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)
255 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
257 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
264 * Returns false when the remaining initialisation should be deferred until
265 * later in the boot cycle when it can be parallelised.
267 static inline bool update_defer_init(pg_data_t *pgdat,
268 unsigned long pfn, unsigned long zone_end,
269 unsigned long *nr_initialised)
271 /* Always populate low zones for address-contrained allocations */
272 if (zone_end < pgdat_end_pfn(pgdat))
275 /* Initialise at least 2G of the highest zone */
277 if (*nr_initialised > (2UL << (30 - PAGE_SHIFT)) &&
278 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
279 pgdat->first_deferred_pfn = pfn;
286 static inline void reset_deferred_meminit(pg_data_t *pgdat)
290 static inline bool early_page_uninitialised(unsigned long pfn)
295 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
300 static inline bool update_defer_init(pg_data_t *pgdat,
301 unsigned long pfn, unsigned long zone_end,
302 unsigned long *nr_initialised)
309 void set_pageblock_migratetype(struct page *page, int migratetype)
311 if (unlikely(page_group_by_mobility_disabled &&
312 migratetype < MIGRATE_PCPTYPES))
313 migratetype = MIGRATE_UNMOVABLE;
315 set_pageblock_flags_group(page, (unsigned long)migratetype,
316 PB_migrate, PB_migrate_end);
319 #ifdef CONFIG_DEBUG_VM
320 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
324 unsigned long pfn = page_to_pfn(page);
325 unsigned long sp, start_pfn;
328 seq = zone_span_seqbegin(zone);
329 start_pfn = zone->zone_start_pfn;
330 sp = zone->spanned_pages;
331 if (!zone_spans_pfn(zone, pfn))
333 } while (zone_span_seqretry(zone, seq));
336 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
337 pfn, zone_to_nid(zone), zone->name,
338 start_pfn, start_pfn + sp);
343 static int page_is_consistent(struct zone *zone, struct page *page)
345 if (!pfn_valid_within(page_to_pfn(page)))
347 if (zone != page_zone(page))
353 * Temporary debugging check for pages not lying within a given zone.
355 static int bad_range(struct zone *zone, struct page *page)
357 if (page_outside_zone_boundaries(zone, page))
359 if (!page_is_consistent(zone, page))
365 static inline int bad_range(struct zone *zone, struct page *page)
371 static void bad_page(struct page *page, const char *reason,
372 unsigned long bad_flags)
374 static unsigned long resume;
375 static unsigned long nr_shown;
376 static unsigned long nr_unshown;
378 /* Don't complain about poisoned pages */
379 if (PageHWPoison(page)) {
380 page_mapcount_reset(page); /* remove PageBuddy */
385 * Allow a burst of 60 reports, then keep quiet for that minute;
386 * or allow a steady drip of one report per second.
388 if (nr_shown == 60) {
389 if (time_before(jiffies, resume)) {
395 "BUG: Bad page state: %lu messages suppressed\n",
402 resume = jiffies + 60 * HZ;
404 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
405 current->comm, page_to_pfn(page));
406 dump_page_badflags(page, reason, bad_flags);
411 /* Leave bad fields for debug, except PageBuddy could make trouble */
412 page_mapcount_reset(page); /* remove PageBuddy */
413 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
417 * Higher-order pages are called "compound pages". They are structured thusly:
419 * The first PAGE_SIZE page is called the "head page".
421 * The remaining PAGE_SIZE pages are called "tail pages".
423 * All pages have PG_compound set. All tail pages have their ->first_page
424 * pointing at the head page.
426 * The first tail page's ->lru.next holds the address of the compound page's
427 * put_page() function. Its ->lru.prev holds the order of allocation.
428 * This usage means that zero-order pages may not be compound.
431 static void free_compound_page(struct page *page)
433 __free_pages_ok(page, compound_order(page));
436 void prep_compound_page(struct page *page, unsigned long order)
439 int nr_pages = 1 << order;
441 set_compound_page_dtor(page, free_compound_page);
442 set_compound_order(page, order);
444 for (i = 1; i < nr_pages; i++) {
445 struct page *p = page + i;
446 set_page_count(p, 0);
447 p->first_page = page;
448 /* Make sure p->first_page is always valid for PageTail() */
454 #ifdef CONFIG_DEBUG_PAGEALLOC
455 unsigned int _debug_guardpage_minorder;
456 bool _debug_pagealloc_enabled __read_mostly;
457 bool _debug_guardpage_enabled __read_mostly;
459 static int __init early_debug_pagealloc(char *buf)
464 if (strcmp(buf, "on") == 0)
465 _debug_pagealloc_enabled = true;
469 early_param("debug_pagealloc", early_debug_pagealloc);
471 static bool need_debug_guardpage(void)
473 /* If we don't use debug_pagealloc, we don't need guard page */
474 if (!debug_pagealloc_enabled())
480 static void init_debug_guardpage(void)
482 if (!debug_pagealloc_enabled())
485 _debug_guardpage_enabled = true;
488 struct page_ext_operations debug_guardpage_ops = {
489 .need = need_debug_guardpage,
490 .init = init_debug_guardpage,
493 static int __init debug_guardpage_minorder_setup(char *buf)
497 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
498 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
501 _debug_guardpage_minorder = res;
502 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
505 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
507 static inline void set_page_guard(struct zone *zone, struct page *page,
508 unsigned int order, int migratetype)
510 struct page_ext *page_ext;
512 if (!debug_guardpage_enabled())
515 page_ext = lookup_page_ext(page);
516 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
518 INIT_LIST_HEAD(&page->lru);
519 set_page_private(page, order);
520 /* Guard pages are not available for any usage */
521 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
524 static inline void clear_page_guard(struct zone *zone, struct page *page,
525 unsigned int order, int migratetype)
527 struct page_ext *page_ext;
529 if (!debug_guardpage_enabled())
532 page_ext = lookup_page_ext(page);
533 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
535 set_page_private(page, 0);
536 if (!is_migrate_isolate(migratetype))
537 __mod_zone_freepage_state(zone, (1 << order), migratetype);
540 struct page_ext_operations debug_guardpage_ops = { NULL, };
541 static inline void set_page_guard(struct zone *zone, struct page *page,
542 unsigned int order, int migratetype) {}
543 static inline void clear_page_guard(struct zone *zone, struct page *page,
544 unsigned int order, int migratetype) {}
547 static inline void set_page_order(struct page *page, unsigned int order)
549 set_page_private(page, order);
550 __SetPageBuddy(page);
553 static inline void rmv_page_order(struct page *page)
555 __ClearPageBuddy(page);
556 set_page_private(page, 0);
560 * This function checks whether a page is free && is the buddy
561 * we can do coalesce a page and its buddy if
562 * (a) the buddy is not in a hole &&
563 * (b) the buddy is in the buddy system &&
564 * (c) a page and its buddy have the same order &&
565 * (d) a page and its buddy are in the same zone.
567 * For recording whether a page is in the buddy system, we set ->_mapcount
568 * PAGE_BUDDY_MAPCOUNT_VALUE.
569 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
570 * serialized by zone->lock.
572 * For recording page's order, we use page_private(page).
574 static inline int page_is_buddy(struct page *page, struct page *buddy,
577 if (!pfn_valid_within(page_to_pfn(buddy)))
580 if (page_is_guard(buddy) && page_order(buddy) == order) {
581 if (page_zone_id(page) != page_zone_id(buddy))
584 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
589 if (PageBuddy(buddy) && page_order(buddy) == order) {
591 * zone check is done late to avoid uselessly
592 * calculating zone/node ids for pages that could
595 if (page_zone_id(page) != page_zone_id(buddy))
598 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
606 * Freeing function for a buddy system allocator.
608 * The concept of a buddy system is to maintain direct-mapped table
609 * (containing bit values) for memory blocks of various "orders".
610 * The bottom level table contains the map for the smallest allocatable
611 * units of memory (here, pages), and each level above it describes
612 * pairs of units from the levels below, hence, "buddies".
613 * At a high level, all that happens here is marking the table entry
614 * at the bottom level available, and propagating the changes upward
615 * as necessary, plus some accounting needed to play nicely with other
616 * parts of the VM system.
617 * At each level, we keep a list of pages, which are heads of continuous
618 * free pages of length of (1 << order) and marked with _mapcount
619 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
621 * So when we are allocating or freeing one, we can derive the state of the
622 * other. That is, if we allocate a small block, and both were
623 * free, the remainder of the region must be split into blocks.
624 * If a block is freed, and its buddy is also free, then this
625 * triggers coalescing into a block of larger size.
630 static inline void __free_one_page(struct page *page,
632 struct zone *zone, unsigned int order,
635 unsigned long page_idx;
636 unsigned long combined_idx;
637 unsigned long uninitialized_var(buddy_idx);
639 int max_order = MAX_ORDER;
641 VM_BUG_ON(!zone_is_initialized(zone));
642 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
644 VM_BUG_ON(migratetype == -1);
645 if (is_migrate_isolate(migratetype)) {
647 * We restrict max order of merging to prevent merge
648 * between freepages on isolate pageblock and normal
649 * pageblock. Without this, pageblock isolation
650 * could cause incorrect freepage accounting.
652 max_order = min(MAX_ORDER, pageblock_order + 1);
654 __mod_zone_freepage_state(zone, 1 << order, migratetype);
657 page_idx = pfn & ((1 << max_order) - 1);
659 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
660 VM_BUG_ON_PAGE(bad_range(zone, page), page);
662 while (order < max_order - 1) {
663 buddy_idx = __find_buddy_index(page_idx, order);
664 buddy = page + (buddy_idx - page_idx);
665 if (!page_is_buddy(page, buddy, order))
668 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
669 * merge with it and move up one order.
671 if (page_is_guard(buddy)) {
672 clear_page_guard(zone, buddy, order, migratetype);
674 list_del(&buddy->lru);
675 zone->free_area[order].nr_free--;
676 rmv_page_order(buddy);
678 combined_idx = buddy_idx & page_idx;
679 page = page + (combined_idx - page_idx);
680 page_idx = combined_idx;
683 set_page_order(page, order);
686 * If this is not the largest possible page, check if the buddy
687 * of the next-highest order is free. If it is, it's possible
688 * that pages are being freed that will coalesce soon. In case,
689 * that is happening, add the free page to the tail of the list
690 * so it's less likely to be used soon and more likely to be merged
691 * as a higher order page
693 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
694 struct page *higher_page, *higher_buddy;
695 combined_idx = buddy_idx & page_idx;
696 higher_page = page + (combined_idx - page_idx);
697 buddy_idx = __find_buddy_index(combined_idx, order + 1);
698 higher_buddy = higher_page + (buddy_idx - combined_idx);
699 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
700 list_add_tail(&page->lru,
701 &zone->free_area[order].free_list[migratetype]);
706 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
708 zone->free_area[order].nr_free++;
711 static inline int free_pages_check(struct page *page)
713 const char *bad_reason = NULL;
714 unsigned long bad_flags = 0;
716 if (unlikely(page_mapcount(page)))
717 bad_reason = "nonzero mapcount";
718 if (unlikely(page->mapping != NULL))
719 bad_reason = "non-NULL mapping";
720 if (unlikely(atomic_read(&page->_count) != 0))
721 bad_reason = "nonzero _count";
722 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
723 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
724 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
727 if (unlikely(page->mem_cgroup))
728 bad_reason = "page still charged to cgroup";
730 if (unlikely(bad_reason)) {
731 bad_page(page, bad_reason, bad_flags);
734 page_cpupid_reset_last(page);
735 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
736 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
741 * Frees a number of pages from the PCP lists
742 * Assumes all pages on list are in same zone, and of same order.
743 * count is the number of pages to free.
745 * If the zone was previously in an "all pages pinned" state then look to
746 * see if this freeing clears that state.
748 * And clear the zone's pages_scanned counter, to hold off the "all pages are
749 * pinned" detection logic.
751 static void free_pcppages_bulk(struct zone *zone, int count,
752 struct per_cpu_pages *pcp)
757 unsigned long nr_scanned;
759 spin_lock(&zone->lock);
760 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
762 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
766 struct list_head *list;
769 * Remove pages from lists in a round-robin fashion. A
770 * batch_free count is maintained that is incremented when an
771 * empty list is encountered. This is so more pages are freed
772 * off fuller lists instead of spinning excessively around empty
777 if (++migratetype == MIGRATE_PCPTYPES)
779 list = &pcp->lists[migratetype];
780 } while (list_empty(list));
782 /* This is the only non-empty list. Free them all. */
783 if (batch_free == MIGRATE_PCPTYPES)
784 batch_free = to_free;
787 int mt; /* migratetype of the to-be-freed page */
789 page = list_entry(list->prev, struct page, lru);
790 /* must delete as __free_one_page list manipulates */
791 list_del(&page->lru);
792 mt = get_freepage_migratetype(page);
793 if (unlikely(has_isolate_pageblock(zone)))
794 mt = get_pageblock_migratetype(page);
796 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
797 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
798 trace_mm_page_pcpu_drain(page, 0, mt);
799 } while (--to_free && --batch_free && !list_empty(list));
801 spin_unlock(&zone->lock);
804 static void free_one_page(struct zone *zone,
805 struct page *page, unsigned long pfn,
809 unsigned long nr_scanned;
810 spin_lock(&zone->lock);
811 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
813 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
815 if (unlikely(has_isolate_pageblock(zone) ||
816 is_migrate_isolate(migratetype))) {
817 migratetype = get_pfnblock_migratetype(page, pfn);
819 __free_one_page(page, pfn, zone, order, migratetype);
820 spin_unlock(&zone->lock);
823 static int free_tail_pages_check(struct page *head_page, struct page *page)
825 if (!IS_ENABLED(CONFIG_DEBUG_VM))
827 if (unlikely(!PageTail(page))) {
828 bad_page(page, "PageTail not set", 0);
831 if (unlikely(page->first_page != head_page)) {
832 bad_page(page, "first_page not consistent", 0);
838 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
839 unsigned long zone, int nid)
841 struct zone *z = &NODE_DATA(nid)->node_zones[zone];
843 set_page_links(page, zone, nid, pfn);
844 mminit_verify_page_links(page, zone, nid, pfn);
845 init_page_count(page);
846 page_mapcount_reset(page);
847 page_cpupid_reset_last(page);
850 * Mark the block movable so that blocks are reserved for
851 * movable at startup. This will force kernel allocations
852 * to reserve their blocks rather than leaking throughout
853 * the address space during boot when many long-lived
854 * kernel allocations are made. Later some blocks near
855 * the start are marked MIGRATE_RESERVE by
856 * setup_zone_migrate_reserve()
858 * bitmap is created for zone's valid pfn range. but memmap
859 * can be created for invalid pages (for alignment)
860 * check here not to call set_pageblock_migratetype() against
863 if ((z->zone_start_pfn <= pfn)
864 && (pfn < zone_end_pfn(z))
865 && !(pfn & (pageblock_nr_pages - 1)))
866 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
868 INIT_LIST_HEAD(&page->lru);
869 #ifdef WANT_PAGE_VIRTUAL
870 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
871 if (!is_highmem_idx(zone))
872 set_page_address(page, __va(pfn << PAGE_SHIFT));
876 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
879 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
882 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
883 static void init_reserved_page(unsigned long pfn)
888 if (!early_page_uninitialised(pfn))
891 nid = early_pfn_to_nid(pfn);
892 pgdat = NODE_DATA(nid);
894 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
895 struct zone *zone = &pgdat->node_zones[zid];
897 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
900 __init_single_pfn(pfn, zid, nid);
903 static inline void init_reserved_page(unsigned long pfn)
906 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
909 * Initialised pages do not have PageReserved set. This function is
910 * called for each range allocated by the bootmem allocator and
911 * marks the pages PageReserved. The remaining valid pages are later
912 * sent to the buddy page allocator.
914 void __meminit reserve_bootmem_region(unsigned long start, unsigned long end)
916 unsigned long start_pfn = PFN_DOWN(start);
917 unsigned long end_pfn = PFN_UP(end);
919 for (; start_pfn < end_pfn; start_pfn++) {
920 if (pfn_valid(start_pfn)) {
921 struct page *page = pfn_to_page(start_pfn);
923 init_reserved_page(start_pfn);
924 SetPageReserved(page);
929 static bool free_pages_prepare(struct page *page, unsigned int order)
931 bool compound = PageCompound(page);
934 VM_BUG_ON_PAGE(PageTail(page), page);
935 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
937 trace_mm_page_free(page, order);
938 kmemcheck_free_shadow(page, order);
939 kasan_free_pages(page, order);
942 page->mapping = NULL;
943 bad += free_pages_check(page);
944 for (i = 1; i < (1 << order); i++) {
946 bad += free_tail_pages_check(page, page + i);
947 bad += free_pages_check(page + i);
952 reset_page_owner(page, order);
954 if (!PageHighMem(page)) {
955 debug_check_no_locks_freed(page_address(page),
957 debug_check_no_obj_freed(page_address(page),
960 arch_free_page(page, order);
961 kernel_map_pages(page, 1 << order, 0);
966 static void __free_pages_ok(struct page *page, unsigned int order)
970 unsigned long pfn = page_to_pfn(page);
972 if (!free_pages_prepare(page, order))
975 migratetype = get_pfnblock_migratetype(page, pfn);
976 local_irq_save(flags);
977 __count_vm_events(PGFREE, 1 << order);
978 set_freepage_migratetype(page, migratetype);
979 free_one_page(page_zone(page), page, pfn, order, migratetype);
980 local_irq_restore(flags);
983 static void __defer_init __free_pages_boot_core(struct page *page,
984 unsigned long pfn, unsigned int order)
986 unsigned int nr_pages = 1 << order;
987 struct page *p = page;
991 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
993 __ClearPageReserved(p);
994 set_page_count(p, 0);
996 __ClearPageReserved(p);
997 set_page_count(p, 0);
999 page_zone(page)->managed_pages += nr_pages;
1000 set_page_refcounted(page);
1001 __free_pages(page, order);
1004 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1005 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1006 /* Only safe to use early in boot when initialisation is single-threaded */
1007 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1009 int __meminit early_pfn_to_nid(unsigned long pfn)
1013 /* The system will behave unpredictably otherwise */
1014 BUG_ON(system_state != SYSTEM_BOOTING);
1016 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1019 /* just returns 0 */
1024 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1025 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1026 struct mminit_pfnnid_cache *state)
1030 nid = __early_pfn_to_nid(pfn, state);
1031 if (nid >= 0 && nid != node)
1036 /* Only safe to use early in boot when initialisation is single-threaded */
1037 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1039 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1044 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1048 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1049 struct mminit_pfnnid_cache *state)
1056 void __defer_init __free_pages_bootmem(struct page *page, unsigned long pfn,
1059 if (early_page_uninitialised(pfn))
1061 return __free_pages_boot_core(page, pfn, order);
1064 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1065 /* Initialise remaining memory on a node */
1066 void __defermem_init deferred_init_memmap(int nid)
1068 struct mminit_pfnnid_cache nid_init_state = { };
1069 unsigned long start = jiffies;
1070 unsigned long nr_pages = 0;
1071 unsigned long walk_start, walk_end;
1074 pg_data_t *pgdat = NODE_DATA(nid);
1075 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1077 if (first_init_pfn == ULONG_MAX)
1080 /* Sanity check boundaries */
1081 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1082 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1083 pgdat->first_deferred_pfn = ULONG_MAX;
1085 /* Only the highest zone is deferred so find it */
1086 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1087 zone = pgdat->node_zones + zid;
1088 if (first_init_pfn < zone_end_pfn(zone))
1092 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1093 unsigned long pfn, end_pfn;
1094 struct page *page = NULL;
1096 end_pfn = min(walk_end, zone_end_pfn(zone));
1097 pfn = first_init_pfn;
1098 if (pfn < walk_start)
1100 if (pfn < zone->zone_start_pfn)
1101 pfn = zone->zone_start_pfn;
1103 for (; pfn < end_pfn; pfn++) {
1104 if (!pfn_valid_within(pfn))
1108 * Ensure pfn_valid is checked every
1109 * MAX_ORDER_NR_PAGES for memory holes
1111 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1112 if (!pfn_valid(pfn)) {
1118 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1123 /* Minimise pfn page lookups and scheduler checks */
1124 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1127 page = pfn_to_page(pfn);
1132 VM_BUG_ON(page_zone(page) != zone);
1136 __init_single_page(page, pfn, zid, nid);
1137 __free_pages_boot_core(page, pfn, 0);
1140 first_init_pfn = max(end_pfn, first_init_pfn);
1143 /* Sanity check that the next zone really is unpopulated */
1144 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1146 pr_info("kswapd %d initialised %lu pages in %ums\n", nid, nr_pages,
1147 jiffies_to_msecs(jiffies - start));
1149 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1152 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1153 void __init init_cma_reserved_pageblock(struct page *page)
1155 unsigned i = pageblock_nr_pages;
1156 struct page *p = page;
1159 __ClearPageReserved(p);
1160 set_page_count(p, 0);
1163 set_pageblock_migratetype(page, MIGRATE_CMA);
1165 if (pageblock_order >= MAX_ORDER) {
1166 i = pageblock_nr_pages;
1169 set_page_refcounted(p);
1170 __free_pages(p, MAX_ORDER - 1);
1171 p += MAX_ORDER_NR_PAGES;
1172 } while (i -= MAX_ORDER_NR_PAGES);
1174 set_page_refcounted(page);
1175 __free_pages(page, pageblock_order);
1178 adjust_managed_page_count(page, pageblock_nr_pages);
1183 * The order of subdivision here is critical for the IO subsystem.
1184 * Please do not alter this order without good reasons and regression
1185 * testing. Specifically, as large blocks of memory are subdivided,
1186 * the order in which smaller blocks are delivered depends on the order
1187 * they're subdivided in this function. This is the primary factor
1188 * influencing the order in which pages are delivered to the IO
1189 * subsystem according to empirical testing, and this is also justified
1190 * by considering the behavior of a buddy system containing a single
1191 * large block of memory acted on by a series of small allocations.
1192 * This behavior is a critical factor in sglist merging's success.
1196 static inline void expand(struct zone *zone, struct page *page,
1197 int low, int high, struct free_area *area,
1200 unsigned long size = 1 << high;
1202 while (high > low) {
1206 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1208 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1209 debug_guardpage_enabled() &&
1210 high < debug_guardpage_minorder()) {
1212 * Mark as guard pages (or page), that will allow to
1213 * merge back to allocator when buddy will be freed.
1214 * Corresponding page table entries will not be touched,
1215 * pages will stay not present in virtual address space
1217 set_page_guard(zone, &page[size], high, migratetype);
1220 list_add(&page[size].lru, &area->free_list[migratetype]);
1222 set_page_order(&page[size], high);
1227 * This page is about to be returned from the page allocator
1229 static inline int check_new_page(struct page *page)
1231 const char *bad_reason = NULL;
1232 unsigned long bad_flags = 0;
1234 if (unlikely(page_mapcount(page)))
1235 bad_reason = "nonzero mapcount";
1236 if (unlikely(page->mapping != NULL))
1237 bad_reason = "non-NULL mapping";
1238 if (unlikely(atomic_read(&page->_count) != 0))
1239 bad_reason = "nonzero _count";
1240 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1241 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1242 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1245 if (unlikely(page->mem_cgroup))
1246 bad_reason = "page still charged to cgroup";
1248 if (unlikely(bad_reason)) {
1249 bad_page(page, bad_reason, bad_flags);
1255 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1260 for (i = 0; i < (1 << order); i++) {
1261 struct page *p = page + i;
1262 if (unlikely(check_new_page(p)))
1266 set_page_private(page, 0);
1267 set_page_refcounted(page);
1269 arch_alloc_page(page, order);
1270 kernel_map_pages(page, 1 << order, 1);
1271 kasan_alloc_pages(page, order);
1273 if (gfp_flags & __GFP_ZERO)
1274 for (i = 0; i < (1 << order); i++)
1275 clear_highpage(page + i);
1277 if (order && (gfp_flags & __GFP_COMP))
1278 prep_compound_page(page, order);
1280 set_page_owner(page, order, gfp_flags);
1283 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was necessary to
1284 * allocate the page. The expectation is that the caller is taking
1285 * steps that will free more memory. The caller should avoid the page
1286 * being used for !PFMEMALLOC purposes.
1288 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
1294 * Go through the free lists for the given migratetype and remove
1295 * the smallest available page from the freelists
1298 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1301 unsigned int current_order;
1302 struct free_area *area;
1305 /* Find a page of the appropriate size in the preferred list */
1306 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1307 area = &(zone->free_area[current_order]);
1308 if (list_empty(&area->free_list[migratetype]))
1311 page = list_entry(area->free_list[migratetype].next,
1313 list_del(&page->lru);
1314 rmv_page_order(page);
1316 expand(zone, page, order, current_order, area, migratetype);
1317 set_freepage_migratetype(page, migratetype);
1326 * This array describes the order lists are fallen back to when
1327 * the free lists for the desirable migrate type are depleted
1329 static int fallbacks[MIGRATE_TYPES][4] = {
1330 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1331 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1332 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
1334 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
1336 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
1337 #ifdef CONFIG_MEMORY_ISOLATION
1338 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
1343 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1346 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1349 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1350 unsigned int order) { return NULL; }
1354 * Move the free pages in a range to the free lists of the requested type.
1355 * Note that start_page and end_pages are not aligned on a pageblock
1356 * boundary. If alignment is required, use move_freepages_block()
1358 int move_freepages(struct zone *zone,
1359 struct page *start_page, struct page *end_page,
1363 unsigned long order;
1364 int pages_moved = 0;
1366 #ifndef CONFIG_HOLES_IN_ZONE
1368 * page_zone is not safe to call in this context when
1369 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1370 * anyway as we check zone boundaries in move_freepages_block().
1371 * Remove at a later date when no bug reports exist related to
1372 * grouping pages by mobility
1374 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1377 for (page = start_page; page <= end_page;) {
1378 /* Make sure we are not inadvertently changing nodes */
1379 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1381 if (!pfn_valid_within(page_to_pfn(page))) {
1386 if (!PageBuddy(page)) {
1391 order = page_order(page);
1392 list_move(&page->lru,
1393 &zone->free_area[order].free_list[migratetype]);
1394 set_freepage_migratetype(page, migratetype);
1396 pages_moved += 1 << order;
1402 int move_freepages_block(struct zone *zone, struct page *page,
1405 unsigned long start_pfn, end_pfn;
1406 struct page *start_page, *end_page;
1408 start_pfn = page_to_pfn(page);
1409 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1410 start_page = pfn_to_page(start_pfn);
1411 end_page = start_page + pageblock_nr_pages - 1;
1412 end_pfn = start_pfn + pageblock_nr_pages - 1;
1414 /* Do not cross zone boundaries */
1415 if (!zone_spans_pfn(zone, start_pfn))
1417 if (!zone_spans_pfn(zone, end_pfn))
1420 return move_freepages(zone, start_page, end_page, migratetype);
1423 static void change_pageblock_range(struct page *pageblock_page,
1424 int start_order, int migratetype)
1426 int nr_pageblocks = 1 << (start_order - pageblock_order);
1428 while (nr_pageblocks--) {
1429 set_pageblock_migratetype(pageblock_page, migratetype);
1430 pageblock_page += pageblock_nr_pages;
1435 * When we are falling back to another migratetype during allocation, try to
1436 * steal extra free pages from the same pageblocks to satisfy further
1437 * allocations, instead of polluting multiple pageblocks.
1439 * If we are stealing a relatively large buddy page, it is likely there will
1440 * be more free pages in the pageblock, so try to steal them all. For
1441 * reclaimable and unmovable allocations, we steal regardless of page size,
1442 * as fragmentation caused by those allocations polluting movable pageblocks
1443 * is worse than movable allocations stealing from unmovable and reclaimable
1446 static bool can_steal_fallback(unsigned int order, int start_mt)
1449 * Leaving this order check is intended, although there is
1450 * relaxed order check in next check. The reason is that
1451 * we can actually steal whole pageblock if this condition met,
1452 * but, below check doesn't guarantee it and that is just heuristic
1453 * so could be changed anytime.
1455 if (order >= pageblock_order)
1458 if (order >= pageblock_order / 2 ||
1459 start_mt == MIGRATE_RECLAIMABLE ||
1460 start_mt == MIGRATE_UNMOVABLE ||
1461 page_group_by_mobility_disabled)
1468 * This function implements actual steal behaviour. If order is large enough,
1469 * we can steal whole pageblock. If not, we first move freepages in this
1470 * pageblock and check whether half of pages are moved or not. If half of
1471 * pages are moved, we can change migratetype of pageblock and permanently
1472 * use it's pages as requested migratetype in the future.
1474 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1477 int current_order = page_order(page);
1480 /* Take ownership for orders >= pageblock_order */
1481 if (current_order >= pageblock_order) {
1482 change_pageblock_range(page, current_order, start_type);
1486 pages = move_freepages_block(zone, page, start_type);
1488 /* Claim the whole block if over half of it is free */
1489 if (pages >= (1 << (pageblock_order-1)) ||
1490 page_group_by_mobility_disabled)
1491 set_pageblock_migratetype(page, start_type);
1495 * Check whether there is a suitable fallback freepage with requested order.
1496 * If only_stealable is true, this function returns fallback_mt only if
1497 * we can steal other freepages all together. This would help to reduce
1498 * fragmentation due to mixed migratetype pages in one pageblock.
1500 int find_suitable_fallback(struct free_area *area, unsigned int order,
1501 int migratetype, bool only_stealable, bool *can_steal)
1506 if (area->nr_free == 0)
1511 fallback_mt = fallbacks[migratetype][i];
1512 if (fallback_mt == MIGRATE_RESERVE)
1515 if (list_empty(&area->free_list[fallback_mt]))
1518 if (can_steal_fallback(order, migratetype))
1521 if (!only_stealable)
1531 /* Remove an element from the buddy allocator from the fallback list */
1532 static inline struct page *
1533 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1535 struct free_area *area;
1536 unsigned int current_order;
1541 /* Find the largest possible block of pages in the other list */
1542 for (current_order = MAX_ORDER-1;
1543 current_order >= order && current_order <= MAX_ORDER-1;
1545 area = &(zone->free_area[current_order]);
1546 fallback_mt = find_suitable_fallback(area, current_order,
1547 start_migratetype, false, &can_steal);
1548 if (fallback_mt == -1)
1551 page = list_entry(area->free_list[fallback_mt].next,
1554 steal_suitable_fallback(zone, page, start_migratetype);
1556 /* Remove the page from the freelists */
1558 list_del(&page->lru);
1559 rmv_page_order(page);
1561 expand(zone, page, order, current_order, area,
1564 * The freepage_migratetype may differ from pageblock's
1565 * migratetype depending on the decisions in
1566 * try_to_steal_freepages(). This is OK as long as it
1567 * does not differ for MIGRATE_CMA pageblocks. For CMA
1568 * we need to make sure unallocated pages flushed from
1569 * pcp lists are returned to the correct freelist.
1571 set_freepage_migratetype(page, start_migratetype);
1573 trace_mm_page_alloc_extfrag(page, order, current_order,
1574 start_migratetype, fallback_mt);
1583 * Do the hard work of removing an element from the buddy allocator.
1584 * Call me with the zone->lock already held.
1586 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1592 page = __rmqueue_smallest(zone, order, migratetype);
1594 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1595 if (migratetype == MIGRATE_MOVABLE)
1596 page = __rmqueue_cma_fallback(zone, order);
1599 page = __rmqueue_fallback(zone, order, migratetype);
1602 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1603 * is used because __rmqueue_smallest is an inline function
1604 * and we want just one call site
1607 migratetype = MIGRATE_RESERVE;
1612 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1617 * Obtain a specified number of elements from the buddy allocator, all under
1618 * a single hold of the lock, for efficiency. Add them to the supplied list.
1619 * Returns the number of new pages which were placed at *list.
1621 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1622 unsigned long count, struct list_head *list,
1623 int migratetype, bool cold)
1627 spin_lock(&zone->lock);
1628 for (i = 0; i < count; ++i) {
1629 struct page *page = __rmqueue(zone, order, migratetype);
1630 if (unlikely(page == NULL))
1634 * Split buddy pages returned by expand() are received here
1635 * in physical page order. The page is added to the callers and
1636 * list and the list head then moves forward. From the callers
1637 * perspective, the linked list is ordered by page number in
1638 * some conditions. This is useful for IO devices that can
1639 * merge IO requests if the physical pages are ordered
1643 list_add(&page->lru, list);
1645 list_add_tail(&page->lru, list);
1647 if (is_migrate_cma(get_freepage_migratetype(page)))
1648 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1651 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1652 spin_unlock(&zone->lock);
1658 * Called from the vmstat counter updater to drain pagesets of this
1659 * currently executing processor on remote nodes after they have
1662 * Note that this function must be called with the thread pinned to
1663 * a single processor.
1665 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1667 unsigned long flags;
1668 int to_drain, batch;
1670 local_irq_save(flags);
1671 batch = READ_ONCE(pcp->batch);
1672 to_drain = min(pcp->count, batch);
1674 free_pcppages_bulk(zone, to_drain, pcp);
1675 pcp->count -= to_drain;
1677 local_irq_restore(flags);
1682 * Drain pcplists of the indicated processor and zone.
1684 * The processor must either be the current processor and the
1685 * thread pinned to the current processor or a processor that
1688 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1690 unsigned long flags;
1691 struct per_cpu_pageset *pset;
1692 struct per_cpu_pages *pcp;
1694 local_irq_save(flags);
1695 pset = per_cpu_ptr(zone->pageset, cpu);
1699 free_pcppages_bulk(zone, pcp->count, pcp);
1702 local_irq_restore(flags);
1706 * Drain pcplists of all zones on the indicated processor.
1708 * The processor must either be the current processor and the
1709 * thread pinned to the current processor or a processor that
1712 static void drain_pages(unsigned int cpu)
1716 for_each_populated_zone(zone) {
1717 drain_pages_zone(cpu, zone);
1722 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1724 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1725 * the single zone's pages.
1727 void drain_local_pages(struct zone *zone)
1729 int cpu = smp_processor_id();
1732 drain_pages_zone(cpu, zone);
1738 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1740 * When zone parameter is non-NULL, spill just the single zone's pages.
1742 * Note that this code is protected against sending an IPI to an offline
1743 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1744 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1745 * nothing keeps CPUs from showing up after we populated the cpumask and
1746 * before the call to on_each_cpu_mask().
1748 void drain_all_pages(struct zone *zone)
1753 * Allocate in the BSS so we wont require allocation in
1754 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1756 static cpumask_t cpus_with_pcps;
1759 * We don't care about racing with CPU hotplug event
1760 * as offline notification will cause the notified
1761 * cpu to drain that CPU pcps and on_each_cpu_mask
1762 * disables preemption as part of its processing
1764 for_each_online_cpu(cpu) {
1765 struct per_cpu_pageset *pcp;
1767 bool has_pcps = false;
1770 pcp = per_cpu_ptr(zone->pageset, cpu);
1774 for_each_populated_zone(z) {
1775 pcp = per_cpu_ptr(z->pageset, cpu);
1776 if (pcp->pcp.count) {
1784 cpumask_set_cpu(cpu, &cpus_with_pcps);
1786 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1788 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
1792 #ifdef CONFIG_HIBERNATION
1794 void mark_free_pages(struct zone *zone)
1796 unsigned long pfn, max_zone_pfn;
1797 unsigned long flags;
1798 unsigned int order, t;
1799 struct list_head *curr;
1801 if (zone_is_empty(zone))
1804 spin_lock_irqsave(&zone->lock, flags);
1806 max_zone_pfn = zone_end_pfn(zone);
1807 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1808 if (pfn_valid(pfn)) {
1809 struct page *page = pfn_to_page(pfn);
1811 if (!swsusp_page_is_forbidden(page))
1812 swsusp_unset_page_free(page);
1815 for_each_migratetype_order(order, t) {
1816 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1819 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1820 for (i = 0; i < (1UL << order); i++)
1821 swsusp_set_page_free(pfn_to_page(pfn + i));
1824 spin_unlock_irqrestore(&zone->lock, flags);
1826 #endif /* CONFIG_PM */
1829 * Free a 0-order page
1830 * cold == true ? free a cold page : free a hot page
1832 void free_hot_cold_page(struct page *page, bool cold)
1834 struct zone *zone = page_zone(page);
1835 struct per_cpu_pages *pcp;
1836 unsigned long flags;
1837 unsigned long pfn = page_to_pfn(page);
1840 if (!free_pages_prepare(page, 0))
1843 migratetype = get_pfnblock_migratetype(page, pfn);
1844 set_freepage_migratetype(page, migratetype);
1845 local_irq_save(flags);
1846 __count_vm_event(PGFREE);
1849 * We only track unmovable, reclaimable and movable on pcp lists.
1850 * Free ISOLATE pages back to the allocator because they are being
1851 * offlined but treat RESERVE as movable pages so we can get those
1852 * areas back if necessary. Otherwise, we may have to free
1853 * excessively into the page allocator
1855 if (migratetype >= MIGRATE_PCPTYPES) {
1856 if (unlikely(is_migrate_isolate(migratetype))) {
1857 free_one_page(zone, page, pfn, 0, migratetype);
1860 migratetype = MIGRATE_MOVABLE;
1863 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1865 list_add(&page->lru, &pcp->lists[migratetype]);
1867 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1869 if (pcp->count >= pcp->high) {
1870 unsigned long batch = READ_ONCE(pcp->batch);
1871 free_pcppages_bulk(zone, batch, pcp);
1872 pcp->count -= batch;
1876 local_irq_restore(flags);
1880 * Free a list of 0-order pages
1882 void free_hot_cold_page_list(struct list_head *list, bool cold)
1884 struct page *page, *next;
1886 list_for_each_entry_safe(page, next, list, lru) {
1887 trace_mm_page_free_batched(page, cold);
1888 free_hot_cold_page(page, cold);
1893 * split_page takes a non-compound higher-order page, and splits it into
1894 * n (1<<order) sub-pages: page[0..n]
1895 * Each sub-page must be freed individually.
1897 * Note: this is probably too low level an operation for use in drivers.
1898 * Please consult with lkml before using this in your driver.
1900 void split_page(struct page *page, unsigned int order)
1904 VM_BUG_ON_PAGE(PageCompound(page), page);
1905 VM_BUG_ON_PAGE(!page_count(page), page);
1907 #ifdef CONFIG_KMEMCHECK
1909 * Split shadow pages too, because free(page[0]) would
1910 * otherwise free the whole shadow.
1912 if (kmemcheck_page_is_tracked(page))
1913 split_page(virt_to_page(page[0].shadow), order);
1916 set_page_owner(page, 0, 0);
1917 for (i = 1; i < (1 << order); i++) {
1918 set_page_refcounted(page + i);
1919 set_page_owner(page + i, 0, 0);
1922 EXPORT_SYMBOL_GPL(split_page);
1924 int __isolate_free_page(struct page *page, unsigned int order)
1926 unsigned long watermark;
1930 BUG_ON(!PageBuddy(page));
1932 zone = page_zone(page);
1933 mt = get_pageblock_migratetype(page);
1935 if (!is_migrate_isolate(mt)) {
1936 /* Obey watermarks as if the page was being allocated */
1937 watermark = low_wmark_pages(zone) + (1 << order);
1938 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1941 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1944 /* Remove page from free list */
1945 list_del(&page->lru);
1946 zone->free_area[order].nr_free--;
1947 rmv_page_order(page);
1949 /* Set the pageblock if the isolated page is at least a pageblock */
1950 if (order >= pageblock_order - 1) {
1951 struct page *endpage = page + (1 << order) - 1;
1952 for (; page < endpage; page += pageblock_nr_pages) {
1953 int mt = get_pageblock_migratetype(page);
1954 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1955 set_pageblock_migratetype(page,
1960 set_page_owner(page, order, 0);
1961 return 1UL << order;
1965 * Similar to split_page except the page is already free. As this is only
1966 * being used for migration, the migratetype of the block also changes.
1967 * As this is called with interrupts disabled, the caller is responsible
1968 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1971 * Note: this is probably too low level an operation for use in drivers.
1972 * Please consult with lkml before using this in your driver.
1974 int split_free_page(struct page *page)
1979 order = page_order(page);
1981 nr_pages = __isolate_free_page(page, order);
1985 /* Split into individual pages */
1986 set_page_refcounted(page);
1987 split_page(page, order);
1992 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
1995 struct page *buffered_rmqueue(struct zone *preferred_zone,
1996 struct zone *zone, unsigned int order,
1997 gfp_t gfp_flags, int migratetype)
1999 unsigned long flags;
2001 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2003 if (likely(order == 0)) {
2004 struct per_cpu_pages *pcp;
2005 struct list_head *list;
2007 local_irq_save(flags);
2008 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2009 list = &pcp->lists[migratetype];
2010 if (list_empty(list)) {
2011 pcp->count += rmqueue_bulk(zone, 0,
2014 if (unlikely(list_empty(list)))
2019 page = list_entry(list->prev, struct page, lru);
2021 page = list_entry(list->next, struct page, lru);
2023 list_del(&page->lru);
2026 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2028 * __GFP_NOFAIL is not to be used in new code.
2030 * All __GFP_NOFAIL callers should be fixed so that they
2031 * properly detect and handle allocation failures.
2033 * We most definitely don't want callers attempting to
2034 * allocate greater than order-1 page units with
2037 WARN_ON_ONCE(order > 1);
2039 spin_lock_irqsave(&zone->lock, flags);
2040 page = __rmqueue(zone, order, migratetype);
2041 spin_unlock(&zone->lock);
2044 __mod_zone_freepage_state(zone, -(1 << order),
2045 get_freepage_migratetype(page));
2048 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2049 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2050 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2051 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2053 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2054 zone_statistics(preferred_zone, zone, gfp_flags);
2055 local_irq_restore(flags);
2057 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2061 local_irq_restore(flags);
2065 #ifdef CONFIG_FAIL_PAGE_ALLOC
2068 struct fault_attr attr;
2070 u32 ignore_gfp_highmem;
2071 u32 ignore_gfp_wait;
2073 } fail_page_alloc = {
2074 .attr = FAULT_ATTR_INITIALIZER,
2075 .ignore_gfp_wait = 1,
2076 .ignore_gfp_highmem = 1,
2080 static int __init setup_fail_page_alloc(char *str)
2082 return setup_fault_attr(&fail_page_alloc.attr, str);
2084 __setup("fail_page_alloc=", setup_fail_page_alloc);
2086 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2088 if (order < fail_page_alloc.min_order)
2090 if (gfp_mask & __GFP_NOFAIL)
2092 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2094 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
2097 return should_fail(&fail_page_alloc.attr, 1 << order);
2100 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2102 static int __init fail_page_alloc_debugfs(void)
2104 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2107 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2108 &fail_page_alloc.attr);
2110 return PTR_ERR(dir);
2112 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2113 &fail_page_alloc.ignore_gfp_wait))
2115 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2116 &fail_page_alloc.ignore_gfp_highmem))
2118 if (!debugfs_create_u32("min-order", mode, dir,
2119 &fail_page_alloc.min_order))
2124 debugfs_remove_recursive(dir);
2129 late_initcall(fail_page_alloc_debugfs);
2131 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2133 #else /* CONFIG_FAIL_PAGE_ALLOC */
2135 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2140 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2143 * Return true if free pages are above 'mark'. This takes into account the order
2144 * of the allocation.
2146 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2147 unsigned long mark, int classzone_idx, int alloc_flags,
2150 /* free_pages may go negative - that's OK */
2155 free_pages -= (1 << order) - 1;
2156 if (alloc_flags & ALLOC_HIGH)
2158 if (alloc_flags & ALLOC_HARDER)
2161 /* If allocation can't use CMA areas don't use free CMA pages */
2162 if (!(alloc_flags & ALLOC_CMA))
2163 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
2166 if (free_pages - free_cma <= min + z->lowmem_reserve[classzone_idx])
2168 for (o = 0; o < order; o++) {
2169 /* At the next order, this order's pages become unavailable */
2170 free_pages -= z->free_area[o].nr_free << o;
2172 /* Require fewer higher order pages to be free */
2175 if (free_pages <= min)
2181 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2182 int classzone_idx, int alloc_flags)
2184 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2185 zone_page_state(z, NR_FREE_PAGES));
2188 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2189 unsigned long mark, int classzone_idx, int alloc_flags)
2191 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2193 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2194 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2196 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2202 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
2203 * skip over zones that are not allowed by the cpuset, or that have
2204 * been recently (in last second) found to be nearly full. See further
2205 * comments in mmzone.h. Reduces cache footprint of zonelist scans
2206 * that have to skip over a lot of full or unallowed zones.
2208 * If the zonelist cache is present in the passed zonelist, then
2209 * returns a pointer to the allowed node mask (either the current
2210 * tasks mems_allowed, or node_states[N_MEMORY].)
2212 * If the zonelist cache is not available for this zonelist, does
2213 * nothing and returns NULL.
2215 * If the fullzones BITMAP in the zonelist cache is stale (more than
2216 * a second since last zap'd) then we zap it out (clear its bits.)
2218 * We hold off even calling zlc_setup, until after we've checked the
2219 * first zone in the zonelist, on the theory that most allocations will
2220 * be satisfied from that first zone, so best to examine that zone as
2221 * quickly as we can.
2223 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
2225 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2226 nodemask_t *allowednodes; /* zonelist_cache approximation */
2228 zlc = zonelist->zlcache_ptr;
2232 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
2233 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2234 zlc->last_full_zap = jiffies;
2237 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
2238 &cpuset_current_mems_allowed :
2239 &node_states[N_MEMORY];
2240 return allowednodes;
2244 * Given 'z' scanning a zonelist, run a couple of quick checks to see
2245 * if it is worth looking at further for free memory:
2246 * 1) Check that the zone isn't thought to be full (doesn't have its
2247 * bit set in the zonelist_cache fullzones BITMAP).
2248 * 2) Check that the zones node (obtained from the zonelist_cache
2249 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
2250 * Return true (non-zero) if zone is worth looking at further, or
2251 * else return false (zero) if it is not.
2253 * This check -ignores- the distinction between various watermarks,
2254 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
2255 * found to be full for any variation of these watermarks, it will
2256 * be considered full for up to one second by all requests, unless
2257 * we are so low on memory on all allowed nodes that we are forced
2258 * into the second scan of the zonelist.
2260 * In the second scan we ignore this zonelist cache and exactly
2261 * apply the watermarks to all zones, even it is slower to do so.
2262 * We are low on memory in the second scan, and should leave no stone
2263 * unturned looking for a free page.
2265 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
2266 nodemask_t *allowednodes)
2268 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2269 int i; /* index of *z in zonelist zones */
2270 int n; /* node that zone *z is on */
2272 zlc = zonelist->zlcache_ptr;
2276 i = z - zonelist->_zonerefs;
2279 /* This zone is worth trying if it is allowed but not full */
2280 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
2284 * Given 'z' scanning a zonelist, set the corresponding bit in
2285 * zlc->fullzones, so that subsequent attempts to allocate a page
2286 * from that zone don't waste time re-examining it.
2288 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
2290 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2291 int i; /* index of *z in zonelist zones */
2293 zlc = zonelist->zlcache_ptr;
2297 i = z - zonelist->_zonerefs;
2299 set_bit(i, zlc->fullzones);
2303 * clear all zones full, called after direct reclaim makes progress so that
2304 * a zone that was recently full is not skipped over for up to a second
2306 static void zlc_clear_zones_full(struct zonelist *zonelist)
2308 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2310 zlc = zonelist->zlcache_ptr;
2314 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2317 static bool zone_local(struct zone *local_zone, struct zone *zone)
2319 return local_zone->node == zone->node;
2322 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2324 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2328 #else /* CONFIG_NUMA */
2330 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
2335 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
2336 nodemask_t *allowednodes)
2341 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
2345 static void zlc_clear_zones_full(struct zonelist *zonelist)
2349 static bool zone_local(struct zone *local_zone, struct zone *zone)
2354 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2359 #endif /* CONFIG_NUMA */
2361 static void reset_alloc_batches(struct zone *preferred_zone)
2363 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2366 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2367 high_wmark_pages(zone) - low_wmark_pages(zone) -
2368 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2369 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2370 } while (zone++ != preferred_zone);
2374 * get_page_from_freelist goes through the zonelist trying to allocate
2377 static struct page *
2378 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2379 const struct alloc_context *ac)
2381 struct zonelist *zonelist = ac->zonelist;
2383 struct page *page = NULL;
2385 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
2386 int zlc_active = 0; /* set if using zonelist_cache */
2387 int did_zlc_setup = 0; /* just call zlc_setup() one time */
2388 bool consider_zone_dirty = (alloc_flags & ALLOC_WMARK_LOW) &&
2389 (gfp_mask & __GFP_WRITE);
2390 int nr_fair_skipped = 0;
2391 bool zonelist_rescan;
2394 zonelist_rescan = false;
2397 * Scan zonelist, looking for a zone with enough free.
2398 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2400 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2404 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2405 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2407 if (cpusets_enabled() &&
2408 (alloc_flags & ALLOC_CPUSET) &&
2409 !cpuset_zone_allowed(zone, gfp_mask))
2412 * Distribute pages in proportion to the individual
2413 * zone size to ensure fair page aging. The zone a
2414 * page was allocated in should have no effect on the
2415 * time the page has in memory before being reclaimed.
2417 if (alloc_flags & ALLOC_FAIR) {
2418 if (!zone_local(ac->preferred_zone, zone))
2420 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2426 * When allocating a page cache page for writing, we
2427 * want to get it from a zone that is within its dirty
2428 * limit, such that no single zone holds more than its
2429 * proportional share of globally allowed dirty pages.
2430 * The dirty limits take into account the zone's
2431 * lowmem reserves and high watermark so that kswapd
2432 * should be able to balance it without having to
2433 * write pages from its LRU list.
2435 * This may look like it could increase pressure on
2436 * lower zones by failing allocations in higher zones
2437 * before they are full. But the pages that do spill
2438 * over are limited as the lower zones are protected
2439 * by this very same mechanism. It should not become
2440 * a practical burden to them.
2442 * XXX: For now, allow allocations to potentially
2443 * exceed the per-zone dirty limit in the slowpath
2444 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2445 * which is important when on a NUMA setup the allowed
2446 * zones are together not big enough to reach the
2447 * global limit. The proper fix for these situations
2448 * will require awareness of zones in the
2449 * dirty-throttling and the flusher threads.
2451 if (consider_zone_dirty && !zone_dirty_ok(zone))
2454 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2455 if (!zone_watermark_ok(zone, order, mark,
2456 ac->classzone_idx, alloc_flags)) {
2459 /* Checked here to keep the fast path fast */
2460 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2461 if (alloc_flags & ALLOC_NO_WATERMARKS)
2464 if (IS_ENABLED(CONFIG_NUMA) &&
2465 !did_zlc_setup && nr_online_nodes > 1) {
2467 * we do zlc_setup if there are multiple nodes
2468 * and before considering the first zone allowed
2471 allowednodes = zlc_setup(zonelist, alloc_flags);
2476 if (zone_reclaim_mode == 0 ||
2477 !zone_allows_reclaim(ac->preferred_zone, zone))
2478 goto this_zone_full;
2481 * As we may have just activated ZLC, check if the first
2482 * eligible zone has failed zone_reclaim recently.
2484 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2485 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2488 ret = zone_reclaim(zone, gfp_mask, order);
2490 case ZONE_RECLAIM_NOSCAN:
2493 case ZONE_RECLAIM_FULL:
2494 /* scanned but unreclaimable */
2497 /* did we reclaim enough */
2498 if (zone_watermark_ok(zone, order, mark,
2499 ac->classzone_idx, alloc_flags))
2503 * Failed to reclaim enough to meet watermark.
2504 * Only mark the zone full if checking the min
2505 * watermark or if we failed to reclaim just
2506 * 1<<order pages or else the page allocator
2507 * fastpath will prematurely mark zones full
2508 * when the watermark is between the low and
2511 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2512 ret == ZONE_RECLAIM_SOME)
2513 goto this_zone_full;
2520 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2521 gfp_mask, ac->migratetype);
2523 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2528 if (IS_ENABLED(CONFIG_NUMA) && zlc_active)
2529 zlc_mark_zone_full(zonelist, z);
2533 * The first pass makes sure allocations are spread fairly within the
2534 * local node. However, the local node might have free pages left
2535 * after the fairness batches are exhausted, and remote zones haven't
2536 * even been considered yet. Try once more without fairness, and
2537 * include remote zones now, before entering the slowpath and waking
2538 * kswapd: prefer spilling to a remote zone over swapping locally.
2540 if (alloc_flags & ALLOC_FAIR) {
2541 alloc_flags &= ~ALLOC_FAIR;
2542 if (nr_fair_skipped) {
2543 zonelist_rescan = true;
2544 reset_alloc_batches(ac->preferred_zone);
2546 if (nr_online_nodes > 1)
2547 zonelist_rescan = true;
2550 if (unlikely(IS_ENABLED(CONFIG_NUMA) && zlc_active)) {
2551 /* Disable zlc cache for second zonelist scan */
2553 zonelist_rescan = true;
2556 if (zonelist_rescan)
2563 * Large machines with many possible nodes should not always dump per-node
2564 * meminfo in irq context.
2566 static inline bool should_suppress_show_mem(void)
2571 ret = in_interrupt();
2576 static DEFINE_RATELIMIT_STATE(nopage_rs,
2577 DEFAULT_RATELIMIT_INTERVAL,
2578 DEFAULT_RATELIMIT_BURST);
2580 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2582 unsigned int filter = SHOW_MEM_FILTER_NODES;
2584 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2585 debug_guardpage_minorder() > 0)
2589 * This documents exceptions given to allocations in certain
2590 * contexts that are allowed to allocate outside current's set
2593 if (!(gfp_mask & __GFP_NOMEMALLOC))
2594 if (test_thread_flag(TIF_MEMDIE) ||
2595 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2596 filter &= ~SHOW_MEM_FILTER_NODES;
2597 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2598 filter &= ~SHOW_MEM_FILTER_NODES;
2601 struct va_format vaf;
2604 va_start(args, fmt);
2609 pr_warn("%pV", &vaf);
2614 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2615 current->comm, order, gfp_mask);
2618 if (!should_suppress_show_mem())
2622 static inline struct page *
2623 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2624 const struct alloc_context *ac, unsigned long *did_some_progress)
2628 *did_some_progress = 0;
2631 * Acquire the oom lock. If that fails, somebody else is
2632 * making progress for us.
2634 if (!mutex_trylock(&oom_lock)) {
2635 *did_some_progress = 1;
2636 schedule_timeout_uninterruptible(1);
2641 * Go through the zonelist yet one more time, keep very high watermark
2642 * here, this is only to catch a parallel oom killing, we must fail if
2643 * we're still under heavy pressure.
2645 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2646 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2650 if (!(gfp_mask & __GFP_NOFAIL)) {
2651 /* Coredumps can quickly deplete all memory reserves */
2652 if (current->flags & PF_DUMPCORE)
2654 /* The OOM killer will not help higher order allocs */
2655 if (order > PAGE_ALLOC_COSTLY_ORDER)
2657 /* The OOM killer does not needlessly kill tasks for lowmem */
2658 if (ac->high_zoneidx < ZONE_NORMAL)
2660 /* The OOM killer does not compensate for IO-less reclaim */
2661 if (!(gfp_mask & __GFP_FS)) {
2663 * XXX: Page reclaim didn't yield anything,
2664 * and the OOM killer can't be invoked, but
2665 * keep looping as per tradition.
2667 *did_some_progress = 1;
2670 if (pm_suspended_storage())
2672 /* The OOM killer may not free memory on a specific node */
2673 if (gfp_mask & __GFP_THISNODE)
2676 /* Exhausted what can be done so it's blamo time */
2677 if (out_of_memory(ac->zonelist, gfp_mask, order, ac->nodemask, false)
2678 || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL))
2679 *did_some_progress = 1;
2681 mutex_unlock(&oom_lock);
2685 #ifdef CONFIG_COMPACTION
2686 /* Try memory compaction for high-order allocations before reclaim */
2687 static struct page *
2688 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2689 int alloc_flags, const struct alloc_context *ac,
2690 enum migrate_mode mode, int *contended_compaction,
2691 bool *deferred_compaction)
2693 unsigned long compact_result;
2699 current->flags |= PF_MEMALLOC;
2700 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2701 mode, contended_compaction);
2702 current->flags &= ~PF_MEMALLOC;
2704 switch (compact_result) {
2705 case COMPACT_DEFERRED:
2706 *deferred_compaction = true;
2708 case COMPACT_SKIPPED:
2715 * At least in one zone compaction wasn't deferred or skipped, so let's
2716 * count a compaction stall
2718 count_vm_event(COMPACTSTALL);
2720 page = get_page_from_freelist(gfp_mask, order,
2721 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2724 struct zone *zone = page_zone(page);
2726 zone->compact_blockskip_flush = false;
2727 compaction_defer_reset(zone, order, true);
2728 count_vm_event(COMPACTSUCCESS);
2733 * It's bad if compaction run occurs and fails. The most likely reason
2734 * is that pages exist, but not enough to satisfy watermarks.
2736 count_vm_event(COMPACTFAIL);
2743 static inline struct page *
2744 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2745 int alloc_flags, const struct alloc_context *ac,
2746 enum migrate_mode mode, int *contended_compaction,
2747 bool *deferred_compaction)
2751 #endif /* CONFIG_COMPACTION */
2753 /* Perform direct synchronous page reclaim */
2755 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2756 const struct alloc_context *ac)
2758 struct reclaim_state reclaim_state;
2763 /* We now go into synchronous reclaim */
2764 cpuset_memory_pressure_bump();
2765 current->flags |= PF_MEMALLOC;
2766 lockdep_set_current_reclaim_state(gfp_mask);
2767 reclaim_state.reclaimed_slab = 0;
2768 current->reclaim_state = &reclaim_state;
2770 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2773 current->reclaim_state = NULL;
2774 lockdep_clear_current_reclaim_state();
2775 current->flags &= ~PF_MEMALLOC;
2782 /* The really slow allocator path where we enter direct reclaim */
2783 static inline struct page *
2784 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2785 int alloc_flags, const struct alloc_context *ac,
2786 unsigned long *did_some_progress)
2788 struct page *page = NULL;
2789 bool drained = false;
2791 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2792 if (unlikely(!(*did_some_progress)))
2795 /* After successful reclaim, reconsider all zones for allocation */
2796 if (IS_ENABLED(CONFIG_NUMA))
2797 zlc_clear_zones_full(ac->zonelist);
2800 page = get_page_from_freelist(gfp_mask, order,
2801 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2804 * If an allocation failed after direct reclaim, it could be because
2805 * pages are pinned on the per-cpu lists. Drain them and try again
2807 if (!page && !drained) {
2808 drain_all_pages(NULL);
2817 * This is called in the allocator slow-path if the allocation request is of
2818 * sufficient urgency to ignore watermarks and take other desperate measures
2820 static inline struct page *
2821 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2822 const struct alloc_context *ac)
2827 page = get_page_from_freelist(gfp_mask, order,
2828 ALLOC_NO_WATERMARKS, ac);
2830 if (!page && gfp_mask & __GFP_NOFAIL)
2831 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC,
2833 } while (!page && (gfp_mask & __GFP_NOFAIL));
2838 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2843 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2844 ac->high_zoneidx, ac->nodemask)
2845 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2849 gfp_to_alloc_flags(gfp_t gfp_mask)
2851 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2852 const bool atomic = !(gfp_mask & (__GFP_WAIT | __GFP_NO_KSWAPD));
2854 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2855 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2858 * The caller may dip into page reserves a bit more if the caller
2859 * cannot run direct reclaim, or if the caller has realtime scheduling
2860 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2861 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2863 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2867 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2868 * if it can't schedule.
2870 if (!(gfp_mask & __GFP_NOMEMALLOC))
2871 alloc_flags |= ALLOC_HARDER;
2873 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2874 * comment for __cpuset_node_allowed().
2876 alloc_flags &= ~ALLOC_CPUSET;
2877 } else if (unlikely(rt_task(current)) && !in_interrupt())
2878 alloc_flags |= ALLOC_HARDER;
2880 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2881 if (gfp_mask & __GFP_MEMALLOC)
2882 alloc_flags |= ALLOC_NO_WATERMARKS;
2883 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2884 alloc_flags |= ALLOC_NO_WATERMARKS;
2885 else if (!in_interrupt() &&
2886 ((current->flags & PF_MEMALLOC) ||
2887 unlikely(test_thread_flag(TIF_MEMDIE))))
2888 alloc_flags |= ALLOC_NO_WATERMARKS;
2891 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2892 alloc_flags |= ALLOC_CMA;
2897 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2899 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2902 static inline struct page *
2903 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2904 struct alloc_context *ac)
2906 const gfp_t wait = gfp_mask & __GFP_WAIT;
2907 struct page *page = NULL;
2909 unsigned long pages_reclaimed = 0;
2910 unsigned long did_some_progress;
2911 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2912 bool deferred_compaction = false;
2913 int contended_compaction = COMPACT_CONTENDED_NONE;
2916 * In the slowpath, we sanity check order to avoid ever trying to
2917 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2918 * be using allocators in order of preference for an area that is
2921 if (order >= MAX_ORDER) {
2922 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2927 * If this allocation cannot block and it is for a specific node, then
2928 * fail early. There's no need to wakeup kswapd or retry for a
2929 * speculative node-specific allocation.
2931 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !wait)
2935 if (!(gfp_mask & __GFP_NO_KSWAPD))
2936 wake_all_kswapds(order, ac);
2939 * OK, we're below the kswapd watermark and have kicked background
2940 * reclaim. Now things get more complex, so set up alloc_flags according
2941 * to how we want to proceed.
2943 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2946 * Find the true preferred zone if the allocation is unconstrained by
2949 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
2950 struct zoneref *preferred_zoneref;
2951 preferred_zoneref = first_zones_zonelist(ac->zonelist,
2952 ac->high_zoneidx, NULL, &ac->preferred_zone);
2953 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
2956 /* This is the last chance, in general, before the goto nopage. */
2957 page = get_page_from_freelist(gfp_mask, order,
2958 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2962 /* Allocate without watermarks if the context allows */
2963 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2965 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2966 * the allocation is high priority and these type of
2967 * allocations are system rather than user orientated
2969 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
2971 page = __alloc_pages_high_priority(gfp_mask, order, ac);
2978 /* Atomic allocations - we can't balance anything */
2981 * All existing users of the deprecated __GFP_NOFAIL are
2982 * blockable, so warn of any new users that actually allow this
2983 * type of allocation to fail.
2985 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
2989 /* Avoid recursion of direct reclaim */
2990 if (current->flags & PF_MEMALLOC)
2993 /* Avoid allocations with no watermarks from looping endlessly */
2994 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2998 * Try direct compaction. The first pass is asynchronous. Subsequent
2999 * attempts after direct reclaim are synchronous
3001 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3003 &contended_compaction,
3004 &deferred_compaction);
3008 /* Checks for THP-specific high-order allocations */
3009 if ((gfp_mask & GFP_TRANSHUGE) == GFP_TRANSHUGE) {
3011 * If compaction is deferred for high-order allocations, it is
3012 * because sync compaction recently failed. If this is the case
3013 * and the caller requested a THP allocation, we do not want
3014 * to heavily disrupt the system, so we fail the allocation
3015 * instead of entering direct reclaim.
3017 if (deferred_compaction)
3021 * In all zones where compaction was attempted (and not
3022 * deferred or skipped), lock contention has been detected.
3023 * For THP allocation we do not want to disrupt the others
3024 * so we fallback to base pages instead.
3026 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3030 * If compaction was aborted due to need_resched(), we do not
3031 * want to further increase allocation latency, unless it is
3032 * khugepaged trying to collapse.
3034 if (contended_compaction == COMPACT_CONTENDED_SCHED
3035 && !(current->flags & PF_KTHREAD))
3040 * It can become very expensive to allocate transparent hugepages at
3041 * fault, so use asynchronous memory compaction for THP unless it is
3042 * khugepaged trying to collapse.
3044 if ((gfp_mask & GFP_TRANSHUGE) != GFP_TRANSHUGE ||
3045 (current->flags & PF_KTHREAD))
3046 migration_mode = MIGRATE_SYNC_LIGHT;
3048 /* Try direct reclaim and then allocating */
3049 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3050 &did_some_progress);
3054 /* Do not loop if specifically requested */
3055 if (gfp_mask & __GFP_NORETRY)
3058 /* Keep reclaiming pages as long as there is reasonable progress */
3059 pages_reclaimed += did_some_progress;
3060 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3061 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3062 /* Wait for some write requests to complete then retry */
3063 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3067 /* Reclaim has failed us, start killing things */
3068 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3072 /* Retry as long as the OOM killer is making progress */
3073 if (did_some_progress)
3078 * High-order allocations do not necessarily loop after
3079 * direct reclaim and reclaim/compaction depends on compaction
3080 * being called after reclaim so call directly if necessary
3082 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3084 &contended_compaction,
3085 &deferred_compaction);
3089 warn_alloc_failed(gfp_mask, order, NULL);
3095 * This is the 'heart' of the zoned buddy allocator.
3098 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3099 struct zonelist *zonelist, nodemask_t *nodemask)
3101 struct zoneref *preferred_zoneref;
3102 struct page *page = NULL;
3103 unsigned int cpuset_mems_cookie;
3104 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3105 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3106 struct alloc_context ac = {
3107 .high_zoneidx = gfp_zone(gfp_mask),
3108 .nodemask = nodemask,
3109 .migratetype = gfpflags_to_migratetype(gfp_mask),
3112 gfp_mask &= gfp_allowed_mask;
3114 lockdep_trace_alloc(gfp_mask);
3116 might_sleep_if(gfp_mask & __GFP_WAIT);
3118 if (should_fail_alloc_page(gfp_mask, order))
3122 * Check the zones suitable for the gfp_mask contain at least one
3123 * valid zone. It's possible to have an empty zonelist as a result
3124 * of __GFP_THISNODE and a memoryless node
3126 if (unlikely(!zonelist->_zonerefs->zone))
3129 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3130 alloc_flags |= ALLOC_CMA;
3133 cpuset_mems_cookie = read_mems_allowed_begin();
3135 /* We set it here, as __alloc_pages_slowpath might have changed it */
3136 ac.zonelist = zonelist;
3137 /* The preferred zone is used for statistics later */
3138 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3139 ac.nodemask ? : &cpuset_current_mems_allowed,
3140 &ac.preferred_zone);
3141 if (!ac.preferred_zone)
3143 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3145 /* First allocation attempt */
3146 alloc_mask = gfp_mask|__GFP_HARDWALL;
3147 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3148 if (unlikely(!page)) {
3150 * Runtime PM, block IO and its error handling path
3151 * can deadlock because I/O on the device might not
3154 alloc_mask = memalloc_noio_flags(gfp_mask);
3156 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3159 if (kmemcheck_enabled && page)
3160 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3162 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3166 * When updating a task's mems_allowed, it is possible to race with
3167 * parallel threads in such a way that an allocation can fail while
3168 * the mask is being updated. If a page allocation is about to fail,
3169 * check if the cpuset changed during allocation and if so, retry.
3171 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3176 EXPORT_SYMBOL(__alloc_pages_nodemask);
3179 * Common helper functions.
3181 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3186 * __get_free_pages() returns a 32-bit address, which cannot represent
3189 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3191 page = alloc_pages(gfp_mask, order);
3194 return (unsigned long) page_address(page);
3196 EXPORT_SYMBOL(__get_free_pages);
3198 unsigned long get_zeroed_page(gfp_t gfp_mask)
3200 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3202 EXPORT_SYMBOL(get_zeroed_page);
3204 void __free_pages(struct page *page, unsigned int order)
3206 if (put_page_testzero(page)) {
3208 free_hot_cold_page(page, false);
3210 __free_pages_ok(page, order);
3214 EXPORT_SYMBOL(__free_pages);
3216 void free_pages(unsigned long addr, unsigned int order)
3219 VM_BUG_ON(!virt_addr_valid((void *)addr));
3220 __free_pages(virt_to_page((void *)addr), order);
3224 EXPORT_SYMBOL(free_pages);
3228 * An arbitrary-length arbitrary-offset area of memory which resides
3229 * within a 0 or higher order page. Multiple fragments within that page
3230 * are individually refcounted, in the page's reference counter.
3232 * The page_frag functions below provide a simple allocation framework for
3233 * page fragments. This is used by the network stack and network device
3234 * drivers to provide a backing region of memory for use as either an
3235 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3237 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3240 struct page *page = NULL;
3241 gfp_t gfp = gfp_mask;
3243 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3244 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3246 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3247 PAGE_FRAG_CACHE_MAX_ORDER);
3248 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3250 if (unlikely(!page))
3251 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3253 nc->va = page ? page_address(page) : NULL;
3258 void *__alloc_page_frag(struct page_frag_cache *nc,
3259 unsigned int fragsz, gfp_t gfp_mask)
3261 unsigned int size = PAGE_SIZE;
3265 if (unlikely(!nc->va)) {
3267 page = __page_frag_refill(nc, gfp_mask);
3271 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3272 /* if size can vary use size else just use PAGE_SIZE */
3275 /* Even if we own the page, we do not use atomic_set().
3276 * This would break get_page_unless_zero() users.
3278 atomic_add(size - 1, &page->_count);
3280 /* reset page count bias and offset to start of new frag */
3281 nc->pfmemalloc = page->pfmemalloc;
3282 nc->pagecnt_bias = size;
3286 offset = nc->offset - fragsz;
3287 if (unlikely(offset < 0)) {
3288 page = virt_to_page(nc->va);
3290 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3293 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3294 /* if size can vary use size else just use PAGE_SIZE */
3297 /* OK, page count is 0, we can safely set it */
3298 atomic_set(&page->_count, size);
3300 /* reset page count bias and offset to start of new frag */
3301 nc->pagecnt_bias = size;
3302 offset = size - fragsz;
3306 nc->offset = offset;
3308 return nc->va + offset;
3310 EXPORT_SYMBOL(__alloc_page_frag);
3313 * Frees a page fragment allocated out of either a compound or order 0 page.
3315 void __free_page_frag(void *addr)
3317 struct page *page = virt_to_head_page(addr);
3319 if (unlikely(put_page_testzero(page)))
3320 __free_pages_ok(page, compound_order(page));
3322 EXPORT_SYMBOL(__free_page_frag);
3325 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3326 * of the current memory cgroup.
3328 * It should be used when the caller would like to use kmalloc, but since the
3329 * allocation is large, it has to fall back to the page allocator.
3331 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3334 struct mem_cgroup *memcg = NULL;
3336 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
3338 page = alloc_pages(gfp_mask, order);
3339 memcg_kmem_commit_charge(page, memcg, order);
3343 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3346 struct mem_cgroup *memcg = NULL;
3348 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
3350 page = alloc_pages_node(nid, gfp_mask, order);
3351 memcg_kmem_commit_charge(page, memcg, order);
3356 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3359 void __free_kmem_pages(struct page *page, unsigned int order)
3361 memcg_kmem_uncharge_pages(page, order);
3362 __free_pages(page, order);
3365 void free_kmem_pages(unsigned long addr, unsigned int order)
3368 VM_BUG_ON(!virt_addr_valid((void *)addr));
3369 __free_kmem_pages(virt_to_page((void *)addr), order);
3373 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
3376 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3377 unsigned long used = addr + PAGE_ALIGN(size);
3379 split_page(virt_to_page((void *)addr), order);
3380 while (used < alloc_end) {
3385 return (void *)addr;
3389 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3390 * @size: the number of bytes to allocate
3391 * @gfp_mask: GFP flags for the allocation
3393 * This function is similar to alloc_pages(), except that it allocates the
3394 * minimum number of pages to satisfy the request. alloc_pages() can only
3395 * allocate memory in power-of-two pages.
3397 * This function is also limited by MAX_ORDER.
3399 * Memory allocated by this function must be released by free_pages_exact().
3401 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3403 unsigned int order = get_order(size);
3406 addr = __get_free_pages(gfp_mask, order);
3407 return make_alloc_exact(addr, order, size);
3409 EXPORT_SYMBOL(alloc_pages_exact);
3412 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3414 * @nid: the preferred node ID where memory should be allocated
3415 * @size: the number of bytes to allocate
3416 * @gfp_mask: GFP flags for the allocation
3418 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3420 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3423 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3425 unsigned order = get_order(size);
3426 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3429 return make_alloc_exact((unsigned long)page_address(p), order, size);
3433 * free_pages_exact - release memory allocated via alloc_pages_exact()
3434 * @virt: the value returned by alloc_pages_exact.
3435 * @size: size of allocation, same value as passed to alloc_pages_exact().
3437 * Release the memory allocated by a previous call to alloc_pages_exact.
3439 void free_pages_exact(void *virt, size_t size)
3441 unsigned long addr = (unsigned long)virt;
3442 unsigned long end = addr + PAGE_ALIGN(size);
3444 while (addr < end) {
3449 EXPORT_SYMBOL(free_pages_exact);
3452 * nr_free_zone_pages - count number of pages beyond high watermark
3453 * @offset: The zone index of the highest zone
3455 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3456 * high watermark within all zones at or below a given zone index. For each
3457 * zone, the number of pages is calculated as:
3458 * managed_pages - high_pages
3460 static unsigned long nr_free_zone_pages(int offset)
3465 /* Just pick one node, since fallback list is circular */
3466 unsigned long sum = 0;
3468 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3470 for_each_zone_zonelist(zone, z, zonelist, offset) {
3471 unsigned long size = zone->managed_pages;
3472 unsigned long high = high_wmark_pages(zone);
3481 * nr_free_buffer_pages - count number of pages beyond high watermark
3483 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3484 * watermark within ZONE_DMA and ZONE_NORMAL.
3486 unsigned long nr_free_buffer_pages(void)
3488 return nr_free_zone_pages(gfp_zone(GFP_USER));
3490 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3493 * nr_free_pagecache_pages - count number of pages beyond high watermark
3495 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3496 * high watermark within all zones.
3498 unsigned long nr_free_pagecache_pages(void)
3500 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3503 static inline void show_node(struct zone *zone)
3505 if (IS_ENABLED(CONFIG_NUMA))
3506 printk("Node %d ", zone_to_nid(zone));
3509 void si_meminfo(struct sysinfo *val)
3511 val->totalram = totalram_pages;
3512 val->sharedram = global_page_state(NR_SHMEM);
3513 val->freeram = global_page_state(NR_FREE_PAGES);
3514 val->bufferram = nr_blockdev_pages();
3515 val->totalhigh = totalhigh_pages;
3516 val->freehigh = nr_free_highpages();
3517 val->mem_unit = PAGE_SIZE;
3520 EXPORT_SYMBOL(si_meminfo);
3523 void si_meminfo_node(struct sysinfo *val, int nid)
3525 int zone_type; /* needs to be signed */
3526 unsigned long managed_pages = 0;
3527 pg_data_t *pgdat = NODE_DATA(nid);
3529 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3530 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3531 val->totalram = managed_pages;
3532 val->sharedram = node_page_state(nid, NR_SHMEM);
3533 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3534 #ifdef CONFIG_HIGHMEM
3535 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3536 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3542 val->mem_unit = PAGE_SIZE;
3547 * Determine whether the node should be displayed or not, depending on whether
3548 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3550 bool skip_free_areas_node(unsigned int flags, int nid)
3553 unsigned int cpuset_mems_cookie;
3555 if (!(flags & SHOW_MEM_FILTER_NODES))
3559 cpuset_mems_cookie = read_mems_allowed_begin();
3560 ret = !node_isset(nid, cpuset_current_mems_allowed);
3561 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3566 #define K(x) ((x) << (PAGE_SHIFT-10))
3568 static void show_migration_types(unsigned char type)
3570 static const char types[MIGRATE_TYPES] = {
3571 [MIGRATE_UNMOVABLE] = 'U',
3572 [MIGRATE_RECLAIMABLE] = 'E',
3573 [MIGRATE_MOVABLE] = 'M',
3574 [MIGRATE_RESERVE] = 'R',
3576 [MIGRATE_CMA] = 'C',
3578 #ifdef CONFIG_MEMORY_ISOLATION
3579 [MIGRATE_ISOLATE] = 'I',
3582 char tmp[MIGRATE_TYPES + 1];
3586 for (i = 0; i < MIGRATE_TYPES; i++) {
3587 if (type & (1 << i))
3592 printk("(%s) ", tmp);
3596 * Show free area list (used inside shift_scroll-lock stuff)
3597 * We also calculate the percentage fragmentation. We do this by counting the
3598 * memory on each free list with the exception of the first item on the list.
3601 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3604 void show_free_areas(unsigned int filter)
3606 unsigned long free_pcp = 0;
3610 for_each_populated_zone(zone) {
3611 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3614 for_each_online_cpu(cpu)
3615 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3618 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3619 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3620 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3621 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3622 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3623 " free:%lu free_pcp:%lu free_cma:%lu\n",
3624 global_page_state(NR_ACTIVE_ANON),
3625 global_page_state(NR_INACTIVE_ANON),
3626 global_page_state(NR_ISOLATED_ANON),
3627 global_page_state(NR_ACTIVE_FILE),
3628 global_page_state(NR_INACTIVE_FILE),
3629 global_page_state(NR_ISOLATED_FILE),
3630 global_page_state(NR_UNEVICTABLE),
3631 global_page_state(NR_FILE_DIRTY),
3632 global_page_state(NR_WRITEBACK),
3633 global_page_state(NR_UNSTABLE_NFS),
3634 global_page_state(NR_SLAB_RECLAIMABLE),
3635 global_page_state(NR_SLAB_UNRECLAIMABLE),
3636 global_page_state(NR_FILE_MAPPED),
3637 global_page_state(NR_SHMEM),
3638 global_page_state(NR_PAGETABLE),
3639 global_page_state(NR_BOUNCE),
3640 global_page_state(NR_FREE_PAGES),
3642 global_page_state(NR_FREE_CMA_PAGES));
3644 for_each_populated_zone(zone) {
3647 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3651 for_each_online_cpu(cpu)
3652 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3660 " active_anon:%lukB"
3661 " inactive_anon:%lukB"
3662 " active_file:%lukB"
3663 " inactive_file:%lukB"
3664 " unevictable:%lukB"
3665 " isolated(anon):%lukB"
3666 " isolated(file):%lukB"
3674 " slab_reclaimable:%lukB"
3675 " slab_unreclaimable:%lukB"
3676 " kernel_stack:%lukB"
3683 " writeback_tmp:%lukB"
3684 " pages_scanned:%lu"
3685 " all_unreclaimable? %s"
3688 K(zone_page_state(zone, NR_FREE_PAGES)),
3689 K(min_wmark_pages(zone)),
3690 K(low_wmark_pages(zone)),
3691 K(high_wmark_pages(zone)),
3692 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3693 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3694 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3695 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3696 K(zone_page_state(zone, NR_UNEVICTABLE)),
3697 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3698 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3699 K(zone->present_pages),
3700 K(zone->managed_pages),
3701 K(zone_page_state(zone, NR_MLOCK)),
3702 K(zone_page_state(zone, NR_FILE_DIRTY)),
3703 K(zone_page_state(zone, NR_WRITEBACK)),
3704 K(zone_page_state(zone, NR_FILE_MAPPED)),
3705 K(zone_page_state(zone, NR_SHMEM)),
3706 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3707 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3708 zone_page_state(zone, NR_KERNEL_STACK) *
3710 K(zone_page_state(zone, NR_PAGETABLE)),
3711 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3712 K(zone_page_state(zone, NR_BOUNCE)),
3714 K(this_cpu_read(zone->pageset->pcp.count)),
3715 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3716 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3717 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3718 (!zone_reclaimable(zone) ? "yes" : "no")
3720 printk("lowmem_reserve[]:");
3721 for (i = 0; i < MAX_NR_ZONES; i++)
3722 printk(" %ld", zone->lowmem_reserve[i]);
3726 for_each_populated_zone(zone) {
3727 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3728 unsigned char types[MAX_ORDER];
3730 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3733 printk("%s: ", zone->name);
3735 spin_lock_irqsave(&zone->lock, flags);
3736 for (order = 0; order < MAX_ORDER; order++) {
3737 struct free_area *area = &zone->free_area[order];
3740 nr[order] = area->nr_free;
3741 total += nr[order] << order;
3744 for (type = 0; type < MIGRATE_TYPES; type++) {
3745 if (!list_empty(&area->free_list[type]))
3746 types[order] |= 1 << type;
3749 spin_unlock_irqrestore(&zone->lock, flags);
3750 for (order = 0; order < MAX_ORDER; order++) {
3751 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3753 show_migration_types(types[order]);
3755 printk("= %lukB\n", K(total));
3758 hugetlb_show_meminfo();
3760 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3762 show_swap_cache_info();
3765 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3767 zoneref->zone = zone;
3768 zoneref->zone_idx = zone_idx(zone);
3772 * Builds allocation fallback zone lists.
3774 * Add all populated zones of a node to the zonelist.
3776 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3780 enum zone_type zone_type = MAX_NR_ZONES;
3784 zone = pgdat->node_zones + zone_type;
3785 if (populated_zone(zone)) {
3786 zoneref_set_zone(zone,
3787 &zonelist->_zonerefs[nr_zones++]);
3788 check_highest_zone(zone_type);
3790 } while (zone_type);
3798 * 0 = automatic detection of better ordering.
3799 * 1 = order by ([node] distance, -zonetype)
3800 * 2 = order by (-zonetype, [node] distance)
3802 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3803 * the same zonelist. So only NUMA can configure this param.
3805 #define ZONELIST_ORDER_DEFAULT 0
3806 #define ZONELIST_ORDER_NODE 1
3807 #define ZONELIST_ORDER_ZONE 2
3809 /* zonelist order in the kernel.
3810 * set_zonelist_order() will set this to NODE or ZONE.
3812 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3813 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3817 /* The value user specified ....changed by config */
3818 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3819 /* string for sysctl */
3820 #define NUMA_ZONELIST_ORDER_LEN 16
3821 char numa_zonelist_order[16] = "default";
3824 * interface for configure zonelist ordering.
3825 * command line option "numa_zonelist_order"
3826 * = "[dD]efault - default, automatic configuration.
3827 * = "[nN]ode - order by node locality, then by zone within node
3828 * = "[zZ]one - order by zone, then by locality within zone
3831 static int __parse_numa_zonelist_order(char *s)
3833 if (*s == 'd' || *s == 'D') {
3834 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3835 } else if (*s == 'n' || *s == 'N') {
3836 user_zonelist_order = ZONELIST_ORDER_NODE;
3837 } else if (*s == 'z' || *s == 'Z') {
3838 user_zonelist_order = ZONELIST_ORDER_ZONE;
3841 "Ignoring invalid numa_zonelist_order value: "
3848 static __init int setup_numa_zonelist_order(char *s)
3855 ret = __parse_numa_zonelist_order(s);
3857 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3861 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3864 * sysctl handler for numa_zonelist_order
3866 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3867 void __user *buffer, size_t *length,
3870 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3872 static DEFINE_MUTEX(zl_order_mutex);
3874 mutex_lock(&zl_order_mutex);
3876 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3880 strcpy(saved_string, (char *)table->data);
3882 ret = proc_dostring(table, write, buffer, length, ppos);
3886 int oldval = user_zonelist_order;
3888 ret = __parse_numa_zonelist_order((char *)table->data);
3891 * bogus value. restore saved string
3893 strncpy((char *)table->data, saved_string,
3894 NUMA_ZONELIST_ORDER_LEN);
3895 user_zonelist_order = oldval;
3896 } else if (oldval != user_zonelist_order) {
3897 mutex_lock(&zonelists_mutex);
3898 build_all_zonelists(NULL, NULL);
3899 mutex_unlock(&zonelists_mutex);
3903 mutex_unlock(&zl_order_mutex);
3908 #define MAX_NODE_LOAD (nr_online_nodes)
3909 static int node_load[MAX_NUMNODES];
3912 * find_next_best_node - find the next node that should appear in a given node's fallback list
3913 * @node: node whose fallback list we're appending
3914 * @used_node_mask: nodemask_t of already used nodes
3916 * We use a number of factors to determine which is the next node that should
3917 * appear on a given node's fallback list. The node should not have appeared
3918 * already in @node's fallback list, and it should be the next closest node
3919 * according to the distance array (which contains arbitrary distance values
3920 * from each node to each node in the system), and should also prefer nodes
3921 * with no CPUs, since presumably they'll have very little allocation pressure
3922 * on them otherwise.
3923 * It returns -1 if no node is found.
3925 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3928 int min_val = INT_MAX;
3929 int best_node = NUMA_NO_NODE;
3930 const struct cpumask *tmp = cpumask_of_node(0);
3932 /* Use the local node if we haven't already */
3933 if (!node_isset(node, *used_node_mask)) {
3934 node_set(node, *used_node_mask);
3938 for_each_node_state(n, N_MEMORY) {
3940 /* Don't want a node to appear more than once */
3941 if (node_isset(n, *used_node_mask))
3944 /* Use the distance array to find the distance */
3945 val = node_distance(node, n);
3947 /* Penalize nodes under us ("prefer the next node") */
3950 /* Give preference to headless and unused nodes */
3951 tmp = cpumask_of_node(n);
3952 if (!cpumask_empty(tmp))
3953 val += PENALTY_FOR_NODE_WITH_CPUS;
3955 /* Slight preference for less loaded node */
3956 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3957 val += node_load[n];
3959 if (val < min_val) {
3966 node_set(best_node, *used_node_mask);
3973 * Build zonelists ordered by node and zones within node.
3974 * This results in maximum locality--normal zone overflows into local
3975 * DMA zone, if any--but risks exhausting DMA zone.
3977 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3980 struct zonelist *zonelist;
3982 zonelist = &pgdat->node_zonelists[0];
3983 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3985 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3986 zonelist->_zonerefs[j].zone = NULL;
3987 zonelist->_zonerefs[j].zone_idx = 0;
3991 * Build gfp_thisnode zonelists
3993 static void build_thisnode_zonelists(pg_data_t *pgdat)
3996 struct zonelist *zonelist;
3998 zonelist = &pgdat->node_zonelists[1];
3999 j = build_zonelists_node(pgdat, zonelist, 0);
4000 zonelist->_zonerefs[j].zone = NULL;
4001 zonelist->_zonerefs[j].zone_idx = 0;
4005 * Build zonelists ordered by zone and nodes within zones.
4006 * This results in conserving DMA zone[s] until all Normal memory is
4007 * exhausted, but results in overflowing to remote node while memory
4008 * may still exist in local DMA zone.
4010 static int node_order[MAX_NUMNODES];
4012 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4015 int zone_type; /* needs to be signed */
4017 struct zonelist *zonelist;
4019 zonelist = &pgdat->node_zonelists[0];
4021 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4022 for (j = 0; j < nr_nodes; j++) {
4023 node = node_order[j];
4024 z = &NODE_DATA(node)->node_zones[zone_type];
4025 if (populated_zone(z)) {
4027 &zonelist->_zonerefs[pos++]);
4028 check_highest_zone(zone_type);
4032 zonelist->_zonerefs[pos].zone = NULL;
4033 zonelist->_zonerefs[pos].zone_idx = 0;
4036 #if defined(CONFIG_64BIT)
4038 * Devices that require DMA32/DMA are relatively rare and do not justify a
4039 * penalty to every machine in case the specialised case applies. Default
4040 * to Node-ordering on 64-bit NUMA machines
4042 static int default_zonelist_order(void)
4044 return ZONELIST_ORDER_NODE;
4048 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4049 * by the kernel. If processes running on node 0 deplete the low memory zone
4050 * then reclaim will occur more frequency increasing stalls and potentially
4051 * be easier to OOM if a large percentage of the zone is under writeback or
4052 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4053 * Hence, default to zone ordering on 32-bit.
4055 static int default_zonelist_order(void)
4057 return ZONELIST_ORDER_ZONE;
4059 #endif /* CONFIG_64BIT */
4061 static void set_zonelist_order(void)
4063 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4064 current_zonelist_order = default_zonelist_order();
4066 current_zonelist_order = user_zonelist_order;
4069 static void build_zonelists(pg_data_t *pgdat)
4073 nodemask_t used_mask;
4074 int local_node, prev_node;
4075 struct zonelist *zonelist;
4076 int order = current_zonelist_order;
4078 /* initialize zonelists */
4079 for (i = 0; i < MAX_ZONELISTS; i++) {
4080 zonelist = pgdat->node_zonelists + i;
4081 zonelist->_zonerefs[0].zone = NULL;
4082 zonelist->_zonerefs[0].zone_idx = 0;
4085 /* NUMA-aware ordering of nodes */
4086 local_node = pgdat->node_id;
4087 load = nr_online_nodes;
4088 prev_node = local_node;
4089 nodes_clear(used_mask);
4091 memset(node_order, 0, sizeof(node_order));
4094 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4096 * We don't want to pressure a particular node.
4097 * So adding penalty to the first node in same
4098 * distance group to make it round-robin.
4100 if (node_distance(local_node, node) !=
4101 node_distance(local_node, prev_node))
4102 node_load[node] = load;
4106 if (order == ZONELIST_ORDER_NODE)
4107 build_zonelists_in_node_order(pgdat, node);
4109 node_order[j++] = node; /* remember order */
4112 if (order == ZONELIST_ORDER_ZONE) {
4113 /* calculate node order -- i.e., DMA last! */
4114 build_zonelists_in_zone_order(pgdat, j);
4117 build_thisnode_zonelists(pgdat);
4120 /* Construct the zonelist performance cache - see further mmzone.h */
4121 static void build_zonelist_cache(pg_data_t *pgdat)
4123 struct zonelist *zonelist;
4124 struct zonelist_cache *zlc;
4127 zonelist = &pgdat->node_zonelists[0];
4128 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
4129 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
4130 for (z = zonelist->_zonerefs; z->zone; z++)
4131 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
4134 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4136 * Return node id of node used for "local" allocations.
4137 * I.e., first node id of first zone in arg node's generic zonelist.
4138 * Used for initializing percpu 'numa_mem', which is used primarily
4139 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4141 int local_memory_node(int node)
4145 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4146 gfp_zone(GFP_KERNEL),
4153 #else /* CONFIG_NUMA */
4155 static void set_zonelist_order(void)
4157 current_zonelist_order = ZONELIST_ORDER_ZONE;
4160 static void build_zonelists(pg_data_t *pgdat)
4162 int node, local_node;
4164 struct zonelist *zonelist;
4166 local_node = pgdat->node_id;
4168 zonelist = &pgdat->node_zonelists[0];
4169 j = build_zonelists_node(pgdat, zonelist, 0);
4172 * Now we build the zonelist so that it contains the zones
4173 * of all the other nodes.
4174 * We don't want to pressure a particular node, so when
4175 * building the zones for node N, we make sure that the
4176 * zones coming right after the local ones are those from
4177 * node N+1 (modulo N)
4179 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4180 if (!node_online(node))
4182 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4184 for (node = 0; node < local_node; node++) {
4185 if (!node_online(node))
4187 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4190 zonelist->_zonerefs[j].zone = NULL;
4191 zonelist->_zonerefs[j].zone_idx = 0;
4194 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
4195 static void build_zonelist_cache(pg_data_t *pgdat)
4197 pgdat->node_zonelists[0].zlcache_ptr = NULL;
4200 #endif /* CONFIG_NUMA */
4203 * Boot pageset table. One per cpu which is going to be used for all
4204 * zones and all nodes. The parameters will be set in such a way
4205 * that an item put on a list will immediately be handed over to
4206 * the buddy list. This is safe since pageset manipulation is done
4207 * with interrupts disabled.
4209 * The boot_pagesets must be kept even after bootup is complete for
4210 * unused processors and/or zones. They do play a role for bootstrapping
4211 * hotplugged processors.
4213 * zoneinfo_show() and maybe other functions do
4214 * not check if the processor is online before following the pageset pointer.
4215 * Other parts of the kernel may not check if the zone is available.
4217 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4218 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4219 static void setup_zone_pageset(struct zone *zone);
4222 * Global mutex to protect against size modification of zonelists
4223 * as well as to serialize pageset setup for the new populated zone.
4225 DEFINE_MUTEX(zonelists_mutex);
4227 /* return values int ....just for stop_machine() */
4228 static int __build_all_zonelists(void *data)
4232 pg_data_t *self = data;
4235 memset(node_load, 0, sizeof(node_load));
4238 if (self && !node_online(self->node_id)) {
4239 build_zonelists(self);
4240 build_zonelist_cache(self);
4243 for_each_online_node(nid) {
4244 pg_data_t *pgdat = NODE_DATA(nid);
4246 build_zonelists(pgdat);
4247 build_zonelist_cache(pgdat);
4251 * Initialize the boot_pagesets that are going to be used
4252 * for bootstrapping processors. The real pagesets for
4253 * each zone will be allocated later when the per cpu
4254 * allocator is available.
4256 * boot_pagesets are used also for bootstrapping offline
4257 * cpus if the system is already booted because the pagesets
4258 * are needed to initialize allocators on a specific cpu too.
4259 * F.e. the percpu allocator needs the page allocator which
4260 * needs the percpu allocator in order to allocate its pagesets
4261 * (a chicken-egg dilemma).
4263 for_each_possible_cpu(cpu) {
4264 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4266 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4268 * We now know the "local memory node" for each node--
4269 * i.e., the node of the first zone in the generic zonelist.
4270 * Set up numa_mem percpu variable for on-line cpus. During
4271 * boot, only the boot cpu should be on-line; we'll init the
4272 * secondary cpus' numa_mem as they come on-line. During
4273 * node/memory hotplug, we'll fixup all on-line cpus.
4275 if (cpu_online(cpu))
4276 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4283 static noinline void __init
4284 build_all_zonelists_init(void)
4286 __build_all_zonelists(NULL);
4287 mminit_verify_zonelist();
4288 cpuset_init_current_mems_allowed();
4292 * Called with zonelists_mutex held always
4293 * unless system_state == SYSTEM_BOOTING.
4295 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4296 * [we're only called with non-NULL zone through __meminit paths] and
4297 * (2) call of __init annotated helper build_all_zonelists_init
4298 * [protected by SYSTEM_BOOTING].
4300 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4302 set_zonelist_order();
4304 if (system_state == SYSTEM_BOOTING) {
4305 build_all_zonelists_init();
4307 #ifdef CONFIG_MEMORY_HOTPLUG
4309 setup_zone_pageset(zone);
4311 /* we have to stop all cpus to guarantee there is no user
4313 stop_machine(__build_all_zonelists, pgdat, NULL);
4314 /* cpuset refresh routine should be here */
4316 vm_total_pages = nr_free_pagecache_pages();
4318 * Disable grouping by mobility if the number of pages in the
4319 * system is too low to allow the mechanism to work. It would be
4320 * more accurate, but expensive to check per-zone. This check is
4321 * made on memory-hotadd so a system can start with mobility
4322 * disabled and enable it later
4324 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4325 page_group_by_mobility_disabled = 1;
4327 page_group_by_mobility_disabled = 0;
4329 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4330 "Total pages: %ld\n",
4332 zonelist_order_name[current_zonelist_order],
4333 page_group_by_mobility_disabled ? "off" : "on",
4336 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4341 * Helper functions to size the waitqueue hash table.
4342 * Essentially these want to choose hash table sizes sufficiently
4343 * large so that collisions trying to wait on pages are rare.
4344 * But in fact, the number of active page waitqueues on typical
4345 * systems is ridiculously low, less than 200. So this is even
4346 * conservative, even though it seems large.
4348 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4349 * waitqueues, i.e. the size of the waitq table given the number of pages.
4351 #define PAGES_PER_WAITQUEUE 256
4353 #ifndef CONFIG_MEMORY_HOTPLUG
4354 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4356 unsigned long size = 1;
4358 pages /= PAGES_PER_WAITQUEUE;
4360 while (size < pages)
4364 * Once we have dozens or even hundreds of threads sleeping
4365 * on IO we've got bigger problems than wait queue collision.
4366 * Limit the size of the wait table to a reasonable size.
4368 size = min(size, 4096UL);
4370 return max(size, 4UL);
4374 * A zone's size might be changed by hot-add, so it is not possible to determine
4375 * a suitable size for its wait_table. So we use the maximum size now.
4377 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4379 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4380 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4381 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4383 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4384 * or more by the traditional way. (See above). It equals:
4386 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4387 * ia64(16K page size) : = ( 8G + 4M)byte.
4388 * powerpc (64K page size) : = (32G +16M)byte.
4390 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4397 * This is an integer logarithm so that shifts can be used later
4398 * to extract the more random high bits from the multiplicative
4399 * hash function before the remainder is taken.
4401 static inline unsigned long wait_table_bits(unsigned long size)
4407 * Check if a pageblock contains reserved pages
4409 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
4413 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4414 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
4421 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4422 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4423 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4424 * higher will lead to a bigger reserve which will get freed as contiguous
4425 * blocks as reclaim kicks in
4427 static void setup_zone_migrate_reserve(struct zone *zone)
4429 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
4431 unsigned long block_migratetype;
4436 * Get the start pfn, end pfn and the number of blocks to reserve
4437 * We have to be careful to be aligned to pageblock_nr_pages to
4438 * make sure that we always check pfn_valid for the first page in
4441 start_pfn = zone->zone_start_pfn;
4442 end_pfn = zone_end_pfn(zone);
4443 start_pfn = roundup(start_pfn, pageblock_nr_pages);
4444 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
4448 * Reserve blocks are generally in place to help high-order atomic
4449 * allocations that are short-lived. A min_free_kbytes value that
4450 * would result in more than 2 reserve blocks for atomic allocations
4451 * is assumed to be in place to help anti-fragmentation for the
4452 * future allocation of hugepages at runtime.
4454 reserve = min(2, reserve);
4455 old_reserve = zone->nr_migrate_reserve_block;
4457 /* When memory hot-add, we almost always need to do nothing */
4458 if (reserve == old_reserve)
4460 zone->nr_migrate_reserve_block = reserve;
4462 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
4463 if (!early_page_nid_uninitialised(pfn, zone_to_nid(zone)))
4466 if (!pfn_valid(pfn))
4468 page = pfn_to_page(pfn);
4470 /* Watch out for overlapping nodes */
4471 if (page_to_nid(page) != zone_to_nid(zone))
4474 block_migratetype = get_pageblock_migratetype(page);
4476 /* Only test what is necessary when the reserves are not met */
4479 * Blocks with reserved pages will never free, skip
4482 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
4483 if (pageblock_is_reserved(pfn, block_end_pfn))
4486 /* If this block is reserved, account for it */
4487 if (block_migratetype == MIGRATE_RESERVE) {
4492 /* Suitable for reserving if this block is movable */
4493 if (block_migratetype == MIGRATE_MOVABLE) {
4494 set_pageblock_migratetype(page,
4496 move_freepages_block(zone, page,
4501 } else if (!old_reserve) {
4503 * At boot time we don't need to scan the whole zone
4504 * for turning off MIGRATE_RESERVE.
4510 * If the reserve is met and this is a previous reserved block,
4513 if (block_migratetype == MIGRATE_RESERVE) {
4514 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4515 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4521 * Initially all pages are reserved - free ones are freed
4522 * up by free_all_bootmem() once the early boot process is
4523 * done. Non-atomic initialization, single-pass.
4525 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4526 unsigned long start_pfn, enum memmap_context context)
4528 pg_data_t *pgdat = NODE_DATA(nid);
4529 unsigned long end_pfn = start_pfn + size;
4532 unsigned long nr_initialised = 0;
4534 if (highest_memmap_pfn < end_pfn - 1)
4535 highest_memmap_pfn = end_pfn - 1;
4537 z = &pgdat->node_zones[zone];
4538 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4540 * There can be holes in boot-time mem_map[]s
4541 * handed to this function. They do not
4542 * exist on hotplugged memory.
4544 if (context == MEMMAP_EARLY) {
4545 if (!early_pfn_valid(pfn))
4547 if (!early_pfn_in_nid(pfn, nid))
4549 if (!update_defer_init(pgdat, pfn, end_pfn,
4553 __init_single_pfn(pfn, zone, nid);
4557 static void __meminit zone_init_free_lists(struct zone *zone)
4559 unsigned int order, t;
4560 for_each_migratetype_order(order, t) {
4561 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4562 zone->free_area[order].nr_free = 0;
4566 #ifndef __HAVE_ARCH_MEMMAP_INIT
4567 #define memmap_init(size, nid, zone, start_pfn) \
4568 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4571 static int zone_batchsize(struct zone *zone)
4577 * The per-cpu-pages pools are set to around 1000th of the
4578 * size of the zone. But no more than 1/2 of a meg.
4580 * OK, so we don't know how big the cache is. So guess.
4582 batch = zone->managed_pages / 1024;
4583 if (batch * PAGE_SIZE > 512 * 1024)
4584 batch = (512 * 1024) / PAGE_SIZE;
4585 batch /= 4; /* We effectively *= 4 below */
4590 * Clamp the batch to a 2^n - 1 value. Having a power
4591 * of 2 value was found to be more likely to have
4592 * suboptimal cache aliasing properties in some cases.
4594 * For example if 2 tasks are alternately allocating
4595 * batches of pages, one task can end up with a lot
4596 * of pages of one half of the possible page colors
4597 * and the other with pages of the other colors.
4599 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4604 /* The deferral and batching of frees should be suppressed under NOMMU
4607 * The problem is that NOMMU needs to be able to allocate large chunks
4608 * of contiguous memory as there's no hardware page translation to
4609 * assemble apparent contiguous memory from discontiguous pages.
4611 * Queueing large contiguous runs of pages for batching, however,
4612 * causes the pages to actually be freed in smaller chunks. As there
4613 * can be a significant delay between the individual batches being
4614 * recycled, this leads to the once large chunks of space being
4615 * fragmented and becoming unavailable for high-order allocations.
4622 * pcp->high and pcp->batch values are related and dependent on one another:
4623 * ->batch must never be higher then ->high.
4624 * The following function updates them in a safe manner without read side
4627 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4628 * those fields changing asynchronously (acording the the above rule).
4630 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4631 * outside of boot time (or some other assurance that no concurrent updaters
4634 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4635 unsigned long batch)
4637 /* start with a fail safe value for batch */
4641 /* Update high, then batch, in order */
4648 /* a companion to pageset_set_high() */
4649 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4651 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4654 static void pageset_init(struct per_cpu_pageset *p)
4656 struct per_cpu_pages *pcp;
4659 memset(p, 0, sizeof(*p));
4663 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4664 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4667 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4670 pageset_set_batch(p, batch);
4674 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4675 * to the value high for the pageset p.
4677 static void pageset_set_high(struct per_cpu_pageset *p,
4680 unsigned long batch = max(1UL, high / 4);
4681 if ((high / 4) > (PAGE_SHIFT * 8))
4682 batch = PAGE_SHIFT * 8;
4684 pageset_update(&p->pcp, high, batch);
4687 static void pageset_set_high_and_batch(struct zone *zone,
4688 struct per_cpu_pageset *pcp)
4690 if (percpu_pagelist_fraction)
4691 pageset_set_high(pcp,
4692 (zone->managed_pages /
4693 percpu_pagelist_fraction));
4695 pageset_set_batch(pcp, zone_batchsize(zone));
4698 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4700 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4703 pageset_set_high_and_batch(zone, pcp);
4706 static void __meminit setup_zone_pageset(struct zone *zone)
4709 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4710 for_each_possible_cpu(cpu)
4711 zone_pageset_init(zone, cpu);
4715 * Allocate per cpu pagesets and initialize them.
4716 * Before this call only boot pagesets were available.
4718 void __init setup_per_cpu_pageset(void)
4722 for_each_populated_zone(zone)
4723 setup_zone_pageset(zone);
4726 static noinline __init_refok
4727 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4733 * The per-page waitqueue mechanism uses hashed waitqueues
4736 zone->wait_table_hash_nr_entries =
4737 wait_table_hash_nr_entries(zone_size_pages);
4738 zone->wait_table_bits =
4739 wait_table_bits(zone->wait_table_hash_nr_entries);
4740 alloc_size = zone->wait_table_hash_nr_entries
4741 * sizeof(wait_queue_head_t);
4743 if (!slab_is_available()) {
4744 zone->wait_table = (wait_queue_head_t *)
4745 memblock_virt_alloc_node_nopanic(
4746 alloc_size, zone->zone_pgdat->node_id);
4749 * This case means that a zone whose size was 0 gets new memory
4750 * via memory hot-add.
4751 * But it may be the case that a new node was hot-added. In
4752 * this case vmalloc() will not be able to use this new node's
4753 * memory - this wait_table must be initialized to use this new
4754 * node itself as well.
4755 * To use this new node's memory, further consideration will be
4758 zone->wait_table = vmalloc(alloc_size);
4760 if (!zone->wait_table)
4763 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4764 init_waitqueue_head(zone->wait_table + i);
4769 static __meminit void zone_pcp_init(struct zone *zone)
4772 * per cpu subsystem is not up at this point. The following code
4773 * relies on the ability of the linker to provide the
4774 * offset of a (static) per cpu variable into the per cpu area.
4776 zone->pageset = &boot_pageset;
4778 if (populated_zone(zone))
4779 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4780 zone->name, zone->present_pages,
4781 zone_batchsize(zone));
4784 int __meminit init_currently_empty_zone(struct zone *zone,
4785 unsigned long zone_start_pfn,
4787 enum memmap_context context)
4789 struct pglist_data *pgdat = zone->zone_pgdat;
4791 ret = zone_wait_table_init(zone, size);
4794 pgdat->nr_zones = zone_idx(zone) + 1;
4796 zone->zone_start_pfn = zone_start_pfn;
4798 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4799 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4801 (unsigned long)zone_idx(zone),
4802 zone_start_pfn, (zone_start_pfn + size));
4804 zone_init_free_lists(zone);
4809 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4810 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4813 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4815 int __meminit __early_pfn_to_nid(unsigned long pfn,
4816 struct mminit_pfnnid_cache *state)
4818 unsigned long start_pfn, end_pfn;
4821 if (state->last_start <= pfn && pfn < state->last_end)
4822 return state->last_nid;
4824 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4826 state->last_start = start_pfn;
4827 state->last_end = end_pfn;
4828 state->last_nid = nid;
4833 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4836 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4837 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4838 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4840 * If an architecture guarantees that all ranges registered contain no holes
4841 * and may be freed, this this function may be used instead of calling
4842 * memblock_free_early_nid() manually.
4844 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4846 unsigned long start_pfn, end_pfn;
4849 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4850 start_pfn = min(start_pfn, max_low_pfn);
4851 end_pfn = min(end_pfn, max_low_pfn);
4853 if (start_pfn < end_pfn)
4854 memblock_free_early_nid(PFN_PHYS(start_pfn),
4855 (end_pfn - start_pfn) << PAGE_SHIFT,
4861 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4862 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4864 * If an architecture guarantees that all ranges registered contain no holes and may
4865 * be freed, this function may be used instead of calling memory_present() manually.
4867 void __init sparse_memory_present_with_active_regions(int nid)
4869 unsigned long start_pfn, end_pfn;
4872 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4873 memory_present(this_nid, start_pfn, end_pfn);
4877 * get_pfn_range_for_nid - Return the start and end page frames for a node
4878 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4879 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4880 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4882 * It returns the start and end page frame of a node based on information
4883 * provided by memblock_set_node(). If called for a node
4884 * with no available memory, a warning is printed and the start and end
4887 void __meminit get_pfn_range_for_nid(unsigned int nid,
4888 unsigned long *start_pfn, unsigned long *end_pfn)
4890 unsigned long this_start_pfn, this_end_pfn;
4896 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4897 *start_pfn = min(*start_pfn, this_start_pfn);
4898 *end_pfn = max(*end_pfn, this_end_pfn);
4901 if (*start_pfn == -1UL)
4906 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4907 * assumption is made that zones within a node are ordered in monotonic
4908 * increasing memory addresses so that the "highest" populated zone is used
4910 static void __init find_usable_zone_for_movable(void)
4913 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4914 if (zone_index == ZONE_MOVABLE)
4917 if (arch_zone_highest_possible_pfn[zone_index] >
4918 arch_zone_lowest_possible_pfn[zone_index])
4922 VM_BUG_ON(zone_index == -1);
4923 movable_zone = zone_index;
4927 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4928 * because it is sized independent of architecture. Unlike the other zones,
4929 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4930 * in each node depending on the size of each node and how evenly kernelcore
4931 * is distributed. This helper function adjusts the zone ranges
4932 * provided by the architecture for a given node by using the end of the
4933 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4934 * zones within a node are in order of monotonic increases memory addresses
4936 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4937 unsigned long zone_type,
4938 unsigned long node_start_pfn,
4939 unsigned long node_end_pfn,
4940 unsigned long *zone_start_pfn,
4941 unsigned long *zone_end_pfn)
4943 /* Only adjust if ZONE_MOVABLE is on this node */
4944 if (zone_movable_pfn[nid]) {
4945 /* Size ZONE_MOVABLE */
4946 if (zone_type == ZONE_MOVABLE) {
4947 *zone_start_pfn = zone_movable_pfn[nid];
4948 *zone_end_pfn = min(node_end_pfn,
4949 arch_zone_highest_possible_pfn[movable_zone]);
4951 /* Adjust for ZONE_MOVABLE starting within this range */
4952 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4953 *zone_end_pfn > zone_movable_pfn[nid]) {
4954 *zone_end_pfn = zone_movable_pfn[nid];
4956 /* Check if this whole range is within ZONE_MOVABLE */
4957 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4958 *zone_start_pfn = *zone_end_pfn;
4963 * Return the number of pages a zone spans in a node, including holes
4964 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4966 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4967 unsigned long zone_type,
4968 unsigned long node_start_pfn,
4969 unsigned long node_end_pfn,
4970 unsigned long *ignored)
4972 unsigned long zone_start_pfn, zone_end_pfn;
4974 /* Get the start and end of the zone */
4975 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4976 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4977 adjust_zone_range_for_zone_movable(nid, zone_type,
4978 node_start_pfn, node_end_pfn,
4979 &zone_start_pfn, &zone_end_pfn);
4981 /* Check that this node has pages within the zone's required range */
4982 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4985 /* Move the zone boundaries inside the node if necessary */
4986 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4987 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4989 /* Return the spanned pages */
4990 return zone_end_pfn - zone_start_pfn;
4994 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4995 * then all holes in the requested range will be accounted for.
4997 unsigned long __meminit __absent_pages_in_range(int nid,
4998 unsigned long range_start_pfn,
4999 unsigned long range_end_pfn)
5001 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5002 unsigned long start_pfn, end_pfn;
5005 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5006 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5007 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5008 nr_absent -= end_pfn - start_pfn;
5014 * absent_pages_in_range - Return number of page frames in holes within a range
5015 * @start_pfn: The start PFN to start searching for holes
5016 * @end_pfn: The end PFN to stop searching for holes
5018 * It returns the number of pages frames in memory holes within a range.
5020 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5021 unsigned long end_pfn)
5023 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5026 /* Return the number of page frames in holes in a zone on a node */
5027 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5028 unsigned long zone_type,
5029 unsigned long node_start_pfn,
5030 unsigned long node_end_pfn,
5031 unsigned long *ignored)
5033 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5034 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5035 unsigned long zone_start_pfn, zone_end_pfn;
5037 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5038 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5040 adjust_zone_range_for_zone_movable(nid, zone_type,
5041 node_start_pfn, node_end_pfn,
5042 &zone_start_pfn, &zone_end_pfn);
5043 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5046 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5047 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5048 unsigned long zone_type,
5049 unsigned long node_start_pfn,
5050 unsigned long node_end_pfn,
5051 unsigned long *zones_size)
5053 return zones_size[zone_type];
5056 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5057 unsigned long zone_type,
5058 unsigned long node_start_pfn,
5059 unsigned long node_end_pfn,
5060 unsigned long *zholes_size)
5065 return zholes_size[zone_type];
5068 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5070 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5071 unsigned long node_start_pfn,
5072 unsigned long node_end_pfn,
5073 unsigned long *zones_size,
5074 unsigned long *zholes_size)
5076 unsigned long realtotalpages = 0, totalpages = 0;
5079 for (i = 0; i < MAX_NR_ZONES; i++) {
5080 struct zone *zone = pgdat->node_zones + i;
5081 unsigned long size, real_size;
5083 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5087 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5088 node_start_pfn, node_end_pfn,
5090 zone->spanned_pages = size;
5091 zone->present_pages = real_size;
5094 realtotalpages += real_size;
5097 pgdat->node_spanned_pages = totalpages;
5098 pgdat->node_present_pages = realtotalpages;
5099 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5103 #ifndef CONFIG_SPARSEMEM
5105 * Calculate the size of the zone->blockflags rounded to an unsigned long
5106 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5107 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5108 * round what is now in bits to nearest long in bits, then return it in
5111 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5113 unsigned long usemapsize;
5115 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5116 usemapsize = roundup(zonesize, pageblock_nr_pages);
5117 usemapsize = usemapsize >> pageblock_order;
5118 usemapsize *= NR_PAGEBLOCK_BITS;
5119 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5121 return usemapsize / 8;
5124 static void __init setup_usemap(struct pglist_data *pgdat,
5126 unsigned long zone_start_pfn,
5127 unsigned long zonesize)
5129 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5130 zone->pageblock_flags = NULL;
5132 zone->pageblock_flags =
5133 memblock_virt_alloc_node_nopanic(usemapsize,
5137 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5138 unsigned long zone_start_pfn, unsigned long zonesize) {}
5139 #endif /* CONFIG_SPARSEMEM */
5141 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5143 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5144 void __paginginit set_pageblock_order(void)
5148 /* Check that pageblock_nr_pages has not already been setup */
5149 if (pageblock_order)
5152 if (HPAGE_SHIFT > PAGE_SHIFT)
5153 order = HUGETLB_PAGE_ORDER;
5155 order = MAX_ORDER - 1;
5158 * Assume the largest contiguous order of interest is a huge page.
5159 * This value may be variable depending on boot parameters on IA64 and
5162 pageblock_order = order;
5164 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5167 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5168 * is unused as pageblock_order is set at compile-time. See
5169 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5172 void __paginginit set_pageblock_order(void)
5176 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5178 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5179 unsigned long present_pages)
5181 unsigned long pages = spanned_pages;
5184 * Provide a more accurate estimation if there are holes within
5185 * the zone and SPARSEMEM is in use. If there are holes within the
5186 * zone, each populated memory region may cost us one or two extra
5187 * memmap pages due to alignment because memmap pages for each
5188 * populated regions may not naturally algined on page boundary.
5189 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5191 if (spanned_pages > present_pages + (present_pages >> 4) &&
5192 IS_ENABLED(CONFIG_SPARSEMEM))
5193 pages = present_pages;
5195 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5199 * Set up the zone data structures:
5200 * - mark all pages reserved
5201 * - mark all memory queues empty
5202 * - clear the memory bitmaps
5204 * NOTE: pgdat should get zeroed by caller.
5206 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
5207 unsigned long node_start_pfn, unsigned long node_end_pfn)
5210 int nid = pgdat->node_id;
5211 unsigned long zone_start_pfn = pgdat->node_start_pfn;
5214 pgdat_resize_init(pgdat);
5215 #ifdef CONFIG_NUMA_BALANCING
5216 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5217 pgdat->numabalancing_migrate_nr_pages = 0;
5218 pgdat->numabalancing_migrate_next_window = jiffies;
5220 init_waitqueue_head(&pgdat->kswapd_wait);
5221 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5222 pgdat_page_ext_init(pgdat);
5224 for (j = 0; j < MAX_NR_ZONES; j++) {
5225 struct zone *zone = pgdat->node_zones + j;
5226 unsigned long size, realsize, freesize, memmap_pages;
5228 size = zone->spanned_pages;
5229 realsize = freesize = zone->present_pages;
5232 * Adjust freesize so that it accounts for how much memory
5233 * is used by this zone for memmap. This affects the watermark
5234 * and per-cpu initialisations
5236 memmap_pages = calc_memmap_size(size, realsize);
5237 if (!is_highmem_idx(j)) {
5238 if (freesize >= memmap_pages) {
5239 freesize -= memmap_pages;
5242 " %s zone: %lu pages used for memmap\n",
5243 zone_names[j], memmap_pages);
5246 " %s zone: %lu pages exceeds freesize %lu\n",
5247 zone_names[j], memmap_pages, freesize);
5250 /* Account for reserved pages */
5251 if (j == 0 && freesize > dma_reserve) {
5252 freesize -= dma_reserve;
5253 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5254 zone_names[0], dma_reserve);
5257 if (!is_highmem_idx(j))
5258 nr_kernel_pages += freesize;
5259 /* Charge for highmem memmap if there are enough kernel pages */
5260 else if (nr_kernel_pages > memmap_pages * 2)
5261 nr_kernel_pages -= memmap_pages;
5262 nr_all_pages += freesize;
5265 * Set an approximate value for lowmem here, it will be adjusted
5266 * when the bootmem allocator frees pages into the buddy system.
5267 * And all highmem pages will be managed by the buddy system.
5269 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5272 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5274 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5276 zone->name = zone_names[j];
5277 spin_lock_init(&zone->lock);
5278 spin_lock_init(&zone->lru_lock);
5279 zone_seqlock_init(zone);
5280 zone->zone_pgdat = pgdat;
5281 zone_pcp_init(zone);
5283 /* For bootup, initialized properly in watermark setup */
5284 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5286 lruvec_init(&zone->lruvec);
5290 set_pageblock_order();
5291 setup_usemap(pgdat, zone, zone_start_pfn, size);
5292 ret = init_currently_empty_zone(zone, zone_start_pfn,
5293 size, MEMMAP_EARLY);
5295 memmap_init(size, nid, j, zone_start_pfn);
5296 zone_start_pfn += size;
5300 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5302 /* Skip empty nodes */
5303 if (!pgdat->node_spanned_pages)
5306 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5307 /* ia64 gets its own node_mem_map, before this, without bootmem */
5308 if (!pgdat->node_mem_map) {
5309 unsigned long size, start, end;
5313 * The zone's endpoints aren't required to be MAX_ORDER
5314 * aligned but the node_mem_map endpoints must be in order
5315 * for the buddy allocator to function correctly.
5317 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5318 end = pgdat_end_pfn(pgdat);
5319 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5320 size = (end - start) * sizeof(struct page);
5321 map = alloc_remap(pgdat->node_id, size);
5323 map = memblock_virt_alloc_node_nopanic(size,
5325 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
5327 #ifndef CONFIG_NEED_MULTIPLE_NODES
5329 * With no DISCONTIG, the global mem_map is just set as node 0's
5331 if (pgdat == NODE_DATA(0)) {
5332 mem_map = NODE_DATA(0)->node_mem_map;
5333 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5334 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5335 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
5336 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5339 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5342 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5343 unsigned long node_start_pfn, unsigned long *zholes_size)
5345 pg_data_t *pgdat = NODE_DATA(nid);
5346 unsigned long start_pfn = 0;
5347 unsigned long end_pfn = 0;
5349 /* pg_data_t should be reset to zero when it's allocated */
5350 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5352 reset_deferred_meminit(pgdat);
5353 pgdat->node_id = nid;
5354 pgdat->node_start_pfn = node_start_pfn;
5355 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5356 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5357 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5358 (u64)start_pfn << PAGE_SHIFT, ((u64)end_pfn << PAGE_SHIFT) - 1);
5360 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5361 zones_size, zholes_size);
5363 alloc_node_mem_map(pgdat);
5364 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5365 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5366 nid, (unsigned long)pgdat,
5367 (unsigned long)pgdat->node_mem_map);
5370 free_area_init_core(pgdat, start_pfn, end_pfn);
5373 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5375 #if MAX_NUMNODES > 1
5377 * Figure out the number of possible node ids.
5379 void __init setup_nr_node_ids(void)
5382 unsigned int highest = 0;
5384 for_each_node_mask(node, node_possible_map)
5386 nr_node_ids = highest + 1;
5391 * node_map_pfn_alignment - determine the maximum internode alignment
5393 * This function should be called after node map is populated and sorted.
5394 * It calculates the maximum power of two alignment which can distinguish
5397 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5398 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5399 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5400 * shifted, 1GiB is enough and this function will indicate so.
5402 * This is used to test whether pfn -> nid mapping of the chosen memory
5403 * model has fine enough granularity to avoid incorrect mapping for the
5404 * populated node map.
5406 * Returns the determined alignment in pfn's. 0 if there is no alignment
5407 * requirement (single node).
5409 unsigned long __init node_map_pfn_alignment(void)
5411 unsigned long accl_mask = 0, last_end = 0;
5412 unsigned long start, end, mask;
5416 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5417 if (!start || last_nid < 0 || last_nid == nid) {
5424 * Start with a mask granular enough to pin-point to the
5425 * start pfn and tick off bits one-by-one until it becomes
5426 * too coarse to separate the current node from the last.
5428 mask = ~((1 << __ffs(start)) - 1);
5429 while (mask && last_end <= (start & (mask << 1)))
5432 /* accumulate all internode masks */
5436 /* convert mask to number of pages */
5437 return ~accl_mask + 1;
5440 /* Find the lowest pfn for a node */
5441 static unsigned long __init find_min_pfn_for_node(int nid)
5443 unsigned long min_pfn = ULONG_MAX;
5444 unsigned long start_pfn;
5447 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5448 min_pfn = min(min_pfn, start_pfn);
5450 if (min_pfn == ULONG_MAX) {
5452 "Could not find start_pfn for node %d\n", nid);
5460 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5462 * It returns the minimum PFN based on information provided via
5463 * memblock_set_node().
5465 unsigned long __init find_min_pfn_with_active_regions(void)
5467 return find_min_pfn_for_node(MAX_NUMNODES);
5471 * early_calculate_totalpages()
5472 * Sum pages in active regions for movable zone.
5473 * Populate N_MEMORY for calculating usable_nodes.
5475 static unsigned long __init early_calculate_totalpages(void)
5477 unsigned long totalpages = 0;
5478 unsigned long start_pfn, end_pfn;
5481 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5482 unsigned long pages = end_pfn - start_pfn;
5484 totalpages += pages;
5486 node_set_state(nid, N_MEMORY);
5492 * Find the PFN the Movable zone begins in each node. Kernel memory
5493 * is spread evenly between nodes as long as the nodes have enough
5494 * memory. When they don't, some nodes will have more kernelcore than
5497 static void __init find_zone_movable_pfns_for_nodes(void)
5500 unsigned long usable_startpfn;
5501 unsigned long kernelcore_node, kernelcore_remaining;
5502 /* save the state before borrow the nodemask */
5503 nodemask_t saved_node_state = node_states[N_MEMORY];
5504 unsigned long totalpages = early_calculate_totalpages();
5505 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5506 struct memblock_region *r;
5508 /* Need to find movable_zone earlier when movable_node is specified. */
5509 find_usable_zone_for_movable();
5512 * If movable_node is specified, ignore kernelcore and movablecore
5515 if (movable_node_is_enabled()) {
5516 for_each_memblock(memory, r) {
5517 if (!memblock_is_hotpluggable(r))
5522 usable_startpfn = PFN_DOWN(r->base);
5523 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5524 min(usable_startpfn, zone_movable_pfn[nid]) :
5532 * If movablecore=nn[KMG] was specified, calculate what size of
5533 * kernelcore that corresponds so that memory usable for
5534 * any allocation type is evenly spread. If both kernelcore
5535 * and movablecore are specified, then the value of kernelcore
5536 * will be used for required_kernelcore if it's greater than
5537 * what movablecore would have allowed.
5539 if (required_movablecore) {
5540 unsigned long corepages;
5543 * Round-up so that ZONE_MOVABLE is at least as large as what
5544 * was requested by the user
5546 required_movablecore =
5547 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5548 corepages = totalpages - required_movablecore;
5550 required_kernelcore = max(required_kernelcore, corepages);
5553 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5554 if (!required_kernelcore)
5557 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5558 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5561 /* Spread kernelcore memory as evenly as possible throughout nodes */
5562 kernelcore_node = required_kernelcore / usable_nodes;
5563 for_each_node_state(nid, N_MEMORY) {
5564 unsigned long start_pfn, end_pfn;
5567 * Recalculate kernelcore_node if the division per node
5568 * now exceeds what is necessary to satisfy the requested
5569 * amount of memory for the kernel
5571 if (required_kernelcore < kernelcore_node)
5572 kernelcore_node = required_kernelcore / usable_nodes;
5575 * As the map is walked, we track how much memory is usable
5576 * by the kernel using kernelcore_remaining. When it is
5577 * 0, the rest of the node is usable by ZONE_MOVABLE
5579 kernelcore_remaining = kernelcore_node;
5581 /* Go through each range of PFNs within this node */
5582 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5583 unsigned long size_pages;
5585 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5586 if (start_pfn >= end_pfn)
5589 /* Account for what is only usable for kernelcore */
5590 if (start_pfn < usable_startpfn) {
5591 unsigned long kernel_pages;
5592 kernel_pages = min(end_pfn, usable_startpfn)
5595 kernelcore_remaining -= min(kernel_pages,
5596 kernelcore_remaining);
5597 required_kernelcore -= min(kernel_pages,
5598 required_kernelcore);
5600 /* Continue if range is now fully accounted */
5601 if (end_pfn <= usable_startpfn) {
5604 * Push zone_movable_pfn to the end so
5605 * that if we have to rebalance
5606 * kernelcore across nodes, we will
5607 * not double account here
5609 zone_movable_pfn[nid] = end_pfn;
5612 start_pfn = usable_startpfn;
5616 * The usable PFN range for ZONE_MOVABLE is from
5617 * start_pfn->end_pfn. Calculate size_pages as the
5618 * number of pages used as kernelcore
5620 size_pages = end_pfn - start_pfn;
5621 if (size_pages > kernelcore_remaining)
5622 size_pages = kernelcore_remaining;
5623 zone_movable_pfn[nid] = start_pfn + size_pages;
5626 * Some kernelcore has been met, update counts and
5627 * break if the kernelcore for this node has been
5630 required_kernelcore -= min(required_kernelcore,
5632 kernelcore_remaining -= size_pages;
5633 if (!kernelcore_remaining)
5639 * If there is still required_kernelcore, we do another pass with one
5640 * less node in the count. This will push zone_movable_pfn[nid] further
5641 * along on the nodes that still have memory until kernelcore is
5645 if (usable_nodes && required_kernelcore > usable_nodes)
5649 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5650 for (nid = 0; nid < MAX_NUMNODES; nid++)
5651 zone_movable_pfn[nid] =
5652 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5655 /* restore the node_state */
5656 node_states[N_MEMORY] = saved_node_state;
5659 /* Any regular or high memory on that node ? */
5660 static void check_for_memory(pg_data_t *pgdat, int nid)
5662 enum zone_type zone_type;
5664 if (N_MEMORY == N_NORMAL_MEMORY)
5667 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5668 struct zone *zone = &pgdat->node_zones[zone_type];
5669 if (populated_zone(zone)) {
5670 node_set_state(nid, N_HIGH_MEMORY);
5671 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5672 zone_type <= ZONE_NORMAL)
5673 node_set_state(nid, N_NORMAL_MEMORY);
5680 * free_area_init_nodes - Initialise all pg_data_t and zone data
5681 * @max_zone_pfn: an array of max PFNs for each zone
5683 * This will call free_area_init_node() for each active node in the system.
5684 * Using the page ranges provided by memblock_set_node(), the size of each
5685 * zone in each node and their holes is calculated. If the maximum PFN
5686 * between two adjacent zones match, it is assumed that the zone is empty.
5687 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5688 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5689 * starts where the previous one ended. For example, ZONE_DMA32 starts
5690 * at arch_max_dma_pfn.
5692 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5694 unsigned long start_pfn, end_pfn;
5697 /* Record where the zone boundaries are */
5698 memset(arch_zone_lowest_possible_pfn, 0,
5699 sizeof(arch_zone_lowest_possible_pfn));
5700 memset(arch_zone_highest_possible_pfn, 0,
5701 sizeof(arch_zone_highest_possible_pfn));
5702 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5703 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5704 for (i = 1; i < MAX_NR_ZONES; i++) {
5705 if (i == ZONE_MOVABLE)
5707 arch_zone_lowest_possible_pfn[i] =
5708 arch_zone_highest_possible_pfn[i-1];
5709 arch_zone_highest_possible_pfn[i] =
5710 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5712 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5713 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5715 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5716 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5717 find_zone_movable_pfns_for_nodes();
5719 /* Print out the zone ranges */
5720 pr_info("Zone ranges:\n");
5721 for (i = 0; i < MAX_NR_ZONES; i++) {
5722 if (i == ZONE_MOVABLE)
5724 pr_info(" %-8s ", zone_names[i]);
5725 if (arch_zone_lowest_possible_pfn[i] ==
5726 arch_zone_highest_possible_pfn[i])
5729 pr_cont("[mem %#018Lx-%#018Lx]\n",
5730 (u64)arch_zone_lowest_possible_pfn[i]
5732 ((u64)arch_zone_highest_possible_pfn[i]
5733 << PAGE_SHIFT) - 1);
5736 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5737 pr_info("Movable zone start for each node\n");
5738 for (i = 0; i < MAX_NUMNODES; i++) {
5739 if (zone_movable_pfn[i])
5740 pr_info(" Node %d: %#018Lx\n", i,
5741 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5744 /* Print out the early node map */
5745 pr_info("Early memory node ranges\n");
5746 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5747 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5748 (u64)start_pfn << PAGE_SHIFT,
5749 ((u64)end_pfn << PAGE_SHIFT) - 1);
5751 /* Initialise every node */
5752 mminit_verify_pageflags_layout();
5753 setup_nr_node_ids();
5754 for_each_online_node(nid) {
5755 pg_data_t *pgdat = NODE_DATA(nid);
5756 free_area_init_node(nid, NULL,
5757 find_min_pfn_for_node(nid), NULL);
5759 /* Any memory on that node */
5760 if (pgdat->node_present_pages)
5761 node_set_state(nid, N_MEMORY);
5762 check_for_memory(pgdat, nid);
5766 static int __init cmdline_parse_core(char *p, unsigned long *core)
5768 unsigned long long coremem;
5772 coremem = memparse(p, &p);
5773 *core = coremem >> PAGE_SHIFT;
5775 /* Paranoid check that UL is enough for the coremem value */
5776 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5782 * kernelcore=size sets the amount of memory for use for allocations that
5783 * cannot be reclaimed or migrated.
5785 static int __init cmdline_parse_kernelcore(char *p)
5787 return cmdline_parse_core(p, &required_kernelcore);
5791 * movablecore=size sets the amount of memory for use for allocations that
5792 * can be reclaimed or migrated.
5794 static int __init cmdline_parse_movablecore(char *p)
5796 return cmdline_parse_core(p, &required_movablecore);
5799 early_param("kernelcore", cmdline_parse_kernelcore);
5800 early_param("movablecore", cmdline_parse_movablecore);
5802 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5804 void adjust_managed_page_count(struct page *page, long count)
5806 spin_lock(&managed_page_count_lock);
5807 page_zone(page)->managed_pages += count;
5808 totalram_pages += count;
5809 #ifdef CONFIG_HIGHMEM
5810 if (PageHighMem(page))
5811 totalhigh_pages += count;
5813 spin_unlock(&managed_page_count_lock);
5815 EXPORT_SYMBOL(adjust_managed_page_count);
5817 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5820 unsigned long pages = 0;
5822 start = (void *)PAGE_ALIGN((unsigned long)start);
5823 end = (void *)((unsigned long)end & PAGE_MASK);
5824 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5825 if ((unsigned int)poison <= 0xFF)
5826 memset(pos, poison, PAGE_SIZE);
5827 free_reserved_page(virt_to_page(pos));
5831 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5832 s, pages << (PAGE_SHIFT - 10), start, end);
5836 EXPORT_SYMBOL(free_reserved_area);
5838 #ifdef CONFIG_HIGHMEM
5839 void free_highmem_page(struct page *page)
5841 __free_reserved_page(page);
5843 page_zone(page)->managed_pages++;
5849 void __init mem_init_print_info(const char *str)
5851 unsigned long physpages, codesize, datasize, rosize, bss_size;
5852 unsigned long init_code_size, init_data_size;
5854 physpages = get_num_physpages();
5855 codesize = _etext - _stext;
5856 datasize = _edata - _sdata;
5857 rosize = __end_rodata - __start_rodata;
5858 bss_size = __bss_stop - __bss_start;
5859 init_data_size = __init_end - __init_begin;
5860 init_code_size = _einittext - _sinittext;
5863 * Detect special cases and adjust section sizes accordingly:
5864 * 1) .init.* may be embedded into .data sections
5865 * 2) .init.text.* may be out of [__init_begin, __init_end],
5866 * please refer to arch/tile/kernel/vmlinux.lds.S.
5867 * 3) .rodata.* may be embedded into .text or .data sections.
5869 #define adj_init_size(start, end, size, pos, adj) \
5871 if (start <= pos && pos < end && size > adj) \
5875 adj_init_size(__init_begin, __init_end, init_data_size,
5876 _sinittext, init_code_size);
5877 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5878 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5879 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5880 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5882 #undef adj_init_size
5884 pr_info("Memory: %luK/%luK available "
5885 "(%luK kernel code, %luK rwdata, %luK rodata, "
5886 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5887 #ifdef CONFIG_HIGHMEM
5891 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5892 codesize >> 10, datasize >> 10, rosize >> 10,
5893 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5894 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5895 totalcma_pages << (PAGE_SHIFT-10),
5896 #ifdef CONFIG_HIGHMEM
5897 totalhigh_pages << (PAGE_SHIFT-10),
5899 str ? ", " : "", str ? str : "");
5903 * set_dma_reserve - set the specified number of pages reserved in the first zone
5904 * @new_dma_reserve: The number of pages to mark reserved
5906 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5907 * In the DMA zone, a significant percentage may be consumed by kernel image
5908 * and other unfreeable allocations which can skew the watermarks badly. This
5909 * function may optionally be used to account for unfreeable pages in the
5910 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5911 * smaller per-cpu batchsize.
5913 void __init set_dma_reserve(unsigned long new_dma_reserve)
5915 dma_reserve = new_dma_reserve;
5918 void __init free_area_init(unsigned long *zones_size)
5920 free_area_init_node(0, zones_size,
5921 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5924 static int page_alloc_cpu_notify(struct notifier_block *self,
5925 unsigned long action, void *hcpu)
5927 int cpu = (unsigned long)hcpu;
5929 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5930 lru_add_drain_cpu(cpu);
5934 * Spill the event counters of the dead processor
5935 * into the current processors event counters.
5936 * This artificially elevates the count of the current
5939 vm_events_fold_cpu(cpu);
5942 * Zero the differential counters of the dead processor
5943 * so that the vm statistics are consistent.
5945 * This is only okay since the processor is dead and cannot
5946 * race with what we are doing.
5948 cpu_vm_stats_fold(cpu);
5953 void __init page_alloc_init(void)
5955 hotcpu_notifier(page_alloc_cpu_notify, 0);
5959 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5960 * or min_free_kbytes changes.
5962 static void calculate_totalreserve_pages(void)
5964 struct pglist_data *pgdat;
5965 unsigned long reserve_pages = 0;
5966 enum zone_type i, j;
5968 for_each_online_pgdat(pgdat) {
5969 for (i = 0; i < MAX_NR_ZONES; i++) {
5970 struct zone *zone = pgdat->node_zones + i;
5973 /* Find valid and maximum lowmem_reserve in the zone */
5974 for (j = i; j < MAX_NR_ZONES; j++) {
5975 if (zone->lowmem_reserve[j] > max)
5976 max = zone->lowmem_reserve[j];
5979 /* we treat the high watermark as reserved pages. */
5980 max += high_wmark_pages(zone);
5982 if (max > zone->managed_pages)
5983 max = zone->managed_pages;
5984 reserve_pages += max;
5986 * Lowmem reserves are not available to
5987 * GFP_HIGHUSER page cache allocations and
5988 * kswapd tries to balance zones to their high
5989 * watermark. As a result, neither should be
5990 * regarded as dirtyable memory, to prevent a
5991 * situation where reclaim has to clean pages
5992 * in order to balance the zones.
5994 zone->dirty_balance_reserve = max;
5997 dirty_balance_reserve = reserve_pages;
5998 totalreserve_pages = reserve_pages;
6002 * setup_per_zone_lowmem_reserve - called whenever
6003 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
6004 * has a correct pages reserved value, so an adequate number of
6005 * pages are left in the zone after a successful __alloc_pages().
6007 static void setup_per_zone_lowmem_reserve(void)
6009 struct pglist_data *pgdat;
6010 enum zone_type j, idx;
6012 for_each_online_pgdat(pgdat) {
6013 for (j = 0; j < MAX_NR_ZONES; j++) {
6014 struct zone *zone = pgdat->node_zones + j;
6015 unsigned long managed_pages = zone->managed_pages;
6017 zone->lowmem_reserve[j] = 0;
6021 struct zone *lower_zone;
6025 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6026 sysctl_lowmem_reserve_ratio[idx] = 1;
6028 lower_zone = pgdat->node_zones + idx;
6029 lower_zone->lowmem_reserve[j] = managed_pages /
6030 sysctl_lowmem_reserve_ratio[idx];
6031 managed_pages += lower_zone->managed_pages;
6036 /* update totalreserve_pages */
6037 calculate_totalreserve_pages();
6040 static void __setup_per_zone_wmarks(void)
6042 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6043 unsigned long lowmem_pages = 0;
6045 unsigned long flags;
6047 /* Calculate total number of !ZONE_HIGHMEM pages */
6048 for_each_zone(zone) {
6049 if (!is_highmem(zone))
6050 lowmem_pages += zone->managed_pages;
6053 for_each_zone(zone) {
6056 spin_lock_irqsave(&zone->lock, flags);
6057 tmp = (u64)pages_min * zone->managed_pages;
6058 do_div(tmp, lowmem_pages);
6059 if (is_highmem(zone)) {
6061 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6062 * need highmem pages, so cap pages_min to a small
6065 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6066 * deltas control asynch page reclaim, and so should
6067 * not be capped for highmem.
6069 unsigned long min_pages;
6071 min_pages = zone->managed_pages / 1024;
6072 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6073 zone->watermark[WMARK_MIN] = min_pages;
6076 * If it's a lowmem zone, reserve a number of pages
6077 * proportionate to the zone's size.
6079 zone->watermark[WMARK_MIN] = tmp;
6082 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
6083 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
6085 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6086 high_wmark_pages(zone) - low_wmark_pages(zone) -
6087 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6089 setup_zone_migrate_reserve(zone);
6090 spin_unlock_irqrestore(&zone->lock, flags);
6093 /* update totalreserve_pages */
6094 calculate_totalreserve_pages();
6098 * setup_per_zone_wmarks - called when min_free_kbytes changes
6099 * or when memory is hot-{added|removed}
6101 * Ensures that the watermark[min,low,high] values for each zone are set
6102 * correctly with respect to min_free_kbytes.
6104 void setup_per_zone_wmarks(void)
6106 mutex_lock(&zonelists_mutex);
6107 __setup_per_zone_wmarks();
6108 mutex_unlock(&zonelists_mutex);
6112 * The inactive anon list should be small enough that the VM never has to
6113 * do too much work, but large enough that each inactive page has a chance
6114 * to be referenced again before it is swapped out.
6116 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6117 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6118 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6119 * the anonymous pages are kept on the inactive list.
6122 * memory ratio inactive anon
6123 * -------------------------------------
6132 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6134 unsigned int gb, ratio;
6136 /* Zone size in gigabytes */
6137 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6139 ratio = int_sqrt(10 * gb);
6143 zone->inactive_ratio = ratio;
6146 static void __meminit setup_per_zone_inactive_ratio(void)
6151 calculate_zone_inactive_ratio(zone);
6155 * Initialise min_free_kbytes.
6157 * For small machines we want it small (128k min). For large machines
6158 * we want it large (64MB max). But it is not linear, because network
6159 * bandwidth does not increase linearly with machine size. We use
6161 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6162 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6178 int __meminit init_per_zone_wmark_min(void)
6180 unsigned long lowmem_kbytes;
6181 int new_min_free_kbytes;
6183 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6184 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6186 if (new_min_free_kbytes > user_min_free_kbytes) {
6187 min_free_kbytes = new_min_free_kbytes;
6188 if (min_free_kbytes < 128)
6189 min_free_kbytes = 128;
6190 if (min_free_kbytes > 65536)
6191 min_free_kbytes = 65536;
6193 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6194 new_min_free_kbytes, user_min_free_kbytes);
6196 setup_per_zone_wmarks();
6197 refresh_zone_stat_thresholds();
6198 setup_per_zone_lowmem_reserve();
6199 setup_per_zone_inactive_ratio();
6202 module_init(init_per_zone_wmark_min)
6205 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6206 * that we can call two helper functions whenever min_free_kbytes
6209 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6210 void __user *buffer, size_t *length, loff_t *ppos)
6214 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6219 user_min_free_kbytes = min_free_kbytes;
6220 setup_per_zone_wmarks();
6226 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6227 void __user *buffer, size_t *length, loff_t *ppos)
6232 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6237 zone->min_unmapped_pages = (zone->managed_pages *
6238 sysctl_min_unmapped_ratio) / 100;
6242 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6243 void __user *buffer, size_t *length, loff_t *ppos)
6248 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6253 zone->min_slab_pages = (zone->managed_pages *
6254 sysctl_min_slab_ratio) / 100;
6260 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6261 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6262 * whenever sysctl_lowmem_reserve_ratio changes.
6264 * The reserve ratio obviously has absolutely no relation with the
6265 * minimum watermarks. The lowmem reserve ratio can only make sense
6266 * if in function of the boot time zone sizes.
6268 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6269 void __user *buffer, size_t *length, loff_t *ppos)
6271 proc_dointvec_minmax(table, write, buffer, length, ppos);
6272 setup_per_zone_lowmem_reserve();
6277 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6278 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6279 * pagelist can have before it gets flushed back to buddy allocator.
6281 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6282 void __user *buffer, size_t *length, loff_t *ppos)
6285 int old_percpu_pagelist_fraction;
6288 mutex_lock(&pcp_batch_high_lock);
6289 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6291 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6292 if (!write || ret < 0)
6295 /* Sanity checking to avoid pcp imbalance */
6296 if (percpu_pagelist_fraction &&
6297 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6298 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6304 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6307 for_each_populated_zone(zone) {
6310 for_each_possible_cpu(cpu)
6311 pageset_set_high_and_batch(zone,
6312 per_cpu_ptr(zone->pageset, cpu));
6315 mutex_unlock(&pcp_batch_high_lock);
6320 int hashdist = HASHDIST_DEFAULT;
6322 static int __init set_hashdist(char *str)
6326 hashdist = simple_strtoul(str, &str, 0);
6329 __setup("hashdist=", set_hashdist);
6333 * allocate a large system hash table from bootmem
6334 * - it is assumed that the hash table must contain an exact power-of-2
6335 * quantity of entries
6336 * - limit is the number of hash buckets, not the total allocation size
6338 void *__init alloc_large_system_hash(const char *tablename,
6339 unsigned long bucketsize,
6340 unsigned long numentries,
6343 unsigned int *_hash_shift,
6344 unsigned int *_hash_mask,
6345 unsigned long low_limit,
6346 unsigned long high_limit)
6348 unsigned long long max = high_limit;
6349 unsigned long log2qty, size;
6352 /* allow the kernel cmdline to have a say */
6354 /* round applicable memory size up to nearest megabyte */
6355 numentries = nr_kernel_pages;
6357 /* It isn't necessary when PAGE_SIZE >= 1MB */
6358 if (PAGE_SHIFT < 20)
6359 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6361 /* limit to 1 bucket per 2^scale bytes of low memory */
6362 if (scale > PAGE_SHIFT)
6363 numentries >>= (scale - PAGE_SHIFT);
6365 numentries <<= (PAGE_SHIFT - scale);
6367 /* Make sure we've got at least a 0-order allocation.. */
6368 if (unlikely(flags & HASH_SMALL)) {
6369 /* Makes no sense without HASH_EARLY */
6370 WARN_ON(!(flags & HASH_EARLY));
6371 if (!(numentries >> *_hash_shift)) {
6372 numentries = 1UL << *_hash_shift;
6373 BUG_ON(!numentries);
6375 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6376 numentries = PAGE_SIZE / bucketsize;
6378 numentries = roundup_pow_of_two(numentries);
6380 /* limit allocation size to 1/16 total memory by default */
6382 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6383 do_div(max, bucketsize);
6385 max = min(max, 0x80000000ULL);
6387 if (numentries < low_limit)
6388 numentries = low_limit;
6389 if (numentries > max)
6392 log2qty = ilog2(numentries);
6395 size = bucketsize << log2qty;
6396 if (flags & HASH_EARLY)
6397 table = memblock_virt_alloc_nopanic(size, 0);
6399 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6402 * If bucketsize is not a power-of-two, we may free
6403 * some pages at the end of hash table which
6404 * alloc_pages_exact() automatically does
6406 if (get_order(size) < MAX_ORDER) {
6407 table = alloc_pages_exact(size, GFP_ATOMIC);
6408 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6411 } while (!table && size > PAGE_SIZE && --log2qty);
6414 panic("Failed to allocate %s hash table\n", tablename);
6416 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6419 ilog2(size) - PAGE_SHIFT,
6423 *_hash_shift = log2qty;
6425 *_hash_mask = (1 << log2qty) - 1;
6430 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6431 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6434 #ifdef CONFIG_SPARSEMEM
6435 return __pfn_to_section(pfn)->pageblock_flags;
6437 return zone->pageblock_flags;
6438 #endif /* CONFIG_SPARSEMEM */
6441 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6443 #ifdef CONFIG_SPARSEMEM
6444 pfn &= (PAGES_PER_SECTION-1);
6445 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6447 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6448 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6449 #endif /* CONFIG_SPARSEMEM */
6453 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6454 * @page: The page within the block of interest
6455 * @pfn: The target page frame number
6456 * @end_bitidx: The last bit of interest to retrieve
6457 * @mask: mask of bits that the caller is interested in
6459 * Return: pageblock_bits flags
6461 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6462 unsigned long end_bitidx,
6466 unsigned long *bitmap;
6467 unsigned long bitidx, word_bitidx;
6470 zone = page_zone(page);
6471 bitmap = get_pageblock_bitmap(zone, pfn);
6472 bitidx = pfn_to_bitidx(zone, pfn);
6473 word_bitidx = bitidx / BITS_PER_LONG;
6474 bitidx &= (BITS_PER_LONG-1);
6476 word = bitmap[word_bitidx];
6477 bitidx += end_bitidx;
6478 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6482 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6483 * @page: The page within the block of interest
6484 * @flags: The flags to set
6485 * @pfn: The target page frame number
6486 * @end_bitidx: The last bit of interest
6487 * @mask: mask of bits that the caller is interested in
6489 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6491 unsigned long end_bitidx,
6495 unsigned long *bitmap;
6496 unsigned long bitidx, word_bitidx;
6497 unsigned long old_word, word;
6499 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6501 zone = page_zone(page);
6502 bitmap = get_pageblock_bitmap(zone, pfn);
6503 bitidx = pfn_to_bitidx(zone, pfn);
6504 word_bitidx = bitidx / BITS_PER_LONG;
6505 bitidx &= (BITS_PER_LONG-1);
6507 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6509 bitidx += end_bitidx;
6510 mask <<= (BITS_PER_LONG - bitidx - 1);
6511 flags <<= (BITS_PER_LONG - bitidx - 1);
6513 word = READ_ONCE(bitmap[word_bitidx]);
6515 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6516 if (word == old_word)
6523 * This function checks whether pageblock includes unmovable pages or not.
6524 * If @count is not zero, it is okay to include less @count unmovable pages
6526 * PageLRU check without isolation or lru_lock could race so that
6527 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6528 * expect this function should be exact.
6530 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6531 bool skip_hwpoisoned_pages)
6533 unsigned long pfn, iter, found;
6537 * For avoiding noise data, lru_add_drain_all() should be called
6538 * If ZONE_MOVABLE, the zone never contains unmovable pages
6540 if (zone_idx(zone) == ZONE_MOVABLE)
6542 mt = get_pageblock_migratetype(page);
6543 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6546 pfn = page_to_pfn(page);
6547 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6548 unsigned long check = pfn + iter;
6550 if (!pfn_valid_within(check))
6553 page = pfn_to_page(check);
6556 * Hugepages are not in LRU lists, but they're movable.
6557 * We need not scan over tail pages bacause we don't
6558 * handle each tail page individually in migration.
6560 if (PageHuge(page)) {
6561 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6566 * We can't use page_count without pin a page
6567 * because another CPU can free compound page.
6568 * This check already skips compound tails of THP
6569 * because their page->_count is zero at all time.
6571 if (!atomic_read(&page->_count)) {
6572 if (PageBuddy(page))
6573 iter += (1 << page_order(page)) - 1;
6578 * The HWPoisoned page may be not in buddy system, and
6579 * page_count() is not 0.
6581 if (skip_hwpoisoned_pages && PageHWPoison(page))
6587 * If there are RECLAIMABLE pages, we need to check
6588 * it. But now, memory offline itself doesn't call
6589 * shrink_node_slabs() and it still to be fixed.
6592 * If the page is not RAM, page_count()should be 0.
6593 * we don't need more check. This is an _used_ not-movable page.
6595 * The problematic thing here is PG_reserved pages. PG_reserved
6596 * is set to both of a memory hole page and a _used_ kernel
6605 bool is_pageblock_removable_nolock(struct page *page)
6611 * We have to be careful here because we are iterating over memory
6612 * sections which are not zone aware so we might end up outside of
6613 * the zone but still within the section.
6614 * We have to take care about the node as well. If the node is offline
6615 * its NODE_DATA will be NULL - see page_zone.
6617 if (!node_online(page_to_nid(page)))
6620 zone = page_zone(page);
6621 pfn = page_to_pfn(page);
6622 if (!zone_spans_pfn(zone, pfn))
6625 return !has_unmovable_pages(zone, page, 0, true);
6630 static unsigned long pfn_max_align_down(unsigned long pfn)
6632 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6633 pageblock_nr_pages) - 1);
6636 static unsigned long pfn_max_align_up(unsigned long pfn)
6638 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6639 pageblock_nr_pages));
6642 /* [start, end) must belong to a single zone. */
6643 static int __alloc_contig_migrate_range(struct compact_control *cc,
6644 unsigned long start, unsigned long end)
6646 /* This function is based on compact_zone() from compaction.c. */
6647 unsigned long nr_reclaimed;
6648 unsigned long pfn = start;
6649 unsigned int tries = 0;
6654 while (pfn < end || !list_empty(&cc->migratepages)) {
6655 if (fatal_signal_pending(current)) {
6660 if (list_empty(&cc->migratepages)) {
6661 cc->nr_migratepages = 0;
6662 pfn = isolate_migratepages_range(cc, pfn, end);
6668 } else if (++tries == 5) {
6669 ret = ret < 0 ? ret : -EBUSY;
6673 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6675 cc->nr_migratepages -= nr_reclaimed;
6677 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6678 NULL, 0, cc->mode, MR_CMA);
6681 putback_movable_pages(&cc->migratepages);
6688 * alloc_contig_range() -- tries to allocate given range of pages
6689 * @start: start PFN to allocate
6690 * @end: one-past-the-last PFN to allocate
6691 * @migratetype: migratetype of the underlaying pageblocks (either
6692 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6693 * in range must have the same migratetype and it must
6694 * be either of the two.
6696 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6697 * aligned, however it's the caller's responsibility to guarantee that
6698 * we are the only thread that changes migrate type of pageblocks the
6701 * The PFN range must belong to a single zone.
6703 * Returns zero on success or negative error code. On success all
6704 * pages which PFN is in [start, end) are allocated for the caller and
6705 * need to be freed with free_contig_range().
6707 int alloc_contig_range(unsigned long start, unsigned long end,
6708 unsigned migratetype)
6710 unsigned long outer_start, outer_end;
6713 struct compact_control cc = {
6714 .nr_migratepages = 0,
6716 .zone = page_zone(pfn_to_page(start)),
6717 .mode = MIGRATE_SYNC,
6718 .ignore_skip_hint = true,
6720 INIT_LIST_HEAD(&cc.migratepages);
6723 * What we do here is we mark all pageblocks in range as
6724 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6725 * have different sizes, and due to the way page allocator
6726 * work, we align the range to biggest of the two pages so
6727 * that page allocator won't try to merge buddies from
6728 * different pageblocks and change MIGRATE_ISOLATE to some
6729 * other migration type.
6731 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6732 * migrate the pages from an unaligned range (ie. pages that
6733 * we are interested in). This will put all the pages in
6734 * range back to page allocator as MIGRATE_ISOLATE.
6736 * When this is done, we take the pages in range from page
6737 * allocator removing them from the buddy system. This way
6738 * page allocator will never consider using them.
6740 * This lets us mark the pageblocks back as
6741 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6742 * aligned range but not in the unaligned, original range are
6743 * put back to page allocator so that buddy can use them.
6746 ret = start_isolate_page_range(pfn_max_align_down(start),
6747 pfn_max_align_up(end), migratetype,
6752 ret = __alloc_contig_migrate_range(&cc, start, end);
6757 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6758 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6759 * more, all pages in [start, end) are free in page allocator.
6760 * What we are going to do is to allocate all pages from
6761 * [start, end) (that is remove them from page allocator).
6763 * The only problem is that pages at the beginning and at the
6764 * end of interesting range may be not aligned with pages that
6765 * page allocator holds, ie. they can be part of higher order
6766 * pages. Because of this, we reserve the bigger range and
6767 * once this is done free the pages we are not interested in.
6769 * We don't have to hold zone->lock here because the pages are
6770 * isolated thus they won't get removed from buddy.
6773 lru_add_drain_all();
6774 drain_all_pages(cc.zone);
6777 outer_start = start;
6778 while (!PageBuddy(pfn_to_page(outer_start))) {
6779 if (++order >= MAX_ORDER) {
6783 outer_start &= ~0UL << order;
6786 /* Make sure the range is really isolated. */
6787 if (test_pages_isolated(outer_start, end, false)) {
6788 pr_info("%s: [%lx, %lx) PFNs busy\n",
6789 __func__, outer_start, end);
6794 /* Grab isolated pages from freelists. */
6795 outer_end = isolate_freepages_range(&cc, outer_start, end);
6801 /* Free head and tail (if any) */
6802 if (start != outer_start)
6803 free_contig_range(outer_start, start - outer_start);
6804 if (end != outer_end)
6805 free_contig_range(end, outer_end - end);
6808 undo_isolate_page_range(pfn_max_align_down(start),
6809 pfn_max_align_up(end), migratetype);
6813 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6815 unsigned int count = 0;
6817 for (; nr_pages--; pfn++) {
6818 struct page *page = pfn_to_page(pfn);
6820 count += page_count(page) != 1;
6823 WARN(count != 0, "%d pages are still in use!\n", count);
6827 #ifdef CONFIG_MEMORY_HOTPLUG
6829 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6830 * page high values need to be recalulated.
6832 void __meminit zone_pcp_update(struct zone *zone)
6835 mutex_lock(&pcp_batch_high_lock);
6836 for_each_possible_cpu(cpu)
6837 pageset_set_high_and_batch(zone,
6838 per_cpu_ptr(zone->pageset, cpu));
6839 mutex_unlock(&pcp_batch_high_lock);
6843 void zone_pcp_reset(struct zone *zone)
6845 unsigned long flags;
6847 struct per_cpu_pageset *pset;
6849 /* avoid races with drain_pages() */
6850 local_irq_save(flags);
6851 if (zone->pageset != &boot_pageset) {
6852 for_each_online_cpu(cpu) {
6853 pset = per_cpu_ptr(zone->pageset, cpu);
6854 drain_zonestat(zone, pset);
6856 free_percpu(zone->pageset);
6857 zone->pageset = &boot_pageset;
6859 local_irq_restore(flags);
6862 #ifdef CONFIG_MEMORY_HOTREMOVE
6864 * All pages in the range must be isolated before calling this.
6867 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6871 unsigned int order, i;
6873 unsigned long flags;
6874 /* find the first valid pfn */
6875 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6880 zone = page_zone(pfn_to_page(pfn));
6881 spin_lock_irqsave(&zone->lock, flags);
6883 while (pfn < end_pfn) {
6884 if (!pfn_valid(pfn)) {
6888 page = pfn_to_page(pfn);
6890 * The HWPoisoned page may be not in buddy system, and
6891 * page_count() is not 0.
6893 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6895 SetPageReserved(page);
6899 BUG_ON(page_count(page));
6900 BUG_ON(!PageBuddy(page));
6901 order = page_order(page);
6902 #ifdef CONFIG_DEBUG_VM
6903 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6904 pfn, 1 << order, end_pfn);
6906 list_del(&page->lru);
6907 rmv_page_order(page);
6908 zone->free_area[order].nr_free--;
6909 for (i = 0; i < (1 << order); i++)
6910 SetPageReserved((page+i));
6911 pfn += (1 << order);
6913 spin_unlock_irqrestore(&zone->lock, flags);
6917 #ifdef CONFIG_MEMORY_FAILURE
6918 bool is_free_buddy_page(struct page *page)
6920 struct zone *zone = page_zone(page);
6921 unsigned long pfn = page_to_pfn(page);
6922 unsigned long flags;
6925 spin_lock_irqsave(&zone->lock, flags);
6926 for (order = 0; order < MAX_ORDER; order++) {
6927 struct page *page_head = page - (pfn & ((1 << order) - 1));
6929 if (PageBuddy(page_head) && page_order(page_head) >= order)
6932 spin_unlock_irqrestore(&zone->lock, flags);
6934 return order < MAX_ORDER;