2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/notifier.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/stop_machine.h>
47 #include <linux/sort.h>
48 #include <linux/pfn.h>
49 #include <linux/backing-dev.h>
50 #include <linux/fault-inject.h>
51 #include <linux/page-isolation.h>
52 #include <linux/page_ext.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <linux/prefetch.h>
58 #include <linux/mm_inline.h>
59 #include <linux/migrate.h>
60 #include <linux/page_ext.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/page_owner.h>
64 #include <linux/kthread.h>
66 #include <asm/sections.h>
67 #include <asm/tlbflush.h>
68 #include <asm/div64.h>
71 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
72 static DEFINE_MUTEX(pcp_batch_high_lock);
73 #define MIN_PERCPU_PAGELIST_FRACTION (8)
75 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
76 DEFINE_PER_CPU(int, numa_node);
77 EXPORT_PER_CPU_SYMBOL(numa_node);
80 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
82 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
83 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
84 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
85 * defined in <linux/topology.h>.
87 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
88 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
89 int _node_numa_mem_[MAX_NUMNODES];
93 * Array of node states.
95 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
96 [N_POSSIBLE] = NODE_MASK_ALL,
97 [N_ONLINE] = { { [0] = 1UL } },
99 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
100 #ifdef CONFIG_HIGHMEM
101 [N_HIGH_MEMORY] = { { [0] = 1UL } },
103 #ifdef CONFIG_MOVABLE_NODE
104 [N_MEMORY] = { { [0] = 1UL } },
106 [N_CPU] = { { [0] = 1UL } },
109 EXPORT_SYMBOL(node_states);
111 /* Protect totalram_pages and zone->managed_pages */
112 static DEFINE_SPINLOCK(managed_page_count_lock);
114 unsigned long totalram_pages __read_mostly;
115 unsigned long totalreserve_pages __read_mostly;
116 unsigned long totalcma_pages __read_mostly;
118 * When calculating the number of globally allowed dirty pages, there
119 * is a certain number of per-zone reserves that should not be
120 * considered dirtyable memory. This is the sum of those reserves
121 * over all existing zones that contribute dirtyable memory.
123 unsigned long dirty_balance_reserve __read_mostly;
125 int percpu_pagelist_fraction;
126 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
128 #ifdef CONFIG_PM_SLEEP
130 * The following functions are used by the suspend/hibernate code to temporarily
131 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
132 * while devices are suspended. To avoid races with the suspend/hibernate code,
133 * they should always be called with pm_mutex held (gfp_allowed_mask also should
134 * only be modified with pm_mutex held, unless the suspend/hibernate code is
135 * guaranteed not to run in parallel with that modification).
138 static gfp_t saved_gfp_mask;
140 void pm_restore_gfp_mask(void)
142 WARN_ON(!mutex_is_locked(&pm_mutex));
143 if (saved_gfp_mask) {
144 gfp_allowed_mask = saved_gfp_mask;
149 void pm_restrict_gfp_mask(void)
151 WARN_ON(!mutex_is_locked(&pm_mutex));
152 WARN_ON(saved_gfp_mask);
153 saved_gfp_mask = gfp_allowed_mask;
154 gfp_allowed_mask &= ~GFP_IOFS;
157 bool pm_suspended_storage(void)
159 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
163 #endif /* CONFIG_PM_SLEEP */
165 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
166 int pageblock_order __read_mostly;
169 static void __free_pages_ok(struct page *page, unsigned int order);
172 * results with 256, 32 in the lowmem_reserve sysctl:
173 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
174 * 1G machine -> (16M dma, 784M normal, 224M high)
175 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
176 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
177 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
179 * TBD: should special case ZONE_DMA32 machines here - in those we normally
180 * don't need any ZONE_NORMAL reservation
182 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
183 #ifdef CONFIG_ZONE_DMA
186 #ifdef CONFIG_ZONE_DMA32
189 #ifdef CONFIG_HIGHMEM
195 EXPORT_SYMBOL(totalram_pages);
197 static char * const zone_names[MAX_NR_ZONES] = {
198 #ifdef CONFIG_ZONE_DMA
201 #ifdef CONFIG_ZONE_DMA32
205 #ifdef CONFIG_HIGHMEM
211 int min_free_kbytes = 1024;
212 int user_min_free_kbytes = -1;
214 static unsigned long __meminitdata nr_kernel_pages;
215 static unsigned long __meminitdata nr_all_pages;
216 static unsigned long __meminitdata dma_reserve;
218 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
219 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
220 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
221 static unsigned long __initdata required_kernelcore;
222 static unsigned long __initdata required_movablecore;
223 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
225 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
227 EXPORT_SYMBOL(movable_zone);
228 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
231 int nr_node_ids __read_mostly = MAX_NUMNODES;
232 int nr_online_nodes __read_mostly = 1;
233 EXPORT_SYMBOL(nr_node_ids);
234 EXPORT_SYMBOL(nr_online_nodes);
237 int page_group_by_mobility_disabled __read_mostly;
239 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
240 static inline void reset_deferred_meminit(pg_data_t *pgdat)
242 pgdat->first_deferred_pfn = ULONG_MAX;
245 /* Returns true if the struct page for the pfn is uninitialised */
246 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
248 if (pfn >= NODE_DATA(early_pfn_to_nid(pfn))->first_deferred_pfn)
254 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
256 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
263 * Returns false when the remaining initialisation should be deferred until
264 * later in the boot cycle when it can be parallelised.
266 static inline bool update_defer_init(pg_data_t *pgdat,
267 unsigned long pfn, unsigned long zone_end,
268 unsigned long *nr_initialised)
270 /* Always populate low zones for address-contrained allocations */
271 if (zone_end < pgdat_end_pfn(pgdat))
274 /* Initialise at least 2G of the highest zone */
276 if (*nr_initialised > (2UL << (30 - PAGE_SHIFT)) &&
277 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
278 pgdat->first_deferred_pfn = pfn;
285 static inline void reset_deferred_meminit(pg_data_t *pgdat)
289 static inline bool early_page_uninitialised(unsigned long pfn)
294 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
299 static inline bool update_defer_init(pg_data_t *pgdat,
300 unsigned long pfn, unsigned long zone_end,
301 unsigned long *nr_initialised)
308 void set_pageblock_migratetype(struct page *page, int migratetype)
310 if (unlikely(page_group_by_mobility_disabled &&
311 migratetype < MIGRATE_PCPTYPES))
312 migratetype = MIGRATE_UNMOVABLE;
314 set_pageblock_flags_group(page, (unsigned long)migratetype,
315 PB_migrate, PB_migrate_end);
318 #ifdef CONFIG_DEBUG_VM
319 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
323 unsigned long pfn = page_to_pfn(page);
324 unsigned long sp, start_pfn;
327 seq = zone_span_seqbegin(zone);
328 start_pfn = zone->zone_start_pfn;
329 sp = zone->spanned_pages;
330 if (!zone_spans_pfn(zone, pfn))
332 } while (zone_span_seqretry(zone, seq));
335 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
336 pfn, zone_to_nid(zone), zone->name,
337 start_pfn, start_pfn + sp);
342 static int page_is_consistent(struct zone *zone, struct page *page)
344 if (!pfn_valid_within(page_to_pfn(page)))
346 if (zone != page_zone(page))
352 * Temporary debugging check for pages not lying within a given zone.
354 static int bad_range(struct zone *zone, struct page *page)
356 if (page_outside_zone_boundaries(zone, page))
358 if (!page_is_consistent(zone, page))
364 static inline int bad_range(struct zone *zone, struct page *page)
370 static void bad_page(struct page *page, const char *reason,
371 unsigned long bad_flags)
373 static unsigned long resume;
374 static unsigned long nr_shown;
375 static unsigned long nr_unshown;
377 /* Don't complain about poisoned pages */
378 if (PageHWPoison(page)) {
379 page_mapcount_reset(page); /* remove PageBuddy */
384 * Allow a burst of 60 reports, then keep quiet for that minute;
385 * or allow a steady drip of one report per second.
387 if (nr_shown == 60) {
388 if (time_before(jiffies, resume)) {
394 "BUG: Bad page state: %lu messages suppressed\n",
401 resume = jiffies + 60 * HZ;
403 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
404 current->comm, page_to_pfn(page));
405 dump_page_badflags(page, reason, bad_flags);
410 /* Leave bad fields for debug, except PageBuddy could make trouble */
411 page_mapcount_reset(page); /* remove PageBuddy */
412 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
416 * Higher-order pages are called "compound pages". They are structured thusly:
418 * The first PAGE_SIZE page is called the "head page".
420 * The remaining PAGE_SIZE pages are called "tail pages".
422 * All pages have PG_compound set. All tail pages have their ->first_page
423 * pointing at the head page.
425 * The first tail page's ->lru.next holds the address of the compound page's
426 * put_page() function. Its ->lru.prev holds the order of allocation.
427 * This usage means that zero-order pages may not be compound.
430 static void free_compound_page(struct page *page)
432 __free_pages_ok(page, compound_order(page));
435 void prep_compound_page(struct page *page, unsigned long order)
438 int nr_pages = 1 << order;
440 set_compound_page_dtor(page, free_compound_page);
441 set_compound_order(page, order);
443 for (i = 1; i < nr_pages; i++) {
444 struct page *p = page + i;
445 set_page_count(p, 0);
446 p->first_page = page;
447 /* Make sure p->first_page is always valid for PageTail() */
453 #ifdef CONFIG_DEBUG_PAGEALLOC
454 unsigned int _debug_guardpage_minorder;
455 bool _debug_pagealloc_enabled __read_mostly;
456 bool _debug_guardpage_enabled __read_mostly;
458 static int __init early_debug_pagealloc(char *buf)
463 if (strcmp(buf, "on") == 0)
464 _debug_pagealloc_enabled = true;
468 early_param("debug_pagealloc", early_debug_pagealloc);
470 static bool need_debug_guardpage(void)
472 /* If we don't use debug_pagealloc, we don't need guard page */
473 if (!debug_pagealloc_enabled())
479 static void init_debug_guardpage(void)
481 if (!debug_pagealloc_enabled())
484 _debug_guardpage_enabled = true;
487 struct page_ext_operations debug_guardpage_ops = {
488 .need = need_debug_guardpage,
489 .init = init_debug_guardpage,
492 static int __init debug_guardpage_minorder_setup(char *buf)
496 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
497 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
500 _debug_guardpage_minorder = res;
501 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
504 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
506 static inline void set_page_guard(struct zone *zone, struct page *page,
507 unsigned int order, int migratetype)
509 struct page_ext *page_ext;
511 if (!debug_guardpage_enabled())
514 page_ext = lookup_page_ext(page);
515 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
517 INIT_LIST_HEAD(&page->lru);
518 set_page_private(page, order);
519 /* Guard pages are not available for any usage */
520 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
523 static inline void clear_page_guard(struct zone *zone, struct page *page,
524 unsigned int order, int migratetype)
526 struct page_ext *page_ext;
528 if (!debug_guardpage_enabled())
531 page_ext = lookup_page_ext(page);
532 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
534 set_page_private(page, 0);
535 if (!is_migrate_isolate(migratetype))
536 __mod_zone_freepage_state(zone, (1 << order), migratetype);
539 struct page_ext_operations debug_guardpage_ops = { NULL, };
540 static inline void set_page_guard(struct zone *zone, struct page *page,
541 unsigned int order, int migratetype) {}
542 static inline void clear_page_guard(struct zone *zone, struct page *page,
543 unsigned int order, int migratetype) {}
546 static inline void set_page_order(struct page *page, unsigned int order)
548 set_page_private(page, order);
549 __SetPageBuddy(page);
552 static inline void rmv_page_order(struct page *page)
554 __ClearPageBuddy(page);
555 set_page_private(page, 0);
559 * This function checks whether a page is free && is the buddy
560 * we can do coalesce a page and its buddy if
561 * (a) the buddy is not in a hole &&
562 * (b) the buddy is in the buddy system &&
563 * (c) a page and its buddy have the same order &&
564 * (d) a page and its buddy are in the same zone.
566 * For recording whether a page is in the buddy system, we set ->_mapcount
567 * PAGE_BUDDY_MAPCOUNT_VALUE.
568 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
569 * serialized by zone->lock.
571 * For recording page's order, we use page_private(page).
573 static inline int page_is_buddy(struct page *page, struct page *buddy,
576 if (!pfn_valid_within(page_to_pfn(buddy)))
579 if (page_is_guard(buddy) && page_order(buddy) == order) {
580 if (page_zone_id(page) != page_zone_id(buddy))
583 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
588 if (PageBuddy(buddy) && page_order(buddy) == order) {
590 * zone check is done late to avoid uselessly
591 * calculating zone/node ids for pages that could
594 if (page_zone_id(page) != page_zone_id(buddy))
597 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
605 * Freeing function for a buddy system allocator.
607 * The concept of a buddy system is to maintain direct-mapped table
608 * (containing bit values) for memory blocks of various "orders".
609 * The bottom level table contains the map for the smallest allocatable
610 * units of memory (here, pages), and each level above it describes
611 * pairs of units from the levels below, hence, "buddies".
612 * At a high level, all that happens here is marking the table entry
613 * at the bottom level available, and propagating the changes upward
614 * as necessary, plus some accounting needed to play nicely with other
615 * parts of the VM system.
616 * At each level, we keep a list of pages, which are heads of continuous
617 * free pages of length of (1 << order) and marked with _mapcount
618 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
620 * So when we are allocating or freeing one, we can derive the state of the
621 * other. That is, if we allocate a small block, and both were
622 * free, the remainder of the region must be split into blocks.
623 * If a block is freed, and its buddy is also free, then this
624 * triggers coalescing into a block of larger size.
629 static inline void __free_one_page(struct page *page,
631 struct zone *zone, unsigned int order,
634 unsigned long page_idx;
635 unsigned long combined_idx;
636 unsigned long uninitialized_var(buddy_idx);
638 int max_order = MAX_ORDER;
640 VM_BUG_ON(!zone_is_initialized(zone));
641 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
643 VM_BUG_ON(migratetype == -1);
644 if (is_migrate_isolate(migratetype)) {
646 * We restrict max order of merging to prevent merge
647 * between freepages on isolate pageblock and normal
648 * pageblock. Without this, pageblock isolation
649 * could cause incorrect freepage accounting.
651 max_order = min(MAX_ORDER, pageblock_order + 1);
653 __mod_zone_freepage_state(zone, 1 << order, migratetype);
656 page_idx = pfn & ((1 << max_order) - 1);
658 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
659 VM_BUG_ON_PAGE(bad_range(zone, page), page);
661 while (order < max_order - 1) {
662 buddy_idx = __find_buddy_index(page_idx, order);
663 buddy = page + (buddy_idx - page_idx);
664 if (!page_is_buddy(page, buddy, order))
667 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
668 * merge with it and move up one order.
670 if (page_is_guard(buddy)) {
671 clear_page_guard(zone, buddy, order, migratetype);
673 list_del(&buddy->lru);
674 zone->free_area[order].nr_free--;
675 rmv_page_order(buddy);
677 combined_idx = buddy_idx & page_idx;
678 page = page + (combined_idx - page_idx);
679 page_idx = combined_idx;
682 set_page_order(page, order);
685 * If this is not the largest possible page, check if the buddy
686 * of the next-highest order is free. If it is, it's possible
687 * that pages are being freed that will coalesce soon. In case,
688 * that is happening, add the free page to the tail of the list
689 * so it's less likely to be used soon and more likely to be merged
690 * as a higher order page
692 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
693 struct page *higher_page, *higher_buddy;
694 combined_idx = buddy_idx & page_idx;
695 higher_page = page + (combined_idx - page_idx);
696 buddy_idx = __find_buddy_index(combined_idx, order + 1);
697 higher_buddy = higher_page + (buddy_idx - combined_idx);
698 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
699 list_add_tail(&page->lru,
700 &zone->free_area[order].free_list[migratetype]);
705 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
707 zone->free_area[order].nr_free++;
710 static inline int free_pages_check(struct page *page)
712 const char *bad_reason = NULL;
713 unsigned long bad_flags = 0;
715 if (unlikely(page_mapcount(page)))
716 bad_reason = "nonzero mapcount";
717 if (unlikely(page->mapping != NULL))
718 bad_reason = "non-NULL mapping";
719 if (unlikely(atomic_read(&page->_count) != 0))
720 bad_reason = "nonzero _count";
721 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
722 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
723 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
726 if (unlikely(page->mem_cgroup))
727 bad_reason = "page still charged to cgroup";
729 if (unlikely(bad_reason)) {
730 bad_page(page, bad_reason, bad_flags);
733 page_cpupid_reset_last(page);
734 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
735 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
740 * Frees a number of pages from the PCP lists
741 * Assumes all pages on list are in same zone, and of same order.
742 * count is the number of pages to free.
744 * If the zone was previously in an "all pages pinned" state then look to
745 * see if this freeing clears that state.
747 * And clear the zone's pages_scanned counter, to hold off the "all pages are
748 * pinned" detection logic.
750 static void free_pcppages_bulk(struct zone *zone, int count,
751 struct per_cpu_pages *pcp)
756 unsigned long nr_scanned;
758 spin_lock(&zone->lock);
759 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
761 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
765 struct list_head *list;
768 * Remove pages from lists in a round-robin fashion. A
769 * batch_free count is maintained that is incremented when an
770 * empty list is encountered. This is so more pages are freed
771 * off fuller lists instead of spinning excessively around empty
776 if (++migratetype == MIGRATE_PCPTYPES)
778 list = &pcp->lists[migratetype];
779 } while (list_empty(list));
781 /* This is the only non-empty list. Free them all. */
782 if (batch_free == MIGRATE_PCPTYPES)
783 batch_free = to_free;
786 int mt; /* migratetype of the to-be-freed page */
788 page = list_entry(list->prev, struct page, lru);
789 /* must delete as __free_one_page list manipulates */
790 list_del(&page->lru);
791 mt = get_freepage_migratetype(page);
792 if (unlikely(has_isolate_pageblock(zone)))
793 mt = get_pageblock_migratetype(page);
795 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
796 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
797 trace_mm_page_pcpu_drain(page, 0, mt);
798 } while (--to_free && --batch_free && !list_empty(list));
800 spin_unlock(&zone->lock);
803 static void free_one_page(struct zone *zone,
804 struct page *page, unsigned long pfn,
808 unsigned long nr_scanned;
809 spin_lock(&zone->lock);
810 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
812 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
814 if (unlikely(has_isolate_pageblock(zone) ||
815 is_migrate_isolate(migratetype))) {
816 migratetype = get_pfnblock_migratetype(page, pfn);
818 __free_one_page(page, pfn, zone, order, migratetype);
819 spin_unlock(&zone->lock);
822 static int free_tail_pages_check(struct page *head_page, struct page *page)
824 if (!IS_ENABLED(CONFIG_DEBUG_VM))
826 if (unlikely(!PageTail(page))) {
827 bad_page(page, "PageTail not set", 0);
830 if (unlikely(page->first_page != head_page)) {
831 bad_page(page, "first_page not consistent", 0);
837 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
838 unsigned long zone, int nid)
840 set_page_links(page, zone, nid, pfn);
841 init_page_count(page);
842 page_mapcount_reset(page);
843 page_cpupid_reset_last(page);
845 INIT_LIST_HEAD(&page->lru);
846 #ifdef WANT_PAGE_VIRTUAL
847 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
848 if (!is_highmem_idx(zone))
849 set_page_address(page, __va(pfn << PAGE_SHIFT));
853 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
856 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
859 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
860 static void init_reserved_page(unsigned long pfn)
865 if (!early_page_uninitialised(pfn))
868 nid = early_pfn_to_nid(pfn);
869 pgdat = NODE_DATA(nid);
871 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
872 struct zone *zone = &pgdat->node_zones[zid];
874 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
877 __init_single_pfn(pfn, zid, nid);
880 static inline void init_reserved_page(unsigned long pfn)
883 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
886 * Initialised pages do not have PageReserved set. This function is
887 * called for each range allocated by the bootmem allocator and
888 * marks the pages PageReserved. The remaining valid pages are later
889 * sent to the buddy page allocator.
891 void __meminit reserve_bootmem_region(unsigned long start, unsigned long end)
893 unsigned long start_pfn = PFN_DOWN(start);
894 unsigned long end_pfn = PFN_UP(end);
896 for (; start_pfn < end_pfn; start_pfn++) {
897 if (pfn_valid(start_pfn)) {
898 struct page *page = pfn_to_page(start_pfn);
900 init_reserved_page(start_pfn);
901 SetPageReserved(page);
906 static bool free_pages_prepare(struct page *page, unsigned int order)
908 bool compound = PageCompound(page);
911 VM_BUG_ON_PAGE(PageTail(page), page);
912 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
914 trace_mm_page_free(page, order);
915 kmemcheck_free_shadow(page, order);
916 kasan_free_pages(page, order);
919 page->mapping = NULL;
920 bad += free_pages_check(page);
921 for (i = 1; i < (1 << order); i++) {
923 bad += free_tail_pages_check(page, page + i);
924 bad += free_pages_check(page + i);
929 reset_page_owner(page, order);
931 if (!PageHighMem(page)) {
932 debug_check_no_locks_freed(page_address(page),
934 debug_check_no_obj_freed(page_address(page),
937 arch_free_page(page, order);
938 kernel_map_pages(page, 1 << order, 0);
943 static void __free_pages_ok(struct page *page, unsigned int order)
947 unsigned long pfn = page_to_pfn(page);
949 if (!free_pages_prepare(page, order))
952 migratetype = get_pfnblock_migratetype(page, pfn);
953 local_irq_save(flags);
954 __count_vm_events(PGFREE, 1 << order);
955 set_freepage_migratetype(page, migratetype);
956 free_one_page(page_zone(page), page, pfn, order, migratetype);
957 local_irq_restore(flags);
960 static void __init __free_pages_boot_core(struct page *page,
961 unsigned long pfn, unsigned int order)
963 unsigned int nr_pages = 1 << order;
964 struct page *p = page;
968 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
970 __ClearPageReserved(p);
971 set_page_count(p, 0);
973 __ClearPageReserved(p);
974 set_page_count(p, 0);
976 page_zone(page)->managed_pages += nr_pages;
977 set_page_refcounted(page);
978 __free_pages(page, order);
981 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
982 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
984 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
986 int __meminit early_pfn_to_nid(unsigned long pfn)
988 static DEFINE_SPINLOCK(early_pfn_lock);
991 spin_lock(&early_pfn_lock);
992 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
995 spin_unlock(&early_pfn_lock);
1001 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1002 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1003 struct mminit_pfnnid_cache *state)
1007 nid = __early_pfn_to_nid(pfn, state);
1008 if (nid >= 0 && nid != node)
1013 /* Only safe to use early in boot when initialisation is single-threaded */
1014 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1016 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1021 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1025 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1026 struct mminit_pfnnid_cache *state)
1033 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1036 if (early_page_uninitialised(pfn))
1038 return __free_pages_boot_core(page, pfn, order);
1041 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1042 static void __init deferred_free_range(struct page *page,
1043 unsigned long pfn, int nr_pages)
1050 /* Free a large naturally-aligned chunk if possible */
1051 if (nr_pages == MAX_ORDER_NR_PAGES &&
1052 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1053 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1054 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1058 for (i = 0; i < nr_pages; i++, page++, pfn++)
1059 __free_pages_boot_core(page, pfn, 0);
1062 /* Completion tracking for deferred_init_memmap() threads */
1063 static atomic_t pgdat_init_n_undone __initdata;
1064 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1066 static inline void __init pgdat_init_report_one_done(void)
1068 if (atomic_dec_and_test(&pgdat_init_n_undone))
1069 complete(&pgdat_init_all_done_comp);
1072 /* Initialise remaining memory on a node */
1073 static int __init deferred_init_memmap(void *data)
1075 pg_data_t *pgdat = data;
1076 int nid = pgdat->node_id;
1077 struct mminit_pfnnid_cache nid_init_state = { };
1078 unsigned long start = jiffies;
1079 unsigned long nr_pages = 0;
1080 unsigned long walk_start, walk_end;
1083 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1084 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1086 if (first_init_pfn == ULONG_MAX) {
1087 pgdat_init_report_one_done();
1091 /* Bind memory initialisation thread to a local node if possible */
1092 if (!cpumask_empty(cpumask))
1093 set_cpus_allowed_ptr(current, cpumask);
1095 /* Sanity check boundaries */
1096 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1097 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1098 pgdat->first_deferred_pfn = ULONG_MAX;
1100 /* Only the highest zone is deferred so find it */
1101 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1102 zone = pgdat->node_zones + zid;
1103 if (first_init_pfn < zone_end_pfn(zone))
1107 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1108 unsigned long pfn, end_pfn;
1109 struct page *page = NULL;
1110 struct page *free_base_page = NULL;
1111 unsigned long free_base_pfn = 0;
1114 end_pfn = min(walk_end, zone_end_pfn(zone));
1115 pfn = first_init_pfn;
1116 if (pfn < walk_start)
1118 if (pfn < zone->zone_start_pfn)
1119 pfn = zone->zone_start_pfn;
1121 for (; pfn < end_pfn; pfn++) {
1122 if (!pfn_valid_within(pfn))
1126 * Ensure pfn_valid is checked every
1127 * MAX_ORDER_NR_PAGES for memory holes
1129 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1130 if (!pfn_valid(pfn)) {
1136 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1141 /* Minimise pfn page lookups and scheduler checks */
1142 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1145 nr_pages += nr_to_free;
1146 deferred_free_range(free_base_page,
1147 free_base_pfn, nr_to_free);
1148 free_base_page = NULL;
1149 free_base_pfn = nr_to_free = 0;
1151 page = pfn_to_page(pfn);
1156 VM_BUG_ON(page_zone(page) != zone);
1160 __init_single_page(page, pfn, zid, nid);
1161 if (!free_base_page) {
1162 free_base_page = page;
1163 free_base_pfn = pfn;
1168 /* Where possible, batch up pages for a single free */
1171 /* Free the current block of pages to allocator */
1172 nr_pages += nr_to_free;
1173 deferred_free_range(free_base_page, free_base_pfn,
1175 free_base_page = NULL;
1176 free_base_pfn = nr_to_free = 0;
1179 first_init_pfn = max(end_pfn, first_init_pfn);
1182 /* Sanity check that the next zone really is unpopulated */
1183 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1185 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1186 jiffies_to_msecs(jiffies - start));
1188 pgdat_init_report_one_done();
1192 void __init page_alloc_init_late(void)
1196 /* There will be num_node_state(N_MEMORY) threads */
1197 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1198 for_each_node_state(nid, N_MEMORY) {
1199 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1202 /* Block until all are initialised */
1203 wait_for_completion(&pgdat_init_all_done_comp);
1205 /* Reinit limits that are based on free pages after the kernel is up */
1206 files_maxfiles_init();
1208 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1211 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1212 void __init init_cma_reserved_pageblock(struct page *page)
1214 unsigned i = pageblock_nr_pages;
1215 struct page *p = page;
1218 __ClearPageReserved(p);
1219 set_page_count(p, 0);
1222 set_pageblock_migratetype(page, MIGRATE_CMA);
1224 if (pageblock_order >= MAX_ORDER) {
1225 i = pageblock_nr_pages;
1228 set_page_refcounted(p);
1229 __free_pages(p, MAX_ORDER - 1);
1230 p += MAX_ORDER_NR_PAGES;
1231 } while (i -= MAX_ORDER_NR_PAGES);
1233 set_page_refcounted(page);
1234 __free_pages(page, pageblock_order);
1237 adjust_managed_page_count(page, pageblock_nr_pages);
1242 * The order of subdivision here is critical for the IO subsystem.
1243 * Please do not alter this order without good reasons and regression
1244 * testing. Specifically, as large blocks of memory are subdivided,
1245 * the order in which smaller blocks are delivered depends on the order
1246 * they're subdivided in this function. This is the primary factor
1247 * influencing the order in which pages are delivered to the IO
1248 * subsystem according to empirical testing, and this is also justified
1249 * by considering the behavior of a buddy system containing a single
1250 * large block of memory acted on by a series of small allocations.
1251 * This behavior is a critical factor in sglist merging's success.
1255 static inline void expand(struct zone *zone, struct page *page,
1256 int low, int high, struct free_area *area,
1259 unsigned long size = 1 << high;
1261 while (high > low) {
1265 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1267 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1268 debug_guardpage_enabled() &&
1269 high < debug_guardpage_minorder()) {
1271 * Mark as guard pages (or page), that will allow to
1272 * merge back to allocator when buddy will be freed.
1273 * Corresponding page table entries will not be touched,
1274 * pages will stay not present in virtual address space
1276 set_page_guard(zone, &page[size], high, migratetype);
1279 list_add(&page[size].lru, &area->free_list[migratetype]);
1281 set_page_order(&page[size], high);
1286 * This page is about to be returned from the page allocator
1288 static inline int check_new_page(struct page *page)
1290 const char *bad_reason = NULL;
1291 unsigned long bad_flags = 0;
1293 if (unlikely(page_mapcount(page)))
1294 bad_reason = "nonzero mapcount";
1295 if (unlikely(page->mapping != NULL))
1296 bad_reason = "non-NULL mapping";
1297 if (unlikely(atomic_read(&page->_count) != 0))
1298 bad_reason = "nonzero _count";
1299 if (unlikely(page->flags & __PG_HWPOISON)) {
1300 bad_reason = "HWPoisoned (hardware-corrupted)";
1301 bad_flags = __PG_HWPOISON;
1303 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1304 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1305 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1308 if (unlikely(page->mem_cgroup))
1309 bad_reason = "page still charged to cgroup";
1311 if (unlikely(bad_reason)) {
1312 bad_page(page, bad_reason, bad_flags);
1318 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1323 for (i = 0; i < (1 << order); i++) {
1324 struct page *p = page + i;
1325 if (unlikely(check_new_page(p)))
1329 set_page_private(page, 0);
1330 set_page_refcounted(page);
1332 arch_alloc_page(page, order);
1333 kernel_map_pages(page, 1 << order, 1);
1334 kasan_alloc_pages(page, order);
1336 if (gfp_flags & __GFP_ZERO)
1337 for (i = 0; i < (1 << order); i++)
1338 clear_highpage(page + i);
1340 if (order && (gfp_flags & __GFP_COMP))
1341 prep_compound_page(page, order);
1343 set_page_owner(page, order, gfp_flags);
1346 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was necessary to
1347 * allocate the page. The expectation is that the caller is taking
1348 * steps that will free more memory. The caller should avoid the page
1349 * being used for !PFMEMALLOC purposes.
1351 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
1357 * Go through the free lists for the given migratetype and remove
1358 * the smallest available page from the freelists
1361 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1364 unsigned int current_order;
1365 struct free_area *area;
1368 /* Find a page of the appropriate size in the preferred list */
1369 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1370 area = &(zone->free_area[current_order]);
1371 if (list_empty(&area->free_list[migratetype]))
1374 page = list_entry(area->free_list[migratetype].next,
1376 list_del(&page->lru);
1377 rmv_page_order(page);
1379 expand(zone, page, order, current_order, area, migratetype);
1380 set_freepage_migratetype(page, migratetype);
1389 * This array describes the order lists are fallen back to when
1390 * the free lists for the desirable migrate type are depleted
1392 static int fallbacks[MIGRATE_TYPES][4] = {
1393 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1394 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1395 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
1397 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
1399 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
1400 #ifdef CONFIG_MEMORY_ISOLATION
1401 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
1406 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1409 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1412 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1413 unsigned int order) { return NULL; }
1417 * Move the free pages in a range to the free lists of the requested type.
1418 * Note that start_page and end_pages are not aligned on a pageblock
1419 * boundary. If alignment is required, use move_freepages_block()
1421 int move_freepages(struct zone *zone,
1422 struct page *start_page, struct page *end_page,
1426 unsigned long order;
1427 int pages_moved = 0;
1429 #ifndef CONFIG_HOLES_IN_ZONE
1431 * page_zone is not safe to call in this context when
1432 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1433 * anyway as we check zone boundaries in move_freepages_block().
1434 * Remove at a later date when no bug reports exist related to
1435 * grouping pages by mobility
1437 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1440 for (page = start_page; page <= end_page;) {
1441 /* Make sure we are not inadvertently changing nodes */
1442 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1444 if (!pfn_valid_within(page_to_pfn(page))) {
1449 if (!PageBuddy(page)) {
1454 order = page_order(page);
1455 list_move(&page->lru,
1456 &zone->free_area[order].free_list[migratetype]);
1457 set_freepage_migratetype(page, migratetype);
1459 pages_moved += 1 << order;
1465 int move_freepages_block(struct zone *zone, struct page *page,
1468 unsigned long start_pfn, end_pfn;
1469 struct page *start_page, *end_page;
1471 start_pfn = page_to_pfn(page);
1472 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1473 start_page = pfn_to_page(start_pfn);
1474 end_page = start_page + pageblock_nr_pages - 1;
1475 end_pfn = start_pfn + pageblock_nr_pages - 1;
1477 /* Do not cross zone boundaries */
1478 if (!zone_spans_pfn(zone, start_pfn))
1480 if (!zone_spans_pfn(zone, end_pfn))
1483 return move_freepages(zone, start_page, end_page, migratetype);
1486 static void change_pageblock_range(struct page *pageblock_page,
1487 int start_order, int migratetype)
1489 int nr_pageblocks = 1 << (start_order - pageblock_order);
1491 while (nr_pageblocks--) {
1492 set_pageblock_migratetype(pageblock_page, migratetype);
1493 pageblock_page += pageblock_nr_pages;
1498 * When we are falling back to another migratetype during allocation, try to
1499 * steal extra free pages from the same pageblocks to satisfy further
1500 * allocations, instead of polluting multiple pageblocks.
1502 * If we are stealing a relatively large buddy page, it is likely there will
1503 * be more free pages in the pageblock, so try to steal them all. For
1504 * reclaimable and unmovable allocations, we steal regardless of page size,
1505 * as fragmentation caused by those allocations polluting movable pageblocks
1506 * is worse than movable allocations stealing from unmovable and reclaimable
1509 static bool can_steal_fallback(unsigned int order, int start_mt)
1512 * Leaving this order check is intended, although there is
1513 * relaxed order check in next check. The reason is that
1514 * we can actually steal whole pageblock if this condition met,
1515 * but, below check doesn't guarantee it and that is just heuristic
1516 * so could be changed anytime.
1518 if (order >= pageblock_order)
1521 if (order >= pageblock_order / 2 ||
1522 start_mt == MIGRATE_RECLAIMABLE ||
1523 start_mt == MIGRATE_UNMOVABLE ||
1524 page_group_by_mobility_disabled)
1531 * This function implements actual steal behaviour. If order is large enough,
1532 * we can steal whole pageblock. If not, we first move freepages in this
1533 * pageblock and check whether half of pages are moved or not. If half of
1534 * pages are moved, we can change migratetype of pageblock and permanently
1535 * use it's pages as requested migratetype in the future.
1537 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1540 int current_order = page_order(page);
1543 /* Take ownership for orders >= pageblock_order */
1544 if (current_order >= pageblock_order) {
1545 change_pageblock_range(page, current_order, start_type);
1549 pages = move_freepages_block(zone, page, start_type);
1551 /* Claim the whole block if over half of it is free */
1552 if (pages >= (1 << (pageblock_order-1)) ||
1553 page_group_by_mobility_disabled)
1554 set_pageblock_migratetype(page, start_type);
1558 * Check whether there is a suitable fallback freepage with requested order.
1559 * If only_stealable is true, this function returns fallback_mt only if
1560 * we can steal other freepages all together. This would help to reduce
1561 * fragmentation due to mixed migratetype pages in one pageblock.
1563 int find_suitable_fallback(struct free_area *area, unsigned int order,
1564 int migratetype, bool only_stealable, bool *can_steal)
1569 if (area->nr_free == 0)
1574 fallback_mt = fallbacks[migratetype][i];
1575 if (fallback_mt == MIGRATE_RESERVE)
1578 if (list_empty(&area->free_list[fallback_mt]))
1581 if (can_steal_fallback(order, migratetype))
1584 if (!only_stealable)
1594 /* Remove an element from the buddy allocator from the fallback list */
1595 static inline struct page *
1596 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1598 struct free_area *area;
1599 unsigned int current_order;
1604 /* Find the largest possible block of pages in the other list */
1605 for (current_order = MAX_ORDER-1;
1606 current_order >= order && current_order <= MAX_ORDER-1;
1608 area = &(zone->free_area[current_order]);
1609 fallback_mt = find_suitable_fallback(area, current_order,
1610 start_migratetype, false, &can_steal);
1611 if (fallback_mt == -1)
1614 page = list_entry(area->free_list[fallback_mt].next,
1617 steal_suitable_fallback(zone, page, start_migratetype);
1619 /* Remove the page from the freelists */
1621 list_del(&page->lru);
1622 rmv_page_order(page);
1624 expand(zone, page, order, current_order, area,
1627 * The freepage_migratetype may differ from pageblock's
1628 * migratetype depending on the decisions in
1629 * try_to_steal_freepages(). This is OK as long as it
1630 * does not differ for MIGRATE_CMA pageblocks. For CMA
1631 * we need to make sure unallocated pages flushed from
1632 * pcp lists are returned to the correct freelist.
1634 set_freepage_migratetype(page, start_migratetype);
1636 trace_mm_page_alloc_extfrag(page, order, current_order,
1637 start_migratetype, fallback_mt);
1646 * Do the hard work of removing an element from the buddy allocator.
1647 * Call me with the zone->lock already held.
1649 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1655 page = __rmqueue_smallest(zone, order, migratetype);
1657 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1658 if (migratetype == MIGRATE_MOVABLE)
1659 page = __rmqueue_cma_fallback(zone, order);
1662 page = __rmqueue_fallback(zone, order, migratetype);
1665 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1666 * is used because __rmqueue_smallest is an inline function
1667 * and we want just one call site
1670 migratetype = MIGRATE_RESERVE;
1675 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1680 * Obtain a specified number of elements from the buddy allocator, all under
1681 * a single hold of the lock, for efficiency. Add them to the supplied list.
1682 * Returns the number of new pages which were placed at *list.
1684 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1685 unsigned long count, struct list_head *list,
1686 int migratetype, bool cold)
1690 spin_lock(&zone->lock);
1691 for (i = 0; i < count; ++i) {
1692 struct page *page = __rmqueue(zone, order, migratetype);
1693 if (unlikely(page == NULL))
1697 * Split buddy pages returned by expand() are received here
1698 * in physical page order. The page is added to the callers and
1699 * list and the list head then moves forward. From the callers
1700 * perspective, the linked list is ordered by page number in
1701 * some conditions. This is useful for IO devices that can
1702 * merge IO requests if the physical pages are ordered
1706 list_add(&page->lru, list);
1708 list_add_tail(&page->lru, list);
1710 if (is_migrate_cma(get_freepage_migratetype(page)))
1711 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1714 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1715 spin_unlock(&zone->lock);
1721 * Called from the vmstat counter updater to drain pagesets of this
1722 * currently executing processor on remote nodes after they have
1725 * Note that this function must be called with the thread pinned to
1726 * a single processor.
1728 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1730 unsigned long flags;
1731 int to_drain, batch;
1733 local_irq_save(flags);
1734 batch = READ_ONCE(pcp->batch);
1735 to_drain = min(pcp->count, batch);
1737 free_pcppages_bulk(zone, to_drain, pcp);
1738 pcp->count -= to_drain;
1740 local_irq_restore(flags);
1745 * Drain pcplists of the indicated processor and zone.
1747 * The processor must either be the current processor and the
1748 * thread pinned to the current processor or a processor that
1751 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1753 unsigned long flags;
1754 struct per_cpu_pageset *pset;
1755 struct per_cpu_pages *pcp;
1757 local_irq_save(flags);
1758 pset = per_cpu_ptr(zone->pageset, cpu);
1762 free_pcppages_bulk(zone, pcp->count, pcp);
1765 local_irq_restore(flags);
1769 * Drain pcplists of all zones on the indicated processor.
1771 * The processor must either be the current processor and the
1772 * thread pinned to the current processor or a processor that
1775 static void drain_pages(unsigned int cpu)
1779 for_each_populated_zone(zone) {
1780 drain_pages_zone(cpu, zone);
1785 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1787 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1788 * the single zone's pages.
1790 void drain_local_pages(struct zone *zone)
1792 int cpu = smp_processor_id();
1795 drain_pages_zone(cpu, zone);
1801 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1803 * When zone parameter is non-NULL, spill just the single zone's pages.
1805 * Note that this code is protected against sending an IPI to an offline
1806 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1807 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1808 * nothing keeps CPUs from showing up after we populated the cpumask and
1809 * before the call to on_each_cpu_mask().
1811 void drain_all_pages(struct zone *zone)
1816 * Allocate in the BSS so we wont require allocation in
1817 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1819 static cpumask_t cpus_with_pcps;
1822 * We don't care about racing with CPU hotplug event
1823 * as offline notification will cause the notified
1824 * cpu to drain that CPU pcps and on_each_cpu_mask
1825 * disables preemption as part of its processing
1827 for_each_online_cpu(cpu) {
1828 struct per_cpu_pageset *pcp;
1830 bool has_pcps = false;
1833 pcp = per_cpu_ptr(zone->pageset, cpu);
1837 for_each_populated_zone(z) {
1838 pcp = per_cpu_ptr(z->pageset, cpu);
1839 if (pcp->pcp.count) {
1847 cpumask_set_cpu(cpu, &cpus_with_pcps);
1849 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1851 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
1855 #ifdef CONFIG_HIBERNATION
1857 void mark_free_pages(struct zone *zone)
1859 unsigned long pfn, max_zone_pfn;
1860 unsigned long flags;
1861 unsigned int order, t;
1862 struct list_head *curr;
1864 if (zone_is_empty(zone))
1867 spin_lock_irqsave(&zone->lock, flags);
1869 max_zone_pfn = zone_end_pfn(zone);
1870 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1871 if (pfn_valid(pfn)) {
1872 struct page *page = pfn_to_page(pfn);
1874 if (!swsusp_page_is_forbidden(page))
1875 swsusp_unset_page_free(page);
1878 for_each_migratetype_order(order, t) {
1879 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1882 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1883 for (i = 0; i < (1UL << order); i++)
1884 swsusp_set_page_free(pfn_to_page(pfn + i));
1887 spin_unlock_irqrestore(&zone->lock, flags);
1889 #endif /* CONFIG_PM */
1892 * Free a 0-order page
1893 * cold == true ? free a cold page : free a hot page
1895 void free_hot_cold_page(struct page *page, bool cold)
1897 struct zone *zone = page_zone(page);
1898 struct per_cpu_pages *pcp;
1899 unsigned long flags;
1900 unsigned long pfn = page_to_pfn(page);
1903 if (!free_pages_prepare(page, 0))
1906 migratetype = get_pfnblock_migratetype(page, pfn);
1907 set_freepage_migratetype(page, migratetype);
1908 local_irq_save(flags);
1909 __count_vm_event(PGFREE);
1912 * We only track unmovable, reclaimable and movable on pcp lists.
1913 * Free ISOLATE pages back to the allocator because they are being
1914 * offlined but treat RESERVE as movable pages so we can get those
1915 * areas back if necessary. Otherwise, we may have to free
1916 * excessively into the page allocator
1918 if (migratetype >= MIGRATE_PCPTYPES) {
1919 if (unlikely(is_migrate_isolate(migratetype))) {
1920 free_one_page(zone, page, pfn, 0, migratetype);
1923 migratetype = MIGRATE_MOVABLE;
1926 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1928 list_add(&page->lru, &pcp->lists[migratetype]);
1930 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1932 if (pcp->count >= pcp->high) {
1933 unsigned long batch = READ_ONCE(pcp->batch);
1934 free_pcppages_bulk(zone, batch, pcp);
1935 pcp->count -= batch;
1939 local_irq_restore(flags);
1943 * Free a list of 0-order pages
1945 void free_hot_cold_page_list(struct list_head *list, bool cold)
1947 struct page *page, *next;
1949 list_for_each_entry_safe(page, next, list, lru) {
1950 trace_mm_page_free_batched(page, cold);
1951 free_hot_cold_page(page, cold);
1956 * split_page takes a non-compound higher-order page, and splits it into
1957 * n (1<<order) sub-pages: page[0..n]
1958 * Each sub-page must be freed individually.
1960 * Note: this is probably too low level an operation for use in drivers.
1961 * Please consult with lkml before using this in your driver.
1963 void split_page(struct page *page, unsigned int order)
1968 VM_BUG_ON_PAGE(PageCompound(page), page);
1969 VM_BUG_ON_PAGE(!page_count(page), page);
1971 #ifdef CONFIG_KMEMCHECK
1973 * Split shadow pages too, because free(page[0]) would
1974 * otherwise free the whole shadow.
1976 if (kmemcheck_page_is_tracked(page))
1977 split_page(virt_to_page(page[0].shadow), order);
1980 gfp_mask = get_page_owner_gfp(page);
1981 set_page_owner(page, 0, gfp_mask);
1982 for (i = 1; i < (1 << order); i++) {
1983 set_page_refcounted(page + i);
1984 set_page_owner(page + i, 0, gfp_mask);
1987 EXPORT_SYMBOL_GPL(split_page);
1989 int __isolate_free_page(struct page *page, unsigned int order)
1991 unsigned long watermark;
1995 BUG_ON(!PageBuddy(page));
1997 zone = page_zone(page);
1998 mt = get_pageblock_migratetype(page);
2000 if (!is_migrate_isolate(mt)) {
2001 /* Obey watermarks as if the page was being allocated */
2002 watermark = low_wmark_pages(zone) + (1 << order);
2003 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2006 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2009 /* Remove page from free list */
2010 list_del(&page->lru);
2011 zone->free_area[order].nr_free--;
2012 rmv_page_order(page);
2014 set_page_owner(page, order, __GFP_MOVABLE);
2016 /* Set the pageblock if the isolated page is at least a pageblock */
2017 if (order >= pageblock_order - 1) {
2018 struct page *endpage = page + (1 << order) - 1;
2019 for (; page < endpage; page += pageblock_nr_pages) {
2020 int mt = get_pageblock_migratetype(page);
2021 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2022 set_pageblock_migratetype(page,
2028 return 1UL << order;
2032 * Similar to split_page except the page is already free. As this is only
2033 * being used for migration, the migratetype of the block also changes.
2034 * As this is called with interrupts disabled, the caller is responsible
2035 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2038 * Note: this is probably too low level an operation for use in drivers.
2039 * Please consult with lkml before using this in your driver.
2041 int split_free_page(struct page *page)
2046 order = page_order(page);
2048 nr_pages = __isolate_free_page(page, order);
2052 /* Split into individual pages */
2053 set_page_refcounted(page);
2054 split_page(page, order);
2059 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2062 struct page *buffered_rmqueue(struct zone *preferred_zone,
2063 struct zone *zone, unsigned int order,
2064 gfp_t gfp_flags, int migratetype)
2066 unsigned long flags;
2068 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2070 if (likely(order == 0)) {
2071 struct per_cpu_pages *pcp;
2072 struct list_head *list;
2074 local_irq_save(flags);
2075 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2076 list = &pcp->lists[migratetype];
2077 if (list_empty(list)) {
2078 pcp->count += rmqueue_bulk(zone, 0,
2081 if (unlikely(list_empty(list)))
2086 page = list_entry(list->prev, struct page, lru);
2088 page = list_entry(list->next, struct page, lru);
2090 list_del(&page->lru);
2093 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2095 * __GFP_NOFAIL is not to be used in new code.
2097 * All __GFP_NOFAIL callers should be fixed so that they
2098 * properly detect and handle allocation failures.
2100 * We most definitely don't want callers attempting to
2101 * allocate greater than order-1 page units with
2104 WARN_ON_ONCE(order > 1);
2106 spin_lock_irqsave(&zone->lock, flags);
2107 page = __rmqueue(zone, order, migratetype);
2108 spin_unlock(&zone->lock);
2111 __mod_zone_freepage_state(zone, -(1 << order),
2112 get_freepage_migratetype(page));
2115 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2116 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2117 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2118 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2120 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2121 zone_statistics(preferred_zone, zone, gfp_flags);
2122 local_irq_restore(flags);
2124 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2128 local_irq_restore(flags);
2132 #ifdef CONFIG_FAIL_PAGE_ALLOC
2135 struct fault_attr attr;
2137 u32 ignore_gfp_highmem;
2138 u32 ignore_gfp_wait;
2140 } fail_page_alloc = {
2141 .attr = FAULT_ATTR_INITIALIZER,
2142 .ignore_gfp_wait = 1,
2143 .ignore_gfp_highmem = 1,
2147 static int __init setup_fail_page_alloc(char *str)
2149 return setup_fault_attr(&fail_page_alloc.attr, str);
2151 __setup("fail_page_alloc=", setup_fail_page_alloc);
2153 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2155 if (order < fail_page_alloc.min_order)
2157 if (gfp_mask & __GFP_NOFAIL)
2159 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2161 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
2164 return should_fail(&fail_page_alloc.attr, 1 << order);
2167 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2169 static int __init fail_page_alloc_debugfs(void)
2171 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2174 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2175 &fail_page_alloc.attr);
2177 return PTR_ERR(dir);
2179 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2180 &fail_page_alloc.ignore_gfp_wait))
2182 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2183 &fail_page_alloc.ignore_gfp_highmem))
2185 if (!debugfs_create_u32("min-order", mode, dir,
2186 &fail_page_alloc.min_order))
2191 debugfs_remove_recursive(dir);
2196 late_initcall(fail_page_alloc_debugfs);
2198 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2200 #else /* CONFIG_FAIL_PAGE_ALLOC */
2202 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2207 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2210 * Return true if free pages are above 'mark'. This takes into account the order
2211 * of the allocation.
2213 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2214 unsigned long mark, int classzone_idx, int alloc_flags,
2217 /* free_pages may go negative - that's OK */
2222 free_pages -= (1 << order) - 1;
2223 if (alloc_flags & ALLOC_HIGH)
2225 if (alloc_flags & ALLOC_HARDER)
2228 /* If allocation can't use CMA areas don't use free CMA pages */
2229 if (!(alloc_flags & ALLOC_CMA))
2230 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
2233 if (free_pages - free_cma <= min + z->lowmem_reserve[classzone_idx])
2235 for (o = 0; o < order; o++) {
2236 /* At the next order, this order's pages become unavailable */
2237 free_pages -= z->free_area[o].nr_free << o;
2239 /* Require fewer higher order pages to be free */
2242 if (free_pages <= min)
2248 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2249 int classzone_idx, int alloc_flags)
2251 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2252 zone_page_state(z, NR_FREE_PAGES));
2255 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2256 unsigned long mark, int classzone_idx, int alloc_flags)
2258 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2260 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2261 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2263 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2269 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
2270 * skip over zones that are not allowed by the cpuset, or that have
2271 * been recently (in last second) found to be nearly full. See further
2272 * comments in mmzone.h. Reduces cache footprint of zonelist scans
2273 * that have to skip over a lot of full or unallowed zones.
2275 * If the zonelist cache is present in the passed zonelist, then
2276 * returns a pointer to the allowed node mask (either the current
2277 * tasks mems_allowed, or node_states[N_MEMORY].)
2279 * If the zonelist cache is not available for this zonelist, does
2280 * nothing and returns NULL.
2282 * If the fullzones BITMAP in the zonelist cache is stale (more than
2283 * a second since last zap'd) then we zap it out (clear its bits.)
2285 * We hold off even calling zlc_setup, until after we've checked the
2286 * first zone in the zonelist, on the theory that most allocations will
2287 * be satisfied from that first zone, so best to examine that zone as
2288 * quickly as we can.
2290 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
2292 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2293 nodemask_t *allowednodes; /* zonelist_cache approximation */
2295 zlc = zonelist->zlcache_ptr;
2299 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
2300 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2301 zlc->last_full_zap = jiffies;
2304 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
2305 &cpuset_current_mems_allowed :
2306 &node_states[N_MEMORY];
2307 return allowednodes;
2311 * Given 'z' scanning a zonelist, run a couple of quick checks to see
2312 * if it is worth looking at further for free memory:
2313 * 1) Check that the zone isn't thought to be full (doesn't have its
2314 * bit set in the zonelist_cache fullzones BITMAP).
2315 * 2) Check that the zones node (obtained from the zonelist_cache
2316 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
2317 * Return true (non-zero) if zone is worth looking at further, or
2318 * else return false (zero) if it is not.
2320 * This check -ignores- the distinction between various watermarks,
2321 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
2322 * found to be full for any variation of these watermarks, it will
2323 * be considered full for up to one second by all requests, unless
2324 * we are so low on memory on all allowed nodes that we are forced
2325 * into the second scan of the zonelist.
2327 * In the second scan we ignore this zonelist cache and exactly
2328 * apply the watermarks to all zones, even it is slower to do so.
2329 * We are low on memory in the second scan, and should leave no stone
2330 * unturned looking for a free page.
2332 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
2333 nodemask_t *allowednodes)
2335 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2336 int i; /* index of *z in zonelist zones */
2337 int n; /* node that zone *z is on */
2339 zlc = zonelist->zlcache_ptr;
2343 i = z - zonelist->_zonerefs;
2346 /* This zone is worth trying if it is allowed but not full */
2347 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
2351 * Given 'z' scanning a zonelist, set the corresponding bit in
2352 * zlc->fullzones, so that subsequent attempts to allocate a page
2353 * from that zone don't waste time re-examining it.
2355 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
2357 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2358 int i; /* index of *z in zonelist zones */
2360 zlc = zonelist->zlcache_ptr;
2364 i = z - zonelist->_zonerefs;
2366 set_bit(i, zlc->fullzones);
2370 * clear all zones full, called after direct reclaim makes progress so that
2371 * a zone that was recently full is not skipped over for up to a second
2373 static void zlc_clear_zones_full(struct zonelist *zonelist)
2375 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2377 zlc = zonelist->zlcache_ptr;
2381 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2384 static bool zone_local(struct zone *local_zone, struct zone *zone)
2386 return local_zone->node == zone->node;
2389 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2391 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2395 #else /* CONFIG_NUMA */
2397 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
2402 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
2403 nodemask_t *allowednodes)
2408 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
2412 static void zlc_clear_zones_full(struct zonelist *zonelist)
2416 static bool zone_local(struct zone *local_zone, struct zone *zone)
2421 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2426 #endif /* CONFIG_NUMA */
2428 static void reset_alloc_batches(struct zone *preferred_zone)
2430 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2433 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2434 high_wmark_pages(zone) - low_wmark_pages(zone) -
2435 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2436 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2437 } while (zone++ != preferred_zone);
2441 * get_page_from_freelist goes through the zonelist trying to allocate
2444 static struct page *
2445 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2446 const struct alloc_context *ac)
2448 struct zonelist *zonelist = ac->zonelist;
2450 struct page *page = NULL;
2452 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
2453 int zlc_active = 0; /* set if using zonelist_cache */
2454 int did_zlc_setup = 0; /* just call zlc_setup() one time */
2455 bool consider_zone_dirty = (alloc_flags & ALLOC_WMARK_LOW) &&
2456 (gfp_mask & __GFP_WRITE);
2457 int nr_fair_skipped = 0;
2458 bool zonelist_rescan;
2461 zonelist_rescan = false;
2464 * Scan zonelist, looking for a zone with enough free.
2465 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2467 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2471 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2472 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2474 if (cpusets_enabled() &&
2475 (alloc_flags & ALLOC_CPUSET) &&
2476 !cpuset_zone_allowed(zone, gfp_mask))
2479 * Distribute pages in proportion to the individual
2480 * zone size to ensure fair page aging. The zone a
2481 * page was allocated in should have no effect on the
2482 * time the page has in memory before being reclaimed.
2484 if (alloc_flags & ALLOC_FAIR) {
2485 if (!zone_local(ac->preferred_zone, zone))
2487 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2493 * When allocating a page cache page for writing, we
2494 * want to get it from a zone that is within its dirty
2495 * limit, such that no single zone holds more than its
2496 * proportional share of globally allowed dirty pages.
2497 * The dirty limits take into account the zone's
2498 * lowmem reserves and high watermark so that kswapd
2499 * should be able to balance it without having to
2500 * write pages from its LRU list.
2502 * This may look like it could increase pressure on
2503 * lower zones by failing allocations in higher zones
2504 * before they are full. But the pages that do spill
2505 * over are limited as the lower zones are protected
2506 * by this very same mechanism. It should not become
2507 * a practical burden to them.
2509 * XXX: For now, allow allocations to potentially
2510 * exceed the per-zone dirty limit in the slowpath
2511 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2512 * which is important when on a NUMA setup the allowed
2513 * zones are together not big enough to reach the
2514 * global limit. The proper fix for these situations
2515 * will require awareness of zones in the
2516 * dirty-throttling and the flusher threads.
2518 if (consider_zone_dirty && !zone_dirty_ok(zone))
2521 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2522 if (!zone_watermark_ok(zone, order, mark,
2523 ac->classzone_idx, alloc_flags)) {
2526 /* Checked here to keep the fast path fast */
2527 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2528 if (alloc_flags & ALLOC_NO_WATERMARKS)
2531 if (IS_ENABLED(CONFIG_NUMA) &&
2532 !did_zlc_setup && nr_online_nodes > 1) {
2534 * we do zlc_setup if there are multiple nodes
2535 * and before considering the first zone allowed
2538 allowednodes = zlc_setup(zonelist, alloc_flags);
2543 if (zone_reclaim_mode == 0 ||
2544 !zone_allows_reclaim(ac->preferred_zone, zone))
2545 goto this_zone_full;
2548 * As we may have just activated ZLC, check if the first
2549 * eligible zone has failed zone_reclaim recently.
2551 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2552 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2555 ret = zone_reclaim(zone, gfp_mask, order);
2557 case ZONE_RECLAIM_NOSCAN:
2560 case ZONE_RECLAIM_FULL:
2561 /* scanned but unreclaimable */
2564 /* did we reclaim enough */
2565 if (zone_watermark_ok(zone, order, mark,
2566 ac->classzone_idx, alloc_flags))
2570 * Failed to reclaim enough to meet watermark.
2571 * Only mark the zone full if checking the min
2572 * watermark or if we failed to reclaim just
2573 * 1<<order pages or else the page allocator
2574 * fastpath will prematurely mark zones full
2575 * when the watermark is between the low and
2578 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2579 ret == ZONE_RECLAIM_SOME)
2580 goto this_zone_full;
2587 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2588 gfp_mask, ac->migratetype);
2590 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2595 if (IS_ENABLED(CONFIG_NUMA) && zlc_active)
2596 zlc_mark_zone_full(zonelist, z);
2600 * The first pass makes sure allocations are spread fairly within the
2601 * local node. However, the local node might have free pages left
2602 * after the fairness batches are exhausted, and remote zones haven't
2603 * even been considered yet. Try once more without fairness, and
2604 * include remote zones now, before entering the slowpath and waking
2605 * kswapd: prefer spilling to a remote zone over swapping locally.
2607 if (alloc_flags & ALLOC_FAIR) {
2608 alloc_flags &= ~ALLOC_FAIR;
2609 if (nr_fair_skipped) {
2610 zonelist_rescan = true;
2611 reset_alloc_batches(ac->preferred_zone);
2613 if (nr_online_nodes > 1)
2614 zonelist_rescan = true;
2617 if (unlikely(IS_ENABLED(CONFIG_NUMA) && zlc_active)) {
2618 /* Disable zlc cache for second zonelist scan */
2620 zonelist_rescan = true;
2623 if (zonelist_rescan)
2630 * Large machines with many possible nodes should not always dump per-node
2631 * meminfo in irq context.
2633 static inline bool should_suppress_show_mem(void)
2638 ret = in_interrupt();
2643 static DEFINE_RATELIMIT_STATE(nopage_rs,
2644 DEFAULT_RATELIMIT_INTERVAL,
2645 DEFAULT_RATELIMIT_BURST);
2647 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2649 unsigned int filter = SHOW_MEM_FILTER_NODES;
2651 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2652 debug_guardpage_minorder() > 0)
2656 * This documents exceptions given to allocations in certain
2657 * contexts that are allowed to allocate outside current's set
2660 if (!(gfp_mask & __GFP_NOMEMALLOC))
2661 if (test_thread_flag(TIF_MEMDIE) ||
2662 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2663 filter &= ~SHOW_MEM_FILTER_NODES;
2664 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2665 filter &= ~SHOW_MEM_FILTER_NODES;
2668 struct va_format vaf;
2671 va_start(args, fmt);
2676 pr_warn("%pV", &vaf);
2681 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2682 current->comm, order, gfp_mask);
2685 if (!should_suppress_show_mem())
2689 static inline struct page *
2690 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2691 const struct alloc_context *ac, unsigned long *did_some_progress)
2695 *did_some_progress = 0;
2698 * Acquire the oom lock. If that fails, somebody else is
2699 * making progress for us.
2701 if (!mutex_trylock(&oom_lock)) {
2702 *did_some_progress = 1;
2703 schedule_timeout_uninterruptible(1);
2708 * Go through the zonelist yet one more time, keep very high watermark
2709 * here, this is only to catch a parallel oom killing, we must fail if
2710 * we're still under heavy pressure.
2712 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2713 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2717 if (!(gfp_mask & __GFP_NOFAIL)) {
2718 /* Coredumps can quickly deplete all memory reserves */
2719 if (current->flags & PF_DUMPCORE)
2721 /* The OOM killer will not help higher order allocs */
2722 if (order > PAGE_ALLOC_COSTLY_ORDER)
2724 /* The OOM killer does not needlessly kill tasks for lowmem */
2725 if (ac->high_zoneidx < ZONE_NORMAL)
2727 /* The OOM killer does not compensate for IO-less reclaim */
2728 if (!(gfp_mask & __GFP_FS)) {
2730 * XXX: Page reclaim didn't yield anything,
2731 * and the OOM killer can't be invoked, but
2732 * keep looping as per tradition.
2734 *did_some_progress = 1;
2737 if (pm_suspended_storage())
2739 /* The OOM killer may not free memory on a specific node */
2740 if (gfp_mask & __GFP_THISNODE)
2743 /* Exhausted what can be done so it's blamo time */
2744 if (out_of_memory(ac->zonelist, gfp_mask, order, ac->nodemask, false)
2745 || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL))
2746 *did_some_progress = 1;
2748 mutex_unlock(&oom_lock);
2752 #ifdef CONFIG_COMPACTION
2753 /* Try memory compaction for high-order allocations before reclaim */
2754 static struct page *
2755 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2756 int alloc_flags, const struct alloc_context *ac,
2757 enum migrate_mode mode, int *contended_compaction,
2758 bool *deferred_compaction)
2760 unsigned long compact_result;
2766 current->flags |= PF_MEMALLOC;
2767 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2768 mode, contended_compaction);
2769 current->flags &= ~PF_MEMALLOC;
2771 switch (compact_result) {
2772 case COMPACT_DEFERRED:
2773 *deferred_compaction = true;
2775 case COMPACT_SKIPPED:
2782 * At least in one zone compaction wasn't deferred or skipped, so let's
2783 * count a compaction stall
2785 count_vm_event(COMPACTSTALL);
2787 page = get_page_from_freelist(gfp_mask, order,
2788 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2791 struct zone *zone = page_zone(page);
2793 zone->compact_blockskip_flush = false;
2794 compaction_defer_reset(zone, order, true);
2795 count_vm_event(COMPACTSUCCESS);
2800 * It's bad if compaction run occurs and fails. The most likely reason
2801 * is that pages exist, but not enough to satisfy watermarks.
2803 count_vm_event(COMPACTFAIL);
2810 static inline struct page *
2811 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2812 int alloc_flags, const struct alloc_context *ac,
2813 enum migrate_mode mode, int *contended_compaction,
2814 bool *deferred_compaction)
2818 #endif /* CONFIG_COMPACTION */
2820 /* Perform direct synchronous page reclaim */
2822 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2823 const struct alloc_context *ac)
2825 struct reclaim_state reclaim_state;
2830 /* We now go into synchronous reclaim */
2831 cpuset_memory_pressure_bump();
2832 current->flags |= PF_MEMALLOC;
2833 lockdep_set_current_reclaim_state(gfp_mask);
2834 reclaim_state.reclaimed_slab = 0;
2835 current->reclaim_state = &reclaim_state;
2837 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2840 current->reclaim_state = NULL;
2841 lockdep_clear_current_reclaim_state();
2842 current->flags &= ~PF_MEMALLOC;
2849 /* The really slow allocator path where we enter direct reclaim */
2850 static inline struct page *
2851 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2852 int alloc_flags, const struct alloc_context *ac,
2853 unsigned long *did_some_progress)
2855 struct page *page = NULL;
2856 bool drained = false;
2858 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2859 if (unlikely(!(*did_some_progress)))
2862 /* After successful reclaim, reconsider all zones for allocation */
2863 if (IS_ENABLED(CONFIG_NUMA))
2864 zlc_clear_zones_full(ac->zonelist);
2867 page = get_page_from_freelist(gfp_mask, order,
2868 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2871 * If an allocation failed after direct reclaim, it could be because
2872 * pages are pinned on the per-cpu lists. Drain them and try again
2874 if (!page && !drained) {
2875 drain_all_pages(NULL);
2884 * This is called in the allocator slow-path if the allocation request is of
2885 * sufficient urgency to ignore watermarks and take other desperate measures
2887 static inline struct page *
2888 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2889 const struct alloc_context *ac)
2894 page = get_page_from_freelist(gfp_mask, order,
2895 ALLOC_NO_WATERMARKS, ac);
2897 if (!page && gfp_mask & __GFP_NOFAIL)
2898 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC,
2900 } while (!page && (gfp_mask & __GFP_NOFAIL));
2905 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2910 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2911 ac->high_zoneidx, ac->nodemask)
2912 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2916 gfp_to_alloc_flags(gfp_t gfp_mask)
2918 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2919 const bool atomic = !(gfp_mask & (__GFP_WAIT | __GFP_NO_KSWAPD));
2921 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2922 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2925 * The caller may dip into page reserves a bit more if the caller
2926 * cannot run direct reclaim, or if the caller has realtime scheduling
2927 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2928 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2930 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2934 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2935 * if it can't schedule.
2937 if (!(gfp_mask & __GFP_NOMEMALLOC))
2938 alloc_flags |= ALLOC_HARDER;
2940 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2941 * comment for __cpuset_node_allowed().
2943 alloc_flags &= ~ALLOC_CPUSET;
2944 } else if (unlikely(rt_task(current)) && !in_interrupt())
2945 alloc_flags |= ALLOC_HARDER;
2947 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2948 if (gfp_mask & __GFP_MEMALLOC)
2949 alloc_flags |= ALLOC_NO_WATERMARKS;
2950 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2951 alloc_flags |= ALLOC_NO_WATERMARKS;
2952 else if (!in_interrupt() &&
2953 ((current->flags & PF_MEMALLOC) ||
2954 unlikely(test_thread_flag(TIF_MEMDIE))))
2955 alloc_flags |= ALLOC_NO_WATERMARKS;
2958 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2959 alloc_flags |= ALLOC_CMA;
2964 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2966 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2969 static inline struct page *
2970 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2971 struct alloc_context *ac)
2973 const gfp_t wait = gfp_mask & __GFP_WAIT;
2974 struct page *page = NULL;
2976 unsigned long pages_reclaimed = 0;
2977 unsigned long did_some_progress;
2978 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2979 bool deferred_compaction = false;
2980 int contended_compaction = COMPACT_CONTENDED_NONE;
2983 * In the slowpath, we sanity check order to avoid ever trying to
2984 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2985 * be using allocators in order of preference for an area that is
2988 if (order >= MAX_ORDER) {
2989 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2994 * If this allocation cannot block and it is for a specific node, then
2995 * fail early. There's no need to wakeup kswapd or retry for a
2996 * speculative node-specific allocation.
2998 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !wait)
3002 if (!(gfp_mask & __GFP_NO_KSWAPD))
3003 wake_all_kswapds(order, ac);
3006 * OK, we're below the kswapd watermark and have kicked background
3007 * reclaim. Now things get more complex, so set up alloc_flags according
3008 * to how we want to proceed.
3010 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3013 * Find the true preferred zone if the allocation is unconstrained by
3016 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3017 struct zoneref *preferred_zoneref;
3018 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3019 ac->high_zoneidx, NULL, &ac->preferred_zone);
3020 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3023 /* This is the last chance, in general, before the goto nopage. */
3024 page = get_page_from_freelist(gfp_mask, order,
3025 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3029 /* Allocate without watermarks if the context allows */
3030 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3032 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3033 * the allocation is high priority and these type of
3034 * allocations are system rather than user orientated
3036 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3038 page = __alloc_pages_high_priority(gfp_mask, order, ac);
3045 /* Atomic allocations - we can't balance anything */
3048 * All existing users of the deprecated __GFP_NOFAIL are
3049 * blockable, so warn of any new users that actually allow this
3050 * type of allocation to fail.
3052 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3056 /* Avoid recursion of direct reclaim */
3057 if (current->flags & PF_MEMALLOC)
3060 /* Avoid allocations with no watermarks from looping endlessly */
3061 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3065 * Try direct compaction. The first pass is asynchronous. Subsequent
3066 * attempts after direct reclaim are synchronous
3068 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3070 &contended_compaction,
3071 &deferred_compaction);
3075 /* Checks for THP-specific high-order allocations */
3076 if ((gfp_mask & GFP_TRANSHUGE) == GFP_TRANSHUGE) {
3078 * If compaction is deferred for high-order allocations, it is
3079 * because sync compaction recently failed. If this is the case
3080 * and the caller requested a THP allocation, we do not want
3081 * to heavily disrupt the system, so we fail the allocation
3082 * instead of entering direct reclaim.
3084 if (deferred_compaction)
3088 * In all zones where compaction was attempted (and not
3089 * deferred or skipped), lock contention has been detected.
3090 * For THP allocation we do not want to disrupt the others
3091 * so we fallback to base pages instead.
3093 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3097 * If compaction was aborted due to need_resched(), we do not
3098 * want to further increase allocation latency, unless it is
3099 * khugepaged trying to collapse.
3101 if (contended_compaction == COMPACT_CONTENDED_SCHED
3102 && !(current->flags & PF_KTHREAD))
3107 * It can become very expensive to allocate transparent hugepages at
3108 * fault, so use asynchronous memory compaction for THP unless it is
3109 * khugepaged trying to collapse.
3111 if ((gfp_mask & GFP_TRANSHUGE) != GFP_TRANSHUGE ||
3112 (current->flags & PF_KTHREAD))
3113 migration_mode = MIGRATE_SYNC_LIGHT;
3115 /* Try direct reclaim and then allocating */
3116 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3117 &did_some_progress);
3121 /* Do not loop if specifically requested */
3122 if (gfp_mask & __GFP_NORETRY)
3125 /* Keep reclaiming pages as long as there is reasonable progress */
3126 pages_reclaimed += did_some_progress;
3127 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3128 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3129 /* Wait for some write requests to complete then retry */
3130 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3134 /* Reclaim has failed us, start killing things */
3135 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3139 /* Retry as long as the OOM killer is making progress */
3140 if (did_some_progress)
3145 * High-order allocations do not necessarily loop after
3146 * direct reclaim and reclaim/compaction depends on compaction
3147 * being called after reclaim so call directly if necessary
3149 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3151 &contended_compaction,
3152 &deferred_compaction);
3156 warn_alloc_failed(gfp_mask, order, NULL);
3162 * This is the 'heart' of the zoned buddy allocator.
3165 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3166 struct zonelist *zonelist, nodemask_t *nodemask)
3168 struct zoneref *preferred_zoneref;
3169 struct page *page = NULL;
3170 unsigned int cpuset_mems_cookie;
3171 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3172 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3173 struct alloc_context ac = {
3174 .high_zoneidx = gfp_zone(gfp_mask),
3175 .nodemask = nodemask,
3176 .migratetype = gfpflags_to_migratetype(gfp_mask),
3179 gfp_mask &= gfp_allowed_mask;
3181 lockdep_trace_alloc(gfp_mask);
3183 might_sleep_if(gfp_mask & __GFP_WAIT);
3185 if (should_fail_alloc_page(gfp_mask, order))
3189 * Check the zones suitable for the gfp_mask contain at least one
3190 * valid zone. It's possible to have an empty zonelist as a result
3191 * of __GFP_THISNODE and a memoryless node
3193 if (unlikely(!zonelist->_zonerefs->zone))
3196 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3197 alloc_flags |= ALLOC_CMA;
3200 cpuset_mems_cookie = read_mems_allowed_begin();
3202 /* We set it here, as __alloc_pages_slowpath might have changed it */
3203 ac.zonelist = zonelist;
3204 /* The preferred zone is used for statistics later */
3205 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3206 ac.nodemask ? : &cpuset_current_mems_allowed,
3207 &ac.preferred_zone);
3208 if (!ac.preferred_zone)
3210 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3212 /* First allocation attempt */
3213 alloc_mask = gfp_mask|__GFP_HARDWALL;
3214 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3215 if (unlikely(!page)) {
3217 * Runtime PM, block IO and its error handling path
3218 * can deadlock because I/O on the device might not
3221 alloc_mask = memalloc_noio_flags(gfp_mask);
3223 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3226 if (kmemcheck_enabled && page)
3227 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3229 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3233 * When updating a task's mems_allowed, it is possible to race with
3234 * parallel threads in such a way that an allocation can fail while
3235 * the mask is being updated. If a page allocation is about to fail,
3236 * check if the cpuset changed during allocation and if so, retry.
3238 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3243 EXPORT_SYMBOL(__alloc_pages_nodemask);
3246 * Common helper functions.
3248 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3253 * __get_free_pages() returns a 32-bit address, which cannot represent
3256 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3258 page = alloc_pages(gfp_mask, order);
3261 return (unsigned long) page_address(page);
3263 EXPORT_SYMBOL(__get_free_pages);
3265 unsigned long get_zeroed_page(gfp_t gfp_mask)
3267 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3269 EXPORT_SYMBOL(get_zeroed_page);
3271 void __free_pages(struct page *page, unsigned int order)
3273 if (put_page_testzero(page)) {
3275 free_hot_cold_page(page, false);
3277 __free_pages_ok(page, order);
3281 EXPORT_SYMBOL(__free_pages);
3283 void free_pages(unsigned long addr, unsigned int order)
3286 VM_BUG_ON(!virt_addr_valid((void *)addr));
3287 __free_pages(virt_to_page((void *)addr), order);
3291 EXPORT_SYMBOL(free_pages);
3295 * An arbitrary-length arbitrary-offset area of memory which resides
3296 * within a 0 or higher order page. Multiple fragments within that page
3297 * are individually refcounted, in the page's reference counter.
3299 * The page_frag functions below provide a simple allocation framework for
3300 * page fragments. This is used by the network stack and network device
3301 * drivers to provide a backing region of memory for use as either an
3302 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3304 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3307 struct page *page = NULL;
3308 gfp_t gfp = gfp_mask;
3310 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3311 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3313 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3314 PAGE_FRAG_CACHE_MAX_ORDER);
3315 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3317 if (unlikely(!page))
3318 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3320 nc->va = page ? page_address(page) : NULL;
3325 void *__alloc_page_frag(struct page_frag_cache *nc,
3326 unsigned int fragsz, gfp_t gfp_mask)
3328 unsigned int size = PAGE_SIZE;
3332 if (unlikely(!nc->va)) {
3334 page = __page_frag_refill(nc, gfp_mask);
3338 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3339 /* if size can vary use size else just use PAGE_SIZE */
3342 /* Even if we own the page, we do not use atomic_set().
3343 * This would break get_page_unless_zero() users.
3345 atomic_add(size - 1, &page->_count);
3347 /* reset page count bias and offset to start of new frag */
3348 nc->pfmemalloc = page->pfmemalloc;
3349 nc->pagecnt_bias = size;
3353 offset = nc->offset - fragsz;
3354 if (unlikely(offset < 0)) {
3355 page = virt_to_page(nc->va);
3357 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3360 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3361 /* if size can vary use size else just use PAGE_SIZE */
3364 /* OK, page count is 0, we can safely set it */
3365 atomic_set(&page->_count, size);
3367 /* reset page count bias and offset to start of new frag */
3368 nc->pagecnt_bias = size;
3369 offset = size - fragsz;
3373 nc->offset = offset;
3375 return nc->va + offset;
3377 EXPORT_SYMBOL(__alloc_page_frag);
3380 * Frees a page fragment allocated out of either a compound or order 0 page.
3382 void __free_page_frag(void *addr)
3384 struct page *page = virt_to_head_page(addr);
3386 if (unlikely(put_page_testzero(page)))
3387 __free_pages_ok(page, compound_order(page));
3389 EXPORT_SYMBOL(__free_page_frag);
3392 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3393 * of the current memory cgroup.
3395 * It should be used when the caller would like to use kmalloc, but since the
3396 * allocation is large, it has to fall back to the page allocator.
3398 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3401 struct mem_cgroup *memcg = NULL;
3403 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
3405 page = alloc_pages(gfp_mask, order);
3406 memcg_kmem_commit_charge(page, memcg, order);
3410 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3413 struct mem_cgroup *memcg = NULL;
3415 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
3417 page = alloc_pages_node(nid, gfp_mask, order);
3418 memcg_kmem_commit_charge(page, memcg, order);
3423 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3426 void __free_kmem_pages(struct page *page, unsigned int order)
3428 memcg_kmem_uncharge_pages(page, order);
3429 __free_pages(page, order);
3432 void free_kmem_pages(unsigned long addr, unsigned int order)
3435 VM_BUG_ON(!virt_addr_valid((void *)addr));
3436 __free_kmem_pages(virt_to_page((void *)addr), order);
3440 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
3443 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3444 unsigned long used = addr + PAGE_ALIGN(size);
3446 split_page(virt_to_page((void *)addr), order);
3447 while (used < alloc_end) {
3452 return (void *)addr;
3456 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3457 * @size: the number of bytes to allocate
3458 * @gfp_mask: GFP flags for the allocation
3460 * This function is similar to alloc_pages(), except that it allocates the
3461 * minimum number of pages to satisfy the request. alloc_pages() can only
3462 * allocate memory in power-of-two pages.
3464 * This function is also limited by MAX_ORDER.
3466 * Memory allocated by this function must be released by free_pages_exact().
3468 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3470 unsigned int order = get_order(size);
3473 addr = __get_free_pages(gfp_mask, order);
3474 return make_alloc_exact(addr, order, size);
3476 EXPORT_SYMBOL(alloc_pages_exact);
3479 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3481 * @nid: the preferred node ID where memory should be allocated
3482 * @size: the number of bytes to allocate
3483 * @gfp_mask: GFP flags for the allocation
3485 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3487 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3490 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3492 unsigned order = get_order(size);
3493 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3496 return make_alloc_exact((unsigned long)page_address(p), order, size);
3500 * free_pages_exact - release memory allocated via alloc_pages_exact()
3501 * @virt: the value returned by alloc_pages_exact.
3502 * @size: size of allocation, same value as passed to alloc_pages_exact().
3504 * Release the memory allocated by a previous call to alloc_pages_exact.
3506 void free_pages_exact(void *virt, size_t size)
3508 unsigned long addr = (unsigned long)virt;
3509 unsigned long end = addr + PAGE_ALIGN(size);
3511 while (addr < end) {
3516 EXPORT_SYMBOL(free_pages_exact);
3519 * nr_free_zone_pages - count number of pages beyond high watermark
3520 * @offset: The zone index of the highest zone
3522 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3523 * high watermark within all zones at or below a given zone index. For each
3524 * zone, the number of pages is calculated as:
3525 * managed_pages - high_pages
3527 static unsigned long nr_free_zone_pages(int offset)
3532 /* Just pick one node, since fallback list is circular */
3533 unsigned long sum = 0;
3535 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3537 for_each_zone_zonelist(zone, z, zonelist, offset) {
3538 unsigned long size = zone->managed_pages;
3539 unsigned long high = high_wmark_pages(zone);
3548 * nr_free_buffer_pages - count number of pages beyond high watermark
3550 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3551 * watermark within ZONE_DMA and ZONE_NORMAL.
3553 unsigned long nr_free_buffer_pages(void)
3555 return nr_free_zone_pages(gfp_zone(GFP_USER));
3557 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3560 * nr_free_pagecache_pages - count number of pages beyond high watermark
3562 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3563 * high watermark within all zones.
3565 unsigned long nr_free_pagecache_pages(void)
3567 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3570 static inline void show_node(struct zone *zone)
3572 if (IS_ENABLED(CONFIG_NUMA))
3573 printk("Node %d ", zone_to_nid(zone));
3576 void si_meminfo(struct sysinfo *val)
3578 val->totalram = totalram_pages;
3579 val->sharedram = global_page_state(NR_SHMEM);
3580 val->freeram = global_page_state(NR_FREE_PAGES);
3581 val->bufferram = nr_blockdev_pages();
3582 val->totalhigh = totalhigh_pages;
3583 val->freehigh = nr_free_highpages();
3584 val->mem_unit = PAGE_SIZE;
3587 EXPORT_SYMBOL(si_meminfo);
3590 void si_meminfo_node(struct sysinfo *val, int nid)
3592 int zone_type; /* needs to be signed */
3593 unsigned long managed_pages = 0;
3594 pg_data_t *pgdat = NODE_DATA(nid);
3596 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3597 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3598 val->totalram = managed_pages;
3599 val->sharedram = node_page_state(nid, NR_SHMEM);
3600 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3601 #ifdef CONFIG_HIGHMEM
3602 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3603 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3609 val->mem_unit = PAGE_SIZE;
3614 * Determine whether the node should be displayed or not, depending on whether
3615 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3617 bool skip_free_areas_node(unsigned int flags, int nid)
3620 unsigned int cpuset_mems_cookie;
3622 if (!(flags & SHOW_MEM_FILTER_NODES))
3626 cpuset_mems_cookie = read_mems_allowed_begin();
3627 ret = !node_isset(nid, cpuset_current_mems_allowed);
3628 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3633 #define K(x) ((x) << (PAGE_SHIFT-10))
3635 static void show_migration_types(unsigned char type)
3637 static const char types[MIGRATE_TYPES] = {
3638 [MIGRATE_UNMOVABLE] = 'U',
3639 [MIGRATE_RECLAIMABLE] = 'E',
3640 [MIGRATE_MOVABLE] = 'M',
3641 [MIGRATE_RESERVE] = 'R',
3643 [MIGRATE_CMA] = 'C',
3645 #ifdef CONFIG_MEMORY_ISOLATION
3646 [MIGRATE_ISOLATE] = 'I',
3649 char tmp[MIGRATE_TYPES + 1];
3653 for (i = 0; i < MIGRATE_TYPES; i++) {
3654 if (type & (1 << i))
3659 printk("(%s) ", tmp);
3663 * Show free area list (used inside shift_scroll-lock stuff)
3664 * We also calculate the percentage fragmentation. We do this by counting the
3665 * memory on each free list with the exception of the first item on the list.
3668 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3671 void show_free_areas(unsigned int filter)
3673 unsigned long free_pcp = 0;
3677 for_each_populated_zone(zone) {
3678 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3681 for_each_online_cpu(cpu)
3682 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3685 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3686 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3687 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3688 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3689 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3690 " free:%lu free_pcp:%lu free_cma:%lu\n",
3691 global_page_state(NR_ACTIVE_ANON),
3692 global_page_state(NR_INACTIVE_ANON),
3693 global_page_state(NR_ISOLATED_ANON),
3694 global_page_state(NR_ACTIVE_FILE),
3695 global_page_state(NR_INACTIVE_FILE),
3696 global_page_state(NR_ISOLATED_FILE),
3697 global_page_state(NR_UNEVICTABLE),
3698 global_page_state(NR_FILE_DIRTY),
3699 global_page_state(NR_WRITEBACK),
3700 global_page_state(NR_UNSTABLE_NFS),
3701 global_page_state(NR_SLAB_RECLAIMABLE),
3702 global_page_state(NR_SLAB_UNRECLAIMABLE),
3703 global_page_state(NR_FILE_MAPPED),
3704 global_page_state(NR_SHMEM),
3705 global_page_state(NR_PAGETABLE),
3706 global_page_state(NR_BOUNCE),
3707 global_page_state(NR_FREE_PAGES),
3709 global_page_state(NR_FREE_CMA_PAGES));
3711 for_each_populated_zone(zone) {
3714 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3718 for_each_online_cpu(cpu)
3719 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3727 " active_anon:%lukB"
3728 " inactive_anon:%lukB"
3729 " active_file:%lukB"
3730 " inactive_file:%lukB"
3731 " unevictable:%lukB"
3732 " isolated(anon):%lukB"
3733 " isolated(file):%lukB"
3741 " slab_reclaimable:%lukB"
3742 " slab_unreclaimable:%lukB"
3743 " kernel_stack:%lukB"
3750 " writeback_tmp:%lukB"
3751 " pages_scanned:%lu"
3752 " all_unreclaimable? %s"
3755 K(zone_page_state(zone, NR_FREE_PAGES)),
3756 K(min_wmark_pages(zone)),
3757 K(low_wmark_pages(zone)),
3758 K(high_wmark_pages(zone)),
3759 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3760 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3761 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3762 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3763 K(zone_page_state(zone, NR_UNEVICTABLE)),
3764 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3765 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3766 K(zone->present_pages),
3767 K(zone->managed_pages),
3768 K(zone_page_state(zone, NR_MLOCK)),
3769 K(zone_page_state(zone, NR_FILE_DIRTY)),
3770 K(zone_page_state(zone, NR_WRITEBACK)),
3771 K(zone_page_state(zone, NR_FILE_MAPPED)),
3772 K(zone_page_state(zone, NR_SHMEM)),
3773 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3774 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3775 zone_page_state(zone, NR_KERNEL_STACK) *
3777 K(zone_page_state(zone, NR_PAGETABLE)),
3778 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3779 K(zone_page_state(zone, NR_BOUNCE)),
3781 K(this_cpu_read(zone->pageset->pcp.count)),
3782 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3783 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3784 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3785 (!zone_reclaimable(zone) ? "yes" : "no")
3787 printk("lowmem_reserve[]:");
3788 for (i = 0; i < MAX_NR_ZONES; i++)
3789 printk(" %ld", zone->lowmem_reserve[i]);
3793 for_each_populated_zone(zone) {
3794 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3795 unsigned char types[MAX_ORDER];
3797 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3800 printk("%s: ", zone->name);
3802 spin_lock_irqsave(&zone->lock, flags);
3803 for (order = 0; order < MAX_ORDER; order++) {
3804 struct free_area *area = &zone->free_area[order];
3807 nr[order] = area->nr_free;
3808 total += nr[order] << order;
3811 for (type = 0; type < MIGRATE_TYPES; type++) {
3812 if (!list_empty(&area->free_list[type]))
3813 types[order] |= 1 << type;
3816 spin_unlock_irqrestore(&zone->lock, flags);
3817 for (order = 0; order < MAX_ORDER; order++) {
3818 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3820 show_migration_types(types[order]);
3822 printk("= %lukB\n", K(total));
3825 hugetlb_show_meminfo();
3827 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3829 show_swap_cache_info();
3832 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3834 zoneref->zone = zone;
3835 zoneref->zone_idx = zone_idx(zone);
3839 * Builds allocation fallback zone lists.
3841 * Add all populated zones of a node to the zonelist.
3843 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3847 enum zone_type zone_type = MAX_NR_ZONES;
3851 zone = pgdat->node_zones + zone_type;
3852 if (populated_zone(zone)) {
3853 zoneref_set_zone(zone,
3854 &zonelist->_zonerefs[nr_zones++]);
3855 check_highest_zone(zone_type);
3857 } while (zone_type);
3865 * 0 = automatic detection of better ordering.
3866 * 1 = order by ([node] distance, -zonetype)
3867 * 2 = order by (-zonetype, [node] distance)
3869 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3870 * the same zonelist. So only NUMA can configure this param.
3872 #define ZONELIST_ORDER_DEFAULT 0
3873 #define ZONELIST_ORDER_NODE 1
3874 #define ZONELIST_ORDER_ZONE 2
3876 /* zonelist order in the kernel.
3877 * set_zonelist_order() will set this to NODE or ZONE.
3879 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3880 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3884 /* The value user specified ....changed by config */
3885 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3886 /* string for sysctl */
3887 #define NUMA_ZONELIST_ORDER_LEN 16
3888 char numa_zonelist_order[16] = "default";
3891 * interface for configure zonelist ordering.
3892 * command line option "numa_zonelist_order"
3893 * = "[dD]efault - default, automatic configuration.
3894 * = "[nN]ode - order by node locality, then by zone within node
3895 * = "[zZ]one - order by zone, then by locality within zone
3898 static int __parse_numa_zonelist_order(char *s)
3900 if (*s == 'd' || *s == 'D') {
3901 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3902 } else if (*s == 'n' || *s == 'N') {
3903 user_zonelist_order = ZONELIST_ORDER_NODE;
3904 } else if (*s == 'z' || *s == 'Z') {
3905 user_zonelist_order = ZONELIST_ORDER_ZONE;
3908 "Ignoring invalid numa_zonelist_order value: "
3915 static __init int setup_numa_zonelist_order(char *s)
3922 ret = __parse_numa_zonelist_order(s);
3924 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3928 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3931 * sysctl handler for numa_zonelist_order
3933 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3934 void __user *buffer, size_t *length,
3937 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3939 static DEFINE_MUTEX(zl_order_mutex);
3941 mutex_lock(&zl_order_mutex);
3943 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3947 strcpy(saved_string, (char *)table->data);
3949 ret = proc_dostring(table, write, buffer, length, ppos);
3953 int oldval = user_zonelist_order;
3955 ret = __parse_numa_zonelist_order((char *)table->data);
3958 * bogus value. restore saved string
3960 strncpy((char *)table->data, saved_string,
3961 NUMA_ZONELIST_ORDER_LEN);
3962 user_zonelist_order = oldval;
3963 } else if (oldval != user_zonelist_order) {
3964 mutex_lock(&zonelists_mutex);
3965 build_all_zonelists(NULL, NULL);
3966 mutex_unlock(&zonelists_mutex);
3970 mutex_unlock(&zl_order_mutex);
3975 #define MAX_NODE_LOAD (nr_online_nodes)
3976 static int node_load[MAX_NUMNODES];
3979 * find_next_best_node - find the next node that should appear in a given node's fallback list
3980 * @node: node whose fallback list we're appending
3981 * @used_node_mask: nodemask_t of already used nodes
3983 * We use a number of factors to determine which is the next node that should
3984 * appear on a given node's fallback list. The node should not have appeared
3985 * already in @node's fallback list, and it should be the next closest node
3986 * according to the distance array (which contains arbitrary distance values
3987 * from each node to each node in the system), and should also prefer nodes
3988 * with no CPUs, since presumably they'll have very little allocation pressure
3989 * on them otherwise.
3990 * It returns -1 if no node is found.
3992 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3995 int min_val = INT_MAX;
3996 int best_node = NUMA_NO_NODE;
3997 const struct cpumask *tmp = cpumask_of_node(0);
3999 /* Use the local node if we haven't already */
4000 if (!node_isset(node, *used_node_mask)) {
4001 node_set(node, *used_node_mask);
4005 for_each_node_state(n, N_MEMORY) {
4007 /* Don't want a node to appear more than once */
4008 if (node_isset(n, *used_node_mask))
4011 /* Use the distance array to find the distance */
4012 val = node_distance(node, n);
4014 /* Penalize nodes under us ("prefer the next node") */
4017 /* Give preference to headless and unused nodes */
4018 tmp = cpumask_of_node(n);
4019 if (!cpumask_empty(tmp))
4020 val += PENALTY_FOR_NODE_WITH_CPUS;
4022 /* Slight preference for less loaded node */
4023 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4024 val += node_load[n];
4026 if (val < min_val) {
4033 node_set(best_node, *used_node_mask);
4040 * Build zonelists ordered by node and zones within node.
4041 * This results in maximum locality--normal zone overflows into local
4042 * DMA zone, if any--but risks exhausting DMA zone.
4044 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4047 struct zonelist *zonelist;
4049 zonelist = &pgdat->node_zonelists[0];
4050 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4052 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4053 zonelist->_zonerefs[j].zone = NULL;
4054 zonelist->_zonerefs[j].zone_idx = 0;
4058 * Build gfp_thisnode zonelists
4060 static void build_thisnode_zonelists(pg_data_t *pgdat)
4063 struct zonelist *zonelist;
4065 zonelist = &pgdat->node_zonelists[1];
4066 j = build_zonelists_node(pgdat, zonelist, 0);
4067 zonelist->_zonerefs[j].zone = NULL;
4068 zonelist->_zonerefs[j].zone_idx = 0;
4072 * Build zonelists ordered by zone and nodes within zones.
4073 * This results in conserving DMA zone[s] until all Normal memory is
4074 * exhausted, but results in overflowing to remote node while memory
4075 * may still exist in local DMA zone.
4077 static int node_order[MAX_NUMNODES];
4079 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4082 int zone_type; /* needs to be signed */
4084 struct zonelist *zonelist;
4086 zonelist = &pgdat->node_zonelists[0];
4088 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4089 for (j = 0; j < nr_nodes; j++) {
4090 node = node_order[j];
4091 z = &NODE_DATA(node)->node_zones[zone_type];
4092 if (populated_zone(z)) {
4094 &zonelist->_zonerefs[pos++]);
4095 check_highest_zone(zone_type);
4099 zonelist->_zonerefs[pos].zone = NULL;
4100 zonelist->_zonerefs[pos].zone_idx = 0;
4103 #if defined(CONFIG_64BIT)
4105 * Devices that require DMA32/DMA are relatively rare and do not justify a
4106 * penalty to every machine in case the specialised case applies. Default
4107 * to Node-ordering on 64-bit NUMA machines
4109 static int default_zonelist_order(void)
4111 return ZONELIST_ORDER_NODE;
4115 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4116 * by the kernel. If processes running on node 0 deplete the low memory zone
4117 * then reclaim will occur more frequency increasing stalls and potentially
4118 * be easier to OOM if a large percentage of the zone is under writeback or
4119 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4120 * Hence, default to zone ordering on 32-bit.
4122 static int default_zonelist_order(void)
4124 return ZONELIST_ORDER_ZONE;
4126 #endif /* CONFIG_64BIT */
4128 static void set_zonelist_order(void)
4130 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4131 current_zonelist_order = default_zonelist_order();
4133 current_zonelist_order = user_zonelist_order;
4136 static void build_zonelists(pg_data_t *pgdat)
4140 nodemask_t used_mask;
4141 int local_node, prev_node;
4142 struct zonelist *zonelist;
4143 int order = current_zonelist_order;
4145 /* initialize zonelists */
4146 for (i = 0; i < MAX_ZONELISTS; i++) {
4147 zonelist = pgdat->node_zonelists + i;
4148 zonelist->_zonerefs[0].zone = NULL;
4149 zonelist->_zonerefs[0].zone_idx = 0;
4152 /* NUMA-aware ordering of nodes */
4153 local_node = pgdat->node_id;
4154 load = nr_online_nodes;
4155 prev_node = local_node;
4156 nodes_clear(used_mask);
4158 memset(node_order, 0, sizeof(node_order));
4161 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4163 * We don't want to pressure a particular node.
4164 * So adding penalty to the first node in same
4165 * distance group to make it round-robin.
4167 if (node_distance(local_node, node) !=
4168 node_distance(local_node, prev_node))
4169 node_load[node] = load;
4173 if (order == ZONELIST_ORDER_NODE)
4174 build_zonelists_in_node_order(pgdat, node);
4176 node_order[j++] = node; /* remember order */
4179 if (order == ZONELIST_ORDER_ZONE) {
4180 /* calculate node order -- i.e., DMA last! */
4181 build_zonelists_in_zone_order(pgdat, j);
4184 build_thisnode_zonelists(pgdat);
4187 /* Construct the zonelist performance cache - see further mmzone.h */
4188 static void build_zonelist_cache(pg_data_t *pgdat)
4190 struct zonelist *zonelist;
4191 struct zonelist_cache *zlc;
4194 zonelist = &pgdat->node_zonelists[0];
4195 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
4196 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
4197 for (z = zonelist->_zonerefs; z->zone; z++)
4198 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
4201 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4203 * Return node id of node used for "local" allocations.
4204 * I.e., first node id of first zone in arg node's generic zonelist.
4205 * Used for initializing percpu 'numa_mem', which is used primarily
4206 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4208 int local_memory_node(int node)
4212 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4213 gfp_zone(GFP_KERNEL),
4220 #else /* CONFIG_NUMA */
4222 static void set_zonelist_order(void)
4224 current_zonelist_order = ZONELIST_ORDER_ZONE;
4227 static void build_zonelists(pg_data_t *pgdat)
4229 int node, local_node;
4231 struct zonelist *zonelist;
4233 local_node = pgdat->node_id;
4235 zonelist = &pgdat->node_zonelists[0];
4236 j = build_zonelists_node(pgdat, zonelist, 0);
4239 * Now we build the zonelist so that it contains the zones
4240 * of all the other nodes.
4241 * We don't want to pressure a particular node, so when
4242 * building the zones for node N, we make sure that the
4243 * zones coming right after the local ones are those from
4244 * node N+1 (modulo N)
4246 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4247 if (!node_online(node))
4249 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4251 for (node = 0; node < local_node; node++) {
4252 if (!node_online(node))
4254 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4257 zonelist->_zonerefs[j].zone = NULL;
4258 zonelist->_zonerefs[j].zone_idx = 0;
4261 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
4262 static void build_zonelist_cache(pg_data_t *pgdat)
4264 pgdat->node_zonelists[0].zlcache_ptr = NULL;
4267 #endif /* CONFIG_NUMA */
4270 * Boot pageset table. One per cpu which is going to be used for all
4271 * zones and all nodes. The parameters will be set in such a way
4272 * that an item put on a list will immediately be handed over to
4273 * the buddy list. This is safe since pageset manipulation is done
4274 * with interrupts disabled.
4276 * The boot_pagesets must be kept even after bootup is complete for
4277 * unused processors and/or zones. They do play a role for bootstrapping
4278 * hotplugged processors.
4280 * zoneinfo_show() and maybe other functions do
4281 * not check if the processor is online before following the pageset pointer.
4282 * Other parts of the kernel may not check if the zone is available.
4284 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4285 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4286 static void setup_zone_pageset(struct zone *zone);
4289 * Global mutex to protect against size modification of zonelists
4290 * as well as to serialize pageset setup for the new populated zone.
4292 DEFINE_MUTEX(zonelists_mutex);
4294 /* return values int ....just for stop_machine() */
4295 static int __build_all_zonelists(void *data)
4299 pg_data_t *self = data;
4302 memset(node_load, 0, sizeof(node_load));
4305 if (self && !node_online(self->node_id)) {
4306 build_zonelists(self);
4307 build_zonelist_cache(self);
4310 for_each_online_node(nid) {
4311 pg_data_t *pgdat = NODE_DATA(nid);
4313 build_zonelists(pgdat);
4314 build_zonelist_cache(pgdat);
4318 * Initialize the boot_pagesets that are going to be used
4319 * for bootstrapping processors. The real pagesets for
4320 * each zone will be allocated later when the per cpu
4321 * allocator is available.
4323 * boot_pagesets are used also for bootstrapping offline
4324 * cpus if the system is already booted because the pagesets
4325 * are needed to initialize allocators on a specific cpu too.
4326 * F.e. the percpu allocator needs the page allocator which
4327 * needs the percpu allocator in order to allocate its pagesets
4328 * (a chicken-egg dilemma).
4330 for_each_possible_cpu(cpu) {
4331 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4333 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4335 * We now know the "local memory node" for each node--
4336 * i.e., the node of the first zone in the generic zonelist.
4337 * Set up numa_mem percpu variable for on-line cpus. During
4338 * boot, only the boot cpu should be on-line; we'll init the
4339 * secondary cpus' numa_mem as they come on-line. During
4340 * node/memory hotplug, we'll fixup all on-line cpus.
4342 if (cpu_online(cpu))
4343 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4350 static noinline void __init
4351 build_all_zonelists_init(void)
4353 __build_all_zonelists(NULL);
4354 mminit_verify_zonelist();
4355 cpuset_init_current_mems_allowed();
4359 * Called with zonelists_mutex held always
4360 * unless system_state == SYSTEM_BOOTING.
4362 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4363 * [we're only called with non-NULL zone through __meminit paths] and
4364 * (2) call of __init annotated helper build_all_zonelists_init
4365 * [protected by SYSTEM_BOOTING].
4367 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4369 set_zonelist_order();
4371 if (system_state == SYSTEM_BOOTING) {
4372 build_all_zonelists_init();
4374 #ifdef CONFIG_MEMORY_HOTPLUG
4376 setup_zone_pageset(zone);
4378 /* we have to stop all cpus to guarantee there is no user
4380 stop_machine(__build_all_zonelists, pgdat, NULL);
4381 /* cpuset refresh routine should be here */
4383 vm_total_pages = nr_free_pagecache_pages();
4385 * Disable grouping by mobility if the number of pages in the
4386 * system is too low to allow the mechanism to work. It would be
4387 * more accurate, but expensive to check per-zone. This check is
4388 * made on memory-hotadd so a system can start with mobility
4389 * disabled and enable it later
4391 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4392 page_group_by_mobility_disabled = 1;
4394 page_group_by_mobility_disabled = 0;
4396 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4397 "Total pages: %ld\n",
4399 zonelist_order_name[current_zonelist_order],
4400 page_group_by_mobility_disabled ? "off" : "on",
4403 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4408 * Helper functions to size the waitqueue hash table.
4409 * Essentially these want to choose hash table sizes sufficiently
4410 * large so that collisions trying to wait on pages are rare.
4411 * But in fact, the number of active page waitqueues on typical
4412 * systems is ridiculously low, less than 200. So this is even
4413 * conservative, even though it seems large.
4415 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4416 * waitqueues, i.e. the size of the waitq table given the number of pages.
4418 #define PAGES_PER_WAITQUEUE 256
4420 #ifndef CONFIG_MEMORY_HOTPLUG
4421 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4423 unsigned long size = 1;
4425 pages /= PAGES_PER_WAITQUEUE;
4427 while (size < pages)
4431 * Once we have dozens or even hundreds of threads sleeping
4432 * on IO we've got bigger problems than wait queue collision.
4433 * Limit the size of the wait table to a reasonable size.
4435 size = min(size, 4096UL);
4437 return max(size, 4UL);
4441 * A zone's size might be changed by hot-add, so it is not possible to determine
4442 * a suitable size for its wait_table. So we use the maximum size now.
4444 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4446 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4447 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4448 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4450 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4451 * or more by the traditional way. (See above). It equals:
4453 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4454 * ia64(16K page size) : = ( 8G + 4M)byte.
4455 * powerpc (64K page size) : = (32G +16M)byte.
4457 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4464 * This is an integer logarithm so that shifts can be used later
4465 * to extract the more random high bits from the multiplicative
4466 * hash function before the remainder is taken.
4468 static inline unsigned long wait_table_bits(unsigned long size)
4474 * Check if a pageblock contains reserved pages
4476 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
4480 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4481 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
4488 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4489 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4490 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4491 * higher will lead to a bigger reserve which will get freed as contiguous
4492 * blocks as reclaim kicks in
4494 static void setup_zone_migrate_reserve(struct zone *zone)
4496 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
4498 unsigned long block_migratetype;
4503 * Get the start pfn, end pfn and the number of blocks to reserve
4504 * We have to be careful to be aligned to pageblock_nr_pages to
4505 * make sure that we always check pfn_valid for the first page in
4508 start_pfn = zone->zone_start_pfn;
4509 end_pfn = zone_end_pfn(zone);
4510 start_pfn = roundup(start_pfn, pageblock_nr_pages);
4511 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
4515 * Reserve blocks are generally in place to help high-order atomic
4516 * allocations that are short-lived. A min_free_kbytes value that
4517 * would result in more than 2 reserve blocks for atomic allocations
4518 * is assumed to be in place to help anti-fragmentation for the
4519 * future allocation of hugepages at runtime.
4521 reserve = min(2, reserve);
4522 old_reserve = zone->nr_migrate_reserve_block;
4524 /* When memory hot-add, we almost always need to do nothing */
4525 if (reserve == old_reserve)
4527 zone->nr_migrate_reserve_block = reserve;
4529 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
4530 if (!early_page_nid_uninitialised(pfn, zone_to_nid(zone)))
4533 if (!pfn_valid(pfn))
4535 page = pfn_to_page(pfn);
4537 /* Watch out for overlapping nodes */
4538 if (page_to_nid(page) != zone_to_nid(zone))
4541 block_migratetype = get_pageblock_migratetype(page);
4543 /* Only test what is necessary when the reserves are not met */
4546 * Blocks with reserved pages will never free, skip
4549 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
4550 if (pageblock_is_reserved(pfn, block_end_pfn))
4553 /* If this block is reserved, account for it */
4554 if (block_migratetype == MIGRATE_RESERVE) {
4559 /* Suitable for reserving if this block is movable */
4560 if (block_migratetype == MIGRATE_MOVABLE) {
4561 set_pageblock_migratetype(page,
4563 move_freepages_block(zone, page,
4568 } else if (!old_reserve) {
4570 * At boot time we don't need to scan the whole zone
4571 * for turning off MIGRATE_RESERVE.
4577 * If the reserve is met and this is a previous reserved block,
4580 if (block_migratetype == MIGRATE_RESERVE) {
4581 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4582 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4588 * Initially all pages are reserved - free ones are freed
4589 * up by free_all_bootmem() once the early boot process is
4590 * done. Non-atomic initialization, single-pass.
4592 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4593 unsigned long start_pfn, enum memmap_context context)
4595 pg_data_t *pgdat = NODE_DATA(nid);
4596 unsigned long end_pfn = start_pfn + size;
4599 unsigned long nr_initialised = 0;
4601 if (highest_memmap_pfn < end_pfn - 1)
4602 highest_memmap_pfn = end_pfn - 1;
4604 z = &pgdat->node_zones[zone];
4605 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4607 * There can be holes in boot-time mem_map[]s
4608 * handed to this function. They do not
4609 * exist on hotplugged memory.
4611 if (context == MEMMAP_EARLY) {
4612 if (!early_pfn_valid(pfn))
4614 if (!early_pfn_in_nid(pfn, nid))
4616 if (!update_defer_init(pgdat, pfn, end_pfn,
4622 * Mark the block movable so that blocks are reserved for
4623 * movable at startup. This will force kernel allocations
4624 * to reserve their blocks rather than leaking throughout
4625 * the address space during boot when many long-lived
4626 * kernel allocations are made. Later some blocks near
4627 * the start are marked MIGRATE_RESERVE by
4628 * setup_zone_migrate_reserve()
4630 * bitmap is created for zone's valid pfn range. but memmap
4631 * can be created for invalid pages (for alignment)
4632 * check here not to call set_pageblock_migratetype() against
4635 if (!(pfn & (pageblock_nr_pages - 1))) {
4636 struct page *page = pfn_to_page(pfn);
4638 __init_single_page(page, pfn, zone, nid);
4639 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4641 __init_single_pfn(pfn, zone, nid);
4646 static void __meminit zone_init_free_lists(struct zone *zone)
4648 unsigned int order, t;
4649 for_each_migratetype_order(order, t) {
4650 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4651 zone->free_area[order].nr_free = 0;
4655 #ifndef __HAVE_ARCH_MEMMAP_INIT
4656 #define memmap_init(size, nid, zone, start_pfn) \
4657 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4660 static int zone_batchsize(struct zone *zone)
4666 * The per-cpu-pages pools are set to around 1000th of the
4667 * size of the zone. But no more than 1/2 of a meg.
4669 * OK, so we don't know how big the cache is. So guess.
4671 batch = zone->managed_pages / 1024;
4672 if (batch * PAGE_SIZE > 512 * 1024)
4673 batch = (512 * 1024) / PAGE_SIZE;
4674 batch /= 4; /* We effectively *= 4 below */
4679 * Clamp the batch to a 2^n - 1 value. Having a power
4680 * of 2 value was found to be more likely to have
4681 * suboptimal cache aliasing properties in some cases.
4683 * For example if 2 tasks are alternately allocating
4684 * batches of pages, one task can end up with a lot
4685 * of pages of one half of the possible page colors
4686 * and the other with pages of the other colors.
4688 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4693 /* The deferral and batching of frees should be suppressed under NOMMU
4696 * The problem is that NOMMU needs to be able to allocate large chunks
4697 * of contiguous memory as there's no hardware page translation to
4698 * assemble apparent contiguous memory from discontiguous pages.
4700 * Queueing large contiguous runs of pages for batching, however,
4701 * causes the pages to actually be freed in smaller chunks. As there
4702 * can be a significant delay between the individual batches being
4703 * recycled, this leads to the once large chunks of space being
4704 * fragmented and becoming unavailable for high-order allocations.
4711 * pcp->high and pcp->batch values are related and dependent on one another:
4712 * ->batch must never be higher then ->high.
4713 * The following function updates them in a safe manner without read side
4716 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4717 * those fields changing asynchronously (acording the the above rule).
4719 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4720 * outside of boot time (or some other assurance that no concurrent updaters
4723 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4724 unsigned long batch)
4726 /* start with a fail safe value for batch */
4730 /* Update high, then batch, in order */
4737 /* a companion to pageset_set_high() */
4738 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4740 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4743 static void pageset_init(struct per_cpu_pageset *p)
4745 struct per_cpu_pages *pcp;
4748 memset(p, 0, sizeof(*p));
4752 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4753 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4756 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4759 pageset_set_batch(p, batch);
4763 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4764 * to the value high for the pageset p.
4766 static void pageset_set_high(struct per_cpu_pageset *p,
4769 unsigned long batch = max(1UL, high / 4);
4770 if ((high / 4) > (PAGE_SHIFT * 8))
4771 batch = PAGE_SHIFT * 8;
4773 pageset_update(&p->pcp, high, batch);
4776 static void pageset_set_high_and_batch(struct zone *zone,
4777 struct per_cpu_pageset *pcp)
4779 if (percpu_pagelist_fraction)
4780 pageset_set_high(pcp,
4781 (zone->managed_pages /
4782 percpu_pagelist_fraction));
4784 pageset_set_batch(pcp, zone_batchsize(zone));
4787 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4789 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4792 pageset_set_high_and_batch(zone, pcp);
4795 static void __meminit setup_zone_pageset(struct zone *zone)
4798 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4799 for_each_possible_cpu(cpu)
4800 zone_pageset_init(zone, cpu);
4804 * Allocate per cpu pagesets and initialize them.
4805 * Before this call only boot pagesets were available.
4807 void __init setup_per_cpu_pageset(void)
4811 for_each_populated_zone(zone)
4812 setup_zone_pageset(zone);
4815 static noinline __init_refok
4816 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4822 * The per-page waitqueue mechanism uses hashed waitqueues
4825 zone->wait_table_hash_nr_entries =
4826 wait_table_hash_nr_entries(zone_size_pages);
4827 zone->wait_table_bits =
4828 wait_table_bits(zone->wait_table_hash_nr_entries);
4829 alloc_size = zone->wait_table_hash_nr_entries
4830 * sizeof(wait_queue_head_t);
4832 if (!slab_is_available()) {
4833 zone->wait_table = (wait_queue_head_t *)
4834 memblock_virt_alloc_node_nopanic(
4835 alloc_size, zone->zone_pgdat->node_id);
4838 * This case means that a zone whose size was 0 gets new memory
4839 * via memory hot-add.
4840 * But it may be the case that a new node was hot-added. In
4841 * this case vmalloc() will not be able to use this new node's
4842 * memory - this wait_table must be initialized to use this new
4843 * node itself as well.
4844 * To use this new node's memory, further consideration will be
4847 zone->wait_table = vmalloc(alloc_size);
4849 if (!zone->wait_table)
4852 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4853 init_waitqueue_head(zone->wait_table + i);
4858 static __meminit void zone_pcp_init(struct zone *zone)
4861 * per cpu subsystem is not up at this point. The following code
4862 * relies on the ability of the linker to provide the
4863 * offset of a (static) per cpu variable into the per cpu area.
4865 zone->pageset = &boot_pageset;
4867 if (populated_zone(zone))
4868 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4869 zone->name, zone->present_pages,
4870 zone_batchsize(zone));
4873 int __meminit init_currently_empty_zone(struct zone *zone,
4874 unsigned long zone_start_pfn,
4876 enum memmap_context context)
4878 struct pglist_data *pgdat = zone->zone_pgdat;
4880 ret = zone_wait_table_init(zone, size);
4883 pgdat->nr_zones = zone_idx(zone) + 1;
4885 zone->zone_start_pfn = zone_start_pfn;
4887 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4888 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4890 (unsigned long)zone_idx(zone),
4891 zone_start_pfn, (zone_start_pfn + size));
4893 zone_init_free_lists(zone);
4898 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4899 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4902 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4904 int __meminit __early_pfn_to_nid(unsigned long pfn,
4905 struct mminit_pfnnid_cache *state)
4907 unsigned long start_pfn, end_pfn;
4910 if (state->last_start <= pfn && pfn < state->last_end)
4911 return state->last_nid;
4913 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4915 state->last_start = start_pfn;
4916 state->last_end = end_pfn;
4917 state->last_nid = nid;
4922 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4925 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4926 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4927 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4929 * If an architecture guarantees that all ranges registered contain no holes
4930 * and may be freed, this this function may be used instead of calling
4931 * memblock_free_early_nid() manually.
4933 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4935 unsigned long start_pfn, end_pfn;
4938 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4939 start_pfn = min(start_pfn, max_low_pfn);
4940 end_pfn = min(end_pfn, max_low_pfn);
4942 if (start_pfn < end_pfn)
4943 memblock_free_early_nid(PFN_PHYS(start_pfn),
4944 (end_pfn - start_pfn) << PAGE_SHIFT,
4950 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4951 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4953 * If an architecture guarantees that all ranges registered contain no holes and may
4954 * be freed, this function may be used instead of calling memory_present() manually.
4956 void __init sparse_memory_present_with_active_regions(int nid)
4958 unsigned long start_pfn, end_pfn;
4961 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4962 memory_present(this_nid, start_pfn, end_pfn);
4966 * get_pfn_range_for_nid - Return the start and end page frames for a node
4967 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4968 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4969 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4971 * It returns the start and end page frame of a node based on information
4972 * provided by memblock_set_node(). If called for a node
4973 * with no available memory, a warning is printed and the start and end
4976 void __meminit get_pfn_range_for_nid(unsigned int nid,
4977 unsigned long *start_pfn, unsigned long *end_pfn)
4979 unsigned long this_start_pfn, this_end_pfn;
4985 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4986 *start_pfn = min(*start_pfn, this_start_pfn);
4987 *end_pfn = max(*end_pfn, this_end_pfn);
4990 if (*start_pfn == -1UL)
4995 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4996 * assumption is made that zones within a node are ordered in monotonic
4997 * increasing memory addresses so that the "highest" populated zone is used
4999 static void __init find_usable_zone_for_movable(void)
5002 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5003 if (zone_index == ZONE_MOVABLE)
5006 if (arch_zone_highest_possible_pfn[zone_index] >
5007 arch_zone_lowest_possible_pfn[zone_index])
5011 VM_BUG_ON(zone_index == -1);
5012 movable_zone = zone_index;
5016 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5017 * because it is sized independent of architecture. Unlike the other zones,
5018 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5019 * in each node depending on the size of each node and how evenly kernelcore
5020 * is distributed. This helper function adjusts the zone ranges
5021 * provided by the architecture for a given node by using the end of the
5022 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5023 * zones within a node are in order of monotonic increases memory addresses
5025 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5026 unsigned long zone_type,
5027 unsigned long node_start_pfn,
5028 unsigned long node_end_pfn,
5029 unsigned long *zone_start_pfn,
5030 unsigned long *zone_end_pfn)
5032 /* Only adjust if ZONE_MOVABLE is on this node */
5033 if (zone_movable_pfn[nid]) {
5034 /* Size ZONE_MOVABLE */
5035 if (zone_type == ZONE_MOVABLE) {
5036 *zone_start_pfn = zone_movable_pfn[nid];
5037 *zone_end_pfn = min(node_end_pfn,
5038 arch_zone_highest_possible_pfn[movable_zone]);
5040 /* Adjust for ZONE_MOVABLE starting within this range */
5041 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
5042 *zone_end_pfn > zone_movable_pfn[nid]) {
5043 *zone_end_pfn = zone_movable_pfn[nid];
5045 /* Check if this whole range is within ZONE_MOVABLE */
5046 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5047 *zone_start_pfn = *zone_end_pfn;
5052 * Return the number of pages a zone spans in a node, including holes
5053 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5055 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5056 unsigned long zone_type,
5057 unsigned long node_start_pfn,
5058 unsigned long node_end_pfn,
5059 unsigned long *ignored)
5061 unsigned long zone_start_pfn, zone_end_pfn;
5063 /* Get the start and end of the zone */
5064 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5065 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5066 adjust_zone_range_for_zone_movable(nid, zone_type,
5067 node_start_pfn, node_end_pfn,
5068 &zone_start_pfn, &zone_end_pfn);
5070 /* Check that this node has pages within the zone's required range */
5071 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
5074 /* Move the zone boundaries inside the node if necessary */
5075 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
5076 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
5078 /* Return the spanned pages */
5079 return zone_end_pfn - zone_start_pfn;
5083 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5084 * then all holes in the requested range will be accounted for.
5086 unsigned long __meminit __absent_pages_in_range(int nid,
5087 unsigned long range_start_pfn,
5088 unsigned long range_end_pfn)
5090 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5091 unsigned long start_pfn, end_pfn;
5094 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5095 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5096 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5097 nr_absent -= end_pfn - start_pfn;
5103 * absent_pages_in_range - Return number of page frames in holes within a range
5104 * @start_pfn: The start PFN to start searching for holes
5105 * @end_pfn: The end PFN to stop searching for holes
5107 * It returns the number of pages frames in memory holes within a range.
5109 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5110 unsigned long end_pfn)
5112 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5115 /* Return the number of page frames in holes in a zone on a node */
5116 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5117 unsigned long zone_type,
5118 unsigned long node_start_pfn,
5119 unsigned long node_end_pfn,
5120 unsigned long *ignored)
5122 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5123 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5124 unsigned long zone_start_pfn, zone_end_pfn;
5126 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5127 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5129 adjust_zone_range_for_zone_movable(nid, zone_type,
5130 node_start_pfn, node_end_pfn,
5131 &zone_start_pfn, &zone_end_pfn);
5132 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5135 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5136 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5137 unsigned long zone_type,
5138 unsigned long node_start_pfn,
5139 unsigned long node_end_pfn,
5140 unsigned long *zones_size)
5142 return zones_size[zone_type];
5145 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5146 unsigned long zone_type,
5147 unsigned long node_start_pfn,
5148 unsigned long node_end_pfn,
5149 unsigned long *zholes_size)
5154 return zholes_size[zone_type];
5157 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5159 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5160 unsigned long node_start_pfn,
5161 unsigned long node_end_pfn,
5162 unsigned long *zones_size,
5163 unsigned long *zholes_size)
5165 unsigned long realtotalpages = 0, totalpages = 0;
5168 for (i = 0; i < MAX_NR_ZONES; i++) {
5169 struct zone *zone = pgdat->node_zones + i;
5170 unsigned long size, real_size;
5172 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5176 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5177 node_start_pfn, node_end_pfn,
5179 zone->spanned_pages = size;
5180 zone->present_pages = real_size;
5183 realtotalpages += real_size;
5186 pgdat->node_spanned_pages = totalpages;
5187 pgdat->node_present_pages = realtotalpages;
5188 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5192 #ifndef CONFIG_SPARSEMEM
5194 * Calculate the size of the zone->blockflags rounded to an unsigned long
5195 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5196 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5197 * round what is now in bits to nearest long in bits, then return it in
5200 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5202 unsigned long usemapsize;
5204 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5205 usemapsize = roundup(zonesize, pageblock_nr_pages);
5206 usemapsize = usemapsize >> pageblock_order;
5207 usemapsize *= NR_PAGEBLOCK_BITS;
5208 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5210 return usemapsize / 8;
5213 static void __init setup_usemap(struct pglist_data *pgdat,
5215 unsigned long zone_start_pfn,
5216 unsigned long zonesize)
5218 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5219 zone->pageblock_flags = NULL;
5221 zone->pageblock_flags =
5222 memblock_virt_alloc_node_nopanic(usemapsize,
5226 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5227 unsigned long zone_start_pfn, unsigned long zonesize) {}
5228 #endif /* CONFIG_SPARSEMEM */
5230 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5232 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5233 void __paginginit set_pageblock_order(void)
5237 /* Check that pageblock_nr_pages has not already been setup */
5238 if (pageblock_order)
5241 if (HPAGE_SHIFT > PAGE_SHIFT)
5242 order = HUGETLB_PAGE_ORDER;
5244 order = MAX_ORDER - 1;
5247 * Assume the largest contiguous order of interest is a huge page.
5248 * This value may be variable depending on boot parameters on IA64 and
5251 pageblock_order = order;
5253 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5256 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5257 * is unused as pageblock_order is set at compile-time. See
5258 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5261 void __paginginit set_pageblock_order(void)
5265 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5267 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5268 unsigned long present_pages)
5270 unsigned long pages = spanned_pages;
5273 * Provide a more accurate estimation if there are holes within
5274 * the zone and SPARSEMEM is in use. If there are holes within the
5275 * zone, each populated memory region may cost us one or two extra
5276 * memmap pages due to alignment because memmap pages for each
5277 * populated regions may not naturally algined on page boundary.
5278 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5280 if (spanned_pages > present_pages + (present_pages >> 4) &&
5281 IS_ENABLED(CONFIG_SPARSEMEM))
5282 pages = present_pages;
5284 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5288 * Set up the zone data structures:
5289 * - mark all pages reserved
5290 * - mark all memory queues empty
5291 * - clear the memory bitmaps
5293 * NOTE: pgdat should get zeroed by caller.
5295 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
5296 unsigned long node_start_pfn, unsigned long node_end_pfn)
5299 int nid = pgdat->node_id;
5300 unsigned long zone_start_pfn = pgdat->node_start_pfn;
5303 pgdat_resize_init(pgdat);
5304 #ifdef CONFIG_NUMA_BALANCING
5305 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5306 pgdat->numabalancing_migrate_nr_pages = 0;
5307 pgdat->numabalancing_migrate_next_window = jiffies;
5309 init_waitqueue_head(&pgdat->kswapd_wait);
5310 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5311 pgdat_page_ext_init(pgdat);
5313 for (j = 0; j < MAX_NR_ZONES; j++) {
5314 struct zone *zone = pgdat->node_zones + j;
5315 unsigned long size, realsize, freesize, memmap_pages;
5317 size = zone->spanned_pages;
5318 realsize = freesize = zone->present_pages;
5321 * Adjust freesize so that it accounts for how much memory
5322 * is used by this zone for memmap. This affects the watermark
5323 * and per-cpu initialisations
5325 memmap_pages = calc_memmap_size(size, realsize);
5326 if (!is_highmem_idx(j)) {
5327 if (freesize >= memmap_pages) {
5328 freesize -= memmap_pages;
5331 " %s zone: %lu pages used for memmap\n",
5332 zone_names[j], memmap_pages);
5335 " %s zone: %lu pages exceeds freesize %lu\n",
5336 zone_names[j], memmap_pages, freesize);
5339 /* Account for reserved pages */
5340 if (j == 0 && freesize > dma_reserve) {
5341 freesize -= dma_reserve;
5342 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5343 zone_names[0], dma_reserve);
5346 if (!is_highmem_idx(j))
5347 nr_kernel_pages += freesize;
5348 /* Charge for highmem memmap if there are enough kernel pages */
5349 else if (nr_kernel_pages > memmap_pages * 2)
5350 nr_kernel_pages -= memmap_pages;
5351 nr_all_pages += freesize;
5354 * Set an approximate value for lowmem here, it will be adjusted
5355 * when the bootmem allocator frees pages into the buddy system.
5356 * And all highmem pages will be managed by the buddy system.
5358 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5361 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5363 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5365 zone->name = zone_names[j];
5366 spin_lock_init(&zone->lock);
5367 spin_lock_init(&zone->lru_lock);
5368 zone_seqlock_init(zone);
5369 zone->zone_pgdat = pgdat;
5370 zone_pcp_init(zone);
5372 /* For bootup, initialized properly in watermark setup */
5373 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5375 lruvec_init(&zone->lruvec);
5379 set_pageblock_order();
5380 setup_usemap(pgdat, zone, zone_start_pfn, size);
5381 ret = init_currently_empty_zone(zone, zone_start_pfn,
5382 size, MEMMAP_EARLY);
5384 memmap_init(size, nid, j, zone_start_pfn);
5385 zone_start_pfn += size;
5389 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5391 /* Skip empty nodes */
5392 if (!pgdat->node_spanned_pages)
5395 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5396 /* ia64 gets its own node_mem_map, before this, without bootmem */
5397 if (!pgdat->node_mem_map) {
5398 unsigned long size, start, end;
5402 * The zone's endpoints aren't required to be MAX_ORDER
5403 * aligned but the node_mem_map endpoints must be in order
5404 * for the buddy allocator to function correctly.
5406 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5407 end = pgdat_end_pfn(pgdat);
5408 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5409 size = (end - start) * sizeof(struct page);
5410 map = alloc_remap(pgdat->node_id, size);
5412 map = memblock_virt_alloc_node_nopanic(size,
5414 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
5416 #ifndef CONFIG_NEED_MULTIPLE_NODES
5418 * With no DISCONTIG, the global mem_map is just set as node 0's
5420 if (pgdat == NODE_DATA(0)) {
5421 mem_map = NODE_DATA(0)->node_mem_map;
5422 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5423 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5424 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
5425 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5428 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5431 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5432 unsigned long node_start_pfn, unsigned long *zholes_size)
5434 pg_data_t *pgdat = NODE_DATA(nid);
5435 unsigned long start_pfn = 0;
5436 unsigned long end_pfn = 0;
5438 /* pg_data_t should be reset to zero when it's allocated */
5439 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5441 reset_deferred_meminit(pgdat);
5442 pgdat->node_id = nid;
5443 pgdat->node_start_pfn = node_start_pfn;
5444 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5445 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5446 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5447 (u64)start_pfn << PAGE_SHIFT, ((u64)end_pfn << PAGE_SHIFT) - 1);
5449 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5450 zones_size, zholes_size);
5452 alloc_node_mem_map(pgdat);
5453 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5454 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5455 nid, (unsigned long)pgdat,
5456 (unsigned long)pgdat->node_mem_map);
5459 free_area_init_core(pgdat, start_pfn, end_pfn);
5462 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5464 #if MAX_NUMNODES > 1
5466 * Figure out the number of possible node ids.
5468 void __init setup_nr_node_ids(void)
5471 unsigned int highest = 0;
5473 for_each_node_mask(node, node_possible_map)
5475 nr_node_ids = highest + 1;
5480 * node_map_pfn_alignment - determine the maximum internode alignment
5482 * This function should be called after node map is populated and sorted.
5483 * It calculates the maximum power of two alignment which can distinguish
5486 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5487 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5488 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5489 * shifted, 1GiB is enough and this function will indicate so.
5491 * This is used to test whether pfn -> nid mapping of the chosen memory
5492 * model has fine enough granularity to avoid incorrect mapping for the
5493 * populated node map.
5495 * Returns the determined alignment in pfn's. 0 if there is no alignment
5496 * requirement (single node).
5498 unsigned long __init node_map_pfn_alignment(void)
5500 unsigned long accl_mask = 0, last_end = 0;
5501 unsigned long start, end, mask;
5505 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5506 if (!start || last_nid < 0 || last_nid == nid) {
5513 * Start with a mask granular enough to pin-point to the
5514 * start pfn and tick off bits one-by-one until it becomes
5515 * too coarse to separate the current node from the last.
5517 mask = ~((1 << __ffs(start)) - 1);
5518 while (mask && last_end <= (start & (mask << 1)))
5521 /* accumulate all internode masks */
5525 /* convert mask to number of pages */
5526 return ~accl_mask + 1;
5529 /* Find the lowest pfn for a node */
5530 static unsigned long __init find_min_pfn_for_node(int nid)
5532 unsigned long min_pfn = ULONG_MAX;
5533 unsigned long start_pfn;
5536 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5537 min_pfn = min(min_pfn, start_pfn);
5539 if (min_pfn == ULONG_MAX) {
5541 "Could not find start_pfn for node %d\n", nid);
5549 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5551 * It returns the minimum PFN based on information provided via
5552 * memblock_set_node().
5554 unsigned long __init find_min_pfn_with_active_regions(void)
5556 return find_min_pfn_for_node(MAX_NUMNODES);
5560 * early_calculate_totalpages()
5561 * Sum pages in active regions for movable zone.
5562 * Populate N_MEMORY for calculating usable_nodes.
5564 static unsigned long __init early_calculate_totalpages(void)
5566 unsigned long totalpages = 0;
5567 unsigned long start_pfn, end_pfn;
5570 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5571 unsigned long pages = end_pfn - start_pfn;
5573 totalpages += pages;
5575 node_set_state(nid, N_MEMORY);
5581 * Find the PFN the Movable zone begins in each node. Kernel memory
5582 * is spread evenly between nodes as long as the nodes have enough
5583 * memory. When they don't, some nodes will have more kernelcore than
5586 static void __init find_zone_movable_pfns_for_nodes(void)
5589 unsigned long usable_startpfn;
5590 unsigned long kernelcore_node, kernelcore_remaining;
5591 /* save the state before borrow the nodemask */
5592 nodemask_t saved_node_state = node_states[N_MEMORY];
5593 unsigned long totalpages = early_calculate_totalpages();
5594 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5595 struct memblock_region *r;
5597 /* Need to find movable_zone earlier when movable_node is specified. */
5598 find_usable_zone_for_movable();
5601 * If movable_node is specified, ignore kernelcore and movablecore
5604 if (movable_node_is_enabled()) {
5605 for_each_memblock(memory, r) {
5606 if (!memblock_is_hotpluggable(r))
5611 usable_startpfn = PFN_DOWN(r->base);
5612 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5613 min(usable_startpfn, zone_movable_pfn[nid]) :
5621 * If movablecore=nn[KMG] was specified, calculate what size of
5622 * kernelcore that corresponds so that memory usable for
5623 * any allocation type is evenly spread. If both kernelcore
5624 * and movablecore are specified, then the value of kernelcore
5625 * will be used for required_kernelcore if it's greater than
5626 * what movablecore would have allowed.
5628 if (required_movablecore) {
5629 unsigned long corepages;
5632 * Round-up so that ZONE_MOVABLE is at least as large as what
5633 * was requested by the user
5635 required_movablecore =
5636 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5637 corepages = totalpages - required_movablecore;
5639 required_kernelcore = max(required_kernelcore, corepages);
5642 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5643 if (!required_kernelcore)
5646 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5647 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5650 /* Spread kernelcore memory as evenly as possible throughout nodes */
5651 kernelcore_node = required_kernelcore / usable_nodes;
5652 for_each_node_state(nid, N_MEMORY) {
5653 unsigned long start_pfn, end_pfn;
5656 * Recalculate kernelcore_node if the division per node
5657 * now exceeds what is necessary to satisfy the requested
5658 * amount of memory for the kernel
5660 if (required_kernelcore < kernelcore_node)
5661 kernelcore_node = required_kernelcore / usable_nodes;
5664 * As the map is walked, we track how much memory is usable
5665 * by the kernel using kernelcore_remaining. When it is
5666 * 0, the rest of the node is usable by ZONE_MOVABLE
5668 kernelcore_remaining = kernelcore_node;
5670 /* Go through each range of PFNs within this node */
5671 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5672 unsigned long size_pages;
5674 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5675 if (start_pfn >= end_pfn)
5678 /* Account for what is only usable for kernelcore */
5679 if (start_pfn < usable_startpfn) {
5680 unsigned long kernel_pages;
5681 kernel_pages = min(end_pfn, usable_startpfn)
5684 kernelcore_remaining -= min(kernel_pages,
5685 kernelcore_remaining);
5686 required_kernelcore -= min(kernel_pages,
5687 required_kernelcore);
5689 /* Continue if range is now fully accounted */
5690 if (end_pfn <= usable_startpfn) {
5693 * Push zone_movable_pfn to the end so
5694 * that if we have to rebalance
5695 * kernelcore across nodes, we will
5696 * not double account here
5698 zone_movable_pfn[nid] = end_pfn;
5701 start_pfn = usable_startpfn;
5705 * The usable PFN range for ZONE_MOVABLE is from
5706 * start_pfn->end_pfn. Calculate size_pages as the
5707 * number of pages used as kernelcore
5709 size_pages = end_pfn - start_pfn;
5710 if (size_pages > kernelcore_remaining)
5711 size_pages = kernelcore_remaining;
5712 zone_movable_pfn[nid] = start_pfn + size_pages;
5715 * Some kernelcore has been met, update counts and
5716 * break if the kernelcore for this node has been
5719 required_kernelcore -= min(required_kernelcore,
5721 kernelcore_remaining -= size_pages;
5722 if (!kernelcore_remaining)
5728 * If there is still required_kernelcore, we do another pass with one
5729 * less node in the count. This will push zone_movable_pfn[nid] further
5730 * along on the nodes that still have memory until kernelcore is
5734 if (usable_nodes && required_kernelcore > usable_nodes)
5738 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5739 for (nid = 0; nid < MAX_NUMNODES; nid++)
5740 zone_movable_pfn[nid] =
5741 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5744 /* restore the node_state */
5745 node_states[N_MEMORY] = saved_node_state;
5748 /* Any regular or high memory on that node ? */
5749 static void check_for_memory(pg_data_t *pgdat, int nid)
5751 enum zone_type zone_type;
5753 if (N_MEMORY == N_NORMAL_MEMORY)
5756 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5757 struct zone *zone = &pgdat->node_zones[zone_type];
5758 if (populated_zone(zone)) {
5759 node_set_state(nid, N_HIGH_MEMORY);
5760 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5761 zone_type <= ZONE_NORMAL)
5762 node_set_state(nid, N_NORMAL_MEMORY);
5769 * free_area_init_nodes - Initialise all pg_data_t and zone data
5770 * @max_zone_pfn: an array of max PFNs for each zone
5772 * This will call free_area_init_node() for each active node in the system.
5773 * Using the page ranges provided by memblock_set_node(), the size of each
5774 * zone in each node and their holes is calculated. If the maximum PFN
5775 * between two adjacent zones match, it is assumed that the zone is empty.
5776 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5777 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5778 * starts where the previous one ended. For example, ZONE_DMA32 starts
5779 * at arch_max_dma_pfn.
5781 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5783 unsigned long start_pfn, end_pfn;
5786 /* Record where the zone boundaries are */
5787 memset(arch_zone_lowest_possible_pfn, 0,
5788 sizeof(arch_zone_lowest_possible_pfn));
5789 memset(arch_zone_highest_possible_pfn, 0,
5790 sizeof(arch_zone_highest_possible_pfn));
5791 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5792 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5793 for (i = 1; i < MAX_NR_ZONES; i++) {
5794 if (i == ZONE_MOVABLE)
5796 arch_zone_lowest_possible_pfn[i] =
5797 arch_zone_highest_possible_pfn[i-1];
5798 arch_zone_highest_possible_pfn[i] =
5799 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5801 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5802 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5804 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5805 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5806 find_zone_movable_pfns_for_nodes();
5808 /* Print out the zone ranges */
5809 pr_info("Zone ranges:\n");
5810 for (i = 0; i < MAX_NR_ZONES; i++) {
5811 if (i == ZONE_MOVABLE)
5813 pr_info(" %-8s ", zone_names[i]);
5814 if (arch_zone_lowest_possible_pfn[i] ==
5815 arch_zone_highest_possible_pfn[i])
5818 pr_cont("[mem %#018Lx-%#018Lx]\n",
5819 (u64)arch_zone_lowest_possible_pfn[i]
5821 ((u64)arch_zone_highest_possible_pfn[i]
5822 << PAGE_SHIFT) - 1);
5825 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5826 pr_info("Movable zone start for each node\n");
5827 for (i = 0; i < MAX_NUMNODES; i++) {
5828 if (zone_movable_pfn[i])
5829 pr_info(" Node %d: %#018Lx\n", i,
5830 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5833 /* Print out the early node map */
5834 pr_info("Early memory node ranges\n");
5835 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5836 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5837 (u64)start_pfn << PAGE_SHIFT,
5838 ((u64)end_pfn << PAGE_SHIFT) - 1);
5840 /* Initialise every node */
5841 mminit_verify_pageflags_layout();
5842 setup_nr_node_ids();
5843 for_each_online_node(nid) {
5844 pg_data_t *pgdat = NODE_DATA(nid);
5845 free_area_init_node(nid, NULL,
5846 find_min_pfn_for_node(nid), NULL);
5848 /* Any memory on that node */
5849 if (pgdat->node_present_pages)
5850 node_set_state(nid, N_MEMORY);
5851 check_for_memory(pgdat, nid);
5855 static int __init cmdline_parse_core(char *p, unsigned long *core)
5857 unsigned long long coremem;
5861 coremem = memparse(p, &p);
5862 *core = coremem >> PAGE_SHIFT;
5864 /* Paranoid check that UL is enough for the coremem value */
5865 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5871 * kernelcore=size sets the amount of memory for use for allocations that
5872 * cannot be reclaimed or migrated.
5874 static int __init cmdline_parse_kernelcore(char *p)
5876 return cmdline_parse_core(p, &required_kernelcore);
5880 * movablecore=size sets the amount of memory for use for allocations that
5881 * can be reclaimed or migrated.
5883 static int __init cmdline_parse_movablecore(char *p)
5885 return cmdline_parse_core(p, &required_movablecore);
5888 early_param("kernelcore", cmdline_parse_kernelcore);
5889 early_param("movablecore", cmdline_parse_movablecore);
5891 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5893 void adjust_managed_page_count(struct page *page, long count)
5895 spin_lock(&managed_page_count_lock);
5896 page_zone(page)->managed_pages += count;
5897 totalram_pages += count;
5898 #ifdef CONFIG_HIGHMEM
5899 if (PageHighMem(page))
5900 totalhigh_pages += count;
5902 spin_unlock(&managed_page_count_lock);
5904 EXPORT_SYMBOL(adjust_managed_page_count);
5906 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5909 unsigned long pages = 0;
5911 start = (void *)PAGE_ALIGN((unsigned long)start);
5912 end = (void *)((unsigned long)end & PAGE_MASK);
5913 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5914 if ((unsigned int)poison <= 0xFF)
5915 memset(pos, poison, PAGE_SIZE);
5916 free_reserved_page(virt_to_page(pos));
5920 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5921 s, pages << (PAGE_SHIFT - 10), start, end);
5925 EXPORT_SYMBOL(free_reserved_area);
5927 #ifdef CONFIG_HIGHMEM
5928 void free_highmem_page(struct page *page)
5930 __free_reserved_page(page);
5932 page_zone(page)->managed_pages++;
5938 void __init mem_init_print_info(const char *str)
5940 unsigned long physpages, codesize, datasize, rosize, bss_size;
5941 unsigned long init_code_size, init_data_size;
5943 physpages = get_num_physpages();
5944 codesize = _etext - _stext;
5945 datasize = _edata - _sdata;
5946 rosize = __end_rodata - __start_rodata;
5947 bss_size = __bss_stop - __bss_start;
5948 init_data_size = __init_end - __init_begin;
5949 init_code_size = _einittext - _sinittext;
5952 * Detect special cases and adjust section sizes accordingly:
5953 * 1) .init.* may be embedded into .data sections
5954 * 2) .init.text.* may be out of [__init_begin, __init_end],
5955 * please refer to arch/tile/kernel/vmlinux.lds.S.
5956 * 3) .rodata.* may be embedded into .text or .data sections.
5958 #define adj_init_size(start, end, size, pos, adj) \
5960 if (start <= pos && pos < end && size > adj) \
5964 adj_init_size(__init_begin, __init_end, init_data_size,
5965 _sinittext, init_code_size);
5966 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5967 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5968 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5969 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5971 #undef adj_init_size
5973 pr_info("Memory: %luK/%luK available "
5974 "(%luK kernel code, %luK rwdata, %luK rodata, "
5975 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5976 #ifdef CONFIG_HIGHMEM
5980 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5981 codesize >> 10, datasize >> 10, rosize >> 10,
5982 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5983 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5984 totalcma_pages << (PAGE_SHIFT-10),
5985 #ifdef CONFIG_HIGHMEM
5986 totalhigh_pages << (PAGE_SHIFT-10),
5988 str ? ", " : "", str ? str : "");
5992 * set_dma_reserve - set the specified number of pages reserved in the first zone
5993 * @new_dma_reserve: The number of pages to mark reserved
5995 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5996 * In the DMA zone, a significant percentage may be consumed by kernel image
5997 * and other unfreeable allocations which can skew the watermarks badly. This
5998 * function may optionally be used to account for unfreeable pages in the
5999 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6000 * smaller per-cpu batchsize.
6002 void __init set_dma_reserve(unsigned long new_dma_reserve)
6004 dma_reserve = new_dma_reserve;
6007 void __init free_area_init(unsigned long *zones_size)
6009 free_area_init_node(0, zones_size,
6010 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6013 static int page_alloc_cpu_notify(struct notifier_block *self,
6014 unsigned long action, void *hcpu)
6016 int cpu = (unsigned long)hcpu;
6018 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6019 lru_add_drain_cpu(cpu);
6023 * Spill the event counters of the dead processor
6024 * into the current processors event counters.
6025 * This artificially elevates the count of the current
6028 vm_events_fold_cpu(cpu);
6031 * Zero the differential counters of the dead processor
6032 * so that the vm statistics are consistent.
6034 * This is only okay since the processor is dead and cannot
6035 * race with what we are doing.
6037 cpu_vm_stats_fold(cpu);
6042 void __init page_alloc_init(void)
6044 hotcpu_notifier(page_alloc_cpu_notify, 0);
6048 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
6049 * or min_free_kbytes changes.
6051 static void calculate_totalreserve_pages(void)
6053 struct pglist_data *pgdat;
6054 unsigned long reserve_pages = 0;
6055 enum zone_type i, j;
6057 for_each_online_pgdat(pgdat) {
6058 for (i = 0; i < MAX_NR_ZONES; i++) {
6059 struct zone *zone = pgdat->node_zones + i;
6062 /* Find valid and maximum lowmem_reserve in the zone */
6063 for (j = i; j < MAX_NR_ZONES; j++) {
6064 if (zone->lowmem_reserve[j] > max)
6065 max = zone->lowmem_reserve[j];
6068 /* we treat the high watermark as reserved pages. */
6069 max += high_wmark_pages(zone);
6071 if (max > zone->managed_pages)
6072 max = zone->managed_pages;
6073 reserve_pages += max;
6075 * Lowmem reserves are not available to
6076 * GFP_HIGHUSER page cache allocations and
6077 * kswapd tries to balance zones to their high
6078 * watermark. As a result, neither should be
6079 * regarded as dirtyable memory, to prevent a
6080 * situation where reclaim has to clean pages
6081 * in order to balance the zones.
6083 zone->dirty_balance_reserve = max;
6086 dirty_balance_reserve = reserve_pages;
6087 totalreserve_pages = reserve_pages;
6091 * setup_per_zone_lowmem_reserve - called whenever
6092 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
6093 * has a correct pages reserved value, so an adequate number of
6094 * pages are left in the zone after a successful __alloc_pages().
6096 static void setup_per_zone_lowmem_reserve(void)
6098 struct pglist_data *pgdat;
6099 enum zone_type j, idx;
6101 for_each_online_pgdat(pgdat) {
6102 for (j = 0; j < MAX_NR_ZONES; j++) {
6103 struct zone *zone = pgdat->node_zones + j;
6104 unsigned long managed_pages = zone->managed_pages;
6106 zone->lowmem_reserve[j] = 0;
6110 struct zone *lower_zone;
6114 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6115 sysctl_lowmem_reserve_ratio[idx] = 1;
6117 lower_zone = pgdat->node_zones + idx;
6118 lower_zone->lowmem_reserve[j] = managed_pages /
6119 sysctl_lowmem_reserve_ratio[idx];
6120 managed_pages += lower_zone->managed_pages;
6125 /* update totalreserve_pages */
6126 calculate_totalreserve_pages();
6129 static void __setup_per_zone_wmarks(void)
6131 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6132 unsigned long lowmem_pages = 0;
6134 unsigned long flags;
6136 /* Calculate total number of !ZONE_HIGHMEM pages */
6137 for_each_zone(zone) {
6138 if (!is_highmem(zone))
6139 lowmem_pages += zone->managed_pages;
6142 for_each_zone(zone) {
6145 spin_lock_irqsave(&zone->lock, flags);
6146 tmp = (u64)pages_min * zone->managed_pages;
6147 do_div(tmp, lowmem_pages);
6148 if (is_highmem(zone)) {
6150 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6151 * need highmem pages, so cap pages_min to a small
6154 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6155 * deltas control asynch page reclaim, and so should
6156 * not be capped for highmem.
6158 unsigned long min_pages;
6160 min_pages = zone->managed_pages / 1024;
6161 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6162 zone->watermark[WMARK_MIN] = min_pages;
6165 * If it's a lowmem zone, reserve a number of pages
6166 * proportionate to the zone's size.
6168 zone->watermark[WMARK_MIN] = tmp;
6171 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
6172 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
6174 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6175 high_wmark_pages(zone) - low_wmark_pages(zone) -
6176 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6178 setup_zone_migrate_reserve(zone);
6179 spin_unlock_irqrestore(&zone->lock, flags);
6182 /* update totalreserve_pages */
6183 calculate_totalreserve_pages();
6187 * setup_per_zone_wmarks - called when min_free_kbytes changes
6188 * or when memory is hot-{added|removed}
6190 * Ensures that the watermark[min,low,high] values for each zone are set
6191 * correctly with respect to min_free_kbytes.
6193 void setup_per_zone_wmarks(void)
6195 mutex_lock(&zonelists_mutex);
6196 __setup_per_zone_wmarks();
6197 mutex_unlock(&zonelists_mutex);
6201 * The inactive anon list should be small enough that the VM never has to
6202 * do too much work, but large enough that each inactive page has a chance
6203 * to be referenced again before it is swapped out.
6205 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6206 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6207 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6208 * the anonymous pages are kept on the inactive list.
6211 * memory ratio inactive anon
6212 * -------------------------------------
6221 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6223 unsigned int gb, ratio;
6225 /* Zone size in gigabytes */
6226 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6228 ratio = int_sqrt(10 * gb);
6232 zone->inactive_ratio = ratio;
6235 static void __meminit setup_per_zone_inactive_ratio(void)
6240 calculate_zone_inactive_ratio(zone);
6244 * Initialise min_free_kbytes.
6246 * For small machines we want it small (128k min). For large machines
6247 * we want it large (64MB max). But it is not linear, because network
6248 * bandwidth does not increase linearly with machine size. We use
6250 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6251 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6267 int __meminit init_per_zone_wmark_min(void)
6269 unsigned long lowmem_kbytes;
6270 int new_min_free_kbytes;
6272 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6273 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6275 if (new_min_free_kbytes > user_min_free_kbytes) {
6276 min_free_kbytes = new_min_free_kbytes;
6277 if (min_free_kbytes < 128)
6278 min_free_kbytes = 128;
6279 if (min_free_kbytes > 65536)
6280 min_free_kbytes = 65536;
6282 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6283 new_min_free_kbytes, user_min_free_kbytes);
6285 setup_per_zone_wmarks();
6286 refresh_zone_stat_thresholds();
6287 setup_per_zone_lowmem_reserve();
6288 setup_per_zone_inactive_ratio();
6291 module_init(init_per_zone_wmark_min)
6294 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6295 * that we can call two helper functions whenever min_free_kbytes
6298 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6299 void __user *buffer, size_t *length, loff_t *ppos)
6303 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6308 user_min_free_kbytes = min_free_kbytes;
6309 setup_per_zone_wmarks();
6315 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6316 void __user *buffer, size_t *length, loff_t *ppos)
6321 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6326 zone->min_unmapped_pages = (zone->managed_pages *
6327 sysctl_min_unmapped_ratio) / 100;
6331 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6332 void __user *buffer, size_t *length, loff_t *ppos)
6337 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6342 zone->min_slab_pages = (zone->managed_pages *
6343 sysctl_min_slab_ratio) / 100;
6349 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6350 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6351 * whenever sysctl_lowmem_reserve_ratio changes.
6353 * The reserve ratio obviously has absolutely no relation with the
6354 * minimum watermarks. The lowmem reserve ratio can only make sense
6355 * if in function of the boot time zone sizes.
6357 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6358 void __user *buffer, size_t *length, loff_t *ppos)
6360 proc_dointvec_minmax(table, write, buffer, length, ppos);
6361 setup_per_zone_lowmem_reserve();
6366 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6367 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6368 * pagelist can have before it gets flushed back to buddy allocator.
6370 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6371 void __user *buffer, size_t *length, loff_t *ppos)
6374 int old_percpu_pagelist_fraction;
6377 mutex_lock(&pcp_batch_high_lock);
6378 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6380 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6381 if (!write || ret < 0)
6384 /* Sanity checking to avoid pcp imbalance */
6385 if (percpu_pagelist_fraction &&
6386 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6387 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6393 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6396 for_each_populated_zone(zone) {
6399 for_each_possible_cpu(cpu)
6400 pageset_set_high_and_batch(zone,
6401 per_cpu_ptr(zone->pageset, cpu));
6404 mutex_unlock(&pcp_batch_high_lock);
6409 int hashdist = HASHDIST_DEFAULT;
6411 static int __init set_hashdist(char *str)
6415 hashdist = simple_strtoul(str, &str, 0);
6418 __setup("hashdist=", set_hashdist);
6422 * allocate a large system hash table from bootmem
6423 * - it is assumed that the hash table must contain an exact power-of-2
6424 * quantity of entries
6425 * - limit is the number of hash buckets, not the total allocation size
6427 void *__init alloc_large_system_hash(const char *tablename,
6428 unsigned long bucketsize,
6429 unsigned long numentries,
6432 unsigned int *_hash_shift,
6433 unsigned int *_hash_mask,
6434 unsigned long low_limit,
6435 unsigned long high_limit)
6437 unsigned long long max = high_limit;
6438 unsigned long log2qty, size;
6441 /* allow the kernel cmdline to have a say */
6443 /* round applicable memory size up to nearest megabyte */
6444 numentries = nr_kernel_pages;
6446 /* It isn't necessary when PAGE_SIZE >= 1MB */
6447 if (PAGE_SHIFT < 20)
6448 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6450 /* limit to 1 bucket per 2^scale bytes of low memory */
6451 if (scale > PAGE_SHIFT)
6452 numentries >>= (scale - PAGE_SHIFT);
6454 numentries <<= (PAGE_SHIFT - scale);
6456 /* Make sure we've got at least a 0-order allocation.. */
6457 if (unlikely(flags & HASH_SMALL)) {
6458 /* Makes no sense without HASH_EARLY */
6459 WARN_ON(!(flags & HASH_EARLY));
6460 if (!(numentries >> *_hash_shift)) {
6461 numentries = 1UL << *_hash_shift;
6462 BUG_ON(!numentries);
6464 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6465 numentries = PAGE_SIZE / bucketsize;
6467 numentries = roundup_pow_of_two(numentries);
6469 /* limit allocation size to 1/16 total memory by default */
6471 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6472 do_div(max, bucketsize);
6474 max = min(max, 0x80000000ULL);
6476 if (numentries < low_limit)
6477 numentries = low_limit;
6478 if (numentries > max)
6481 log2qty = ilog2(numentries);
6484 size = bucketsize << log2qty;
6485 if (flags & HASH_EARLY)
6486 table = memblock_virt_alloc_nopanic(size, 0);
6488 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6491 * If bucketsize is not a power-of-two, we may free
6492 * some pages at the end of hash table which
6493 * alloc_pages_exact() automatically does
6495 if (get_order(size) < MAX_ORDER) {
6496 table = alloc_pages_exact(size, GFP_ATOMIC);
6497 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6500 } while (!table && size > PAGE_SIZE && --log2qty);
6503 panic("Failed to allocate %s hash table\n", tablename);
6505 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6508 ilog2(size) - PAGE_SHIFT,
6512 *_hash_shift = log2qty;
6514 *_hash_mask = (1 << log2qty) - 1;
6519 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6520 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6523 #ifdef CONFIG_SPARSEMEM
6524 return __pfn_to_section(pfn)->pageblock_flags;
6526 return zone->pageblock_flags;
6527 #endif /* CONFIG_SPARSEMEM */
6530 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6532 #ifdef CONFIG_SPARSEMEM
6533 pfn &= (PAGES_PER_SECTION-1);
6534 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6536 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6537 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6538 #endif /* CONFIG_SPARSEMEM */
6542 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6543 * @page: The page within the block of interest
6544 * @pfn: The target page frame number
6545 * @end_bitidx: The last bit of interest to retrieve
6546 * @mask: mask of bits that the caller is interested in
6548 * Return: pageblock_bits flags
6550 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6551 unsigned long end_bitidx,
6555 unsigned long *bitmap;
6556 unsigned long bitidx, word_bitidx;
6559 zone = page_zone(page);
6560 bitmap = get_pageblock_bitmap(zone, pfn);
6561 bitidx = pfn_to_bitidx(zone, pfn);
6562 word_bitidx = bitidx / BITS_PER_LONG;
6563 bitidx &= (BITS_PER_LONG-1);
6565 word = bitmap[word_bitidx];
6566 bitidx += end_bitidx;
6567 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6571 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6572 * @page: The page within the block of interest
6573 * @flags: The flags to set
6574 * @pfn: The target page frame number
6575 * @end_bitidx: The last bit of interest
6576 * @mask: mask of bits that the caller is interested in
6578 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6580 unsigned long end_bitidx,
6584 unsigned long *bitmap;
6585 unsigned long bitidx, word_bitidx;
6586 unsigned long old_word, word;
6588 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6590 zone = page_zone(page);
6591 bitmap = get_pageblock_bitmap(zone, pfn);
6592 bitidx = pfn_to_bitidx(zone, pfn);
6593 word_bitidx = bitidx / BITS_PER_LONG;
6594 bitidx &= (BITS_PER_LONG-1);
6596 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6598 bitidx += end_bitidx;
6599 mask <<= (BITS_PER_LONG - bitidx - 1);
6600 flags <<= (BITS_PER_LONG - bitidx - 1);
6602 word = READ_ONCE(bitmap[word_bitidx]);
6604 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6605 if (word == old_word)
6612 * This function checks whether pageblock includes unmovable pages or not.
6613 * If @count is not zero, it is okay to include less @count unmovable pages
6615 * PageLRU check without isolation or lru_lock could race so that
6616 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6617 * expect this function should be exact.
6619 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6620 bool skip_hwpoisoned_pages)
6622 unsigned long pfn, iter, found;
6626 * For avoiding noise data, lru_add_drain_all() should be called
6627 * If ZONE_MOVABLE, the zone never contains unmovable pages
6629 if (zone_idx(zone) == ZONE_MOVABLE)
6631 mt = get_pageblock_migratetype(page);
6632 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6635 pfn = page_to_pfn(page);
6636 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6637 unsigned long check = pfn + iter;
6639 if (!pfn_valid_within(check))
6642 page = pfn_to_page(check);
6645 * Hugepages are not in LRU lists, but they're movable.
6646 * We need not scan over tail pages bacause we don't
6647 * handle each tail page individually in migration.
6649 if (PageHuge(page)) {
6650 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6655 * We can't use page_count without pin a page
6656 * because another CPU can free compound page.
6657 * This check already skips compound tails of THP
6658 * because their page->_count is zero at all time.
6660 if (!atomic_read(&page->_count)) {
6661 if (PageBuddy(page))
6662 iter += (1 << page_order(page)) - 1;
6667 * The HWPoisoned page may be not in buddy system, and
6668 * page_count() is not 0.
6670 if (skip_hwpoisoned_pages && PageHWPoison(page))
6676 * If there are RECLAIMABLE pages, we need to check
6677 * it. But now, memory offline itself doesn't call
6678 * shrink_node_slabs() and it still to be fixed.
6681 * If the page is not RAM, page_count()should be 0.
6682 * we don't need more check. This is an _used_ not-movable page.
6684 * The problematic thing here is PG_reserved pages. PG_reserved
6685 * is set to both of a memory hole page and a _used_ kernel
6694 bool is_pageblock_removable_nolock(struct page *page)
6700 * We have to be careful here because we are iterating over memory
6701 * sections which are not zone aware so we might end up outside of
6702 * the zone but still within the section.
6703 * We have to take care about the node as well. If the node is offline
6704 * its NODE_DATA will be NULL - see page_zone.
6706 if (!node_online(page_to_nid(page)))
6709 zone = page_zone(page);
6710 pfn = page_to_pfn(page);
6711 if (!zone_spans_pfn(zone, pfn))
6714 return !has_unmovable_pages(zone, page, 0, true);
6719 static unsigned long pfn_max_align_down(unsigned long pfn)
6721 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6722 pageblock_nr_pages) - 1);
6725 static unsigned long pfn_max_align_up(unsigned long pfn)
6727 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6728 pageblock_nr_pages));
6731 /* [start, end) must belong to a single zone. */
6732 static int __alloc_contig_migrate_range(struct compact_control *cc,
6733 unsigned long start, unsigned long end)
6735 /* This function is based on compact_zone() from compaction.c. */
6736 unsigned long nr_reclaimed;
6737 unsigned long pfn = start;
6738 unsigned int tries = 0;
6743 while (pfn < end || !list_empty(&cc->migratepages)) {
6744 if (fatal_signal_pending(current)) {
6749 if (list_empty(&cc->migratepages)) {
6750 cc->nr_migratepages = 0;
6751 pfn = isolate_migratepages_range(cc, pfn, end);
6757 } else if (++tries == 5) {
6758 ret = ret < 0 ? ret : -EBUSY;
6762 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6764 cc->nr_migratepages -= nr_reclaimed;
6766 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6767 NULL, 0, cc->mode, MR_CMA);
6770 putback_movable_pages(&cc->migratepages);
6777 * alloc_contig_range() -- tries to allocate given range of pages
6778 * @start: start PFN to allocate
6779 * @end: one-past-the-last PFN to allocate
6780 * @migratetype: migratetype of the underlaying pageblocks (either
6781 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6782 * in range must have the same migratetype and it must
6783 * be either of the two.
6785 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6786 * aligned, however it's the caller's responsibility to guarantee that
6787 * we are the only thread that changes migrate type of pageblocks the
6790 * The PFN range must belong to a single zone.
6792 * Returns zero on success or negative error code. On success all
6793 * pages which PFN is in [start, end) are allocated for the caller and
6794 * need to be freed with free_contig_range().
6796 int alloc_contig_range(unsigned long start, unsigned long end,
6797 unsigned migratetype)
6799 unsigned long outer_start, outer_end;
6802 struct compact_control cc = {
6803 .nr_migratepages = 0,
6805 .zone = page_zone(pfn_to_page(start)),
6806 .mode = MIGRATE_SYNC,
6807 .ignore_skip_hint = true,
6809 INIT_LIST_HEAD(&cc.migratepages);
6812 * What we do here is we mark all pageblocks in range as
6813 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6814 * have different sizes, and due to the way page allocator
6815 * work, we align the range to biggest of the two pages so
6816 * that page allocator won't try to merge buddies from
6817 * different pageblocks and change MIGRATE_ISOLATE to some
6818 * other migration type.
6820 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6821 * migrate the pages from an unaligned range (ie. pages that
6822 * we are interested in). This will put all the pages in
6823 * range back to page allocator as MIGRATE_ISOLATE.
6825 * When this is done, we take the pages in range from page
6826 * allocator removing them from the buddy system. This way
6827 * page allocator will never consider using them.
6829 * This lets us mark the pageblocks back as
6830 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6831 * aligned range but not in the unaligned, original range are
6832 * put back to page allocator so that buddy can use them.
6835 ret = start_isolate_page_range(pfn_max_align_down(start),
6836 pfn_max_align_up(end), migratetype,
6841 ret = __alloc_contig_migrate_range(&cc, start, end);
6846 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6847 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6848 * more, all pages in [start, end) are free in page allocator.
6849 * What we are going to do is to allocate all pages from
6850 * [start, end) (that is remove them from page allocator).
6852 * The only problem is that pages at the beginning and at the
6853 * end of interesting range may be not aligned with pages that
6854 * page allocator holds, ie. they can be part of higher order
6855 * pages. Because of this, we reserve the bigger range and
6856 * once this is done free the pages we are not interested in.
6858 * We don't have to hold zone->lock here because the pages are
6859 * isolated thus they won't get removed from buddy.
6862 lru_add_drain_all();
6863 drain_all_pages(cc.zone);
6866 outer_start = start;
6867 while (!PageBuddy(pfn_to_page(outer_start))) {
6868 if (++order >= MAX_ORDER) {
6872 outer_start &= ~0UL << order;
6875 /* Make sure the range is really isolated. */
6876 if (test_pages_isolated(outer_start, end, false)) {
6877 pr_info("%s: [%lx, %lx) PFNs busy\n",
6878 __func__, outer_start, end);
6883 /* Grab isolated pages from freelists. */
6884 outer_end = isolate_freepages_range(&cc, outer_start, end);
6890 /* Free head and tail (if any) */
6891 if (start != outer_start)
6892 free_contig_range(outer_start, start - outer_start);
6893 if (end != outer_end)
6894 free_contig_range(end, outer_end - end);
6897 undo_isolate_page_range(pfn_max_align_down(start),
6898 pfn_max_align_up(end), migratetype);
6902 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6904 unsigned int count = 0;
6906 for (; nr_pages--; pfn++) {
6907 struct page *page = pfn_to_page(pfn);
6909 count += page_count(page) != 1;
6912 WARN(count != 0, "%d pages are still in use!\n", count);
6916 #ifdef CONFIG_MEMORY_HOTPLUG
6918 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6919 * page high values need to be recalulated.
6921 void __meminit zone_pcp_update(struct zone *zone)
6924 mutex_lock(&pcp_batch_high_lock);
6925 for_each_possible_cpu(cpu)
6926 pageset_set_high_and_batch(zone,
6927 per_cpu_ptr(zone->pageset, cpu));
6928 mutex_unlock(&pcp_batch_high_lock);
6932 void zone_pcp_reset(struct zone *zone)
6934 unsigned long flags;
6936 struct per_cpu_pageset *pset;
6938 /* avoid races with drain_pages() */
6939 local_irq_save(flags);
6940 if (zone->pageset != &boot_pageset) {
6941 for_each_online_cpu(cpu) {
6942 pset = per_cpu_ptr(zone->pageset, cpu);
6943 drain_zonestat(zone, pset);
6945 free_percpu(zone->pageset);
6946 zone->pageset = &boot_pageset;
6948 local_irq_restore(flags);
6951 #ifdef CONFIG_MEMORY_HOTREMOVE
6953 * All pages in the range must be isolated before calling this.
6956 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6960 unsigned int order, i;
6962 unsigned long flags;
6963 /* find the first valid pfn */
6964 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6969 zone = page_zone(pfn_to_page(pfn));
6970 spin_lock_irqsave(&zone->lock, flags);
6972 while (pfn < end_pfn) {
6973 if (!pfn_valid(pfn)) {
6977 page = pfn_to_page(pfn);
6979 * The HWPoisoned page may be not in buddy system, and
6980 * page_count() is not 0.
6982 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6984 SetPageReserved(page);
6988 BUG_ON(page_count(page));
6989 BUG_ON(!PageBuddy(page));
6990 order = page_order(page);
6991 #ifdef CONFIG_DEBUG_VM
6992 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6993 pfn, 1 << order, end_pfn);
6995 list_del(&page->lru);
6996 rmv_page_order(page);
6997 zone->free_area[order].nr_free--;
6998 for (i = 0; i < (1 << order); i++)
6999 SetPageReserved((page+i));
7000 pfn += (1 << order);
7002 spin_unlock_irqrestore(&zone->lock, flags);
7006 #ifdef CONFIG_MEMORY_FAILURE
7007 bool is_free_buddy_page(struct page *page)
7009 struct zone *zone = page_zone(page);
7010 unsigned long pfn = page_to_pfn(page);
7011 unsigned long flags;
7014 spin_lock_irqsave(&zone->lock, flags);
7015 for (order = 0; order < MAX_ORDER; order++) {
7016 struct page *page_head = page - (pfn & ((1 << order) - 1));
7018 if (PageBuddy(page_head) && page_order(page_head) >= order)
7021 spin_unlock_irqrestore(&zone->lock, flags);
7023 return order < MAX_ORDER;