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);
792 mt = get_freepage_migratetype(page);
793 /* MIGRATE_ISOLATE page should not go to pcplists */
794 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
795 /* Pageblock could have been isolated meanwhile */
796 if (unlikely(has_isolate_pageblock(zone)))
797 mt = get_pageblock_migratetype(page);
799 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
800 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
801 trace_mm_page_pcpu_drain(page, 0, mt);
802 } while (--to_free && --batch_free && !list_empty(list));
804 spin_unlock(&zone->lock);
807 static void free_one_page(struct zone *zone,
808 struct page *page, unsigned long pfn,
812 unsigned long nr_scanned;
813 spin_lock(&zone->lock);
814 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
816 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
818 if (unlikely(has_isolate_pageblock(zone) ||
819 is_migrate_isolate(migratetype))) {
820 migratetype = get_pfnblock_migratetype(page, pfn);
822 __free_one_page(page, pfn, zone, order, migratetype);
823 spin_unlock(&zone->lock);
826 static int free_tail_pages_check(struct page *head_page, struct page *page)
828 if (!IS_ENABLED(CONFIG_DEBUG_VM))
830 if (unlikely(!PageTail(page))) {
831 bad_page(page, "PageTail not set", 0);
834 if (unlikely(page->first_page != head_page)) {
835 bad_page(page, "first_page not consistent", 0);
841 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
842 unsigned long zone, int nid)
844 set_page_links(page, zone, nid, pfn);
845 init_page_count(page);
846 page_mapcount_reset(page);
847 page_cpupid_reset_last(page);
849 INIT_LIST_HEAD(&page->lru);
850 #ifdef WANT_PAGE_VIRTUAL
851 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
852 if (!is_highmem_idx(zone))
853 set_page_address(page, __va(pfn << PAGE_SHIFT));
857 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
860 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
863 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
864 static void init_reserved_page(unsigned long pfn)
869 if (!early_page_uninitialised(pfn))
872 nid = early_pfn_to_nid(pfn);
873 pgdat = NODE_DATA(nid);
875 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
876 struct zone *zone = &pgdat->node_zones[zid];
878 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
881 __init_single_pfn(pfn, zid, nid);
884 static inline void init_reserved_page(unsigned long pfn)
887 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
890 * Initialised pages do not have PageReserved set. This function is
891 * called for each range allocated by the bootmem allocator and
892 * marks the pages PageReserved. The remaining valid pages are later
893 * sent to the buddy page allocator.
895 void __meminit reserve_bootmem_region(unsigned long start, unsigned long end)
897 unsigned long start_pfn = PFN_DOWN(start);
898 unsigned long end_pfn = PFN_UP(end);
900 for (; start_pfn < end_pfn; start_pfn++) {
901 if (pfn_valid(start_pfn)) {
902 struct page *page = pfn_to_page(start_pfn);
904 init_reserved_page(start_pfn);
905 SetPageReserved(page);
910 static bool free_pages_prepare(struct page *page, unsigned int order)
912 bool compound = PageCompound(page);
915 VM_BUG_ON_PAGE(PageTail(page), page);
916 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
918 trace_mm_page_free(page, order);
919 kmemcheck_free_shadow(page, order);
920 kasan_free_pages(page, order);
923 page->mapping = NULL;
924 bad += free_pages_check(page);
925 for (i = 1; i < (1 << order); i++) {
927 bad += free_tail_pages_check(page, page + i);
928 bad += free_pages_check(page + i);
933 reset_page_owner(page, order);
935 if (!PageHighMem(page)) {
936 debug_check_no_locks_freed(page_address(page),
938 debug_check_no_obj_freed(page_address(page),
941 arch_free_page(page, order);
942 kernel_map_pages(page, 1 << order, 0);
947 static void __free_pages_ok(struct page *page, unsigned int order)
951 unsigned long pfn = page_to_pfn(page);
953 if (!free_pages_prepare(page, order))
956 migratetype = get_pfnblock_migratetype(page, pfn);
957 local_irq_save(flags);
958 __count_vm_events(PGFREE, 1 << order);
959 set_freepage_migratetype(page, migratetype);
960 free_one_page(page_zone(page), page, pfn, order, migratetype);
961 local_irq_restore(flags);
964 static void __init __free_pages_boot_core(struct page *page,
965 unsigned long pfn, unsigned int order)
967 unsigned int nr_pages = 1 << order;
968 struct page *p = page;
972 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
974 __ClearPageReserved(p);
975 set_page_count(p, 0);
977 __ClearPageReserved(p);
978 set_page_count(p, 0);
980 page_zone(page)->managed_pages += nr_pages;
981 set_page_refcounted(page);
982 __free_pages(page, order);
985 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
986 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
988 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
990 int __meminit early_pfn_to_nid(unsigned long pfn)
992 static DEFINE_SPINLOCK(early_pfn_lock);
995 spin_lock(&early_pfn_lock);
996 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
999 spin_unlock(&early_pfn_lock);
1005 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1006 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1007 struct mminit_pfnnid_cache *state)
1011 nid = __early_pfn_to_nid(pfn, state);
1012 if (nid >= 0 && nid != node)
1017 /* Only safe to use early in boot when initialisation is single-threaded */
1018 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1020 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1025 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1029 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1030 struct mminit_pfnnid_cache *state)
1037 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1040 if (early_page_uninitialised(pfn))
1042 return __free_pages_boot_core(page, pfn, order);
1045 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1046 static void __init deferred_free_range(struct page *page,
1047 unsigned long pfn, int nr_pages)
1054 /* Free a large naturally-aligned chunk if possible */
1055 if (nr_pages == MAX_ORDER_NR_PAGES &&
1056 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1057 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1058 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1062 for (i = 0; i < nr_pages; i++, page++, pfn++)
1063 __free_pages_boot_core(page, pfn, 0);
1066 /* Completion tracking for deferred_init_memmap() threads */
1067 static atomic_t pgdat_init_n_undone __initdata;
1068 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1070 static inline void __init pgdat_init_report_one_done(void)
1072 if (atomic_dec_and_test(&pgdat_init_n_undone))
1073 complete(&pgdat_init_all_done_comp);
1076 /* Initialise remaining memory on a node */
1077 static int __init deferred_init_memmap(void *data)
1079 pg_data_t *pgdat = data;
1080 int nid = pgdat->node_id;
1081 struct mminit_pfnnid_cache nid_init_state = { };
1082 unsigned long start = jiffies;
1083 unsigned long nr_pages = 0;
1084 unsigned long walk_start, walk_end;
1087 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1088 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1090 if (first_init_pfn == ULONG_MAX) {
1091 pgdat_init_report_one_done();
1095 /* Bind memory initialisation thread to a local node if possible */
1096 if (!cpumask_empty(cpumask))
1097 set_cpus_allowed_ptr(current, cpumask);
1099 /* Sanity check boundaries */
1100 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1101 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1102 pgdat->first_deferred_pfn = ULONG_MAX;
1104 /* Only the highest zone is deferred so find it */
1105 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1106 zone = pgdat->node_zones + zid;
1107 if (first_init_pfn < zone_end_pfn(zone))
1111 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1112 unsigned long pfn, end_pfn;
1113 struct page *page = NULL;
1114 struct page *free_base_page = NULL;
1115 unsigned long free_base_pfn = 0;
1118 end_pfn = min(walk_end, zone_end_pfn(zone));
1119 pfn = first_init_pfn;
1120 if (pfn < walk_start)
1122 if (pfn < zone->zone_start_pfn)
1123 pfn = zone->zone_start_pfn;
1125 for (; pfn < end_pfn; pfn++) {
1126 if (!pfn_valid_within(pfn))
1130 * Ensure pfn_valid is checked every
1131 * MAX_ORDER_NR_PAGES for memory holes
1133 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1134 if (!pfn_valid(pfn)) {
1140 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1145 /* Minimise pfn page lookups and scheduler checks */
1146 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1149 nr_pages += nr_to_free;
1150 deferred_free_range(free_base_page,
1151 free_base_pfn, nr_to_free);
1152 free_base_page = NULL;
1153 free_base_pfn = nr_to_free = 0;
1155 page = pfn_to_page(pfn);
1160 VM_BUG_ON(page_zone(page) != zone);
1164 __init_single_page(page, pfn, zid, nid);
1165 if (!free_base_page) {
1166 free_base_page = page;
1167 free_base_pfn = pfn;
1172 /* Where possible, batch up pages for a single free */
1175 /* Free the current block of pages to allocator */
1176 nr_pages += nr_to_free;
1177 deferred_free_range(free_base_page, free_base_pfn,
1179 free_base_page = NULL;
1180 free_base_pfn = nr_to_free = 0;
1183 first_init_pfn = max(end_pfn, first_init_pfn);
1186 /* Sanity check that the next zone really is unpopulated */
1187 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1189 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1190 jiffies_to_msecs(jiffies - start));
1192 pgdat_init_report_one_done();
1196 void __init page_alloc_init_late(void)
1200 /* There will be num_node_state(N_MEMORY) threads */
1201 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1202 for_each_node_state(nid, N_MEMORY) {
1203 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1206 /* Block until all are initialised */
1207 wait_for_completion(&pgdat_init_all_done_comp);
1209 /* Reinit limits that are based on free pages after the kernel is up */
1210 files_maxfiles_init();
1212 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1215 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1216 void __init init_cma_reserved_pageblock(struct page *page)
1218 unsigned i = pageblock_nr_pages;
1219 struct page *p = page;
1222 __ClearPageReserved(p);
1223 set_page_count(p, 0);
1226 set_pageblock_migratetype(page, MIGRATE_CMA);
1228 if (pageblock_order >= MAX_ORDER) {
1229 i = pageblock_nr_pages;
1232 set_page_refcounted(p);
1233 __free_pages(p, MAX_ORDER - 1);
1234 p += MAX_ORDER_NR_PAGES;
1235 } while (i -= MAX_ORDER_NR_PAGES);
1237 set_page_refcounted(page);
1238 __free_pages(page, pageblock_order);
1241 adjust_managed_page_count(page, pageblock_nr_pages);
1246 * The order of subdivision here is critical for the IO subsystem.
1247 * Please do not alter this order without good reasons and regression
1248 * testing. Specifically, as large blocks of memory are subdivided,
1249 * the order in which smaller blocks are delivered depends on the order
1250 * they're subdivided in this function. This is the primary factor
1251 * influencing the order in which pages are delivered to the IO
1252 * subsystem according to empirical testing, and this is also justified
1253 * by considering the behavior of a buddy system containing a single
1254 * large block of memory acted on by a series of small allocations.
1255 * This behavior is a critical factor in sglist merging's success.
1259 static inline void expand(struct zone *zone, struct page *page,
1260 int low, int high, struct free_area *area,
1263 unsigned long size = 1 << high;
1265 while (high > low) {
1269 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1271 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1272 debug_guardpage_enabled() &&
1273 high < debug_guardpage_minorder()) {
1275 * Mark as guard pages (or page), that will allow to
1276 * merge back to allocator when buddy will be freed.
1277 * Corresponding page table entries will not be touched,
1278 * pages will stay not present in virtual address space
1280 set_page_guard(zone, &page[size], high, migratetype);
1283 list_add(&page[size].lru, &area->free_list[migratetype]);
1285 set_page_order(&page[size], high);
1290 * This page is about to be returned from the page allocator
1292 static inline int check_new_page(struct page *page)
1294 const char *bad_reason = NULL;
1295 unsigned long bad_flags = 0;
1297 if (unlikely(page_mapcount(page)))
1298 bad_reason = "nonzero mapcount";
1299 if (unlikely(page->mapping != NULL))
1300 bad_reason = "non-NULL mapping";
1301 if (unlikely(atomic_read(&page->_count) != 0))
1302 bad_reason = "nonzero _count";
1303 if (unlikely(page->flags & __PG_HWPOISON)) {
1304 bad_reason = "HWPoisoned (hardware-corrupted)";
1305 bad_flags = __PG_HWPOISON;
1307 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1308 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1309 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1312 if (unlikely(page->mem_cgroup))
1313 bad_reason = "page still charged to cgroup";
1315 if (unlikely(bad_reason)) {
1316 bad_page(page, bad_reason, bad_flags);
1322 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1327 for (i = 0; i < (1 << order); i++) {
1328 struct page *p = page + i;
1329 if (unlikely(check_new_page(p)))
1333 set_page_private(page, 0);
1334 set_page_refcounted(page);
1336 arch_alloc_page(page, order);
1337 kernel_map_pages(page, 1 << order, 1);
1338 kasan_alloc_pages(page, order);
1340 if (gfp_flags & __GFP_ZERO)
1341 for (i = 0; i < (1 << order); i++)
1342 clear_highpage(page + i);
1344 if (order && (gfp_flags & __GFP_COMP))
1345 prep_compound_page(page, order);
1347 set_page_owner(page, order, gfp_flags);
1350 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1351 * allocate the page. The expectation is that the caller is taking
1352 * steps that will free more memory. The caller should avoid the page
1353 * being used for !PFMEMALLOC purposes.
1355 if (alloc_flags & ALLOC_NO_WATERMARKS)
1356 set_page_pfmemalloc(page);
1358 clear_page_pfmemalloc(page);
1364 * Go through the free lists for the given migratetype and remove
1365 * the smallest available page from the freelists
1368 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1371 unsigned int current_order;
1372 struct free_area *area;
1375 /* Find a page of the appropriate size in the preferred list */
1376 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1377 area = &(zone->free_area[current_order]);
1378 if (list_empty(&area->free_list[migratetype]))
1381 page = list_entry(area->free_list[migratetype].next,
1383 list_del(&page->lru);
1384 rmv_page_order(page);
1386 expand(zone, page, order, current_order, area, migratetype);
1387 set_freepage_migratetype(page, migratetype);
1396 * This array describes the order lists are fallen back to when
1397 * the free lists for the desirable migrate type are depleted
1399 static int fallbacks[MIGRATE_TYPES][4] = {
1400 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1401 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1402 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
1404 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
1406 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
1407 #ifdef CONFIG_MEMORY_ISOLATION
1408 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
1413 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1416 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1419 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1420 unsigned int order) { return NULL; }
1424 * Move the free pages in a range to the free lists of the requested type.
1425 * Note that start_page and end_pages are not aligned on a pageblock
1426 * boundary. If alignment is required, use move_freepages_block()
1428 int move_freepages(struct zone *zone,
1429 struct page *start_page, struct page *end_page,
1433 unsigned long order;
1434 int pages_moved = 0;
1436 #ifndef CONFIG_HOLES_IN_ZONE
1438 * page_zone is not safe to call in this context when
1439 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1440 * anyway as we check zone boundaries in move_freepages_block().
1441 * Remove at a later date when no bug reports exist related to
1442 * grouping pages by mobility
1444 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1447 for (page = start_page; page <= end_page;) {
1448 /* Make sure we are not inadvertently changing nodes */
1449 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1451 if (!pfn_valid_within(page_to_pfn(page))) {
1456 if (!PageBuddy(page)) {
1461 order = page_order(page);
1462 list_move(&page->lru,
1463 &zone->free_area[order].free_list[migratetype]);
1464 set_freepage_migratetype(page, migratetype);
1466 pages_moved += 1 << order;
1472 int move_freepages_block(struct zone *zone, struct page *page,
1475 unsigned long start_pfn, end_pfn;
1476 struct page *start_page, *end_page;
1478 start_pfn = page_to_pfn(page);
1479 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1480 start_page = pfn_to_page(start_pfn);
1481 end_page = start_page + pageblock_nr_pages - 1;
1482 end_pfn = start_pfn + pageblock_nr_pages - 1;
1484 /* Do not cross zone boundaries */
1485 if (!zone_spans_pfn(zone, start_pfn))
1487 if (!zone_spans_pfn(zone, end_pfn))
1490 return move_freepages(zone, start_page, end_page, migratetype);
1493 static void change_pageblock_range(struct page *pageblock_page,
1494 int start_order, int migratetype)
1496 int nr_pageblocks = 1 << (start_order - pageblock_order);
1498 while (nr_pageblocks--) {
1499 set_pageblock_migratetype(pageblock_page, migratetype);
1500 pageblock_page += pageblock_nr_pages;
1505 * When we are falling back to another migratetype during allocation, try to
1506 * steal extra free pages from the same pageblocks to satisfy further
1507 * allocations, instead of polluting multiple pageblocks.
1509 * If we are stealing a relatively large buddy page, it is likely there will
1510 * be more free pages in the pageblock, so try to steal them all. For
1511 * reclaimable and unmovable allocations, we steal regardless of page size,
1512 * as fragmentation caused by those allocations polluting movable pageblocks
1513 * is worse than movable allocations stealing from unmovable and reclaimable
1516 static bool can_steal_fallback(unsigned int order, int start_mt)
1519 * Leaving this order check is intended, although there is
1520 * relaxed order check in next check. The reason is that
1521 * we can actually steal whole pageblock if this condition met,
1522 * but, below check doesn't guarantee it and that is just heuristic
1523 * so could be changed anytime.
1525 if (order >= pageblock_order)
1528 if (order >= pageblock_order / 2 ||
1529 start_mt == MIGRATE_RECLAIMABLE ||
1530 start_mt == MIGRATE_UNMOVABLE ||
1531 page_group_by_mobility_disabled)
1538 * This function implements actual steal behaviour. If order is large enough,
1539 * we can steal whole pageblock. If not, we first move freepages in this
1540 * pageblock and check whether half of pages are moved or not. If half of
1541 * pages are moved, we can change migratetype of pageblock and permanently
1542 * use it's pages as requested migratetype in the future.
1544 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1547 int current_order = page_order(page);
1550 /* Take ownership for orders >= pageblock_order */
1551 if (current_order >= pageblock_order) {
1552 change_pageblock_range(page, current_order, start_type);
1556 pages = move_freepages_block(zone, page, start_type);
1558 /* Claim the whole block if over half of it is free */
1559 if (pages >= (1 << (pageblock_order-1)) ||
1560 page_group_by_mobility_disabled)
1561 set_pageblock_migratetype(page, start_type);
1565 * Check whether there is a suitable fallback freepage with requested order.
1566 * If only_stealable is true, this function returns fallback_mt only if
1567 * we can steal other freepages all together. This would help to reduce
1568 * fragmentation due to mixed migratetype pages in one pageblock.
1570 int find_suitable_fallback(struct free_area *area, unsigned int order,
1571 int migratetype, bool only_stealable, bool *can_steal)
1576 if (area->nr_free == 0)
1581 fallback_mt = fallbacks[migratetype][i];
1582 if (fallback_mt == MIGRATE_RESERVE)
1585 if (list_empty(&area->free_list[fallback_mt]))
1588 if (can_steal_fallback(order, migratetype))
1591 if (!only_stealable)
1601 /* Remove an element from the buddy allocator from the fallback list */
1602 static inline struct page *
1603 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1605 struct free_area *area;
1606 unsigned int current_order;
1611 /* Find the largest possible block of pages in the other list */
1612 for (current_order = MAX_ORDER-1;
1613 current_order >= order && current_order <= MAX_ORDER-1;
1615 area = &(zone->free_area[current_order]);
1616 fallback_mt = find_suitable_fallback(area, current_order,
1617 start_migratetype, false, &can_steal);
1618 if (fallback_mt == -1)
1621 page = list_entry(area->free_list[fallback_mt].next,
1624 steal_suitable_fallback(zone, page, start_migratetype);
1626 /* Remove the page from the freelists */
1628 list_del(&page->lru);
1629 rmv_page_order(page);
1631 expand(zone, page, order, current_order, area,
1634 * The freepage_migratetype may differ from pageblock's
1635 * migratetype depending on the decisions in
1636 * try_to_steal_freepages(). This is OK as long as it
1637 * does not differ for MIGRATE_CMA pageblocks. For CMA
1638 * we need to make sure unallocated pages flushed from
1639 * pcp lists are returned to the correct freelist.
1641 set_freepage_migratetype(page, start_migratetype);
1643 trace_mm_page_alloc_extfrag(page, order, current_order,
1644 start_migratetype, fallback_mt);
1653 * Do the hard work of removing an element from the buddy allocator.
1654 * Call me with the zone->lock already held.
1656 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1662 page = __rmqueue_smallest(zone, order, migratetype);
1664 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1665 if (migratetype == MIGRATE_MOVABLE)
1666 page = __rmqueue_cma_fallback(zone, order);
1669 page = __rmqueue_fallback(zone, order, migratetype);
1672 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1673 * is used because __rmqueue_smallest is an inline function
1674 * and we want just one call site
1677 migratetype = MIGRATE_RESERVE;
1682 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1687 * Obtain a specified number of elements from the buddy allocator, all under
1688 * a single hold of the lock, for efficiency. Add them to the supplied list.
1689 * Returns the number of new pages which were placed at *list.
1691 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1692 unsigned long count, struct list_head *list,
1693 int migratetype, bool cold)
1697 spin_lock(&zone->lock);
1698 for (i = 0; i < count; ++i) {
1699 struct page *page = __rmqueue(zone, order, migratetype);
1700 if (unlikely(page == NULL))
1704 * Split buddy pages returned by expand() are received here
1705 * in physical page order. The page is added to the callers and
1706 * list and the list head then moves forward. From the callers
1707 * perspective, the linked list is ordered by page number in
1708 * some conditions. This is useful for IO devices that can
1709 * merge IO requests if the physical pages are ordered
1713 list_add(&page->lru, list);
1715 list_add_tail(&page->lru, list);
1717 if (is_migrate_cma(get_freepage_migratetype(page)))
1718 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1721 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1722 spin_unlock(&zone->lock);
1728 * Called from the vmstat counter updater to drain pagesets of this
1729 * currently executing processor on remote nodes after they have
1732 * Note that this function must be called with the thread pinned to
1733 * a single processor.
1735 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1737 unsigned long flags;
1738 int to_drain, batch;
1740 local_irq_save(flags);
1741 batch = READ_ONCE(pcp->batch);
1742 to_drain = min(pcp->count, batch);
1744 free_pcppages_bulk(zone, to_drain, pcp);
1745 pcp->count -= to_drain;
1747 local_irq_restore(flags);
1752 * Drain pcplists of the indicated processor and zone.
1754 * The processor must either be the current processor and the
1755 * thread pinned to the current processor or a processor that
1758 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1760 unsigned long flags;
1761 struct per_cpu_pageset *pset;
1762 struct per_cpu_pages *pcp;
1764 local_irq_save(flags);
1765 pset = per_cpu_ptr(zone->pageset, cpu);
1769 free_pcppages_bulk(zone, pcp->count, pcp);
1772 local_irq_restore(flags);
1776 * Drain pcplists of all zones on the indicated processor.
1778 * The processor must either be the current processor and the
1779 * thread pinned to the current processor or a processor that
1782 static void drain_pages(unsigned int cpu)
1786 for_each_populated_zone(zone) {
1787 drain_pages_zone(cpu, zone);
1792 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1794 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1795 * the single zone's pages.
1797 void drain_local_pages(struct zone *zone)
1799 int cpu = smp_processor_id();
1802 drain_pages_zone(cpu, zone);
1808 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1810 * When zone parameter is non-NULL, spill just the single zone's pages.
1812 * Note that this code is protected against sending an IPI to an offline
1813 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1814 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1815 * nothing keeps CPUs from showing up after we populated the cpumask and
1816 * before the call to on_each_cpu_mask().
1818 void drain_all_pages(struct zone *zone)
1823 * Allocate in the BSS so we wont require allocation in
1824 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1826 static cpumask_t cpus_with_pcps;
1829 * We don't care about racing with CPU hotplug event
1830 * as offline notification will cause the notified
1831 * cpu to drain that CPU pcps and on_each_cpu_mask
1832 * disables preemption as part of its processing
1834 for_each_online_cpu(cpu) {
1835 struct per_cpu_pageset *pcp;
1837 bool has_pcps = false;
1840 pcp = per_cpu_ptr(zone->pageset, cpu);
1844 for_each_populated_zone(z) {
1845 pcp = per_cpu_ptr(z->pageset, cpu);
1846 if (pcp->pcp.count) {
1854 cpumask_set_cpu(cpu, &cpus_with_pcps);
1856 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1858 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
1862 #ifdef CONFIG_HIBERNATION
1864 void mark_free_pages(struct zone *zone)
1866 unsigned long pfn, max_zone_pfn;
1867 unsigned long flags;
1868 unsigned int order, t;
1869 struct list_head *curr;
1871 if (zone_is_empty(zone))
1874 spin_lock_irqsave(&zone->lock, flags);
1876 max_zone_pfn = zone_end_pfn(zone);
1877 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1878 if (pfn_valid(pfn)) {
1879 struct page *page = pfn_to_page(pfn);
1881 if (!swsusp_page_is_forbidden(page))
1882 swsusp_unset_page_free(page);
1885 for_each_migratetype_order(order, t) {
1886 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1889 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1890 for (i = 0; i < (1UL << order); i++)
1891 swsusp_set_page_free(pfn_to_page(pfn + i));
1894 spin_unlock_irqrestore(&zone->lock, flags);
1896 #endif /* CONFIG_PM */
1899 * Free a 0-order page
1900 * cold == true ? free a cold page : free a hot page
1902 void free_hot_cold_page(struct page *page, bool cold)
1904 struct zone *zone = page_zone(page);
1905 struct per_cpu_pages *pcp;
1906 unsigned long flags;
1907 unsigned long pfn = page_to_pfn(page);
1910 if (!free_pages_prepare(page, 0))
1913 migratetype = get_pfnblock_migratetype(page, pfn);
1914 set_freepage_migratetype(page, migratetype);
1915 local_irq_save(flags);
1916 __count_vm_event(PGFREE);
1919 * We only track unmovable, reclaimable and movable on pcp lists.
1920 * Free ISOLATE pages back to the allocator because they are being
1921 * offlined but treat RESERVE as movable pages so we can get those
1922 * areas back if necessary. Otherwise, we may have to free
1923 * excessively into the page allocator
1925 if (migratetype >= MIGRATE_PCPTYPES) {
1926 if (unlikely(is_migrate_isolate(migratetype))) {
1927 free_one_page(zone, page, pfn, 0, migratetype);
1930 migratetype = MIGRATE_MOVABLE;
1933 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1935 list_add(&page->lru, &pcp->lists[migratetype]);
1937 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1939 if (pcp->count >= pcp->high) {
1940 unsigned long batch = READ_ONCE(pcp->batch);
1941 free_pcppages_bulk(zone, batch, pcp);
1942 pcp->count -= batch;
1946 local_irq_restore(flags);
1950 * Free a list of 0-order pages
1952 void free_hot_cold_page_list(struct list_head *list, bool cold)
1954 struct page *page, *next;
1956 list_for_each_entry_safe(page, next, list, lru) {
1957 trace_mm_page_free_batched(page, cold);
1958 free_hot_cold_page(page, cold);
1963 * split_page takes a non-compound higher-order page, and splits it into
1964 * n (1<<order) sub-pages: page[0..n]
1965 * Each sub-page must be freed individually.
1967 * Note: this is probably too low level an operation for use in drivers.
1968 * Please consult with lkml before using this in your driver.
1970 void split_page(struct page *page, unsigned int order)
1975 VM_BUG_ON_PAGE(PageCompound(page), page);
1976 VM_BUG_ON_PAGE(!page_count(page), page);
1978 #ifdef CONFIG_KMEMCHECK
1980 * Split shadow pages too, because free(page[0]) would
1981 * otherwise free the whole shadow.
1983 if (kmemcheck_page_is_tracked(page))
1984 split_page(virt_to_page(page[0].shadow), order);
1987 gfp_mask = get_page_owner_gfp(page);
1988 set_page_owner(page, 0, gfp_mask);
1989 for (i = 1; i < (1 << order); i++) {
1990 set_page_refcounted(page + i);
1991 set_page_owner(page + i, 0, gfp_mask);
1994 EXPORT_SYMBOL_GPL(split_page);
1996 int __isolate_free_page(struct page *page, unsigned int order)
1998 unsigned long watermark;
2002 BUG_ON(!PageBuddy(page));
2004 zone = page_zone(page);
2005 mt = get_pageblock_migratetype(page);
2007 if (!is_migrate_isolate(mt)) {
2008 /* Obey watermarks as if the page was being allocated */
2009 watermark = low_wmark_pages(zone) + (1 << order);
2010 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2013 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2016 /* Remove page from free list */
2017 list_del(&page->lru);
2018 zone->free_area[order].nr_free--;
2019 rmv_page_order(page);
2021 set_page_owner(page, order, __GFP_MOVABLE);
2023 /* Set the pageblock if the isolated page is at least a pageblock */
2024 if (order >= pageblock_order - 1) {
2025 struct page *endpage = page + (1 << order) - 1;
2026 for (; page < endpage; page += pageblock_nr_pages) {
2027 int mt = get_pageblock_migratetype(page);
2028 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2029 set_pageblock_migratetype(page,
2035 return 1UL << order;
2039 * Similar to split_page except the page is already free. As this is only
2040 * being used for migration, the migratetype of the block also changes.
2041 * As this is called with interrupts disabled, the caller is responsible
2042 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2045 * Note: this is probably too low level an operation for use in drivers.
2046 * Please consult with lkml before using this in your driver.
2048 int split_free_page(struct page *page)
2053 order = page_order(page);
2055 nr_pages = __isolate_free_page(page, order);
2059 /* Split into individual pages */
2060 set_page_refcounted(page);
2061 split_page(page, order);
2066 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2069 struct page *buffered_rmqueue(struct zone *preferred_zone,
2070 struct zone *zone, unsigned int order,
2071 gfp_t gfp_flags, int migratetype)
2073 unsigned long flags;
2075 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2077 if (likely(order == 0)) {
2078 struct per_cpu_pages *pcp;
2079 struct list_head *list;
2081 local_irq_save(flags);
2082 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2083 list = &pcp->lists[migratetype];
2084 if (list_empty(list)) {
2085 pcp->count += rmqueue_bulk(zone, 0,
2088 if (unlikely(list_empty(list)))
2093 page = list_entry(list->prev, struct page, lru);
2095 page = list_entry(list->next, struct page, lru);
2097 list_del(&page->lru);
2100 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2102 * __GFP_NOFAIL is not to be used in new code.
2104 * All __GFP_NOFAIL callers should be fixed so that they
2105 * properly detect and handle allocation failures.
2107 * We most definitely don't want callers attempting to
2108 * allocate greater than order-1 page units with
2111 WARN_ON_ONCE(order > 1);
2113 spin_lock_irqsave(&zone->lock, flags);
2114 page = __rmqueue(zone, order, migratetype);
2115 spin_unlock(&zone->lock);
2118 __mod_zone_freepage_state(zone, -(1 << order),
2119 get_freepage_migratetype(page));
2122 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2123 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2124 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2125 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2127 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2128 zone_statistics(preferred_zone, zone, gfp_flags);
2129 local_irq_restore(flags);
2131 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2135 local_irq_restore(flags);
2139 #ifdef CONFIG_FAIL_PAGE_ALLOC
2142 struct fault_attr attr;
2144 u32 ignore_gfp_highmem;
2145 u32 ignore_gfp_wait;
2147 } fail_page_alloc = {
2148 .attr = FAULT_ATTR_INITIALIZER,
2149 .ignore_gfp_wait = 1,
2150 .ignore_gfp_highmem = 1,
2154 static int __init setup_fail_page_alloc(char *str)
2156 return setup_fault_attr(&fail_page_alloc.attr, str);
2158 __setup("fail_page_alloc=", setup_fail_page_alloc);
2160 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2162 if (order < fail_page_alloc.min_order)
2164 if (gfp_mask & __GFP_NOFAIL)
2166 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2168 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
2171 return should_fail(&fail_page_alloc.attr, 1 << order);
2174 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2176 static int __init fail_page_alloc_debugfs(void)
2178 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2181 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2182 &fail_page_alloc.attr);
2184 return PTR_ERR(dir);
2186 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2187 &fail_page_alloc.ignore_gfp_wait))
2189 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2190 &fail_page_alloc.ignore_gfp_highmem))
2192 if (!debugfs_create_u32("min-order", mode, dir,
2193 &fail_page_alloc.min_order))
2198 debugfs_remove_recursive(dir);
2203 late_initcall(fail_page_alloc_debugfs);
2205 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2207 #else /* CONFIG_FAIL_PAGE_ALLOC */
2209 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2214 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2217 * Return true if free pages are above 'mark'. This takes into account the order
2218 * of the allocation.
2220 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2221 unsigned long mark, int classzone_idx, int alloc_flags,
2224 /* free_pages may go negative - that's OK */
2229 free_pages -= (1 << order) - 1;
2230 if (alloc_flags & ALLOC_HIGH)
2232 if (alloc_flags & ALLOC_HARDER)
2235 /* If allocation can't use CMA areas don't use free CMA pages */
2236 if (!(alloc_flags & ALLOC_CMA))
2237 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
2240 if (free_pages - free_cma <= min + z->lowmem_reserve[classzone_idx])
2242 for (o = 0; o < order; o++) {
2243 /* At the next order, this order's pages become unavailable */
2244 free_pages -= z->free_area[o].nr_free << o;
2246 /* Require fewer higher order pages to be free */
2249 if (free_pages <= min)
2255 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2256 int classzone_idx, int alloc_flags)
2258 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2259 zone_page_state(z, NR_FREE_PAGES));
2262 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2263 unsigned long mark, int classzone_idx, int alloc_flags)
2265 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2267 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2268 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2270 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2276 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
2277 * skip over zones that are not allowed by the cpuset, or that have
2278 * been recently (in last second) found to be nearly full. See further
2279 * comments in mmzone.h. Reduces cache footprint of zonelist scans
2280 * that have to skip over a lot of full or unallowed zones.
2282 * If the zonelist cache is present in the passed zonelist, then
2283 * returns a pointer to the allowed node mask (either the current
2284 * tasks mems_allowed, or node_states[N_MEMORY].)
2286 * If the zonelist cache is not available for this zonelist, does
2287 * nothing and returns NULL.
2289 * If the fullzones BITMAP in the zonelist cache is stale (more than
2290 * a second since last zap'd) then we zap it out (clear its bits.)
2292 * We hold off even calling zlc_setup, until after we've checked the
2293 * first zone in the zonelist, on the theory that most allocations will
2294 * be satisfied from that first zone, so best to examine that zone as
2295 * quickly as we can.
2297 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
2299 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2300 nodemask_t *allowednodes; /* zonelist_cache approximation */
2302 zlc = zonelist->zlcache_ptr;
2306 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
2307 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2308 zlc->last_full_zap = jiffies;
2311 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
2312 &cpuset_current_mems_allowed :
2313 &node_states[N_MEMORY];
2314 return allowednodes;
2318 * Given 'z' scanning a zonelist, run a couple of quick checks to see
2319 * if it is worth looking at further for free memory:
2320 * 1) Check that the zone isn't thought to be full (doesn't have its
2321 * bit set in the zonelist_cache fullzones BITMAP).
2322 * 2) Check that the zones node (obtained from the zonelist_cache
2323 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
2324 * Return true (non-zero) if zone is worth looking at further, or
2325 * else return false (zero) if it is not.
2327 * This check -ignores- the distinction between various watermarks,
2328 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
2329 * found to be full for any variation of these watermarks, it will
2330 * be considered full for up to one second by all requests, unless
2331 * we are so low on memory on all allowed nodes that we are forced
2332 * into the second scan of the zonelist.
2334 * In the second scan we ignore this zonelist cache and exactly
2335 * apply the watermarks to all zones, even it is slower to do so.
2336 * We are low on memory in the second scan, and should leave no stone
2337 * unturned looking for a free page.
2339 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
2340 nodemask_t *allowednodes)
2342 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2343 int i; /* index of *z in zonelist zones */
2344 int n; /* node that zone *z is on */
2346 zlc = zonelist->zlcache_ptr;
2350 i = z - zonelist->_zonerefs;
2353 /* This zone is worth trying if it is allowed but not full */
2354 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
2358 * Given 'z' scanning a zonelist, set the corresponding bit in
2359 * zlc->fullzones, so that subsequent attempts to allocate a page
2360 * from that zone don't waste time re-examining it.
2362 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
2364 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2365 int i; /* index of *z in zonelist zones */
2367 zlc = zonelist->zlcache_ptr;
2371 i = z - zonelist->_zonerefs;
2373 set_bit(i, zlc->fullzones);
2377 * clear all zones full, called after direct reclaim makes progress so that
2378 * a zone that was recently full is not skipped over for up to a second
2380 static void zlc_clear_zones_full(struct zonelist *zonelist)
2382 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2384 zlc = zonelist->zlcache_ptr;
2388 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2391 static bool zone_local(struct zone *local_zone, struct zone *zone)
2393 return local_zone->node == zone->node;
2396 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2398 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2402 #else /* CONFIG_NUMA */
2404 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
2409 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
2410 nodemask_t *allowednodes)
2415 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
2419 static void zlc_clear_zones_full(struct zonelist *zonelist)
2423 static bool zone_local(struct zone *local_zone, struct zone *zone)
2428 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2433 #endif /* CONFIG_NUMA */
2435 static void reset_alloc_batches(struct zone *preferred_zone)
2437 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2440 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2441 high_wmark_pages(zone) - low_wmark_pages(zone) -
2442 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2443 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2444 } while (zone++ != preferred_zone);
2448 * get_page_from_freelist goes through the zonelist trying to allocate
2451 static struct page *
2452 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2453 const struct alloc_context *ac)
2455 struct zonelist *zonelist = ac->zonelist;
2457 struct page *page = NULL;
2459 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
2460 int zlc_active = 0; /* set if using zonelist_cache */
2461 int did_zlc_setup = 0; /* just call zlc_setup() one time */
2462 bool consider_zone_dirty = (alloc_flags & ALLOC_WMARK_LOW) &&
2463 (gfp_mask & __GFP_WRITE);
2464 int nr_fair_skipped = 0;
2465 bool zonelist_rescan;
2468 zonelist_rescan = false;
2471 * Scan zonelist, looking for a zone with enough free.
2472 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2474 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2478 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2479 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2481 if (cpusets_enabled() &&
2482 (alloc_flags & ALLOC_CPUSET) &&
2483 !cpuset_zone_allowed(zone, gfp_mask))
2486 * Distribute pages in proportion to the individual
2487 * zone size to ensure fair page aging. The zone a
2488 * page was allocated in should have no effect on the
2489 * time the page has in memory before being reclaimed.
2491 if (alloc_flags & ALLOC_FAIR) {
2492 if (!zone_local(ac->preferred_zone, zone))
2494 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2500 * When allocating a page cache page for writing, we
2501 * want to get it from a zone that is within its dirty
2502 * limit, such that no single zone holds more than its
2503 * proportional share of globally allowed dirty pages.
2504 * The dirty limits take into account the zone's
2505 * lowmem reserves and high watermark so that kswapd
2506 * should be able to balance it without having to
2507 * write pages from its LRU list.
2509 * This may look like it could increase pressure on
2510 * lower zones by failing allocations in higher zones
2511 * before they are full. But the pages that do spill
2512 * over are limited as the lower zones are protected
2513 * by this very same mechanism. It should not become
2514 * a practical burden to them.
2516 * XXX: For now, allow allocations to potentially
2517 * exceed the per-zone dirty limit in the slowpath
2518 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2519 * which is important when on a NUMA setup the allowed
2520 * zones are together not big enough to reach the
2521 * global limit. The proper fix for these situations
2522 * will require awareness of zones in the
2523 * dirty-throttling and the flusher threads.
2525 if (consider_zone_dirty && !zone_dirty_ok(zone))
2528 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2529 if (!zone_watermark_ok(zone, order, mark,
2530 ac->classzone_idx, alloc_flags)) {
2533 /* Checked here to keep the fast path fast */
2534 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2535 if (alloc_flags & ALLOC_NO_WATERMARKS)
2538 if (IS_ENABLED(CONFIG_NUMA) &&
2539 !did_zlc_setup && nr_online_nodes > 1) {
2541 * we do zlc_setup if there are multiple nodes
2542 * and before considering the first zone allowed
2545 allowednodes = zlc_setup(zonelist, alloc_flags);
2550 if (zone_reclaim_mode == 0 ||
2551 !zone_allows_reclaim(ac->preferred_zone, zone))
2552 goto this_zone_full;
2555 * As we may have just activated ZLC, check if the first
2556 * eligible zone has failed zone_reclaim recently.
2558 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2559 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2562 ret = zone_reclaim(zone, gfp_mask, order);
2564 case ZONE_RECLAIM_NOSCAN:
2567 case ZONE_RECLAIM_FULL:
2568 /* scanned but unreclaimable */
2571 /* did we reclaim enough */
2572 if (zone_watermark_ok(zone, order, mark,
2573 ac->classzone_idx, alloc_flags))
2577 * Failed to reclaim enough to meet watermark.
2578 * Only mark the zone full if checking the min
2579 * watermark or if we failed to reclaim just
2580 * 1<<order pages or else the page allocator
2581 * fastpath will prematurely mark zones full
2582 * when the watermark is between the low and
2585 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2586 ret == ZONE_RECLAIM_SOME)
2587 goto this_zone_full;
2594 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2595 gfp_mask, ac->migratetype);
2597 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2602 if (IS_ENABLED(CONFIG_NUMA) && zlc_active)
2603 zlc_mark_zone_full(zonelist, z);
2607 * The first pass makes sure allocations are spread fairly within the
2608 * local node. However, the local node might have free pages left
2609 * after the fairness batches are exhausted, and remote zones haven't
2610 * even been considered yet. Try once more without fairness, and
2611 * include remote zones now, before entering the slowpath and waking
2612 * kswapd: prefer spilling to a remote zone over swapping locally.
2614 if (alloc_flags & ALLOC_FAIR) {
2615 alloc_flags &= ~ALLOC_FAIR;
2616 if (nr_fair_skipped) {
2617 zonelist_rescan = true;
2618 reset_alloc_batches(ac->preferred_zone);
2620 if (nr_online_nodes > 1)
2621 zonelist_rescan = true;
2624 if (unlikely(IS_ENABLED(CONFIG_NUMA) && zlc_active)) {
2625 /* Disable zlc cache for second zonelist scan */
2627 zonelist_rescan = true;
2630 if (zonelist_rescan)
2637 * Large machines with many possible nodes should not always dump per-node
2638 * meminfo in irq context.
2640 static inline bool should_suppress_show_mem(void)
2645 ret = in_interrupt();
2650 static DEFINE_RATELIMIT_STATE(nopage_rs,
2651 DEFAULT_RATELIMIT_INTERVAL,
2652 DEFAULT_RATELIMIT_BURST);
2654 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2656 unsigned int filter = SHOW_MEM_FILTER_NODES;
2658 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2659 debug_guardpage_minorder() > 0)
2663 * This documents exceptions given to allocations in certain
2664 * contexts that are allowed to allocate outside current's set
2667 if (!(gfp_mask & __GFP_NOMEMALLOC))
2668 if (test_thread_flag(TIF_MEMDIE) ||
2669 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2670 filter &= ~SHOW_MEM_FILTER_NODES;
2671 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2672 filter &= ~SHOW_MEM_FILTER_NODES;
2675 struct va_format vaf;
2678 va_start(args, fmt);
2683 pr_warn("%pV", &vaf);
2688 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2689 current->comm, order, gfp_mask);
2692 if (!should_suppress_show_mem())
2696 static inline struct page *
2697 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2698 const struct alloc_context *ac, unsigned long *did_some_progress)
2700 struct oom_control oc = {
2701 .zonelist = ac->zonelist,
2702 .nodemask = ac->nodemask,
2703 .gfp_mask = gfp_mask,
2708 *did_some_progress = 0;
2711 * Acquire the oom lock. If that fails, somebody else is
2712 * making progress for us.
2714 if (!mutex_trylock(&oom_lock)) {
2715 *did_some_progress = 1;
2716 schedule_timeout_uninterruptible(1);
2721 * Go through the zonelist yet one more time, keep very high watermark
2722 * here, this is only to catch a parallel oom killing, we must fail if
2723 * we're still under heavy pressure.
2725 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2726 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2730 if (!(gfp_mask & __GFP_NOFAIL)) {
2731 /* Coredumps can quickly deplete all memory reserves */
2732 if (current->flags & PF_DUMPCORE)
2734 /* The OOM killer will not help higher order allocs */
2735 if (order > PAGE_ALLOC_COSTLY_ORDER)
2737 /* The OOM killer does not needlessly kill tasks for lowmem */
2738 if (ac->high_zoneidx < ZONE_NORMAL)
2740 /* The OOM killer does not compensate for IO-less reclaim */
2741 if (!(gfp_mask & __GFP_FS)) {
2743 * XXX: Page reclaim didn't yield anything,
2744 * and the OOM killer can't be invoked, but
2745 * keep looping as per tradition.
2747 *did_some_progress = 1;
2750 if (pm_suspended_storage())
2752 /* The OOM killer may not free memory on a specific node */
2753 if (gfp_mask & __GFP_THISNODE)
2756 /* Exhausted what can be done so it's blamo time */
2757 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL))
2758 *did_some_progress = 1;
2760 mutex_unlock(&oom_lock);
2764 #ifdef CONFIG_COMPACTION
2765 /* Try memory compaction for high-order allocations before reclaim */
2766 static struct page *
2767 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2768 int alloc_flags, const struct alloc_context *ac,
2769 enum migrate_mode mode, int *contended_compaction,
2770 bool *deferred_compaction)
2772 unsigned long compact_result;
2778 current->flags |= PF_MEMALLOC;
2779 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2780 mode, contended_compaction);
2781 current->flags &= ~PF_MEMALLOC;
2783 switch (compact_result) {
2784 case COMPACT_DEFERRED:
2785 *deferred_compaction = true;
2787 case COMPACT_SKIPPED:
2794 * At least in one zone compaction wasn't deferred or skipped, so let's
2795 * count a compaction stall
2797 count_vm_event(COMPACTSTALL);
2799 page = get_page_from_freelist(gfp_mask, order,
2800 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2803 struct zone *zone = page_zone(page);
2805 zone->compact_blockskip_flush = false;
2806 compaction_defer_reset(zone, order, true);
2807 count_vm_event(COMPACTSUCCESS);
2812 * It's bad if compaction run occurs and fails. The most likely reason
2813 * is that pages exist, but not enough to satisfy watermarks.
2815 count_vm_event(COMPACTFAIL);
2822 static inline struct page *
2823 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2824 int alloc_flags, const struct alloc_context *ac,
2825 enum migrate_mode mode, int *contended_compaction,
2826 bool *deferred_compaction)
2830 #endif /* CONFIG_COMPACTION */
2832 /* Perform direct synchronous page reclaim */
2834 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2835 const struct alloc_context *ac)
2837 struct reclaim_state reclaim_state;
2842 /* We now go into synchronous reclaim */
2843 cpuset_memory_pressure_bump();
2844 current->flags |= PF_MEMALLOC;
2845 lockdep_set_current_reclaim_state(gfp_mask);
2846 reclaim_state.reclaimed_slab = 0;
2847 current->reclaim_state = &reclaim_state;
2849 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2852 current->reclaim_state = NULL;
2853 lockdep_clear_current_reclaim_state();
2854 current->flags &= ~PF_MEMALLOC;
2861 /* The really slow allocator path where we enter direct reclaim */
2862 static inline struct page *
2863 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2864 int alloc_flags, const struct alloc_context *ac,
2865 unsigned long *did_some_progress)
2867 struct page *page = NULL;
2868 bool drained = false;
2870 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2871 if (unlikely(!(*did_some_progress)))
2874 /* After successful reclaim, reconsider all zones for allocation */
2875 if (IS_ENABLED(CONFIG_NUMA))
2876 zlc_clear_zones_full(ac->zonelist);
2879 page = get_page_from_freelist(gfp_mask, order,
2880 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2883 * If an allocation failed after direct reclaim, it could be because
2884 * pages are pinned on the per-cpu lists. Drain them and try again
2886 if (!page && !drained) {
2887 drain_all_pages(NULL);
2896 * This is called in the allocator slow-path if the allocation request is of
2897 * sufficient urgency to ignore watermarks and take other desperate measures
2899 static inline struct page *
2900 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2901 const struct alloc_context *ac)
2906 page = get_page_from_freelist(gfp_mask, order,
2907 ALLOC_NO_WATERMARKS, ac);
2909 if (!page && gfp_mask & __GFP_NOFAIL)
2910 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC,
2912 } while (!page && (gfp_mask & __GFP_NOFAIL));
2917 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2922 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2923 ac->high_zoneidx, ac->nodemask)
2924 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2928 gfp_to_alloc_flags(gfp_t gfp_mask)
2930 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2931 const bool atomic = !(gfp_mask & (__GFP_WAIT | __GFP_NO_KSWAPD));
2933 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2934 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2937 * The caller may dip into page reserves a bit more if the caller
2938 * cannot run direct reclaim, or if the caller has realtime scheduling
2939 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2940 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2942 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2946 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2947 * if it can't schedule.
2949 if (!(gfp_mask & __GFP_NOMEMALLOC))
2950 alloc_flags |= ALLOC_HARDER;
2952 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2953 * comment for __cpuset_node_allowed().
2955 alloc_flags &= ~ALLOC_CPUSET;
2956 } else if (unlikely(rt_task(current)) && !in_interrupt())
2957 alloc_flags |= ALLOC_HARDER;
2959 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2960 if (gfp_mask & __GFP_MEMALLOC)
2961 alloc_flags |= ALLOC_NO_WATERMARKS;
2962 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2963 alloc_flags |= ALLOC_NO_WATERMARKS;
2964 else if (!in_interrupt() &&
2965 ((current->flags & PF_MEMALLOC) ||
2966 unlikely(test_thread_flag(TIF_MEMDIE))))
2967 alloc_flags |= ALLOC_NO_WATERMARKS;
2970 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2971 alloc_flags |= ALLOC_CMA;
2976 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2978 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2981 static inline struct page *
2982 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2983 struct alloc_context *ac)
2985 const gfp_t wait = gfp_mask & __GFP_WAIT;
2986 struct page *page = NULL;
2988 unsigned long pages_reclaimed = 0;
2989 unsigned long did_some_progress;
2990 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2991 bool deferred_compaction = false;
2992 int contended_compaction = COMPACT_CONTENDED_NONE;
2995 * In the slowpath, we sanity check order to avoid ever trying to
2996 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2997 * be using allocators in order of preference for an area that is
3000 if (order >= MAX_ORDER) {
3001 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3006 * If this allocation cannot block and it is for a specific node, then
3007 * fail early. There's no need to wakeup kswapd or retry for a
3008 * speculative node-specific allocation.
3010 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !wait)
3014 if (!(gfp_mask & __GFP_NO_KSWAPD))
3015 wake_all_kswapds(order, ac);
3018 * OK, we're below the kswapd watermark and have kicked background
3019 * reclaim. Now things get more complex, so set up alloc_flags according
3020 * to how we want to proceed.
3022 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3025 * Find the true preferred zone if the allocation is unconstrained by
3028 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3029 struct zoneref *preferred_zoneref;
3030 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3031 ac->high_zoneidx, NULL, &ac->preferred_zone);
3032 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3035 /* This is the last chance, in general, before the goto nopage. */
3036 page = get_page_from_freelist(gfp_mask, order,
3037 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3041 /* Allocate without watermarks if the context allows */
3042 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3044 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3045 * the allocation is high priority and these type of
3046 * allocations are system rather than user orientated
3048 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3050 page = __alloc_pages_high_priority(gfp_mask, order, ac);
3057 /* Atomic allocations - we can't balance anything */
3060 * All existing users of the deprecated __GFP_NOFAIL are
3061 * blockable, so warn of any new users that actually allow this
3062 * type of allocation to fail.
3064 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3068 /* Avoid recursion of direct reclaim */
3069 if (current->flags & PF_MEMALLOC)
3072 /* Avoid allocations with no watermarks from looping endlessly */
3073 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3077 * Try direct compaction. The first pass is asynchronous. Subsequent
3078 * attempts after direct reclaim are synchronous
3080 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3082 &contended_compaction,
3083 &deferred_compaction);
3087 /* Checks for THP-specific high-order allocations */
3088 if ((gfp_mask & GFP_TRANSHUGE) == GFP_TRANSHUGE) {
3090 * If compaction is deferred for high-order allocations, it is
3091 * because sync compaction recently failed. If this is the case
3092 * and the caller requested a THP allocation, we do not want
3093 * to heavily disrupt the system, so we fail the allocation
3094 * instead of entering direct reclaim.
3096 if (deferred_compaction)
3100 * In all zones where compaction was attempted (and not
3101 * deferred or skipped), lock contention has been detected.
3102 * For THP allocation we do not want to disrupt the others
3103 * so we fallback to base pages instead.
3105 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3109 * If compaction was aborted due to need_resched(), we do not
3110 * want to further increase allocation latency, unless it is
3111 * khugepaged trying to collapse.
3113 if (contended_compaction == COMPACT_CONTENDED_SCHED
3114 && !(current->flags & PF_KTHREAD))
3119 * It can become very expensive to allocate transparent hugepages at
3120 * fault, so use asynchronous memory compaction for THP unless it is
3121 * khugepaged trying to collapse.
3123 if ((gfp_mask & GFP_TRANSHUGE) != GFP_TRANSHUGE ||
3124 (current->flags & PF_KTHREAD))
3125 migration_mode = MIGRATE_SYNC_LIGHT;
3127 /* Try direct reclaim and then allocating */
3128 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3129 &did_some_progress);
3133 /* Do not loop if specifically requested */
3134 if (gfp_mask & __GFP_NORETRY)
3137 /* Keep reclaiming pages as long as there is reasonable progress */
3138 pages_reclaimed += did_some_progress;
3139 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3140 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3141 /* Wait for some write requests to complete then retry */
3142 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3146 /* Reclaim has failed us, start killing things */
3147 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3151 /* Retry as long as the OOM killer is making progress */
3152 if (did_some_progress)
3157 * High-order allocations do not necessarily loop after
3158 * direct reclaim and reclaim/compaction depends on compaction
3159 * being called after reclaim so call directly if necessary
3161 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3163 &contended_compaction,
3164 &deferred_compaction);
3168 warn_alloc_failed(gfp_mask, order, NULL);
3174 * This is the 'heart' of the zoned buddy allocator.
3177 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3178 struct zonelist *zonelist, nodemask_t *nodemask)
3180 struct zoneref *preferred_zoneref;
3181 struct page *page = NULL;
3182 unsigned int cpuset_mems_cookie;
3183 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3184 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3185 struct alloc_context ac = {
3186 .high_zoneidx = gfp_zone(gfp_mask),
3187 .nodemask = nodemask,
3188 .migratetype = gfpflags_to_migratetype(gfp_mask),
3191 gfp_mask &= gfp_allowed_mask;
3193 lockdep_trace_alloc(gfp_mask);
3195 might_sleep_if(gfp_mask & __GFP_WAIT);
3197 if (should_fail_alloc_page(gfp_mask, order))
3201 * Check the zones suitable for the gfp_mask contain at least one
3202 * valid zone. It's possible to have an empty zonelist as a result
3203 * of __GFP_THISNODE and a memoryless node
3205 if (unlikely(!zonelist->_zonerefs->zone))
3208 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3209 alloc_flags |= ALLOC_CMA;
3212 cpuset_mems_cookie = read_mems_allowed_begin();
3214 /* We set it here, as __alloc_pages_slowpath might have changed it */
3215 ac.zonelist = zonelist;
3216 /* The preferred zone is used for statistics later */
3217 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3218 ac.nodemask ? : &cpuset_current_mems_allowed,
3219 &ac.preferred_zone);
3220 if (!ac.preferred_zone)
3222 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3224 /* First allocation attempt */
3225 alloc_mask = gfp_mask|__GFP_HARDWALL;
3226 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3227 if (unlikely(!page)) {
3229 * Runtime PM, block IO and its error handling path
3230 * can deadlock because I/O on the device might not
3233 alloc_mask = memalloc_noio_flags(gfp_mask);
3235 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3238 if (kmemcheck_enabled && page)
3239 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3241 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3245 * When updating a task's mems_allowed, it is possible to race with
3246 * parallel threads in such a way that an allocation can fail while
3247 * the mask is being updated. If a page allocation is about to fail,
3248 * check if the cpuset changed during allocation and if so, retry.
3250 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3255 EXPORT_SYMBOL(__alloc_pages_nodemask);
3258 * Common helper functions.
3260 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3265 * __get_free_pages() returns a 32-bit address, which cannot represent
3268 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3270 page = alloc_pages(gfp_mask, order);
3273 return (unsigned long) page_address(page);
3275 EXPORT_SYMBOL(__get_free_pages);
3277 unsigned long get_zeroed_page(gfp_t gfp_mask)
3279 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3281 EXPORT_SYMBOL(get_zeroed_page);
3283 void __free_pages(struct page *page, unsigned int order)
3285 if (put_page_testzero(page)) {
3287 free_hot_cold_page(page, false);
3289 __free_pages_ok(page, order);
3293 EXPORT_SYMBOL(__free_pages);
3295 void free_pages(unsigned long addr, unsigned int order)
3298 VM_BUG_ON(!virt_addr_valid((void *)addr));
3299 __free_pages(virt_to_page((void *)addr), order);
3303 EXPORT_SYMBOL(free_pages);
3307 * An arbitrary-length arbitrary-offset area of memory which resides
3308 * within a 0 or higher order page. Multiple fragments within that page
3309 * are individually refcounted, in the page's reference counter.
3311 * The page_frag functions below provide a simple allocation framework for
3312 * page fragments. This is used by the network stack and network device
3313 * drivers to provide a backing region of memory for use as either an
3314 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3316 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3319 struct page *page = NULL;
3320 gfp_t gfp = gfp_mask;
3322 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3323 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3325 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3326 PAGE_FRAG_CACHE_MAX_ORDER);
3327 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3329 if (unlikely(!page))
3330 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3332 nc->va = page ? page_address(page) : NULL;
3337 void *__alloc_page_frag(struct page_frag_cache *nc,
3338 unsigned int fragsz, gfp_t gfp_mask)
3340 unsigned int size = PAGE_SIZE;
3344 if (unlikely(!nc->va)) {
3346 page = __page_frag_refill(nc, gfp_mask);
3350 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3351 /* if size can vary use size else just use PAGE_SIZE */
3354 /* Even if we own the page, we do not use atomic_set().
3355 * This would break get_page_unless_zero() users.
3357 atomic_add(size - 1, &page->_count);
3359 /* reset page count bias and offset to start of new frag */
3360 nc->pfmemalloc = page_is_pfmemalloc(page);
3361 nc->pagecnt_bias = size;
3365 offset = nc->offset - fragsz;
3366 if (unlikely(offset < 0)) {
3367 page = virt_to_page(nc->va);
3369 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3372 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3373 /* if size can vary use size else just use PAGE_SIZE */
3376 /* OK, page count is 0, we can safely set it */
3377 atomic_set(&page->_count, size);
3379 /* reset page count bias and offset to start of new frag */
3380 nc->pagecnt_bias = size;
3381 offset = size - fragsz;
3385 nc->offset = offset;
3387 return nc->va + offset;
3389 EXPORT_SYMBOL(__alloc_page_frag);
3392 * Frees a page fragment allocated out of either a compound or order 0 page.
3394 void __free_page_frag(void *addr)
3396 struct page *page = virt_to_head_page(addr);
3398 if (unlikely(put_page_testzero(page)))
3399 __free_pages_ok(page, compound_order(page));
3401 EXPORT_SYMBOL(__free_page_frag);
3404 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3405 * of the current memory cgroup.
3407 * It should be used when the caller would like to use kmalloc, but since the
3408 * allocation is large, it has to fall back to the page allocator.
3410 struct page *alloc_kmem_pages(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(gfp_mask, order);
3418 memcg_kmem_commit_charge(page, memcg, order);
3422 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3425 struct mem_cgroup *memcg = NULL;
3427 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
3429 page = alloc_pages_node(nid, gfp_mask, order);
3430 memcg_kmem_commit_charge(page, memcg, order);
3435 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3438 void __free_kmem_pages(struct page *page, unsigned int order)
3440 memcg_kmem_uncharge_pages(page, order);
3441 __free_pages(page, order);
3444 void free_kmem_pages(unsigned long addr, unsigned int order)
3447 VM_BUG_ON(!virt_addr_valid((void *)addr));
3448 __free_kmem_pages(virt_to_page((void *)addr), order);
3452 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
3455 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3456 unsigned long used = addr + PAGE_ALIGN(size);
3458 split_page(virt_to_page((void *)addr), order);
3459 while (used < alloc_end) {
3464 return (void *)addr;
3468 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3469 * @size: the number of bytes to allocate
3470 * @gfp_mask: GFP flags for the allocation
3472 * This function is similar to alloc_pages(), except that it allocates the
3473 * minimum number of pages to satisfy the request. alloc_pages() can only
3474 * allocate memory in power-of-two pages.
3476 * This function is also limited by MAX_ORDER.
3478 * Memory allocated by this function must be released by free_pages_exact().
3480 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3482 unsigned int order = get_order(size);
3485 addr = __get_free_pages(gfp_mask, order);
3486 return make_alloc_exact(addr, order, size);
3488 EXPORT_SYMBOL(alloc_pages_exact);
3491 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3493 * @nid: the preferred node ID where memory should be allocated
3494 * @size: the number of bytes to allocate
3495 * @gfp_mask: GFP flags for the allocation
3497 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3499 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3502 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3504 unsigned order = get_order(size);
3505 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3508 return make_alloc_exact((unsigned long)page_address(p), order, size);
3512 * free_pages_exact - release memory allocated via alloc_pages_exact()
3513 * @virt: the value returned by alloc_pages_exact.
3514 * @size: size of allocation, same value as passed to alloc_pages_exact().
3516 * Release the memory allocated by a previous call to alloc_pages_exact.
3518 void free_pages_exact(void *virt, size_t size)
3520 unsigned long addr = (unsigned long)virt;
3521 unsigned long end = addr + PAGE_ALIGN(size);
3523 while (addr < end) {
3528 EXPORT_SYMBOL(free_pages_exact);
3531 * nr_free_zone_pages - count number of pages beyond high watermark
3532 * @offset: The zone index of the highest zone
3534 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3535 * high watermark within all zones at or below a given zone index. For each
3536 * zone, the number of pages is calculated as:
3537 * managed_pages - high_pages
3539 static unsigned long nr_free_zone_pages(int offset)
3544 /* Just pick one node, since fallback list is circular */
3545 unsigned long sum = 0;
3547 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3549 for_each_zone_zonelist(zone, z, zonelist, offset) {
3550 unsigned long size = zone->managed_pages;
3551 unsigned long high = high_wmark_pages(zone);
3560 * nr_free_buffer_pages - count number of pages beyond high watermark
3562 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3563 * watermark within ZONE_DMA and ZONE_NORMAL.
3565 unsigned long nr_free_buffer_pages(void)
3567 return nr_free_zone_pages(gfp_zone(GFP_USER));
3569 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3572 * nr_free_pagecache_pages - count number of pages beyond high watermark
3574 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3575 * high watermark within all zones.
3577 unsigned long nr_free_pagecache_pages(void)
3579 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3582 static inline void show_node(struct zone *zone)
3584 if (IS_ENABLED(CONFIG_NUMA))
3585 printk("Node %d ", zone_to_nid(zone));
3588 void si_meminfo(struct sysinfo *val)
3590 val->totalram = totalram_pages;
3591 val->sharedram = global_page_state(NR_SHMEM);
3592 val->freeram = global_page_state(NR_FREE_PAGES);
3593 val->bufferram = nr_blockdev_pages();
3594 val->totalhigh = totalhigh_pages;
3595 val->freehigh = nr_free_highpages();
3596 val->mem_unit = PAGE_SIZE;
3599 EXPORT_SYMBOL(si_meminfo);
3602 void si_meminfo_node(struct sysinfo *val, int nid)
3604 int zone_type; /* needs to be signed */
3605 unsigned long managed_pages = 0;
3606 pg_data_t *pgdat = NODE_DATA(nid);
3608 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3609 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3610 val->totalram = managed_pages;
3611 val->sharedram = node_page_state(nid, NR_SHMEM);
3612 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3613 #ifdef CONFIG_HIGHMEM
3614 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3615 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3621 val->mem_unit = PAGE_SIZE;
3626 * Determine whether the node should be displayed or not, depending on whether
3627 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3629 bool skip_free_areas_node(unsigned int flags, int nid)
3632 unsigned int cpuset_mems_cookie;
3634 if (!(flags & SHOW_MEM_FILTER_NODES))
3638 cpuset_mems_cookie = read_mems_allowed_begin();
3639 ret = !node_isset(nid, cpuset_current_mems_allowed);
3640 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3645 #define K(x) ((x) << (PAGE_SHIFT-10))
3647 static void show_migration_types(unsigned char type)
3649 static const char types[MIGRATE_TYPES] = {
3650 [MIGRATE_UNMOVABLE] = 'U',
3651 [MIGRATE_RECLAIMABLE] = 'E',
3652 [MIGRATE_MOVABLE] = 'M',
3653 [MIGRATE_RESERVE] = 'R',
3655 [MIGRATE_CMA] = 'C',
3657 #ifdef CONFIG_MEMORY_ISOLATION
3658 [MIGRATE_ISOLATE] = 'I',
3661 char tmp[MIGRATE_TYPES + 1];
3665 for (i = 0; i < MIGRATE_TYPES; i++) {
3666 if (type & (1 << i))
3671 printk("(%s) ", tmp);
3675 * Show free area list (used inside shift_scroll-lock stuff)
3676 * We also calculate the percentage fragmentation. We do this by counting the
3677 * memory on each free list with the exception of the first item on the list.
3680 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3683 void show_free_areas(unsigned int filter)
3685 unsigned long free_pcp = 0;
3689 for_each_populated_zone(zone) {
3690 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3693 for_each_online_cpu(cpu)
3694 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3697 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3698 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3699 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3700 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3701 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3702 " free:%lu free_pcp:%lu free_cma:%lu\n",
3703 global_page_state(NR_ACTIVE_ANON),
3704 global_page_state(NR_INACTIVE_ANON),
3705 global_page_state(NR_ISOLATED_ANON),
3706 global_page_state(NR_ACTIVE_FILE),
3707 global_page_state(NR_INACTIVE_FILE),
3708 global_page_state(NR_ISOLATED_FILE),
3709 global_page_state(NR_UNEVICTABLE),
3710 global_page_state(NR_FILE_DIRTY),
3711 global_page_state(NR_WRITEBACK),
3712 global_page_state(NR_UNSTABLE_NFS),
3713 global_page_state(NR_SLAB_RECLAIMABLE),
3714 global_page_state(NR_SLAB_UNRECLAIMABLE),
3715 global_page_state(NR_FILE_MAPPED),
3716 global_page_state(NR_SHMEM),
3717 global_page_state(NR_PAGETABLE),
3718 global_page_state(NR_BOUNCE),
3719 global_page_state(NR_FREE_PAGES),
3721 global_page_state(NR_FREE_CMA_PAGES));
3723 for_each_populated_zone(zone) {
3726 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3730 for_each_online_cpu(cpu)
3731 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3739 " active_anon:%lukB"
3740 " inactive_anon:%lukB"
3741 " active_file:%lukB"
3742 " inactive_file:%lukB"
3743 " unevictable:%lukB"
3744 " isolated(anon):%lukB"
3745 " isolated(file):%lukB"
3753 " slab_reclaimable:%lukB"
3754 " slab_unreclaimable:%lukB"
3755 " kernel_stack:%lukB"
3762 " writeback_tmp:%lukB"
3763 " pages_scanned:%lu"
3764 " all_unreclaimable? %s"
3767 K(zone_page_state(zone, NR_FREE_PAGES)),
3768 K(min_wmark_pages(zone)),
3769 K(low_wmark_pages(zone)),
3770 K(high_wmark_pages(zone)),
3771 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3772 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3773 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3774 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3775 K(zone_page_state(zone, NR_UNEVICTABLE)),
3776 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3777 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3778 K(zone->present_pages),
3779 K(zone->managed_pages),
3780 K(zone_page_state(zone, NR_MLOCK)),
3781 K(zone_page_state(zone, NR_FILE_DIRTY)),
3782 K(zone_page_state(zone, NR_WRITEBACK)),
3783 K(zone_page_state(zone, NR_FILE_MAPPED)),
3784 K(zone_page_state(zone, NR_SHMEM)),
3785 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3786 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3787 zone_page_state(zone, NR_KERNEL_STACK) *
3789 K(zone_page_state(zone, NR_PAGETABLE)),
3790 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3791 K(zone_page_state(zone, NR_BOUNCE)),
3793 K(this_cpu_read(zone->pageset->pcp.count)),
3794 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3795 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3796 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3797 (!zone_reclaimable(zone) ? "yes" : "no")
3799 printk("lowmem_reserve[]:");
3800 for (i = 0; i < MAX_NR_ZONES; i++)
3801 printk(" %ld", zone->lowmem_reserve[i]);
3805 for_each_populated_zone(zone) {
3806 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3807 unsigned char types[MAX_ORDER];
3809 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3812 printk("%s: ", zone->name);
3814 spin_lock_irqsave(&zone->lock, flags);
3815 for (order = 0; order < MAX_ORDER; order++) {
3816 struct free_area *area = &zone->free_area[order];
3819 nr[order] = area->nr_free;
3820 total += nr[order] << order;
3823 for (type = 0; type < MIGRATE_TYPES; type++) {
3824 if (!list_empty(&area->free_list[type]))
3825 types[order] |= 1 << type;
3828 spin_unlock_irqrestore(&zone->lock, flags);
3829 for (order = 0; order < MAX_ORDER; order++) {
3830 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3832 show_migration_types(types[order]);
3834 printk("= %lukB\n", K(total));
3837 hugetlb_show_meminfo();
3839 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3841 show_swap_cache_info();
3844 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3846 zoneref->zone = zone;
3847 zoneref->zone_idx = zone_idx(zone);
3851 * Builds allocation fallback zone lists.
3853 * Add all populated zones of a node to the zonelist.
3855 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3859 enum zone_type zone_type = MAX_NR_ZONES;
3863 zone = pgdat->node_zones + zone_type;
3864 if (populated_zone(zone)) {
3865 zoneref_set_zone(zone,
3866 &zonelist->_zonerefs[nr_zones++]);
3867 check_highest_zone(zone_type);
3869 } while (zone_type);
3877 * 0 = automatic detection of better ordering.
3878 * 1 = order by ([node] distance, -zonetype)
3879 * 2 = order by (-zonetype, [node] distance)
3881 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3882 * the same zonelist. So only NUMA can configure this param.
3884 #define ZONELIST_ORDER_DEFAULT 0
3885 #define ZONELIST_ORDER_NODE 1
3886 #define ZONELIST_ORDER_ZONE 2
3888 /* zonelist order in the kernel.
3889 * set_zonelist_order() will set this to NODE or ZONE.
3891 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3892 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3896 /* The value user specified ....changed by config */
3897 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3898 /* string for sysctl */
3899 #define NUMA_ZONELIST_ORDER_LEN 16
3900 char numa_zonelist_order[16] = "default";
3903 * interface for configure zonelist ordering.
3904 * command line option "numa_zonelist_order"
3905 * = "[dD]efault - default, automatic configuration.
3906 * = "[nN]ode - order by node locality, then by zone within node
3907 * = "[zZ]one - order by zone, then by locality within zone
3910 static int __parse_numa_zonelist_order(char *s)
3912 if (*s == 'd' || *s == 'D') {
3913 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3914 } else if (*s == 'n' || *s == 'N') {
3915 user_zonelist_order = ZONELIST_ORDER_NODE;
3916 } else if (*s == 'z' || *s == 'Z') {
3917 user_zonelist_order = ZONELIST_ORDER_ZONE;
3920 "Ignoring invalid numa_zonelist_order value: "
3927 static __init int setup_numa_zonelist_order(char *s)
3934 ret = __parse_numa_zonelist_order(s);
3936 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3940 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3943 * sysctl handler for numa_zonelist_order
3945 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3946 void __user *buffer, size_t *length,
3949 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3951 static DEFINE_MUTEX(zl_order_mutex);
3953 mutex_lock(&zl_order_mutex);
3955 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3959 strcpy(saved_string, (char *)table->data);
3961 ret = proc_dostring(table, write, buffer, length, ppos);
3965 int oldval = user_zonelist_order;
3967 ret = __parse_numa_zonelist_order((char *)table->data);
3970 * bogus value. restore saved string
3972 strncpy((char *)table->data, saved_string,
3973 NUMA_ZONELIST_ORDER_LEN);
3974 user_zonelist_order = oldval;
3975 } else if (oldval != user_zonelist_order) {
3976 mutex_lock(&zonelists_mutex);
3977 build_all_zonelists(NULL, NULL);
3978 mutex_unlock(&zonelists_mutex);
3982 mutex_unlock(&zl_order_mutex);
3987 #define MAX_NODE_LOAD (nr_online_nodes)
3988 static int node_load[MAX_NUMNODES];
3991 * find_next_best_node - find the next node that should appear in a given node's fallback list
3992 * @node: node whose fallback list we're appending
3993 * @used_node_mask: nodemask_t of already used nodes
3995 * We use a number of factors to determine which is the next node that should
3996 * appear on a given node's fallback list. The node should not have appeared
3997 * already in @node's fallback list, and it should be the next closest node
3998 * according to the distance array (which contains arbitrary distance values
3999 * from each node to each node in the system), and should also prefer nodes
4000 * with no CPUs, since presumably they'll have very little allocation pressure
4001 * on them otherwise.
4002 * It returns -1 if no node is found.
4004 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4007 int min_val = INT_MAX;
4008 int best_node = NUMA_NO_NODE;
4009 const struct cpumask *tmp = cpumask_of_node(0);
4011 /* Use the local node if we haven't already */
4012 if (!node_isset(node, *used_node_mask)) {
4013 node_set(node, *used_node_mask);
4017 for_each_node_state(n, N_MEMORY) {
4019 /* Don't want a node to appear more than once */
4020 if (node_isset(n, *used_node_mask))
4023 /* Use the distance array to find the distance */
4024 val = node_distance(node, n);
4026 /* Penalize nodes under us ("prefer the next node") */
4029 /* Give preference to headless and unused nodes */
4030 tmp = cpumask_of_node(n);
4031 if (!cpumask_empty(tmp))
4032 val += PENALTY_FOR_NODE_WITH_CPUS;
4034 /* Slight preference for less loaded node */
4035 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4036 val += node_load[n];
4038 if (val < min_val) {
4045 node_set(best_node, *used_node_mask);
4052 * Build zonelists ordered by node and zones within node.
4053 * This results in maximum locality--normal zone overflows into local
4054 * DMA zone, if any--but risks exhausting DMA zone.
4056 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4059 struct zonelist *zonelist;
4061 zonelist = &pgdat->node_zonelists[0];
4062 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4064 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4065 zonelist->_zonerefs[j].zone = NULL;
4066 zonelist->_zonerefs[j].zone_idx = 0;
4070 * Build gfp_thisnode zonelists
4072 static void build_thisnode_zonelists(pg_data_t *pgdat)
4075 struct zonelist *zonelist;
4077 zonelist = &pgdat->node_zonelists[1];
4078 j = build_zonelists_node(pgdat, zonelist, 0);
4079 zonelist->_zonerefs[j].zone = NULL;
4080 zonelist->_zonerefs[j].zone_idx = 0;
4084 * Build zonelists ordered by zone and nodes within zones.
4085 * This results in conserving DMA zone[s] until all Normal memory is
4086 * exhausted, but results in overflowing to remote node while memory
4087 * may still exist in local DMA zone.
4089 static int node_order[MAX_NUMNODES];
4091 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4094 int zone_type; /* needs to be signed */
4096 struct zonelist *zonelist;
4098 zonelist = &pgdat->node_zonelists[0];
4100 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4101 for (j = 0; j < nr_nodes; j++) {
4102 node = node_order[j];
4103 z = &NODE_DATA(node)->node_zones[zone_type];
4104 if (populated_zone(z)) {
4106 &zonelist->_zonerefs[pos++]);
4107 check_highest_zone(zone_type);
4111 zonelist->_zonerefs[pos].zone = NULL;
4112 zonelist->_zonerefs[pos].zone_idx = 0;
4115 #if defined(CONFIG_64BIT)
4117 * Devices that require DMA32/DMA are relatively rare and do not justify a
4118 * penalty to every machine in case the specialised case applies. Default
4119 * to Node-ordering on 64-bit NUMA machines
4121 static int default_zonelist_order(void)
4123 return ZONELIST_ORDER_NODE;
4127 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4128 * by the kernel. If processes running on node 0 deplete the low memory zone
4129 * then reclaim will occur more frequency increasing stalls and potentially
4130 * be easier to OOM if a large percentage of the zone is under writeback or
4131 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4132 * Hence, default to zone ordering on 32-bit.
4134 static int default_zonelist_order(void)
4136 return ZONELIST_ORDER_ZONE;
4138 #endif /* CONFIG_64BIT */
4140 static void set_zonelist_order(void)
4142 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4143 current_zonelist_order = default_zonelist_order();
4145 current_zonelist_order = user_zonelist_order;
4148 static void build_zonelists(pg_data_t *pgdat)
4152 nodemask_t used_mask;
4153 int local_node, prev_node;
4154 struct zonelist *zonelist;
4155 int order = current_zonelist_order;
4157 /* initialize zonelists */
4158 for (i = 0; i < MAX_ZONELISTS; i++) {
4159 zonelist = pgdat->node_zonelists + i;
4160 zonelist->_zonerefs[0].zone = NULL;
4161 zonelist->_zonerefs[0].zone_idx = 0;
4164 /* NUMA-aware ordering of nodes */
4165 local_node = pgdat->node_id;
4166 load = nr_online_nodes;
4167 prev_node = local_node;
4168 nodes_clear(used_mask);
4170 memset(node_order, 0, sizeof(node_order));
4173 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4175 * We don't want to pressure a particular node.
4176 * So adding penalty to the first node in same
4177 * distance group to make it round-robin.
4179 if (node_distance(local_node, node) !=
4180 node_distance(local_node, prev_node))
4181 node_load[node] = load;
4185 if (order == ZONELIST_ORDER_NODE)
4186 build_zonelists_in_node_order(pgdat, node);
4188 node_order[j++] = node; /* remember order */
4191 if (order == ZONELIST_ORDER_ZONE) {
4192 /* calculate node order -- i.e., DMA last! */
4193 build_zonelists_in_zone_order(pgdat, j);
4196 build_thisnode_zonelists(pgdat);
4199 /* Construct the zonelist performance cache - see further mmzone.h */
4200 static void build_zonelist_cache(pg_data_t *pgdat)
4202 struct zonelist *zonelist;
4203 struct zonelist_cache *zlc;
4206 zonelist = &pgdat->node_zonelists[0];
4207 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
4208 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
4209 for (z = zonelist->_zonerefs; z->zone; z++)
4210 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
4213 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4215 * Return node id of node used for "local" allocations.
4216 * I.e., first node id of first zone in arg node's generic zonelist.
4217 * Used for initializing percpu 'numa_mem', which is used primarily
4218 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4220 int local_memory_node(int node)
4224 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4225 gfp_zone(GFP_KERNEL),
4232 #else /* CONFIG_NUMA */
4234 static void set_zonelist_order(void)
4236 current_zonelist_order = ZONELIST_ORDER_ZONE;
4239 static void build_zonelists(pg_data_t *pgdat)
4241 int node, local_node;
4243 struct zonelist *zonelist;
4245 local_node = pgdat->node_id;
4247 zonelist = &pgdat->node_zonelists[0];
4248 j = build_zonelists_node(pgdat, zonelist, 0);
4251 * Now we build the zonelist so that it contains the zones
4252 * of all the other nodes.
4253 * We don't want to pressure a particular node, so when
4254 * building the zones for node N, we make sure that the
4255 * zones coming right after the local ones are those from
4256 * node N+1 (modulo N)
4258 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4259 if (!node_online(node))
4261 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4263 for (node = 0; node < local_node; node++) {
4264 if (!node_online(node))
4266 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4269 zonelist->_zonerefs[j].zone = NULL;
4270 zonelist->_zonerefs[j].zone_idx = 0;
4273 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
4274 static void build_zonelist_cache(pg_data_t *pgdat)
4276 pgdat->node_zonelists[0].zlcache_ptr = NULL;
4279 #endif /* CONFIG_NUMA */
4282 * Boot pageset table. One per cpu which is going to be used for all
4283 * zones and all nodes. The parameters will be set in such a way
4284 * that an item put on a list will immediately be handed over to
4285 * the buddy list. This is safe since pageset manipulation is done
4286 * with interrupts disabled.
4288 * The boot_pagesets must be kept even after bootup is complete for
4289 * unused processors and/or zones. They do play a role for bootstrapping
4290 * hotplugged processors.
4292 * zoneinfo_show() and maybe other functions do
4293 * not check if the processor is online before following the pageset pointer.
4294 * Other parts of the kernel may not check if the zone is available.
4296 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4297 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4298 static void setup_zone_pageset(struct zone *zone);
4301 * Global mutex to protect against size modification of zonelists
4302 * as well as to serialize pageset setup for the new populated zone.
4304 DEFINE_MUTEX(zonelists_mutex);
4306 /* return values int ....just for stop_machine() */
4307 static int __build_all_zonelists(void *data)
4311 pg_data_t *self = data;
4314 memset(node_load, 0, sizeof(node_load));
4317 if (self && !node_online(self->node_id)) {
4318 build_zonelists(self);
4319 build_zonelist_cache(self);
4322 for_each_online_node(nid) {
4323 pg_data_t *pgdat = NODE_DATA(nid);
4325 build_zonelists(pgdat);
4326 build_zonelist_cache(pgdat);
4330 * Initialize the boot_pagesets that are going to be used
4331 * for bootstrapping processors. The real pagesets for
4332 * each zone will be allocated later when the per cpu
4333 * allocator is available.
4335 * boot_pagesets are used also for bootstrapping offline
4336 * cpus if the system is already booted because the pagesets
4337 * are needed to initialize allocators on a specific cpu too.
4338 * F.e. the percpu allocator needs the page allocator which
4339 * needs the percpu allocator in order to allocate its pagesets
4340 * (a chicken-egg dilemma).
4342 for_each_possible_cpu(cpu) {
4343 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4345 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4347 * We now know the "local memory node" for each node--
4348 * i.e., the node of the first zone in the generic zonelist.
4349 * Set up numa_mem percpu variable for on-line cpus. During
4350 * boot, only the boot cpu should be on-line; we'll init the
4351 * secondary cpus' numa_mem as they come on-line. During
4352 * node/memory hotplug, we'll fixup all on-line cpus.
4354 if (cpu_online(cpu))
4355 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4362 static noinline void __init
4363 build_all_zonelists_init(void)
4365 __build_all_zonelists(NULL);
4366 mminit_verify_zonelist();
4367 cpuset_init_current_mems_allowed();
4371 * Called with zonelists_mutex held always
4372 * unless system_state == SYSTEM_BOOTING.
4374 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4375 * [we're only called with non-NULL zone through __meminit paths] and
4376 * (2) call of __init annotated helper build_all_zonelists_init
4377 * [protected by SYSTEM_BOOTING].
4379 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4381 set_zonelist_order();
4383 if (system_state == SYSTEM_BOOTING) {
4384 build_all_zonelists_init();
4386 #ifdef CONFIG_MEMORY_HOTPLUG
4388 setup_zone_pageset(zone);
4390 /* we have to stop all cpus to guarantee there is no user
4392 stop_machine(__build_all_zonelists, pgdat, NULL);
4393 /* cpuset refresh routine should be here */
4395 vm_total_pages = nr_free_pagecache_pages();
4397 * Disable grouping by mobility if the number of pages in the
4398 * system is too low to allow the mechanism to work. It would be
4399 * more accurate, but expensive to check per-zone. This check is
4400 * made on memory-hotadd so a system can start with mobility
4401 * disabled and enable it later
4403 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4404 page_group_by_mobility_disabled = 1;
4406 page_group_by_mobility_disabled = 0;
4408 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4409 "Total pages: %ld\n",
4411 zonelist_order_name[current_zonelist_order],
4412 page_group_by_mobility_disabled ? "off" : "on",
4415 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4420 * Helper functions to size the waitqueue hash table.
4421 * Essentially these want to choose hash table sizes sufficiently
4422 * large so that collisions trying to wait on pages are rare.
4423 * But in fact, the number of active page waitqueues on typical
4424 * systems is ridiculously low, less than 200. So this is even
4425 * conservative, even though it seems large.
4427 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4428 * waitqueues, i.e. the size of the waitq table given the number of pages.
4430 #define PAGES_PER_WAITQUEUE 256
4432 #ifndef CONFIG_MEMORY_HOTPLUG
4433 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4435 unsigned long size = 1;
4437 pages /= PAGES_PER_WAITQUEUE;
4439 while (size < pages)
4443 * Once we have dozens or even hundreds of threads sleeping
4444 * on IO we've got bigger problems than wait queue collision.
4445 * Limit the size of the wait table to a reasonable size.
4447 size = min(size, 4096UL);
4449 return max(size, 4UL);
4453 * A zone's size might be changed by hot-add, so it is not possible to determine
4454 * a suitable size for its wait_table. So we use the maximum size now.
4456 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4458 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4459 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4460 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4462 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4463 * or more by the traditional way. (See above). It equals:
4465 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4466 * ia64(16K page size) : = ( 8G + 4M)byte.
4467 * powerpc (64K page size) : = (32G +16M)byte.
4469 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4476 * This is an integer logarithm so that shifts can be used later
4477 * to extract the more random high bits from the multiplicative
4478 * hash function before the remainder is taken.
4480 static inline unsigned long wait_table_bits(unsigned long size)
4486 * Check if a pageblock contains reserved pages
4488 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
4492 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4493 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
4500 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4501 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4502 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4503 * higher will lead to a bigger reserve which will get freed as contiguous
4504 * blocks as reclaim kicks in
4506 static void setup_zone_migrate_reserve(struct zone *zone)
4508 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
4510 unsigned long block_migratetype;
4515 * Get the start pfn, end pfn and the number of blocks to reserve
4516 * We have to be careful to be aligned to pageblock_nr_pages to
4517 * make sure that we always check pfn_valid for the first page in
4520 start_pfn = zone->zone_start_pfn;
4521 end_pfn = zone_end_pfn(zone);
4522 start_pfn = roundup(start_pfn, pageblock_nr_pages);
4523 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
4527 * Reserve blocks are generally in place to help high-order atomic
4528 * allocations that are short-lived. A min_free_kbytes value that
4529 * would result in more than 2 reserve blocks for atomic allocations
4530 * is assumed to be in place to help anti-fragmentation for the
4531 * future allocation of hugepages at runtime.
4533 reserve = min(2, reserve);
4534 old_reserve = zone->nr_migrate_reserve_block;
4536 /* When memory hot-add, we almost always need to do nothing */
4537 if (reserve == old_reserve)
4539 zone->nr_migrate_reserve_block = reserve;
4541 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
4542 if (!early_page_nid_uninitialised(pfn, zone_to_nid(zone)))
4545 if (!pfn_valid(pfn))
4547 page = pfn_to_page(pfn);
4549 /* Watch out for overlapping nodes */
4550 if (page_to_nid(page) != zone_to_nid(zone))
4553 block_migratetype = get_pageblock_migratetype(page);
4555 /* Only test what is necessary when the reserves are not met */
4558 * Blocks with reserved pages will never free, skip
4561 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
4562 if (pageblock_is_reserved(pfn, block_end_pfn))
4565 /* If this block is reserved, account for it */
4566 if (block_migratetype == MIGRATE_RESERVE) {
4571 /* Suitable for reserving if this block is movable */
4572 if (block_migratetype == MIGRATE_MOVABLE) {
4573 set_pageblock_migratetype(page,
4575 move_freepages_block(zone, page,
4580 } else if (!old_reserve) {
4582 * At boot time we don't need to scan the whole zone
4583 * for turning off MIGRATE_RESERVE.
4589 * If the reserve is met and this is a previous reserved block,
4592 if (block_migratetype == MIGRATE_RESERVE) {
4593 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4594 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4600 * Initially all pages are reserved - free ones are freed
4601 * up by free_all_bootmem() once the early boot process is
4602 * done. Non-atomic initialization, single-pass.
4604 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4605 unsigned long start_pfn, enum memmap_context context)
4607 pg_data_t *pgdat = NODE_DATA(nid);
4608 unsigned long end_pfn = start_pfn + size;
4611 unsigned long nr_initialised = 0;
4613 if (highest_memmap_pfn < end_pfn - 1)
4614 highest_memmap_pfn = end_pfn - 1;
4616 z = &pgdat->node_zones[zone];
4617 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4619 * There can be holes in boot-time mem_map[]s
4620 * handed to this function. They do not
4621 * exist on hotplugged memory.
4623 if (context == MEMMAP_EARLY) {
4624 if (!early_pfn_valid(pfn))
4626 if (!early_pfn_in_nid(pfn, nid))
4628 if (!update_defer_init(pgdat, pfn, end_pfn,
4634 * Mark the block movable so that blocks are reserved for
4635 * movable at startup. This will force kernel allocations
4636 * to reserve their blocks rather than leaking throughout
4637 * the address space during boot when many long-lived
4638 * kernel allocations are made. Later some blocks near
4639 * the start are marked MIGRATE_RESERVE by
4640 * setup_zone_migrate_reserve()
4642 * bitmap is created for zone's valid pfn range. but memmap
4643 * can be created for invalid pages (for alignment)
4644 * check here not to call set_pageblock_migratetype() against
4647 if (!(pfn & (pageblock_nr_pages - 1))) {
4648 struct page *page = pfn_to_page(pfn);
4650 __init_single_page(page, pfn, zone, nid);
4651 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4653 __init_single_pfn(pfn, zone, nid);
4658 static void __meminit zone_init_free_lists(struct zone *zone)
4660 unsigned int order, t;
4661 for_each_migratetype_order(order, t) {
4662 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4663 zone->free_area[order].nr_free = 0;
4667 #ifndef __HAVE_ARCH_MEMMAP_INIT
4668 #define memmap_init(size, nid, zone, start_pfn) \
4669 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4672 static int zone_batchsize(struct zone *zone)
4678 * The per-cpu-pages pools are set to around 1000th of the
4679 * size of the zone. But no more than 1/2 of a meg.
4681 * OK, so we don't know how big the cache is. So guess.
4683 batch = zone->managed_pages / 1024;
4684 if (batch * PAGE_SIZE > 512 * 1024)
4685 batch = (512 * 1024) / PAGE_SIZE;
4686 batch /= 4; /* We effectively *= 4 below */
4691 * Clamp the batch to a 2^n - 1 value. Having a power
4692 * of 2 value was found to be more likely to have
4693 * suboptimal cache aliasing properties in some cases.
4695 * For example if 2 tasks are alternately allocating
4696 * batches of pages, one task can end up with a lot
4697 * of pages of one half of the possible page colors
4698 * and the other with pages of the other colors.
4700 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4705 /* The deferral and batching of frees should be suppressed under NOMMU
4708 * The problem is that NOMMU needs to be able to allocate large chunks
4709 * of contiguous memory as there's no hardware page translation to
4710 * assemble apparent contiguous memory from discontiguous pages.
4712 * Queueing large contiguous runs of pages for batching, however,
4713 * causes the pages to actually be freed in smaller chunks. As there
4714 * can be a significant delay between the individual batches being
4715 * recycled, this leads to the once large chunks of space being
4716 * fragmented and becoming unavailable for high-order allocations.
4723 * pcp->high and pcp->batch values are related and dependent on one another:
4724 * ->batch must never be higher then ->high.
4725 * The following function updates them in a safe manner without read side
4728 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4729 * those fields changing asynchronously (acording the the above rule).
4731 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4732 * outside of boot time (or some other assurance that no concurrent updaters
4735 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4736 unsigned long batch)
4738 /* start with a fail safe value for batch */
4742 /* Update high, then batch, in order */
4749 /* a companion to pageset_set_high() */
4750 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4752 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4755 static void pageset_init(struct per_cpu_pageset *p)
4757 struct per_cpu_pages *pcp;
4760 memset(p, 0, sizeof(*p));
4764 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4765 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4768 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4771 pageset_set_batch(p, batch);
4775 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4776 * to the value high for the pageset p.
4778 static void pageset_set_high(struct per_cpu_pageset *p,
4781 unsigned long batch = max(1UL, high / 4);
4782 if ((high / 4) > (PAGE_SHIFT * 8))
4783 batch = PAGE_SHIFT * 8;
4785 pageset_update(&p->pcp, high, batch);
4788 static void pageset_set_high_and_batch(struct zone *zone,
4789 struct per_cpu_pageset *pcp)
4791 if (percpu_pagelist_fraction)
4792 pageset_set_high(pcp,
4793 (zone->managed_pages /
4794 percpu_pagelist_fraction));
4796 pageset_set_batch(pcp, zone_batchsize(zone));
4799 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4801 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4804 pageset_set_high_and_batch(zone, pcp);
4807 static void __meminit setup_zone_pageset(struct zone *zone)
4810 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4811 for_each_possible_cpu(cpu)
4812 zone_pageset_init(zone, cpu);
4816 * Allocate per cpu pagesets and initialize them.
4817 * Before this call only boot pagesets were available.
4819 void __init setup_per_cpu_pageset(void)
4823 for_each_populated_zone(zone)
4824 setup_zone_pageset(zone);
4827 static noinline __init_refok
4828 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4834 * The per-page waitqueue mechanism uses hashed waitqueues
4837 zone->wait_table_hash_nr_entries =
4838 wait_table_hash_nr_entries(zone_size_pages);
4839 zone->wait_table_bits =
4840 wait_table_bits(zone->wait_table_hash_nr_entries);
4841 alloc_size = zone->wait_table_hash_nr_entries
4842 * sizeof(wait_queue_head_t);
4844 if (!slab_is_available()) {
4845 zone->wait_table = (wait_queue_head_t *)
4846 memblock_virt_alloc_node_nopanic(
4847 alloc_size, zone->zone_pgdat->node_id);
4850 * This case means that a zone whose size was 0 gets new memory
4851 * via memory hot-add.
4852 * But it may be the case that a new node was hot-added. In
4853 * this case vmalloc() will not be able to use this new node's
4854 * memory - this wait_table must be initialized to use this new
4855 * node itself as well.
4856 * To use this new node's memory, further consideration will be
4859 zone->wait_table = vmalloc(alloc_size);
4861 if (!zone->wait_table)
4864 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4865 init_waitqueue_head(zone->wait_table + i);
4870 static __meminit void zone_pcp_init(struct zone *zone)
4873 * per cpu subsystem is not up at this point. The following code
4874 * relies on the ability of the linker to provide the
4875 * offset of a (static) per cpu variable into the per cpu area.
4877 zone->pageset = &boot_pageset;
4879 if (populated_zone(zone))
4880 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4881 zone->name, zone->present_pages,
4882 zone_batchsize(zone));
4885 int __meminit init_currently_empty_zone(struct zone *zone,
4886 unsigned long zone_start_pfn,
4888 enum memmap_context context)
4890 struct pglist_data *pgdat = zone->zone_pgdat;
4892 ret = zone_wait_table_init(zone, size);
4895 pgdat->nr_zones = zone_idx(zone) + 1;
4897 zone->zone_start_pfn = zone_start_pfn;
4899 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4900 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4902 (unsigned long)zone_idx(zone),
4903 zone_start_pfn, (zone_start_pfn + size));
4905 zone_init_free_lists(zone);
4910 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4911 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4914 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4916 int __meminit __early_pfn_to_nid(unsigned long pfn,
4917 struct mminit_pfnnid_cache *state)
4919 unsigned long start_pfn, end_pfn;
4922 if (state->last_start <= pfn && pfn < state->last_end)
4923 return state->last_nid;
4925 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4927 state->last_start = start_pfn;
4928 state->last_end = end_pfn;
4929 state->last_nid = nid;
4934 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4937 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4938 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4939 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4941 * If an architecture guarantees that all ranges registered contain no holes
4942 * and may be freed, this this function may be used instead of calling
4943 * memblock_free_early_nid() manually.
4945 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4947 unsigned long start_pfn, end_pfn;
4950 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4951 start_pfn = min(start_pfn, max_low_pfn);
4952 end_pfn = min(end_pfn, max_low_pfn);
4954 if (start_pfn < end_pfn)
4955 memblock_free_early_nid(PFN_PHYS(start_pfn),
4956 (end_pfn - start_pfn) << PAGE_SHIFT,
4962 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4963 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4965 * If an architecture guarantees that all ranges registered contain no holes and may
4966 * be freed, this function may be used instead of calling memory_present() manually.
4968 void __init sparse_memory_present_with_active_regions(int nid)
4970 unsigned long start_pfn, end_pfn;
4973 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4974 memory_present(this_nid, start_pfn, end_pfn);
4978 * get_pfn_range_for_nid - Return the start and end page frames for a node
4979 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4980 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4981 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4983 * It returns the start and end page frame of a node based on information
4984 * provided by memblock_set_node(). If called for a node
4985 * with no available memory, a warning is printed and the start and end
4988 void __meminit get_pfn_range_for_nid(unsigned int nid,
4989 unsigned long *start_pfn, unsigned long *end_pfn)
4991 unsigned long this_start_pfn, this_end_pfn;
4997 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4998 *start_pfn = min(*start_pfn, this_start_pfn);
4999 *end_pfn = max(*end_pfn, this_end_pfn);
5002 if (*start_pfn == -1UL)
5007 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5008 * assumption is made that zones within a node are ordered in monotonic
5009 * increasing memory addresses so that the "highest" populated zone is used
5011 static void __init find_usable_zone_for_movable(void)
5014 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5015 if (zone_index == ZONE_MOVABLE)
5018 if (arch_zone_highest_possible_pfn[zone_index] >
5019 arch_zone_lowest_possible_pfn[zone_index])
5023 VM_BUG_ON(zone_index == -1);
5024 movable_zone = zone_index;
5028 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5029 * because it is sized independent of architecture. Unlike the other zones,
5030 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5031 * in each node depending on the size of each node and how evenly kernelcore
5032 * is distributed. This helper function adjusts the zone ranges
5033 * provided by the architecture for a given node by using the end of the
5034 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5035 * zones within a node are in order of monotonic increases memory addresses
5037 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5038 unsigned long zone_type,
5039 unsigned long node_start_pfn,
5040 unsigned long node_end_pfn,
5041 unsigned long *zone_start_pfn,
5042 unsigned long *zone_end_pfn)
5044 /* Only adjust if ZONE_MOVABLE is on this node */
5045 if (zone_movable_pfn[nid]) {
5046 /* Size ZONE_MOVABLE */
5047 if (zone_type == ZONE_MOVABLE) {
5048 *zone_start_pfn = zone_movable_pfn[nid];
5049 *zone_end_pfn = min(node_end_pfn,
5050 arch_zone_highest_possible_pfn[movable_zone]);
5052 /* Adjust for ZONE_MOVABLE starting within this range */
5053 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
5054 *zone_end_pfn > zone_movable_pfn[nid]) {
5055 *zone_end_pfn = zone_movable_pfn[nid];
5057 /* Check if this whole range is within ZONE_MOVABLE */
5058 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5059 *zone_start_pfn = *zone_end_pfn;
5064 * Return the number of pages a zone spans in a node, including holes
5065 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5067 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5068 unsigned long zone_type,
5069 unsigned long node_start_pfn,
5070 unsigned long node_end_pfn,
5071 unsigned long *ignored)
5073 unsigned long zone_start_pfn, zone_end_pfn;
5075 /* When hotadd a new node, the node should be empty */
5076 if (!node_start_pfn && !node_end_pfn)
5079 /* Get the start and end of the zone */
5080 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5081 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5082 adjust_zone_range_for_zone_movable(nid, zone_type,
5083 node_start_pfn, node_end_pfn,
5084 &zone_start_pfn, &zone_end_pfn);
5086 /* Check that this node has pages within the zone's required range */
5087 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
5090 /* Move the zone boundaries inside the node if necessary */
5091 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
5092 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
5094 /* Return the spanned pages */
5095 return zone_end_pfn - zone_start_pfn;
5099 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5100 * then all holes in the requested range will be accounted for.
5102 unsigned long __meminit __absent_pages_in_range(int nid,
5103 unsigned long range_start_pfn,
5104 unsigned long range_end_pfn)
5106 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5107 unsigned long start_pfn, end_pfn;
5110 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5111 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5112 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5113 nr_absent -= end_pfn - start_pfn;
5119 * absent_pages_in_range - Return number of page frames in holes within a range
5120 * @start_pfn: The start PFN to start searching for holes
5121 * @end_pfn: The end PFN to stop searching for holes
5123 * It returns the number of pages frames in memory holes within a range.
5125 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5126 unsigned long end_pfn)
5128 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5131 /* Return the number of page frames in holes in a zone on a node */
5132 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5133 unsigned long zone_type,
5134 unsigned long node_start_pfn,
5135 unsigned long node_end_pfn,
5136 unsigned long *ignored)
5138 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5139 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5140 unsigned long zone_start_pfn, zone_end_pfn;
5142 /* When hotadd a new node, the node should be empty */
5143 if (!node_start_pfn && !node_end_pfn)
5146 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5147 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5149 adjust_zone_range_for_zone_movable(nid, zone_type,
5150 node_start_pfn, node_end_pfn,
5151 &zone_start_pfn, &zone_end_pfn);
5152 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5155 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5156 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5157 unsigned long zone_type,
5158 unsigned long node_start_pfn,
5159 unsigned long node_end_pfn,
5160 unsigned long *zones_size)
5162 return zones_size[zone_type];
5165 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5166 unsigned long zone_type,
5167 unsigned long node_start_pfn,
5168 unsigned long node_end_pfn,
5169 unsigned long *zholes_size)
5174 return zholes_size[zone_type];
5177 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5179 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5180 unsigned long node_start_pfn,
5181 unsigned long node_end_pfn,
5182 unsigned long *zones_size,
5183 unsigned long *zholes_size)
5185 unsigned long realtotalpages = 0, totalpages = 0;
5188 for (i = 0; i < MAX_NR_ZONES; i++) {
5189 struct zone *zone = pgdat->node_zones + i;
5190 unsigned long size, real_size;
5192 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5196 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5197 node_start_pfn, node_end_pfn,
5199 zone->spanned_pages = size;
5200 zone->present_pages = real_size;
5203 realtotalpages += real_size;
5206 pgdat->node_spanned_pages = totalpages;
5207 pgdat->node_present_pages = realtotalpages;
5208 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5212 #ifndef CONFIG_SPARSEMEM
5214 * Calculate the size of the zone->blockflags rounded to an unsigned long
5215 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5216 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5217 * round what is now in bits to nearest long in bits, then return it in
5220 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5222 unsigned long usemapsize;
5224 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5225 usemapsize = roundup(zonesize, pageblock_nr_pages);
5226 usemapsize = usemapsize >> pageblock_order;
5227 usemapsize *= NR_PAGEBLOCK_BITS;
5228 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5230 return usemapsize / 8;
5233 static void __init setup_usemap(struct pglist_data *pgdat,
5235 unsigned long zone_start_pfn,
5236 unsigned long zonesize)
5238 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5239 zone->pageblock_flags = NULL;
5241 zone->pageblock_flags =
5242 memblock_virt_alloc_node_nopanic(usemapsize,
5246 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5247 unsigned long zone_start_pfn, unsigned long zonesize) {}
5248 #endif /* CONFIG_SPARSEMEM */
5250 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5252 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5253 void __paginginit set_pageblock_order(void)
5257 /* Check that pageblock_nr_pages has not already been setup */
5258 if (pageblock_order)
5261 if (HPAGE_SHIFT > PAGE_SHIFT)
5262 order = HUGETLB_PAGE_ORDER;
5264 order = MAX_ORDER - 1;
5267 * Assume the largest contiguous order of interest is a huge page.
5268 * This value may be variable depending on boot parameters on IA64 and
5271 pageblock_order = order;
5273 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5276 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5277 * is unused as pageblock_order is set at compile-time. See
5278 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5281 void __paginginit set_pageblock_order(void)
5285 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5287 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5288 unsigned long present_pages)
5290 unsigned long pages = spanned_pages;
5293 * Provide a more accurate estimation if there are holes within
5294 * the zone and SPARSEMEM is in use. If there are holes within the
5295 * zone, each populated memory region may cost us one or two extra
5296 * memmap pages due to alignment because memmap pages for each
5297 * populated regions may not naturally algined on page boundary.
5298 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5300 if (spanned_pages > present_pages + (present_pages >> 4) &&
5301 IS_ENABLED(CONFIG_SPARSEMEM))
5302 pages = present_pages;
5304 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5308 * Set up the zone data structures:
5309 * - mark all pages reserved
5310 * - mark all memory queues empty
5311 * - clear the memory bitmaps
5313 * NOTE: pgdat should get zeroed by caller.
5315 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5318 int nid = pgdat->node_id;
5319 unsigned long zone_start_pfn = pgdat->node_start_pfn;
5322 pgdat_resize_init(pgdat);
5323 #ifdef CONFIG_NUMA_BALANCING
5324 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5325 pgdat->numabalancing_migrate_nr_pages = 0;
5326 pgdat->numabalancing_migrate_next_window = jiffies;
5328 init_waitqueue_head(&pgdat->kswapd_wait);
5329 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5330 pgdat_page_ext_init(pgdat);
5332 for (j = 0; j < MAX_NR_ZONES; j++) {
5333 struct zone *zone = pgdat->node_zones + j;
5334 unsigned long size, realsize, freesize, memmap_pages;
5336 size = zone->spanned_pages;
5337 realsize = freesize = zone->present_pages;
5340 * Adjust freesize so that it accounts for how much memory
5341 * is used by this zone for memmap. This affects the watermark
5342 * and per-cpu initialisations
5344 memmap_pages = calc_memmap_size(size, realsize);
5345 if (!is_highmem_idx(j)) {
5346 if (freesize >= memmap_pages) {
5347 freesize -= memmap_pages;
5350 " %s zone: %lu pages used for memmap\n",
5351 zone_names[j], memmap_pages);
5354 " %s zone: %lu pages exceeds freesize %lu\n",
5355 zone_names[j], memmap_pages, freesize);
5358 /* Account for reserved pages */
5359 if (j == 0 && freesize > dma_reserve) {
5360 freesize -= dma_reserve;
5361 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5362 zone_names[0], dma_reserve);
5365 if (!is_highmem_idx(j))
5366 nr_kernel_pages += freesize;
5367 /* Charge for highmem memmap if there are enough kernel pages */
5368 else if (nr_kernel_pages > memmap_pages * 2)
5369 nr_kernel_pages -= memmap_pages;
5370 nr_all_pages += freesize;
5373 * Set an approximate value for lowmem here, it will be adjusted
5374 * when the bootmem allocator frees pages into the buddy system.
5375 * And all highmem pages will be managed by the buddy system.
5377 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5380 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5382 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5384 zone->name = zone_names[j];
5385 spin_lock_init(&zone->lock);
5386 spin_lock_init(&zone->lru_lock);
5387 zone_seqlock_init(zone);
5388 zone->zone_pgdat = pgdat;
5389 zone_pcp_init(zone);
5391 /* For bootup, initialized properly in watermark setup */
5392 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5394 lruvec_init(&zone->lruvec);
5398 set_pageblock_order();
5399 setup_usemap(pgdat, zone, zone_start_pfn, size);
5400 ret = init_currently_empty_zone(zone, zone_start_pfn,
5401 size, MEMMAP_EARLY);
5403 memmap_init(size, nid, j, zone_start_pfn);
5404 zone_start_pfn += size;
5408 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5410 /* Skip empty nodes */
5411 if (!pgdat->node_spanned_pages)
5414 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5415 /* ia64 gets its own node_mem_map, before this, without bootmem */
5416 if (!pgdat->node_mem_map) {
5417 unsigned long size, start, end;
5421 * The zone's endpoints aren't required to be MAX_ORDER
5422 * aligned but the node_mem_map endpoints must be in order
5423 * for the buddy allocator to function correctly.
5425 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5426 end = pgdat_end_pfn(pgdat);
5427 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5428 size = (end - start) * sizeof(struct page);
5429 map = alloc_remap(pgdat->node_id, size);
5431 map = memblock_virt_alloc_node_nopanic(size,
5433 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
5435 #ifndef CONFIG_NEED_MULTIPLE_NODES
5437 * With no DISCONTIG, the global mem_map is just set as node 0's
5439 if (pgdat == NODE_DATA(0)) {
5440 mem_map = NODE_DATA(0)->node_mem_map;
5441 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5442 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5443 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
5444 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5447 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5450 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5451 unsigned long node_start_pfn, unsigned long *zholes_size)
5453 pg_data_t *pgdat = NODE_DATA(nid);
5454 unsigned long start_pfn = 0;
5455 unsigned long end_pfn = 0;
5457 /* pg_data_t should be reset to zero when it's allocated */
5458 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5460 reset_deferred_meminit(pgdat);
5461 pgdat->node_id = nid;
5462 pgdat->node_start_pfn = node_start_pfn;
5463 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5464 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5465 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5466 (u64)start_pfn << PAGE_SHIFT, ((u64)end_pfn << PAGE_SHIFT) - 1);
5468 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5469 zones_size, zholes_size);
5471 alloc_node_mem_map(pgdat);
5472 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5473 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5474 nid, (unsigned long)pgdat,
5475 (unsigned long)pgdat->node_mem_map);
5478 free_area_init_core(pgdat);
5481 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5483 #if MAX_NUMNODES > 1
5485 * Figure out the number of possible node ids.
5487 void __init setup_nr_node_ids(void)
5489 unsigned int highest;
5491 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5492 nr_node_ids = highest + 1;
5497 * node_map_pfn_alignment - determine the maximum internode alignment
5499 * This function should be called after node map is populated and sorted.
5500 * It calculates the maximum power of two alignment which can distinguish
5503 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5504 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5505 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5506 * shifted, 1GiB is enough and this function will indicate so.
5508 * This is used to test whether pfn -> nid mapping of the chosen memory
5509 * model has fine enough granularity to avoid incorrect mapping for the
5510 * populated node map.
5512 * Returns the determined alignment in pfn's. 0 if there is no alignment
5513 * requirement (single node).
5515 unsigned long __init node_map_pfn_alignment(void)
5517 unsigned long accl_mask = 0, last_end = 0;
5518 unsigned long start, end, mask;
5522 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5523 if (!start || last_nid < 0 || last_nid == nid) {
5530 * Start with a mask granular enough to pin-point to the
5531 * start pfn and tick off bits one-by-one until it becomes
5532 * too coarse to separate the current node from the last.
5534 mask = ~((1 << __ffs(start)) - 1);
5535 while (mask && last_end <= (start & (mask << 1)))
5538 /* accumulate all internode masks */
5542 /* convert mask to number of pages */
5543 return ~accl_mask + 1;
5546 /* Find the lowest pfn for a node */
5547 static unsigned long __init find_min_pfn_for_node(int nid)
5549 unsigned long min_pfn = ULONG_MAX;
5550 unsigned long start_pfn;
5553 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5554 min_pfn = min(min_pfn, start_pfn);
5556 if (min_pfn == ULONG_MAX) {
5558 "Could not find start_pfn for node %d\n", nid);
5566 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5568 * It returns the minimum PFN based on information provided via
5569 * memblock_set_node().
5571 unsigned long __init find_min_pfn_with_active_regions(void)
5573 return find_min_pfn_for_node(MAX_NUMNODES);
5577 * early_calculate_totalpages()
5578 * Sum pages in active regions for movable zone.
5579 * Populate N_MEMORY for calculating usable_nodes.
5581 static unsigned long __init early_calculate_totalpages(void)
5583 unsigned long totalpages = 0;
5584 unsigned long start_pfn, end_pfn;
5587 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5588 unsigned long pages = end_pfn - start_pfn;
5590 totalpages += pages;
5592 node_set_state(nid, N_MEMORY);
5598 * Find the PFN the Movable zone begins in each node. Kernel memory
5599 * is spread evenly between nodes as long as the nodes have enough
5600 * memory. When they don't, some nodes will have more kernelcore than
5603 static void __init find_zone_movable_pfns_for_nodes(void)
5606 unsigned long usable_startpfn;
5607 unsigned long kernelcore_node, kernelcore_remaining;
5608 /* save the state before borrow the nodemask */
5609 nodemask_t saved_node_state = node_states[N_MEMORY];
5610 unsigned long totalpages = early_calculate_totalpages();
5611 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5612 struct memblock_region *r;
5614 /* Need to find movable_zone earlier when movable_node is specified. */
5615 find_usable_zone_for_movable();
5618 * If movable_node is specified, ignore kernelcore and movablecore
5621 if (movable_node_is_enabled()) {
5622 for_each_memblock(memory, r) {
5623 if (!memblock_is_hotpluggable(r))
5628 usable_startpfn = PFN_DOWN(r->base);
5629 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5630 min(usable_startpfn, zone_movable_pfn[nid]) :
5638 * If movablecore=nn[KMG] was specified, calculate what size of
5639 * kernelcore that corresponds so that memory usable for
5640 * any allocation type is evenly spread. If both kernelcore
5641 * and movablecore are specified, then the value of kernelcore
5642 * will be used for required_kernelcore if it's greater than
5643 * what movablecore would have allowed.
5645 if (required_movablecore) {
5646 unsigned long corepages;
5649 * Round-up so that ZONE_MOVABLE is at least as large as what
5650 * was requested by the user
5652 required_movablecore =
5653 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5654 corepages = totalpages - required_movablecore;
5656 required_kernelcore = max(required_kernelcore, corepages);
5659 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5660 if (!required_kernelcore)
5663 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5664 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5667 /* Spread kernelcore memory as evenly as possible throughout nodes */
5668 kernelcore_node = required_kernelcore / usable_nodes;
5669 for_each_node_state(nid, N_MEMORY) {
5670 unsigned long start_pfn, end_pfn;
5673 * Recalculate kernelcore_node if the division per node
5674 * now exceeds what is necessary to satisfy the requested
5675 * amount of memory for the kernel
5677 if (required_kernelcore < kernelcore_node)
5678 kernelcore_node = required_kernelcore / usable_nodes;
5681 * As the map is walked, we track how much memory is usable
5682 * by the kernel using kernelcore_remaining. When it is
5683 * 0, the rest of the node is usable by ZONE_MOVABLE
5685 kernelcore_remaining = kernelcore_node;
5687 /* Go through each range of PFNs within this node */
5688 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5689 unsigned long size_pages;
5691 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5692 if (start_pfn >= end_pfn)
5695 /* Account for what is only usable for kernelcore */
5696 if (start_pfn < usable_startpfn) {
5697 unsigned long kernel_pages;
5698 kernel_pages = min(end_pfn, usable_startpfn)
5701 kernelcore_remaining -= min(kernel_pages,
5702 kernelcore_remaining);
5703 required_kernelcore -= min(kernel_pages,
5704 required_kernelcore);
5706 /* Continue if range is now fully accounted */
5707 if (end_pfn <= usable_startpfn) {
5710 * Push zone_movable_pfn to the end so
5711 * that if we have to rebalance
5712 * kernelcore across nodes, we will
5713 * not double account here
5715 zone_movable_pfn[nid] = end_pfn;
5718 start_pfn = usable_startpfn;
5722 * The usable PFN range for ZONE_MOVABLE is from
5723 * start_pfn->end_pfn. Calculate size_pages as the
5724 * number of pages used as kernelcore
5726 size_pages = end_pfn - start_pfn;
5727 if (size_pages > kernelcore_remaining)
5728 size_pages = kernelcore_remaining;
5729 zone_movable_pfn[nid] = start_pfn + size_pages;
5732 * Some kernelcore has been met, update counts and
5733 * break if the kernelcore for this node has been
5736 required_kernelcore -= min(required_kernelcore,
5738 kernelcore_remaining -= size_pages;
5739 if (!kernelcore_remaining)
5745 * If there is still required_kernelcore, we do another pass with one
5746 * less node in the count. This will push zone_movable_pfn[nid] further
5747 * along on the nodes that still have memory until kernelcore is
5751 if (usable_nodes && required_kernelcore > usable_nodes)
5755 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5756 for (nid = 0; nid < MAX_NUMNODES; nid++)
5757 zone_movable_pfn[nid] =
5758 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5761 /* restore the node_state */
5762 node_states[N_MEMORY] = saved_node_state;
5765 /* Any regular or high memory on that node ? */
5766 static void check_for_memory(pg_data_t *pgdat, int nid)
5768 enum zone_type zone_type;
5770 if (N_MEMORY == N_NORMAL_MEMORY)
5773 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5774 struct zone *zone = &pgdat->node_zones[zone_type];
5775 if (populated_zone(zone)) {
5776 node_set_state(nid, N_HIGH_MEMORY);
5777 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5778 zone_type <= ZONE_NORMAL)
5779 node_set_state(nid, N_NORMAL_MEMORY);
5786 * free_area_init_nodes - Initialise all pg_data_t and zone data
5787 * @max_zone_pfn: an array of max PFNs for each zone
5789 * This will call free_area_init_node() for each active node in the system.
5790 * Using the page ranges provided by memblock_set_node(), the size of each
5791 * zone in each node and their holes is calculated. If the maximum PFN
5792 * between two adjacent zones match, it is assumed that the zone is empty.
5793 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5794 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5795 * starts where the previous one ended. For example, ZONE_DMA32 starts
5796 * at arch_max_dma_pfn.
5798 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5800 unsigned long start_pfn, end_pfn;
5803 /* Record where the zone boundaries are */
5804 memset(arch_zone_lowest_possible_pfn, 0,
5805 sizeof(arch_zone_lowest_possible_pfn));
5806 memset(arch_zone_highest_possible_pfn, 0,
5807 sizeof(arch_zone_highest_possible_pfn));
5808 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5809 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5810 for (i = 1; i < MAX_NR_ZONES; i++) {
5811 if (i == ZONE_MOVABLE)
5813 arch_zone_lowest_possible_pfn[i] =
5814 arch_zone_highest_possible_pfn[i-1];
5815 arch_zone_highest_possible_pfn[i] =
5816 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5818 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5819 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5821 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5822 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5823 find_zone_movable_pfns_for_nodes();
5825 /* Print out the zone ranges */
5826 pr_info("Zone ranges:\n");
5827 for (i = 0; i < MAX_NR_ZONES; i++) {
5828 if (i == ZONE_MOVABLE)
5830 pr_info(" %-8s ", zone_names[i]);
5831 if (arch_zone_lowest_possible_pfn[i] ==
5832 arch_zone_highest_possible_pfn[i])
5835 pr_cont("[mem %#018Lx-%#018Lx]\n",
5836 (u64)arch_zone_lowest_possible_pfn[i]
5838 ((u64)arch_zone_highest_possible_pfn[i]
5839 << PAGE_SHIFT) - 1);
5842 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5843 pr_info("Movable zone start for each node\n");
5844 for (i = 0; i < MAX_NUMNODES; i++) {
5845 if (zone_movable_pfn[i])
5846 pr_info(" Node %d: %#018Lx\n", i,
5847 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5850 /* Print out the early node map */
5851 pr_info("Early memory node ranges\n");
5852 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5853 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5854 (u64)start_pfn << PAGE_SHIFT,
5855 ((u64)end_pfn << PAGE_SHIFT) - 1);
5857 /* Initialise every node */
5858 mminit_verify_pageflags_layout();
5859 setup_nr_node_ids();
5860 for_each_online_node(nid) {
5861 pg_data_t *pgdat = NODE_DATA(nid);
5862 free_area_init_node(nid, NULL,
5863 find_min_pfn_for_node(nid), NULL);
5865 /* Any memory on that node */
5866 if (pgdat->node_present_pages)
5867 node_set_state(nid, N_MEMORY);
5868 check_for_memory(pgdat, nid);
5872 static int __init cmdline_parse_core(char *p, unsigned long *core)
5874 unsigned long long coremem;
5878 coremem = memparse(p, &p);
5879 *core = coremem >> PAGE_SHIFT;
5881 /* Paranoid check that UL is enough for the coremem value */
5882 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5888 * kernelcore=size sets the amount of memory for use for allocations that
5889 * cannot be reclaimed or migrated.
5891 static int __init cmdline_parse_kernelcore(char *p)
5893 return cmdline_parse_core(p, &required_kernelcore);
5897 * movablecore=size sets the amount of memory for use for allocations that
5898 * can be reclaimed or migrated.
5900 static int __init cmdline_parse_movablecore(char *p)
5902 return cmdline_parse_core(p, &required_movablecore);
5905 early_param("kernelcore", cmdline_parse_kernelcore);
5906 early_param("movablecore", cmdline_parse_movablecore);
5908 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5910 void adjust_managed_page_count(struct page *page, long count)
5912 spin_lock(&managed_page_count_lock);
5913 page_zone(page)->managed_pages += count;
5914 totalram_pages += count;
5915 #ifdef CONFIG_HIGHMEM
5916 if (PageHighMem(page))
5917 totalhigh_pages += count;
5919 spin_unlock(&managed_page_count_lock);
5921 EXPORT_SYMBOL(adjust_managed_page_count);
5923 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5926 unsigned long pages = 0;
5928 start = (void *)PAGE_ALIGN((unsigned long)start);
5929 end = (void *)((unsigned long)end & PAGE_MASK);
5930 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5931 if ((unsigned int)poison <= 0xFF)
5932 memset(pos, poison, PAGE_SIZE);
5933 free_reserved_page(virt_to_page(pos));
5937 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5938 s, pages << (PAGE_SHIFT - 10), start, end);
5942 EXPORT_SYMBOL(free_reserved_area);
5944 #ifdef CONFIG_HIGHMEM
5945 void free_highmem_page(struct page *page)
5947 __free_reserved_page(page);
5949 page_zone(page)->managed_pages++;
5955 void __init mem_init_print_info(const char *str)
5957 unsigned long physpages, codesize, datasize, rosize, bss_size;
5958 unsigned long init_code_size, init_data_size;
5960 physpages = get_num_physpages();
5961 codesize = _etext - _stext;
5962 datasize = _edata - _sdata;
5963 rosize = __end_rodata - __start_rodata;
5964 bss_size = __bss_stop - __bss_start;
5965 init_data_size = __init_end - __init_begin;
5966 init_code_size = _einittext - _sinittext;
5969 * Detect special cases and adjust section sizes accordingly:
5970 * 1) .init.* may be embedded into .data sections
5971 * 2) .init.text.* may be out of [__init_begin, __init_end],
5972 * please refer to arch/tile/kernel/vmlinux.lds.S.
5973 * 3) .rodata.* may be embedded into .text or .data sections.
5975 #define adj_init_size(start, end, size, pos, adj) \
5977 if (start <= pos && pos < end && size > adj) \
5981 adj_init_size(__init_begin, __init_end, init_data_size,
5982 _sinittext, init_code_size);
5983 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5984 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5985 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5986 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5988 #undef adj_init_size
5990 pr_info("Memory: %luK/%luK available "
5991 "(%luK kernel code, %luK rwdata, %luK rodata, "
5992 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5993 #ifdef CONFIG_HIGHMEM
5997 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5998 codesize >> 10, datasize >> 10, rosize >> 10,
5999 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6000 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
6001 totalcma_pages << (PAGE_SHIFT-10),
6002 #ifdef CONFIG_HIGHMEM
6003 totalhigh_pages << (PAGE_SHIFT-10),
6005 str ? ", " : "", str ? str : "");
6009 * set_dma_reserve - set the specified number of pages reserved in the first zone
6010 * @new_dma_reserve: The number of pages to mark reserved
6012 * The per-cpu batchsize and zone watermarks are determined by present_pages.
6013 * In the DMA zone, a significant percentage may be consumed by kernel image
6014 * and other unfreeable allocations which can skew the watermarks badly. This
6015 * function may optionally be used to account for unfreeable pages in the
6016 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6017 * smaller per-cpu batchsize.
6019 void __init set_dma_reserve(unsigned long new_dma_reserve)
6021 dma_reserve = new_dma_reserve;
6024 void __init free_area_init(unsigned long *zones_size)
6026 free_area_init_node(0, zones_size,
6027 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6030 static int page_alloc_cpu_notify(struct notifier_block *self,
6031 unsigned long action, void *hcpu)
6033 int cpu = (unsigned long)hcpu;
6035 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6036 lru_add_drain_cpu(cpu);
6040 * Spill the event counters of the dead processor
6041 * into the current processors event counters.
6042 * This artificially elevates the count of the current
6045 vm_events_fold_cpu(cpu);
6048 * Zero the differential counters of the dead processor
6049 * so that the vm statistics are consistent.
6051 * This is only okay since the processor is dead and cannot
6052 * race with what we are doing.
6054 cpu_vm_stats_fold(cpu);
6059 void __init page_alloc_init(void)
6061 hotcpu_notifier(page_alloc_cpu_notify, 0);
6065 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
6066 * or min_free_kbytes changes.
6068 static void calculate_totalreserve_pages(void)
6070 struct pglist_data *pgdat;
6071 unsigned long reserve_pages = 0;
6072 enum zone_type i, j;
6074 for_each_online_pgdat(pgdat) {
6075 for (i = 0; i < MAX_NR_ZONES; i++) {
6076 struct zone *zone = pgdat->node_zones + i;
6079 /* Find valid and maximum lowmem_reserve in the zone */
6080 for (j = i; j < MAX_NR_ZONES; j++) {
6081 if (zone->lowmem_reserve[j] > max)
6082 max = zone->lowmem_reserve[j];
6085 /* we treat the high watermark as reserved pages. */
6086 max += high_wmark_pages(zone);
6088 if (max > zone->managed_pages)
6089 max = zone->managed_pages;
6090 reserve_pages += max;
6092 * Lowmem reserves are not available to
6093 * GFP_HIGHUSER page cache allocations and
6094 * kswapd tries to balance zones to their high
6095 * watermark. As a result, neither should be
6096 * regarded as dirtyable memory, to prevent a
6097 * situation where reclaim has to clean pages
6098 * in order to balance the zones.
6100 zone->dirty_balance_reserve = max;
6103 dirty_balance_reserve = reserve_pages;
6104 totalreserve_pages = reserve_pages;
6108 * setup_per_zone_lowmem_reserve - called whenever
6109 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
6110 * has a correct pages reserved value, so an adequate number of
6111 * pages are left in the zone after a successful __alloc_pages().
6113 static void setup_per_zone_lowmem_reserve(void)
6115 struct pglist_data *pgdat;
6116 enum zone_type j, idx;
6118 for_each_online_pgdat(pgdat) {
6119 for (j = 0; j < MAX_NR_ZONES; j++) {
6120 struct zone *zone = pgdat->node_zones + j;
6121 unsigned long managed_pages = zone->managed_pages;
6123 zone->lowmem_reserve[j] = 0;
6127 struct zone *lower_zone;
6131 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6132 sysctl_lowmem_reserve_ratio[idx] = 1;
6134 lower_zone = pgdat->node_zones + idx;
6135 lower_zone->lowmem_reserve[j] = managed_pages /
6136 sysctl_lowmem_reserve_ratio[idx];
6137 managed_pages += lower_zone->managed_pages;
6142 /* update totalreserve_pages */
6143 calculate_totalreserve_pages();
6146 static void __setup_per_zone_wmarks(void)
6148 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6149 unsigned long lowmem_pages = 0;
6151 unsigned long flags;
6153 /* Calculate total number of !ZONE_HIGHMEM pages */
6154 for_each_zone(zone) {
6155 if (!is_highmem(zone))
6156 lowmem_pages += zone->managed_pages;
6159 for_each_zone(zone) {
6162 spin_lock_irqsave(&zone->lock, flags);
6163 tmp = (u64)pages_min * zone->managed_pages;
6164 do_div(tmp, lowmem_pages);
6165 if (is_highmem(zone)) {
6167 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6168 * need highmem pages, so cap pages_min to a small
6171 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6172 * deltas control asynch page reclaim, and so should
6173 * not be capped for highmem.
6175 unsigned long min_pages;
6177 min_pages = zone->managed_pages / 1024;
6178 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6179 zone->watermark[WMARK_MIN] = min_pages;
6182 * If it's a lowmem zone, reserve a number of pages
6183 * proportionate to the zone's size.
6185 zone->watermark[WMARK_MIN] = tmp;
6188 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
6189 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
6191 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6192 high_wmark_pages(zone) - low_wmark_pages(zone) -
6193 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6195 setup_zone_migrate_reserve(zone);
6196 spin_unlock_irqrestore(&zone->lock, flags);
6199 /* update totalreserve_pages */
6200 calculate_totalreserve_pages();
6204 * setup_per_zone_wmarks - called when min_free_kbytes changes
6205 * or when memory is hot-{added|removed}
6207 * Ensures that the watermark[min,low,high] values for each zone are set
6208 * correctly with respect to min_free_kbytes.
6210 void setup_per_zone_wmarks(void)
6212 mutex_lock(&zonelists_mutex);
6213 __setup_per_zone_wmarks();
6214 mutex_unlock(&zonelists_mutex);
6218 * The inactive anon list should be small enough that the VM never has to
6219 * do too much work, but large enough that each inactive page has a chance
6220 * to be referenced again before it is swapped out.
6222 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6223 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6224 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6225 * the anonymous pages are kept on the inactive list.
6228 * memory ratio inactive anon
6229 * -------------------------------------
6238 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6240 unsigned int gb, ratio;
6242 /* Zone size in gigabytes */
6243 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6245 ratio = int_sqrt(10 * gb);
6249 zone->inactive_ratio = ratio;
6252 static void __meminit setup_per_zone_inactive_ratio(void)
6257 calculate_zone_inactive_ratio(zone);
6261 * Initialise min_free_kbytes.
6263 * For small machines we want it small (128k min). For large machines
6264 * we want it large (64MB max). But it is not linear, because network
6265 * bandwidth does not increase linearly with machine size. We use
6267 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6268 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6284 int __meminit init_per_zone_wmark_min(void)
6286 unsigned long lowmem_kbytes;
6287 int new_min_free_kbytes;
6289 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6290 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6292 if (new_min_free_kbytes > user_min_free_kbytes) {
6293 min_free_kbytes = new_min_free_kbytes;
6294 if (min_free_kbytes < 128)
6295 min_free_kbytes = 128;
6296 if (min_free_kbytes > 65536)
6297 min_free_kbytes = 65536;
6299 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6300 new_min_free_kbytes, user_min_free_kbytes);
6302 setup_per_zone_wmarks();
6303 refresh_zone_stat_thresholds();
6304 setup_per_zone_lowmem_reserve();
6305 setup_per_zone_inactive_ratio();
6308 module_init(init_per_zone_wmark_min)
6311 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6312 * that we can call two helper functions whenever min_free_kbytes
6315 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6316 void __user *buffer, size_t *length, loff_t *ppos)
6320 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6325 user_min_free_kbytes = min_free_kbytes;
6326 setup_per_zone_wmarks();
6332 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6333 void __user *buffer, size_t *length, loff_t *ppos)
6338 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6343 zone->min_unmapped_pages = (zone->managed_pages *
6344 sysctl_min_unmapped_ratio) / 100;
6348 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6349 void __user *buffer, size_t *length, loff_t *ppos)
6354 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6359 zone->min_slab_pages = (zone->managed_pages *
6360 sysctl_min_slab_ratio) / 100;
6366 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6367 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6368 * whenever sysctl_lowmem_reserve_ratio changes.
6370 * The reserve ratio obviously has absolutely no relation with the
6371 * minimum watermarks. The lowmem reserve ratio can only make sense
6372 * if in function of the boot time zone sizes.
6374 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6375 void __user *buffer, size_t *length, loff_t *ppos)
6377 proc_dointvec_minmax(table, write, buffer, length, ppos);
6378 setup_per_zone_lowmem_reserve();
6383 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6384 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6385 * pagelist can have before it gets flushed back to buddy allocator.
6387 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6388 void __user *buffer, size_t *length, loff_t *ppos)
6391 int old_percpu_pagelist_fraction;
6394 mutex_lock(&pcp_batch_high_lock);
6395 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6397 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6398 if (!write || ret < 0)
6401 /* Sanity checking to avoid pcp imbalance */
6402 if (percpu_pagelist_fraction &&
6403 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6404 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6410 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6413 for_each_populated_zone(zone) {
6416 for_each_possible_cpu(cpu)
6417 pageset_set_high_and_batch(zone,
6418 per_cpu_ptr(zone->pageset, cpu));
6421 mutex_unlock(&pcp_batch_high_lock);
6426 int hashdist = HASHDIST_DEFAULT;
6428 static int __init set_hashdist(char *str)
6432 hashdist = simple_strtoul(str, &str, 0);
6435 __setup("hashdist=", set_hashdist);
6439 * allocate a large system hash table from bootmem
6440 * - it is assumed that the hash table must contain an exact power-of-2
6441 * quantity of entries
6442 * - limit is the number of hash buckets, not the total allocation size
6444 void *__init alloc_large_system_hash(const char *tablename,
6445 unsigned long bucketsize,
6446 unsigned long numentries,
6449 unsigned int *_hash_shift,
6450 unsigned int *_hash_mask,
6451 unsigned long low_limit,
6452 unsigned long high_limit)
6454 unsigned long long max = high_limit;
6455 unsigned long log2qty, size;
6458 /* allow the kernel cmdline to have a say */
6460 /* round applicable memory size up to nearest megabyte */
6461 numentries = nr_kernel_pages;
6463 /* It isn't necessary when PAGE_SIZE >= 1MB */
6464 if (PAGE_SHIFT < 20)
6465 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6467 /* limit to 1 bucket per 2^scale bytes of low memory */
6468 if (scale > PAGE_SHIFT)
6469 numentries >>= (scale - PAGE_SHIFT);
6471 numentries <<= (PAGE_SHIFT - scale);
6473 /* Make sure we've got at least a 0-order allocation.. */
6474 if (unlikely(flags & HASH_SMALL)) {
6475 /* Makes no sense without HASH_EARLY */
6476 WARN_ON(!(flags & HASH_EARLY));
6477 if (!(numentries >> *_hash_shift)) {
6478 numentries = 1UL << *_hash_shift;
6479 BUG_ON(!numentries);
6481 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6482 numentries = PAGE_SIZE / bucketsize;
6484 numentries = roundup_pow_of_two(numentries);
6486 /* limit allocation size to 1/16 total memory by default */
6488 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6489 do_div(max, bucketsize);
6491 max = min(max, 0x80000000ULL);
6493 if (numentries < low_limit)
6494 numentries = low_limit;
6495 if (numentries > max)
6498 log2qty = ilog2(numentries);
6501 size = bucketsize << log2qty;
6502 if (flags & HASH_EARLY)
6503 table = memblock_virt_alloc_nopanic(size, 0);
6505 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6508 * If bucketsize is not a power-of-two, we may free
6509 * some pages at the end of hash table which
6510 * alloc_pages_exact() automatically does
6512 if (get_order(size) < MAX_ORDER) {
6513 table = alloc_pages_exact(size, GFP_ATOMIC);
6514 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6517 } while (!table && size > PAGE_SIZE && --log2qty);
6520 panic("Failed to allocate %s hash table\n", tablename);
6522 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6525 ilog2(size) - PAGE_SHIFT,
6529 *_hash_shift = log2qty;
6531 *_hash_mask = (1 << log2qty) - 1;
6536 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6537 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6540 #ifdef CONFIG_SPARSEMEM
6541 return __pfn_to_section(pfn)->pageblock_flags;
6543 return zone->pageblock_flags;
6544 #endif /* CONFIG_SPARSEMEM */
6547 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6549 #ifdef CONFIG_SPARSEMEM
6550 pfn &= (PAGES_PER_SECTION-1);
6551 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6553 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6554 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6555 #endif /* CONFIG_SPARSEMEM */
6559 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6560 * @page: The page within the block of interest
6561 * @pfn: The target page frame number
6562 * @end_bitidx: The last bit of interest to retrieve
6563 * @mask: mask of bits that the caller is interested in
6565 * Return: pageblock_bits flags
6567 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6568 unsigned long end_bitidx,
6572 unsigned long *bitmap;
6573 unsigned long bitidx, word_bitidx;
6576 zone = page_zone(page);
6577 bitmap = get_pageblock_bitmap(zone, pfn);
6578 bitidx = pfn_to_bitidx(zone, pfn);
6579 word_bitidx = bitidx / BITS_PER_LONG;
6580 bitidx &= (BITS_PER_LONG-1);
6582 word = bitmap[word_bitidx];
6583 bitidx += end_bitidx;
6584 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6588 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6589 * @page: The page within the block of interest
6590 * @flags: The flags to set
6591 * @pfn: The target page frame number
6592 * @end_bitidx: The last bit of interest
6593 * @mask: mask of bits that the caller is interested in
6595 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6597 unsigned long end_bitidx,
6601 unsigned long *bitmap;
6602 unsigned long bitidx, word_bitidx;
6603 unsigned long old_word, word;
6605 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6607 zone = page_zone(page);
6608 bitmap = get_pageblock_bitmap(zone, pfn);
6609 bitidx = pfn_to_bitidx(zone, pfn);
6610 word_bitidx = bitidx / BITS_PER_LONG;
6611 bitidx &= (BITS_PER_LONG-1);
6613 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6615 bitidx += end_bitidx;
6616 mask <<= (BITS_PER_LONG - bitidx - 1);
6617 flags <<= (BITS_PER_LONG - bitidx - 1);
6619 word = READ_ONCE(bitmap[word_bitidx]);
6621 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6622 if (word == old_word)
6629 * This function checks whether pageblock includes unmovable pages or not.
6630 * If @count is not zero, it is okay to include less @count unmovable pages
6632 * PageLRU check without isolation or lru_lock could race so that
6633 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6634 * expect this function should be exact.
6636 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6637 bool skip_hwpoisoned_pages)
6639 unsigned long pfn, iter, found;
6643 * For avoiding noise data, lru_add_drain_all() should be called
6644 * If ZONE_MOVABLE, the zone never contains unmovable pages
6646 if (zone_idx(zone) == ZONE_MOVABLE)
6648 mt = get_pageblock_migratetype(page);
6649 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6652 pfn = page_to_pfn(page);
6653 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6654 unsigned long check = pfn + iter;
6656 if (!pfn_valid_within(check))
6659 page = pfn_to_page(check);
6662 * Hugepages are not in LRU lists, but they're movable.
6663 * We need not scan over tail pages bacause we don't
6664 * handle each tail page individually in migration.
6666 if (PageHuge(page)) {
6667 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6672 * We can't use page_count without pin a page
6673 * because another CPU can free compound page.
6674 * This check already skips compound tails of THP
6675 * because their page->_count is zero at all time.
6677 if (!atomic_read(&page->_count)) {
6678 if (PageBuddy(page))
6679 iter += (1 << page_order(page)) - 1;
6684 * The HWPoisoned page may be not in buddy system, and
6685 * page_count() is not 0.
6687 if (skip_hwpoisoned_pages && PageHWPoison(page))
6693 * If there are RECLAIMABLE pages, we need to check
6694 * it. But now, memory offline itself doesn't call
6695 * shrink_node_slabs() and it still to be fixed.
6698 * If the page is not RAM, page_count()should be 0.
6699 * we don't need more check. This is an _used_ not-movable page.
6701 * The problematic thing here is PG_reserved pages. PG_reserved
6702 * is set to both of a memory hole page and a _used_ kernel
6711 bool is_pageblock_removable_nolock(struct page *page)
6717 * We have to be careful here because we are iterating over memory
6718 * sections which are not zone aware so we might end up outside of
6719 * the zone but still within the section.
6720 * We have to take care about the node as well. If the node is offline
6721 * its NODE_DATA will be NULL - see page_zone.
6723 if (!node_online(page_to_nid(page)))
6726 zone = page_zone(page);
6727 pfn = page_to_pfn(page);
6728 if (!zone_spans_pfn(zone, pfn))
6731 return !has_unmovable_pages(zone, page, 0, true);
6736 static unsigned long pfn_max_align_down(unsigned long pfn)
6738 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6739 pageblock_nr_pages) - 1);
6742 static unsigned long pfn_max_align_up(unsigned long pfn)
6744 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6745 pageblock_nr_pages));
6748 /* [start, end) must belong to a single zone. */
6749 static int __alloc_contig_migrate_range(struct compact_control *cc,
6750 unsigned long start, unsigned long end)
6752 /* This function is based on compact_zone() from compaction.c. */
6753 unsigned long nr_reclaimed;
6754 unsigned long pfn = start;
6755 unsigned int tries = 0;
6760 while (pfn < end || !list_empty(&cc->migratepages)) {
6761 if (fatal_signal_pending(current)) {
6766 if (list_empty(&cc->migratepages)) {
6767 cc->nr_migratepages = 0;
6768 pfn = isolate_migratepages_range(cc, pfn, end);
6774 } else if (++tries == 5) {
6775 ret = ret < 0 ? ret : -EBUSY;
6779 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6781 cc->nr_migratepages -= nr_reclaimed;
6783 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6784 NULL, 0, cc->mode, MR_CMA);
6787 putback_movable_pages(&cc->migratepages);
6794 * alloc_contig_range() -- tries to allocate given range of pages
6795 * @start: start PFN to allocate
6796 * @end: one-past-the-last PFN to allocate
6797 * @migratetype: migratetype of the underlaying pageblocks (either
6798 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6799 * in range must have the same migratetype and it must
6800 * be either of the two.
6802 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6803 * aligned, however it's the caller's responsibility to guarantee that
6804 * we are the only thread that changes migrate type of pageblocks the
6807 * The PFN range must belong to a single zone.
6809 * Returns zero on success or negative error code. On success all
6810 * pages which PFN is in [start, end) are allocated for the caller and
6811 * need to be freed with free_contig_range().
6813 int alloc_contig_range(unsigned long start, unsigned long end,
6814 unsigned migratetype)
6816 unsigned long outer_start, outer_end;
6819 struct compact_control cc = {
6820 .nr_migratepages = 0,
6822 .zone = page_zone(pfn_to_page(start)),
6823 .mode = MIGRATE_SYNC,
6824 .ignore_skip_hint = true,
6826 INIT_LIST_HEAD(&cc.migratepages);
6829 * What we do here is we mark all pageblocks in range as
6830 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6831 * have different sizes, and due to the way page allocator
6832 * work, we align the range to biggest of the two pages so
6833 * that page allocator won't try to merge buddies from
6834 * different pageblocks and change MIGRATE_ISOLATE to some
6835 * other migration type.
6837 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6838 * migrate the pages from an unaligned range (ie. pages that
6839 * we are interested in). This will put all the pages in
6840 * range back to page allocator as MIGRATE_ISOLATE.
6842 * When this is done, we take the pages in range from page
6843 * allocator removing them from the buddy system. This way
6844 * page allocator will never consider using them.
6846 * This lets us mark the pageblocks back as
6847 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6848 * aligned range but not in the unaligned, original range are
6849 * put back to page allocator so that buddy can use them.
6852 ret = start_isolate_page_range(pfn_max_align_down(start),
6853 pfn_max_align_up(end), migratetype,
6858 ret = __alloc_contig_migrate_range(&cc, start, end);
6863 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6864 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6865 * more, all pages in [start, end) are free in page allocator.
6866 * What we are going to do is to allocate all pages from
6867 * [start, end) (that is remove them from page allocator).
6869 * The only problem is that pages at the beginning and at the
6870 * end of interesting range may be not aligned with pages that
6871 * page allocator holds, ie. they can be part of higher order
6872 * pages. Because of this, we reserve the bigger range and
6873 * once this is done free the pages we are not interested in.
6875 * We don't have to hold zone->lock here because the pages are
6876 * isolated thus they won't get removed from buddy.
6879 lru_add_drain_all();
6880 drain_all_pages(cc.zone);
6883 outer_start = start;
6884 while (!PageBuddy(pfn_to_page(outer_start))) {
6885 if (++order >= MAX_ORDER) {
6889 outer_start &= ~0UL << order;
6892 /* Make sure the range is really isolated. */
6893 if (test_pages_isolated(outer_start, end, false)) {
6894 pr_info("%s: [%lx, %lx) PFNs busy\n",
6895 __func__, outer_start, end);
6900 /* Grab isolated pages from freelists. */
6901 outer_end = isolate_freepages_range(&cc, outer_start, end);
6907 /* Free head and tail (if any) */
6908 if (start != outer_start)
6909 free_contig_range(outer_start, start - outer_start);
6910 if (end != outer_end)
6911 free_contig_range(end, outer_end - end);
6914 undo_isolate_page_range(pfn_max_align_down(start),
6915 pfn_max_align_up(end), migratetype);
6919 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6921 unsigned int count = 0;
6923 for (; nr_pages--; pfn++) {
6924 struct page *page = pfn_to_page(pfn);
6926 count += page_count(page) != 1;
6929 WARN(count != 0, "%d pages are still in use!\n", count);
6933 #ifdef CONFIG_MEMORY_HOTPLUG
6935 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6936 * page high values need to be recalulated.
6938 void __meminit zone_pcp_update(struct zone *zone)
6941 mutex_lock(&pcp_batch_high_lock);
6942 for_each_possible_cpu(cpu)
6943 pageset_set_high_and_batch(zone,
6944 per_cpu_ptr(zone->pageset, cpu));
6945 mutex_unlock(&pcp_batch_high_lock);
6949 void zone_pcp_reset(struct zone *zone)
6951 unsigned long flags;
6953 struct per_cpu_pageset *pset;
6955 /* avoid races with drain_pages() */
6956 local_irq_save(flags);
6957 if (zone->pageset != &boot_pageset) {
6958 for_each_online_cpu(cpu) {
6959 pset = per_cpu_ptr(zone->pageset, cpu);
6960 drain_zonestat(zone, pset);
6962 free_percpu(zone->pageset);
6963 zone->pageset = &boot_pageset;
6965 local_irq_restore(flags);
6968 #ifdef CONFIG_MEMORY_HOTREMOVE
6970 * All pages in the range must be isolated before calling this.
6973 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6977 unsigned int order, i;
6979 unsigned long flags;
6980 /* find the first valid pfn */
6981 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6986 zone = page_zone(pfn_to_page(pfn));
6987 spin_lock_irqsave(&zone->lock, flags);
6989 while (pfn < end_pfn) {
6990 if (!pfn_valid(pfn)) {
6994 page = pfn_to_page(pfn);
6996 * The HWPoisoned page may be not in buddy system, and
6997 * page_count() is not 0.
6999 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7001 SetPageReserved(page);
7005 BUG_ON(page_count(page));
7006 BUG_ON(!PageBuddy(page));
7007 order = page_order(page);
7008 #ifdef CONFIG_DEBUG_VM
7009 printk(KERN_INFO "remove from free list %lx %d %lx\n",
7010 pfn, 1 << order, end_pfn);
7012 list_del(&page->lru);
7013 rmv_page_order(page);
7014 zone->free_area[order].nr_free--;
7015 for (i = 0; i < (1 << order); i++)
7016 SetPageReserved((page+i));
7017 pfn += (1 << order);
7019 spin_unlock_irqrestore(&zone->lock, flags);
7023 #ifdef CONFIG_MEMORY_FAILURE
7024 bool is_free_buddy_page(struct page *page)
7026 struct zone *zone = page_zone(page);
7027 unsigned long pfn = page_to_pfn(page);
7028 unsigned long flags;
7031 spin_lock_irqsave(&zone->lock, flags);
7032 for (order = 0; order < MAX_ORDER; order++) {
7033 struct page *page_head = page - (pfn & ((1 << order) - 1));
7035 if (PageBuddy(page_head) && page_order(page_head) >= order)
7038 spin_unlock_irqrestore(&zone->lock, flags);
7040 return order < MAX_ORDER;