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>
65 #include <asm/sections.h>
66 #include <asm/tlbflush.h>
67 #include <asm/div64.h>
70 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
71 static DEFINE_MUTEX(pcp_batch_high_lock);
72 #define MIN_PERCPU_PAGELIST_FRACTION (8)
74 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
75 DEFINE_PER_CPU(int, numa_node);
76 EXPORT_PER_CPU_SYMBOL(numa_node);
79 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
81 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
82 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
83 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
84 * defined in <linux/topology.h>.
86 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
87 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
88 int _node_numa_mem_[MAX_NUMNODES];
92 * Array of node states.
94 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
95 [N_POSSIBLE] = NODE_MASK_ALL,
96 [N_ONLINE] = { { [0] = 1UL } },
98 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
100 [N_HIGH_MEMORY] = { { [0] = 1UL } },
102 #ifdef CONFIG_MOVABLE_NODE
103 [N_MEMORY] = { { [0] = 1UL } },
105 [N_CPU] = { { [0] = 1UL } },
108 EXPORT_SYMBOL(node_states);
110 /* Protect totalram_pages and zone->managed_pages */
111 static DEFINE_SPINLOCK(managed_page_count_lock);
113 unsigned long totalram_pages __read_mostly;
114 unsigned long totalreserve_pages __read_mostly;
115 unsigned long totalcma_pages __read_mostly;
117 * When calculating the number of globally allowed dirty pages, there
118 * is a certain number of per-zone reserves that should not be
119 * considered dirtyable memory. This is the sum of those reserves
120 * over all existing zones that contribute dirtyable memory.
122 unsigned long dirty_balance_reserve __read_mostly;
124 int percpu_pagelist_fraction;
125 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
127 #ifdef CONFIG_PM_SLEEP
129 * The following functions are used by the suspend/hibernate code to temporarily
130 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
131 * while devices are suspended. To avoid races with the suspend/hibernate code,
132 * they should always be called with pm_mutex held (gfp_allowed_mask also should
133 * only be modified with pm_mutex held, unless the suspend/hibernate code is
134 * guaranteed not to run in parallel with that modification).
137 static gfp_t saved_gfp_mask;
139 void pm_restore_gfp_mask(void)
141 WARN_ON(!mutex_is_locked(&pm_mutex));
142 if (saved_gfp_mask) {
143 gfp_allowed_mask = saved_gfp_mask;
148 void pm_restrict_gfp_mask(void)
150 WARN_ON(!mutex_is_locked(&pm_mutex));
151 WARN_ON(saved_gfp_mask);
152 saved_gfp_mask = gfp_allowed_mask;
153 gfp_allowed_mask &= ~GFP_IOFS;
156 bool pm_suspended_storage(void)
158 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
162 #endif /* CONFIG_PM_SLEEP */
164 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
165 int pageblock_order __read_mostly;
168 static void __free_pages_ok(struct page *page, unsigned int order);
171 * results with 256, 32 in the lowmem_reserve sysctl:
172 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
173 * 1G machine -> (16M dma, 784M normal, 224M high)
174 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
175 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
176 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
178 * TBD: should special case ZONE_DMA32 machines here - in those we normally
179 * don't need any ZONE_NORMAL reservation
181 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
182 #ifdef CONFIG_ZONE_DMA
185 #ifdef CONFIG_ZONE_DMA32
188 #ifdef CONFIG_HIGHMEM
194 EXPORT_SYMBOL(totalram_pages);
196 static char * const zone_names[MAX_NR_ZONES] = {
197 #ifdef CONFIG_ZONE_DMA
200 #ifdef CONFIG_ZONE_DMA32
204 #ifdef CONFIG_HIGHMEM
210 int min_free_kbytes = 1024;
211 int user_min_free_kbytes = -1;
213 static unsigned long __meminitdata nr_kernel_pages;
214 static unsigned long __meminitdata nr_all_pages;
215 static unsigned long __meminitdata dma_reserve;
217 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
218 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
219 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
220 static unsigned long __initdata required_kernelcore;
221 static unsigned long __initdata required_movablecore;
222 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
224 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
226 EXPORT_SYMBOL(movable_zone);
227 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
230 int nr_node_ids __read_mostly = MAX_NUMNODES;
231 int nr_online_nodes __read_mostly = 1;
232 EXPORT_SYMBOL(nr_node_ids);
233 EXPORT_SYMBOL(nr_online_nodes);
236 int page_group_by_mobility_disabled __read_mostly;
238 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
239 static inline void reset_deferred_meminit(pg_data_t *pgdat)
241 pgdat->first_deferred_pfn = ULONG_MAX;
244 /* Returns true if the struct page for the pfn is uninitialised */
245 static inline bool __defermem_init early_page_uninitialised(unsigned long pfn)
247 int nid = early_pfn_to_nid(pfn);
249 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
256 * Returns false when the remaining initialisation should be deferred until
257 * later in the boot cycle when it can be parallelised.
259 static inline bool update_defer_init(pg_data_t *pgdat,
260 unsigned long pfn, unsigned long zone_end,
261 unsigned long *nr_initialised)
263 /* Always populate low zones for address-contrained allocations */
264 if (zone_end < pgdat_end_pfn(pgdat))
267 /* Initialise at least 2G of the highest zone */
269 if (*nr_initialised > (2UL << (30 - PAGE_SHIFT)) &&
270 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
271 pgdat->first_deferred_pfn = pfn;
278 static inline void reset_deferred_meminit(pg_data_t *pgdat)
282 static inline bool early_page_uninitialised(unsigned long pfn)
287 static inline bool update_defer_init(pg_data_t *pgdat,
288 unsigned long pfn, unsigned long zone_end,
289 unsigned long *nr_initialised)
296 void set_pageblock_migratetype(struct page *page, int migratetype)
298 if (unlikely(page_group_by_mobility_disabled &&
299 migratetype < MIGRATE_PCPTYPES))
300 migratetype = MIGRATE_UNMOVABLE;
302 set_pageblock_flags_group(page, (unsigned long)migratetype,
303 PB_migrate, PB_migrate_end);
306 #ifdef CONFIG_DEBUG_VM
307 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
311 unsigned long pfn = page_to_pfn(page);
312 unsigned long sp, start_pfn;
315 seq = zone_span_seqbegin(zone);
316 start_pfn = zone->zone_start_pfn;
317 sp = zone->spanned_pages;
318 if (!zone_spans_pfn(zone, pfn))
320 } while (zone_span_seqretry(zone, seq));
323 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
324 pfn, zone_to_nid(zone), zone->name,
325 start_pfn, start_pfn + sp);
330 static int page_is_consistent(struct zone *zone, struct page *page)
332 if (!pfn_valid_within(page_to_pfn(page)))
334 if (zone != page_zone(page))
340 * Temporary debugging check for pages not lying within a given zone.
342 static int bad_range(struct zone *zone, struct page *page)
344 if (page_outside_zone_boundaries(zone, page))
346 if (!page_is_consistent(zone, page))
352 static inline int bad_range(struct zone *zone, struct page *page)
358 static void bad_page(struct page *page, const char *reason,
359 unsigned long bad_flags)
361 static unsigned long resume;
362 static unsigned long nr_shown;
363 static unsigned long nr_unshown;
365 /* Don't complain about poisoned pages */
366 if (PageHWPoison(page)) {
367 page_mapcount_reset(page); /* remove PageBuddy */
372 * Allow a burst of 60 reports, then keep quiet for that minute;
373 * or allow a steady drip of one report per second.
375 if (nr_shown == 60) {
376 if (time_before(jiffies, resume)) {
382 "BUG: Bad page state: %lu messages suppressed\n",
389 resume = jiffies + 60 * HZ;
391 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
392 current->comm, page_to_pfn(page));
393 dump_page_badflags(page, reason, bad_flags);
398 /* Leave bad fields for debug, except PageBuddy could make trouble */
399 page_mapcount_reset(page); /* remove PageBuddy */
400 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
404 * Higher-order pages are called "compound pages". They are structured thusly:
406 * The first PAGE_SIZE page is called the "head page".
408 * The remaining PAGE_SIZE pages are called "tail pages".
410 * All pages have PG_compound set. All tail pages have their ->first_page
411 * pointing at the head page.
413 * The first tail page's ->lru.next holds the address of the compound page's
414 * put_page() function. Its ->lru.prev holds the order of allocation.
415 * This usage means that zero-order pages may not be compound.
418 static void free_compound_page(struct page *page)
420 __free_pages_ok(page, compound_order(page));
423 void prep_compound_page(struct page *page, unsigned long order)
426 int nr_pages = 1 << order;
428 set_compound_page_dtor(page, free_compound_page);
429 set_compound_order(page, order);
431 for (i = 1; i < nr_pages; i++) {
432 struct page *p = page + i;
433 set_page_count(p, 0);
434 p->first_page = page;
435 /* Make sure p->first_page is always valid for PageTail() */
441 #ifdef CONFIG_DEBUG_PAGEALLOC
442 unsigned int _debug_guardpage_minorder;
443 bool _debug_pagealloc_enabled __read_mostly;
444 bool _debug_guardpage_enabled __read_mostly;
446 static int __init early_debug_pagealloc(char *buf)
451 if (strcmp(buf, "on") == 0)
452 _debug_pagealloc_enabled = true;
456 early_param("debug_pagealloc", early_debug_pagealloc);
458 static bool need_debug_guardpage(void)
460 /* If we don't use debug_pagealloc, we don't need guard page */
461 if (!debug_pagealloc_enabled())
467 static void init_debug_guardpage(void)
469 if (!debug_pagealloc_enabled())
472 _debug_guardpage_enabled = true;
475 struct page_ext_operations debug_guardpage_ops = {
476 .need = need_debug_guardpage,
477 .init = init_debug_guardpage,
480 static int __init debug_guardpage_minorder_setup(char *buf)
484 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
485 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
488 _debug_guardpage_minorder = res;
489 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
492 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
494 static inline void set_page_guard(struct zone *zone, struct page *page,
495 unsigned int order, int migratetype)
497 struct page_ext *page_ext;
499 if (!debug_guardpage_enabled())
502 page_ext = lookup_page_ext(page);
503 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
505 INIT_LIST_HEAD(&page->lru);
506 set_page_private(page, order);
507 /* Guard pages are not available for any usage */
508 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
511 static inline void clear_page_guard(struct zone *zone, struct page *page,
512 unsigned int order, int migratetype)
514 struct page_ext *page_ext;
516 if (!debug_guardpage_enabled())
519 page_ext = lookup_page_ext(page);
520 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
522 set_page_private(page, 0);
523 if (!is_migrate_isolate(migratetype))
524 __mod_zone_freepage_state(zone, (1 << order), migratetype);
527 struct page_ext_operations debug_guardpage_ops = { NULL, };
528 static inline void set_page_guard(struct zone *zone, struct page *page,
529 unsigned int order, int migratetype) {}
530 static inline void clear_page_guard(struct zone *zone, struct page *page,
531 unsigned int order, int migratetype) {}
534 static inline void set_page_order(struct page *page, unsigned int order)
536 set_page_private(page, order);
537 __SetPageBuddy(page);
540 static inline void rmv_page_order(struct page *page)
542 __ClearPageBuddy(page);
543 set_page_private(page, 0);
547 * This function checks whether a page is free && is the buddy
548 * we can do coalesce a page and its buddy if
549 * (a) the buddy is not in a hole &&
550 * (b) the buddy is in the buddy system &&
551 * (c) a page and its buddy have the same order &&
552 * (d) a page and its buddy are in the same zone.
554 * For recording whether a page is in the buddy system, we set ->_mapcount
555 * PAGE_BUDDY_MAPCOUNT_VALUE.
556 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
557 * serialized by zone->lock.
559 * For recording page's order, we use page_private(page).
561 static inline int page_is_buddy(struct page *page, struct page *buddy,
564 if (!pfn_valid_within(page_to_pfn(buddy)))
567 if (page_is_guard(buddy) && page_order(buddy) == order) {
568 if (page_zone_id(page) != page_zone_id(buddy))
571 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
576 if (PageBuddy(buddy) && page_order(buddy) == order) {
578 * zone check is done late to avoid uselessly
579 * calculating zone/node ids for pages that could
582 if (page_zone_id(page) != page_zone_id(buddy))
585 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
593 * Freeing function for a buddy system allocator.
595 * The concept of a buddy system is to maintain direct-mapped table
596 * (containing bit values) for memory blocks of various "orders".
597 * The bottom level table contains the map for the smallest allocatable
598 * units of memory (here, pages), and each level above it describes
599 * pairs of units from the levels below, hence, "buddies".
600 * At a high level, all that happens here is marking the table entry
601 * at the bottom level available, and propagating the changes upward
602 * as necessary, plus some accounting needed to play nicely with other
603 * parts of the VM system.
604 * At each level, we keep a list of pages, which are heads of continuous
605 * free pages of length of (1 << order) and marked with _mapcount
606 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
608 * So when we are allocating or freeing one, we can derive the state of the
609 * other. That is, if we allocate a small block, and both were
610 * free, the remainder of the region must be split into blocks.
611 * If a block is freed, and its buddy is also free, then this
612 * triggers coalescing into a block of larger size.
617 static inline void __free_one_page(struct page *page,
619 struct zone *zone, unsigned int order,
622 unsigned long page_idx;
623 unsigned long combined_idx;
624 unsigned long uninitialized_var(buddy_idx);
626 int max_order = MAX_ORDER;
628 VM_BUG_ON(!zone_is_initialized(zone));
629 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
631 VM_BUG_ON(migratetype == -1);
632 if (is_migrate_isolate(migratetype)) {
634 * We restrict max order of merging to prevent merge
635 * between freepages on isolate pageblock and normal
636 * pageblock. Without this, pageblock isolation
637 * could cause incorrect freepage accounting.
639 max_order = min(MAX_ORDER, pageblock_order + 1);
641 __mod_zone_freepage_state(zone, 1 << order, migratetype);
644 page_idx = pfn & ((1 << max_order) - 1);
646 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
647 VM_BUG_ON_PAGE(bad_range(zone, page), page);
649 while (order < max_order - 1) {
650 buddy_idx = __find_buddy_index(page_idx, order);
651 buddy = page + (buddy_idx - page_idx);
652 if (!page_is_buddy(page, buddy, order))
655 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
656 * merge with it and move up one order.
658 if (page_is_guard(buddy)) {
659 clear_page_guard(zone, buddy, order, migratetype);
661 list_del(&buddy->lru);
662 zone->free_area[order].nr_free--;
663 rmv_page_order(buddy);
665 combined_idx = buddy_idx & page_idx;
666 page = page + (combined_idx - page_idx);
667 page_idx = combined_idx;
670 set_page_order(page, order);
673 * If this is not the largest possible page, check if the buddy
674 * of the next-highest order is free. If it is, it's possible
675 * that pages are being freed that will coalesce soon. In case,
676 * that is happening, add the free page to the tail of the list
677 * so it's less likely to be used soon and more likely to be merged
678 * as a higher order page
680 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
681 struct page *higher_page, *higher_buddy;
682 combined_idx = buddy_idx & page_idx;
683 higher_page = page + (combined_idx - page_idx);
684 buddy_idx = __find_buddy_index(combined_idx, order + 1);
685 higher_buddy = higher_page + (buddy_idx - combined_idx);
686 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
687 list_add_tail(&page->lru,
688 &zone->free_area[order].free_list[migratetype]);
693 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
695 zone->free_area[order].nr_free++;
698 static inline int free_pages_check(struct page *page)
700 const char *bad_reason = NULL;
701 unsigned long bad_flags = 0;
703 if (unlikely(page_mapcount(page)))
704 bad_reason = "nonzero mapcount";
705 if (unlikely(page->mapping != NULL))
706 bad_reason = "non-NULL mapping";
707 if (unlikely(atomic_read(&page->_count) != 0))
708 bad_reason = "nonzero _count";
709 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
710 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
711 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
714 if (unlikely(page->mem_cgroup))
715 bad_reason = "page still charged to cgroup";
717 if (unlikely(bad_reason)) {
718 bad_page(page, bad_reason, bad_flags);
721 page_cpupid_reset_last(page);
722 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
723 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
728 * Frees a number of pages from the PCP lists
729 * Assumes all pages on list are in same zone, and of same order.
730 * count is the number of pages to free.
732 * If the zone was previously in an "all pages pinned" state then look to
733 * see if this freeing clears that state.
735 * And clear the zone's pages_scanned counter, to hold off the "all pages are
736 * pinned" detection logic.
738 static void free_pcppages_bulk(struct zone *zone, int count,
739 struct per_cpu_pages *pcp)
744 unsigned long nr_scanned;
746 spin_lock(&zone->lock);
747 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
749 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
753 struct list_head *list;
756 * Remove pages from lists in a round-robin fashion. A
757 * batch_free count is maintained that is incremented when an
758 * empty list is encountered. This is so more pages are freed
759 * off fuller lists instead of spinning excessively around empty
764 if (++migratetype == MIGRATE_PCPTYPES)
766 list = &pcp->lists[migratetype];
767 } while (list_empty(list));
769 /* This is the only non-empty list. Free them all. */
770 if (batch_free == MIGRATE_PCPTYPES)
771 batch_free = to_free;
774 int mt; /* migratetype of the to-be-freed page */
776 page = list_entry(list->prev, struct page, lru);
777 /* must delete as __free_one_page list manipulates */
778 list_del(&page->lru);
779 mt = get_freepage_migratetype(page);
780 if (unlikely(has_isolate_pageblock(zone)))
781 mt = get_pageblock_migratetype(page);
783 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
784 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
785 trace_mm_page_pcpu_drain(page, 0, mt);
786 } while (--to_free && --batch_free && !list_empty(list));
788 spin_unlock(&zone->lock);
791 static void free_one_page(struct zone *zone,
792 struct page *page, unsigned long pfn,
796 unsigned long nr_scanned;
797 spin_lock(&zone->lock);
798 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
800 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
802 if (unlikely(has_isolate_pageblock(zone) ||
803 is_migrate_isolate(migratetype))) {
804 migratetype = get_pfnblock_migratetype(page, pfn);
806 __free_one_page(page, pfn, zone, order, migratetype);
807 spin_unlock(&zone->lock);
810 static int free_tail_pages_check(struct page *head_page, struct page *page)
812 if (!IS_ENABLED(CONFIG_DEBUG_VM))
814 if (unlikely(!PageTail(page))) {
815 bad_page(page, "PageTail not set", 0);
818 if (unlikely(page->first_page != head_page)) {
819 bad_page(page, "first_page not consistent", 0);
825 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
826 unsigned long zone, int nid)
828 struct zone *z = &NODE_DATA(nid)->node_zones[zone];
830 set_page_links(page, zone, nid, pfn);
831 mminit_verify_page_links(page, zone, nid, pfn);
832 init_page_count(page);
833 page_mapcount_reset(page);
834 page_cpupid_reset_last(page);
837 * Mark the block movable so that blocks are reserved for
838 * movable at startup. This will force kernel allocations
839 * to reserve their blocks rather than leaking throughout
840 * the address space during boot when many long-lived
841 * kernel allocations are made. Later some blocks near
842 * the start are marked MIGRATE_RESERVE by
843 * setup_zone_migrate_reserve()
845 * bitmap is created for zone's valid pfn range. but memmap
846 * can be created for invalid pages (for alignment)
847 * check here not to call set_pageblock_migratetype() against
850 if ((z->zone_start_pfn <= pfn)
851 && (pfn < zone_end_pfn(z))
852 && !(pfn & (pageblock_nr_pages - 1)))
853 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
855 INIT_LIST_HEAD(&page->lru);
856 #ifdef WANT_PAGE_VIRTUAL
857 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
858 if (!is_highmem_idx(zone))
859 set_page_address(page, __va(pfn << PAGE_SHIFT));
863 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
866 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
870 * Initialised pages do not have PageReserved set. This function is
871 * called for each range allocated by the bootmem allocator and
872 * marks the pages PageReserved. The remaining valid pages are later
873 * sent to the buddy page allocator.
875 void reserve_bootmem_region(unsigned long start, unsigned long end)
877 unsigned long start_pfn = PFN_DOWN(start);
878 unsigned long end_pfn = PFN_UP(end);
880 for (; start_pfn < end_pfn; start_pfn++)
881 if (pfn_valid(start_pfn))
882 SetPageReserved(pfn_to_page(start_pfn));
885 static bool free_pages_prepare(struct page *page, unsigned int order)
887 bool compound = PageCompound(page);
890 VM_BUG_ON_PAGE(PageTail(page), page);
891 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
893 trace_mm_page_free(page, order);
894 kmemcheck_free_shadow(page, order);
895 kasan_free_pages(page, order);
898 page->mapping = NULL;
899 bad += free_pages_check(page);
900 for (i = 1; i < (1 << order); i++) {
902 bad += free_tail_pages_check(page, page + i);
903 bad += free_pages_check(page + i);
908 reset_page_owner(page, order);
910 if (!PageHighMem(page)) {
911 debug_check_no_locks_freed(page_address(page),
913 debug_check_no_obj_freed(page_address(page),
916 arch_free_page(page, order);
917 kernel_map_pages(page, 1 << order, 0);
922 static void __free_pages_ok(struct page *page, unsigned int order)
926 unsigned long pfn = page_to_pfn(page);
928 if (!free_pages_prepare(page, order))
931 migratetype = get_pfnblock_migratetype(page, pfn);
932 local_irq_save(flags);
933 __count_vm_events(PGFREE, 1 << order);
934 set_freepage_migratetype(page, migratetype);
935 free_one_page(page_zone(page), page, pfn, order, migratetype);
936 local_irq_restore(flags);
939 static void __defer_init __free_pages_boot_core(struct page *page,
940 unsigned long pfn, unsigned int order)
942 unsigned int nr_pages = 1 << order;
943 struct page *p = page;
947 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
949 __ClearPageReserved(p);
950 set_page_count(p, 0);
952 __ClearPageReserved(p);
953 set_page_count(p, 0);
955 page_zone(page)->managed_pages += nr_pages;
956 set_page_refcounted(page);
957 __free_pages(page, order);
960 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
961 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
962 /* Only safe to use early in boot when initialisation is single-threaded */
963 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
965 int __meminit early_pfn_to_nid(unsigned long pfn)
969 /* The system will behave unpredictably otherwise */
970 BUG_ON(system_state != SYSTEM_BOOTING);
972 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
980 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
981 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
982 struct mminit_pfnnid_cache *state)
986 nid = __early_pfn_to_nid(pfn, state);
987 if (nid >= 0 && nid != node)
992 /* Only safe to use early in boot when initialisation is single-threaded */
993 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
995 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1000 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1004 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1005 struct mminit_pfnnid_cache *state)
1012 void __defer_init __free_pages_bootmem(struct page *page, unsigned long pfn,
1015 if (early_page_uninitialised(pfn))
1017 return __free_pages_boot_core(page, pfn, order);
1021 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1022 void __init init_cma_reserved_pageblock(struct page *page)
1024 unsigned i = pageblock_nr_pages;
1025 struct page *p = page;
1028 __ClearPageReserved(p);
1029 set_page_count(p, 0);
1032 set_pageblock_migratetype(page, MIGRATE_CMA);
1034 if (pageblock_order >= MAX_ORDER) {
1035 i = pageblock_nr_pages;
1038 set_page_refcounted(p);
1039 __free_pages(p, MAX_ORDER - 1);
1040 p += MAX_ORDER_NR_PAGES;
1041 } while (i -= MAX_ORDER_NR_PAGES);
1043 set_page_refcounted(page);
1044 __free_pages(page, pageblock_order);
1047 adjust_managed_page_count(page, pageblock_nr_pages);
1052 * The order of subdivision here is critical for the IO subsystem.
1053 * Please do not alter this order without good reasons and regression
1054 * testing. Specifically, as large blocks of memory are subdivided,
1055 * the order in which smaller blocks are delivered depends on the order
1056 * they're subdivided in this function. This is the primary factor
1057 * influencing the order in which pages are delivered to the IO
1058 * subsystem according to empirical testing, and this is also justified
1059 * by considering the behavior of a buddy system containing a single
1060 * large block of memory acted on by a series of small allocations.
1061 * This behavior is a critical factor in sglist merging's success.
1065 static inline void expand(struct zone *zone, struct page *page,
1066 int low, int high, struct free_area *area,
1069 unsigned long size = 1 << high;
1071 while (high > low) {
1075 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1077 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1078 debug_guardpage_enabled() &&
1079 high < debug_guardpage_minorder()) {
1081 * Mark as guard pages (or page), that will allow to
1082 * merge back to allocator when buddy will be freed.
1083 * Corresponding page table entries will not be touched,
1084 * pages will stay not present in virtual address space
1086 set_page_guard(zone, &page[size], high, migratetype);
1089 list_add(&page[size].lru, &area->free_list[migratetype]);
1091 set_page_order(&page[size], high);
1096 * This page is about to be returned from the page allocator
1098 static inline int check_new_page(struct page *page)
1100 const char *bad_reason = NULL;
1101 unsigned long bad_flags = 0;
1103 if (unlikely(page_mapcount(page)))
1104 bad_reason = "nonzero mapcount";
1105 if (unlikely(page->mapping != NULL))
1106 bad_reason = "non-NULL mapping";
1107 if (unlikely(atomic_read(&page->_count) != 0))
1108 bad_reason = "nonzero _count";
1109 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1110 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1111 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1114 if (unlikely(page->mem_cgroup))
1115 bad_reason = "page still charged to cgroup";
1117 if (unlikely(bad_reason)) {
1118 bad_page(page, bad_reason, bad_flags);
1124 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1129 for (i = 0; i < (1 << order); i++) {
1130 struct page *p = page + i;
1131 if (unlikely(check_new_page(p)))
1135 set_page_private(page, 0);
1136 set_page_refcounted(page);
1138 arch_alloc_page(page, order);
1139 kernel_map_pages(page, 1 << order, 1);
1140 kasan_alloc_pages(page, order);
1142 if (gfp_flags & __GFP_ZERO)
1143 for (i = 0; i < (1 << order); i++)
1144 clear_highpage(page + i);
1146 if (order && (gfp_flags & __GFP_COMP))
1147 prep_compound_page(page, order);
1149 set_page_owner(page, order, gfp_flags);
1152 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was necessary to
1153 * allocate the page. The expectation is that the caller is taking
1154 * steps that will free more memory. The caller should avoid the page
1155 * being used for !PFMEMALLOC purposes.
1157 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
1163 * Go through the free lists for the given migratetype and remove
1164 * the smallest available page from the freelists
1167 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1170 unsigned int current_order;
1171 struct free_area *area;
1174 /* Find a page of the appropriate size in the preferred list */
1175 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1176 area = &(zone->free_area[current_order]);
1177 if (list_empty(&area->free_list[migratetype]))
1180 page = list_entry(area->free_list[migratetype].next,
1182 list_del(&page->lru);
1183 rmv_page_order(page);
1185 expand(zone, page, order, current_order, area, migratetype);
1186 set_freepage_migratetype(page, migratetype);
1195 * This array describes the order lists are fallen back to when
1196 * the free lists for the desirable migrate type are depleted
1198 static int fallbacks[MIGRATE_TYPES][4] = {
1199 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1200 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1201 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
1203 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
1205 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
1206 #ifdef CONFIG_MEMORY_ISOLATION
1207 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
1212 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1215 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1218 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1219 unsigned int order) { return NULL; }
1223 * Move the free pages in a range to the free lists of the requested type.
1224 * Note that start_page and end_pages are not aligned on a pageblock
1225 * boundary. If alignment is required, use move_freepages_block()
1227 int move_freepages(struct zone *zone,
1228 struct page *start_page, struct page *end_page,
1232 unsigned long order;
1233 int pages_moved = 0;
1235 #ifndef CONFIG_HOLES_IN_ZONE
1237 * page_zone is not safe to call in this context when
1238 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1239 * anyway as we check zone boundaries in move_freepages_block().
1240 * Remove at a later date when no bug reports exist related to
1241 * grouping pages by mobility
1243 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1246 for (page = start_page; page <= end_page;) {
1247 /* Make sure we are not inadvertently changing nodes */
1248 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1250 if (!pfn_valid_within(page_to_pfn(page))) {
1255 if (!PageBuddy(page)) {
1260 order = page_order(page);
1261 list_move(&page->lru,
1262 &zone->free_area[order].free_list[migratetype]);
1263 set_freepage_migratetype(page, migratetype);
1265 pages_moved += 1 << order;
1271 int move_freepages_block(struct zone *zone, struct page *page,
1274 unsigned long start_pfn, end_pfn;
1275 struct page *start_page, *end_page;
1277 start_pfn = page_to_pfn(page);
1278 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1279 start_page = pfn_to_page(start_pfn);
1280 end_page = start_page + pageblock_nr_pages - 1;
1281 end_pfn = start_pfn + pageblock_nr_pages - 1;
1283 /* Do not cross zone boundaries */
1284 if (!zone_spans_pfn(zone, start_pfn))
1286 if (!zone_spans_pfn(zone, end_pfn))
1289 return move_freepages(zone, start_page, end_page, migratetype);
1292 static void change_pageblock_range(struct page *pageblock_page,
1293 int start_order, int migratetype)
1295 int nr_pageblocks = 1 << (start_order - pageblock_order);
1297 while (nr_pageblocks--) {
1298 set_pageblock_migratetype(pageblock_page, migratetype);
1299 pageblock_page += pageblock_nr_pages;
1304 * When we are falling back to another migratetype during allocation, try to
1305 * steal extra free pages from the same pageblocks to satisfy further
1306 * allocations, instead of polluting multiple pageblocks.
1308 * If we are stealing a relatively large buddy page, it is likely there will
1309 * be more free pages in the pageblock, so try to steal them all. For
1310 * reclaimable and unmovable allocations, we steal regardless of page size,
1311 * as fragmentation caused by those allocations polluting movable pageblocks
1312 * is worse than movable allocations stealing from unmovable and reclaimable
1315 static bool can_steal_fallback(unsigned int order, int start_mt)
1318 * Leaving this order check is intended, although there is
1319 * relaxed order check in next check. The reason is that
1320 * we can actually steal whole pageblock if this condition met,
1321 * but, below check doesn't guarantee it and that is just heuristic
1322 * so could be changed anytime.
1324 if (order >= pageblock_order)
1327 if (order >= pageblock_order / 2 ||
1328 start_mt == MIGRATE_RECLAIMABLE ||
1329 start_mt == MIGRATE_UNMOVABLE ||
1330 page_group_by_mobility_disabled)
1337 * This function implements actual steal behaviour. If order is large enough,
1338 * we can steal whole pageblock. If not, we first move freepages in this
1339 * pageblock and check whether half of pages are moved or not. If half of
1340 * pages are moved, we can change migratetype of pageblock and permanently
1341 * use it's pages as requested migratetype in the future.
1343 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1346 int current_order = page_order(page);
1349 /* Take ownership for orders >= pageblock_order */
1350 if (current_order >= pageblock_order) {
1351 change_pageblock_range(page, current_order, start_type);
1355 pages = move_freepages_block(zone, page, start_type);
1357 /* Claim the whole block if over half of it is free */
1358 if (pages >= (1 << (pageblock_order-1)) ||
1359 page_group_by_mobility_disabled)
1360 set_pageblock_migratetype(page, start_type);
1364 * Check whether there is a suitable fallback freepage with requested order.
1365 * If only_stealable is true, this function returns fallback_mt only if
1366 * we can steal other freepages all together. This would help to reduce
1367 * fragmentation due to mixed migratetype pages in one pageblock.
1369 int find_suitable_fallback(struct free_area *area, unsigned int order,
1370 int migratetype, bool only_stealable, bool *can_steal)
1375 if (area->nr_free == 0)
1380 fallback_mt = fallbacks[migratetype][i];
1381 if (fallback_mt == MIGRATE_RESERVE)
1384 if (list_empty(&area->free_list[fallback_mt]))
1387 if (can_steal_fallback(order, migratetype))
1390 if (!only_stealable)
1400 /* Remove an element from the buddy allocator from the fallback list */
1401 static inline struct page *
1402 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1404 struct free_area *area;
1405 unsigned int current_order;
1410 /* Find the largest possible block of pages in the other list */
1411 for (current_order = MAX_ORDER-1;
1412 current_order >= order && current_order <= MAX_ORDER-1;
1414 area = &(zone->free_area[current_order]);
1415 fallback_mt = find_suitable_fallback(area, current_order,
1416 start_migratetype, false, &can_steal);
1417 if (fallback_mt == -1)
1420 page = list_entry(area->free_list[fallback_mt].next,
1423 steal_suitable_fallback(zone, page, start_migratetype);
1425 /* Remove the page from the freelists */
1427 list_del(&page->lru);
1428 rmv_page_order(page);
1430 expand(zone, page, order, current_order, area,
1433 * The freepage_migratetype may differ from pageblock's
1434 * migratetype depending on the decisions in
1435 * try_to_steal_freepages(). This is OK as long as it
1436 * does not differ for MIGRATE_CMA pageblocks. For CMA
1437 * we need to make sure unallocated pages flushed from
1438 * pcp lists are returned to the correct freelist.
1440 set_freepage_migratetype(page, start_migratetype);
1442 trace_mm_page_alloc_extfrag(page, order, current_order,
1443 start_migratetype, fallback_mt);
1452 * Do the hard work of removing an element from the buddy allocator.
1453 * Call me with the zone->lock already held.
1455 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1461 page = __rmqueue_smallest(zone, order, migratetype);
1463 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1464 if (migratetype == MIGRATE_MOVABLE)
1465 page = __rmqueue_cma_fallback(zone, order);
1468 page = __rmqueue_fallback(zone, order, migratetype);
1471 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1472 * is used because __rmqueue_smallest is an inline function
1473 * and we want just one call site
1476 migratetype = MIGRATE_RESERVE;
1481 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1486 * Obtain a specified number of elements from the buddy allocator, all under
1487 * a single hold of the lock, for efficiency. Add them to the supplied list.
1488 * Returns the number of new pages which were placed at *list.
1490 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1491 unsigned long count, struct list_head *list,
1492 int migratetype, bool cold)
1496 spin_lock(&zone->lock);
1497 for (i = 0; i < count; ++i) {
1498 struct page *page = __rmqueue(zone, order, migratetype);
1499 if (unlikely(page == NULL))
1503 * Split buddy pages returned by expand() are received here
1504 * in physical page order. The page is added to the callers and
1505 * list and the list head then moves forward. From the callers
1506 * perspective, the linked list is ordered by page number in
1507 * some conditions. This is useful for IO devices that can
1508 * merge IO requests if the physical pages are ordered
1512 list_add(&page->lru, list);
1514 list_add_tail(&page->lru, list);
1516 if (is_migrate_cma(get_freepage_migratetype(page)))
1517 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1520 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1521 spin_unlock(&zone->lock);
1527 * Called from the vmstat counter updater to drain pagesets of this
1528 * currently executing processor on remote nodes after they have
1531 * Note that this function must be called with the thread pinned to
1532 * a single processor.
1534 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1536 unsigned long flags;
1537 int to_drain, batch;
1539 local_irq_save(flags);
1540 batch = READ_ONCE(pcp->batch);
1541 to_drain = min(pcp->count, batch);
1543 free_pcppages_bulk(zone, to_drain, pcp);
1544 pcp->count -= to_drain;
1546 local_irq_restore(flags);
1551 * Drain pcplists of the indicated processor and zone.
1553 * The processor must either be the current processor and the
1554 * thread pinned to the current processor or a processor that
1557 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1559 unsigned long flags;
1560 struct per_cpu_pageset *pset;
1561 struct per_cpu_pages *pcp;
1563 local_irq_save(flags);
1564 pset = per_cpu_ptr(zone->pageset, cpu);
1568 free_pcppages_bulk(zone, pcp->count, pcp);
1571 local_irq_restore(flags);
1575 * Drain pcplists of all zones on the indicated processor.
1577 * The processor must either be the current processor and the
1578 * thread pinned to the current processor or a processor that
1581 static void drain_pages(unsigned int cpu)
1585 for_each_populated_zone(zone) {
1586 drain_pages_zone(cpu, zone);
1591 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1593 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1594 * the single zone's pages.
1596 void drain_local_pages(struct zone *zone)
1598 int cpu = smp_processor_id();
1601 drain_pages_zone(cpu, zone);
1607 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1609 * When zone parameter is non-NULL, spill just the single zone's pages.
1611 * Note that this code is protected against sending an IPI to an offline
1612 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1613 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1614 * nothing keeps CPUs from showing up after we populated the cpumask and
1615 * before the call to on_each_cpu_mask().
1617 void drain_all_pages(struct zone *zone)
1622 * Allocate in the BSS so we wont require allocation in
1623 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1625 static cpumask_t cpus_with_pcps;
1628 * We don't care about racing with CPU hotplug event
1629 * as offline notification will cause the notified
1630 * cpu to drain that CPU pcps and on_each_cpu_mask
1631 * disables preemption as part of its processing
1633 for_each_online_cpu(cpu) {
1634 struct per_cpu_pageset *pcp;
1636 bool has_pcps = false;
1639 pcp = per_cpu_ptr(zone->pageset, cpu);
1643 for_each_populated_zone(z) {
1644 pcp = per_cpu_ptr(z->pageset, cpu);
1645 if (pcp->pcp.count) {
1653 cpumask_set_cpu(cpu, &cpus_with_pcps);
1655 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1657 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
1661 #ifdef CONFIG_HIBERNATION
1663 void mark_free_pages(struct zone *zone)
1665 unsigned long pfn, max_zone_pfn;
1666 unsigned long flags;
1667 unsigned int order, t;
1668 struct list_head *curr;
1670 if (zone_is_empty(zone))
1673 spin_lock_irqsave(&zone->lock, flags);
1675 max_zone_pfn = zone_end_pfn(zone);
1676 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1677 if (pfn_valid(pfn)) {
1678 struct page *page = pfn_to_page(pfn);
1680 if (!swsusp_page_is_forbidden(page))
1681 swsusp_unset_page_free(page);
1684 for_each_migratetype_order(order, t) {
1685 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1688 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1689 for (i = 0; i < (1UL << order); i++)
1690 swsusp_set_page_free(pfn_to_page(pfn + i));
1693 spin_unlock_irqrestore(&zone->lock, flags);
1695 #endif /* CONFIG_PM */
1698 * Free a 0-order page
1699 * cold == true ? free a cold page : free a hot page
1701 void free_hot_cold_page(struct page *page, bool cold)
1703 struct zone *zone = page_zone(page);
1704 struct per_cpu_pages *pcp;
1705 unsigned long flags;
1706 unsigned long pfn = page_to_pfn(page);
1709 if (!free_pages_prepare(page, 0))
1712 migratetype = get_pfnblock_migratetype(page, pfn);
1713 set_freepage_migratetype(page, migratetype);
1714 local_irq_save(flags);
1715 __count_vm_event(PGFREE);
1718 * We only track unmovable, reclaimable and movable on pcp lists.
1719 * Free ISOLATE pages back to the allocator because they are being
1720 * offlined but treat RESERVE as movable pages so we can get those
1721 * areas back if necessary. Otherwise, we may have to free
1722 * excessively into the page allocator
1724 if (migratetype >= MIGRATE_PCPTYPES) {
1725 if (unlikely(is_migrate_isolate(migratetype))) {
1726 free_one_page(zone, page, pfn, 0, migratetype);
1729 migratetype = MIGRATE_MOVABLE;
1732 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1734 list_add(&page->lru, &pcp->lists[migratetype]);
1736 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1738 if (pcp->count >= pcp->high) {
1739 unsigned long batch = READ_ONCE(pcp->batch);
1740 free_pcppages_bulk(zone, batch, pcp);
1741 pcp->count -= batch;
1745 local_irq_restore(flags);
1749 * Free a list of 0-order pages
1751 void free_hot_cold_page_list(struct list_head *list, bool cold)
1753 struct page *page, *next;
1755 list_for_each_entry_safe(page, next, list, lru) {
1756 trace_mm_page_free_batched(page, cold);
1757 free_hot_cold_page(page, cold);
1762 * split_page takes a non-compound higher-order page, and splits it into
1763 * n (1<<order) sub-pages: page[0..n]
1764 * Each sub-page must be freed individually.
1766 * Note: this is probably too low level an operation for use in drivers.
1767 * Please consult with lkml before using this in your driver.
1769 void split_page(struct page *page, unsigned int order)
1773 VM_BUG_ON_PAGE(PageCompound(page), page);
1774 VM_BUG_ON_PAGE(!page_count(page), page);
1776 #ifdef CONFIG_KMEMCHECK
1778 * Split shadow pages too, because free(page[0]) would
1779 * otherwise free the whole shadow.
1781 if (kmemcheck_page_is_tracked(page))
1782 split_page(virt_to_page(page[0].shadow), order);
1785 set_page_owner(page, 0, 0);
1786 for (i = 1; i < (1 << order); i++) {
1787 set_page_refcounted(page + i);
1788 set_page_owner(page + i, 0, 0);
1791 EXPORT_SYMBOL_GPL(split_page);
1793 int __isolate_free_page(struct page *page, unsigned int order)
1795 unsigned long watermark;
1799 BUG_ON(!PageBuddy(page));
1801 zone = page_zone(page);
1802 mt = get_pageblock_migratetype(page);
1804 if (!is_migrate_isolate(mt)) {
1805 /* Obey watermarks as if the page was being allocated */
1806 watermark = low_wmark_pages(zone) + (1 << order);
1807 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1810 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1813 /* Remove page from free list */
1814 list_del(&page->lru);
1815 zone->free_area[order].nr_free--;
1816 rmv_page_order(page);
1818 /* Set the pageblock if the isolated page is at least a pageblock */
1819 if (order >= pageblock_order - 1) {
1820 struct page *endpage = page + (1 << order) - 1;
1821 for (; page < endpage; page += pageblock_nr_pages) {
1822 int mt = get_pageblock_migratetype(page);
1823 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1824 set_pageblock_migratetype(page,
1829 set_page_owner(page, order, 0);
1830 return 1UL << order;
1834 * Similar to split_page except the page is already free. As this is only
1835 * being used for migration, the migratetype of the block also changes.
1836 * As this is called with interrupts disabled, the caller is responsible
1837 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1840 * Note: this is probably too low level an operation for use in drivers.
1841 * Please consult with lkml before using this in your driver.
1843 int split_free_page(struct page *page)
1848 order = page_order(page);
1850 nr_pages = __isolate_free_page(page, order);
1854 /* Split into individual pages */
1855 set_page_refcounted(page);
1856 split_page(page, order);
1861 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
1864 struct page *buffered_rmqueue(struct zone *preferred_zone,
1865 struct zone *zone, unsigned int order,
1866 gfp_t gfp_flags, int migratetype)
1868 unsigned long flags;
1870 bool cold = ((gfp_flags & __GFP_COLD) != 0);
1872 if (likely(order == 0)) {
1873 struct per_cpu_pages *pcp;
1874 struct list_head *list;
1876 local_irq_save(flags);
1877 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1878 list = &pcp->lists[migratetype];
1879 if (list_empty(list)) {
1880 pcp->count += rmqueue_bulk(zone, 0,
1883 if (unlikely(list_empty(list)))
1888 page = list_entry(list->prev, struct page, lru);
1890 page = list_entry(list->next, struct page, lru);
1892 list_del(&page->lru);
1895 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1897 * __GFP_NOFAIL is not to be used in new code.
1899 * All __GFP_NOFAIL callers should be fixed so that they
1900 * properly detect and handle allocation failures.
1902 * We most definitely don't want callers attempting to
1903 * allocate greater than order-1 page units with
1906 WARN_ON_ONCE(order > 1);
1908 spin_lock_irqsave(&zone->lock, flags);
1909 page = __rmqueue(zone, order, migratetype);
1910 spin_unlock(&zone->lock);
1913 __mod_zone_freepage_state(zone, -(1 << order),
1914 get_freepage_migratetype(page));
1917 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
1918 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
1919 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
1920 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
1922 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1923 zone_statistics(preferred_zone, zone, gfp_flags);
1924 local_irq_restore(flags);
1926 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1930 local_irq_restore(flags);
1934 #ifdef CONFIG_FAIL_PAGE_ALLOC
1937 struct fault_attr attr;
1939 u32 ignore_gfp_highmem;
1940 u32 ignore_gfp_wait;
1942 } fail_page_alloc = {
1943 .attr = FAULT_ATTR_INITIALIZER,
1944 .ignore_gfp_wait = 1,
1945 .ignore_gfp_highmem = 1,
1949 static int __init setup_fail_page_alloc(char *str)
1951 return setup_fault_attr(&fail_page_alloc.attr, str);
1953 __setup("fail_page_alloc=", setup_fail_page_alloc);
1955 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1957 if (order < fail_page_alloc.min_order)
1959 if (gfp_mask & __GFP_NOFAIL)
1961 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1963 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1966 return should_fail(&fail_page_alloc.attr, 1 << order);
1969 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1971 static int __init fail_page_alloc_debugfs(void)
1973 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1976 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1977 &fail_page_alloc.attr);
1979 return PTR_ERR(dir);
1981 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1982 &fail_page_alloc.ignore_gfp_wait))
1984 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1985 &fail_page_alloc.ignore_gfp_highmem))
1987 if (!debugfs_create_u32("min-order", mode, dir,
1988 &fail_page_alloc.min_order))
1993 debugfs_remove_recursive(dir);
1998 late_initcall(fail_page_alloc_debugfs);
2000 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2002 #else /* CONFIG_FAIL_PAGE_ALLOC */
2004 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2009 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2012 * Return true if free pages are above 'mark'. This takes into account the order
2013 * of the allocation.
2015 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2016 unsigned long mark, int classzone_idx, int alloc_flags,
2019 /* free_pages may go negative - that's OK */
2024 free_pages -= (1 << order) - 1;
2025 if (alloc_flags & ALLOC_HIGH)
2027 if (alloc_flags & ALLOC_HARDER)
2030 /* If allocation can't use CMA areas don't use free CMA pages */
2031 if (!(alloc_flags & ALLOC_CMA))
2032 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
2035 if (free_pages - free_cma <= min + z->lowmem_reserve[classzone_idx])
2037 for (o = 0; o < order; o++) {
2038 /* At the next order, this order's pages become unavailable */
2039 free_pages -= z->free_area[o].nr_free << o;
2041 /* Require fewer higher order pages to be free */
2044 if (free_pages <= min)
2050 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2051 int classzone_idx, int alloc_flags)
2053 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2054 zone_page_state(z, NR_FREE_PAGES));
2057 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2058 unsigned long mark, int classzone_idx, int alloc_flags)
2060 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2062 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2063 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2065 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2071 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
2072 * skip over zones that are not allowed by the cpuset, or that have
2073 * been recently (in last second) found to be nearly full. See further
2074 * comments in mmzone.h. Reduces cache footprint of zonelist scans
2075 * that have to skip over a lot of full or unallowed zones.
2077 * If the zonelist cache is present in the passed zonelist, then
2078 * returns a pointer to the allowed node mask (either the current
2079 * tasks mems_allowed, or node_states[N_MEMORY].)
2081 * If the zonelist cache is not available for this zonelist, does
2082 * nothing and returns NULL.
2084 * If the fullzones BITMAP in the zonelist cache is stale (more than
2085 * a second since last zap'd) then we zap it out (clear its bits.)
2087 * We hold off even calling zlc_setup, until after we've checked the
2088 * first zone in the zonelist, on the theory that most allocations will
2089 * be satisfied from that first zone, so best to examine that zone as
2090 * quickly as we can.
2092 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
2094 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2095 nodemask_t *allowednodes; /* zonelist_cache approximation */
2097 zlc = zonelist->zlcache_ptr;
2101 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
2102 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2103 zlc->last_full_zap = jiffies;
2106 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
2107 &cpuset_current_mems_allowed :
2108 &node_states[N_MEMORY];
2109 return allowednodes;
2113 * Given 'z' scanning a zonelist, run a couple of quick checks to see
2114 * if it is worth looking at further for free memory:
2115 * 1) Check that the zone isn't thought to be full (doesn't have its
2116 * bit set in the zonelist_cache fullzones BITMAP).
2117 * 2) Check that the zones node (obtained from the zonelist_cache
2118 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
2119 * Return true (non-zero) if zone is worth looking at further, or
2120 * else return false (zero) if it is not.
2122 * This check -ignores- the distinction between various watermarks,
2123 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
2124 * found to be full for any variation of these watermarks, it will
2125 * be considered full for up to one second by all requests, unless
2126 * we are so low on memory on all allowed nodes that we are forced
2127 * into the second scan of the zonelist.
2129 * In the second scan we ignore this zonelist cache and exactly
2130 * apply the watermarks to all zones, even it is slower to do so.
2131 * We are low on memory in the second scan, and should leave no stone
2132 * unturned looking for a free page.
2134 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
2135 nodemask_t *allowednodes)
2137 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2138 int i; /* index of *z in zonelist zones */
2139 int n; /* node that zone *z is on */
2141 zlc = zonelist->zlcache_ptr;
2145 i = z - zonelist->_zonerefs;
2148 /* This zone is worth trying if it is allowed but not full */
2149 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
2153 * Given 'z' scanning a zonelist, set the corresponding bit in
2154 * zlc->fullzones, so that subsequent attempts to allocate a page
2155 * from that zone don't waste time re-examining it.
2157 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
2159 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2160 int i; /* index of *z in zonelist zones */
2162 zlc = zonelist->zlcache_ptr;
2166 i = z - zonelist->_zonerefs;
2168 set_bit(i, zlc->fullzones);
2172 * clear all zones full, called after direct reclaim makes progress so that
2173 * a zone that was recently full is not skipped over for up to a second
2175 static void zlc_clear_zones_full(struct zonelist *zonelist)
2177 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2179 zlc = zonelist->zlcache_ptr;
2183 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2186 static bool zone_local(struct zone *local_zone, struct zone *zone)
2188 return local_zone->node == zone->node;
2191 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2193 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2197 #else /* CONFIG_NUMA */
2199 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
2204 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
2205 nodemask_t *allowednodes)
2210 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
2214 static void zlc_clear_zones_full(struct zonelist *zonelist)
2218 static bool zone_local(struct zone *local_zone, struct zone *zone)
2223 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2228 #endif /* CONFIG_NUMA */
2230 static void reset_alloc_batches(struct zone *preferred_zone)
2232 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2235 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2236 high_wmark_pages(zone) - low_wmark_pages(zone) -
2237 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2238 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2239 } while (zone++ != preferred_zone);
2243 * get_page_from_freelist goes through the zonelist trying to allocate
2246 static struct page *
2247 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2248 const struct alloc_context *ac)
2250 struct zonelist *zonelist = ac->zonelist;
2252 struct page *page = NULL;
2254 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
2255 int zlc_active = 0; /* set if using zonelist_cache */
2256 int did_zlc_setup = 0; /* just call zlc_setup() one time */
2257 bool consider_zone_dirty = (alloc_flags & ALLOC_WMARK_LOW) &&
2258 (gfp_mask & __GFP_WRITE);
2259 int nr_fair_skipped = 0;
2260 bool zonelist_rescan;
2263 zonelist_rescan = false;
2266 * Scan zonelist, looking for a zone with enough free.
2267 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2269 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2273 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2274 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2276 if (cpusets_enabled() &&
2277 (alloc_flags & ALLOC_CPUSET) &&
2278 !cpuset_zone_allowed(zone, gfp_mask))
2281 * Distribute pages in proportion to the individual
2282 * zone size to ensure fair page aging. The zone a
2283 * page was allocated in should have no effect on the
2284 * time the page has in memory before being reclaimed.
2286 if (alloc_flags & ALLOC_FAIR) {
2287 if (!zone_local(ac->preferred_zone, zone))
2289 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2295 * When allocating a page cache page for writing, we
2296 * want to get it from a zone that is within its dirty
2297 * limit, such that no single zone holds more than its
2298 * proportional share of globally allowed dirty pages.
2299 * The dirty limits take into account the zone's
2300 * lowmem reserves and high watermark so that kswapd
2301 * should be able to balance it without having to
2302 * write pages from its LRU list.
2304 * This may look like it could increase pressure on
2305 * lower zones by failing allocations in higher zones
2306 * before they are full. But the pages that do spill
2307 * over are limited as the lower zones are protected
2308 * by this very same mechanism. It should not become
2309 * a practical burden to them.
2311 * XXX: For now, allow allocations to potentially
2312 * exceed the per-zone dirty limit in the slowpath
2313 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2314 * which is important when on a NUMA setup the allowed
2315 * zones are together not big enough to reach the
2316 * global limit. The proper fix for these situations
2317 * will require awareness of zones in the
2318 * dirty-throttling and the flusher threads.
2320 if (consider_zone_dirty && !zone_dirty_ok(zone))
2323 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2324 if (!zone_watermark_ok(zone, order, mark,
2325 ac->classzone_idx, alloc_flags)) {
2328 /* Checked here to keep the fast path fast */
2329 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2330 if (alloc_flags & ALLOC_NO_WATERMARKS)
2333 if (IS_ENABLED(CONFIG_NUMA) &&
2334 !did_zlc_setup && nr_online_nodes > 1) {
2336 * we do zlc_setup if there are multiple nodes
2337 * and before considering the first zone allowed
2340 allowednodes = zlc_setup(zonelist, alloc_flags);
2345 if (zone_reclaim_mode == 0 ||
2346 !zone_allows_reclaim(ac->preferred_zone, zone))
2347 goto this_zone_full;
2350 * As we may have just activated ZLC, check if the first
2351 * eligible zone has failed zone_reclaim recently.
2353 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2354 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2357 ret = zone_reclaim(zone, gfp_mask, order);
2359 case ZONE_RECLAIM_NOSCAN:
2362 case ZONE_RECLAIM_FULL:
2363 /* scanned but unreclaimable */
2366 /* did we reclaim enough */
2367 if (zone_watermark_ok(zone, order, mark,
2368 ac->classzone_idx, alloc_flags))
2372 * Failed to reclaim enough to meet watermark.
2373 * Only mark the zone full if checking the min
2374 * watermark or if we failed to reclaim just
2375 * 1<<order pages or else the page allocator
2376 * fastpath will prematurely mark zones full
2377 * when the watermark is between the low and
2380 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2381 ret == ZONE_RECLAIM_SOME)
2382 goto this_zone_full;
2389 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2390 gfp_mask, ac->migratetype);
2392 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2397 if (IS_ENABLED(CONFIG_NUMA) && zlc_active)
2398 zlc_mark_zone_full(zonelist, z);
2402 * The first pass makes sure allocations are spread fairly within the
2403 * local node. However, the local node might have free pages left
2404 * after the fairness batches are exhausted, and remote zones haven't
2405 * even been considered yet. Try once more without fairness, and
2406 * include remote zones now, before entering the slowpath and waking
2407 * kswapd: prefer spilling to a remote zone over swapping locally.
2409 if (alloc_flags & ALLOC_FAIR) {
2410 alloc_flags &= ~ALLOC_FAIR;
2411 if (nr_fair_skipped) {
2412 zonelist_rescan = true;
2413 reset_alloc_batches(ac->preferred_zone);
2415 if (nr_online_nodes > 1)
2416 zonelist_rescan = true;
2419 if (unlikely(IS_ENABLED(CONFIG_NUMA) && zlc_active)) {
2420 /* Disable zlc cache for second zonelist scan */
2422 zonelist_rescan = true;
2425 if (zonelist_rescan)
2432 * Large machines with many possible nodes should not always dump per-node
2433 * meminfo in irq context.
2435 static inline bool should_suppress_show_mem(void)
2440 ret = in_interrupt();
2445 static DEFINE_RATELIMIT_STATE(nopage_rs,
2446 DEFAULT_RATELIMIT_INTERVAL,
2447 DEFAULT_RATELIMIT_BURST);
2449 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2451 unsigned int filter = SHOW_MEM_FILTER_NODES;
2453 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2454 debug_guardpage_minorder() > 0)
2458 * This documents exceptions given to allocations in certain
2459 * contexts that are allowed to allocate outside current's set
2462 if (!(gfp_mask & __GFP_NOMEMALLOC))
2463 if (test_thread_flag(TIF_MEMDIE) ||
2464 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2465 filter &= ~SHOW_MEM_FILTER_NODES;
2466 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2467 filter &= ~SHOW_MEM_FILTER_NODES;
2470 struct va_format vaf;
2473 va_start(args, fmt);
2478 pr_warn("%pV", &vaf);
2483 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2484 current->comm, order, gfp_mask);
2487 if (!should_suppress_show_mem())
2491 static inline struct page *
2492 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2493 const struct alloc_context *ac, unsigned long *did_some_progress)
2497 *did_some_progress = 0;
2500 * Acquire the oom lock. If that fails, somebody else is
2501 * making progress for us.
2503 if (!mutex_trylock(&oom_lock)) {
2504 *did_some_progress = 1;
2505 schedule_timeout_uninterruptible(1);
2510 * Go through the zonelist yet one more time, keep very high watermark
2511 * here, this is only to catch a parallel oom killing, we must fail if
2512 * we're still under heavy pressure.
2514 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2515 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2519 if (!(gfp_mask & __GFP_NOFAIL)) {
2520 /* Coredumps can quickly deplete all memory reserves */
2521 if (current->flags & PF_DUMPCORE)
2523 /* The OOM killer will not help higher order allocs */
2524 if (order > PAGE_ALLOC_COSTLY_ORDER)
2526 /* The OOM killer does not needlessly kill tasks for lowmem */
2527 if (ac->high_zoneidx < ZONE_NORMAL)
2529 /* The OOM killer does not compensate for IO-less reclaim */
2530 if (!(gfp_mask & __GFP_FS)) {
2532 * XXX: Page reclaim didn't yield anything,
2533 * and the OOM killer can't be invoked, but
2534 * keep looping as per tradition.
2536 *did_some_progress = 1;
2539 if (pm_suspended_storage())
2541 /* The OOM killer may not free memory on a specific node */
2542 if (gfp_mask & __GFP_THISNODE)
2545 /* Exhausted what can be done so it's blamo time */
2546 if (out_of_memory(ac->zonelist, gfp_mask, order, ac->nodemask, false)
2547 || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL))
2548 *did_some_progress = 1;
2550 mutex_unlock(&oom_lock);
2554 #ifdef CONFIG_COMPACTION
2555 /* Try memory compaction for high-order allocations before reclaim */
2556 static struct page *
2557 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2558 int alloc_flags, const struct alloc_context *ac,
2559 enum migrate_mode mode, int *contended_compaction,
2560 bool *deferred_compaction)
2562 unsigned long compact_result;
2568 current->flags |= PF_MEMALLOC;
2569 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2570 mode, contended_compaction);
2571 current->flags &= ~PF_MEMALLOC;
2573 switch (compact_result) {
2574 case COMPACT_DEFERRED:
2575 *deferred_compaction = true;
2577 case COMPACT_SKIPPED:
2584 * At least in one zone compaction wasn't deferred or skipped, so let's
2585 * count a compaction stall
2587 count_vm_event(COMPACTSTALL);
2589 page = get_page_from_freelist(gfp_mask, order,
2590 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2593 struct zone *zone = page_zone(page);
2595 zone->compact_blockskip_flush = false;
2596 compaction_defer_reset(zone, order, true);
2597 count_vm_event(COMPACTSUCCESS);
2602 * It's bad if compaction run occurs and fails. The most likely reason
2603 * is that pages exist, but not enough to satisfy watermarks.
2605 count_vm_event(COMPACTFAIL);
2612 static inline struct page *
2613 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2614 int alloc_flags, const struct alloc_context *ac,
2615 enum migrate_mode mode, int *contended_compaction,
2616 bool *deferred_compaction)
2620 #endif /* CONFIG_COMPACTION */
2622 /* Perform direct synchronous page reclaim */
2624 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2625 const struct alloc_context *ac)
2627 struct reclaim_state reclaim_state;
2632 /* We now go into synchronous reclaim */
2633 cpuset_memory_pressure_bump();
2634 current->flags |= PF_MEMALLOC;
2635 lockdep_set_current_reclaim_state(gfp_mask);
2636 reclaim_state.reclaimed_slab = 0;
2637 current->reclaim_state = &reclaim_state;
2639 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2642 current->reclaim_state = NULL;
2643 lockdep_clear_current_reclaim_state();
2644 current->flags &= ~PF_MEMALLOC;
2651 /* The really slow allocator path where we enter direct reclaim */
2652 static inline struct page *
2653 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2654 int alloc_flags, const struct alloc_context *ac,
2655 unsigned long *did_some_progress)
2657 struct page *page = NULL;
2658 bool drained = false;
2660 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2661 if (unlikely(!(*did_some_progress)))
2664 /* After successful reclaim, reconsider all zones for allocation */
2665 if (IS_ENABLED(CONFIG_NUMA))
2666 zlc_clear_zones_full(ac->zonelist);
2669 page = get_page_from_freelist(gfp_mask, order,
2670 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2673 * If an allocation failed after direct reclaim, it could be because
2674 * pages are pinned on the per-cpu lists. Drain them and try again
2676 if (!page && !drained) {
2677 drain_all_pages(NULL);
2686 * This is called in the allocator slow-path if the allocation request is of
2687 * sufficient urgency to ignore watermarks and take other desperate measures
2689 static inline struct page *
2690 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2691 const struct alloc_context *ac)
2696 page = get_page_from_freelist(gfp_mask, order,
2697 ALLOC_NO_WATERMARKS, ac);
2699 if (!page && gfp_mask & __GFP_NOFAIL)
2700 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC,
2702 } while (!page && (gfp_mask & __GFP_NOFAIL));
2707 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2712 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2713 ac->high_zoneidx, ac->nodemask)
2714 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2718 gfp_to_alloc_flags(gfp_t gfp_mask)
2720 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2721 const bool atomic = !(gfp_mask & (__GFP_WAIT | __GFP_NO_KSWAPD));
2723 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2724 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2727 * The caller may dip into page reserves a bit more if the caller
2728 * cannot run direct reclaim, or if the caller has realtime scheduling
2729 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2730 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2732 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2736 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2737 * if it can't schedule.
2739 if (!(gfp_mask & __GFP_NOMEMALLOC))
2740 alloc_flags |= ALLOC_HARDER;
2742 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2743 * comment for __cpuset_node_allowed().
2745 alloc_flags &= ~ALLOC_CPUSET;
2746 } else if (unlikely(rt_task(current)) && !in_interrupt())
2747 alloc_flags |= ALLOC_HARDER;
2749 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2750 if (gfp_mask & __GFP_MEMALLOC)
2751 alloc_flags |= ALLOC_NO_WATERMARKS;
2752 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2753 alloc_flags |= ALLOC_NO_WATERMARKS;
2754 else if (!in_interrupt() &&
2755 ((current->flags & PF_MEMALLOC) ||
2756 unlikely(test_thread_flag(TIF_MEMDIE))))
2757 alloc_flags |= ALLOC_NO_WATERMARKS;
2760 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2761 alloc_flags |= ALLOC_CMA;
2766 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2768 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2771 static inline struct page *
2772 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2773 struct alloc_context *ac)
2775 const gfp_t wait = gfp_mask & __GFP_WAIT;
2776 struct page *page = NULL;
2778 unsigned long pages_reclaimed = 0;
2779 unsigned long did_some_progress;
2780 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2781 bool deferred_compaction = false;
2782 int contended_compaction = COMPACT_CONTENDED_NONE;
2785 * In the slowpath, we sanity check order to avoid ever trying to
2786 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2787 * be using allocators in order of preference for an area that is
2790 if (order >= MAX_ORDER) {
2791 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2796 * If this allocation cannot block and it is for a specific node, then
2797 * fail early. There's no need to wakeup kswapd or retry for a
2798 * speculative node-specific allocation.
2800 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !wait)
2804 if (!(gfp_mask & __GFP_NO_KSWAPD))
2805 wake_all_kswapds(order, ac);
2808 * OK, we're below the kswapd watermark and have kicked background
2809 * reclaim. Now things get more complex, so set up alloc_flags according
2810 * to how we want to proceed.
2812 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2815 * Find the true preferred zone if the allocation is unconstrained by
2818 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
2819 struct zoneref *preferred_zoneref;
2820 preferred_zoneref = first_zones_zonelist(ac->zonelist,
2821 ac->high_zoneidx, NULL, &ac->preferred_zone);
2822 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
2825 /* This is the last chance, in general, before the goto nopage. */
2826 page = get_page_from_freelist(gfp_mask, order,
2827 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2831 /* Allocate without watermarks if the context allows */
2832 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2834 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2835 * the allocation is high priority and these type of
2836 * allocations are system rather than user orientated
2838 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
2840 page = __alloc_pages_high_priority(gfp_mask, order, ac);
2847 /* Atomic allocations - we can't balance anything */
2850 * All existing users of the deprecated __GFP_NOFAIL are
2851 * blockable, so warn of any new users that actually allow this
2852 * type of allocation to fail.
2854 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
2858 /* Avoid recursion of direct reclaim */
2859 if (current->flags & PF_MEMALLOC)
2862 /* Avoid allocations with no watermarks from looping endlessly */
2863 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2867 * Try direct compaction. The first pass is asynchronous. Subsequent
2868 * attempts after direct reclaim are synchronous
2870 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
2872 &contended_compaction,
2873 &deferred_compaction);
2877 /* Checks for THP-specific high-order allocations */
2878 if ((gfp_mask & GFP_TRANSHUGE) == GFP_TRANSHUGE) {
2880 * If compaction is deferred for high-order allocations, it is
2881 * because sync compaction recently failed. If this is the case
2882 * and the caller requested a THP allocation, we do not want
2883 * to heavily disrupt the system, so we fail the allocation
2884 * instead of entering direct reclaim.
2886 if (deferred_compaction)
2890 * In all zones where compaction was attempted (and not
2891 * deferred or skipped), lock contention has been detected.
2892 * For THP allocation we do not want to disrupt the others
2893 * so we fallback to base pages instead.
2895 if (contended_compaction == COMPACT_CONTENDED_LOCK)
2899 * If compaction was aborted due to need_resched(), we do not
2900 * want to further increase allocation latency, unless it is
2901 * khugepaged trying to collapse.
2903 if (contended_compaction == COMPACT_CONTENDED_SCHED
2904 && !(current->flags & PF_KTHREAD))
2909 * It can become very expensive to allocate transparent hugepages at
2910 * fault, so use asynchronous memory compaction for THP unless it is
2911 * khugepaged trying to collapse.
2913 if ((gfp_mask & GFP_TRANSHUGE) != GFP_TRANSHUGE ||
2914 (current->flags & PF_KTHREAD))
2915 migration_mode = MIGRATE_SYNC_LIGHT;
2917 /* Try direct reclaim and then allocating */
2918 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
2919 &did_some_progress);
2923 /* Do not loop if specifically requested */
2924 if (gfp_mask & __GFP_NORETRY)
2927 /* Keep reclaiming pages as long as there is reasonable progress */
2928 pages_reclaimed += did_some_progress;
2929 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
2930 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
2931 /* Wait for some write requests to complete then retry */
2932 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
2936 /* Reclaim has failed us, start killing things */
2937 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
2941 /* Retry as long as the OOM killer is making progress */
2942 if (did_some_progress)
2947 * High-order allocations do not necessarily loop after
2948 * direct reclaim and reclaim/compaction depends on compaction
2949 * being called after reclaim so call directly if necessary
2951 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
2953 &contended_compaction,
2954 &deferred_compaction);
2958 warn_alloc_failed(gfp_mask, order, NULL);
2964 * This is the 'heart' of the zoned buddy allocator.
2967 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2968 struct zonelist *zonelist, nodemask_t *nodemask)
2970 struct zoneref *preferred_zoneref;
2971 struct page *page = NULL;
2972 unsigned int cpuset_mems_cookie;
2973 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
2974 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
2975 struct alloc_context ac = {
2976 .high_zoneidx = gfp_zone(gfp_mask),
2977 .nodemask = nodemask,
2978 .migratetype = gfpflags_to_migratetype(gfp_mask),
2981 gfp_mask &= gfp_allowed_mask;
2983 lockdep_trace_alloc(gfp_mask);
2985 might_sleep_if(gfp_mask & __GFP_WAIT);
2987 if (should_fail_alloc_page(gfp_mask, order))
2991 * Check the zones suitable for the gfp_mask contain at least one
2992 * valid zone. It's possible to have an empty zonelist as a result
2993 * of __GFP_THISNODE and a memoryless node
2995 if (unlikely(!zonelist->_zonerefs->zone))
2998 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
2999 alloc_flags |= ALLOC_CMA;
3002 cpuset_mems_cookie = read_mems_allowed_begin();
3004 /* We set it here, as __alloc_pages_slowpath might have changed it */
3005 ac.zonelist = zonelist;
3006 /* The preferred zone is used for statistics later */
3007 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3008 ac.nodemask ? : &cpuset_current_mems_allowed,
3009 &ac.preferred_zone);
3010 if (!ac.preferred_zone)
3012 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3014 /* First allocation attempt */
3015 alloc_mask = gfp_mask|__GFP_HARDWALL;
3016 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3017 if (unlikely(!page)) {
3019 * Runtime PM, block IO and its error handling path
3020 * can deadlock because I/O on the device might not
3023 alloc_mask = memalloc_noio_flags(gfp_mask);
3025 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3028 if (kmemcheck_enabled && page)
3029 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3031 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3035 * When updating a task's mems_allowed, it is possible to race with
3036 * parallel threads in such a way that an allocation can fail while
3037 * the mask is being updated. If a page allocation is about to fail,
3038 * check if the cpuset changed during allocation and if so, retry.
3040 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3045 EXPORT_SYMBOL(__alloc_pages_nodemask);
3048 * Common helper functions.
3050 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3055 * __get_free_pages() returns a 32-bit address, which cannot represent
3058 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3060 page = alloc_pages(gfp_mask, order);
3063 return (unsigned long) page_address(page);
3065 EXPORT_SYMBOL(__get_free_pages);
3067 unsigned long get_zeroed_page(gfp_t gfp_mask)
3069 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3071 EXPORT_SYMBOL(get_zeroed_page);
3073 void __free_pages(struct page *page, unsigned int order)
3075 if (put_page_testzero(page)) {
3077 free_hot_cold_page(page, false);
3079 __free_pages_ok(page, order);
3083 EXPORT_SYMBOL(__free_pages);
3085 void free_pages(unsigned long addr, unsigned int order)
3088 VM_BUG_ON(!virt_addr_valid((void *)addr));
3089 __free_pages(virt_to_page((void *)addr), order);
3093 EXPORT_SYMBOL(free_pages);
3097 * An arbitrary-length arbitrary-offset area of memory which resides
3098 * within a 0 or higher order page. Multiple fragments within that page
3099 * are individually refcounted, in the page's reference counter.
3101 * The page_frag functions below provide a simple allocation framework for
3102 * page fragments. This is used by the network stack and network device
3103 * drivers to provide a backing region of memory for use as either an
3104 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3106 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3109 struct page *page = NULL;
3110 gfp_t gfp = gfp_mask;
3112 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3113 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3115 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3116 PAGE_FRAG_CACHE_MAX_ORDER);
3117 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3119 if (unlikely(!page))
3120 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3122 nc->va = page ? page_address(page) : NULL;
3127 void *__alloc_page_frag(struct page_frag_cache *nc,
3128 unsigned int fragsz, gfp_t gfp_mask)
3130 unsigned int size = PAGE_SIZE;
3134 if (unlikely(!nc->va)) {
3136 page = __page_frag_refill(nc, gfp_mask);
3140 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3141 /* if size can vary use size else just use PAGE_SIZE */
3144 /* Even if we own the page, we do not use atomic_set().
3145 * This would break get_page_unless_zero() users.
3147 atomic_add(size - 1, &page->_count);
3149 /* reset page count bias and offset to start of new frag */
3150 nc->pfmemalloc = page->pfmemalloc;
3151 nc->pagecnt_bias = size;
3155 offset = nc->offset - fragsz;
3156 if (unlikely(offset < 0)) {
3157 page = virt_to_page(nc->va);
3159 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3162 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3163 /* if size can vary use size else just use PAGE_SIZE */
3166 /* OK, page count is 0, we can safely set it */
3167 atomic_set(&page->_count, size);
3169 /* reset page count bias and offset to start of new frag */
3170 nc->pagecnt_bias = size;
3171 offset = size - fragsz;
3175 nc->offset = offset;
3177 return nc->va + offset;
3179 EXPORT_SYMBOL(__alloc_page_frag);
3182 * Frees a page fragment allocated out of either a compound or order 0 page.
3184 void __free_page_frag(void *addr)
3186 struct page *page = virt_to_head_page(addr);
3188 if (unlikely(put_page_testzero(page)))
3189 __free_pages_ok(page, compound_order(page));
3191 EXPORT_SYMBOL(__free_page_frag);
3194 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3195 * of the current memory cgroup.
3197 * It should be used when the caller would like to use kmalloc, but since the
3198 * allocation is large, it has to fall back to the page allocator.
3200 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3203 struct mem_cgroup *memcg = NULL;
3205 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
3207 page = alloc_pages(gfp_mask, order);
3208 memcg_kmem_commit_charge(page, memcg, order);
3212 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3215 struct mem_cgroup *memcg = NULL;
3217 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
3219 page = alloc_pages_node(nid, gfp_mask, order);
3220 memcg_kmem_commit_charge(page, memcg, order);
3225 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3228 void __free_kmem_pages(struct page *page, unsigned int order)
3230 memcg_kmem_uncharge_pages(page, order);
3231 __free_pages(page, order);
3234 void free_kmem_pages(unsigned long addr, unsigned int order)
3237 VM_BUG_ON(!virt_addr_valid((void *)addr));
3238 __free_kmem_pages(virt_to_page((void *)addr), order);
3242 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
3245 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3246 unsigned long used = addr + PAGE_ALIGN(size);
3248 split_page(virt_to_page((void *)addr), order);
3249 while (used < alloc_end) {
3254 return (void *)addr;
3258 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3259 * @size: the number of bytes to allocate
3260 * @gfp_mask: GFP flags for the allocation
3262 * This function is similar to alloc_pages(), except that it allocates the
3263 * minimum number of pages to satisfy the request. alloc_pages() can only
3264 * allocate memory in power-of-two pages.
3266 * This function is also limited by MAX_ORDER.
3268 * Memory allocated by this function must be released by free_pages_exact().
3270 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3272 unsigned int order = get_order(size);
3275 addr = __get_free_pages(gfp_mask, order);
3276 return make_alloc_exact(addr, order, size);
3278 EXPORT_SYMBOL(alloc_pages_exact);
3281 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3283 * @nid: the preferred node ID where memory should be allocated
3284 * @size: the number of bytes to allocate
3285 * @gfp_mask: GFP flags for the allocation
3287 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3289 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3292 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3294 unsigned order = get_order(size);
3295 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3298 return make_alloc_exact((unsigned long)page_address(p), order, size);
3302 * free_pages_exact - release memory allocated via alloc_pages_exact()
3303 * @virt: the value returned by alloc_pages_exact.
3304 * @size: size of allocation, same value as passed to alloc_pages_exact().
3306 * Release the memory allocated by a previous call to alloc_pages_exact.
3308 void free_pages_exact(void *virt, size_t size)
3310 unsigned long addr = (unsigned long)virt;
3311 unsigned long end = addr + PAGE_ALIGN(size);
3313 while (addr < end) {
3318 EXPORT_SYMBOL(free_pages_exact);
3321 * nr_free_zone_pages - count number of pages beyond high watermark
3322 * @offset: The zone index of the highest zone
3324 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3325 * high watermark within all zones at or below a given zone index. For each
3326 * zone, the number of pages is calculated as:
3327 * managed_pages - high_pages
3329 static unsigned long nr_free_zone_pages(int offset)
3334 /* Just pick one node, since fallback list is circular */
3335 unsigned long sum = 0;
3337 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3339 for_each_zone_zonelist(zone, z, zonelist, offset) {
3340 unsigned long size = zone->managed_pages;
3341 unsigned long high = high_wmark_pages(zone);
3350 * nr_free_buffer_pages - count number of pages beyond high watermark
3352 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3353 * watermark within ZONE_DMA and ZONE_NORMAL.
3355 unsigned long nr_free_buffer_pages(void)
3357 return nr_free_zone_pages(gfp_zone(GFP_USER));
3359 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3362 * nr_free_pagecache_pages - count number of pages beyond high watermark
3364 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3365 * high watermark within all zones.
3367 unsigned long nr_free_pagecache_pages(void)
3369 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3372 static inline void show_node(struct zone *zone)
3374 if (IS_ENABLED(CONFIG_NUMA))
3375 printk("Node %d ", zone_to_nid(zone));
3378 void si_meminfo(struct sysinfo *val)
3380 val->totalram = totalram_pages;
3381 val->sharedram = global_page_state(NR_SHMEM);
3382 val->freeram = global_page_state(NR_FREE_PAGES);
3383 val->bufferram = nr_blockdev_pages();
3384 val->totalhigh = totalhigh_pages;
3385 val->freehigh = nr_free_highpages();
3386 val->mem_unit = PAGE_SIZE;
3389 EXPORT_SYMBOL(si_meminfo);
3392 void si_meminfo_node(struct sysinfo *val, int nid)
3394 int zone_type; /* needs to be signed */
3395 unsigned long managed_pages = 0;
3396 pg_data_t *pgdat = NODE_DATA(nid);
3398 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3399 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3400 val->totalram = managed_pages;
3401 val->sharedram = node_page_state(nid, NR_SHMEM);
3402 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3403 #ifdef CONFIG_HIGHMEM
3404 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3405 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3411 val->mem_unit = PAGE_SIZE;
3416 * Determine whether the node should be displayed or not, depending on whether
3417 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3419 bool skip_free_areas_node(unsigned int flags, int nid)
3422 unsigned int cpuset_mems_cookie;
3424 if (!(flags & SHOW_MEM_FILTER_NODES))
3428 cpuset_mems_cookie = read_mems_allowed_begin();
3429 ret = !node_isset(nid, cpuset_current_mems_allowed);
3430 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3435 #define K(x) ((x) << (PAGE_SHIFT-10))
3437 static void show_migration_types(unsigned char type)
3439 static const char types[MIGRATE_TYPES] = {
3440 [MIGRATE_UNMOVABLE] = 'U',
3441 [MIGRATE_RECLAIMABLE] = 'E',
3442 [MIGRATE_MOVABLE] = 'M',
3443 [MIGRATE_RESERVE] = 'R',
3445 [MIGRATE_CMA] = 'C',
3447 #ifdef CONFIG_MEMORY_ISOLATION
3448 [MIGRATE_ISOLATE] = 'I',
3451 char tmp[MIGRATE_TYPES + 1];
3455 for (i = 0; i < MIGRATE_TYPES; i++) {
3456 if (type & (1 << i))
3461 printk("(%s) ", tmp);
3465 * Show free area list (used inside shift_scroll-lock stuff)
3466 * We also calculate the percentage fragmentation. We do this by counting the
3467 * memory on each free list with the exception of the first item on the list.
3470 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3473 void show_free_areas(unsigned int filter)
3475 unsigned long free_pcp = 0;
3479 for_each_populated_zone(zone) {
3480 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3483 for_each_online_cpu(cpu)
3484 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3487 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3488 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3489 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3490 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3491 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3492 " free:%lu free_pcp:%lu free_cma:%lu\n",
3493 global_page_state(NR_ACTIVE_ANON),
3494 global_page_state(NR_INACTIVE_ANON),
3495 global_page_state(NR_ISOLATED_ANON),
3496 global_page_state(NR_ACTIVE_FILE),
3497 global_page_state(NR_INACTIVE_FILE),
3498 global_page_state(NR_ISOLATED_FILE),
3499 global_page_state(NR_UNEVICTABLE),
3500 global_page_state(NR_FILE_DIRTY),
3501 global_page_state(NR_WRITEBACK),
3502 global_page_state(NR_UNSTABLE_NFS),
3503 global_page_state(NR_SLAB_RECLAIMABLE),
3504 global_page_state(NR_SLAB_UNRECLAIMABLE),
3505 global_page_state(NR_FILE_MAPPED),
3506 global_page_state(NR_SHMEM),
3507 global_page_state(NR_PAGETABLE),
3508 global_page_state(NR_BOUNCE),
3509 global_page_state(NR_FREE_PAGES),
3511 global_page_state(NR_FREE_CMA_PAGES));
3513 for_each_populated_zone(zone) {
3516 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3520 for_each_online_cpu(cpu)
3521 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3529 " active_anon:%lukB"
3530 " inactive_anon:%lukB"
3531 " active_file:%lukB"
3532 " inactive_file:%lukB"
3533 " unevictable:%lukB"
3534 " isolated(anon):%lukB"
3535 " isolated(file):%lukB"
3543 " slab_reclaimable:%lukB"
3544 " slab_unreclaimable:%lukB"
3545 " kernel_stack:%lukB"
3552 " writeback_tmp:%lukB"
3553 " pages_scanned:%lu"
3554 " all_unreclaimable? %s"
3557 K(zone_page_state(zone, NR_FREE_PAGES)),
3558 K(min_wmark_pages(zone)),
3559 K(low_wmark_pages(zone)),
3560 K(high_wmark_pages(zone)),
3561 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3562 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3563 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3564 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3565 K(zone_page_state(zone, NR_UNEVICTABLE)),
3566 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3567 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3568 K(zone->present_pages),
3569 K(zone->managed_pages),
3570 K(zone_page_state(zone, NR_MLOCK)),
3571 K(zone_page_state(zone, NR_FILE_DIRTY)),
3572 K(zone_page_state(zone, NR_WRITEBACK)),
3573 K(zone_page_state(zone, NR_FILE_MAPPED)),
3574 K(zone_page_state(zone, NR_SHMEM)),
3575 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3576 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3577 zone_page_state(zone, NR_KERNEL_STACK) *
3579 K(zone_page_state(zone, NR_PAGETABLE)),
3580 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3581 K(zone_page_state(zone, NR_BOUNCE)),
3583 K(this_cpu_read(zone->pageset->pcp.count)),
3584 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3585 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3586 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3587 (!zone_reclaimable(zone) ? "yes" : "no")
3589 printk("lowmem_reserve[]:");
3590 for (i = 0; i < MAX_NR_ZONES; i++)
3591 printk(" %ld", zone->lowmem_reserve[i]);
3595 for_each_populated_zone(zone) {
3596 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3597 unsigned char types[MAX_ORDER];
3599 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3602 printk("%s: ", zone->name);
3604 spin_lock_irqsave(&zone->lock, flags);
3605 for (order = 0; order < MAX_ORDER; order++) {
3606 struct free_area *area = &zone->free_area[order];
3609 nr[order] = area->nr_free;
3610 total += nr[order] << order;
3613 for (type = 0; type < MIGRATE_TYPES; type++) {
3614 if (!list_empty(&area->free_list[type]))
3615 types[order] |= 1 << type;
3618 spin_unlock_irqrestore(&zone->lock, flags);
3619 for (order = 0; order < MAX_ORDER; order++) {
3620 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3622 show_migration_types(types[order]);
3624 printk("= %lukB\n", K(total));
3627 hugetlb_show_meminfo();
3629 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3631 show_swap_cache_info();
3634 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3636 zoneref->zone = zone;
3637 zoneref->zone_idx = zone_idx(zone);
3641 * Builds allocation fallback zone lists.
3643 * Add all populated zones of a node to the zonelist.
3645 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3649 enum zone_type zone_type = MAX_NR_ZONES;
3653 zone = pgdat->node_zones + zone_type;
3654 if (populated_zone(zone)) {
3655 zoneref_set_zone(zone,
3656 &zonelist->_zonerefs[nr_zones++]);
3657 check_highest_zone(zone_type);
3659 } while (zone_type);
3667 * 0 = automatic detection of better ordering.
3668 * 1 = order by ([node] distance, -zonetype)
3669 * 2 = order by (-zonetype, [node] distance)
3671 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3672 * the same zonelist. So only NUMA can configure this param.
3674 #define ZONELIST_ORDER_DEFAULT 0
3675 #define ZONELIST_ORDER_NODE 1
3676 #define ZONELIST_ORDER_ZONE 2
3678 /* zonelist order in the kernel.
3679 * set_zonelist_order() will set this to NODE or ZONE.
3681 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3682 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3686 /* The value user specified ....changed by config */
3687 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3688 /* string for sysctl */
3689 #define NUMA_ZONELIST_ORDER_LEN 16
3690 char numa_zonelist_order[16] = "default";
3693 * interface for configure zonelist ordering.
3694 * command line option "numa_zonelist_order"
3695 * = "[dD]efault - default, automatic configuration.
3696 * = "[nN]ode - order by node locality, then by zone within node
3697 * = "[zZ]one - order by zone, then by locality within zone
3700 static int __parse_numa_zonelist_order(char *s)
3702 if (*s == 'd' || *s == 'D') {
3703 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3704 } else if (*s == 'n' || *s == 'N') {
3705 user_zonelist_order = ZONELIST_ORDER_NODE;
3706 } else if (*s == 'z' || *s == 'Z') {
3707 user_zonelist_order = ZONELIST_ORDER_ZONE;
3710 "Ignoring invalid numa_zonelist_order value: "
3717 static __init int setup_numa_zonelist_order(char *s)
3724 ret = __parse_numa_zonelist_order(s);
3726 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3730 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3733 * sysctl handler for numa_zonelist_order
3735 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3736 void __user *buffer, size_t *length,
3739 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3741 static DEFINE_MUTEX(zl_order_mutex);
3743 mutex_lock(&zl_order_mutex);
3745 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3749 strcpy(saved_string, (char *)table->data);
3751 ret = proc_dostring(table, write, buffer, length, ppos);
3755 int oldval = user_zonelist_order;
3757 ret = __parse_numa_zonelist_order((char *)table->data);
3760 * bogus value. restore saved string
3762 strncpy((char *)table->data, saved_string,
3763 NUMA_ZONELIST_ORDER_LEN);
3764 user_zonelist_order = oldval;
3765 } else if (oldval != user_zonelist_order) {
3766 mutex_lock(&zonelists_mutex);
3767 build_all_zonelists(NULL, NULL);
3768 mutex_unlock(&zonelists_mutex);
3772 mutex_unlock(&zl_order_mutex);
3777 #define MAX_NODE_LOAD (nr_online_nodes)
3778 static int node_load[MAX_NUMNODES];
3781 * find_next_best_node - find the next node that should appear in a given node's fallback list
3782 * @node: node whose fallback list we're appending
3783 * @used_node_mask: nodemask_t of already used nodes
3785 * We use a number of factors to determine which is the next node that should
3786 * appear on a given node's fallback list. The node should not have appeared
3787 * already in @node's fallback list, and it should be the next closest node
3788 * according to the distance array (which contains arbitrary distance values
3789 * from each node to each node in the system), and should also prefer nodes
3790 * with no CPUs, since presumably they'll have very little allocation pressure
3791 * on them otherwise.
3792 * It returns -1 if no node is found.
3794 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3797 int min_val = INT_MAX;
3798 int best_node = NUMA_NO_NODE;
3799 const struct cpumask *tmp = cpumask_of_node(0);
3801 /* Use the local node if we haven't already */
3802 if (!node_isset(node, *used_node_mask)) {
3803 node_set(node, *used_node_mask);
3807 for_each_node_state(n, N_MEMORY) {
3809 /* Don't want a node to appear more than once */
3810 if (node_isset(n, *used_node_mask))
3813 /* Use the distance array to find the distance */
3814 val = node_distance(node, n);
3816 /* Penalize nodes under us ("prefer the next node") */
3819 /* Give preference to headless and unused nodes */
3820 tmp = cpumask_of_node(n);
3821 if (!cpumask_empty(tmp))
3822 val += PENALTY_FOR_NODE_WITH_CPUS;
3824 /* Slight preference for less loaded node */
3825 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3826 val += node_load[n];
3828 if (val < min_val) {
3835 node_set(best_node, *used_node_mask);
3842 * Build zonelists ordered by node and zones within node.
3843 * This results in maximum locality--normal zone overflows into local
3844 * DMA zone, if any--but risks exhausting DMA zone.
3846 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3849 struct zonelist *zonelist;
3851 zonelist = &pgdat->node_zonelists[0];
3852 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3854 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3855 zonelist->_zonerefs[j].zone = NULL;
3856 zonelist->_zonerefs[j].zone_idx = 0;
3860 * Build gfp_thisnode zonelists
3862 static void build_thisnode_zonelists(pg_data_t *pgdat)
3865 struct zonelist *zonelist;
3867 zonelist = &pgdat->node_zonelists[1];
3868 j = build_zonelists_node(pgdat, zonelist, 0);
3869 zonelist->_zonerefs[j].zone = NULL;
3870 zonelist->_zonerefs[j].zone_idx = 0;
3874 * Build zonelists ordered by zone and nodes within zones.
3875 * This results in conserving DMA zone[s] until all Normal memory is
3876 * exhausted, but results in overflowing to remote node while memory
3877 * may still exist in local DMA zone.
3879 static int node_order[MAX_NUMNODES];
3881 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3884 int zone_type; /* needs to be signed */
3886 struct zonelist *zonelist;
3888 zonelist = &pgdat->node_zonelists[0];
3890 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3891 for (j = 0; j < nr_nodes; j++) {
3892 node = node_order[j];
3893 z = &NODE_DATA(node)->node_zones[zone_type];
3894 if (populated_zone(z)) {
3896 &zonelist->_zonerefs[pos++]);
3897 check_highest_zone(zone_type);
3901 zonelist->_zonerefs[pos].zone = NULL;
3902 zonelist->_zonerefs[pos].zone_idx = 0;
3905 #if defined(CONFIG_64BIT)
3907 * Devices that require DMA32/DMA are relatively rare and do not justify a
3908 * penalty to every machine in case the specialised case applies. Default
3909 * to Node-ordering on 64-bit NUMA machines
3911 static int default_zonelist_order(void)
3913 return ZONELIST_ORDER_NODE;
3917 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
3918 * by the kernel. If processes running on node 0 deplete the low memory zone
3919 * then reclaim will occur more frequency increasing stalls and potentially
3920 * be easier to OOM if a large percentage of the zone is under writeback or
3921 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
3922 * Hence, default to zone ordering on 32-bit.
3924 static int default_zonelist_order(void)
3926 return ZONELIST_ORDER_ZONE;
3928 #endif /* CONFIG_64BIT */
3930 static void set_zonelist_order(void)
3932 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3933 current_zonelist_order = default_zonelist_order();
3935 current_zonelist_order = user_zonelist_order;
3938 static void build_zonelists(pg_data_t *pgdat)
3942 nodemask_t used_mask;
3943 int local_node, prev_node;
3944 struct zonelist *zonelist;
3945 int order = current_zonelist_order;
3947 /* initialize zonelists */
3948 for (i = 0; i < MAX_ZONELISTS; i++) {
3949 zonelist = pgdat->node_zonelists + i;
3950 zonelist->_zonerefs[0].zone = NULL;
3951 zonelist->_zonerefs[0].zone_idx = 0;
3954 /* NUMA-aware ordering of nodes */
3955 local_node = pgdat->node_id;
3956 load = nr_online_nodes;
3957 prev_node = local_node;
3958 nodes_clear(used_mask);
3960 memset(node_order, 0, sizeof(node_order));
3963 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3965 * We don't want to pressure a particular node.
3966 * So adding penalty to the first node in same
3967 * distance group to make it round-robin.
3969 if (node_distance(local_node, node) !=
3970 node_distance(local_node, prev_node))
3971 node_load[node] = load;
3975 if (order == ZONELIST_ORDER_NODE)
3976 build_zonelists_in_node_order(pgdat, node);
3978 node_order[j++] = node; /* remember order */
3981 if (order == ZONELIST_ORDER_ZONE) {
3982 /* calculate node order -- i.e., DMA last! */
3983 build_zonelists_in_zone_order(pgdat, j);
3986 build_thisnode_zonelists(pgdat);
3989 /* Construct the zonelist performance cache - see further mmzone.h */
3990 static void build_zonelist_cache(pg_data_t *pgdat)
3992 struct zonelist *zonelist;
3993 struct zonelist_cache *zlc;
3996 zonelist = &pgdat->node_zonelists[0];
3997 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3998 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3999 for (z = zonelist->_zonerefs; z->zone; z++)
4000 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
4003 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4005 * Return node id of node used for "local" allocations.
4006 * I.e., first node id of first zone in arg node's generic zonelist.
4007 * Used for initializing percpu 'numa_mem', which is used primarily
4008 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4010 int local_memory_node(int node)
4014 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4015 gfp_zone(GFP_KERNEL),
4022 #else /* CONFIG_NUMA */
4024 static void set_zonelist_order(void)
4026 current_zonelist_order = ZONELIST_ORDER_ZONE;
4029 static void build_zonelists(pg_data_t *pgdat)
4031 int node, local_node;
4033 struct zonelist *zonelist;
4035 local_node = pgdat->node_id;
4037 zonelist = &pgdat->node_zonelists[0];
4038 j = build_zonelists_node(pgdat, zonelist, 0);
4041 * Now we build the zonelist so that it contains the zones
4042 * of all the other nodes.
4043 * We don't want to pressure a particular node, so when
4044 * building the zones for node N, we make sure that the
4045 * zones coming right after the local ones are those from
4046 * node N+1 (modulo N)
4048 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4049 if (!node_online(node))
4051 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4053 for (node = 0; node < local_node; node++) {
4054 if (!node_online(node))
4056 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4059 zonelist->_zonerefs[j].zone = NULL;
4060 zonelist->_zonerefs[j].zone_idx = 0;
4063 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
4064 static void build_zonelist_cache(pg_data_t *pgdat)
4066 pgdat->node_zonelists[0].zlcache_ptr = NULL;
4069 #endif /* CONFIG_NUMA */
4072 * Boot pageset table. One per cpu which is going to be used for all
4073 * zones and all nodes. The parameters will be set in such a way
4074 * that an item put on a list will immediately be handed over to
4075 * the buddy list. This is safe since pageset manipulation is done
4076 * with interrupts disabled.
4078 * The boot_pagesets must be kept even after bootup is complete for
4079 * unused processors and/or zones. They do play a role for bootstrapping
4080 * hotplugged processors.
4082 * zoneinfo_show() and maybe other functions do
4083 * not check if the processor is online before following the pageset pointer.
4084 * Other parts of the kernel may not check if the zone is available.
4086 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4087 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4088 static void setup_zone_pageset(struct zone *zone);
4091 * Global mutex to protect against size modification of zonelists
4092 * as well as to serialize pageset setup for the new populated zone.
4094 DEFINE_MUTEX(zonelists_mutex);
4096 /* return values int ....just for stop_machine() */
4097 static int __build_all_zonelists(void *data)
4101 pg_data_t *self = data;
4104 memset(node_load, 0, sizeof(node_load));
4107 if (self && !node_online(self->node_id)) {
4108 build_zonelists(self);
4109 build_zonelist_cache(self);
4112 for_each_online_node(nid) {
4113 pg_data_t *pgdat = NODE_DATA(nid);
4115 build_zonelists(pgdat);
4116 build_zonelist_cache(pgdat);
4120 * Initialize the boot_pagesets that are going to be used
4121 * for bootstrapping processors. The real pagesets for
4122 * each zone will be allocated later when the per cpu
4123 * allocator is available.
4125 * boot_pagesets are used also for bootstrapping offline
4126 * cpus if the system is already booted because the pagesets
4127 * are needed to initialize allocators on a specific cpu too.
4128 * F.e. the percpu allocator needs the page allocator which
4129 * needs the percpu allocator in order to allocate its pagesets
4130 * (a chicken-egg dilemma).
4132 for_each_possible_cpu(cpu) {
4133 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4135 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4137 * We now know the "local memory node" for each node--
4138 * i.e., the node of the first zone in the generic zonelist.
4139 * Set up numa_mem percpu variable for on-line cpus. During
4140 * boot, only the boot cpu should be on-line; we'll init the
4141 * secondary cpus' numa_mem as they come on-line. During
4142 * node/memory hotplug, we'll fixup all on-line cpus.
4144 if (cpu_online(cpu))
4145 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4152 static noinline void __init
4153 build_all_zonelists_init(void)
4155 __build_all_zonelists(NULL);
4156 mminit_verify_zonelist();
4157 cpuset_init_current_mems_allowed();
4161 * Called with zonelists_mutex held always
4162 * unless system_state == SYSTEM_BOOTING.
4164 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4165 * [we're only called with non-NULL zone through __meminit paths] and
4166 * (2) call of __init annotated helper build_all_zonelists_init
4167 * [protected by SYSTEM_BOOTING].
4169 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4171 set_zonelist_order();
4173 if (system_state == SYSTEM_BOOTING) {
4174 build_all_zonelists_init();
4176 #ifdef CONFIG_MEMORY_HOTPLUG
4178 setup_zone_pageset(zone);
4180 /* we have to stop all cpus to guarantee there is no user
4182 stop_machine(__build_all_zonelists, pgdat, NULL);
4183 /* cpuset refresh routine should be here */
4185 vm_total_pages = nr_free_pagecache_pages();
4187 * Disable grouping by mobility if the number of pages in the
4188 * system is too low to allow the mechanism to work. It would be
4189 * more accurate, but expensive to check per-zone. This check is
4190 * made on memory-hotadd so a system can start with mobility
4191 * disabled and enable it later
4193 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4194 page_group_by_mobility_disabled = 1;
4196 page_group_by_mobility_disabled = 0;
4198 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4199 "Total pages: %ld\n",
4201 zonelist_order_name[current_zonelist_order],
4202 page_group_by_mobility_disabled ? "off" : "on",
4205 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4210 * Helper functions to size the waitqueue hash table.
4211 * Essentially these want to choose hash table sizes sufficiently
4212 * large so that collisions trying to wait on pages are rare.
4213 * But in fact, the number of active page waitqueues on typical
4214 * systems is ridiculously low, less than 200. So this is even
4215 * conservative, even though it seems large.
4217 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4218 * waitqueues, i.e. the size of the waitq table given the number of pages.
4220 #define PAGES_PER_WAITQUEUE 256
4222 #ifndef CONFIG_MEMORY_HOTPLUG
4223 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4225 unsigned long size = 1;
4227 pages /= PAGES_PER_WAITQUEUE;
4229 while (size < pages)
4233 * Once we have dozens or even hundreds of threads sleeping
4234 * on IO we've got bigger problems than wait queue collision.
4235 * Limit the size of the wait table to a reasonable size.
4237 size = min(size, 4096UL);
4239 return max(size, 4UL);
4243 * A zone's size might be changed by hot-add, so it is not possible to determine
4244 * a suitable size for its wait_table. So we use the maximum size now.
4246 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4248 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4249 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4250 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4252 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4253 * or more by the traditional way. (See above). It equals:
4255 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4256 * ia64(16K page size) : = ( 8G + 4M)byte.
4257 * powerpc (64K page size) : = (32G +16M)byte.
4259 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4266 * This is an integer logarithm so that shifts can be used later
4267 * to extract the more random high bits from the multiplicative
4268 * hash function before the remainder is taken.
4270 static inline unsigned long wait_table_bits(unsigned long size)
4276 * Check if a pageblock contains reserved pages
4278 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
4282 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4283 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
4290 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4291 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4292 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4293 * higher will lead to a bigger reserve which will get freed as contiguous
4294 * blocks as reclaim kicks in
4296 static void setup_zone_migrate_reserve(struct zone *zone)
4298 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
4300 unsigned long block_migratetype;
4305 * Get the start pfn, end pfn and the number of blocks to reserve
4306 * We have to be careful to be aligned to pageblock_nr_pages to
4307 * make sure that we always check pfn_valid for the first page in
4310 start_pfn = zone->zone_start_pfn;
4311 end_pfn = zone_end_pfn(zone);
4312 start_pfn = roundup(start_pfn, pageblock_nr_pages);
4313 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
4317 * Reserve blocks are generally in place to help high-order atomic
4318 * allocations that are short-lived. A min_free_kbytes value that
4319 * would result in more than 2 reserve blocks for atomic allocations
4320 * is assumed to be in place to help anti-fragmentation for the
4321 * future allocation of hugepages at runtime.
4323 reserve = min(2, reserve);
4324 old_reserve = zone->nr_migrate_reserve_block;
4326 /* When memory hot-add, we almost always need to do nothing */
4327 if (reserve == old_reserve)
4329 zone->nr_migrate_reserve_block = reserve;
4331 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
4332 if (!pfn_valid(pfn))
4334 page = pfn_to_page(pfn);
4336 /* Watch out for overlapping nodes */
4337 if (page_to_nid(page) != zone_to_nid(zone))
4340 block_migratetype = get_pageblock_migratetype(page);
4342 /* Only test what is necessary when the reserves are not met */
4345 * Blocks with reserved pages will never free, skip
4348 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
4349 if (pageblock_is_reserved(pfn, block_end_pfn))
4352 /* If this block is reserved, account for it */
4353 if (block_migratetype == MIGRATE_RESERVE) {
4358 /* Suitable for reserving if this block is movable */
4359 if (block_migratetype == MIGRATE_MOVABLE) {
4360 set_pageblock_migratetype(page,
4362 move_freepages_block(zone, page,
4367 } else if (!old_reserve) {
4369 * At boot time we don't need to scan the whole zone
4370 * for turning off MIGRATE_RESERVE.
4376 * If the reserve is met and this is a previous reserved block,
4379 if (block_migratetype == MIGRATE_RESERVE) {
4380 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4381 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4387 * Initially all pages are reserved - free ones are freed
4388 * up by free_all_bootmem() once the early boot process is
4389 * done. Non-atomic initialization, single-pass.
4391 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4392 unsigned long start_pfn, enum memmap_context context)
4394 pg_data_t *pgdat = NODE_DATA(nid);
4395 unsigned long end_pfn = start_pfn + size;
4398 unsigned long nr_initialised = 0;
4400 if (highest_memmap_pfn < end_pfn - 1)
4401 highest_memmap_pfn = end_pfn - 1;
4403 z = &pgdat->node_zones[zone];
4404 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4406 * There can be holes in boot-time mem_map[]s
4407 * handed to this function. They do not
4408 * exist on hotplugged memory.
4410 if (context == MEMMAP_EARLY) {
4411 if (!early_pfn_valid(pfn))
4413 if (!early_pfn_in_nid(pfn, nid))
4415 if (!update_defer_init(pgdat, pfn, end_pfn,
4419 __init_single_pfn(pfn, zone, nid);
4423 static void __meminit zone_init_free_lists(struct zone *zone)
4425 unsigned int order, t;
4426 for_each_migratetype_order(order, t) {
4427 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4428 zone->free_area[order].nr_free = 0;
4432 #ifndef __HAVE_ARCH_MEMMAP_INIT
4433 #define memmap_init(size, nid, zone, start_pfn) \
4434 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4437 static int zone_batchsize(struct zone *zone)
4443 * The per-cpu-pages pools are set to around 1000th of the
4444 * size of the zone. But no more than 1/2 of a meg.
4446 * OK, so we don't know how big the cache is. So guess.
4448 batch = zone->managed_pages / 1024;
4449 if (batch * PAGE_SIZE > 512 * 1024)
4450 batch = (512 * 1024) / PAGE_SIZE;
4451 batch /= 4; /* We effectively *= 4 below */
4456 * Clamp the batch to a 2^n - 1 value. Having a power
4457 * of 2 value was found to be more likely to have
4458 * suboptimal cache aliasing properties in some cases.
4460 * For example if 2 tasks are alternately allocating
4461 * batches of pages, one task can end up with a lot
4462 * of pages of one half of the possible page colors
4463 * and the other with pages of the other colors.
4465 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4470 /* The deferral and batching of frees should be suppressed under NOMMU
4473 * The problem is that NOMMU needs to be able to allocate large chunks
4474 * of contiguous memory as there's no hardware page translation to
4475 * assemble apparent contiguous memory from discontiguous pages.
4477 * Queueing large contiguous runs of pages for batching, however,
4478 * causes the pages to actually be freed in smaller chunks. As there
4479 * can be a significant delay between the individual batches being
4480 * recycled, this leads to the once large chunks of space being
4481 * fragmented and becoming unavailable for high-order allocations.
4488 * pcp->high and pcp->batch values are related and dependent on one another:
4489 * ->batch must never be higher then ->high.
4490 * The following function updates them in a safe manner without read side
4493 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4494 * those fields changing asynchronously (acording the the above rule).
4496 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4497 * outside of boot time (or some other assurance that no concurrent updaters
4500 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4501 unsigned long batch)
4503 /* start with a fail safe value for batch */
4507 /* Update high, then batch, in order */
4514 /* a companion to pageset_set_high() */
4515 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4517 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4520 static void pageset_init(struct per_cpu_pageset *p)
4522 struct per_cpu_pages *pcp;
4525 memset(p, 0, sizeof(*p));
4529 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4530 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4533 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4536 pageset_set_batch(p, batch);
4540 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4541 * to the value high for the pageset p.
4543 static void pageset_set_high(struct per_cpu_pageset *p,
4546 unsigned long batch = max(1UL, high / 4);
4547 if ((high / 4) > (PAGE_SHIFT * 8))
4548 batch = PAGE_SHIFT * 8;
4550 pageset_update(&p->pcp, high, batch);
4553 static void pageset_set_high_and_batch(struct zone *zone,
4554 struct per_cpu_pageset *pcp)
4556 if (percpu_pagelist_fraction)
4557 pageset_set_high(pcp,
4558 (zone->managed_pages /
4559 percpu_pagelist_fraction));
4561 pageset_set_batch(pcp, zone_batchsize(zone));
4564 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4566 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4569 pageset_set_high_and_batch(zone, pcp);
4572 static void __meminit setup_zone_pageset(struct zone *zone)
4575 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4576 for_each_possible_cpu(cpu)
4577 zone_pageset_init(zone, cpu);
4581 * Allocate per cpu pagesets and initialize them.
4582 * Before this call only boot pagesets were available.
4584 void __init setup_per_cpu_pageset(void)
4588 for_each_populated_zone(zone)
4589 setup_zone_pageset(zone);
4592 static noinline __init_refok
4593 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4599 * The per-page waitqueue mechanism uses hashed waitqueues
4602 zone->wait_table_hash_nr_entries =
4603 wait_table_hash_nr_entries(zone_size_pages);
4604 zone->wait_table_bits =
4605 wait_table_bits(zone->wait_table_hash_nr_entries);
4606 alloc_size = zone->wait_table_hash_nr_entries
4607 * sizeof(wait_queue_head_t);
4609 if (!slab_is_available()) {
4610 zone->wait_table = (wait_queue_head_t *)
4611 memblock_virt_alloc_node_nopanic(
4612 alloc_size, zone->zone_pgdat->node_id);
4615 * This case means that a zone whose size was 0 gets new memory
4616 * via memory hot-add.
4617 * But it may be the case that a new node was hot-added. In
4618 * this case vmalloc() will not be able to use this new node's
4619 * memory - this wait_table must be initialized to use this new
4620 * node itself as well.
4621 * To use this new node's memory, further consideration will be
4624 zone->wait_table = vmalloc(alloc_size);
4626 if (!zone->wait_table)
4629 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4630 init_waitqueue_head(zone->wait_table + i);
4635 static __meminit void zone_pcp_init(struct zone *zone)
4638 * per cpu subsystem is not up at this point. The following code
4639 * relies on the ability of the linker to provide the
4640 * offset of a (static) per cpu variable into the per cpu area.
4642 zone->pageset = &boot_pageset;
4644 if (populated_zone(zone))
4645 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4646 zone->name, zone->present_pages,
4647 zone_batchsize(zone));
4650 int __meminit init_currently_empty_zone(struct zone *zone,
4651 unsigned long zone_start_pfn,
4653 enum memmap_context context)
4655 struct pglist_data *pgdat = zone->zone_pgdat;
4657 ret = zone_wait_table_init(zone, size);
4660 pgdat->nr_zones = zone_idx(zone) + 1;
4662 zone->zone_start_pfn = zone_start_pfn;
4664 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4665 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4667 (unsigned long)zone_idx(zone),
4668 zone_start_pfn, (zone_start_pfn + size));
4670 zone_init_free_lists(zone);
4675 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4676 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4679 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4681 int __meminit __early_pfn_to_nid(unsigned long pfn,
4682 struct mminit_pfnnid_cache *state)
4684 unsigned long start_pfn, end_pfn;
4687 if (state->last_start <= pfn && pfn < state->last_end)
4688 return state->last_nid;
4690 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4692 state->last_start = start_pfn;
4693 state->last_end = end_pfn;
4694 state->last_nid = nid;
4699 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4702 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4703 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4704 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4706 * If an architecture guarantees that all ranges registered contain no holes
4707 * and may be freed, this this function may be used instead of calling
4708 * memblock_free_early_nid() manually.
4710 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4712 unsigned long start_pfn, end_pfn;
4715 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4716 start_pfn = min(start_pfn, max_low_pfn);
4717 end_pfn = min(end_pfn, max_low_pfn);
4719 if (start_pfn < end_pfn)
4720 memblock_free_early_nid(PFN_PHYS(start_pfn),
4721 (end_pfn - start_pfn) << PAGE_SHIFT,
4727 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4728 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4730 * If an architecture guarantees that all ranges registered contain no holes and may
4731 * be freed, this function may be used instead of calling memory_present() manually.
4733 void __init sparse_memory_present_with_active_regions(int nid)
4735 unsigned long start_pfn, end_pfn;
4738 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4739 memory_present(this_nid, start_pfn, end_pfn);
4743 * get_pfn_range_for_nid - Return the start and end page frames for a node
4744 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4745 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4746 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4748 * It returns the start and end page frame of a node based on information
4749 * provided by memblock_set_node(). If called for a node
4750 * with no available memory, a warning is printed and the start and end
4753 void __meminit get_pfn_range_for_nid(unsigned int nid,
4754 unsigned long *start_pfn, unsigned long *end_pfn)
4756 unsigned long this_start_pfn, this_end_pfn;
4762 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4763 *start_pfn = min(*start_pfn, this_start_pfn);
4764 *end_pfn = max(*end_pfn, this_end_pfn);
4767 if (*start_pfn == -1UL)
4772 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4773 * assumption is made that zones within a node are ordered in monotonic
4774 * increasing memory addresses so that the "highest" populated zone is used
4776 static void __init find_usable_zone_for_movable(void)
4779 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4780 if (zone_index == ZONE_MOVABLE)
4783 if (arch_zone_highest_possible_pfn[zone_index] >
4784 arch_zone_lowest_possible_pfn[zone_index])
4788 VM_BUG_ON(zone_index == -1);
4789 movable_zone = zone_index;
4793 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4794 * because it is sized independent of architecture. Unlike the other zones,
4795 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4796 * in each node depending on the size of each node and how evenly kernelcore
4797 * is distributed. This helper function adjusts the zone ranges
4798 * provided by the architecture for a given node by using the end of the
4799 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4800 * zones within a node are in order of monotonic increases memory addresses
4802 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4803 unsigned long zone_type,
4804 unsigned long node_start_pfn,
4805 unsigned long node_end_pfn,
4806 unsigned long *zone_start_pfn,
4807 unsigned long *zone_end_pfn)
4809 /* Only adjust if ZONE_MOVABLE is on this node */
4810 if (zone_movable_pfn[nid]) {
4811 /* Size ZONE_MOVABLE */
4812 if (zone_type == ZONE_MOVABLE) {
4813 *zone_start_pfn = zone_movable_pfn[nid];
4814 *zone_end_pfn = min(node_end_pfn,
4815 arch_zone_highest_possible_pfn[movable_zone]);
4817 /* Adjust for ZONE_MOVABLE starting within this range */
4818 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4819 *zone_end_pfn > zone_movable_pfn[nid]) {
4820 *zone_end_pfn = zone_movable_pfn[nid];
4822 /* Check if this whole range is within ZONE_MOVABLE */
4823 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4824 *zone_start_pfn = *zone_end_pfn;
4829 * Return the number of pages a zone spans in a node, including holes
4830 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4832 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4833 unsigned long zone_type,
4834 unsigned long node_start_pfn,
4835 unsigned long node_end_pfn,
4836 unsigned long *ignored)
4838 unsigned long zone_start_pfn, zone_end_pfn;
4840 /* Get the start and end of the zone */
4841 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4842 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4843 adjust_zone_range_for_zone_movable(nid, zone_type,
4844 node_start_pfn, node_end_pfn,
4845 &zone_start_pfn, &zone_end_pfn);
4847 /* Check that this node has pages within the zone's required range */
4848 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4851 /* Move the zone boundaries inside the node if necessary */
4852 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4853 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4855 /* Return the spanned pages */
4856 return zone_end_pfn - zone_start_pfn;
4860 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4861 * then all holes in the requested range will be accounted for.
4863 unsigned long __meminit __absent_pages_in_range(int nid,
4864 unsigned long range_start_pfn,
4865 unsigned long range_end_pfn)
4867 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4868 unsigned long start_pfn, end_pfn;
4871 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4872 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4873 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4874 nr_absent -= end_pfn - start_pfn;
4880 * absent_pages_in_range - Return number of page frames in holes within a range
4881 * @start_pfn: The start PFN to start searching for holes
4882 * @end_pfn: The end PFN to stop searching for holes
4884 * It returns the number of pages frames in memory holes within a range.
4886 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4887 unsigned long end_pfn)
4889 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4892 /* Return the number of page frames in holes in a zone on a node */
4893 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4894 unsigned long zone_type,
4895 unsigned long node_start_pfn,
4896 unsigned long node_end_pfn,
4897 unsigned long *ignored)
4899 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4900 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4901 unsigned long zone_start_pfn, zone_end_pfn;
4903 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4904 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4906 adjust_zone_range_for_zone_movable(nid, zone_type,
4907 node_start_pfn, node_end_pfn,
4908 &zone_start_pfn, &zone_end_pfn);
4909 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4912 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4913 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4914 unsigned long zone_type,
4915 unsigned long node_start_pfn,
4916 unsigned long node_end_pfn,
4917 unsigned long *zones_size)
4919 return zones_size[zone_type];
4922 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4923 unsigned long zone_type,
4924 unsigned long node_start_pfn,
4925 unsigned long node_end_pfn,
4926 unsigned long *zholes_size)
4931 return zholes_size[zone_type];
4934 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4936 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4937 unsigned long node_start_pfn,
4938 unsigned long node_end_pfn,
4939 unsigned long *zones_size,
4940 unsigned long *zholes_size)
4942 unsigned long realtotalpages = 0, totalpages = 0;
4945 for (i = 0; i < MAX_NR_ZONES; i++) {
4946 struct zone *zone = pgdat->node_zones + i;
4947 unsigned long size, real_size;
4949 size = zone_spanned_pages_in_node(pgdat->node_id, i,
4953 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
4954 node_start_pfn, node_end_pfn,
4956 zone->spanned_pages = size;
4957 zone->present_pages = real_size;
4960 realtotalpages += real_size;
4963 pgdat->node_spanned_pages = totalpages;
4964 pgdat->node_present_pages = realtotalpages;
4965 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4969 #ifndef CONFIG_SPARSEMEM
4971 * Calculate the size of the zone->blockflags rounded to an unsigned long
4972 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4973 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4974 * round what is now in bits to nearest long in bits, then return it in
4977 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4979 unsigned long usemapsize;
4981 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4982 usemapsize = roundup(zonesize, pageblock_nr_pages);
4983 usemapsize = usemapsize >> pageblock_order;
4984 usemapsize *= NR_PAGEBLOCK_BITS;
4985 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4987 return usemapsize / 8;
4990 static void __init setup_usemap(struct pglist_data *pgdat,
4992 unsigned long zone_start_pfn,
4993 unsigned long zonesize)
4995 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4996 zone->pageblock_flags = NULL;
4998 zone->pageblock_flags =
4999 memblock_virt_alloc_node_nopanic(usemapsize,
5003 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5004 unsigned long zone_start_pfn, unsigned long zonesize) {}
5005 #endif /* CONFIG_SPARSEMEM */
5007 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5009 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5010 void __paginginit set_pageblock_order(void)
5014 /* Check that pageblock_nr_pages has not already been setup */
5015 if (pageblock_order)
5018 if (HPAGE_SHIFT > PAGE_SHIFT)
5019 order = HUGETLB_PAGE_ORDER;
5021 order = MAX_ORDER - 1;
5024 * Assume the largest contiguous order of interest is a huge page.
5025 * This value may be variable depending on boot parameters on IA64 and
5028 pageblock_order = order;
5030 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5033 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5034 * is unused as pageblock_order is set at compile-time. See
5035 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5038 void __paginginit set_pageblock_order(void)
5042 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5044 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5045 unsigned long present_pages)
5047 unsigned long pages = spanned_pages;
5050 * Provide a more accurate estimation if there are holes within
5051 * the zone and SPARSEMEM is in use. If there are holes within the
5052 * zone, each populated memory region may cost us one or two extra
5053 * memmap pages due to alignment because memmap pages for each
5054 * populated regions may not naturally algined on page boundary.
5055 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5057 if (spanned_pages > present_pages + (present_pages >> 4) &&
5058 IS_ENABLED(CONFIG_SPARSEMEM))
5059 pages = present_pages;
5061 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5065 * Set up the zone data structures:
5066 * - mark all pages reserved
5067 * - mark all memory queues empty
5068 * - clear the memory bitmaps
5070 * NOTE: pgdat should get zeroed by caller.
5072 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
5073 unsigned long node_start_pfn, unsigned long node_end_pfn)
5076 int nid = pgdat->node_id;
5077 unsigned long zone_start_pfn = pgdat->node_start_pfn;
5080 pgdat_resize_init(pgdat);
5081 #ifdef CONFIG_NUMA_BALANCING
5082 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5083 pgdat->numabalancing_migrate_nr_pages = 0;
5084 pgdat->numabalancing_migrate_next_window = jiffies;
5086 init_waitqueue_head(&pgdat->kswapd_wait);
5087 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5088 pgdat_page_ext_init(pgdat);
5090 for (j = 0; j < MAX_NR_ZONES; j++) {
5091 struct zone *zone = pgdat->node_zones + j;
5092 unsigned long size, realsize, freesize, memmap_pages;
5094 size = zone->spanned_pages;
5095 realsize = freesize = zone->present_pages;
5098 * Adjust freesize so that it accounts for how much memory
5099 * is used by this zone for memmap. This affects the watermark
5100 * and per-cpu initialisations
5102 memmap_pages = calc_memmap_size(size, realsize);
5103 if (!is_highmem_idx(j)) {
5104 if (freesize >= memmap_pages) {
5105 freesize -= memmap_pages;
5108 " %s zone: %lu pages used for memmap\n",
5109 zone_names[j], memmap_pages);
5112 " %s zone: %lu pages exceeds freesize %lu\n",
5113 zone_names[j], memmap_pages, freesize);
5116 /* Account for reserved pages */
5117 if (j == 0 && freesize > dma_reserve) {
5118 freesize -= dma_reserve;
5119 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5120 zone_names[0], dma_reserve);
5123 if (!is_highmem_idx(j))
5124 nr_kernel_pages += freesize;
5125 /* Charge for highmem memmap if there are enough kernel pages */
5126 else if (nr_kernel_pages > memmap_pages * 2)
5127 nr_kernel_pages -= memmap_pages;
5128 nr_all_pages += freesize;
5131 * Set an approximate value for lowmem here, it will be adjusted
5132 * when the bootmem allocator frees pages into the buddy system.
5133 * And all highmem pages will be managed by the buddy system.
5135 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5138 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5140 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5142 zone->name = zone_names[j];
5143 spin_lock_init(&zone->lock);
5144 spin_lock_init(&zone->lru_lock);
5145 zone_seqlock_init(zone);
5146 zone->zone_pgdat = pgdat;
5147 zone_pcp_init(zone);
5149 /* For bootup, initialized properly in watermark setup */
5150 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5152 lruvec_init(&zone->lruvec);
5156 set_pageblock_order();
5157 setup_usemap(pgdat, zone, zone_start_pfn, size);
5158 ret = init_currently_empty_zone(zone, zone_start_pfn,
5159 size, MEMMAP_EARLY);
5161 memmap_init(size, nid, j, zone_start_pfn);
5162 zone_start_pfn += size;
5166 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5168 /* Skip empty nodes */
5169 if (!pgdat->node_spanned_pages)
5172 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5173 /* ia64 gets its own node_mem_map, before this, without bootmem */
5174 if (!pgdat->node_mem_map) {
5175 unsigned long size, start, end;
5179 * The zone's endpoints aren't required to be MAX_ORDER
5180 * aligned but the node_mem_map endpoints must be in order
5181 * for the buddy allocator to function correctly.
5183 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5184 end = pgdat_end_pfn(pgdat);
5185 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5186 size = (end - start) * sizeof(struct page);
5187 map = alloc_remap(pgdat->node_id, size);
5189 map = memblock_virt_alloc_node_nopanic(size,
5191 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
5193 #ifndef CONFIG_NEED_MULTIPLE_NODES
5195 * With no DISCONTIG, the global mem_map is just set as node 0's
5197 if (pgdat == NODE_DATA(0)) {
5198 mem_map = NODE_DATA(0)->node_mem_map;
5199 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5200 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5201 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
5202 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5205 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5208 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5209 unsigned long node_start_pfn, unsigned long *zholes_size)
5211 pg_data_t *pgdat = NODE_DATA(nid);
5212 unsigned long start_pfn = 0;
5213 unsigned long end_pfn = 0;
5215 /* pg_data_t should be reset to zero when it's allocated */
5216 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5218 reset_deferred_meminit(pgdat);
5219 pgdat->node_id = nid;
5220 pgdat->node_start_pfn = node_start_pfn;
5221 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5222 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5223 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5224 (u64)start_pfn << PAGE_SHIFT, ((u64)end_pfn << PAGE_SHIFT) - 1);
5226 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5227 zones_size, zholes_size);
5229 alloc_node_mem_map(pgdat);
5230 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5231 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5232 nid, (unsigned long)pgdat,
5233 (unsigned long)pgdat->node_mem_map);
5236 free_area_init_core(pgdat, start_pfn, end_pfn);
5239 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5241 #if MAX_NUMNODES > 1
5243 * Figure out the number of possible node ids.
5245 void __init setup_nr_node_ids(void)
5248 unsigned int highest = 0;
5250 for_each_node_mask(node, node_possible_map)
5252 nr_node_ids = highest + 1;
5257 * node_map_pfn_alignment - determine the maximum internode alignment
5259 * This function should be called after node map is populated and sorted.
5260 * It calculates the maximum power of two alignment which can distinguish
5263 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5264 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5265 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5266 * shifted, 1GiB is enough and this function will indicate so.
5268 * This is used to test whether pfn -> nid mapping of the chosen memory
5269 * model has fine enough granularity to avoid incorrect mapping for the
5270 * populated node map.
5272 * Returns the determined alignment in pfn's. 0 if there is no alignment
5273 * requirement (single node).
5275 unsigned long __init node_map_pfn_alignment(void)
5277 unsigned long accl_mask = 0, last_end = 0;
5278 unsigned long start, end, mask;
5282 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5283 if (!start || last_nid < 0 || last_nid == nid) {
5290 * Start with a mask granular enough to pin-point to the
5291 * start pfn and tick off bits one-by-one until it becomes
5292 * too coarse to separate the current node from the last.
5294 mask = ~((1 << __ffs(start)) - 1);
5295 while (mask && last_end <= (start & (mask << 1)))
5298 /* accumulate all internode masks */
5302 /* convert mask to number of pages */
5303 return ~accl_mask + 1;
5306 /* Find the lowest pfn for a node */
5307 static unsigned long __init find_min_pfn_for_node(int nid)
5309 unsigned long min_pfn = ULONG_MAX;
5310 unsigned long start_pfn;
5313 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5314 min_pfn = min(min_pfn, start_pfn);
5316 if (min_pfn == ULONG_MAX) {
5318 "Could not find start_pfn for node %d\n", nid);
5326 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5328 * It returns the minimum PFN based on information provided via
5329 * memblock_set_node().
5331 unsigned long __init find_min_pfn_with_active_regions(void)
5333 return find_min_pfn_for_node(MAX_NUMNODES);
5337 * early_calculate_totalpages()
5338 * Sum pages in active regions for movable zone.
5339 * Populate N_MEMORY for calculating usable_nodes.
5341 static unsigned long __init early_calculate_totalpages(void)
5343 unsigned long totalpages = 0;
5344 unsigned long start_pfn, end_pfn;
5347 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5348 unsigned long pages = end_pfn - start_pfn;
5350 totalpages += pages;
5352 node_set_state(nid, N_MEMORY);
5358 * Find the PFN the Movable zone begins in each node. Kernel memory
5359 * is spread evenly between nodes as long as the nodes have enough
5360 * memory. When they don't, some nodes will have more kernelcore than
5363 static void __init find_zone_movable_pfns_for_nodes(void)
5366 unsigned long usable_startpfn;
5367 unsigned long kernelcore_node, kernelcore_remaining;
5368 /* save the state before borrow the nodemask */
5369 nodemask_t saved_node_state = node_states[N_MEMORY];
5370 unsigned long totalpages = early_calculate_totalpages();
5371 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5372 struct memblock_region *r;
5374 /* Need to find movable_zone earlier when movable_node is specified. */
5375 find_usable_zone_for_movable();
5378 * If movable_node is specified, ignore kernelcore and movablecore
5381 if (movable_node_is_enabled()) {
5382 for_each_memblock(memory, r) {
5383 if (!memblock_is_hotpluggable(r))
5388 usable_startpfn = PFN_DOWN(r->base);
5389 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5390 min(usable_startpfn, zone_movable_pfn[nid]) :
5398 * If movablecore=nn[KMG] was specified, calculate what size of
5399 * kernelcore that corresponds so that memory usable for
5400 * any allocation type is evenly spread. If both kernelcore
5401 * and movablecore are specified, then the value of kernelcore
5402 * will be used for required_kernelcore if it's greater than
5403 * what movablecore would have allowed.
5405 if (required_movablecore) {
5406 unsigned long corepages;
5409 * Round-up so that ZONE_MOVABLE is at least as large as what
5410 * was requested by the user
5412 required_movablecore =
5413 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5414 corepages = totalpages - required_movablecore;
5416 required_kernelcore = max(required_kernelcore, corepages);
5419 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5420 if (!required_kernelcore)
5423 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5424 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5427 /* Spread kernelcore memory as evenly as possible throughout nodes */
5428 kernelcore_node = required_kernelcore / usable_nodes;
5429 for_each_node_state(nid, N_MEMORY) {
5430 unsigned long start_pfn, end_pfn;
5433 * Recalculate kernelcore_node if the division per node
5434 * now exceeds what is necessary to satisfy the requested
5435 * amount of memory for the kernel
5437 if (required_kernelcore < kernelcore_node)
5438 kernelcore_node = required_kernelcore / usable_nodes;
5441 * As the map is walked, we track how much memory is usable
5442 * by the kernel using kernelcore_remaining. When it is
5443 * 0, the rest of the node is usable by ZONE_MOVABLE
5445 kernelcore_remaining = kernelcore_node;
5447 /* Go through each range of PFNs within this node */
5448 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5449 unsigned long size_pages;
5451 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5452 if (start_pfn >= end_pfn)
5455 /* Account for what is only usable for kernelcore */
5456 if (start_pfn < usable_startpfn) {
5457 unsigned long kernel_pages;
5458 kernel_pages = min(end_pfn, usable_startpfn)
5461 kernelcore_remaining -= min(kernel_pages,
5462 kernelcore_remaining);
5463 required_kernelcore -= min(kernel_pages,
5464 required_kernelcore);
5466 /* Continue if range is now fully accounted */
5467 if (end_pfn <= usable_startpfn) {
5470 * Push zone_movable_pfn to the end so
5471 * that if we have to rebalance
5472 * kernelcore across nodes, we will
5473 * not double account here
5475 zone_movable_pfn[nid] = end_pfn;
5478 start_pfn = usable_startpfn;
5482 * The usable PFN range for ZONE_MOVABLE is from
5483 * start_pfn->end_pfn. Calculate size_pages as the
5484 * number of pages used as kernelcore
5486 size_pages = end_pfn - start_pfn;
5487 if (size_pages > kernelcore_remaining)
5488 size_pages = kernelcore_remaining;
5489 zone_movable_pfn[nid] = start_pfn + size_pages;
5492 * Some kernelcore has been met, update counts and
5493 * break if the kernelcore for this node has been
5496 required_kernelcore -= min(required_kernelcore,
5498 kernelcore_remaining -= size_pages;
5499 if (!kernelcore_remaining)
5505 * If there is still required_kernelcore, we do another pass with one
5506 * less node in the count. This will push zone_movable_pfn[nid] further
5507 * along on the nodes that still have memory until kernelcore is
5511 if (usable_nodes && required_kernelcore > usable_nodes)
5515 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5516 for (nid = 0; nid < MAX_NUMNODES; nid++)
5517 zone_movable_pfn[nid] =
5518 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5521 /* restore the node_state */
5522 node_states[N_MEMORY] = saved_node_state;
5525 /* Any regular or high memory on that node ? */
5526 static void check_for_memory(pg_data_t *pgdat, int nid)
5528 enum zone_type zone_type;
5530 if (N_MEMORY == N_NORMAL_MEMORY)
5533 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5534 struct zone *zone = &pgdat->node_zones[zone_type];
5535 if (populated_zone(zone)) {
5536 node_set_state(nid, N_HIGH_MEMORY);
5537 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5538 zone_type <= ZONE_NORMAL)
5539 node_set_state(nid, N_NORMAL_MEMORY);
5546 * free_area_init_nodes - Initialise all pg_data_t and zone data
5547 * @max_zone_pfn: an array of max PFNs for each zone
5549 * This will call free_area_init_node() for each active node in the system.
5550 * Using the page ranges provided by memblock_set_node(), the size of each
5551 * zone in each node and their holes is calculated. If the maximum PFN
5552 * between two adjacent zones match, it is assumed that the zone is empty.
5553 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5554 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5555 * starts where the previous one ended. For example, ZONE_DMA32 starts
5556 * at arch_max_dma_pfn.
5558 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5560 unsigned long start_pfn, end_pfn;
5563 /* Record where the zone boundaries are */
5564 memset(arch_zone_lowest_possible_pfn, 0,
5565 sizeof(arch_zone_lowest_possible_pfn));
5566 memset(arch_zone_highest_possible_pfn, 0,
5567 sizeof(arch_zone_highest_possible_pfn));
5568 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5569 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5570 for (i = 1; i < MAX_NR_ZONES; i++) {
5571 if (i == ZONE_MOVABLE)
5573 arch_zone_lowest_possible_pfn[i] =
5574 arch_zone_highest_possible_pfn[i-1];
5575 arch_zone_highest_possible_pfn[i] =
5576 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5578 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5579 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5581 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5582 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5583 find_zone_movable_pfns_for_nodes();
5585 /* Print out the zone ranges */
5586 pr_info("Zone ranges:\n");
5587 for (i = 0; i < MAX_NR_ZONES; i++) {
5588 if (i == ZONE_MOVABLE)
5590 pr_info(" %-8s ", zone_names[i]);
5591 if (arch_zone_lowest_possible_pfn[i] ==
5592 arch_zone_highest_possible_pfn[i])
5595 pr_cont("[mem %#018Lx-%#018Lx]\n",
5596 (u64)arch_zone_lowest_possible_pfn[i]
5598 ((u64)arch_zone_highest_possible_pfn[i]
5599 << PAGE_SHIFT) - 1);
5602 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5603 pr_info("Movable zone start for each node\n");
5604 for (i = 0; i < MAX_NUMNODES; i++) {
5605 if (zone_movable_pfn[i])
5606 pr_info(" Node %d: %#018Lx\n", i,
5607 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5610 /* Print out the early node map */
5611 pr_info("Early memory node ranges\n");
5612 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5613 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5614 (u64)start_pfn << PAGE_SHIFT,
5615 ((u64)end_pfn << PAGE_SHIFT) - 1);
5617 /* Initialise every node */
5618 mminit_verify_pageflags_layout();
5619 setup_nr_node_ids();
5620 for_each_online_node(nid) {
5621 pg_data_t *pgdat = NODE_DATA(nid);
5622 free_area_init_node(nid, NULL,
5623 find_min_pfn_for_node(nid), NULL);
5625 /* Any memory on that node */
5626 if (pgdat->node_present_pages)
5627 node_set_state(nid, N_MEMORY);
5628 check_for_memory(pgdat, nid);
5632 static int __init cmdline_parse_core(char *p, unsigned long *core)
5634 unsigned long long coremem;
5638 coremem = memparse(p, &p);
5639 *core = coremem >> PAGE_SHIFT;
5641 /* Paranoid check that UL is enough for the coremem value */
5642 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5648 * kernelcore=size sets the amount of memory for use for allocations that
5649 * cannot be reclaimed or migrated.
5651 static int __init cmdline_parse_kernelcore(char *p)
5653 return cmdline_parse_core(p, &required_kernelcore);
5657 * movablecore=size sets the amount of memory for use for allocations that
5658 * can be reclaimed or migrated.
5660 static int __init cmdline_parse_movablecore(char *p)
5662 return cmdline_parse_core(p, &required_movablecore);
5665 early_param("kernelcore", cmdline_parse_kernelcore);
5666 early_param("movablecore", cmdline_parse_movablecore);
5668 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5670 void adjust_managed_page_count(struct page *page, long count)
5672 spin_lock(&managed_page_count_lock);
5673 page_zone(page)->managed_pages += count;
5674 totalram_pages += count;
5675 #ifdef CONFIG_HIGHMEM
5676 if (PageHighMem(page))
5677 totalhigh_pages += count;
5679 spin_unlock(&managed_page_count_lock);
5681 EXPORT_SYMBOL(adjust_managed_page_count);
5683 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5686 unsigned long pages = 0;
5688 start = (void *)PAGE_ALIGN((unsigned long)start);
5689 end = (void *)((unsigned long)end & PAGE_MASK);
5690 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5691 if ((unsigned int)poison <= 0xFF)
5692 memset(pos, poison, PAGE_SIZE);
5693 free_reserved_page(virt_to_page(pos));
5697 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5698 s, pages << (PAGE_SHIFT - 10), start, end);
5702 EXPORT_SYMBOL(free_reserved_area);
5704 #ifdef CONFIG_HIGHMEM
5705 void free_highmem_page(struct page *page)
5707 __free_reserved_page(page);
5709 page_zone(page)->managed_pages++;
5715 void __init mem_init_print_info(const char *str)
5717 unsigned long physpages, codesize, datasize, rosize, bss_size;
5718 unsigned long init_code_size, init_data_size;
5720 physpages = get_num_physpages();
5721 codesize = _etext - _stext;
5722 datasize = _edata - _sdata;
5723 rosize = __end_rodata - __start_rodata;
5724 bss_size = __bss_stop - __bss_start;
5725 init_data_size = __init_end - __init_begin;
5726 init_code_size = _einittext - _sinittext;
5729 * Detect special cases and adjust section sizes accordingly:
5730 * 1) .init.* may be embedded into .data sections
5731 * 2) .init.text.* may be out of [__init_begin, __init_end],
5732 * please refer to arch/tile/kernel/vmlinux.lds.S.
5733 * 3) .rodata.* may be embedded into .text or .data sections.
5735 #define adj_init_size(start, end, size, pos, adj) \
5737 if (start <= pos && pos < end && size > adj) \
5741 adj_init_size(__init_begin, __init_end, init_data_size,
5742 _sinittext, init_code_size);
5743 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5744 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5745 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5746 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5748 #undef adj_init_size
5750 pr_info("Memory: %luK/%luK available "
5751 "(%luK kernel code, %luK rwdata, %luK rodata, "
5752 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5753 #ifdef CONFIG_HIGHMEM
5757 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5758 codesize >> 10, datasize >> 10, rosize >> 10,
5759 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5760 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5761 totalcma_pages << (PAGE_SHIFT-10),
5762 #ifdef CONFIG_HIGHMEM
5763 totalhigh_pages << (PAGE_SHIFT-10),
5765 str ? ", " : "", str ? str : "");
5769 * set_dma_reserve - set the specified number of pages reserved in the first zone
5770 * @new_dma_reserve: The number of pages to mark reserved
5772 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5773 * In the DMA zone, a significant percentage may be consumed by kernel image
5774 * and other unfreeable allocations which can skew the watermarks badly. This
5775 * function may optionally be used to account for unfreeable pages in the
5776 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5777 * smaller per-cpu batchsize.
5779 void __init set_dma_reserve(unsigned long new_dma_reserve)
5781 dma_reserve = new_dma_reserve;
5784 void __init free_area_init(unsigned long *zones_size)
5786 free_area_init_node(0, zones_size,
5787 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5790 static int page_alloc_cpu_notify(struct notifier_block *self,
5791 unsigned long action, void *hcpu)
5793 int cpu = (unsigned long)hcpu;
5795 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5796 lru_add_drain_cpu(cpu);
5800 * Spill the event counters of the dead processor
5801 * into the current processors event counters.
5802 * This artificially elevates the count of the current
5805 vm_events_fold_cpu(cpu);
5808 * Zero the differential counters of the dead processor
5809 * so that the vm statistics are consistent.
5811 * This is only okay since the processor is dead and cannot
5812 * race with what we are doing.
5814 cpu_vm_stats_fold(cpu);
5819 void __init page_alloc_init(void)
5821 hotcpu_notifier(page_alloc_cpu_notify, 0);
5825 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5826 * or min_free_kbytes changes.
5828 static void calculate_totalreserve_pages(void)
5830 struct pglist_data *pgdat;
5831 unsigned long reserve_pages = 0;
5832 enum zone_type i, j;
5834 for_each_online_pgdat(pgdat) {
5835 for (i = 0; i < MAX_NR_ZONES; i++) {
5836 struct zone *zone = pgdat->node_zones + i;
5839 /* Find valid and maximum lowmem_reserve in the zone */
5840 for (j = i; j < MAX_NR_ZONES; j++) {
5841 if (zone->lowmem_reserve[j] > max)
5842 max = zone->lowmem_reserve[j];
5845 /* we treat the high watermark as reserved pages. */
5846 max += high_wmark_pages(zone);
5848 if (max > zone->managed_pages)
5849 max = zone->managed_pages;
5850 reserve_pages += max;
5852 * Lowmem reserves are not available to
5853 * GFP_HIGHUSER page cache allocations and
5854 * kswapd tries to balance zones to their high
5855 * watermark. As a result, neither should be
5856 * regarded as dirtyable memory, to prevent a
5857 * situation where reclaim has to clean pages
5858 * in order to balance the zones.
5860 zone->dirty_balance_reserve = max;
5863 dirty_balance_reserve = reserve_pages;
5864 totalreserve_pages = reserve_pages;
5868 * setup_per_zone_lowmem_reserve - called whenever
5869 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5870 * has a correct pages reserved value, so an adequate number of
5871 * pages are left in the zone after a successful __alloc_pages().
5873 static void setup_per_zone_lowmem_reserve(void)
5875 struct pglist_data *pgdat;
5876 enum zone_type j, idx;
5878 for_each_online_pgdat(pgdat) {
5879 for (j = 0; j < MAX_NR_ZONES; j++) {
5880 struct zone *zone = pgdat->node_zones + j;
5881 unsigned long managed_pages = zone->managed_pages;
5883 zone->lowmem_reserve[j] = 0;
5887 struct zone *lower_zone;
5891 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5892 sysctl_lowmem_reserve_ratio[idx] = 1;
5894 lower_zone = pgdat->node_zones + idx;
5895 lower_zone->lowmem_reserve[j] = managed_pages /
5896 sysctl_lowmem_reserve_ratio[idx];
5897 managed_pages += lower_zone->managed_pages;
5902 /* update totalreserve_pages */
5903 calculate_totalreserve_pages();
5906 static void __setup_per_zone_wmarks(void)
5908 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5909 unsigned long lowmem_pages = 0;
5911 unsigned long flags;
5913 /* Calculate total number of !ZONE_HIGHMEM pages */
5914 for_each_zone(zone) {
5915 if (!is_highmem(zone))
5916 lowmem_pages += zone->managed_pages;
5919 for_each_zone(zone) {
5922 spin_lock_irqsave(&zone->lock, flags);
5923 tmp = (u64)pages_min * zone->managed_pages;
5924 do_div(tmp, lowmem_pages);
5925 if (is_highmem(zone)) {
5927 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5928 * need highmem pages, so cap pages_min to a small
5931 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5932 * deltas control asynch page reclaim, and so should
5933 * not be capped for highmem.
5935 unsigned long min_pages;
5937 min_pages = zone->managed_pages / 1024;
5938 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5939 zone->watermark[WMARK_MIN] = min_pages;
5942 * If it's a lowmem zone, reserve a number of pages
5943 * proportionate to the zone's size.
5945 zone->watermark[WMARK_MIN] = tmp;
5948 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5949 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5951 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
5952 high_wmark_pages(zone) - low_wmark_pages(zone) -
5953 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
5955 setup_zone_migrate_reserve(zone);
5956 spin_unlock_irqrestore(&zone->lock, flags);
5959 /* update totalreserve_pages */
5960 calculate_totalreserve_pages();
5964 * setup_per_zone_wmarks - called when min_free_kbytes changes
5965 * or when memory is hot-{added|removed}
5967 * Ensures that the watermark[min,low,high] values for each zone are set
5968 * correctly with respect to min_free_kbytes.
5970 void setup_per_zone_wmarks(void)
5972 mutex_lock(&zonelists_mutex);
5973 __setup_per_zone_wmarks();
5974 mutex_unlock(&zonelists_mutex);
5978 * The inactive anon list should be small enough that the VM never has to
5979 * do too much work, but large enough that each inactive page has a chance
5980 * to be referenced again before it is swapped out.
5982 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5983 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5984 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5985 * the anonymous pages are kept on the inactive list.
5988 * memory ratio inactive anon
5989 * -------------------------------------
5998 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6000 unsigned int gb, ratio;
6002 /* Zone size in gigabytes */
6003 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6005 ratio = int_sqrt(10 * gb);
6009 zone->inactive_ratio = ratio;
6012 static void __meminit setup_per_zone_inactive_ratio(void)
6017 calculate_zone_inactive_ratio(zone);
6021 * Initialise min_free_kbytes.
6023 * For small machines we want it small (128k min). For large machines
6024 * we want it large (64MB max). But it is not linear, because network
6025 * bandwidth does not increase linearly with machine size. We use
6027 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6028 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6044 int __meminit init_per_zone_wmark_min(void)
6046 unsigned long lowmem_kbytes;
6047 int new_min_free_kbytes;
6049 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6050 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6052 if (new_min_free_kbytes > user_min_free_kbytes) {
6053 min_free_kbytes = new_min_free_kbytes;
6054 if (min_free_kbytes < 128)
6055 min_free_kbytes = 128;
6056 if (min_free_kbytes > 65536)
6057 min_free_kbytes = 65536;
6059 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6060 new_min_free_kbytes, user_min_free_kbytes);
6062 setup_per_zone_wmarks();
6063 refresh_zone_stat_thresholds();
6064 setup_per_zone_lowmem_reserve();
6065 setup_per_zone_inactive_ratio();
6068 module_init(init_per_zone_wmark_min)
6071 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6072 * that we can call two helper functions whenever min_free_kbytes
6075 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6076 void __user *buffer, size_t *length, loff_t *ppos)
6080 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6085 user_min_free_kbytes = min_free_kbytes;
6086 setup_per_zone_wmarks();
6092 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6093 void __user *buffer, size_t *length, loff_t *ppos)
6098 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6103 zone->min_unmapped_pages = (zone->managed_pages *
6104 sysctl_min_unmapped_ratio) / 100;
6108 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6109 void __user *buffer, size_t *length, loff_t *ppos)
6114 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6119 zone->min_slab_pages = (zone->managed_pages *
6120 sysctl_min_slab_ratio) / 100;
6126 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6127 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6128 * whenever sysctl_lowmem_reserve_ratio changes.
6130 * The reserve ratio obviously has absolutely no relation with the
6131 * minimum watermarks. The lowmem reserve ratio can only make sense
6132 * if in function of the boot time zone sizes.
6134 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6135 void __user *buffer, size_t *length, loff_t *ppos)
6137 proc_dointvec_minmax(table, write, buffer, length, ppos);
6138 setup_per_zone_lowmem_reserve();
6143 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6144 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6145 * pagelist can have before it gets flushed back to buddy allocator.
6147 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6148 void __user *buffer, size_t *length, loff_t *ppos)
6151 int old_percpu_pagelist_fraction;
6154 mutex_lock(&pcp_batch_high_lock);
6155 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6157 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6158 if (!write || ret < 0)
6161 /* Sanity checking to avoid pcp imbalance */
6162 if (percpu_pagelist_fraction &&
6163 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6164 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6170 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6173 for_each_populated_zone(zone) {
6176 for_each_possible_cpu(cpu)
6177 pageset_set_high_and_batch(zone,
6178 per_cpu_ptr(zone->pageset, cpu));
6181 mutex_unlock(&pcp_batch_high_lock);
6186 int hashdist = HASHDIST_DEFAULT;
6188 static int __init set_hashdist(char *str)
6192 hashdist = simple_strtoul(str, &str, 0);
6195 __setup("hashdist=", set_hashdist);
6199 * allocate a large system hash table from bootmem
6200 * - it is assumed that the hash table must contain an exact power-of-2
6201 * quantity of entries
6202 * - limit is the number of hash buckets, not the total allocation size
6204 void *__init alloc_large_system_hash(const char *tablename,
6205 unsigned long bucketsize,
6206 unsigned long numentries,
6209 unsigned int *_hash_shift,
6210 unsigned int *_hash_mask,
6211 unsigned long low_limit,
6212 unsigned long high_limit)
6214 unsigned long long max = high_limit;
6215 unsigned long log2qty, size;
6218 /* allow the kernel cmdline to have a say */
6220 /* round applicable memory size up to nearest megabyte */
6221 numentries = nr_kernel_pages;
6223 /* It isn't necessary when PAGE_SIZE >= 1MB */
6224 if (PAGE_SHIFT < 20)
6225 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6227 /* limit to 1 bucket per 2^scale bytes of low memory */
6228 if (scale > PAGE_SHIFT)
6229 numentries >>= (scale - PAGE_SHIFT);
6231 numentries <<= (PAGE_SHIFT - scale);
6233 /* Make sure we've got at least a 0-order allocation.. */
6234 if (unlikely(flags & HASH_SMALL)) {
6235 /* Makes no sense without HASH_EARLY */
6236 WARN_ON(!(flags & HASH_EARLY));
6237 if (!(numentries >> *_hash_shift)) {
6238 numentries = 1UL << *_hash_shift;
6239 BUG_ON(!numentries);
6241 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6242 numentries = PAGE_SIZE / bucketsize;
6244 numentries = roundup_pow_of_two(numentries);
6246 /* limit allocation size to 1/16 total memory by default */
6248 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6249 do_div(max, bucketsize);
6251 max = min(max, 0x80000000ULL);
6253 if (numentries < low_limit)
6254 numentries = low_limit;
6255 if (numentries > max)
6258 log2qty = ilog2(numentries);
6261 size = bucketsize << log2qty;
6262 if (flags & HASH_EARLY)
6263 table = memblock_virt_alloc_nopanic(size, 0);
6265 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6268 * If bucketsize is not a power-of-two, we may free
6269 * some pages at the end of hash table which
6270 * alloc_pages_exact() automatically does
6272 if (get_order(size) < MAX_ORDER) {
6273 table = alloc_pages_exact(size, GFP_ATOMIC);
6274 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6277 } while (!table && size > PAGE_SIZE && --log2qty);
6280 panic("Failed to allocate %s hash table\n", tablename);
6282 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6285 ilog2(size) - PAGE_SHIFT,
6289 *_hash_shift = log2qty;
6291 *_hash_mask = (1 << log2qty) - 1;
6296 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6297 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6300 #ifdef CONFIG_SPARSEMEM
6301 return __pfn_to_section(pfn)->pageblock_flags;
6303 return zone->pageblock_flags;
6304 #endif /* CONFIG_SPARSEMEM */
6307 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6309 #ifdef CONFIG_SPARSEMEM
6310 pfn &= (PAGES_PER_SECTION-1);
6311 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6313 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6314 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6315 #endif /* CONFIG_SPARSEMEM */
6319 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6320 * @page: The page within the block of interest
6321 * @pfn: The target page frame number
6322 * @end_bitidx: The last bit of interest to retrieve
6323 * @mask: mask of bits that the caller is interested in
6325 * Return: pageblock_bits flags
6327 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6328 unsigned long end_bitidx,
6332 unsigned long *bitmap;
6333 unsigned long bitidx, word_bitidx;
6336 zone = page_zone(page);
6337 bitmap = get_pageblock_bitmap(zone, pfn);
6338 bitidx = pfn_to_bitidx(zone, pfn);
6339 word_bitidx = bitidx / BITS_PER_LONG;
6340 bitidx &= (BITS_PER_LONG-1);
6342 word = bitmap[word_bitidx];
6343 bitidx += end_bitidx;
6344 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6348 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6349 * @page: The page within the block of interest
6350 * @flags: The flags to set
6351 * @pfn: The target page frame number
6352 * @end_bitidx: The last bit of interest
6353 * @mask: mask of bits that the caller is interested in
6355 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6357 unsigned long end_bitidx,
6361 unsigned long *bitmap;
6362 unsigned long bitidx, word_bitidx;
6363 unsigned long old_word, word;
6365 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6367 zone = page_zone(page);
6368 bitmap = get_pageblock_bitmap(zone, pfn);
6369 bitidx = pfn_to_bitidx(zone, pfn);
6370 word_bitidx = bitidx / BITS_PER_LONG;
6371 bitidx &= (BITS_PER_LONG-1);
6373 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6375 bitidx += end_bitidx;
6376 mask <<= (BITS_PER_LONG - bitidx - 1);
6377 flags <<= (BITS_PER_LONG - bitidx - 1);
6379 word = READ_ONCE(bitmap[word_bitidx]);
6381 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6382 if (word == old_word)
6389 * This function checks whether pageblock includes unmovable pages or not.
6390 * If @count is not zero, it is okay to include less @count unmovable pages
6392 * PageLRU check without isolation or lru_lock could race so that
6393 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6394 * expect this function should be exact.
6396 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6397 bool skip_hwpoisoned_pages)
6399 unsigned long pfn, iter, found;
6403 * For avoiding noise data, lru_add_drain_all() should be called
6404 * If ZONE_MOVABLE, the zone never contains unmovable pages
6406 if (zone_idx(zone) == ZONE_MOVABLE)
6408 mt = get_pageblock_migratetype(page);
6409 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6412 pfn = page_to_pfn(page);
6413 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6414 unsigned long check = pfn + iter;
6416 if (!pfn_valid_within(check))
6419 page = pfn_to_page(check);
6422 * Hugepages are not in LRU lists, but they're movable.
6423 * We need not scan over tail pages bacause we don't
6424 * handle each tail page individually in migration.
6426 if (PageHuge(page)) {
6427 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6432 * We can't use page_count without pin a page
6433 * because another CPU can free compound page.
6434 * This check already skips compound tails of THP
6435 * because their page->_count is zero at all time.
6437 if (!atomic_read(&page->_count)) {
6438 if (PageBuddy(page))
6439 iter += (1 << page_order(page)) - 1;
6444 * The HWPoisoned page may be not in buddy system, and
6445 * page_count() is not 0.
6447 if (skip_hwpoisoned_pages && PageHWPoison(page))
6453 * If there are RECLAIMABLE pages, we need to check
6454 * it. But now, memory offline itself doesn't call
6455 * shrink_node_slabs() and it still to be fixed.
6458 * If the page is not RAM, page_count()should be 0.
6459 * we don't need more check. This is an _used_ not-movable page.
6461 * The problematic thing here is PG_reserved pages. PG_reserved
6462 * is set to both of a memory hole page and a _used_ kernel
6471 bool is_pageblock_removable_nolock(struct page *page)
6477 * We have to be careful here because we are iterating over memory
6478 * sections which are not zone aware so we might end up outside of
6479 * the zone but still within the section.
6480 * We have to take care about the node as well. If the node is offline
6481 * its NODE_DATA will be NULL - see page_zone.
6483 if (!node_online(page_to_nid(page)))
6486 zone = page_zone(page);
6487 pfn = page_to_pfn(page);
6488 if (!zone_spans_pfn(zone, pfn))
6491 return !has_unmovable_pages(zone, page, 0, true);
6496 static unsigned long pfn_max_align_down(unsigned long pfn)
6498 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6499 pageblock_nr_pages) - 1);
6502 static unsigned long pfn_max_align_up(unsigned long pfn)
6504 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6505 pageblock_nr_pages));
6508 /* [start, end) must belong to a single zone. */
6509 static int __alloc_contig_migrate_range(struct compact_control *cc,
6510 unsigned long start, unsigned long end)
6512 /* This function is based on compact_zone() from compaction.c. */
6513 unsigned long nr_reclaimed;
6514 unsigned long pfn = start;
6515 unsigned int tries = 0;
6520 while (pfn < end || !list_empty(&cc->migratepages)) {
6521 if (fatal_signal_pending(current)) {
6526 if (list_empty(&cc->migratepages)) {
6527 cc->nr_migratepages = 0;
6528 pfn = isolate_migratepages_range(cc, pfn, end);
6534 } else if (++tries == 5) {
6535 ret = ret < 0 ? ret : -EBUSY;
6539 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6541 cc->nr_migratepages -= nr_reclaimed;
6543 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6544 NULL, 0, cc->mode, MR_CMA);
6547 putback_movable_pages(&cc->migratepages);
6554 * alloc_contig_range() -- tries to allocate given range of pages
6555 * @start: start PFN to allocate
6556 * @end: one-past-the-last PFN to allocate
6557 * @migratetype: migratetype of the underlaying pageblocks (either
6558 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6559 * in range must have the same migratetype and it must
6560 * be either of the two.
6562 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6563 * aligned, however it's the caller's responsibility to guarantee that
6564 * we are the only thread that changes migrate type of pageblocks the
6567 * The PFN range must belong to a single zone.
6569 * Returns zero on success or negative error code. On success all
6570 * pages which PFN is in [start, end) are allocated for the caller and
6571 * need to be freed with free_contig_range().
6573 int alloc_contig_range(unsigned long start, unsigned long end,
6574 unsigned migratetype)
6576 unsigned long outer_start, outer_end;
6579 struct compact_control cc = {
6580 .nr_migratepages = 0,
6582 .zone = page_zone(pfn_to_page(start)),
6583 .mode = MIGRATE_SYNC,
6584 .ignore_skip_hint = true,
6586 INIT_LIST_HEAD(&cc.migratepages);
6589 * What we do here is we mark all pageblocks in range as
6590 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6591 * have different sizes, and due to the way page allocator
6592 * work, we align the range to biggest of the two pages so
6593 * that page allocator won't try to merge buddies from
6594 * different pageblocks and change MIGRATE_ISOLATE to some
6595 * other migration type.
6597 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6598 * migrate the pages from an unaligned range (ie. pages that
6599 * we are interested in). This will put all the pages in
6600 * range back to page allocator as MIGRATE_ISOLATE.
6602 * When this is done, we take the pages in range from page
6603 * allocator removing them from the buddy system. This way
6604 * page allocator will never consider using them.
6606 * This lets us mark the pageblocks back as
6607 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6608 * aligned range but not in the unaligned, original range are
6609 * put back to page allocator so that buddy can use them.
6612 ret = start_isolate_page_range(pfn_max_align_down(start),
6613 pfn_max_align_up(end), migratetype,
6618 ret = __alloc_contig_migrate_range(&cc, start, end);
6623 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6624 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6625 * more, all pages in [start, end) are free in page allocator.
6626 * What we are going to do is to allocate all pages from
6627 * [start, end) (that is remove them from page allocator).
6629 * The only problem is that pages at the beginning and at the
6630 * end of interesting range may be not aligned with pages that
6631 * page allocator holds, ie. they can be part of higher order
6632 * pages. Because of this, we reserve the bigger range and
6633 * once this is done free the pages we are not interested in.
6635 * We don't have to hold zone->lock here because the pages are
6636 * isolated thus they won't get removed from buddy.
6639 lru_add_drain_all();
6640 drain_all_pages(cc.zone);
6643 outer_start = start;
6644 while (!PageBuddy(pfn_to_page(outer_start))) {
6645 if (++order >= MAX_ORDER) {
6649 outer_start &= ~0UL << order;
6652 /* Make sure the range is really isolated. */
6653 if (test_pages_isolated(outer_start, end, false)) {
6654 pr_info("%s: [%lx, %lx) PFNs busy\n",
6655 __func__, outer_start, end);
6660 /* Grab isolated pages from freelists. */
6661 outer_end = isolate_freepages_range(&cc, outer_start, end);
6667 /* Free head and tail (if any) */
6668 if (start != outer_start)
6669 free_contig_range(outer_start, start - outer_start);
6670 if (end != outer_end)
6671 free_contig_range(end, outer_end - end);
6674 undo_isolate_page_range(pfn_max_align_down(start),
6675 pfn_max_align_up(end), migratetype);
6679 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6681 unsigned int count = 0;
6683 for (; nr_pages--; pfn++) {
6684 struct page *page = pfn_to_page(pfn);
6686 count += page_count(page) != 1;
6689 WARN(count != 0, "%d pages are still in use!\n", count);
6693 #ifdef CONFIG_MEMORY_HOTPLUG
6695 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6696 * page high values need to be recalulated.
6698 void __meminit zone_pcp_update(struct zone *zone)
6701 mutex_lock(&pcp_batch_high_lock);
6702 for_each_possible_cpu(cpu)
6703 pageset_set_high_and_batch(zone,
6704 per_cpu_ptr(zone->pageset, cpu));
6705 mutex_unlock(&pcp_batch_high_lock);
6709 void zone_pcp_reset(struct zone *zone)
6711 unsigned long flags;
6713 struct per_cpu_pageset *pset;
6715 /* avoid races with drain_pages() */
6716 local_irq_save(flags);
6717 if (zone->pageset != &boot_pageset) {
6718 for_each_online_cpu(cpu) {
6719 pset = per_cpu_ptr(zone->pageset, cpu);
6720 drain_zonestat(zone, pset);
6722 free_percpu(zone->pageset);
6723 zone->pageset = &boot_pageset;
6725 local_irq_restore(flags);
6728 #ifdef CONFIG_MEMORY_HOTREMOVE
6730 * All pages in the range must be isolated before calling this.
6733 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6737 unsigned int order, i;
6739 unsigned long flags;
6740 /* find the first valid pfn */
6741 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6746 zone = page_zone(pfn_to_page(pfn));
6747 spin_lock_irqsave(&zone->lock, flags);
6749 while (pfn < end_pfn) {
6750 if (!pfn_valid(pfn)) {
6754 page = pfn_to_page(pfn);
6756 * The HWPoisoned page may be not in buddy system, and
6757 * page_count() is not 0.
6759 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6761 SetPageReserved(page);
6765 BUG_ON(page_count(page));
6766 BUG_ON(!PageBuddy(page));
6767 order = page_order(page);
6768 #ifdef CONFIG_DEBUG_VM
6769 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6770 pfn, 1 << order, end_pfn);
6772 list_del(&page->lru);
6773 rmv_page_order(page);
6774 zone->free_area[order].nr_free--;
6775 for (i = 0; i < (1 << order); i++)
6776 SetPageReserved((page+i));
6777 pfn += (1 << order);
6779 spin_unlock_irqrestore(&zone->lock, flags);
6783 #ifdef CONFIG_MEMORY_FAILURE
6784 bool is_free_buddy_page(struct page *page)
6786 struct zone *zone = page_zone(page);
6787 unsigned long pfn = page_to_pfn(page);
6788 unsigned long flags;
6791 spin_lock_irqsave(&zone->lock, flags);
6792 for (order = 0; order < MAX_ORDER; order++) {
6793 struct page *page_head = page - (pfn & ((1 << order) - 1));
6795 if (PageBuddy(page_head) && page_order(page_head) >= order)
6798 spin_unlock_irqrestore(&zone->lock, flags);
6800 return order < MAX_ORDER;