored)
5019 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5020 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5021 unsigned long zone_start_pfn, zone_end_pfn;
5023 /* When hotadd a new node from cpu_up(), the node should be empty */
5024 if (!node_start_pfn && !node_end_pfn)
5027 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5028 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5030 adjust_zone_range_for_zone_movable(nid, zone_type,
5031 node_start_pfn, node_end_pfn,
5032 &zone_start_pfn, &zone_end_pfn);
5033 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5036 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5037 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5038 unsigned long zone_type,
5039 unsigned long node_start_pfn,
5040 unsigned long node_end_pfn,
5041 unsigned long *zones_size)
5043 return zones_size[zone_type];
5046 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5047 unsigned long zone_type,
5048 unsigned long node_start_pfn,
5049 unsigned long node_end_pfn,
5050 unsigned long *zholes_size)
5055 return zholes_size[zone_type];
5058 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5060 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5061 unsigned long node_start_pfn,
5062 unsigned long node_end_pfn,
5063 unsigned long *zones_size,
5064 unsigned long *zholes_size)
5066 unsigned long realtotalpages = 0, totalpages = 0;
5069 for (i = 0; i < MAX_NR_ZONES; i++) {
5070 struct zone *zone = pgdat->node_zones + i;
5071 unsigned long size, real_size;
5073 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5077 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5078 node_start_pfn, node_end_pfn,
5080 zone->spanned_pages = size;
5081 zone->present_pages = real_size;
5084 realtotalpages += real_size;
5087 pgdat->node_spanned_pages = totalpages;
5088 pgdat->node_present_pages = realtotalpages;
5089 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5093 #ifndef CONFIG_SPARSEMEM
5095 * Calculate the size of the zone->blockflags rounded to an unsigned long
5096 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5097 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5098 * round what is now in bits to nearest long in bits, then return it in
5101 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5103 unsigned long usemapsize;
5105 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5106 usemapsize = roundup(zonesize, pageblock_nr_pages);
5107 usemapsize = usemapsize >> pageblock_order;
5108 usemapsize *= NR_PAGEBLOCK_BITS;
5109 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5111 return usemapsize / 8;
5114 static void __init setup_usemap(struct pglist_data *pgdat,
5116 unsigned long zone_start_pfn,
5117 unsigned long zonesize)
5119 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5120 zone->pageblock_flags = NULL;
5122 zone->pageblock_flags =
5123 memblock_virt_alloc_node_nopanic(usemapsize,
5127 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5128 unsigned long zone_start_pfn, unsigned long zonesize) {}
5129 #endif /* CONFIG_SPARSEMEM */
5131 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5133 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5134 void __paginginit set_pageblock_order(void)
5138 /* Check that pageblock_nr_pages has not already been setup */
5139 if (pageblock_order)
5142 if (HPAGE_SHIFT > PAGE_SHIFT)
5143 order = HUGETLB_PAGE_ORDER;
5145 order = MAX_ORDER - 1;
5148 * Assume the largest contiguous order of interest is a huge page.
5149 * This value may be variable depending on boot parameters on IA64 and
5152 pageblock_order = order;
5154 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5157 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5158 * is unused as pageblock_order is set at compile-time. See
5159 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5162 void __paginginit set_pageblock_order(void)
5166 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5168 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5169 unsigned long present_pages)
5171 unsigned long pages = spanned_pages;
5174 * Provide a more accurate estimation if there are holes within
5175 * the zone and SPARSEMEM is in use. If there are holes within the
5176 * zone, each populated memory region may cost us one or two extra
5177 * memmap pages due to alignment because memmap pages for each
5178 * populated regions may not naturally algined on page boundary.
5179 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5181 if (spanned_pages > present_pages + (present_pages >> 4) &&
5182 IS_ENABLED(CONFIG_SPARSEMEM))
5183 pages = present_pages;
5185 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5189 * Set up the zone data structures:
5190 * - mark all pages reserved
5191 * - mark all memory queues empty
5192 * - clear the memory bitmaps
5194 * NOTE: pgdat should get zeroed by caller.
5196 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5199 int nid = pgdat->node_id;
5200 unsigned long zone_start_pfn = pgdat->node_start_pfn;
5203 pgdat_resize_init(pgdat);
5204 #ifdef CONFIG_NUMA_BALANCING
5205 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5206 pgdat->numabalancing_migrate_nr_pages = 0;
5207 pgdat->numabalancing_migrate_next_window = jiffies;
5209 init_waitqueue_head(&pgdat->kswapd_wait);
5210 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5211 pgdat_page_ext_init(pgdat);
5213 for (j = 0; j < MAX_NR_ZONES; j++) {
5214 struct zone *zone = pgdat->node_zones + j;
5215 unsigned long size, realsize, freesize, memmap_pages;
5217 size = zone->spanned_pages;
5218 realsize = freesize = zone->present_pages;
5221 * Adjust freesize so that it accounts for how much memory
5222 * is used by this zone for memmap. This affects the watermark
5223 * and per-cpu initialisations
5225 memmap_pages = calc_memmap_size(size, realsize);
5226 if (!is_highmem_idx(j)) {
5227 if (freesize >= memmap_pages) {
5228 freesize -= memmap_pages;
5231 " %s zone: %lu pages used for memmap\n",
5232 zone_names[j], memmap_pages);
5235 " %s zone: %lu pages exceeds freesize %lu\n",
5236 zone_names[j], memmap_pages, freesize);
5239 /* Account for reserved pages */
5240 if (j == 0 && freesize > dma_reserve) {
5241 freesize -= dma_reserve;
5242 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5243 zone_names[0], dma_reserve);
5246 if (!is_highmem_idx(j))
5247 nr_kernel_pages += freesize;
5248 /* Charge for highmem memmap if there are enough kernel pages */
5249 else if (nr_kernel_pages > memmap_pages * 2)
5250 nr_kernel_pages -= memmap_pages;
5251 nr_all_pages += freesize;
5254 * Set an approximate value for lowmem here, it will be adjusted
5255 * when the bootmem allocator frees pages into the buddy system.
5256 * And all highmem pages will be managed by the buddy system.
5258 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5261 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5263 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5265 zone->name = zone_names[j];
5266 spin_lock_init(&zone->lock);
5267 spin_lock_init(&zone->lru_lock);
5268 zone_seqlock_init(zone);
5269 zone->zone_pgdat = pgdat;
5270 zone_pcp_init(zone);
5272 /* For bootup, initialized properly in watermark setup */
5273 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5275 lruvec_init(&zone->lruvec);
5279 set_pageblock_order();
5280 setup_usemap(pgdat, zone, zone_start_pfn, size);
5281 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5283 memmap_init(size, nid, j, zone_start_pfn);
5284 zone_start_pfn += size;
5288 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5290 unsigned long __maybe_unused start = 0;
5291 unsigned long __maybe_unused offset = 0;
5293 /* Skip empty nodes */
5294 if (!pgdat->node_spanned_pages)
5297 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5298 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5299 offset = pgdat->node_start_pfn - start;
5300 /* ia64 gets its own node_mem_map, before this, without bootmem */
5301 if (!pgdat->node_mem_map) {
5302 unsigned long size, end;
5306 * The zone's endpoints aren't required to be MAX_ORDER
5307 * aligned but the node_mem_map endpoints must be in order
5308 * for the buddy allocator to function correctly.
5310 end = pgdat_end_pfn(pgdat);
5311 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5312 size = (end - start) * sizeof(struct page);
5313 map = alloc_remap(pgdat->node_id, size);
5315 map = memblock_virt_alloc_node_nopanic(size,
5317 pgdat->node_mem_map = map + offset;
5319 #ifndef CONFIG_NEED_MULTIPLE_NODES
5321 * With no DISCONTIG, the global mem_map is just set as node 0's
5323 if (pgdat == NODE_DATA(0)) {
5324 mem_map = NODE_DATA(0)->node_mem_map;
5325 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5326 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5328 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5331 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5334 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5335 unsigned long node_start_pfn, unsigned long *zholes_size)
5337 pg_data_t *pgdat = NODE_DATA(nid);
5338 unsigned long start_pfn = 0;
5339 unsigned long end_pfn = 0;
5341 /* pg_data_t should be reset to zero when it's allocated */
5342 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5344 reset_deferred_meminit(pgdat);
5345 pgdat->node_id = nid;
5346 pgdat->node_start_pfn = node_start_pfn;
5347 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5348 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5349 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5350 (u64)start_pfn << PAGE_SHIFT,
5351 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5353 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5354 zones_size, zholes_size);
5356 alloc_node_mem_map(pgdat);
5357 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5358 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5359 nid, (unsigned long)pgdat,
5360 (unsigned long)pgdat->node_mem_map);
5363 free_area_init_core(pgdat);
5366 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5368 #if MAX_NUMNODES > 1
5370 * Figure out the number of possible node ids.
5372 void __init setup_nr_node_ids(void)
5374 unsigned int highest;
5376 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5377 nr_node_ids = highest + 1;
5382 * node_map_pfn_alignment - determine the maximum internode alignment
5384 * This function should be called after node map is populated and sorted.
5385 * It calculates the maximum power of two alignment which can distinguish
5388 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5389 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5390 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5391 * shifted, 1GiB is enough and this function will indicate so.
5393 * This is used to test whether pfn -> nid mapping of the chosen memory
5394 * model has fine enough granularity to avoid incorrect mapping for the
5395 * populated node map.
5397 * Returns the determined alignment in pfn's. 0 if there is no alignment
5398 * requirement (single node).
5400 unsigned long __init node_map_pfn_alignment(void)
5402 unsigned long accl_mask = 0, last_end = 0;
5403 unsigned long start, end, mask;
5407 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5408 if (!start || last_nid < 0 || last_nid == nid) {
5415 * Start with a mask granular enough to pin-point to the
5416 * start pfn and tick off bits one-by-one until it becomes
5417 * too coarse to separate the current node from the last.
5419 mask = ~((1 << __ffs(start)) - 1);
5420 while (mask && last_end <= (start & (mask << 1)))
5423 /* accumulate all internode masks */
5427 /* convert mask to number of pages */
5428 return ~accl_mask + 1;
5431 /* Find the lowest pfn for a node */
5432 static unsigned long __init find_min_pfn_for_node(int nid)
5434 unsigned long min_pfn = ULONG_MAX;
5435 unsigned long start_pfn;
5438 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5439 min_pfn = min(min_pfn, start_pfn);
5441 if (min_pfn == ULONG_MAX) {
5443 "Could not find start_pfn for node %d\n", nid);
5451 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5453 * It returns the minimum PFN based on information provided via
5454 * memblock_set_node().
5456 unsigned long __init find_min_pfn_with_active_regions(void)
5458 return find_min_pfn_for_node(MAX_NUMNODES);
5462 * early_calculate_totalpages()
5463 * Sum pages in active regions for movable zone.
5464 * Populate N_MEMORY for calculating usable_nodes.
5466 static unsigned long __init early_calculate_totalpages(void)
5468 unsigned long totalpages = 0;
5469 unsigned long start_pfn, end_pfn;
5472 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5473 unsigned long pages = end_pfn - start_pfn;
5475 totalpages += pages;
5477 node_set_state(nid, N_MEMORY);
5483 * Find the PFN the Movable zone begins in each node. Kernel memory
5484 * is spread evenly between nodes as long as the nodes have enough
5485 * memory. When they don't, some nodes will have more kernelcore than
5488 static void __init find_zone_movable_pfns_for_nodes(void)
5491 unsigned long usable_startpfn;
5492 unsigned long kernelcore_node, kernelcore_remaining;
5493 /* save the state before borrow the nodemask */
5494 nodemask_t saved_node_state = node_states[N_MEMORY];
5495 unsigned long totalpages = early_calculate_totalpages();
5496 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5497 struct memblock_region *r;
5499 /* Need to find movable_zone earlier when movable_node is specified. */
5500 find_usable_zone_for_movable();
5503 * If movable_node is specified, ignore kernelcore and movablecore
5506 if (movable_node_is_enabled()) {
5507 for_each_memblock(memory, r) {
5508 if (!memblock_is_hotpluggable(r))
5513 usable_startpfn = PFN_DOWN(r->base);
5514 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5515 min(usable_startpfn, zone_movable_pfn[nid]) :
5523 * If movablecore=nn[KMG] was specified, calculate what size of
5524 * kernelcore that corresponds so that memory usable for
5525 * any allocation type is evenly spread. If both kernelcore
5526 * and movablecore are specified, then the value of kernelcore
5527 * will be used for required_kernelcore if it's greater than
5528 * what movablecore would have allowed.
5530 if (required_movablecore) {
5531 unsigned long corepages;
5534 * Round-up so that ZONE_MOVABLE is at least as large as what
5535 * was requested by the user
5537 required_movablecore =
5538 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5539 required_movablecore = min(totalpages, required_movablecore);
5540 corepages = totalpages - required_movablecore;
5542 required_kernelcore = max(required_kernelcore, corepages);
5546 * If kernelcore was not specified or kernelcore size is larger
5547 * than totalpages, there is no ZONE_MOVABLE.
5549 if (!required_kernelcore || required_kernelcore >= totalpages)
5552 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5553 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5556 /* Spread kernelcore memory as evenly as possible throughout nodes */
5557 kernelcore_node = required_kernelcore / usable_nodes;
5558 for_each_node_state(nid, N_MEMORY) {
5559 unsigned long start_pfn, end_pfn;
5562 * Recalculate kernelcore_node if the division per node
5563 * now exceeds what is necessary to satisfy the requested
5564 * amount of memory for the kernel
5566 if (required_kernelcore < kernelcore_node)
5567 kernelcore_node = required_kernelcore / usable_nodes;
5570 * As the map is walked, we track how much memory is usable
5571 * by the kernel using kernelcore_remaining. When it is
5572 * 0, the rest of the node is usable by ZONE_MOVABLE
5574 kernelcore_remaining = kernelcore_node;
5576 /* Go through each range of PFNs within this node */
5577 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5578 unsigned long size_pages;
5580 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5581 if (start_pfn >= end_pfn)
5584 /* Account for what is only usable for kernelcore */
5585 if (start_pfn < usable_startpfn) {
5586 unsigned long kernel_pages;
5587 kernel_pages = min(end_pfn, usable_startpfn)
5590 kernelcore_remaining -= min(kernel_pages,
5591 kernelcore_remaining);
5592 required_kernelcore -= min(kernel_pages,
5593 required_kernelcore);
5595 /* Continue if range is now fully accounted */
5596 if (end_pfn <= usable_startpfn) {
5599 * Push zone_movable_pfn to the end so
5600 * that if we have to rebalance
5601 * kernelcore across nodes, we will
5602 * not double account here
5604 zone_movable_pfn[nid] = end_pfn;
5607 start_pfn = usable_startpfn;
5611 * The usable PFN range for ZONE_MOVABLE is from
5612 * start_pfn->end_pfn. Calculate size_pages as the
5613 * number of pages used as kernelcore
5615 size_pages = end_pfn - start_pfn;
5616 if (size_pages > kernelcore_remaining)
5617 size_pages = kernelcore_remaining;
5618 zone_movable_pfn[nid] = start_pfn + size_pages;
5621 * Some kernelcore has been met, update counts and
5622 * break if the kernelcore for this node has been
5625 required_kernelcore -= min(required_kernelcore,
5627 kernelcore_remaining -= size_pages;
5628 if (!kernelcore_remaining)
5634 * If there is still required_kernelcore, we do another pass with one
5635 * less node in the count. This will push zone_movable_pfn[nid] further
5636 * along on the nodes that still have memory until kernelcore is
5640 if (usable_nodes && required_kernelcore > usable_nodes)
5644 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5645 for (nid = 0; nid < MAX_NUMNODES; nid++)
5646 zone_movable_pfn[nid] =
5647 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5650 /* restore the node_state */
5651 node_states[N_MEMORY] = saved_node_state;
5654 /* Any regular or high memory on that node ? */
5655 static void check_for_memory(pg_data_t *pgdat, int nid)
5657 enum zone_type zone_type;
5659 if (N_MEMORY == N_NORMAL_MEMORY)
5662 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5663 struct zone *zone = &pgdat->node_zones[zone_type];
5664 if (populated_zone(zone)) {
5665 node_set_state(nid, N_HIGH_MEMORY);
5666 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5667 zone_type <= ZONE_NORMAL)
5668 node_set_state(nid, N_NORMAL_MEMORY);
5675 * free_area_init_nodes - Initialise all pg_data_t and zone data
5676 * @max_zone_pfn: an array of max PFNs for each zone
5678 * This will call free_area_init_node() for each active node in the system.
5679 * Using the page ranges provided by memblock_set_node(), the size of each
5680 * zone in each node and their holes is calculated. If the maximum PFN
5681 * between two adjacent zones match, it is assumed that the zone is empty.
5682 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5683 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5684 * starts where the previous one ended. For example, ZONE_DMA32 starts
5685 * at arch_max_dma_pfn.
5687 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5689 unsigned long start_pfn, end_pfn;
5692 /* Record where the zone boundaries are */
5693 memset(arch_zone_lowest_possible_pfn, 0,
5694 sizeof(arch_zone_lowest_possible_pfn));
5695 memset(arch_zone_highest_possible_pfn, 0,
5696 sizeof(arch_zone_highest_possible_pfn));
5697 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5698 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5699 for (i = 1; i < MAX_NR_ZONES; i++) {
5700 if (i == ZONE_MOVABLE)
5702 arch_zone_lowest_possible_pfn[i] =
5703 arch_zone_highest_possible_pfn[i-1];
5704 arch_zone_highest_possible_pfn[i] =
5705 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5707 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5708 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5710 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5711 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5712 find_zone_movable_pfns_for_nodes();
5714 /* Print out the zone ranges */
5715 pr_info("Zone ranges:\n");
5716 for (i = 0; i < MAX_NR_ZONES; i++) {
5717 if (i == ZONE_MOVABLE)
5719 pr_info(" %-8s ", zone_names[i]);
5720 if (arch_zone_lowest_possible_pfn[i] ==
5721 arch_zone_highest_possible_pfn[i])
5724 pr_cont("[mem %#018Lx-%#018Lx]\n",
5725 (u64)arch_zone_lowest_possible_pfn[i]
5727 ((u64)arch_zone_highest_possible_pfn[i]
5728 << PAGE_SHIFT) - 1);
5731 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5732 pr_info("Movable zone start for each node\n");
5733 for (i = 0; i < MAX_NUMNODES; i++) {
5734 if (zone_movable_pfn[i])
5735 pr_info(" Node %d: %#018Lx\n", i,
5736 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5739 /* Print out the early node map */
5740 pr_info("Early memory node ranges\n");
5741 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5742 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5743 (u64)start_pfn << PAGE_SHIFT,
5744 ((u64)end_pfn << PAGE_SHIFT) - 1);
5746 /* Initialise every node */
5747 mminit_verify_pageflags_layout();
5748 setup_nr_node_ids();
5749 for_each_online_node(nid) {
5750 pg_data_t *pgdat = NODE_DATA(nid);
5751 free_area_init_node(nid, NULL,
5752 find_min_pfn_for_node(nid), NULL);
5754 /* Any memory on that node */
5755 if (pgdat->node_present_pages)
5756 node_set_state(nid, N_MEMORY);
5757 check_for_memory(pgdat, nid);
5761 static int __init cmdline_parse_core(char *p, unsigned long *core)
5763 unsigned long long coremem;
5767 coremem = memparse(p, &p);
5768 *core = coremem >> PAGE_SHIFT;
5770 /* Paranoid check that UL is enough for the coremem value */
5771 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5777 * kernelcore=size sets the amount of memory for use for allocations that
5778 * cannot be reclaimed or migrated.
5780 static int __init cmdline_parse_kernelcore(char *p)
5782 return cmdline_parse_core(p, &required_kernelcore);
5786 * movablecore=size sets the amount of memory for use for allocations that
5787 * can be reclaimed or migrated.
5789 static int __init cmdline_parse_movablecore(char *p)
5791 return cmdline_parse_core(p, &required_movablecore);
5794 early_param("kernelcore", cmdline_parse_kernelcore);
5795 early_param("movablecore", cmdline_parse_movablecore);
5797 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5799 void adjust_managed_page_count(struct page *page, long count)
5801 spin_lock(&managed_page_count_lock);
5802 page_zone(page)->managed_pages += count;
5803 totalram_pages += count;
5804 #ifdef CONFIG_HIGHMEM
5805 if (PageHighMem(page))
5806 totalhigh_pages += count;
5808 spin_unlock(&managed_page_count_lock);
5810 EXPORT_SYMBOL(adjust_managed_page_count);
5812 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5815 unsigned long pages = 0;
5817 start = (void *)PAGE_ALIGN((unsigned long)start);
5818 end = (void *)((unsigned long)end & PAGE_MASK);
5819 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5820 if ((unsigned int)poison <= 0xFF)
5821 memset(pos, poison, PAGE_SIZE);
5822 free_reserved_page(virt_to_page(pos));
5826 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5827 s, pages << (PAGE_SHIFT - 10), start, end);
5831 EXPORT_SYMBOL(free_reserved_area);
5833 #ifdef CONFIG_HIGHMEM
5834 void free_highmem_page(struct page *page)
5836 __free_reserved_page(page);
5838 page_zone(page)->managed_pages++;
5844 void __init mem_init_print_info(const char *str)
5846 unsigned long physpages, codesize, datasize, rosize, bss_size;
5847 unsigned long init_code_size, init_data_size;
5849 physpages = get_num_physpages();
5850 codesize = _etext - _stext;
5851 datasize = _edata - _sdata;
5852 rosize = __end_rodata - __start_rodata;
5853 bss_size = __bss_stop - __bss_start;
5854 init_data_size = __init_end - __init_begin;
5855 init_code_size = _einittext - _sinittext;
5858 * Detect special cases and adjust section sizes accordingly:
5859 * 1) .init.* may be embedded into .data sections
5860 * 2) .init.text.* may be out of [__init_begin, __init_end],
5861 * please refer to arch/tile/kernel/vmlinux.lds.S.
5862 * 3) .rodata.* may be embedded into .text or .data sections.
5864 #define adj_init_size(start, end, size, pos, adj) \
5866 if (start <= pos && pos < end && size > adj) \
5870 adj_init_size(__init_begin, __init_end, init_data_size,
5871 _sinittext, init_code_size);
5872 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5873 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5874 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5875 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5877 #undef adj_init_size
5879 pr_info("Memory: %luK/%luK available "
5880 "(%luK kernel code, %luK rwdata, %luK rodata, "
5881 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5882 #ifdef CONFIG_HIGHMEM
5886 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5887 codesize >> 10, datasize >> 10, rosize >> 10,
5888 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5889 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5890 totalcma_pages << (PAGE_SHIFT-10),
5891 #ifdef CONFIG_HIGHMEM
5892 totalhigh_pages << (PAGE_SHIFT-10),
5894 str ? ", " : "", str ? str : "");
5898 * set_dma_reserve - set the specified number of pages reserved in the first zone
5899 * @new_dma_reserve: The number of pages to mark reserved
5901 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
5902 * In the DMA zone, a significant percentage may be consumed by kernel image
5903 * and other unfreeable allocations which can skew the watermarks badly. This
5904 * function may optionally be used to account for unfreeable pages in the
5905 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5906 * smaller per-cpu batchsize.
5908 void __init set_dma_reserve(unsigned long new_dma_reserve)
5910 dma_reserve = new_dma_reserve;
5913 void __init free_area_init(unsigned long *zones_size)
5915 free_area_init_node(0, zones_size,
5916 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5919 static int page_alloc_cpu_notify(struct notifier_block *self,
5920 unsigned long action, void *hcpu)
5922 int cpu = (unsigned long)hcpu;
5924 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5925 lru_add_drain_cpu(cpu);
5929 * Spill the event counters of the dead processor
5930 * into the current processors event counters.
5931 * This artificially elevates the count of the current
5934 vm_events_fold_cpu(cpu);
5937 * Zero the differential counters of the dead processor
5938 * so that the vm statistics are consistent.
5940 * This is only okay since the processor is dead and cannot
5941 * race with what we are doing.
5943 cpu_vm_stats_fold(cpu);
5948 void __init page_alloc_init(void)
5950 hotcpu_notifier(page_alloc_cpu_notify, 0);
5954 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5955 * or min_free_kbytes changes.
5957 static void calculate_totalreserve_pages(void)
5959 struct pglist_data *pgdat;
5960 unsigned long reserve_pages = 0;
5961 enum zone_type i, j;
5963 for_each_online_pgdat(pgdat) {
5964 for (i = 0; i < MAX_NR_ZONES; i++) {
5965 struct zone *zone = pgdat->node_zones + i;
5968 /* Find valid and maximum lowmem_reserve in the zone */
5969 for (j = i; j < MAX_NR_ZONES; j++) {
5970 if (zone->lowmem_reserve[j] > max)
5971 max = zone->lowmem_reserve[j];
5974 /* we treat the high watermark as reserved pages. */
5975 max += high_wmark_pages(zone);
5977 if (max > zone->managed_pages)
5978 max = zone->managed_pages;
5979 reserve_pages += max;
5981 * Lowmem reserves are not available to
5982 * GFP_HIGHUSER page cache allocations and
5983 * kswapd tries to balance zones to their high
5984 * watermark. As a result, neither should be
5985 * regarded as dirtyable memory, to prevent a
5986 * situation where reclaim has to clean pages
5987 * in order to balance the zones.
5989 zone->dirty_balance_reserve = max;
5992 dirty_balance_reserve = reserve_pages;
5993 totalreserve_pages = reserve_pages;
5997 * setup_per_zone_lowmem_reserve - called whenever
5998 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
5999 * has a correct pages reserved value, so an adequate number of
6000 * pages are left in the zone after a successful __alloc_pages().
6002 static void setup_per_zone_lowmem_reserve(void)
6004 struct pglist_data *pgdat;
6005 enum zone_type j, idx;
6007 for_each_online_pgdat(pgdat) {
6008 for (j = 0; j < MAX_NR_ZONES; j++) {
6009 struct zone *zone = pgdat->node_zones + j;
6010 unsigned long managed_pages = zone->managed_pages;
6012 zone->lowmem_reserve[j] = 0;
6016 struct zone *lower_zone;
6020 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6021 sysctl_lowmem_reserve_ratio[idx] = 1;
6023 lower_zone = pgdat->node_zones + idx;
6024 lower_zone->lowmem_reserve[j] = managed_pages /
6025 sysctl_lowmem_reserve_ratio[idx];
6026 managed_pages += lower_zone->managed_pages;
6031 /* update totalreserve_pages */
6032 calculate_totalreserve_pages();
6035 static void __setup_per_zone_wmarks(void)
6037 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6038 unsigned long lowmem_pages = 0;
6040 unsigned long flags;
6042 /* Calculate total number of !ZONE_HIGHMEM pages */
6043 for_each_zone(zone) {
6044 if (!is_highmem(zone))
6045 lowmem_pages += zone->managed_pages;
6048 for_each_zone(zone) {
6051 spin_lock_irqsave(&zone->lock, flags);
6052 tmp = (u64)pages_min * zone->managed_pages;
6053 do_div(tmp, lowmem_pages);
6054 if (is_highmem(zone)) {
6056 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6057 * need highmem pages, so cap pages_min to a small
6060 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6061 * deltas control asynch page reclaim, and so should
6062 * not be capped for highmem.
6064 unsigned long min_pages;
6066 min_pages = zone->managed_pages / 1024;
6067 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6068 zone->watermark[WMARK_MIN] = min_pages;
6071 * If it's a lowmem zone, reserve a number of pages
6072 * proportionate to the zone's size.
6074 zone->watermark[WMARK_MIN] = tmp;
6077 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
6078 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
6080 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6081 high_wmark_pages(zone) - low_wmark_pages(zone) -
6082 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6084 spin_unlock_irqrestore(&zone->lock, flags);
6087 /* update totalreserve_pages */
6088 calculate_totalreserve_pages();
6092 * setup_per_zone_wmarks - called when min_free_kbytes changes
6093 * or when memory is hot-{added|removed}
6095 * Ensures that the watermark[min,low,high] values for each zone are set
6096 * correctly with respect to min_free_kbytes.
6098 void setup_per_zone_wmarks(void)
6100 mutex_lock(&zonelists_mutex);
6101 __setup_per_zone_wmarks();
6102 mutex_unlock(&zonelists_mutex);
6106 * The inactive anon list should be small enough that the VM never has to
6107 * do too much work, but large enough that each inactive page has a chance
6108 * to be referenced again before it is swapped out.
6110 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6111 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6112 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6113 * the anonymous pages are kept on the inactive list.
6116 * memory ratio inactive anon
6117 * -------------------------------------
6126 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6128 unsigned int gb, ratio;
6130 /* Zone size in gigabytes */
6131 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6133 ratio = int_sqrt(10 * gb);
6137 zone->inactive_ratio = ratio;
6140 static void __meminit setup_per_zone_inactive_ratio(void)
6145 calculate_zone_inactive_ratio(zone);
6149 * Initialise min_free_kbytes.
6151 * For small machines we want it small (128k min). For large machines
6152 * we want it large (64MB max). But it is not linear, because network
6153 * bandwidth does not increase linearly with machine size. We use
6155 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6156 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6172 int __meminit init_per_zone_wmark_min(void)
6174 unsigned long lowmem_kbytes;
6175 int new_min_free_kbytes;
6177 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6178 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6180 if (new_min_free_kbytes > user_min_free_kbytes) {
6181 min_free_kbytes = new_min_free_kbytes;
6182 if (min_free_kbytes < 128)
6183 min_free_kbytes = 128;
6184 if (min_free_kbytes > 65536)
6185 min_free_kbytes = 65536;
6187 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6188 new_min_free_kbytes, user_min_free_kbytes);
6190 setup_per_zone_wmarks();
6191 refresh_zone_stat_thresholds();
6192 setup_per_zone_lowmem_reserve();
6193 setup_per_zone_inactive_ratio();
6196 module_init(init_per_zone_wmark_min)
6199 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6200 * that we can call two helper functions whenever min_free_kbytes
6203 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6204 void __user *buffer, size_t *length, loff_t *ppos)
6208 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6213 user_min_free_kbytes = min_free_kbytes;
6214 setup_per_zone_wmarks();
6220 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6221 void __user *buffer, size_t *length, loff_t *ppos)
6226 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6231 zone->min_unmapped_pages = (zone->managed_pages *
6232 sysctl_min_unmapped_ratio) / 100;
6236 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6237 void __user *buffer, size_t *length, loff_t *ppos)
6242 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6247 zone->min_slab_pages = (zone->managed_pages *
6248 sysctl_min_slab_ratio) / 100;
6254 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6255 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6256 * whenever sysctl_lowmem_reserve_ratio changes.
6258 * The reserve ratio obviously has absolutely no relation with the
6259 * minimum watermarks. The lowmem reserve ratio can only make sense
6260 * if in function of the boot time zone sizes.
6262 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6263 void __user *buffer, size_t *length, loff_t *ppos)
6265 proc_dointvec_minmax(table, write, buffer, length, ppos);
6266 setup_per_zone_lowmem_reserve();
6271 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6272 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6273 * pagelist can have before it gets flushed back to buddy allocator.
6275 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6276 void __user *buffer, size_t *length, loff_t *ppos)
6279 int old_percpu_pagelist_fraction;
6282 mutex_lock(&pcp_batch_high_lock);
6283 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6285 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6286 if (!write || ret < 0)
6289 /* Sanity checking to avoid pcp imbalance */
6290 if (percpu_pagelist_fraction &&
6291 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6292 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6298 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6301 for_each_populated_zone(zone) {
6304 for_each_possible_cpu(cpu)
6305 pageset_set_high_and_batch(zone,
6306 per_cpu_ptr(zone->pageset, cpu));
6309 mutex_unlock(&pcp_batch_high_lock);
6314 int hashdist = HASHDIST_DEFAULT;
6316 static int __init set_hashdist(char *str)
6320 hashdist = simple_strtoul(str, &str, 0);
6323 __setup("hashdist=", set_hashdist);
6327 * allocate a large system hash table from bootmem
6328 * - it is assumed that the hash table must contain an exact power-of-2
6329 * quantity of entries
6330 * - limit is the number of hash buckets, not the total allocation size
6332 void *__init alloc_large_system_hash(const char *tablename,
6333 unsigned long bucketsize,
6334 unsigned long numentries,
6337 unsigned int *_hash_shift,
6338 unsigned int *_hash_mask,
6339 unsigned long low_limit,
6340 unsigned long high_limit)
6342 unsigned long long max = high_limit;
6343 unsigned long log2qty, size;
6346 /* allow the kernel cmdline to have a say */
6348 /* round applicable memory size up to nearest megabyte */
6349 numentries = nr_kernel_pages;
6351 /* It isn't necessary when PAGE_SIZE >= 1MB */
6352 if (PAGE_SHIFT < 20)
6353 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6355 /* limit to 1 bucket per 2^scale bytes of low memory */
6356 if (scale > PAGE_SHIFT)
6357 numentries >>= (scale - PAGE_SHIFT);
6359 numentries <<= (PAGE_SHIFT - scale);
6361 /* Make sure we've got at least a 0-order allocation.. */
6362 if (unlikely(flags & HASH_SMALL)) {
6363 /* Makes no sense without HASH_EARLY */
6364 WARN_ON(!(flags & HASH_EARLY));
6365 if (!(numentries >> *_hash_shift)) {
6366 numentries = 1UL << *_hash_shift;
6367 BUG_ON(!numentries);
6369 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6370 numentries = PAGE_SIZE / bucketsize;
6372 numentries = roundup_pow_of_two(numentries);
6374 /* limit allocation size to 1/16 total memory by default */
6376 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6377 do_div(max, bucketsize);
6379 max = min(max, 0x80000000ULL);
6381 if (numentries < low_limit)
6382 numentries = low_limit;
6383 if (numentries > max)
6386 log2qty = ilog2(numentries);
6389 size = bucketsize << log2qty;
6390 if (flags & HASH_EARLY)
6391 table = memblock_virt_alloc_nopanic(size, 0);
6393 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6396 * If bucketsize is not a power-of-two, we may free
6397 * some pages at the end of hash table which
6398 * alloc_pages_exact() automatically does
6400 if (get_order(size) < MAX_ORDER) {
6401 table = alloc_pages_exact(size, GFP_ATOMIC);
6402 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6405 } while (!table && size > PAGE_SIZE && --log2qty);
6408 panic("Failed to allocate %s hash table\n", tablename);
6410 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6413 ilog2(size) - PAGE_SHIFT,
6417 *_hash_shift = log2qty;
6419 *_hash_mask = (1 << log2qty) - 1;
6424 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6425 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6428 #ifdef CONFIG_SPARSEMEM
6429 return __pfn_to_section(pfn)->pageblock_flags;
6431 return zone->pageblock_flags;
6432 #endif /* CONFIG_SPARSEMEM */
6435 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6437 #ifdef CONFIG_SPARSEMEM
6438 pfn &= (PAGES_PER_SECTION-1);
6439 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6441 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6442 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6443 #endif /* CONFIG_SPARSEMEM */
6447 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6448 * @page: The page within the block of interest
6449 * @pfn: The target page frame number
6450 * @end_bitidx: The last bit of interest to retrieve
6451 * @mask: mask of bits that the caller is interested in
6453 * Return: pageblock_bits flags
6455 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6456 unsigned long end_bitidx,
6460 unsigned long *bitmap;
6461 unsigned long bitidx, word_bitidx;
6464 zone = page_zone(page);
6465 bitmap = get_pageblock_bitmap(zone, pfn);
6466 bitidx = pfn_to_bitidx(zone, pfn);
6467 word_bitidx = bitidx / BITS_PER_LONG;
6468 bitidx &= (BITS_PER_LONG-1);
6470 word = bitmap[word_bitidx];
6471 bitidx += end_bitidx;
6472 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6476 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6477 * @page: The page within the block of interest
6478 * @flags: The flags to set
6479 * @pfn: The target page frame number
6480 * @end_bitidx: The last bit of interest
6481 * @mask: mask of bits that the caller is interested in
6483 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6485 unsigned long end_bitidx,
6489 unsigned long *bitmap;
6490 unsigned long bitidx, word_bitidx;
6491 unsigned long old_word, word;
6493 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6495 zone = page_zone(page);
6496 bitmap = get_pageblock_bitmap(zone, pfn);
6497 bitidx = pfn_to_bitidx(zone, pfn);
6498 word_bitidx = bitidx / BITS_PER_LONG;
6499 bitidx &= (BITS_PER_LONG-1);
6501 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6503 bitidx += end_bitidx;
6504 mask <<= (BITS_PER_LONG - bitidx - 1);
6505 flags <<= (BITS_PER_LONG - bitidx - 1);
6507 word = READ_ONCE(bitmap[word_bitidx]);
6509 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6510 if (word == old_word)
6517 * This function checks whether pageblock includes unmovable pages or not.
6518 * If @count is not zero, it is okay to include less @count unmovable pages
6520 * PageLRU check without isolation or lru_lock could race so that
6521 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6522 * expect this function should be exact.
6524 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6525 bool skip_hwpoisoned_pages)
6527 unsigned long pfn, iter, found;
6531 * For avoiding noise data, lru_add_drain_all() should be called
6532 * If ZONE_MOVABLE, the zone never contains unmovable pages
6534 if (zone_idx(zone) == ZONE_MOVABLE)
6536 mt = get_pageblock_migratetype(page);
6537 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6540 pfn = page_to_pfn(page);
6541 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6542 unsigned long check = pfn + iter;
6544 if (!pfn_valid_within(check))
6547 page = pfn_to_page(check);
6550 * Hugepages are not in LRU lists, but they're movable.
6551 * We need not scan over tail pages bacause we don't
6552 * handle each tail page individually in migration.
6554 if (PageHuge(page)) {
6555 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6560 * We can't use page_count without pin a page
6561 * because another CPU can free compound page.
6562 * This check already skips compound tails of THP
6563 * because their page->_count is zero at all time.
6565 if (!atomic_read(&page->_count)) {
6566 if (PageBuddy(page))
6567 iter += (1 << page_order(page)) - 1;
6572 * The HWPoisoned page may be not in buddy system, and
6573 * page_count() is not 0.
6575 if (skip_hwpoisoned_pages && PageHWPoison(page))
6581 * If there are RECLAIMABLE pages, we need to check
6582 * it. But now, memory offline itself doesn't call
6583 * shrink_node_slabs() and it still to be fixed.
6586 * If the page is not RAM, page_count()should be 0.
6587 * we don't need more check. This is an _used_ not-movable page.
6589 * The problematic thing here is PG_reserved pages. PG_reserved
6590 * is set to both of a memory hole page and a _used_ kernel
6599 bool is_pageblock_removable_nolock(struct page *page)
6605 * We have to be careful here because we are iterating over memory
6606 * sections which are not zone aware so we might end up outside of
6607 * the zone but still within the section.
6608 * We have to take care about the node as well. If the node is offline
6609 * its NODE_DATA will be NULL - see page_zone.
6611 if (!node_online(page_to_nid(page)))
6614 zone = page_zone(page);
6615 pfn = page_to_pfn(page);
6616 if (!zone_spans_pfn(zone, pfn))
6619 return !has_unmovable_pages(zone, page, 0, true);
6624 static unsigned long pfn_max_align_down(unsigned long pfn)
6626 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6627 pageblock_nr_pages) - 1);
6630 static unsigned long pfn_max_align_up(unsigned long pfn)
6632 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6633 pageblock_nr_pages));
6636 /* [start, end) must belong to a single zone. */
6637 static int __alloc_contig_migrate_range(struct compact_control *cc,
6638 unsigned long start, unsigned long end)
6640 /* This function is based on compact_zone() from compaction.c. */
6641 unsigned long nr_reclaimed;
6642 unsigned long pfn = start;
6643 unsigned int tries = 0;
6648 while (pfn < end || !list_empty(&cc->migratepages)) {
6649 if (fatal_signal_pending(current)) {
6654 if (list_empty(&cc->migratepages)) {
6655 cc->nr_migratepages = 0;
6656 pfn = isolate_migratepages_range(cc, pfn, end);
6662 } else if (++tries == 5) {
6663 ret = ret < 0 ? ret : -EBUSY;
6667 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6669 cc->nr_migratepages -= nr_reclaimed;
6671 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6672 NULL, 0, cc->mode, MR_CMA);
6675 putback_movable_pages(&cc->migratepages);
6682 * alloc_contig_range() -- tries to allocate given range of pages
6683 * @start: start PFN to allocate
6684 * @end: one-past-the-last PFN to allocate
6685 * @migratetype: migratetype of the underlaying pageblocks (either
6686 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6687 * in range must have the same migratetype and it must
6688 * be either of the two.
6690 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6691 * aligned, however it's the caller's responsibility to guarantee that
6692 * we are the only thread that changes migrate type of pageblocks the
6695 * The PFN range must belong to a single zone.
6697 * Returns zero on success or negative error code. On success all
6698 * pages which PFN is in [start, end) are allocated for the caller and
6699 * need to be freed with free_contig_range().
6701 int alloc_contig_range(unsigned long start, unsigned long end,
6702 unsigned migratetype)
6704 unsigned long outer_start, outer_end;
6708 struct compact_control cc = {
6709 .nr_migratepages = 0,
6711 .zone = page_zone(pfn_to_page(start)),
6712 .mode = MIGRATE_SYNC,
6713 .ignore_skip_hint = true,
6715 INIT_LIST_HEAD(&cc.migratepages);
6718 * What we do here is we mark all pageblocks in range as
6719 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6720 * have different sizes, and due to the way page allocator
6721 * work, we align the range to biggest of the two pages so
6722 * that page allocator won't try to merge buddies from
6723 * different pageblocks and change MIGRATE_ISOLATE to some
6724 * other migration type.
6726 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6727 * migrate the pages from an unaligned range (ie. pages that
6728 * we are interested in). This will put all the pages in
6729 * range back to page allocator as MIGRATE_ISOLATE.
6731 * When this is done, we take the pages in range from page
6732 * allocator removing them from the buddy system. This way
6733 * page allocator will never consider using them.
6735 * This lets us mark the pageblocks back as
6736 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6737 * aligned range but not in the unaligned, original range are
6738 * put back to page allocator so that buddy can use them.
6741 ret = start_isolate_page_range(pfn_max_align_down(start),
6742 pfn_max_align_up(end), migratetype,
6747 ret = __alloc_contig_migrate_range(&cc, start, end);
6752 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6753 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6754 * more, all pages in [start, end) are free in page allocator.
6755 * What we are going to do is to allocate all pages from
6756 * [start, end) (that is remove them from page allocator).
6758 * The only problem is that pages at the beginning and at the
6759 * end of interesting range may be not aligned with pages that
6760 * page allocator holds, ie. they can be part of higher order
6761 * pages. Because of this, we reserve the bigger range and
6762 * once this is done free the pages we are not interested in.
6764 * We don't have to hold zone->lock here because the pages are
6765 * isolated thus they won't get removed from buddy.
6768 lru_add_drain_all();
6769 drain_all_pages(cc.zone);
6772 outer_start = start;
6773 while (!PageBuddy(pfn_to_page(outer_start))) {
6774 if (++order >= MAX_ORDER) {
6778 outer_start &= ~0UL << order;
6781 /* Make sure the range is really isolated. */
6782 if (test_pages_isolated(outer_start, end, false)) {
6783 pr_info("%s: [%lx, %lx) PFNs busy\n",
6784 __func__, outer_start, end);
6789 /* Grab isolated pages from freelists. */
6790 outer_end = isolate_freepages_range(&cc, outer_start, end);
6796 /* Free head and tail (if any) */
6797 if (start != outer_start)
6798 free_contig_range(outer_start, start - outer_start);
6799 if (end != outer_end)
6800 free_contig_range(end, outer_end - end);
6803 undo_isolate_page_range(pfn_max_align_down(start),
6804 pfn_max_align_up(end), migratetype);
6808 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6810 unsigned int count = 0;
6812 for (; nr_pages--; pfn++) {
6813 struct page *page = pfn_to_page(pfn);
6815 count += page_count(page) != 1;
6818 WARN(count != 0, "%d pages are still in use!\n", count);
6822 #ifdef CONFIG_MEMORY_HOTPLUG
6824 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6825 * page high values need to be recalulated.
6827 void __meminit zone_pcp_update(struct zone *zone)
6830 mutex_lock(&pcp_batch_high_lock);
6831 for_each_possible_cpu(cpu)
6832 pageset_set_high_and_batch(zone,
6833 per_cpu_ptr(zone->pageset, cpu));
6834 mutex_unlock(&pcp_batch_high_lock);
6838 void zone_pcp_reset(struct zone *zone)
6840 unsigned long flags;
6842 struct per_cpu_pageset *pset;
6844 /* avoid races with drain_pages() */
6845 local_irq_save(flags);
6846 if (zone->pageset != &boot_pageset) {
6847 for_each_online_cpu(cpu) {
6848 pset = per_cpu_ptr(zone->pageset, cpu);
6849 drain_zonestat(zone, pset);
6851 free_percpu(zone->pageset);
6852 zone->pageset = &boot_pageset;
6854 local_irq_restore(flags);
6857 #ifdef CONFIG_MEMORY_HOTREMOVE
6859 * All pages in the range must be isolated before calling this.
6862 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6866 unsigned int order, i;
6868 unsigned long flags;
6869 /* find the first valid pfn */
6870 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6875 zone = page_zone(pfn_to_page(pfn));
6876 spin_lock_irqsave(&zone->lock, flags);
6878 while (pfn < end_pfn) {
6879 if (!pfn_valid(pfn)) {
6883 page = pfn_to_page(pfn);
6885 * The HWPoisoned page may be not in buddy system, and
6886 * page_count() is not 0.
6888 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6890 SetPageReserved(page);
6894 BUG_ON(page_count(page));
6895 BUG_ON(!PageBuddy(page));
6896 order = page_order(page);
6897 #ifdef CONFIG_DEBUG_VM
6898 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6899 pfn, 1 << order, end_pfn);
6901 list_del(&page->lru);
6902 rmv_page_order(page);
6903 zone->free_area[order].nr_free--;
6904 for (i = 0; i < (1 << order); i++)
6905 SetPageReserved((page+i));
6906 pfn += (1 << order);
6908 spin_unlock_irqrestore(&zone->lock, flags);
6912 #ifdef CONFIG_MEMORY_FAILURE
6913 bool is_free_buddy_page(struct page *page)
6915 struct zone *zone = page_zone(page);
6916 unsigned long pfn = page_to_pfn(page);
6917 unsigned long flags;
6920 spin_lock_irqsave(&zone->lock, flags);
6921 for (order = 0; order < MAX_ORDER; order++) {
6922 struct page *page_head = page - (pfn & ((1 << order) - 1));
6924 if (PageBuddy(page_head) && page_order(page_head) >= order)
6927 spin_unlock_irqrestore(&zone->lock, flags);
6929 return order < MAX_ORDER;