2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/notifier.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/stop_machine.h>
47 #include <linux/sort.h>
48 #include <linux/pfn.h>
49 #include <linux/backing-dev.h>
50 #include <linux/fault-inject.h>
51 #include <linux/page-isolation.h>
52 #include <linux/page_ext.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <linux/prefetch.h>
58 #include <linux/mm_inline.h>
59 #include <linux/migrate.h>
60 #include <linux/page_ext.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/page_owner.h>
64 #include <linux/kthread.h>
66 #include <asm/sections.h>
67 #include <asm/tlbflush.h>
68 #include <asm/div64.h>
71 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
72 static DEFINE_MUTEX(pcp_batch_high_lock);
73 #define MIN_PERCPU_PAGELIST_FRACTION (8)
75 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
76 DEFINE_PER_CPU(int, numa_node);
77 EXPORT_PER_CPU_SYMBOL(numa_node);
80 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
82 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
83 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
84 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
85 * defined in <linux/topology.h>.
87 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
88 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
89 int _node_numa_mem_[MAX_NUMNODES];
93 * Array of node states.
95 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
96 [N_POSSIBLE] = NODE_MASK_ALL,
97 [N_ONLINE] = { { [0] = 1UL } },
99 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
100 #ifdef CONFIG_HIGHMEM
101 [N_HIGH_MEMORY] = { { [0] = 1UL } },
103 #ifdef CONFIG_MOVABLE_NODE
104 [N_MEMORY] = { { [0] = 1UL } },
106 [N_CPU] = { { [0] = 1UL } },
109 EXPORT_SYMBOL(node_states);
111 /* Protect totalram_pages and zone->managed_pages */
112 static DEFINE_SPINLOCK(managed_page_count_lock);
114 unsigned long totalram_pages __read_mostly;
115 unsigned long totalreserve_pages __read_mostly;
116 unsigned long totalcma_pages __read_mostly;
118 * When calculating the number of globally allowed dirty pages, there
119 * is a certain number of per-zone reserves that should not be
120 * considered dirtyable memory. This is the sum of those reserves
121 * over all existing zones that contribute dirtyable memory.
123 unsigned long dirty_balance_reserve __read_mostly;
125 int percpu_pagelist_fraction;
126 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
128 #ifdef CONFIG_PM_SLEEP
130 * The following functions are used by the suspend/hibernate code to temporarily
131 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
132 * while devices are suspended. To avoid races with the suspend/hibernate code,
133 * they should always be called with pm_mutex held (gfp_allowed_mask also should
134 * only be modified with pm_mutex held, unless the suspend/hibernate code is
135 * guaranteed not to run in parallel with that modification).
138 static gfp_t saved_gfp_mask;
140 void pm_restore_gfp_mask(void)
142 WARN_ON(!mutex_is_locked(&pm_mutex));
143 if (saved_gfp_mask) {
144 gfp_allowed_mask = saved_gfp_mask;
149 void pm_restrict_gfp_mask(void)
151 WARN_ON(!mutex_is_locked(&pm_mutex));
152 WARN_ON(saved_gfp_mask);
153 saved_gfp_mask = gfp_allowed_mask;
154 gfp_allowed_mask &= ~GFP_IOFS;
157 bool pm_suspended_storage(void)
159 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
163 #endif /* CONFIG_PM_SLEEP */
165 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
166 int pageblock_order __read_mostly;
169 static void __free_pages_ok(struct page *page, unsigned int order);
172 * results with 256, 32 in the lowmem_reserve sysctl:
173 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
174 * 1G machine -> (16M dma, 784M normal, 224M high)
175 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
176 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
177 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
179 * TBD: should special case ZONE_DMA32 machines here - in those we normally
180 * don't need any ZONE_NORMAL reservation
182 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
183 #ifdef CONFIG_ZONE_DMA
186 #ifdef CONFIG_ZONE_DMA32
189 #ifdef CONFIG_HIGHMEM
195 EXPORT_SYMBOL(totalram_pages);
197 static char * const zone_names[MAX_NR_ZONES] = {
198 #ifdef CONFIG_ZONE_DMA
201 #ifdef CONFIG_ZONE_DMA32
205 #ifdef CONFIG_HIGHMEM
211 int min_free_kbytes = 1024;
212 int user_min_free_kbytes = -1;
214 static unsigned long __meminitdata nr_kernel_pages;
215 static unsigned long __meminitdata nr_all_pages;
216 static unsigned long __meminitdata dma_reserve;
218 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
219 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
220 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
221 static unsigned long __initdata required_kernelcore;
222 static unsigned long __initdata required_movablecore;
223 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
225 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
227 EXPORT_SYMBOL(movable_zone);
228 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
231 int nr_node_ids __read_mostly = MAX_NUMNODES;
232 int nr_online_nodes __read_mostly = 1;
233 EXPORT_SYMBOL(nr_node_ids);
234 EXPORT_SYMBOL(nr_online_nodes);
237 int page_group_by_mobility_disabled __read_mostly;
239 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
240 static inline void reset_deferred_meminit(pg_data_t *pgdat)
242 pgdat->first_deferred_pfn = ULONG_MAX;
245 /* Returns true if the struct page for the pfn is uninitialised */
246 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
248 if (pfn >= NODE_DATA(early_pfn_to_nid(pfn))->first_deferred_pfn)
254 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
256 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
263 * Returns false when the remaining initialisation should be deferred until
264 * later in the boot cycle when it can be parallelised.
266 static inline bool update_defer_init(pg_data_t *pgdat,
267 unsigned long pfn, unsigned long zone_end,
268 unsigned long *nr_initialised)
270 /* Always populate low zones for address-contrained allocations */
271 if (zone_end < pgdat_end_pfn(pgdat))
274 /* Initialise at least 2G of the highest zone */
276 if (*nr_initialised > (2UL << (30 - PAGE_SHIFT)) &&
277 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
278 pgdat->first_deferred_pfn = pfn;
285 static inline void reset_deferred_meminit(pg_data_t *pgdat)
289 static inline bool early_page_uninitialised(unsigned long pfn)
294 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
299 static inline bool update_defer_init(pg_data_t *pgdat,
300 unsigned long pfn, unsigned long zone_end,
301 unsigned long *nr_initialised)
308 void set_pageblock_migratetype(struct page *page, int migratetype)
310 if (unlikely(page_group_by_mobility_disabled &&
311 migratetype < MIGRATE_PCPTYPES))
312 migratetype = MIGRATE_UNMOVABLE;
314 set_pageblock_flags_group(page, (unsigned long)migratetype,
315 PB_migrate, PB_migrate_end);
318 #ifdef CONFIG_DEBUG_VM
319 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
323 unsigned long pfn = page_to_pfn(page);
324 unsigned long sp, start_pfn;
327 seq = zone_span_seqbegin(zone);
328 start_pfn = zone->zone_start_pfn;
329 sp = zone->spanned_pages;
330 if (!zone_spans_pfn(zone, pfn))
332 } while (zone_span_seqretry(zone, seq));
335 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
336 pfn, zone_to_nid(zone), zone->name,
337 start_pfn, start_pfn + sp);
342 static int page_is_consistent(struct zone *zone, struct page *page)
344 if (!pfn_valid_within(page_to_pfn(page)))
346 if (zone != page_zone(page))
352 * Temporary debugging check for pages not lying within a given zone.
354 static int bad_range(struct zone *zone, struct page *page)
356 if (page_outside_zone_boundaries(zone, page))
358 if (!page_is_consistent(zone, page))
364 static inline int bad_range(struct zone *zone, struct page *page)
370 static void bad_page(struct page *page, const char *reason,
371 unsigned long bad_flags)
373 static unsigned long resume;
374 static unsigned long nr_shown;
375 static unsigned long nr_unshown;
377 /* Don't complain about poisoned pages */
378 if (PageHWPoison(page)) {
379 page_mapcount_reset(page); /* remove PageBuddy */
384 * Allow a burst of 60 reports, then keep quiet for that minute;
385 * or allow a steady drip of one report per second.
387 if (nr_shown == 60) {
388 if (time_before(jiffies, resume)) {
394 "BUG: Bad page state: %lu messages suppressed\n",
401 resume = jiffies + 60 * HZ;
403 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
404 current->comm, page_to_pfn(page));
405 dump_page_badflags(page, reason, bad_flags);
410 /* Leave bad fields for debug, except PageBuddy could make trouble */
411 page_mapcount_reset(page); /* remove PageBuddy */
412 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
416 * Higher-order pages are called "compound pages". They are structured thusly:
418 * The first PAGE_SIZE page is called the "head page".
420 * The remaining PAGE_SIZE pages are called "tail pages".
422 * All pages have PG_compound set. All tail pages have their ->first_page
423 * pointing at the head page.
425 * The first tail page's ->lru.next holds the address of the compound page's
426 * put_page() function. Its ->lru.prev holds the order of allocation.
427 * This usage means that zero-order pages may not be compound.
430 static void free_compound_page(struct page *page)
432 __free_pages_ok(page, compound_order(page));
435 void prep_compound_page(struct page *page, unsigned long order)
438 int nr_pages = 1 << order;
440 set_compound_page_dtor(page, free_compound_page);
441 set_compound_order(page, order);
443 for (i = 1; i < nr_pages; i++) {
444 struct page *p = page + i;
445 set_page_count(p, 0);
446 p->first_page = page;
447 /* Make sure p->first_page is always valid for PageTail() */
453 #ifdef CONFIG_DEBUG_PAGEALLOC
454 unsigned int _debug_guardpage_minorder;
455 bool _debug_pagealloc_enabled __read_mostly;
456 bool _debug_guardpage_enabled __read_mostly;
458 static int __init early_debug_pagealloc(char *buf)
463 if (strcmp(buf, "on") == 0)
464 _debug_pagealloc_enabled = true;
468 early_param("debug_pagealloc", early_debug_pagealloc);
470 static bool need_debug_guardpage(void)
472 /* If we don't use debug_pagealloc, we don't need guard page */
473 if (!debug_pagealloc_enabled())
479 static void init_debug_guardpage(void)
481 if (!debug_pagealloc_enabled())
484 _debug_guardpage_enabled = true;
487 struct page_ext_operations debug_guardpage_ops = {
488 .need = need_debug_guardpage,
489 .init = init_debug_guardpage,
492 static int __init debug_guardpage_minorder_setup(char *buf)
496 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
497 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
500 _debug_guardpage_minorder = res;
501 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
504 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
506 static inline void set_page_guard(struct zone *zone, struct page *page,
507 unsigned int order, int migratetype)
509 struct page_ext *page_ext;
511 if (!debug_guardpage_enabled())
514 page_ext = lookup_page_ext(page);
515 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
517 INIT_LIST_HEAD(&page->lru);
518 set_page_private(page, order);
519 /* Guard pages are not available for any usage */
520 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
523 static inline void clear_page_guard(struct zone *zone, struct page *page,
524 unsigned int order, int migratetype)
526 struct page_ext *page_ext;
528 if (!debug_guardpage_enabled())
531 page_ext = lookup_page_ext(page);
532 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
534 set_page_private(page, 0);
535 if (!is_migrate_isolate(migratetype))
536 __mod_zone_freepage_state(zone, (1 << order), migratetype);
539 struct page_ext_operations debug_guardpage_ops = { NULL, };
540 static inline void set_page_guard(struct zone *zone, struct page *page,
541 unsigned int order, int migratetype) {}
542 static inline void clear_page_guard(struct zone *zone, struct page *page,
543 unsigned int order, int migratetype) {}
546 static inline void set_page_order(struct page *page, unsigned int order)
548 set_page_private(page, order);
549 __SetPageBuddy(page);
552 static inline void rmv_page_order(struct page *page)
554 __ClearPageBuddy(page);
555 set_page_private(page, 0);
559 * This function checks whether a page is free && is the buddy
560 * we can do coalesce a page and its buddy if
561 * (a) the buddy is not in a hole &&
562 * (b) the buddy is in the buddy system &&
563 * (c) a page and its buddy have the same order &&
564 * (d) a page and its buddy are in the same zone.
566 * For recording whether a page is in the buddy system, we set ->_mapcount
567 * PAGE_BUDDY_MAPCOUNT_VALUE.
568 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
569 * serialized by zone->lock.
571 * For recording page's order, we use page_private(page).
573 static inline int page_is_buddy(struct page *page, struct page *buddy,
576 if (!pfn_valid_within(page_to_pfn(buddy)))
579 if (page_is_guard(buddy) && page_order(buddy) == order) {
580 if (page_zone_id(page) != page_zone_id(buddy))
583 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
588 if (PageBuddy(buddy) && page_order(buddy) == order) {
590 * zone check is done late to avoid uselessly
591 * calculating zone/node ids for pages that could
594 if (page_zone_id(page) != page_zone_id(buddy))
597 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
605 * Freeing function for a buddy system allocator.
607 * The concept of a buddy system is to maintain direct-mapped table
608 * (containing bit values) for memory blocks of various "orders".
609 * The bottom level table contains the map for the smallest allocatable
610 * units of memory (here, pages), and each level above it describes
611 * pairs of units from the levels below, hence, "buddies".
612 * At a high level, all that happens here is marking the table entry
613 * at the bottom level available, and propagating the changes upward
614 * as necessary, plus some accounting needed to play nicely with other
615 * parts of the VM system.
616 * At each level, we keep a list of pages, which are heads of continuous
617 * free pages of length of (1 << order) and marked with _mapcount
618 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
620 * So when we are allocating or freeing one, we can derive the state of the
621 * other. That is, if we allocate a small block, and both were
622 * free, the remainder of the region must be split into blocks.
623 * If a block is freed, and its buddy is also free, then this
624 * triggers coalescing into a block of larger size.
629 static inline void __free_one_page(struct page *page,
631 struct zone *zone, unsigned int order,
634 unsigned long page_idx;
635 unsigned long combined_idx;
636 unsigned long uninitialized_var(buddy_idx);
638 int max_order = MAX_ORDER;
640 VM_BUG_ON(!zone_is_initialized(zone));
641 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
643 VM_BUG_ON(migratetype == -1);
644 if (is_migrate_isolate(migratetype)) {
646 * We restrict max order of merging to prevent merge
647 * between freepages on isolate pageblock and normal
648 * pageblock. Without this, pageblock isolation
649 * could cause incorrect freepage accounting.
651 max_order = min(MAX_ORDER, pageblock_order + 1);
653 __mod_zone_freepage_state(zone, 1 << order, migratetype);
656 page_idx = pfn & ((1 << max_order) - 1);
658 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
659 VM_BUG_ON_PAGE(bad_range(zone, page), page);
661 while (order < max_order - 1) {
662 buddy_idx = __find_buddy_index(page_idx, order);
663 buddy = page + (buddy_idx - page_idx);
664 if (!page_is_buddy(page, buddy, order))
667 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
668 * merge with it and move up one order.
670 if (page_is_guard(buddy)) {
671 clear_page_guard(zone, buddy, order, migratetype);
673 list_del(&buddy->lru);
674 zone->free_area[order].nr_free--;
675 rmv_page_order(buddy);
677 combined_idx = buddy_idx & page_idx;
678 page = page + (combined_idx - page_idx);
679 page_idx = combined_idx;
682 set_page_order(page, order);
685 * If this is not the largest possible page, check if the buddy
686 * of the next-highest order is free. If it is, it's possible
687 * that pages are being freed that will coalesce soon. In case,
688 * that is happening, add the free page to the tail of the list
689 * so it's less likely to be used soon and more likely to be merged
690 * as a higher order page
692 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
693 struct page *higher_page, *higher_buddy;
694 combined_idx = buddy_idx & page_idx;
695 higher_page = page + (combined_idx - page_idx);
696 buddy_idx = __find_buddy_index(combined_idx, order + 1);
697 higher_buddy = higher_page + (buddy_idx - combined_idx);
698 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
699 list_add_tail(&page->lru,
700 &zone->free_area[order].free_list[migratetype]);
705 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
707 zone->free_area[order].nr_free++;
710 static inline int free_pages_check(struct page *page)
712 const char *bad_reason = NULL;
713 unsigned long bad_flags = 0;
715 if (unlikely(page_mapcount(page)))
716 bad_reason = "nonzero mapcount";
717 if (unlikely(page->mapping != NULL))
718 bad_reason = "non-NULL mapping";
719 if (unlikely(atomic_read(&page->_count) != 0))
720 bad_reason = "nonzero _count";
721 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
722 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
723 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
726 if (unlikely(page->mem_cgroup))
727 bad_reason = "page still charged to cgroup";
729 if (unlikely(bad_reason)) {
730 bad_page(page, bad_reason, bad_flags);
733 page_cpupid_reset_last(page);
734 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
735 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
740 * Frees a number of pages from the PCP lists
741 * Assumes all pages on list are in same zone, and of same order.
742 * count is the number of pages to free.
744 * If the zone was previously in an "all pages pinned" state then look to
745 * see if this freeing clears that state.
747 * And clear the zone's pages_scanned counter, to hold off the "all pages are
748 * pinned" detection logic.
750 static void free_pcppages_bulk(struct zone *zone, int count,
751 struct per_cpu_pages *pcp)
756 unsigned long nr_scanned;
758 spin_lock(&zone->lock);
759 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
761 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
765 struct list_head *list;
768 * Remove pages from lists in a round-robin fashion. A
769 * batch_free count is maintained that is incremented when an
770 * empty list is encountered. This is so more pages are freed
771 * off fuller lists instead of spinning excessively around empty
776 if (++migratetype == MIGRATE_PCPTYPES)
778 list = &pcp->lists[migratetype];
779 } while (list_empty(list));
781 /* This is the only non-empty list. Free them all. */
782 if (batch_free == MIGRATE_PCPTYPES)
783 batch_free = to_free;
786 int mt; /* migratetype of the to-be-freed page */
788 page = list_entry(list->prev, struct page, lru);
789 /* must delete as __free_one_page list manipulates */
790 list_del(&page->lru);
791 mt = get_freepage_migratetype(page);
792 if (unlikely(has_isolate_pageblock(zone)))
793 mt = get_pageblock_migratetype(page);
795 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
796 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
797 trace_mm_page_pcpu_drain(page, 0, mt);
798 } while (--to_free && --batch_free && !list_empty(list));
800 spin_unlock(&zone->lock);
803 static void free_one_page(struct zone *zone,
804 struct page *page, unsigned long pfn,
808 unsigned long nr_scanned;
809 spin_lock(&zone->lock);
810 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
812 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
814 if (unlikely(has_isolate_pageblock(zone) ||
815 is_migrate_isolate(migratetype))) {
816 migratetype = get_pfnblock_migratetype(page, pfn);
818 __free_one_page(page, pfn, zone, order, migratetype);
819 spin_unlock(&zone->lock);
822 static int free_tail_pages_check(struct page *head_page, struct page *page)
824 if (!IS_ENABLED(CONFIG_DEBUG_VM))
826 if (unlikely(!PageTail(page))) {
827 bad_page(page, "PageTail not set", 0);
830 if (unlikely(page->first_page != head_page)) {
831 bad_page(page, "first_page not consistent", 0);
837 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
838 unsigned long zone, int nid)
840 set_page_links(page, zone, nid, pfn);
841 init_page_count(page);
842 page_mapcount_reset(page);
843 page_cpupid_reset_last(page);
845 INIT_LIST_HEAD(&page->lru);
846 #ifdef WANT_PAGE_VIRTUAL
847 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
848 if (!is_highmem_idx(zone))
849 set_page_address(page, __va(pfn << PAGE_SHIFT));
853 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
856 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
859 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
860 static void init_reserved_page(unsigned long pfn)
865 if (!early_page_uninitialised(pfn))
868 nid = early_pfn_to_nid(pfn);
869 pgdat = NODE_DATA(nid);
871 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
872 struct zone *zone = &pgdat->node_zones[zid];
874 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
877 __init_single_pfn(pfn, zid, nid);
880 static inline void init_reserved_page(unsigned long pfn)
883 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
886 * Initialised pages do not have PageReserved set. This function is
887 * called for each range allocated by the bootmem allocator and
888 * marks the pages PageReserved. The remaining valid pages are later
889 * sent to the buddy page allocator.
891 void __meminit reserve_bootmem_region(unsigned long start, unsigned long end)
893 unsigned long start_pfn = PFN_DOWN(start);
894 unsigned long end_pfn = PFN_UP(end);
896 for (; start_pfn < end_pfn; start_pfn++) {
897 if (pfn_valid(start_pfn)) {
898 struct page *page = pfn_to_page(start_pfn);
900 init_reserved_page(start_pfn);
901 SetPageReserved(page);
906 static bool free_pages_prepare(struct page *page, unsigned int order)
908 bool compound = PageCompound(page);
911 VM_BUG_ON_PAGE(PageTail(page), page);
912 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
914 trace_mm_page_free(page, order);
915 kmemcheck_free_shadow(page, order);
916 kasan_free_pages(page, order);
919 page->mapping = NULL;
920 bad += free_pages_check(page);
921 for (i = 1; i < (1 << order); i++) {
923 bad += free_tail_pages_check(page, page + i);
924 bad += free_pages_check(page + i);
929 reset_page_owner(page, order);
931 if (!PageHighMem(page)) {
932 debug_check_no_locks_freed(page_address(page),
934 debug_check_no_obj_freed(page_address(page),
937 arch_free_page(page, order);
938 kernel_map_pages(page, 1 << order, 0);
943 static void __free_pages_ok(struct page *page, unsigned int order)
947 unsigned long pfn = page_to_pfn(page);
949 if (!free_pages_prepare(page, order))
952 migratetype = get_pfnblock_migratetype(page, pfn);
953 local_irq_save(flags);
954 __count_vm_events(PGFREE, 1 << order);
955 set_freepage_migratetype(page, migratetype);
956 free_one_page(page_zone(page), page, pfn, order, migratetype);
957 local_irq_restore(flags);
960 static void __init __free_pages_boot_core(struct page *page,
961 unsigned long pfn, unsigned int order)
963 unsigned int nr_pages = 1 << order;
964 struct page *p = page;
968 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
970 __ClearPageReserved(p);
971 set_page_count(p, 0);
973 __ClearPageReserved(p);
974 set_page_count(p, 0);
976 page_zone(page)->managed_pages += nr_pages;
977 set_page_refcounted(page);
978 __free_pages(page, order);
981 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
982 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
984 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
986 int __meminit early_pfn_to_nid(unsigned long pfn)
988 static DEFINE_SPINLOCK(early_pfn_lock);
991 spin_lock(&early_pfn_lock);
992 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
995 spin_unlock(&early_pfn_lock);
1001 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1002 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1003 struct mminit_pfnnid_cache *state)
1007 nid = __early_pfn_to_nid(pfn, state);
1008 if (nid >= 0 && nid != node)
1013 /* Only safe to use early in boot when initialisation is single-threaded */
1014 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1016 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1021 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1025 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1026 struct mminit_pfnnid_cache *state)
1033 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1036 if (early_page_uninitialised(pfn))
1038 return __free_pages_boot_core(page, pfn, order);
1041 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1042 static void __init deferred_free_range(struct page *page,
1043 unsigned long pfn, int nr_pages)
1050 /* Free a large naturally-aligned chunk if possible */
1051 if (nr_pages == MAX_ORDER_NR_PAGES &&
1052 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1053 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1054 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1058 for (i = 0; i < nr_pages; i++, page++, pfn++)
1059 __free_pages_boot_core(page, pfn, 0);
1062 /* Completion tracking for deferred_init_memmap() threads */
1063 static atomic_t pgdat_init_n_undone __initdata;
1064 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1066 static inline void __init pgdat_init_report_one_done(void)
1068 if (atomic_dec_and_test(&pgdat_init_n_undone))
1069 complete(&pgdat_init_all_done_comp);
1072 /* Initialise remaining memory on a node */
1073 static int __init deferred_init_memmap(void *data)
1075 pg_data_t *pgdat = data;
1076 int nid = pgdat->node_id;
1077 struct mminit_pfnnid_cache nid_init_state = { };
1078 unsigned long start = jiffies;
1079 unsigned long nr_pages = 0;
1080 unsigned long walk_start, walk_end;
1083 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1084 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1086 if (first_init_pfn == ULONG_MAX) {
1087 pgdat_init_report_one_done();
1091 /* Bind memory initialisation thread to a local node if possible */
1092 if (!cpumask_empty(cpumask))
1093 set_cpus_allowed_ptr(current, cpumask);
1095 /* Sanity check boundaries */
1096 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1097 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1098 pgdat->first_deferred_pfn = ULONG_MAX;
1100 /* Only the highest zone is deferred so find it */
1101 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1102 zone = pgdat->node_zones + zid;
1103 if (first_init_pfn < zone_end_pfn(zone))
1107 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1108 unsigned long pfn, end_pfn;
1109 struct page *page = NULL;
1110 struct page *free_base_page = NULL;
1111 unsigned long free_base_pfn = 0;
1114 end_pfn = min(walk_end, zone_end_pfn(zone));
1115 pfn = first_init_pfn;
1116 if (pfn < walk_start)
1118 if (pfn < zone->zone_start_pfn)
1119 pfn = zone->zone_start_pfn;
1121 for (; pfn < end_pfn; pfn++) {
1122 if (!pfn_valid_within(pfn))
1126 * Ensure pfn_valid is checked every
1127 * MAX_ORDER_NR_PAGES for memory holes
1129 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1130 if (!pfn_valid(pfn)) {
1136 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1141 /* Minimise pfn page lookups and scheduler checks */
1142 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1145 nr_pages += nr_to_free;
1146 deferred_free_range(free_base_page,
1147 free_base_pfn, nr_to_free);
1148 free_base_page = NULL;
1149 free_base_pfn = nr_to_free = 0;
1151 page = pfn_to_page(pfn);
1156 VM_BUG_ON(page_zone(page) != zone);
1160 __init_single_page(page, pfn, zid, nid);
1161 if (!free_base_page) {
1162 free_base_page = page;
1163 free_base_pfn = pfn;
1168 /* Where possible, batch up pages for a single free */
1171 /* Free the current block of pages to allocator */
1172 nr_pages += nr_to_free;
1173 deferred_free_range(free_base_page, free_base_pfn,
1175 free_base_page = NULL;
1176 free_base_pfn = nr_to_free = 0;
1179 first_init_pfn = max(end_pfn, first_init_pfn);
1182 /* Sanity check that the next zone really is unpopulated */
1183 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1185 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1186 jiffies_to_msecs(jiffies - start));
1188 pgdat_init_report_one_done();
1192 void __init page_alloc_init_late(void)
1196 /* There will be num_node_state(N_MEMORY) threads */
1197 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1198 for_each_node_state(nid, N_MEMORY) {
1199 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1202 /* Block until all are initialised */
1203 wait_for_completion(&pgdat_init_all_done_comp);
1205 /* Reinit limits that are based on free pages after the kernel is up */
1206 files_maxfiles_init();
1208 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1211 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1212 void __init init_cma_reserved_pageblock(struct page *page)
1214 unsigned i = pageblock_nr_pages;
1215 struct page *p = page;
1218 __ClearPageReserved(p);
1219 set_page_count(p, 0);
1222 set_pageblock_migratetype(page, MIGRATE_CMA);
1224 if (pageblock_order >= MAX_ORDER) {
1225 i = pageblock_nr_pages;
1228 set_page_refcounted(p);
1229 __free_pages(p, MAX_ORDER - 1);
1230 p += MAX_ORDER_NR_PAGES;
1231 } while (i -= MAX_ORDER_NR_PAGES);
1233 set_page_refcounted(page);
1234 __free_pages(page, pageblock_order);
1237 adjust_managed_page_count(page, pageblock_nr_pages);
1242 * The order of subdivision here is critical for the IO subsystem.
1243 * Please do not alter this order without good reasons and regression
1244 * testing. Specifically, as large blocks of memory are subdivided,
1245 * the order in which smaller blocks are delivered depends on the order
1246 * they're subdivided in this function. This is the primary factor
1247 * influencing the order in which pages are delivered to the IO
1248 * subsystem according to empirical testing, and this is also justified
1249 * by considering the behavior of a buddy system containing a single
1250 * large block of memory acted on by a series of small allocations.
1251 * This behavior is a critical factor in sglist merging's success.
1255 static inline void expand(struct zone *zone, struct page *page,
1256 int low, int high, struct free_area *area,
1259 unsigned long size = 1 << high;
1261 while (high > low) {
1265 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1267 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1268 debug_guardpage_enabled() &&
1269 high < debug_guardpage_minorder()) {
1271 * Mark as guard pages (or page), that will allow to
1272 * merge back to allocator when buddy will be freed.
1273 * Corresponding page table entries will not be touched,
1274 * pages will stay not present in virtual address space
1276 set_page_guard(zone, &page[size], high, migratetype);
1279 list_add(&page[size].lru, &area->free_list[migratetype]);
1281 set_page_order(&page[size], high);
1286 * This page is about to be returned from the page allocator
1288 static inline int check_new_page(struct page *page)
1290 const char *bad_reason = NULL;
1291 unsigned long bad_flags = 0;
1293 if (unlikely(page_mapcount(page)))
1294 bad_reason = "nonzero mapcount";
1295 if (unlikely(page->mapping != NULL))
1296 bad_reason = "non-NULL mapping";
1297 if (unlikely(atomic_read(&page->_count) != 0))
1298 bad_reason = "nonzero _count";
1299 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1300 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1301 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1304 if (unlikely(page->mem_cgroup))
1305 bad_reason = "page still charged to cgroup";
1307 if (unlikely(bad_reason)) {
1308 bad_page(page, bad_reason, bad_flags);
1314 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1319 for (i = 0; i < (1 << order); i++) {
1320 struct page *p = page + i;
1321 if (unlikely(check_new_page(p)))
1325 set_page_private(page, 0);
1326 set_page_refcounted(page);
1328 arch_alloc_page(page, order);
1329 kernel_map_pages(page, 1 << order, 1);
1330 kasan_alloc_pages(page, order);
1332 if (gfp_flags & __GFP_ZERO)
1333 for (i = 0; i < (1 << order); i++)
1334 clear_highpage(page + i);
1336 if (order && (gfp_flags & __GFP_COMP))
1337 prep_compound_page(page, order);
1339 set_page_owner(page, order, gfp_flags);
1342 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was necessary to
1343 * allocate the page. The expectation is that the caller is taking
1344 * steps that will free more memory. The caller should avoid the page
1345 * being used for !PFMEMALLOC purposes.
1347 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
1353 * Go through the free lists for the given migratetype and remove
1354 * the smallest available page from the freelists
1357 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1360 unsigned int current_order;
1361 struct free_area *area;
1364 /* Find a page of the appropriate size in the preferred list */
1365 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1366 area = &(zone->free_area[current_order]);
1367 if (list_empty(&area->free_list[migratetype]))
1370 page = list_entry(area->free_list[migratetype].next,
1372 list_del(&page->lru);
1373 rmv_page_order(page);
1375 expand(zone, page, order, current_order, area, migratetype);
1376 set_freepage_migratetype(page, migratetype);
1385 * This array describes the order lists are fallen back to when
1386 * the free lists for the desirable migrate type are depleted
1388 static int fallbacks[MIGRATE_TYPES][4] = {
1389 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1390 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1391 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
1393 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
1395 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
1396 #ifdef CONFIG_MEMORY_ISOLATION
1397 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
1402 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1405 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1408 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1409 unsigned int order) { return NULL; }
1413 * Move the free pages in a range to the free lists of the requested type.
1414 * Note that start_page and end_pages are not aligned on a pageblock
1415 * boundary. If alignment is required, use move_freepages_block()
1417 int move_freepages(struct zone *zone,
1418 struct page *start_page, struct page *end_page,
1422 unsigned long order;
1423 int pages_moved = 0;
1425 #ifndef CONFIG_HOLES_IN_ZONE
1427 * page_zone is not safe to call in this context when
1428 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1429 * anyway as we check zone boundaries in move_freepages_block().
1430 * Remove at a later date when no bug reports exist related to
1431 * grouping pages by mobility
1433 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1436 for (page = start_page; page <= end_page;) {
1437 /* Make sure we are not inadvertently changing nodes */
1438 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1440 if (!pfn_valid_within(page_to_pfn(page))) {
1445 if (!PageBuddy(page)) {
1450 order = page_order(page);
1451 list_move(&page->lru,
1452 &zone->free_area[order].free_list[migratetype]);
1453 set_freepage_migratetype(page, migratetype);
1455 pages_moved += 1 << order;
1461 int move_freepages_block(struct zone *zone, struct page *page,
1464 unsigned long start_pfn, end_pfn;
1465 struct page *start_page, *end_page;
1467 start_pfn = page_to_pfn(page);
1468 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1469 start_page = pfn_to_page(start_pfn);
1470 end_page = start_page + pageblock_nr_pages - 1;
1471 end_pfn = start_pfn + pageblock_nr_pages - 1;
1473 /* Do not cross zone boundaries */
1474 if (!zone_spans_pfn(zone, start_pfn))
1476 if (!zone_spans_pfn(zone, end_pfn))
1479 return move_freepages(zone, start_page, end_page, migratetype);
1482 static void change_pageblock_range(struct page *pageblock_page,
1483 int start_order, int migratetype)
1485 int nr_pageblocks = 1 << (start_order - pageblock_order);
1487 while (nr_pageblocks--) {
1488 set_pageblock_migratetype(pageblock_page, migratetype);
1489 pageblock_page += pageblock_nr_pages;
1494 * When we are falling back to another migratetype during allocation, try to
1495 * steal extra free pages from the same pageblocks to satisfy further
1496 * allocations, instead of polluting multiple pageblocks.
1498 * If we are stealing a relatively large buddy page, it is likely there will
1499 * be more free pages in the pageblock, so try to steal them all. For
1500 * reclaimable and unmovable allocations, we steal regardless of page size,
1501 * as fragmentation caused by those allocations polluting movable pageblocks
1502 * is worse than movable allocations stealing from unmovable and reclaimable
1505 static bool can_steal_fallback(unsigned int order, int start_mt)
1508 * Leaving this order check is intended, although there is
1509 * relaxed order check in next check. The reason is that
1510 * we can actually steal whole pageblock if this condition met,
1511 * but, below check doesn't guarantee it and that is just heuristic
1512 * so could be changed anytime.
1514 if (order >= pageblock_order)
1517 if (order >= pageblock_order / 2 ||
1518 start_mt == MIGRATE_RECLAIMABLE ||
1519 start_mt == MIGRATE_UNMOVABLE ||
1520 page_group_by_mobility_disabled)
1527 * This function implements actual steal behaviour. If order is large enough,
1528 * we can steal whole pageblock. If not, we first move freepages in this
1529 * pageblock and check whether half of pages are moved or not. If half of
1530 * pages are moved, we can change migratetype of pageblock and permanently
1531 * use it's pages as requested migratetype in the future.
1533 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1536 int current_order = page_order(page);
1539 /* Take ownership for orders >= pageblock_order */
1540 if (current_order >= pageblock_order) {
1541 change_pageblock_range(page, current_order, start_type);
1545 pages = move_freepages_block(zone, page, start_type);
1547 /* Claim the whole block if over half of it is free */
1548 if (pages >= (1 << (pageblock_order-1)) ||
1549 page_group_by_mobility_disabled)
1550 set_pageblock_migratetype(page, start_type);
1554 * Check whether there is a suitable fallback freepage with requested order.
1555 * If only_stealable is true, this function returns fallback_mt only if
1556 * we can steal other freepages all together. This would help to reduce
1557 * fragmentation due to mixed migratetype pages in one pageblock.
1559 int find_suitable_fallback(struct free_area *area, unsigned int order,
1560 int migratetype, bool only_stealable, bool *can_steal)
1565 if (area->nr_free == 0)
1570 fallback_mt = fallbacks[migratetype][i];
1571 if (fallback_mt == MIGRATE_RESERVE)
1574 if (list_empty(&area->free_list[fallback_mt]))
1577 if (can_steal_fallback(order, migratetype))
1580 if (!only_stealable)
1590 /* Remove an element from the buddy allocator from the fallback list */
1591 static inline struct page *
1592 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1594 struct free_area *area;
1595 unsigned int current_order;
1600 /* Find the largest possible block of pages in the other list */
1601 for (current_order = MAX_ORDER-1;
1602 current_order >= order && current_order <= MAX_ORDER-1;
1604 area = &(zone->free_area[current_order]);
1605 fallback_mt = find_suitable_fallback(area, current_order,
1606 start_migratetype, false, &can_steal);
1607 if (fallback_mt == -1)
1610 page = list_entry(area->free_list[fallback_mt].next,
1613 steal_suitable_fallback(zone, page, start_migratetype);
1615 /* Remove the page from the freelists */
1617 list_del(&page->lru);
1618 rmv_page_order(page);
1620 expand(zone, page, order, current_order, area,
1623 * The freepage_migratetype may differ from pageblock's
1624 * migratetype depending on the decisions in
1625 * try_to_steal_freepages(). This is OK as long as it
1626 * does not differ for MIGRATE_CMA pageblocks. For CMA
1627 * we need to make sure unallocated pages flushed from
1628 * pcp lists are returned to the correct freelist.
1630 set_freepage_migratetype(page, start_migratetype);
1632 trace_mm_page_alloc_extfrag(page, order, current_order,
1633 start_migratetype, fallback_mt);
1642 * Do the hard work of removing an element from the buddy allocator.
1643 * Call me with the zone->lock already held.
1645 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1651 page = __rmqueue_smallest(zone, order, migratetype);
1653 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1654 if (migratetype == MIGRATE_MOVABLE)
1655 page = __rmqueue_cma_fallback(zone, order);
1658 page = __rmqueue_fallback(zone, order, migratetype);
1661 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1662 * is used because __rmqueue_smallest is an inline function
1663 * and we want just one call site
1666 migratetype = MIGRATE_RESERVE;
1671 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1676 * Obtain a specified number of elements from the buddy allocator, all under
1677 * a single hold of the lock, for efficiency. Add them to the supplied list.
1678 * Returns the number of new pages which were placed at *list.
1680 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1681 unsigned long count, struct list_head *list,
1682 int migratetype, bool cold)
1686 spin_lock(&zone->lock);
1687 for (i = 0; i < count; ++i) {
1688 struct page *page = __rmqueue(zone, order, migratetype);
1689 if (unlikely(page == NULL))
1693 * Split buddy pages returned by expand() are received here
1694 * in physical page order. The page is added to the callers and
1695 * list and the list head then moves forward. From the callers
1696 * perspective, the linked list is ordered by page number in
1697 * some conditions. This is useful for IO devices that can
1698 * merge IO requests if the physical pages are ordered
1702 list_add(&page->lru, list);
1704 list_add_tail(&page->lru, list);
1706 if (is_migrate_cma(get_freepage_migratetype(page)))
1707 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1710 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1711 spin_unlock(&zone->lock);
1717 * Called from the vmstat counter updater to drain pagesets of this
1718 * currently executing processor on remote nodes after they have
1721 * Note that this function must be called with the thread pinned to
1722 * a single processor.
1724 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1726 unsigned long flags;
1727 int to_drain, batch;
1729 local_irq_save(flags);
1730 batch = READ_ONCE(pcp->batch);
1731 to_drain = min(pcp->count, batch);
1733 free_pcppages_bulk(zone, to_drain, pcp);
1734 pcp->count -= to_drain;
1736 local_irq_restore(flags);
1741 * Drain pcplists of the indicated processor and zone.
1743 * The processor must either be the current processor and the
1744 * thread pinned to the current processor or a processor that
1747 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1749 unsigned long flags;
1750 struct per_cpu_pageset *pset;
1751 struct per_cpu_pages *pcp;
1753 local_irq_save(flags);
1754 pset = per_cpu_ptr(zone->pageset, cpu);
1758 free_pcppages_bulk(zone, pcp->count, pcp);
1761 local_irq_restore(flags);
1765 * Drain pcplists of all zones on the indicated processor.
1767 * The processor must either be the current processor and the
1768 * thread pinned to the current processor or a processor that
1771 static void drain_pages(unsigned int cpu)
1775 for_each_populated_zone(zone) {
1776 drain_pages_zone(cpu, zone);
1781 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1783 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1784 * the single zone's pages.
1786 void drain_local_pages(struct zone *zone)
1788 int cpu = smp_processor_id();
1791 drain_pages_zone(cpu, zone);
1797 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1799 * When zone parameter is non-NULL, spill just the single zone's pages.
1801 * Note that this code is protected against sending an IPI to an offline
1802 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1803 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1804 * nothing keeps CPUs from showing up after we populated the cpumask and
1805 * before the call to on_each_cpu_mask().
1807 void drain_all_pages(struct zone *zone)
1812 * Allocate in the BSS so we wont require allocation in
1813 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1815 static cpumask_t cpus_with_pcps;
1818 * We don't care about racing with CPU hotplug event
1819 * as offline notification will cause the notified
1820 * cpu to drain that CPU pcps and on_each_cpu_mask
1821 * disables preemption as part of its processing
1823 for_each_online_cpu(cpu) {
1824 struct per_cpu_pageset *pcp;
1826 bool has_pcps = false;
1829 pcp = per_cpu_ptr(zone->pageset, cpu);
1833 for_each_populated_zone(z) {
1834 pcp = per_cpu_ptr(z->pageset, cpu);
1835 if (pcp->pcp.count) {
1843 cpumask_set_cpu(cpu, &cpus_with_pcps);
1845 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1847 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
1851 #ifdef CONFIG_HIBERNATION
1853 void mark_free_pages(struct zone *zone)
1855 unsigned long pfn, max_zone_pfn;
1856 unsigned long flags;
1857 unsigned int order, t;
1858 struct list_head *curr;
1860 if (zone_is_empty(zone))
1863 spin_lock_irqsave(&zone->lock, flags);
1865 max_zone_pfn = zone_end_pfn(zone);
1866 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1867 if (pfn_valid(pfn)) {
1868 struct page *page = pfn_to_page(pfn);
1870 if (!swsusp_page_is_forbidden(page))
1871 swsusp_unset_page_free(page);
1874 for_each_migratetype_order(order, t) {
1875 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1878 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1879 for (i = 0; i < (1UL << order); i++)
1880 swsusp_set_page_free(pfn_to_page(pfn + i));
1883 spin_unlock_irqrestore(&zone->lock, flags);
1885 #endif /* CONFIG_PM */
1888 * Free a 0-order page
1889 * cold == true ? free a cold page : free a hot page
1891 void free_hot_cold_page(struct page *page, bool cold)
1893 struct zone *zone = page_zone(page);
1894 struct per_cpu_pages *pcp;
1895 unsigned long flags;
1896 unsigned long pfn = page_to_pfn(page);
1899 if (!free_pages_prepare(page, 0))
1902 migratetype = get_pfnblock_migratetype(page, pfn);
1903 set_freepage_migratetype(page, migratetype);
1904 local_irq_save(flags);
1905 __count_vm_event(PGFREE);
1908 * We only track unmovable, reclaimable and movable on pcp lists.
1909 * Free ISOLATE pages back to the allocator because they are being
1910 * offlined but treat RESERVE as movable pages so we can get those
1911 * areas back if necessary. Otherwise, we may have to free
1912 * excessively into the page allocator
1914 if (migratetype >= MIGRATE_PCPTYPES) {
1915 if (unlikely(is_migrate_isolate(migratetype))) {
1916 free_one_page(zone, page, pfn, 0, migratetype);
1919 migratetype = MIGRATE_MOVABLE;
1922 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1924 list_add(&page->lru, &pcp->lists[migratetype]);
1926 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1928 if (pcp->count >= pcp->high) {
1929 unsigned long batch = READ_ONCE(pcp->batch);
1930 free_pcppages_bulk(zone, batch, pcp);
1931 pcp->count -= batch;
1935 local_irq_restore(flags);
1939 * Free a list of 0-order pages
1941 void free_hot_cold_page_list(struct list_head *list, bool cold)
1943 struct page *page, *next;
1945 list_for_each_entry_safe(page, next, list, lru) {
1946 trace_mm_page_free_batched(page, cold);
1947 free_hot_cold_page(page, cold);
1952 * split_page takes a non-compound higher-order page, and splits it into
1953 * n (1<<order) sub-pages: page[0..n]
1954 * Each sub-page must be freed individually.
1956 * Note: this is probably too low level an operation for use in drivers.
1957 * Please consult with lkml before using this in your driver.
1959 void split_page(struct page *page, unsigned int order)
1964 VM_BUG_ON_PAGE(PageCompound(page), page);
1965 VM_BUG_ON_PAGE(!page_count(page), page);
1967 #ifdef CONFIG_KMEMCHECK
1969 * Split shadow pages too, because free(page[0]) would
1970 * otherwise free the whole shadow.
1972 if (kmemcheck_page_is_tracked(page))
1973 split_page(virt_to_page(page[0].shadow), order);
1976 gfp_mask = get_page_owner_gfp(page);
1977 set_page_owner(page, 0, gfp_mask);
1978 for (i = 1; i < (1 << order); i++) {
1979 set_page_refcounted(page + i);
1980 set_page_owner(page + i, 0, gfp_mask);
1983 EXPORT_SYMBOL_GPL(split_page);
1985 int __isolate_free_page(struct page *page, unsigned int order)
1987 unsigned long watermark;
1991 BUG_ON(!PageBuddy(page));
1993 zone = page_zone(page);
1994 mt = get_pageblock_migratetype(page);
1996 if (!is_migrate_isolate(mt)) {
1997 /* Obey watermarks as if the page was being allocated */
1998 watermark = low_wmark_pages(zone) + (1 << order);
1999 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2002 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2005 /* Remove page from free list */
2006 list_del(&page->lru);
2007 zone->free_area[order].nr_free--;
2008 rmv_page_order(page);
2010 set_page_owner(page, order, __GFP_MOVABLE);
2012 /* Set the pageblock if the isolated page is at least a pageblock */
2013 if (order >= pageblock_order - 1) {
2014 struct page *endpage = page + (1 << order) - 1;
2015 for (; page < endpage; page += pageblock_nr_pages) {
2016 int mt = get_pageblock_migratetype(page);
2017 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2018 set_pageblock_migratetype(page,
2024 return 1UL << order;
2028 * Similar to split_page except the page is already free. As this is only
2029 * being used for migration, the migratetype of the block also changes.
2030 * As this is called with interrupts disabled, the caller is responsible
2031 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2034 * Note: this is probably too low level an operation for use in drivers.
2035 * Please consult with lkml before using this in your driver.
2037 int split_free_page(struct page *page)
2042 order = page_order(page);
2044 nr_pages = __isolate_free_page(page, order);
2048 /* Split into individual pages */
2049 set_page_refcounted(page);
2050 split_page(page, order);
2055 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2058 struct page *buffered_rmqueue(struct zone *preferred_zone,
2059 struct zone *zone, unsigned int order,
2060 gfp_t gfp_flags, int migratetype)
2062 unsigned long flags;
2064 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2066 if (likely(order == 0)) {
2067 struct per_cpu_pages *pcp;
2068 struct list_head *list;
2070 local_irq_save(flags);
2071 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2072 list = &pcp->lists[migratetype];
2073 if (list_empty(list)) {
2074 pcp->count += rmqueue_bulk(zone, 0,
2077 if (unlikely(list_empty(list)))
2082 page = list_entry(list->prev, struct page, lru);
2084 page = list_entry(list->next, struct page, lru);
2086 list_del(&page->lru);
2089 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2091 * __GFP_NOFAIL is not to be used in new code.
2093 * All __GFP_NOFAIL callers should be fixed so that they
2094 * properly detect and handle allocation failures.
2096 * We most definitely don't want callers attempting to
2097 * allocate greater than order-1 page units with
2100 WARN_ON_ONCE(order > 1);
2102 spin_lock_irqsave(&zone->lock, flags);
2103 page = __rmqueue(zone, order, migratetype);
2104 spin_unlock(&zone->lock);
2107 __mod_zone_freepage_state(zone, -(1 << order),
2108 get_freepage_migratetype(page));
2111 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2112 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2113 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2114 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2116 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2117 zone_statistics(preferred_zone, zone, gfp_flags);
2118 local_irq_restore(flags);
2120 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2124 local_irq_restore(flags);
2128 #ifdef CONFIG_FAIL_PAGE_ALLOC
2131 struct fault_attr attr;
2133 u32 ignore_gfp_highmem;
2134 u32 ignore_gfp_wait;
2136 } fail_page_alloc = {
2137 .attr = FAULT_ATTR_INITIALIZER,
2138 .ignore_gfp_wait = 1,
2139 .ignore_gfp_highmem = 1,
2143 static int __init setup_fail_page_alloc(char *str)
2145 return setup_fault_attr(&fail_page_alloc.attr, str);
2147 __setup("fail_page_alloc=", setup_fail_page_alloc);
2149 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2151 if (order < fail_page_alloc.min_order)
2153 if (gfp_mask & __GFP_NOFAIL)
2155 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2157 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
2160 return should_fail(&fail_page_alloc.attr, 1 << order);
2163 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2165 static int __init fail_page_alloc_debugfs(void)
2167 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2170 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2171 &fail_page_alloc.attr);
2173 return PTR_ERR(dir);
2175 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2176 &fail_page_alloc.ignore_gfp_wait))
2178 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2179 &fail_page_alloc.ignore_gfp_highmem))
2181 if (!debugfs_create_u32("min-order", mode, dir,
2182 &fail_page_alloc.min_order))
2187 debugfs_remove_recursive(dir);
2192 late_initcall(fail_page_alloc_debugfs);
2194 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2196 #else /* CONFIG_FAIL_PAGE_ALLOC */
2198 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2203 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2206 * Return true if free pages are above 'mark'. This takes into account the order
2207 * of the allocation.
2209 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2210 unsigned long mark, int classzone_idx, int alloc_flags,
2213 /* free_pages may go negative - that's OK */
2218 free_pages -= (1 << order) - 1;
2219 if (alloc_flags & ALLOC_HIGH)
2221 if (alloc_flags & ALLOC_HARDER)
2224 /* If allocation can't use CMA areas don't use free CMA pages */
2225 if (!(alloc_flags & ALLOC_CMA))
2226 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
2229 if (free_pages - free_cma <= min + z->lowmem_reserve[classzone_idx])
2231 for (o = 0; o < order; o++) {
2232 /* At the next order, this order's pages become unavailable */
2233 free_pages -= z->free_area[o].nr_free << o;
2235 /* Require fewer higher order pages to be free */
2238 if (free_pages <= min)
2244 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2245 int classzone_idx, int alloc_flags)
2247 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2248 zone_page_state(z, NR_FREE_PAGES));
2251 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2252 unsigned long mark, int classzone_idx, int alloc_flags)
2254 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2256 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2257 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2259 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2265 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
2266 * skip over zones that are not allowed by the cpuset, or that have
2267 * been recently (in last second) found to be nearly full. See further
2268 * comments in mmzone.h. Reduces cache footprint of zonelist scans
2269 * that have to skip over a lot of full or unallowed zones.
2271 * If the zonelist cache is present in the passed zonelist, then
2272 * returns a pointer to the allowed node mask (either the current
2273 * tasks mems_allowed, or node_states[N_MEMORY].)
2275 * If the zonelist cache is not available for this zonelist, does
2276 * nothing and returns NULL.
2278 * If the fullzones BITMAP in the zonelist cache is stale (more than
2279 * a second since last zap'd) then we zap it out (clear its bits.)
2281 * We hold off even calling zlc_setup, until after we've checked the
2282 * first zone in the zonelist, on the theory that most allocations will
2283 * be satisfied from that first zone, so best to examine that zone as
2284 * quickly as we can.
2286 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
2288 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2289 nodemask_t *allowednodes; /* zonelist_cache approximation */
2291 zlc = zonelist->zlcache_ptr;
2295 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
2296 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2297 zlc->last_full_zap = jiffies;
2300 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
2301 &cpuset_current_mems_allowed :
2302 &node_states[N_MEMORY];
2303 return allowednodes;
2307 * Given 'z' scanning a zonelist, run a couple of quick checks to see
2308 * if it is worth looking at further for free memory:
2309 * 1) Check that the zone isn't thought to be full (doesn't have its
2310 * bit set in the zonelist_cache fullzones BITMAP).
2311 * 2) Check that the zones node (obtained from the zonelist_cache
2312 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
2313 * Return true (non-zero) if zone is worth looking at further, or
2314 * else return false (zero) if it is not.
2316 * This check -ignores- the distinction between various watermarks,
2317 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
2318 * found to be full for any variation of these watermarks, it will
2319 * be considered full for up to one second by all requests, unless
2320 * we are so low on memory on all allowed nodes that we are forced
2321 * into the second scan of the zonelist.
2323 * In the second scan we ignore this zonelist cache and exactly
2324 * apply the watermarks to all zones, even it is slower to do so.
2325 * We are low on memory in the second scan, and should leave no stone
2326 * unturned looking for a free page.
2328 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
2329 nodemask_t *allowednodes)
2331 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2332 int i; /* index of *z in zonelist zones */
2333 int n; /* node that zone *z is on */
2335 zlc = zonelist->zlcache_ptr;
2339 i = z - zonelist->_zonerefs;
2342 /* This zone is worth trying if it is allowed but not full */
2343 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
2347 * Given 'z' scanning a zonelist, set the corresponding bit in
2348 * zlc->fullzones, so that subsequent attempts to allocate a page
2349 * from that zone don't waste time re-examining it.
2351 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
2353 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2354 int i; /* index of *z in zonelist zones */
2356 zlc = zonelist->zlcache_ptr;
2360 i = z - zonelist->_zonerefs;
2362 set_bit(i, zlc->fullzones);
2366 * clear all zones full, called after direct reclaim makes progress so that
2367 * a zone that was recently full is not skipped over for up to a second
2369 static void zlc_clear_zones_full(struct zonelist *zonelist)
2371 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2373 zlc = zonelist->zlcache_ptr;
2377 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2380 static bool zone_local(struct zone *local_zone, struct zone *zone)
2382 return local_zone->node == zone->node;
2385 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2387 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2391 #else /* CONFIG_NUMA */
2393 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
2398 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
2399 nodemask_t *allowednodes)
2404 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
2408 static void zlc_clear_zones_full(struct zonelist *zonelist)
2412 static bool zone_local(struct zone *local_zone, struct zone *zone)
2417 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2422 #endif /* CONFIG_NUMA */
2424 static void reset_alloc_batches(struct zone *preferred_zone)
2426 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2429 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2430 high_wmark_pages(zone) - low_wmark_pages(zone) -
2431 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2432 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2433 } while (zone++ != preferred_zone);
2437 * get_page_from_freelist goes through the zonelist trying to allocate
2440 static struct page *
2441 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2442 const struct alloc_context *ac)
2444 struct zonelist *zonelist = ac->zonelist;
2446 struct page *page = NULL;
2448 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
2449 int zlc_active = 0; /* set if using zonelist_cache */
2450 int did_zlc_setup = 0; /* just call zlc_setup() one time */
2451 bool consider_zone_dirty = (alloc_flags & ALLOC_WMARK_LOW) &&
2452 (gfp_mask & __GFP_WRITE);
2453 int nr_fair_skipped = 0;
2454 bool zonelist_rescan;
2457 zonelist_rescan = false;
2460 * Scan zonelist, looking for a zone with enough free.
2461 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2463 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2467 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2468 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2470 if (cpusets_enabled() &&
2471 (alloc_flags & ALLOC_CPUSET) &&
2472 !cpuset_zone_allowed(zone, gfp_mask))
2475 * Distribute pages in proportion to the individual
2476 * zone size to ensure fair page aging. The zone a
2477 * page was allocated in should have no effect on the
2478 * time the page has in memory before being reclaimed.
2480 if (alloc_flags & ALLOC_FAIR) {
2481 if (!zone_local(ac->preferred_zone, zone))
2483 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2489 * When allocating a page cache page for writing, we
2490 * want to get it from a zone that is within its dirty
2491 * limit, such that no single zone holds more than its
2492 * proportional share of globally allowed dirty pages.
2493 * The dirty limits take into account the zone's
2494 * lowmem reserves and high watermark so that kswapd
2495 * should be able to balance it without having to
2496 * write pages from its LRU list.
2498 * This may look like it could increase pressure on
2499 * lower zones by failing allocations in higher zones
2500 * before they are full. But the pages that do spill
2501 * over are limited as the lower zones are protected
2502 * by this very same mechanism. It should not become
2503 * a practical burden to them.
2505 * XXX: For now, allow allocations to potentially
2506 * exceed the per-zone dirty limit in the slowpath
2507 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2508 * which is important when on a NUMA setup the allowed
2509 * zones are together not big enough to reach the
2510 * global limit. The proper fix for these situations
2511 * will require awareness of zones in the
2512 * dirty-throttling and the flusher threads.
2514 if (consider_zone_dirty && !zone_dirty_ok(zone))
2517 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2518 if (!zone_watermark_ok(zone, order, mark,
2519 ac->classzone_idx, alloc_flags)) {
2522 /* Checked here to keep the fast path fast */
2523 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2524 if (alloc_flags & ALLOC_NO_WATERMARKS)
2527 if (IS_ENABLED(CONFIG_NUMA) &&
2528 !did_zlc_setup && nr_online_nodes > 1) {
2530 * we do zlc_setup if there are multiple nodes
2531 * and before considering the first zone allowed
2534 allowednodes = zlc_setup(zonelist, alloc_flags);
2539 if (zone_reclaim_mode == 0 ||
2540 !zone_allows_reclaim(ac->preferred_zone, zone))
2541 goto this_zone_full;
2544 * As we may have just activated ZLC, check if the first
2545 * eligible zone has failed zone_reclaim recently.
2547 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2548 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2551 ret = zone_reclaim(zone, gfp_mask, order);
2553 case ZONE_RECLAIM_NOSCAN:
2556 case ZONE_RECLAIM_FULL:
2557 /* scanned but unreclaimable */
2560 /* did we reclaim enough */
2561 if (zone_watermark_ok(zone, order, mark,
2562 ac->classzone_idx, alloc_flags))
2566 * Failed to reclaim enough to meet watermark.
2567 * Only mark the zone full if checking the min
2568 * watermark or if we failed to reclaim just
2569 * 1<<order pages or else the page allocator
2570 * fastpath will prematurely mark zones full
2571 * when the watermark is between the low and
2574 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2575 ret == ZONE_RECLAIM_SOME)
2576 goto this_zone_full;
2583 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2584 gfp_mask, ac->migratetype);
2586 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2591 if (IS_ENABLED(CONFIG_NUMA) && zlc_active)
2592 zlc_mark_zone_full(zonelist, z);
2596 * The first pass makes sure allocations are spread fairly within the
2597 * local node. However, the local node might have free pages left
2598 * after the fairness batches are exhausted, and remote zones haven't
2599 * even been considered yet. Try once more without fairness, and
2600 * include remote zones now, before entering the slowpath and waking
2601 * kswapd: prefer spilling to a remote zone over swapping locally.
2603 if (alloc_flags & ALLOC_FAIR) {
2604 alloc_flags &= ~ALLOC_FAIR;
2605 if (nr_fair_skipped) {
2606 zonelist_rescan = true;
2607 reset_alloc_batches(ac->preferred_zone);
2609 if (nr_online_nodes > 1)
2610 zonelist_rescan = true;
2613 if (unlikely(IS_ENABLED(CONFIG_NUMA) && zlc_active)) {
2614 /* Disable zlc cache for second zonelist scan */
2616 zonelist_rescan = true;
2619 if (zonelist_rescan)
2626 * Large machines with many possible nodes should not always dump per-node
2627 * meminfo in irq context.
2629 static inline bool should_suppress_show_mem(void)
2634 ret = in_interrupt();
2639 static DEFINE_RATELIMIT_STATE(nopage_rs,
2640 DEFAULT_RATELIMIT_INTERVAL,
2641 DEFAULT_RATELIMIT_BURST);
2643 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2645 unsigned int filter = SHOW_MEM_FILTER_NODES;
2647 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2648 debug_guardpage_minorder() > 0)
2652 * This documents exceptions given to allocations in certain
2653 * contexts that are allowed to allocate outside current's set
2656 if (!(gfp_mask & __GFP_NOMEMALLOC))
2657 if (test_thread_flag(TIF_MEMDIE) ||
2658 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2659 filter &= ~SHOW_MEM_FILTER_NODES;
2660 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2661 filter &= ~SHOW_MEM_FILTER_NODES;
2664 struct va_format vaf;
2667 va_start(args, fmt);
2672 pr_warn("%pV", &vaf);
2677 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2678 current->comm, order, gfp_mask);
2681 if (!should_suppress_show_mem())
2685 static inline struct page *
2686 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2687 const struct alloc_context *ac, unsigned long *did_some_progress)
2691 *did_some_progress = 0;
2694 * Acquire the oom lock. If that fails, somebody else is
2695 * making progress for us.
2697 if (!mutex_trylock(&oom_lock)) {
2698 *did_some_progress = 1;
2699 schedule_timeout_uninterruptible(1);
2704 * Go through the zonelist yet one more time, keep very high watermark
2705 * here, this is only to catch a parallel oom killing, we must fail if
2706 * we're still under heavy pressure.
2708 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2709 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2713 if (!(gfp_mask & __GFP_NOFAIL)) {
2714 /* Coredumps can quickly deplete all memory reserves */
2715 if (current->flags & PF_DUMPCORE)
2717 /* The OOM killer will not help higher order allocs */
2718 if (order > PAGE_ALLOC_COSTLY_ORDER)
2720 /* The OOM killer does not needlessly kill tasks for lowmem */
2721 if (ac->high_zoneidx < ZONE_NORMAL)
2723 /* The OOM killer does not compensate for IO-less reclaim */
2724 if (!(gfp_mask & __GFP_FS)) {
2726 * XXX: Page reclaim didn't yield anything,
2727 * and the OOM killer can't be invoked, but
2728 * keep looping as per tradition.
2730 *did_some_progress = 1;
2733 if (pm_suspended_storage())
2735 /* The OOM killer may not free memory on a specific node */
2736 if (gfp_mask & __GFP_THISNODE)
2739 /* Exhausted what can be done so it's blamo time */
2740 if (out_of_memory(ac->zonelist, gfp_mask, order, ac->nodemask, false)
2741 || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL))
2742 *did_some_progress = 1;
2744 mutex_unlock(&oom_lock);
2748 #ifdef CONFIG_COMPACTION
2749 /* Try memory compaction for high-order allocations before reclaim */
2750 static struct page *
2751 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2752 int alloc_flags, const struct alloc_context *ac,
2753 enum migrate_mode mode, int *contended_compaction,
2754 bool *deferred_compaction)
2756 unsigned long compact_result;
2762 current->flags |= PF_MEMALLOC;
2763 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2764 mode, contended_compaction);
2765 current->flags &= ~PF_MEMALLOC;
2767 switch (compact_result) {
2768 case COMPACT_DEFERRED:
2769 *deferred_compaction = true;
2771 case COMPACT_SKIPPED:
2778 * At least in one zone compaction wasn't deferred or skipped, so let's
2779 * count a compaction stall
2781 count_vm_event(COMPACTSTALL);
2783 page = get_page_from_freelist(gfp_mask, order,
2784 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2787 struct zone *zone = page_zone(page);
2789 zone->compact_blockskip_flush = false;
2790 compaction_defer_reset(zone, order, true);
2791 count_vm_event(COMPACTSUCCESS);
2796 * It's bad if compaction run occurs and fails. The most likely reason
2797 * is that pages exist, but not enough to satisfy watermarks.
2799 count_vm_event(COMPACTFAIL);
2806 static inline struct page *
2807 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2808 int alloc_flags, const struct alloc_context *ac,
2809 enum migrate_mode mode, int *contended_compaction,
2810 bool *deferred_compaction)
2814 #endif /* CONFIG_COMPACTION */
2816 /* Perform direct synchronous page reclaim */
2818 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2819 const struct alloc_context *ac)
2821 struct reclaim_state reclaim_state;
2826 /* We now go into synchronous reclaim */
2827 cpuset_memory_pressure_bump();
2828 current->flags |= PF_MEMALLOC;
2829 lockdep_set_current_reclaim_state(gfp_mask);
2830 reclaim_state.reclaimed_slab = 0;
2831 current->reclaim_state = &reclaim_state;
2833 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2836 current->reclaim_state = NULL;
2837 lockdep_clear_current_reclaim_state();
2838 current->flags &= ~PF_MEMALLOC;
2845 /* The really slow allocator path where we enter direct reclaim */
2846 static inline struct page *
2847 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2848 int alloc_flags, const struct alloc_context *ac,
2849 unsigned long *did_some_progress)
2851 struct page *page = NULL;
2852 bool drained = false;
2854 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2855 if (unlikely(!(*did_some_progress)))
2858 /* After successful reclaim, reconsider all zones for allocation */
2859 if (IS_ENABLED(CONFIG_NUMA))
2860 zlc_clear_zones_full(ac->zonelist);
2863 page = get_page_from_freelist(gfp_mask, order,
2864 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2867 * If an allocation failed after direct reclaim, it could be because
2868 * pages are pinned on the per-cpu lists. Drain them and try again
2870 if (!page && !drained) {
2871 drain_all_pages(NULL);
2880 * This is called in the allocator slow-path if the allocation request is of
2881 * sufficient urgency to ignore watermarks and take other desperate measures
2883 static inline struct page *
2884 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2885 const struct alloc_context *ac)
2890 page = get_page_from_freelist(gfp_mask, order,
2891 ALLOC_NO_WATERMARKS, ac);
2893 if (!page && gfp_mask & __GFP_NOFAIL)
2894 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC,
2896 } while (!page && (gfp_mask & __GFP_NOFAIL));
2901 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2906 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2907 ac->high_zoneidx, ac->nodemask)
2908 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2912 gfp_to_alloc_flags(gfp_t gfp_mask)
2914 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2915 const bool atomic = !(gfp_mask & (__GFP_WAIT | __GFP_NO_KSWAPD));
2917 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2918 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2921 * The caller may dip into page reserves a bit more if the caller
2922 * cannot run direct reclaim, or if the caller has realtime scheduling
2923 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2924 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2926 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2930 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2931 * if it can't schedule.
2933 if (!(gfp_mask & __GFP_NOMEMALLOC))
2934 alloc_flags |= ALLOC_HARDER;
2936 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2937 * comment for __cpuset_node_allowed().
2939 alloc_flags &= ~ALLOC_CPUSET;
2940 } else if (unlikely(rt_task(current)) && !in_interrupt())
2941 alloc_flags |= ALLOC_HARDER;
2943 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2944 if (gfp_mask & __GFP_MEMALLOC)
2945 alloc_flags |= ALLOC_NO_WATERMARKS;
2946 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2947 alloc_flags |= ALLOC_NO_WATERMARKS;
2948 else if (!in_interrupt() &&
2949 ((current->flags & PF_MEMALLOC) ||
2950 unlikely(test_thread_flag(TIF_MEMDIE))))
2951 alloc_flags |= ALLOC_NO_WATERMARKS;
2954 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2955 alloc_flags |= ALLOC_CMA;
2960 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2962 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2965 static inline struct page *
2966 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2967 struct alloc_context *ac)
2969 const gfp_t wait = gfp_mask & __GFP_WAIT;
2970 struct page *page = NULL;
2972 unsigned long pages_reclaimed = 0;
2973 unsigned long did_some_progress;
2974 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2975 bool deferred_compaction = false;
2976 int contended_compaction = COMPACT_CONTENDED_NONE;
2979 * In the slowpath, we sanity check order to avoid ever trying to
2980 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2981 * be using allocators in order of preference for an area that is
2984 if (order >= MAX_ORDER) {
2985 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2990 * If this allocation cannot block and it is for a specific node, then
2991 * fail early. There's no need to wakeup kswapd or retry for a
2992 * speculative node-specific allocation.
2994 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !wait)
2998 if (!(gfp_mask & __GFP_NO_KSWAPD))
2999 wake_all_kswapds(order, ac);
3002 * OK, we're below the kswapd watermark and have kicked background
3003 * reclaim. Now things get more complex, so set up alloc_flags according
3004 * to how we want to proceed.
3006 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3009 * Find the true preferred zone if the allocation is unconstrained by
3012 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3013 struct zoneref *preferred_zoneref;
3014 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3015 ac->high_zoneidx, NULL, &ac->preferred_zone);
3016 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3019 /* This is the last chance, in general, before the goto nopage. */
3020 page = get_page_from_freelist(gfp_mask, order,
3021 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3025 /* Allocate without watermarks if the context allows */
3026 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3028 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3029 * the allocation is high priority and these type of
3030 * allocations are system rather than user orientated
3032 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3034 page = __alloc_pages_high_priority(gfp_mask, order, ac);
3041 /* Atomic allocations - we can't balance anything */
3044 * All existing users of the deprecated __GFP_NOFAIL are
3045 * blockable, so warn of any new users that actually allow this
3046 * type of allocation to fail.
3048 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3052 /* Avoid recursion of direct reclaim */
3053 if (current->flags & PF_MEMALLOC)
3056 /* Avoid allocations with no watermarks from looping endlessly */
3057 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3061 * Try direct compaction. The first pass is asynchronous. Subsequent
3062 * attempts after direct reclaim are synchronous
3064 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3066 &contended_compaction,
3067 &deferred_compaction);
3071 /* Checks for THP-specific high-order allocations */
3072 if ((gfp_mask & GFP_TRANSHUGE) == GFP_TRANSHUGE) {
3074 * If compaction is deferred for high-order allocations, it is
3075 * because sync compaction recently failed. If this is the case
3076 * and the caller requested a THP allocation, we do not want
3077 * to heavily disrupt the system, so we fail the allocation
3078 * instead of entering direct reclaim.
3080 if (deferred_compaction)
3084 * In all zones where compaction was attempted (and not
3085 * deferred or skipped), lock contention has been detected.
3086 * For THP allocation we do not want to disrupt the others
3087 * so we fallback to base pages instead.
3089 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3093 * If compaction was aborted due to need_resched(), we do not
3094 * want to further increase allocation latency, unless it is
3095 * khugepaged trying to collapse.
3097 if (contended_compaction == COMPACT_CONTENDED_SCHED
3098 && !(current->flags & PF_KTHREAD))
3103 * It can become very expensive to allocate transparent hugepages at
3104 * fault, so use asynchronous memory compaction for THP unless it is
3105 * khugepaged trying to collapse.
3107 if ((gfp_mask & GFP_TRANSHUGE) != GFP_TRANSHUGE ||
3108 (current->flags & PF_KTHREAD))
3109 migration_mode = MIGRATE_SYNC_LIGHT;
3111 /* Try direct reclaim and then allocating */
3112 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3113 &did_some_progress);
3117 /* Do not loop if specifically requested */
3118 if (gfp_mask & __GFP_NORETRY)
3121 /* Keep reclaiming pages as long as there is reasonable progress */
3122 pages_reclaimed += did_some_progress;
3123 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3124 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3125 /* Wait for some write requests to complete then retry */
3126 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3130 /* Reclaim has failed us, start killing things */
3131 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3135 /* Retry as long as the OOM killer is making progress */
3136 if (did_some_progress)
3141 * High-order allocations do not necessarily loop after
3142 * direct reclaim and reclaim/compaction depends on compaction
3143 * being called after reclaim so call directly if necessary
3145 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3147 &contended_compaction,
3148 &deferred_compaction);
3152 warn_alloc_failed(gfp_mask, order, NULL);
3158 * This is the 'heart' of the zoned buddy allocator.
3161 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3162 struct zonelist *zonelist, nodemask_t *nodemask)
3164 struct zoneref *preferred_zoneref;
3165 struct page *page = NULL;
3166 unsigned int cpuset_mems_cookie;
3167 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3168 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3169 struct alloc_context ac = {
3170 .high_zoneidx = gfp_zone(gfp_mask),
3171 .nodemask = nodemask,
3172 .migratetype = gfpflags_to_migratetype(gfp_mask),
3175 gfp_mask &= gfp_allowed_mask;
3177 lockdep_trace_alloc(gfp_mask);
3179 might_sleep_if(gfp_mask & __GFP_WAIT);
3181 if (should_fail_alloc_page(gfp_mask, order))
3185 * Check the zones suitable for the gfp_mask contain at least one
3186 * valid zone. It's possible to have an empty zonelist as a result
3187 * of __GFP_THISNODE and a memoryless node
3189 if (unlikely(!zonelist->_zonerefs->zone))
3192 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3193 alloc_flags |= ALLOC_CMA;
3196 cpuset_mems_cookie = read_mems_allowed_begin();
3198 /* We set it here, as __alloc_pages_slowpath might have changed it */
3199 ac.zonelist = zonelist;
3200 /* The preferred zone is used for statistics later */
3201 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3202 ac.nodemask ? : &cpuset_current_mems_allowed,
3203 &ac.preferred_zone);
3204 if (!ac.preferred_zone)
3206 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3208 /* First allocation attempt */
3209 alloc_mask = gfp_mask|__GFP_HARDWALL;
3210 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3211 if (unlikely(!page)) {
3213 * Runtime PM, block IO and its error handling path
3214 * can deadlock because I/O on the device might not
3217 alloc_mask = memalloc_noio_flags(gfp_mask);
3219 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3222 if (kmemcheck_enabled && page)
3223 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3225 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3229 * When updating a task's mems_allowed, it is possible to race with
3230 * parallel threads in such a way that an allocation can fail while
3231 * the mask is being updated. If a page allocation is about to fail,
3232 * check if the cpuset changed during allocation and if so, retry.
3234 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3239 EXPORT_SYMBOL(__alloc_pages_nodemask);
3242 * Common helper functions.
3244 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3249 * __get_free_pages() returns a 32-bit address, which cannot represent
3252 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3254 page = alloc_pages(gfp_mask, order);
3257 return (unsigned long) page_address(page);
3259 EXPORT_SYMBOL(__get_free_pages);
3261 unsigned long get_zeroed_page(gfp_t gfp_mask)
3263 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3265 EXPORT_SYMBOL(get_zeroed_page);
3267 void __free_pages(struct page *page, unsigned int order)
3269 if (put_page_testzero(page)) {
3271 free_hot_cold_page(page, false);
3273 __free_pages_ok(page, order);
3277 EXPORT_SYMBOL(__free_pages);
3279 void free_pages(unsigned long addr, unsigned int order)
3282 VM_BUG_ON(!virt_addr_valid((void *)addr));
3283 __free_pages(virt_to_page((void *)addr), order);
3287 EXPORT_SYMBOL(free_pages);
3291 * An arbitrary-length arbitrary-offset area of memory which resides
3292 * within a 0 or higher order page. Multiple fragments within that page
3293 * are individually refcounted, in the page's reference counter.
3295 * The page_frag functions below provide a simple allocation framework for
3296 * page fragments. This is used by the network stack and network device
3297 * drivers to provide a backing region of memory for use as either an
3298 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3300 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3303 struct page *page = NULL;
3304 gfp_t gfp = gfp_mask;
3306 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3307 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3309 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3310 PAGE_FRAG_CACHE_MAX_ORDER);
3311 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3313 if (unlikely(!page))
3314 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3316 nc->va = page ? page_address(page) : NULL;
3321 void *__alloc_page_frag(struct page_frag_cache *nc,
3322 unsigned int fragsz, gfp_t gfp_mask)
3324 unsigned int size = PAGE_SIZE;
3328 if (unlikely(!nc->va)) {
3330 page = __page_frag_refill(nc, gfp_mask);
3334 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3335 /* if size can vary use size else just use PAGE_SIZE */
3338 /* Even if we own the page, we do not use atomic_set().
3339 * This would break get_page_unless_zero() users.
3341 atomic_add(size - 1, &page->_count);
3343 /* reset page count bias and offset to start of new frag */
3344 nc->pfmemalloc = page->pfmemalloc;
3345 nc->pagecnt_bias = size;
3349 offset = nc->offset - fragsz;
3350 if (unlikely(offset < 0)) {
3351 page = virt_to_page(nc->va);
3353 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3356 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3357 /* if size can vary use size else just use PAGE_SIZE */
3360 /* OK, page count is 0, we can safely set it */
3361 atomic_set(&page->_count, size);
3363 /* reset page count bias and offset to start of new frag */
3364 nc->pagecnt_bias = size;
3365 offset = size - fragsz;
3369 nc->offset = offset;
3371 return nc->va + offset;
3373 EXPORT_SYMBOL(__alloc_page_frag);
3376 * Frees a page fragment allocated out of either a compound or order 0 page.
3378 void __free_page_frag(void *addr)
3380 struct page *page = virt_to_head_page(addr);
3382 if (unlikely(put_page_testzero(page)))
3383 __free_pages_ok(page, compound_order(page));
3385 EXPORT_SYMBOL(__free_page_frag);
3388 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3389 * of the current memory cgroup.
3391 * It should be used when the caller would like to use kmalloc, but since the
3392 * allocation is large, it has to fall back to the page allocator.
3394 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3397 struct mem_cgroup *memcg = NULL;
3399 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
3401 page = alloc_pages(gfp_mask, order);
3402 memcg_kmem_commit_charge(page, memcg, order);
3406 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3409 struct mem_cgroup *memcg = NULL;
3411 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
3413 page = alloc_pages_node(nid, gfp_mask, order);
3414 memcg_kmem_commit_charge(page, memcg, order);
3419 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3422 void __free_kmem_pages(struct page *page, unsigned int order)
3424 memcg_kmem_uncharge_pages(page, order);
3425 __free_pages(page, order);
3428 void free_kmem_pages(unsigned long addr, unsigned int order)
3431 VM_BUG_ON(!virt_addr_valid((void *)addr));
3432 __free_kmem_pages(virt_to_page((void *)addr), order);
3436 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
3439 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3440 unsigned long used = addr + PAGE_ALIGN(size);
3442 split_page(virt_to_page((void *)addr), order);
3443 while (used < alloc_end) {
3448 return (void *)addr;
3452 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3453 * @size: the number of bytes to allocate
3454 * @gfp_mask: GFP flags for the allocation
3456 * This function is similar to alloc_pages(), except that it allocates the
3457 * minimum number of pages to satisfy the request. alloc_pages() can only
3458 * allocate memory in power-of-two pages.
3460 * This function is also limited by MAX_ORDER.
3462 * Memory allocated by this function must be released by free_pages_exact().
3464 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3466 unsigned int order = get_order(size);
3469 addr = __get_free_pages(gfp_mask, order);
3470 return make_alloc_exact(addr, order, size);
3472 EXPORT_SYMBOL(alloc_pages_exact);
3475 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3477 * @nid: the preferred node ID where memory should be allocated
3478 * @size: the number of bytes to allocate
3479 * @gfp_mask: GFP flags for the allocation
3481 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3483 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3486 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3488 unsigned order = get_order(size);
3489 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3492 return make_alloc_exact((unsigned long)page_address(p), order, size);
3496 * free_pages_exact - release memory allocated via alloc_pages_exact()
3497 * @virt: the value returned by alloc_pages_exact.
3498 * @size: size of allocation, same value as passed to alloc_pages_exact().
3500 * Release the memory allocated by a previous call to alloc_pages_exact.
3502 void free_pages_exact(void *virt, size_t size)
3504 unsigned long addr = (unsigned long)virt;
3505 unsigned long end = addr + PAGE_ALIGN(size);
3507 while (addr < end) {
3512 EXPORT_SYMBOL(free_pages_exact);
3515 * nr_free_zone_pages - count number of pages beyond high watermark
3516 * @offset: The zone index of the highest zone
3518 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3519 * high watermark within all zones at or below a given zone index. For each
3520 * zone, the number of pages is calculated as:
3521 * managed_pages - high_pages
3523 static unsigned long nr_free_zone_pages(int offset)
3528 /* Just pick one node, since fallback list is circular */
3529 unsigned long sum = 0;
3531 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3533 for_each_zone_zonelist(zone, z, zonelist, offset) {
3534 unsigned long size = zone->managed_pages;
3535 unsigned long high = high_wmark_pages(zone);
3544 * nr_free_buffer_pages - count number of pages beyond high watermark
3546 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3547 * watermark within ZONE_DMA and ZONE_NORMAL.
3549 unsigned long nr_free_buffer_pages(void)
3551 return nr_free_zone_pages(gfp_zone(GFP_USER));
3553 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3556 * nr_free_pagecache_pages - count number of pages beyond high watermark
3558 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3559 * high watermark within all zones.
3561 unsigned long nr_free_pagecache_pages(void)
3563 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3566 static inline void show_node(struct zone *zone)
3568 if (IS_ENABLED(CONFIG_NUMA))
3569 printk("Node %d ", zone_to_nid(zone));
3572 void si_meminfo(struct sysinfo *val)
3574 val->totalram = totalram_pages;
3575 val->sharedram = global_page_state(NR_SHMEM);
3576 val->freeram = global_page_state(NR_FREE_PAGES);
3577 val->bufferram = nr_blockdev_pages();
3578 val->totalhigh = totalhigh_pages;
3579 val->freehigh = nr_free_highpages();
3580 val->mem_unit = PAGE_SIZE;
3583 EXPORT_SYMBOL(si_meminfo);
3586 void si_meminfo_node(struct sysinfo *val, int nid)
3588 int zone_type; /* needs to be signed */
3589 unsigned long managed_pages = 0;
3590 pg_data_t *pgdat = NODE_DATA(nid);
3592 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3593 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3594 val->totalram = managed_pages;
3595 val->sharedram = node_page_state(nid, NR_SHMEM);
3596 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3597 #ifdef CONFIG_HIGHMEM
3598 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3599 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3605 val->mem_unit = PAGE_SIZE;
3610 * Determine whether the node should be displayed or not, depending on whether
3611 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3613 bool skip_free_areas_node(unsigned int flags, int nid)
3616 unsigned int cpuset_mems_cookie;
3618 if (!(flags & SHOW_MEM_FILTER_NODES))
3622 cpuset_mems_cookie = read_mems_allowed_begin();
3623 ret = !node_isset(nid, cpuset_current_mems_allowed);
3624 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3629 #define K(x) ((x) << (PAGE_SHIFT-10))
3631 static void show_migration_types(unsigned char type)
3633 static const char types[MIGRATE_TYPES] = {
3634 [MIGRATE_UNMOVABLE] = 'U',
3635 [MIGRATE_RECLAIMABLE] = 'E',
3636 [MIGRATE_MOVABLE] = 'M',
3637 [MIGRATE_RESERVE] = 'R',
3639 [MIGRATE_CMA] = 'C',
3641 #ifdef CONFIG_MEMORY_ISOLATION
3642 [MIGRATE_ISOLATE] = 'I',
3645 char tmp[MIGRATE_TYPES + 1];
3649 for (i = 0; i < MIGRATE_TYPES; i++) {
3650 if (type & (1 << i))
3655 printk("(%s) ", tmp);
3659 * Show free area list (used inside shift_scroll-lock stuff)
3660 * We also calculate the percentage fragmentation. We do this by counting the
3661 * memory on each free list with the exception of the first item on the list.
3664 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3667 void show_free_areas(unsigned int filter)
3669 unsigned long free_pcp = 0;
3673 for_each_populated_zone(zone) {
3674 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3677 for_each_online_cpu(cpu)
3678 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3681 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3682 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3683 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3684 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3685 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3686 " free:%lu free_pcp:%lu free_cma:%lu\n",
3687 global_page_state(NR_ACTIVE_ANON),
3688 global_page_state(NR_INACTIVE_ANON),
3689 global_page_state(NR_ISOLATED_ANON),
3690 global_page_state(NR_ACTIVE_FILE),
3691 global_page_state(NR_INACTIVE_FILE),
3692 global_page_state(NR_ISOLATED_FILE),
3693 global_page_state(NR_UNEVICTABLE),
3694 global_page_state(NR_FILE_DIRTY),
3695 global_page_state(NR_WRITEBACK),
3696 global_page_state(NR_UNSTABLE_NFS),
3697 global_page_state(NR_SLAB_RECLAIMABLE),
3698 global_page_state(NR_SLAB_UNRECLAIMABLE),
3699 global_page_state(NR_FILE_MAPPED),
3700 global_page_state(NR_SHMEM),
3701 global_page_state(NR_PAGETABLE),
3702 global_page_state(NR_BOUNCE),
3703 global_page_state(NR_FREE_PAGES),
3705 global_page_state(NR_FREE_CMA_PAGES));
3707 for_each_populated_zone(zone) {
3710 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3714 for_each_online_cpu(cpu)
3715 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3723 " active_anon:%lukB"
3724 " inactive_anon:%lukB"
3725 " active_file:%lukB"
3726 " inactive_file:%lukB"
3727 " unevictable:%lukB"
3728 " isolated(anon):%lukB"
3729 " isolated(file):%lukB"
3737 " slab_reclaimable:%lukB"
3738 " slab_unreclaimable:%lukB"
3739 " kernel_stack:%lukB"
3746 " writeback_tmp:%lukB"
3747 " pages_scanned:%lu"
3748 " all_unreclaimable? %s"
3751 K(zone_page_state(zone, NR_FREE_PAGES)),
3752 K(min_wmark_pages(zone)),
3753 K(low_wmark_pages(zone)),
3754 K(high_wmark_pages(zone)),
3755 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3756 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3757 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3758 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3759 K(zone_page_state(zone, NR_UNEVICTABLE)),
3760 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3761 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3762 K(zone->present_pages),
3763 K(zone->managed_pages),
3764 K(zone_page_state(zone, NR_MLOCK)),
3765 K(zone_page_state(zone, NR_FILE_DIRTY)),
3766 K(zone_page_state(zone, NR_WRITEBACK)),
3767 K(zone_page_state(zone, NR_FILE_MAPPED)),
3768 K(zone_page_state(zone, NR_SHMEM)),
3769 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3770 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3771 zone_page_state(zone, NR_KERNEL_STACK) *
3773 K(zone_page_state(zone, NR_PAGETABLE)),
3774 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3775 K(zone_page_state(zone, NR_BOUNCE)),
3777 K(this_cpu_read(zone->pageset->pcp.count)),
3778 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3779 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3780 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3781 (!zone_reclaimable(zone) ? "yes" : "no")
3783 printk("lowmem_reserve[]:");
3784 for (i = 0; i < MAX_NR_ZONES; i++)
3785 printk(" %ld", zone->lowmem_reserve[i]);
3789 for_each_populated_zone(zone) {
3790 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3791 unsigned char types[MAX_ORDER];
3793 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3796 printk("%s: ", zone->name);
3798 spin_lock_irqsave(&zone->lock, flags);
3799 for (order = 0; order < MAX_ORDER; order++) {
3800 struct free_area *area = &zone->free_area[order];
3803 nr[order] = area->nr_free;
3804 total += nr[order] << order;
3807 for (type = 0; type < MIGRATE_TYPES; type++) {
3808 if (!list_empty(&area->free_list[type]))
3809 types[order] |= 1 << type;
3812 spin_unlock_irqrestore(&zone->lock, flags);
3813 for (order = 0; order < MAX_ORDER; order++) {
3814 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3816 show_migration_types(types[order]);
3818 printk("= %lukB\n", K(total));
3821 hugetlb_show_meminfo();
3823 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3825 show_swap_cache_info();
3828 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3830 zoneref->zone = zone;
3831 zoneref->zone_idx = zone_idx(zone);
3835 * Builds allocation fallback zone lists.
3837 * Add all populated zones of a node to the zonelist.
3839 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3843 enum zone_type zone_type = MAX_NR_ZONES;
3847 zone = pgdat->node_zones + zone_type;
3848 if (populated_zone(zone)) {
3849 zoneref_set_zone(zone,
3850 &zonelist->_zonerefs[nr_zones++]);
3851 check_highest_zone(zone_type);
3853 } while (zone_type);
3861 * 0 = automatic detection of better ordering.
3862 * 1 = order by ([node] distance, -zonetype)
3863 * 2 = order by (-zonetype, [node] distance)
3865 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3866 * the same zonelist. So only NUMA can configure this param.
3868 #define ZONELIST_ORDER_DEFAULT 0
3869 #define ZONELIST_ORDER_NODE 1
3870 #define ZONELIST_ORDER_ZONE 2
3872 /* zonelist order in the kernel.
3873 * set_zonelist_order() will set this to NODE or ZONE.
3875 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3876 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3880 /* The value user specified ....changed by config */
3881 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3882 /* string for sysctl */
3883 #define NUMA_ZONELIST_ORDER_LEN 16
3884 char numa_zonelist_order[16] = "default";
3887 * interface for configure zonelist ordering.
3888 * command line option "numa_zonelist_order"
3889 * = "[dD]efault - default, automatic configuration.
3890 * = "[nN]ode - order by node locality, then by zone within node
3891 * = "[zZ]one - order by zone, then by locality within zone
3894 static int __parse_numa_zonelist_order(char *s)
3896 if (*s == 'd' || *s == 'D') {
3897 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3898 } else if (*s == 'n' || *s == 'N') {
3899 user_zonelist_order = ZONELIST_ORDER_NODE;
3900 } else if (*s == 'z' || *s == 'Z') {
3901 user_zonelist_order = ZONELIST_ORDER_ZONE;
3904 "Ignoring invalid numa_zonelist_order value: "
3911 static __init int setup_numa_zonelist_order(char *s)
3918 ret = __parse_numa_zonelist_order(s);
3920 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3924 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3927 * sysctl handler for numa_zonelist_order
3929 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3930 void __user *buffer, size_t *length,
3933 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3935 static DEFINE_MUTEX(zl_order_mutex);
3937 mutex_lock(&zl_order_mutex);
3939 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3943 strcpy(saved_string, (char *)table->data);
3945 ret = proc_dostring(table, write, buffer, length, ppos);
3949 int oldval = user_zonelist_order;
3951 ret = __parse_numa_zonelist_order((char *)table->data);
3954 * bogus value. restore saved string
3956 strncpy((char *)table->data, saved_string,
3957 NUMA_ZONELIST_ORDER_LEN);
3958 user_zonelist_order = oldval;
3959 } else if (oldval != user_zonelist_order) {
3960 mutex_lock(&zonelists_mutex);
3961 build_all_zonelists(NULL, NULL);
3962 mutex_unlock(&zonelists_mutex);
3966 mutex_unlock(&zl_order_mutex);
3971 #define MAX_NODE_LOAD (nr_online_nodes)
3972 static int node_load[MAX_NUMNODES];
3975 * find_next_best_node - find the next node that should appear in a given node's fallback list
3976 * @node: node whose fallback list we're appending
3977 * @used_node_mask: nodemask_t of already used nodes
3979 * We use a number of factors to determine which is the next node that should
3980 * appear on a given node's fallback list. The node should not have appeared
3981 * already in @node's fallback list, and it should be the next closest node
3982 * according to the distance array (which contains arbitrary distance values
3983 * from each node to each node in the system), and should also prefer nodes
3984 * with no CPUs, since presumably they'll have very little allocation pressure
3985 * on them otherwise.
3986 * It returns -1 if no node is found.
3988 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3991 int min_val = INT_MAX;
3992 int best_node = NUMA_NO_NODE;
3993 const struct cpumask *tmp = cpumask_of_node(0);
3995 /* Use the local node if we haven't already */
3996 if (!node_isset(node, *used_node_mask)) {
3997 node_set(node, *used_node_mask);
4001 for_each_node_state(n, N_MEMORY) {
4003 /* Don't want a node to appear more than once */
4004 if (node_isset(n, *used_node_mask))
4007 /* Use the distance array to find the distance */
4008 val = node_distance(node, n);
4010 /* Penalize nodes under us ("prefer the next node") */
4013 /* Give preference to headless and unused nodes */
4014 tmp = cpumask_of_node(n);
4015 if (!cpumask_empty(tmp))
4016 val += PENALTY_FOR_NODE_WITH_CPUS;
4018 /* Slight preference for less loaded node */
4019 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4020 val += node_load[n];
4022 if (val < min_val) {
4029 node_set(best_node, *used_node_mask);
4036 * Build zonelists ordered by node and zones within node.
4037 * This results in maximum locality--normal zone overflows into local
4038 * DMA zone, if any--but risks exhausting DMA zone.
4040 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4043 struct zonelist *zonelist;
4045 zonelist = &pgdat->node_zonelists[0];
4046 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4048 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4049 zonelist->_zonerefs[j].zone = NULL;
4050 zonelist->_zonerefs[j].zone_idx = 0;
4054 * Build gfp_thisnode zonelists
4056 static void build_thisnode_zonelists(pg_data_t *pgdat)
4059 struct zonelist *zonelist;
4061 zonelist = &pgdat->node_zonelists[1];
4062 j = build_zonelists_node(pgdat, zonelist, 0);
4063 zonelist->_zonerefs[j].zone = NULL;
4064 zonelist->_zonerefs[j].zone_idx = 0;
4068 * Build zonelists ordered by zone and nodes within zones.
4069 * This results in conserving DMA zone[s] until all Normal memory is
4070 * exhausted, but results in overflowing to remote node while memory
4071 * may still exist in local DMA zone.
4073 static int node_order[MAX_NUMNODES];
4075 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4078 int zone_type; /* needs to be signed */
4080 struct zonelist *zonelist;
4082 zonelist = &pgdat->node_zonelists[0];
4084 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4085 for (j = 0; j < nr_nodes; j++) {
4086 node = node_order[j];
4087 z = &NODE_DATA(node)->node_zones[zone_type];
4088 if (populated_zone(z)) {
4090 &zonelist->_zonerefs[pos++]);
4091 check_highest_zone(zone_type);
4095 zonelist->_zonerefs[pos].zone = NULL;
4096 zonelist->_zonerefs[pos].zone_idx = 0;
4099 #if defined(CONFIG_64BIT)
4101 * Devices that require DMA32/DMA are relatively rare and do not justify a
4102 * penalty to every machine in case the specialised case applies. Default
4103 * to Node-ordering on 64-bit NUMA machines
4105 static int default_zonelist_order(void)
4107 return ZONELIST_ORDER_NODE;
4111 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4112 * by the kernel. If processes running on node 0 deplete the low memory zone
4113 * then reclaim will occur more frequency increasing stalls and potentially
4114 * be easier to OOM if a large percentage of the zone is under writeback or
4115 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4116 * Hence, default to zone ordering on 32-bit.
4118 static int default_zonelist_order(void)
4120 return ZONELIST_ORDER_ZONE;
4122 #endif /* CONFIG_64BIT */
4124 static void set_zonelist_order(void)
4126 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4127 current_zonelist_order = default_zonelist_order();
4129 current_zonelist_order = user_zonelist_order;
4132 static void build_zonelists(pg_data_t *pgdat)
4136 nodemask_t used_mask;
4137 int local_node, prev_node;
4138 struct zonelist *zonelist;
4139 int order = current_zonelist_order;
4141 /* initialize zonelists */
4142 for (i = 0; i < MAX_ZONELISTS; i++) {
4143 zonelist = pgdat->node_zonelists + i;
4144 zonelist->_zonerefs[0].zone = NULL;
4145 zonelist->_zonerefs[0].zone_idx = 0;
4148 /* NUMA-aware ordering of nodes */
4149 local_node = pgdat->node_id;
4150 load = nr_online_nodes;
4151 prev_node = local_node;
4152 nodes_clear(used_mask);
4154 memset(node_order, 0, sizeof(node_order));
4157 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4159 * We don't want to pressure a particular node.
4160 * So adding penalty to the first node in same
4161 * distance group to make it round-robin.
4163 if (node_distance(local_node, node) !=
4164 node_distance(local_node, prev_node))
4165 node_load[node] = load;
4169 if (order == ZONELIST_ORDER_NODE)
4170 build_zonelists_in_node_order(pgdat, node);
4172 node_order[j++] = node; /* remember order */
4175 if (order == ZONELIST_ORDER_ZONE) {
4176 /* calculate node order -- i.e., DMA last! */
4177 build_zonelists_in_zone_order(pgdat, j);
4180 build_thisnode_zonelists(pgdat);
4183 /* Construct the zonelist performance cache - see further mmzone.h */
4184 static void build_zonelist_cache(pg_data_t *pgdat)
4186 struct zonelist *zonelist;
4187 struct zonelist_cache *zlc;
4190 zonelist = &pgdat->node_zonelists[0];
4191 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
4192 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
4193 for (z = zonelist->_zonerefs; z->zone; z++)
4194 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
4197 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4199 * Return node id of node used for "local" allocations.
4200 * I.e., first node id of first zone in arg node's generic zonelist.
4201 * Used for initializing percpu 'numa_mem', which is used primarily
4202 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4204 int local_memory_node(int node)
4208 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4209 gfp_zone(GFP_KERNEL),
4216 #else /* CONFIG_NUMA */
4218 static void set_zonelist_order(void)
4220 current_zonelist_order = ZONELIST_ORDER_ZONE;
4223 static void build_zonelists(pg_data_t *pgdat)
4225 int node, local_node;
4227 struct zonelist *zonelist;
4229 local_node = pgdat->node_id;
4231 zonelist = &pgdat->node_zonelists[0];
4232 j = build_zonelists_node(pgdat, zonelist, 0);
4235 * Now we build the zonelist so that it contains the zones
4236 * of all the other nodes.
4237 * We don't want to pressure a particular node, so when
4238 * building the zones for node N, we make sure that the
4239 * zones coming right after the local ones are those from
4240 * node N+1 (modulo N)
4242 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4243 if (!node_online(node))
4245 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4247 for (node = 0; node < local_node; node++) {
4248 if (!node_online(node))
4250 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4253 zonelist->_zonerefs[j].zone = NULL;
4254 zonelist->_zonerefs[j].zone_idx = 0;
4257 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
4258 static void build_zonelist_cache(pg_data_t *pgdat)
4260 pgdat->node_zonelists[0].zlcache_ptr = NULL;
4263 #endif /* CONFIG_NUMA */
4266 * Boot pageset table. One per cpu which is going to be used for all
4267 * zones and all nodes. The parameters will be set in such a way
4268 * that an item put on a list will immediately be handed over to
4269 * the buddy list. This is safe since pageset manipulation is done
4270 * with interrupts disabled.
4272 * The boot_pagesets must be kept even after bootup is complete for
4273 * unused processors and/or zones. They do play a role for bootstrapping
4274 * hotplugged processors.
4276 * zoneinfo_show() and maybe other functions do
4277 * not check if the processor is online before following the pageset pointer.
4278 * Other parts of the kernel may not check if the zone is available.
4280 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4281 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4282 static void setup_zone_pageset(struct zone *zone);
4285 * Global mutex to protect against size modification of zonelists
4286 * as well as to serialize pageset setup for the new populated zone.
4288 DEFINE_MUTEX(zonelists_mutex);
4290 /* return values int ....just for stop_machine() */
4291 static int __build_all_zonelists(void *data)
4295 pg_data_t *self = data;
4298 memset(node_load, 0, sizeof(node_load));
4301 if (self && !node_online(self->node_id)) {
4302 build_zonelists(self);
4303 build_zonelist_cache(self);
4306 for_each_online_node(nid) {
4307 pg_data_t *pgdat = NODE_DATA(nid);
4309 build_zonelists(pgdat);
4310 build_zonelist_cache(pgdat);
4314 * Initialize the boot_pagesets that are going to be used
4315 * for bootstrapping processors. The real pagesets for
4316 * each zone will be allocated later when the per cpu
4317 * allocator is available.
4319 * boot_pagesets are used also for bootstrapping offline
4320 * cpus if the system is already booted because the pagesets
4321 * are needed to initialize allocators on a specific cpu too.
4322 * F.e. the percpu allocator needs the page allocator which
4323 * needs the percpu allocator in order to allocate its pagesets
4324 * (a chicken-egg dilemma).
4326 for_each_possible_cpu(cpu) {
4327 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4329 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4331 * We now know the "local memory node" for each node--
4332 * i.e., the node of the first zone in the generic zonelist.
4333 * Set up numa_mem percpu variable for on-line cpus. During
4334 * boot, only the boot cpu should be on-line; we'll init the
4335 * secondary cpus' numa_mem as they come on-line. During
4336 * node/memory hotplug, we'll fixup all on-line cpus.
4338 if (cpu_online(cpu))
4339 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4346 static noinline void __init
4347 build_all_zonelists_init(void)
4349 __build_all_zonelists(NULL);
4350 mminit_verify_zonelist();
4351 cpuset_init_current_mems_allowed();
4355 * Called with zonelists_mutex held always
4356 * unless system_state == SYSTEM_BOOTING.
4358 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4359 * [we're only called with non-NULL zone through __meminit paths] and
4360 * (2) call of __init annotated helper build_all_zonelists_init
4361 * [protected by SYSTEM_BOOTING].
4363 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4365 set_zonelist_order();
4367 if (system_state == SYSTEM_BOOTING) {
4368 build_all_zonelists_init();
4370 #ifdef CONFIG_MEMORY_HOTPLUG
4372 setup_zone_pageset(zone);
4374 /* we have to stop all cpus to guarantee there is no user
4376 stop_machine(__build_all_zonelists, pgdat, NULL);
4377 /* cpuset refresh routine should be here */
4379 vm_total_pages = nr_free_pagecache_pages();
4381 * Disable grouping by mobility if the number of pages in the
4382 * system is too low to allow the mechanism to work. It would be
4383 * more accurate, but expensive to check per-zone. This check is
4384 * made on memory-hotadd so a system can start with mobility
4385 * disabled and enable it later
4387 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4388 page_group_by_mobility_disabled = 1;
4390 page_group_by_mobility_disabled = 0;
4392 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4393 "Total pages: %ld\n",
4395 zonelist_order_name[current_zonelist_order],
4396 page_group_by_mobility_disabled ? "off" : "on",
4399 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4404 * Helper functions to size the waitqueue hash table.
4405 * Essentially these want to choose hash table sizes sufficiently
4406 * large so that collisions trying to wait on pages are rare.
4407 * But in fact, the number of active page waitqueues on typical
4408 * systems is ridiculously low, less than 200. So this is even
4409 * conservative, even though it seems large.
4411 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4412 * waitqueues, i.e. the size of the waitq table given the number of pages.
4414 #define PAGES_PER_WAITQUEUE 256
4416 #ifndef CONFIG_MEMORY_HOTPLUG
4417 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4419 unsigned long size = 1;
4421 pages /= PAGES_PER_WAITQUEUE;
4423 while (size < pages)
4427 * Once we have dozens or even hundreds of threads sleeping
4428 * on IO we've got bigger problems than wait queue collision.
4429 * Limit the size of the wait table to a reasonable size.
4431 size = min(size, 4096UL);
4433 return max(size, 4UL);
4437 * A zone's size might be changed by hot-add, so it is not possible to determine
4438 * a suitable size for its wait_table. So we use the maximum size now.
4440 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4442 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4443 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4444 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4446 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4447 * or more by the traditional way. (See above). It equals:
4449 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4450 * ia64(16K page size) : = ( 8G + 4M)byte.
4451 * powerpc (64K page size) : = (32G +16M)byte.
4453 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4460 * This is an integer logarithm so that shifts can be used later
4461 * to extract the more random high bits from the multiplicative
4462 * hash function before the remainder is taken.
4464 static inline unsigned long wait_table_bits(unsigned long size)
4470 * Check if a pageblock contains reserved pages
4472 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
4476 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4477 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
4484 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4485 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4486 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4487 * higher will lead to a bigger reserve which will get freed as contiguous
4488 * blocks as reclaim kicks in
4490 static void setup_zone_migrate_reserve(struct zone *zone)
4492 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
4494 unsigned long block_migratetype;
4499 * Get the start pfn, end pfn and the number of blocks to reserve
4500 * We have to be careful to be aligned to pageblock_nr_pages to
4501 * make sure that we always check pfn_valid for the first page in
4504 start_pfn = zone->zone_start_pfn;
4505 end_pfn = zone_end_pfn(zone);
4506 start_pfn = roundup(start_pfn, pageblock_nr_pages);
4507 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
4511 * Reserve blocks are generally in place to help high-order atomic
4512 * allocations that are short-lived. A min_free_kbytes value that
4513 * would result in more than 2 reserve blocks for atomic allocations
4514 * is assumed to be in place to help anti-fragmentation for the
4515 * future allocation of hugepages at runtime.
4517 reserve = min(2, reserve);
4518 old_reserve = zone->nr_migrate_reserve_block;
4520 /* When memory hot-add, we almost always need to do nothing */
4521 if (reserve == old_reserve)
4523 zone->nr_migrate_reserve_block = reserve;
4525 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
4526 if (!early_page_nid_uninitialised(pfn, zone_to_nid(zone)))
4529 if (!pfn_valid(pfn))
4531 page = pfn_to_page(pfn);
4533 /* Watch out for overlapping nodes */
4534 if (page_to_nid(page) != zone_to_nid(zone))
4537 block_migratetype = get_pageblock_migratetype(page);
4539 /* Only test what is necessary when the reserves are not met */
4542 * Blocks with reserved pages will never free, skip
4545 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
4546 if (pageblock_is_reserved(pfn, block_end_pfn))
4549 /* If this block is reserved, account for it */
4550 if (block_migratetype == MIGRATE_RESERVE) {
4555 /* Suitable for reserving if this block is movable */
4556 if (block_migratetype == MIGRATE_MOVABLE) {
4557 set_pageblock_migratetype(page,
4559 move_freepages_block(zone, page,
4564 } else if (!old_reserve) {
4566 * At boot time we don't need to scan the whole zone
4567 * for turning off MIGRATE_RESERVE.
4573 * If the reserve is met and this is a previous reserved block,
4576 if (block_migratetype == MIGRATE_RESERVE) {
4577 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4578 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4584 * Initially all pages are reserved - free ones are freed
4585 * up by free_all_bootmem() once the early boot process is
4586 * done. Non-atomic initialization, single-pass.
4588 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4589 unsigned long start_pfn, enum memmap_context context)
4591 pg_data_t *pgdat = NODE_DATA(nid);
4592 unsigned long end_pfn = start_pfn + size;
4595 unsigned long nr_initialised = 0;
4597 if (highest_memmap_pfn < end_pfn - 1)
4598 highest_memmap_pfn = end_pfn - 1;
4600 z = &pgdat->node_zones[zone];
4601 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4603 * There can be holes in boot-time mem_map[]s
4604 * handed to this function. They do not
4605 * exist on hotplugged memory.
4607 if (context == MEMMAP_EARLY) {
4608 if (!early_pfn_valid(pfn))
4610 if (!early_pfn_in_nid(pfn, nid))
4612 if (!update_defer_init(pgdat, pfn, end_pfn,
4618 * Mark the block movable so that blocks are reserved for
4619 * movable at startup. This will force kernel allocations
4620 * to reserve their blocks rather than leaking throughout
4621 * the address space during boot when many long-lived
4622 * kernel allocations are made. Later some blocks near
4623 * the start are marked MIGRATE_RESERVE by
4624 * setup_zone_migrate_reserve()
4626 * bitmap is created for zone's valid pfn range. but memmap
4627 * can be created for invalid pages (for alignment)
4628 * check here not to call set_pageblock_migratetype() against
4631 if (!(pfn & (pageblock_nr_pages - 1))) {
4632 struct page *page = pfn_to_page(pfn);
4634 __init_single_page(page, pfn, zone, nid);
4635 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4637 __init_single_pfn(pfn, zone, nid);
4642 static void __meminit zone_init_free_lists(struct zone *zone)
4644 unsigned int order, t;
4645 for_each_migratetype_order(order, t) {
4646 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4647 zone->free_area[order].nr_free = 0;
4651 #ifndef __HAVE_ARCH_MEMMAP_INIT
4652 #define memmap_init(size, nid, zone, start_pfn) \
4653 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4656 static int zone_batchsize(struct zone *zone)
4662 * The per-cpu-pages pools are set to around 1000th of the
4663 * size of the zone. But no more than 1/2 of a meg.
4665 * OK, so we don't know how big the cache is. So guess.
4667 batch = zone->managed_pages / 1024;
4668 if (batch * PAGE_SIZE > 512 * 1024)
4669 batch = (512 * 1024) / PAGE_SIZE;
4670 batch /= 4; /* We effectively *= 4 below */
4675 * Clamp the batch to a 2^n - 1 value. Having a power
4676 * of 2 value was found to be more likely to have
4677 * suboptimal cache aliasing properties in some cases.
4679 * For example if 2 tasks are alternately allocating
4680 * batches of pages, one task can end up with a lot
4681 * of pages of one half of the possible page colors
4682 * and the other with pages of the other colors.
4684 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4689 /* The deferral and batching of frees should be suppressed under NOMMU
4692 * The problem is that NOMMU needs to be able to allocate large chunks
4693 * of contiguous memory as there's no hardware page translation to
4694 * assemble apparent contiguous memory from discontiguous pages.
4696 * Queueing large contiguous runs of pages for batching, however,
4697 * causes the pages to actually be freed in smaller chunks. As there
4698 * can be a significant delay between the individual batches being
4699 * recycled, this leads to the once large chunks of space being
4700 * fragmented and becoming unavailable for high-order allocations.
4707 * pcp->high and pcp->batch values are related and dependent on one another:
4708 * ->batch must never be higher then ->high.
4709 * The following function updates them in a safe manner without read side
4712 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4713 * those fields changing asynchronously (acording the the above rule).
4715 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4716 * outside of boot time (or some other assurance that no concurrent updaters
4719 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4720 unsigned long batch)
4722 /* start with a fail safe value for batch */
4726 /* Update high, then batch, in order */
4733 /* a companion to pageset_set_high() */
4734 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4736 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4739 static void pageset_init(struct per_cpu_pageset *p)
4741 struct per_cpu_pages *pcp;
4744 memset(p, 0, sizeof(*p));
4748 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4749 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4752 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4755 pageset_set_batch(p, batch);
4759 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4760 * to the value high for the pageset p.
4762 static void pageset_set_high(struct per_cpu_pageset *p,
4765 unsigned long batch = max(1UL, high / 4);
4766 if ((high / 4) > (PAGE_SHIFT * 8))
4767 batch = PAGE_SHIFT * 8;
4769 pageset_update(&p->pcp, high, batch);
4772 static void pageset_set_high_and_batch(struct zone *zone,
4773 struct per_cpu_pageset *pcp)
4775 if (percpu_pagelist_fraction)
4776 pageset_set_high(pcp,
4777 (zone->managed_pages /
4778 percpu_pagelist_fraction));
4780 pageset_set_batch(pcp, zone_batchsize(zone));
4783 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4785 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4788 pageset_set_high_and_batch(zone, pcp);
4791 static void __meminit setup_zone_pageset(struct zone *zone)
4794 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4795 for_each_possible_cpu(cpu)
4796 zone_pageset_init(zone, cpu);
4800 * Allocate per cpu pagesets and initialize them.
4801 * Before this call only boot pagesets were available.
4803 void __init setup_per_cpu_pageset(void)
4807 for_each_populated_zone(zone)
4808 setup_zone_pageset(zone);
4811 static noinline __init_refok
4812 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4818 * The per-page waitqueue mechanism uses hashed waitqueues
4821 zone->wait_table_hash_nr_entries =
4822 wait_table_hash_nr_entries(zone_size_pages);
4823 zone->wait_table_bits =
4824 wait_table_bits(zone->wait_table_hash_nr_entries);
4825 alloc_size = zone->wait_table_hash_nr_entries
4826 * sizeof(wait_queue_head_t);
4828 if (!slab_is_available()) {
4829 zone->wait_table = (wait_queue_head_t *)
4830 memblock_virt_alloc_node_nopanic(
4831 alloc_size, zone->zone_pgdat->node_id);
4834 * This case means that a zone whose size was 0 gets new memory
4835 * via memory hot-add.
4836 * But it may be the case that a new node was hot-added. In
4837 * this case vmalloc() will not be able to use this new node's
4838 * memory - this wait_table must be initialized to use this new
4839 * node itself as well.
4840 * To use this new node's memory, further consideration will be
4843 zone->wait_table = vmalloc(alloc_size);
4845 if (!zone->wait_table)
4848 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4849 init_waitqueue_head(zone->wait_table + i);
4854 static __meminit void zone_pcp_init(struct zone *zone)
4857 * per cpu subsystem is not up at this point. The following code
4858 * relies on the ability of the linker to provide the
4859 * offset of a (static) per cpu variable into the per cpu area.
4861 zone->pageset = &boot_pageset;
4863 if (populated_zone(zone))
4864 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4865 zone->name, zone->present_pages,
4866 zone_batchsize(zone));
4869 int __meminit init_currently_empty_zone(struct zone *zone,
4870 unsigned long zone_start_pfn,
4872 enum memmap_context context)
4874 struct pglist_data *pgdat = zone->zone_pgdat;
4876 ret = zone_wait_table_init(zone, size);
4879 pgdat->nr_zones = zone_idx(zone) + 1;
4881 zone->zone_start_pfn = zone_start_pfn;
4883 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4884 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4886 (unsigned long)zone_idx(zone),
4887 zone_start_pfn, (zone_start_pfn + size));
4889 zone_init_free_lists(zone);
4894 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4895 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4898 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4900 int __meminit __early_pfn_to_nid(unsigned long pfn,
4901 struct mminit_pfnnid_cache *state)
4903 unsigned long start_pfn, end_pfn;
4906 if (state->last_start <= pfn && pfn < state->last_end)
4907 return state->last_nid;
4909 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4911 state->last_start = start_pfn;
4912 state->last_end = end_pfn;
4913 state->last_nid = nid;
4918 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4921 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4922 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4923 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4925 * If an architecture guarantees that all ranges registered contain no holes
4926 * and may be freed, this this function may be used instead of calling
4927 * memblock_free_early_nid() manually.
4929 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4931 unsigned long start_pfn, end_pfn;
4934 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4935 start_pfn = min(start_pfn, max_low_pfn);
4936 end_pfn = min(end_pfn, max_low_pfn);
4938 if (start_pfn < end_pfn)
4939 memblock_free_early_nid(PFN_PHYS(start_pfn),
4940 (end_pfn - start_pfn) << PAGE_SHIFT,
4946 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4947 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4949 * If an architecture guarantees that all ranges registered contain no holes and may
4950 * be freed, this function may be used instead of calling memory_present() manually.
4952 void __init sparse_memory_present_with_active_regions(int nid)
4954 unsigned long start_pfn, end_pfn;
4957 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4958 memory_present(this_nid, start_pfn, end_pfn);
4962 * get_pfn_range_for_nid - Return the start and end page frames for a node
4963 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4964 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4965 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4967 * It returns the start and end page frame of a node based on information
4968 * provided by memblock_set_node(). If called for a node
4969 * with no available memory, a warning is printed and the start and end
4972 void __meminit get_pfn_range_for_nid(unsigned int nid,
4973 unsigned long *start_pfn, unsigned long *end_pfn)
4975 unsigned long this_start_pfn, this_end_pfn;
4981 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4982 *start_pfn = min(*start_pfn, this_start_pfn);
4983 *end_pfn = max(*end_pfn, this_end_pfn);
4986 if (*start_pfn == -1UL)
4991 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4992 * assumption is made that zones within a node are ordered in monotonic
4993 * increasing memory addresses so that the "highest" populated zone is used
4995 static void __init find_usable_zone_for_movable(void)
4998 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4999 if (zone_index == ZONE_MOVABLE)
5002 if (arch_zone_highest_possible_pfn[zone_index] >
5003 arch_zone_lowest_possible_pfn[zone_index])
5007 VM_BUG_ON(zone_index == -1);
5008 movable_zone = zone_index;
5012 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5013 * because it is sized independent of architecture. Unlike the other zones,
5014 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5015 * in each node depending on the size of each node and how evenly kernelcore
5016 * is distributed. This helper function adjusts the zone ranges
5017 * provided by the architecture for a given node by using the end of the
5018 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5019 * zones within a node are in order of monotonic increases memory addresses
5021 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5022 unsigned long zone_type,
5023 unsigned long node_start_pfn,
5024 unsigned long node_end_pfn,
5025 unsigned long *zone_start_pfn,
5026 unsigned long *zone_end_pfn)
5028 /* Only adjust if ZONE_MOVABLE is on this node */
5029 if (zone_movable_pfn[nid]) {
5030 /* Size ZONE_MOVABLE */
5031 if (zone_type == ZONE_MOVABLE) {
5032 *zone_start_pfn = zone_movable_pfn[nid];
5033 *zone_end_pfn = min(node_end_pfn,
5034 arch_zone_highest_possible_pfn[movable_zone]);
5036 /* Adjust for ZONE_MOVABLE starting within this range */
5037 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
5038 *zone_end_pfn > zone_movable_pfn[nid]) {
5039 *zone_end_pfn = zone_movable_pfn[nid];
5041 /* Check if this whole range is within ZONE_MOVABLE */
5042 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5043 *zone_start_pfn = *zone_end_pfn;
5048 * Return the number of pages a zone spans in a node, including holes
5049 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5051 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5052 unsigned long zone_type,
5053 unsigned long node_start_pfn,
5054 unsigned long node_end_pfn,
5055 unsigned long *ignored)
5057 unsigned long zone_start_pfn, zone_end_pfn;
5059 /* Get the start and end of the zone */
5060 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5061 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5062 adjust_zone_range_for_zone_movable(nid, zone_type,
5063 node_start_pfn, node_end_pfn,
5064 &zone_start_pfn, &zone_end_pfn);
5066 /* Check that this node has pages within the zone's required range */
5067 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
5070 /* Move the zone boundaries inside the node if necessary */
5071 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
5072 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
5074 /* Return the spanned pages */
5075 return zone_end_pfn - zone_start_pfn;
5079 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5080 * then all holes in the requested range will be accounted for.
5082 unsigned long __meminit __absent_pages_in_range(int nid,
5083 unsigned long range_start_pfn,
5084 unsigned long range_end_pfn)
5086 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5087 unsigned long start_pfn, end_pfn;
5090 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5091 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5092 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5093 nr_absent -= end_pfn - start_pfn;
5099 * absent_pages_in_range - Return number of page frames in holes within a range
5100 * @start_pfn: The start PFN to start searching for holes
5101 * @end_pfn: The end PFN to stop searching for holes
5103 * It returns the number of pages frames in memory holes within a range.
5105 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5106 unsigned long end_pfn)
5108 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5111 /* Return the number of page frames in holes in a zone on a node */
5112 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5113 unsigned long zone_type,
5114 unsigned long node_start_pfn,
5115 unsigned long node_end_pfn,
5116 unsigned long *ignored)
5118 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5119 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5120 unsigned long zone_start_pfn, zone_end_pfn;
5122 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5123 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5125 adjust_zone_range_for_zone_movable(nid, zone_type,
5126 node_start_pfn, node_end_pfn,
5127 &zone_start_pfn, &zone_end_pfn);
5128 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5131 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5132 static inline unsigned long __meminit zone_spanned_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 *zones_size)
5138 return zones_size[zone_type];
5141 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5142 unsigned long zone_type,
5143 unsigned long node_start_pfn,
5144 unsigned long node_end_pfn,
5145 unsigned long *zholes_size)
5150 return zholes_size[zone_type];
5153 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5155 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5156 unsigned long node_start_pfn,
5157 unsigned long node_end_pfn,
5158 unsigned long *zones_size,
5159 unsigned long *zholes_size)
5161 unsigned long realtotalpages = 0, totalpages = 0;
5164 for (i = 0; i < MAX_NR_ZONES; i++) {
5165 struct zone *zone = pgdat->node_zones + i;
5166 unsigned long size, real_size;
5168 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5172 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5173 node_start_pfn, node_end_pfn,
5175 zone->spanned_pages = size;
5176 zone->present_pages = real_size;
5179 realtotalpages += real_size;
5182 pgdat->node_spanned_pages = totalpages;
5183 pgdat->node_present_pages = realtotalpages;
5184 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5188 #ifndef CONFIG_SPARSEMEM
5190 * Calculate the size of the zone->blockflags rounded to an unsigned long
5191 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5192 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5193 * round what is now in bits to nearest long in bits, then return it in
5196 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5198 unsigned long usemapsize;
5200 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5201 usemapsize = roundup(zonesize, pageblock_nr_pages);
5202 usemapsize = usemapsize >> pageblock_order;
5203 usemapsize *= NR_PAGEBLOCK_BITS;
5204 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5206 return usemapsize / 8;
5209 static void __init setup_usemap(struct pglist_data *pgdat,
5211 unsigned long zone_start_pfn,
5212 unsigned long zonesize)
5214 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5215 zone->pageblock_flags = NULL;
5217 zone->pageblock_flags =
5218 memblock_virt_alloc_node_nopanic(usemapsize,
5222 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5223 unsigned long zone_start_pfn, unsigned long zonesize) {}
5224 #endif /* CONFIG_SPARSEMEM */
5226 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5228 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5229 void __paginginit set_pageblock_order(void)
5233 /* Check that pageblock_nr_pages has not already been setup */
5234 if (pageblock_order)
5237 if (HPAGE_SHIFT > PAGE_SHIFT)
5238 order = HUGETLB_PAGE_ORDER;
5240 order = MAX_ORDER - 1;
5243 * Assume the largest contiguous order of interest is a huge page.
5244 * This value may be variable depending on boot parameters on IA64 and
5247 pageblock_order = order;
5249 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5252 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5253 * is unused as pageblock_order is set at compile-time. See
5254 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5257 void __paginginit set_pageblock_order(void)
5261 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5263 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5264 unsigned long present_pages)
5266 unsigned long pages = spanned_pages;
5269 * Provide a more accurate estimation if there are holes within
5270 * the zone and SPARSEMEM is in use. If there are holes within the
5271 * zone, each populated memory region may cost us one or two extra
5272 * memmap pages due to alignment because memmap pages for each
5273 * populated regions may not naturally algined on page boundary.
5274 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5276 if (spanned_pages > present_pages + (present_pages >> 4) &&
5277 IS_ENABLED(CONFIG_SPARSEMEM))
5278 pages = present_pages;
5280 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5284 * Set up the zone data structures:
5285 * - mark all pages reserved
5286 * - mark all memory queues empty
5287 * - clear the memory bitmaps
5289 * NOTE: pgdat should get zeroed by caller.
5291 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
5292 unsigned long node_start_pfn, unsigned long node_end_pfn)
5295 int nid = pgdat->node_id;
5296 unsigned long zone_start_pfn = pgdat->node_start_pfn;
5299 pgdat_resize_init(pgdat);
5300 #ifdef CONFIG_NUMA_BALANCING
5301 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5302 pgdat->numabalancing_migrate_nr_pages = 0;
5303 pgdat->numabalancing_migrate_next_window = jiffies;
5305 init_waitqueue_head(&pgdat->kswapd_wait);
5306 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5307 pgdat_page_ext_init(pgdat);
5309 for (j = 0; j < MAX_NR_ZONES; j++) {
5310 struct zone *zone = pgdat->node_zones + j;
5311 unsigned long size, realsize, freesize, memmap_pages;
5313 size = zone->spanned_pages;
5314 realsize = freesize = zone->present_pages;
5317 * Adjust freesize so that it accounts for how much memory
5318 * is used by this zone for memmap. This affects the watermark
5319 * and per-cpu initialisations
5321 memmap_pages = calc_memmap_size(size, realsize);
5322 if (!is_highmem_idx(j)) {
5323 if (freesize >= memmap_pages) {
5324 freesize -= memmap_pages;
5327 " %s zone: %lu pages used for memmap\n",
5328 zone_names[j], memmap_pages);
5331 " %s zone: %lu pages exceeds freesize %lu\n",
5332 zone_names[j], memmap_pages, freesize);
5335 /* Account for reserved pages */
5336 if (j == 0 && freesize > dma_reserve) {
5337 freesize -= dma_reserve;
5338 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5339 zone_names[0], dma_reserve);
5342 if (!is_highmem_idx(j))
5343 nr_kernel_pages += freesize;
5344 /* Charge for highmem memmap if there are enough kernel pages */
5345 else if (nr_kernel_pages > memmap_pages * 2)
5346 nr_kernel_pages -= memmap_pages;
5347 nr_all_pages += freesize;
5350 * Set an approximate value for lowmem here, it will be adjusted
5351 * when the bootmem allocator frees pages into the buddy system.
5352 * And all highmem pages will be managed by the buddy system.
5354 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5357 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5359 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5361 zone->name = zone_names[j];
5362 spin_lock_init(&zone->lock);
5363 spin_lock_init(&zone->lru_lock);
5364 zone_seqlock_init(zone);
5365 zone->zone_pgdat = pgdat;
5366 zone_pcp_init(zone);
5368 /* For bootup, initialized properly in watermark setup */
5369 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5371 lruvec_init(&zone->lruvec);
5375 set_pageblock_order();
5376 setup_usemap(pgdat, zone, zone_start_pfn, size);
5377 ret = init_currently_empty_zone(zone, zone_start_pfn,
5378 size, MEMMAP_EARLY);
5380 memmap_init(size, nid, j, zone_start_pfn);
5381 zone_start_pfn += size;
5385 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5387 /* Skip empty nodes */
5388 if (!pgdat->node_spanned_pages)
5391 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5392 /* ia64 gets its own node_mem_map, before this, without bootmem */
5393 if (!pgdat->node_mem_map) {
5394 unsigned long size, start, end;
5398 * The zone's endpoints aren't required to be MAX_ORDER
5399 * aligned but the node_mem_map endpoints must be in order
5400 * for the buddy allocator to function correctly.
5402 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5403 end = pgdat_end_pfn(pgdat);
5404 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5405 size = (end - start) * sizeof(struct page);
5406 map = alloc_remap(pgdat->node_id, size);
5408 map = memblock_virt_alloc_node_nopanic(size,
5410 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
5412 #ifndef CONFIG_NEED_MULTIPLE_NODES
5414 * With no DISCONTIG, the global mem_map is just set as node 0's
5416 if (pgdat == NODE_DATA(0)) {
5417 mem_map = NODE_DATA(0)->node_mem_map;
5418 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5419 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5420 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
5421 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5424 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5427 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5428 unsigned long node_start_pfn, unsigned long *zholes_size)
5430 pg_data_t *pgdat = NODE_DATA(nid);
5431 unsigned long start_pfn = 0;
5432 unsigned long end_pfn = 0;
5434 /* pg_data_t should be reset to zero when it's allocated */
5435 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5437 reset_deferred_meminit(pgdat);
5438 pgdat->node_id = nid;
5439 pgdat->node_start_pfn = node_start_pfn;
5440 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5441 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5442 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5443 (u64)start_pfn << PAGE_SHIFT, ((u64)end_pfn << PAGE_SHIFT) - 1);
5445 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5446 zones_size, zholes_size);
5448 alloc_node_mem_map(pgdat);
5449 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5450 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5451 nid, (unsigned long)pgdat,
5452 (unsigned long)pgdat->node_mem_map);
5455 free_area_init_core(pgdat, start_pfn, end_pfn);
5458 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5460 #if MAX_NUMNODES > 1
5462 * Figure out the number of possible node ids.
5464 void __init setup_nr_node_ids(void)
5467 unsigned int highest = 0;
5469 for_each_node_mask(node, node_possible_map)
5471 nr_node_ids = highest + 1;
5476 * node_map_pfn_alignment - determine the maximum internode alignment
5478 * This function should be called after node map is populated and sorted.
5479 * It calculates the maximum power of two alignment which can distinguish
5482 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5483 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5484 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5485 * shifted, 1GiB is enough and this function will indicate so.
5487 * This is used to test whether pfn -> nid mapping of the chosen memory
5488 * model has fine enough granularity to avoid incorrect mapping for the
5489 * populated node map.
5491 * Returns the determined alignment in pfn's. 0 if there is no alignment
5492 * requirement (single node).
5494 unsigned long __init node_map_pfn_alignment(void)
5496 unsigned long accl_mask = 0, last_end = 0;
5497 unsigned long start, end, mask;
5501 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5502 if (!start || last_nid < 0 || last_nid == nid) {
5509 * Start with a mask granular enough to pin-point to the
5510 * start pfn and tick off bits one-by-one until it becomes
5511 * too coarse to separate the current node from the last.
5513 mask = ~((1 << __ffs(start)) - 1);
5514 while (mask && last_end <= (start & (mask << 1)))
5517 /* accumulate all internode masks */
5521 /* convert mask to number of pages */
5522 return ~accl_mask + 1;
5525 /* Find the lowest pfn for a node */
5526 static unsigned long __init find_min_pfn_for_node(int nid)
5528 unsigned long min_pfn = ULONG_MAX;
5529 unsigned long start_pfn;
5532 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5533 min_pfn = min(min_pfn, start_pfn);
5535 if (min_pfn == ULONG_MAX) {
5537 "Could not find start_pfn for node %d\n", nid);
5545 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5547 * It returns the minimum PFN based on information provided via
5548 * memblock_set_node().
5550 unsigned long __init find_min_pfn_with_active_regions(void)
5552 return find_min_pfn_for_node(MAX_NUMNODES);
5556 * early_calculate_totalpages()
5557 * Sum pages in active regions for movable zone.
5558 * Populate N_MEMORY for calculating usable_nodes.
5560 static unsigned long __init early_calculate_totalpages(void)
5562 unsigned long totalpages = 0;
5563 unsigned long start_pfn, end_pfn;
5566 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5567 unsigned long pages = end_pfn - start_pfn;
5569 totalpages += pages;
5571 node_set_state(nid, N_MEMORY);
5577 * Find the PFN the Movable zone begins in each node. Kernel memory
5578 * is spread evenly between nodes as long as the nodes have enough
5579 * memory. When they don't, some nodes will have more kernelcore than
5582 static void __init find_zone_movable_pfns_for_nodes(void)
5585 unsigned long usable_startpfn;
5586 unsigned long kernelcore_node, kernelcore_remaining;
5587 /* save the state before borrow the nodemask */
5588 nodemask_t saved_node_state = node_states[N_MEMORY];
5589 unsigned long totalpages = early_calculate_totalpages();
5590 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5591 struct memblock_region *r;
5593 /* Need to find movable_zone earlier when movable_node is specified. */
5594 find_usable_zone_for_movable();
5597 * If movable_node is specified, ignore kernelcore and movablecore
5600 if (movable_node_is_enabled()) {
5601 for_each_memblock(memory, r) {
5602 if (!memblock_is_hotpluggable(r))
5607 usable_startpfn = PFN_DOWN(r->base);
5608 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5609 min(usable_startpfn, zone_movable_pfn[nid]) :
5617 * If movablecore=nn[KMG] was specified, calculate what size of
5618 * kernelcore that corresponds so that memory usable for
5619 * any allocation type is evenly spread. If both kernelcore
5620 * and movablecore are specified, then the value of kernelcore
5621 * will be used for required_kernelcore if it's greater than
5622 * what movablecore would have allowed.
5624 if (required_movablecore) {
5625 unsigned long corepages;
5628 * Round-up so that ZONE_MOVABLE is at least as large as what
5629 * was requested by the user
5631 required_movablecore =
5632 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5633 corepages = totalpages - required_movablecore;
5635 required_kernelcore = max(required_kernelcore, corepages);
5638 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5639 if (!required_kernelcore)
5642 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5643 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5646 /* Spread kernelcore memory as evenly as possible throughout nodes */
5647 kernelcore_node = required_kernelcore / usable_nodes;
5648 for_each_node_state(nid, N_MEMORY) {
5649 unsigned long start_pfn, end_pfn;
5652 * Recalculate kernelcore_node if the division per node
5653 * now exceeds what is necessary to satisfy the requested
5654 * amount of memory for the kernel
5656 if (required_kernelcore < kernelcore_node)
5657 kernelcore_node = required_kernelcore / usable_nodes;
5660 * As the map is walked, we track how much memory is usable
5661 * by the kernel using kernelcore_remaining. When it is
5662 * 0, the rest of the node is usable by ZONE_MOVABLE
5664 kernelcore_remaining = kernelcore_node;
5666 /* Go through each range of PFNs within this node */
5667 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5668 unsigned long size_pages;
5670 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5671 if (start_pfn >= end_pfn)
5674 /* Account for what is only usable for kernelcore */
5675 if (start_pfn < usable_startpfn) {
5676 unsigned long kernel_pages;
5677 kernel_pages = min(end_pfn, usable_startpfn)
5680 kernelcore_remaining -= min(kernel_pages,
5681 kernelcore_remaining);
5682 required_kernelcore -= min(kernel_pages,
5683 required_kernelcore);
5685 /* Continue if range is now fully accounted */
5686 if (end_pfn <= usable_startpfn) {
5689 * Push zone_movable_pfn to the end so
5690 * that if we have to rebalance
5691 * kernelcore across nodes, we will
5692 * not double account here
5694 zone_movable_pfn[nid] = end_pfn;
5697 start_pfn = usable_startpfn;
5701 * The usable PFN range for ZONE_MOVABLE is from
5702 * start_pfn->end_pfn. Calculate size_pages as the
5703 * number of pages used as kernelcore
5705 size_pages = end_pfn - start_pfn;
5706 if (size_pages > kernelcore_remaining)
5707 size_pages = kernelcore_remaining;
5708 zone_movable_pfn[nid] = start_pfn + size_pages;
5711 * Some kernelcore has been met, update counts and
5712 * break if the kernelcore for this node has been
5715 required_kernelcore -= min(required_kernelcore,
5717 kernelcore_remaining -= size_pages;
5718 if (!kernelcore_remaining)
5724 * If there is still required_kernelcore, we do another pass with one
5725 * less node in the count. This will push zone_movable_pfn[nid] further
5726 * along on the nodes that still have memory until kernelcore is
5730 if (usable_nodes && required_kernelcore > usable_nodes)
5734 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5735 for (nid = 0; nid < MAX_NUMNODES; nid++)
5736 zone_movable_pfn[nid] =
5737 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5740 /* restore the node_state */
5741 node_states[N_MEMORY] = saved_node_state;
5744 /* Any regular or high memory on that node ? */
5745 static void check_for_memory(pg_data_t *pgdat, int nid)
5747 enum zone_type zone_type;
5749 if (N_MEMORY == N_NORMAL_MEMORY)
5752 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5753 struct zone *zone = &pgdat->node_zones[zone_type];
5754 if (populated_zone(zone)) {
5755 node_set_state(nid, N_HIGH_MEMORY);
5756 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5757 zone_type <= ZONE_NORMAL)
5758 node_set_state(nid, N_NORMAL_MEMORY);
5765 * free_area_init_nodes - Initialise all pg_data_t and zone data
5766 * @max_zone_pfn: an array of max PFNs for each zone
5768 * This will call free_area_init_node() for each active node in the system.
5769 * Using the page ranges provided by memblock_set_node(), the size of each
5770 * zone in each node and their holes is calculated. If the maximum PFN
5771 * between two adjacent zones match, it is assumed that the zone is empty.
5772 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5773 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5774 * starts where the previous one ended. For example, ZONE_DMA32 starts
5775 * at arch_max_dma_pfn.
5777 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5779 unsigned long start_pfn, end_pfn;
5782 /* Record where the zone boundaries are */
5783 memset(arch_zone_lowest_possible_pfn, 0,
5784 sizeof(arch_zone_lowest_possible_pfn));
5785 memset(arch_zone_highest_possible_pfn, 0,
5786 sizeof(arch_zone_highest_possible_pfn));
5787 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5788 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5789 for (i = 1; i < MAX_NR_ZONES; i++) {
5790 if (i == ZONE_MOVABLE)
5792 arch_zone_lowest_possible_pfn[i] =
5793 arch_zone_highest_possible_pfn[i-1];
5794 arch_zone_highest_possible_pfn[i] =
5795 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5797 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5798 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5800 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5801 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5802 find_zone_movable_pfns_for_nodes();
5804 /* Print out the zone ranges */
5805 pr_info("Zone ranges:\n");
5806 for (i = 0; i < MAX_NR_ZONES; i++) {
5807 if (i == ZONE_MOVABLE)
5809 pr_info(" %-8s ", zone_names[i]);
5810 if (arch_zone_lowest_possible_pfn[i] ==
5811 arch_zone_highest_possible_pfn[i])
5814 pr_cont("[mem %#018Lx-%#018Lx]\n",
5815 (u64)arch_zone_lowest_possible_pfn[i]
5817 ((u64)arch_zone_highest_possible_pfn[i]
5818 << PAGE_SHIFT) - 1);
5821 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5822 pr_info("Movable zone start for each node\n");
5823 for (i = 0; i < MAX_NUMNODES; i++) {
5824 if (zone_movable_pfn[i])
5825 pr_info(" Node %d: %#018Lx\n", i,
5826 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5829 /* Print out the early node map */
5830 pr_info("Early memory node ranges\n");
5831 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5832 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5833 (u64)start_pfn << PAGE_SHIFT,
5834 ((u64)end_pfn << PAGE_SHIFT) - 1);
5836 /* Initialise every node */
5837 mminit_verify_pageflags_layout();
5838 setup_nr_node_ids();
5839 for_each_online_node(nid) {
5840 pg_data_t *pgdat = NODE_DATA(nid);
5841 free_area_init_node(nid, NULL,
5842 find_min_pfn_for_node(nid), NULL);
5844 /* Any memory on that node */
5845 if (pgdat->node_present_pages)
5846 node_set_state(nid, N_MEMORY);
5847 check_for_memory(pgdat, nid);
5851 static int __init cmdline_parse_core(char *p, unsigned long *core)
5853 unsigned long long coremem;
5857 coremem = memparse(p, &p);
5858 *core = coremem >> PAGE_SHIFT;
5860 /* Paranoid check that UL is enough for the coremem value */
5861 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5867 * kernelcore=size sets the amount of memory for use for allocations that
5868 * cannot be reclaimed or migrated.
5870 static int __init cmdline_parse_kernelcore(char *p)
5872 return cmdline_parse_core(p, &required_kernelcore);
5876 * movablecore=size sets the amount of memory for use for allocations that
5877 * can be reclaimed or migrated.
5879 static int __init cmdline_parse_movablecore(char *p)
5881 return cmdline_parse_core(p, &required_movablecore);
5884 early_param("kernelcore", cmdline_parse_kernelcore);
5885 early_param("movablecore", cmdline_parse_movablecore);
5887 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5889 void adjust_managed_page_count(struct page *page, long count)
5891 spin_lock(&managed_page_count_lock);
5892 page_zone(page)->managed_pages += count;
5893 totalram_pages += count;
5894 #ifdef CONFIG_HIGHMEM
5895 if (PageHighMem(page))
5896 totalhigh_pages += count;
5898 spin_unlock(&managed_page_count_lock);
5900 EXPORT_SYMBOL(adjust_managed_page_count);
5902 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5905 unsigned long pages = 0;
5907 start = (void *)PAGE_ALIGN((unsigned long)start);
5908 end = (void *)((unsigned long)end & PAGE_MASK);
5909 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5910 if ((unsigned int)poison <= 0xFF)
5911 memset(pos, poison, PAGE_SIZE);
5912 free_reserved_page(virt_to_page(pos));
5916 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5917 s, pages << (PAGE_SHIFT - 10), start, end);
5921 EXPORT_SYMBOL(free_reserved_area);
5923 #ifdef CONFIG_HIGHMEM
5924 void free_highmem_page(struct page *page)
5926 __free_reserved_page(page);
5928 page_zone(page)->managed_pages++;
5934 void __init mem_init_print_info(const char *str)
5936 unsigned long physpages, codesize, datasize, rosize, bss_size;
5937 unsigned long init_code_size, init_data_size;
5939 physpages = get_num_physpages();
5940 codesize = _etext - _stext;
5941 datasize = _edata - _sdata;
5942 rosize = __end_rodata - __start_rodata;
5943 bss_size = __bss_stop - __bss_start;
5944 init_data_size = __init_end - __init_begin;
5945 init_code_size = _einittext - _sinittext;
5948 * Detect special cases and adjust section sizes accordingly:
5949 * 1) .init.* may be embedded into .data sections
5950 * 2) .init.text.* may be out of [__init_begin, __init_end],
5951 * please refer to arch/tile/kernel/vmlinux.lds.S.
5952 * 3) .rodata.* may be embedded into .text or .data sections.
5954 #define adj_init_size(start, end, size, pos, adj) \
5956 if (start <= pos && pos < end && size > adj) \
5960 adj_init_size(__init_begin, __init_end, init_data_size,
5961 _sinittext, init_code_size);
5962 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5963 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5964 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5965 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5967 #undef adj_init_size
5969 pr_info("Memory: %luK/%luK available "
5970 "(%luK kernel code, %luK rwdata, %luK rodata, "
5971 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5972 #ifdef CONFIG_HIGHMEM
5976 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5977 codesize >> 10, datasize >> 10, rosize >> 10,
5978 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5979 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5980 totalcma_pages << (PAGE_SHIFT-10),
5981 #ifdef CONFIG_HIGHMEM
5982 totalhigh_pages << (PAGE_SHIFT-10),
5984 str ? ", " : "", str ? str : "");
5988 * set_dma_reserve - set the specified number of pages reserved in the first zone
5989 * @new_dma_reserve: The number of pages to mark reserved
5991 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5992 * In the DMA zone, a significant percentage may be consumed by kernel image
5993 * and other unfreeable allocations which can skew the watermarks badly. This
5994 * function may optionally be used to account for unfreeable pages in the
5995 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5996 * smaller per-cpu batchsize.
5998 void __init set_dma_reserve(unsigned long new_dma_reserve)
6000 dma_reserve = new_dma_reserve;
6003 void __init free_area_init(unsigned long *zones_size)
6005 free_area_init_node(0, zones_size,
6006 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6009 static int page_alloc_cpu_notify(struct notifier_block *self,
6010 unsigned long action, void *hcpu)
6012 int cpu = (unsigned long)hcpu;
6014 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6015 lru_add_drain_cpu(cpu);
6019 * Spill the event counters of the dead processor
6020 * into the current processors event counters.
6021 * This artificially elevates the count of the current
6024 vm_events_fold_cpu(cpu);
6027 * Zero the differential counters of the dead processor
6028 * so that the vm statistics are consistent.
6030 * This is only okay since the processor is dead and cannot
6031 * race with what we are doing.
6033 cpu_vm_stats_fold(cpu);
6038 void __init page_alloc_init(void)
6040 hotcpu_notifier(page_alloc_cpu_notify, 0);
6044 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
6045 * or min_free_kbytes changes.
6047 static void calculate_totalreserve_pages(void)
6049 struct pglist_data *pgdat;
6050 unsigned long reserve_pages = 0;
6051 enum zone_type i, j;
6053 for_each_online_pgdat(pgdat) {
6054 for (i = 0; i < MAX_NR_ZONES; i++) {
6055 struct zone *zone = pgdat->node_zones + i;
6058 /* Find valid and maximum lowmem_reserve in the zone */
6059 for (j = i; j < MAX_NR_ZONES; j++) {
6060 if (zone->lowmem_reserve[j] > max)
6061 max = zone->lowmem_reserve[j];
6064 /* we treat the high watermark as reserved pages. */
6065 max += high_wmark_pages(zone);
6067 if (max > zone->managed_pages)
6068 max = zone->managed_pages;
6069 reserve_pages += max;
6071 * Lowmem reserves are not available to
6072 * GFP_HIGHUSER page cache allocations and
6073 * kswapd tries to balance zones to their high
6074 * watermark. As a result, neither should be
6075 * regarded as dirtyable memory, to prevent a
6076 * situation where reclaim has to clean pages
6077 * in order to balance the zones.
6079 zone->dirty_balance_reserve = max;
6082 dirty_balance_reserve = reserve_pages;
6083 totalreserve_pages = reserve_pages;
6087 * setup_per_zone_lowmem_reserve - called whenever
6088 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
6089 * has a correct pages reserved value, so an adequate number of
6090 * pages are left in the zone after a successful __alloc_pages().
6092 static void setup_per_zone_lowmem_reserve(void)
6094 struct pglist_data *pgdat;
6095 enum zone_type j, idx;
6097 for_each_online_pgdat(pgdat) {
6098 for (j = 0; j < MAX_NR_ZONES; j++) {
6099 struct zone *zone = pgdat->node_zones + j;
6100 unsigned long managed_pages = zone->managed_pages;
6102 zone->lowmem_reserve[j] = 0;
6106 struct zone *lower_zone;
6110 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6111 sysctl_lowmem_reserve_ratio[idx] = 1;
6113 lower_zone = pgdat->node_zones + idx;
6114 lower_zone->lowmem_reserve[j] = managed_pages /
6115 sysctl_lowmem_reserve_ratio[idx];
6116 managed_pages += lower_zone->managed_pages;
6121 /* update totalreserve_pages */
6122 calculate_totalreserve_pages();
6125 static void __setup_per_zone_wmarks(void)
6127 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6128 unsigned long lowmem_pages = 0;
6130 unsigned long flags;
6132 /* Calculate total number of !ZONE_HIGHMEM pages */
6133 for_each_zone(zone) {
6134 if (!is_highmem(zone))
6135 lowmem_pages += zone->managed_pages;
6138 for_each_zone(zone) {
6141 spin_lock_irqsave(&zone->lock, flags);
6142 tmp = (u64)pages_min * zone->managed_pages;
6143 do_div(tmp, lowmem_pages);
6144 if (is_highmem(zone)) {
6146 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6147 * need highmem pages, so cap pages_min to a small
6150 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6151 * deltas control asynch page reclaim, and so should
6152 * not be capped for highmem.
6154 unsigned long min_pages;
6156 min_pages = zone->managed_pages / 1024;
6157 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6158 zone->watermark[WMARK_MIN] = min_pages;
6161 * If it's a lowmem zone, reserve a number of pages
6162 * proportionate to the zone's size.
6164 zone->watermark[WMARK_MIN] = tmp;
6167 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
6168 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
6170 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6171 high_wmark_pages(zone) - low_wmark_pages(zone) -
6172 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6174 setup_zone_migrate_reserve(zone);
6175 spin_unlock_irqrestore(&zone->lock, flags);
6178 /* update totalreserve_pages */
6179 calculate_totalreserve_pages();
6183 * setup_per_zone_wmarks - called when min_free_kbytes changes
6184 * or when memory is hot-{added|removed}
6186 * Ensures that the watermark[min,low,high] values for each zone are set
6187 * correctly with respect to min_free_kbytes.
6189 void setup_per_zone_wmarks(void)
6191 mutex_lock(&zonelists_mutex);
6192 __setup_per_zone_wmarks();
6193 mutex_unlock(&zonelists_mutex);
6197 * The inactive anon list should be small enough that the VM never has to
6198 * do too much work, but large enough that each inactive page has a chance
6199 * to be referenced again before it is swapped out.
6201 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6202 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6203 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6204 * the anonymous pages are kept on the inactive list.
6207 * memory ratio inactive anon
6208 * -------------------------------------
6217 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6219 unsigned int gb, ratio;
6221 /* Zone size in gigabytes */
6222 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6224 ratio = int_sqrt(10 * gb);
6228 zone->inactive_ratio = ratio;
6231 static void __meminit setup_per_zone_inactive_ratio(void)
6236 calculate_zone_inactive_ratio(zone);
6240 * Initialise min_free_kbytes.
6242 * For small machines we want it small (128k min). For large machines
6243 * we want it large (64MB max). But it is not linear, because network
6244 * bandwidth does not increase linearly with machine size. We use
6246 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6247 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6263 int __meminit init_per_zone_wmark_min(void)
6265 unsigned long lowmem_kbytes;
6266 int new_min_free_kbytes;
6268 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6269 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6271 if (new_min_free_kbytes > user_min_free_kbytes) {
6272 min_free_kbytes = new_min_free_kbytes;
6273 if (min_free_kbytes < 128)
6274 min_free_kbytes = 128;
6275 if (min_free_kbytes > 65536)
6276 min_free_kbytes = 65536;
6278 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6279 new_min_free_kbytes, user_min_free_kbytes);
6281 setup_per_zone_wmarks();
6282 refresh_zone_stat_thresholds();
6283 setup_per_zone_lowmem_reserve();
6284 setup_per_zone_inactive_ratio();
6287 module_init(init_per_zone_wmark_min)
6290 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6291 * that we can call two helper functions whenever min_free_kbytes
6294 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6295 void __user *buffer, size_t *length, loff_t *ppos)
6299 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6304 user_min_free_kbytes = min_free_kbytes;
6305 setup_per_zone_wmarks();
6311 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6312 void __user *buffer, size_t *length, loff_t *ppos)
6317 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6322 zone->min_unmapped_pages = (zone->managed_pages *
6323 sysctl_min_unmapped_ratio) / 100;
6327 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6328 void __user *buffer, size_t *length, loff_t *ppos)
6333 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6338 zone->min_slab_pages = (zone->managed_pages *
6339 sysctl_min_slab_ratio) / 100;
6345 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6346 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6347 * whenever sysctl_lowmem_reserve_ratio changes.
6349 * The reserve ratio obviously has absolutely no relation with the
6350 * minimum watermarks. The lowmem reserve ratio can only make sense
6351 * if in function of the boot time zone sizes.
6353 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6354 void __user *buffer, size_t *length, loff_t *ppos)
6356 proc_dointvec_minmax(table, write, buffer, length, ppos);
6357 setup_per_zone_lowmem_reserve();
6362 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6363 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6364 * pagelist can have before it gets flushed back to buddy allocator.
6366 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6367 void __user *buffer, size_t *length, loff_t *ppos)
6370 int old_percpu_pagelist_fraction;
6373 mutex_lock(&pcp_batch_high_lock);
6374 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6376 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6377 if (!write || ret < 0)
6380 /* Sanity checking to avoid pcp imbalance */
6381 if (percpu_pagelist_fraction &&
6382 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6383 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6389 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6392 for_each_populated_zone(zone) {
6395 for_each_possible_cpu(cpu)
6396 pageset_set_high_and_batch(zone,
6397 per_cpu_ptr(zone->pageset, cpu));
6400 mutex_unlock(&pcp_batch_high_lock);
6405 int hashdist = HASHDIST_DEFAULT;
6407 static int __init set_hashdist(char *str)
6411 hashdist = simple_strtoul(str, &str, 0);
6414 __setup("hashdist=", set_hashdist);
6418 * allocate a large system hash table from bootmem
6419 * - it is assumed that the hash table must contain an exact power-of-2
6420 * quantity of entries
6421 * - limit is the number of hash buckets, not the total allocation size
6423 void *__init alloc_large_system_hash(const char *tablename,
6424 unsigned long bucketsize,
6425 unsigned long numentries,
6428 unsigned int *_hash_shift,
6429 unsigned int *_hash_mask,
6430 unsigned long low_limit,
6431 unsigned long high_limit)
6433 unsigned long long max = high_limit;
6434 unsigned long log2qty, size;
6437 /* allow the kernel cmdline to have a say */
6439 /* round applicable memory size up to nearest megabyte */
6440 numentries = nr_kernel_pages;
6442 /* It isn't necessary when PAGE_SIZE >= 1MB */
6443 if (PAGE_SHIFT < 20)
6444 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6446 /* limit to 1 bucket per 2^scale bytes of low memory */
6447 if (scale > PAGE_SHIFT)
6448 numentries >>= (scale - PAGE_SHIFT);
6450 numentries <<= (PAGE_SHIFT - scale);
6452 /* Make sure we've got at least a 0-order allocation.. */
6453 if (unlikely(flags & HASH_SMALL)) {
6454 /* Makes no sense without HASH_EARLY */
6455 WARN_ON(!(flags & HASH_EARLY));
6456 if (!(numentries >> *_hash_shift)) {
6457 numentries = 1UL << *_hash_shift;
6458 BUG_ON(!numentries);
6460 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6461 numentries = PAGE_SIZE / bucketsize;
6463 numentries = roundup_pow_of_two(numentries);
6465 /* limit allocation size to 1/16 total memory by default */
6467 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6468 do_div(max, bucketsize);
6470 max = min(max, 0x80000000ULL);
6472 if (numentries < low_limit)
6473 numentries = low_limit;
6474 if (numentries > max)
6477 log2qty = ilog2(numentries);
6480 size = bucketsize << log2qty;
6481 if (flags & HASH_EARLY)
6482 table = memblock_virt_alloc_nopanic(size, 0);
6484 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6487 * If bucketsize is not a power-of-two, we may free
6488 * some pages at the end of hash table which
6489 * alloc_pages_exact() automatically does
6491 if (get_order(size) < MAX_ORDER) {
6492 table = alloc_pages_exact(size, GFP_ATOMIC);
6493 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6496 } while (!table && size > PAGE_SIZE && --log2qty);
6499 panic("Failed to allocate %s hash table\n", tablename);
6501 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6504 ilog2(size) - PAGE_SHIFT,
6508 *_hash_shift = log2qty;
6510 *_hash_mask = (1 << log2qty) - 1;
6515 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6516 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6519 #ifdef CONFIG_SPARSEMEM
6520 return __pfn_to_section(pfn)->pageblock_flags;
6522 return zone->pageblock_flags;
6523 #endif /* CONFIG_SPARSEMEM */
6526 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6528 #ifdef CONFIG_SPARSEMEM
6529 pfn &= (PAGES_PER_SECTION-1);
6530 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6532 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6533 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6534 #endif /* CONFIG_SPARSEMEM */
6538 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6539 * @page: The page within the block of interest
6540 * @pfn: The target page frame number
6541 * @end_bitidx: The last bit of interest to retrieve
6542 * @mask: mask of bits that the caller is interested in
6544 * Return: pageblock_bits flags
6546 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6547 unsigned long end_bitidx,
6551 unsigned long *bitmap;
6552 unsigned long bitidx, word_bitidx;
6555 zone = page_zone(page);
6556 bitmap = get_pageblock_bitmap(zone, pfn);
6557 bitidx = pfn_to_bitidx(zone, pfn);
6558 word_bitidx = bitidx / BITS_PER_LONG;
6559 bitidx &= (BITS_PER_LONG-1);
6561 word = bitmap[word_bitidx];
6562 bitidx += end_bitidx;
6563 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6567 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6568 * @page: The page within the block of interest
6569 * @flags: The flags to set
6570 * @pfn: The target page frame number
6571 * @end_bitidx: The last bit of interest
6572 * @mask: mask of bits that the caller is interested in
6574 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6576 unsigned long end_bitidx,
6580 unsigned long *bitmap;
6581 unsigned long bitidx, word_bitidx;
6582 unsigned long old_word, word;
6584 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6586 zone = page_zone(page);
6587 bitmap = get_pageblock_bitmap(zone, pfn);
6588 bitidx = pfn_to_bitidx(zone, pfn);
6589 word_bitidx = bitidx / BITS_PER_LONG;
6590 bitidx &= (BITS_PER_LONG-1);
6592 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6594 bitidx += end_bitidx;
6595 mask <<= (BITS_PER_LONG - bitidx - 1);
6596 flags <<= (BITS_PER_LONG - bitidx - 1);
6598 word = READ_ONCE(bitmap[word_bitidx]);
6600 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6601 if (word == old_word)
6608 * This function checks whether pageblock includes unmovable pages or not.
6609 * If @count is not zero, it is okay to include less @count unmovable pages
6611 * PageLRU check without isolation or lru_lock could race so that
6612 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6613 * expect this function should be exact.
6615 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6616 bool skip_hwpoisoned_pages)
6618 unsigned long pfn, iter, found;
6622 * For avoiding noise data, lru_add_drain_all() should be called
6623 * If ZONE_MOVABLE, the zone never contains unmovable pages
6625 if (zone_idx(zone) == ZONE_MOVABLE)
6627 mt = get_pageblock_migratetype(page);
6628 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6631 pfn = page_to_pfn(page);
6632 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6633 unsigned long check = pfn + iter;
6635 if (!pfn_valid_within(check))
6638 page = pfn_to_page(check);
6641 * Hugepages are not in LRU lists, but they're movable.
6642 * We need not scan over tail pages bacause we don't
6643 * handle each tail page individually in migration.
6645 if (PageHuge(page)) {
6646 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6651 * We can't use page_count without pin a page
6652 * because another CPU can free compound page.
6653 * This check already skips compound tails of THP
6654 * because their page->_count is zero at all time.
6656 if (!atomic_read(&page->_count)) {
6657 if (PageBuddy(page))
6658 iter += (1 << page_order(page)) - 1;
6663 * The HWPoisoned page may be not in buddy system, and
6664 * page_count() is not 0.
6666 if (skip_hwpoisoned_pages && PageHWPoison(page))
6672 * If there are RECLAIMABLE pages, we need to check
6673 * it. But now, memory offline itself doesn't call
6674 * shrink_node_slabs() and it still to be fixed.
6677 * If the page is not RAM, page_count()should be 0.
6678 * we don't need more check. This is an _used_ not-movable page.
6680 * The problematic thing here is PG_reserved pages. PG_reserved
6681 * is set to both of a memory hole page and a _used_ kernel
6690 bool is_pageblock_removable_nolock(struct page *page)
6696 * We have to be careful here because we are iterating over memory
6697 * sections which are not zone aware so we might end up outside of
6698 * the zone but still within the section.
6699 * We have to take care about the node as well. If the node is offline
6700 * its NODE_DATA will be NULL - see page_zone.
6702 if (!node_online(page_to_nid(page)))
6705 zone = page_zone(page);
6706 pfn = page_to_pfn(page);
6707 if (!zone_spans_pfn(zone, pfn))
6710 return !has_unmovable_pages(zone, page, 0, true);
6715 static unsigned long pfn_max_align_down(unsigned long pfn)
6717 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6718 pageblock_nr_pages) - 1);
6721 static unsigned long pfn_max_align_up(unsigned long pfn)
6723 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6724 pageblock_nr_pages));
6727 /* [start, end) must belong to a single zone. */
6728 static int __alloc_contig_migrate_range(struct compact_control *cc,
6729 unsigned long start, unsigned long end)
6731 /* This function is based on compact_zone() from compaction.c. */
6732 unsigned long nr_reclaimed;
6733 unsigned long pfn = start;
6734 unsigned int tries = 0;
6739 while (pfn < end || !list_empty(&cc->migratepages)) {
6740 if (fatal_signal_pending(current)) {
6745 if (list_empty(&cc->migratepages)) {
6746 cc->nr_migratepages = 0;
6747 pfn = isolate_migratepages_range(cc, pfn, end);
6753 } else if (++tries == 5) {
6754 ret = ret < 0 ? ret : -EBUSY;
6758 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6760 cc->nr_migratepages -= nr_reclaimed;
6762 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6763 NULL, 0, cc->mode, MR_CMA);
6766 putback_movable_pages(&cc->migratepages);
6773 * alloc_contig_range() -- tries to allocate given range of pages
6774 * @start: start PFN to allocate
6775 * @end: one-past-the-last PFN to allocate
6776 * @migratetype: migratetype of the underlaying pageblocks (either
6777 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6778 * in range must have the same migratetype and it must
6779 * be either of the two.
6781 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6782 * aligned, however it's the caller's responsibility to guarantee that
6783 * we are the only thread that changes migrate type of pageblocks the
6786 * The PFN range must belong to a single zone.
6788 * Returns zero on success or negative error code. On success all
6789 * pages which PFN is in [start, end) are allocated for the caller and
6790 * need to be freed with free_contig_range().
6792 int alloc_contig_range(unsigned long start, unsigned long end,
6793 unsigned migratetype)
6795 unsigned long outer_start, outer_end;
6798 struct compact_control cc = {
6799 .nr_migratepages = 0,
6801 .zone = page_zone(pfn_to_page(start)),
6802 .mode = MIGRATE_SYNC,
6803 .ignore_skip_hint = true,
6805 INIT_LIST_HEAD(&cc.migratepages);
6808 * What we do here is we mark all pageblocks in range as
6809 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6810 * have different sizes, and due to the way page allocator
6811 * work, we align the range to biggest of the two pages so
6812 * that page allocator won't try to merge buddies from
6813 * different pageblocks and change MIGRATE_ISOLATE to some
6814 * other migration type.
6816 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6817 * migrate the pages from an unaligned range (ie. pages that
6818 * we are interested in). This will put all the pages in
6819 * range back to page allocator as MIGRATE_ISOLATE.
6821 * When this is done, we take the pages in range from page
6822 * allocator removing them from the buddy system. This way
6823 * page allocator will never consider using them.
6825 * This lets us mark the pageblocks back as
6826 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6827 * aligned range but not in the unaligned, original range are
6828 * put back to page allocator so that buddy can use them.
6831 ret = start_isolate_page_range(pfn_max_align_down(start),
6832 pfn_max_align_up(end), migratetype,
6837 ret = __alloc_contig_migrate_range(&cc, start, end);
6842 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6843 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6844 * more, all pages in [start, end) are free in page allocator.
6845 * What we are going to do is to allocate all pages from
6846 * [start, end) (that is remove them from page allocator).
6848 * The only problem is that pages at the beginning and at the
6849 * end of interesting range may be not aligned with pages that
6850 * page allocator holds, ie. they can be part of higher order
6851 * pages. Because of this, we reserve the bigger range and
6852 * once this is done free the pages we are not interested in.
6854 * We don't have to hold zone->lock here because the pages are
6855 * isolated thus they won't get removed from buddy.
6858 lru_add_drain_all();
6859 drain_all_pages(cc.zone);
6862 outer_start = start;
6863 while (!PageBuddy(pfn_to_page(outer_start))) {
6864 if (++order >= MAX_ORDER) {
6868 outer_start &= ~0UL << order;
6871 /* Make sure the range is really isolated. */
6872 if (test_pages_isolated(outer_start, end, false)) {
6873 pr_info("%s: [%lx, %lx) PFNs busy\n",
6874 __func__, outer_start, end);
6879 /* Grab isolated pages from freelists. */
6880 outer_end = isolate_freepages_range(&cc, outer_start, end);
6886 /* Free head and tail (if any) */
6887 if (start != outer_start)
6888 free_contig_range(outer_start, start - outer_start);
6889 if (end != outer_end)
6890 free_contig_range(end, outer_end - end);
6893 undo_isolate_page_range(pfn_max_align_down(start),
6894 pfn_max_align_up(end), migratetype);
6898 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6900 unsigned int count = 0;
6902 for (; nr_pages--; pfn++) {
6903 struct page *page = pfn_to_page(pfn);
6905 count += page_count(page) != 1;
6908 WARN(count != 0, "%d pages are still in use!\n", count);
6912 #ifdef CONFIG_MEMORY_HOTPLUG
6914 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6915 * page high values need to be recalulated.
6917 void __meminit zone_pcp_update(struct zone *zone)
6920 mutex_lock(&pcp_batch_high_lock);
6921 for_each_possible_cpu(cpu)
6922 pageset_set_high_and_batch(zone,
6923 per_cpu_ptr(zone->pageset, cpu));
6924 mutex_unlock(&pcp_batch_high_lock);
6928 void zone_pcp_reset(struct zone *zone)
6930 unsigned long flags;
6932 struct per_cpu_pageset *pset;
6934 /* avoid races with drain_pages() */
6935 local_irq_save(flags);
6936 if (zone->pageset != &boot_pageset) {
6937 for_each_online_cpu(cpu) {
6938 pset = per_cpu_ptr(zone->pageset, cpu);
6939 drain_zonestat(zone, pset);
6941 free_percpu(zone->pageset);
6942 zone->pageset = &boot_pageset;
6944 local_irq_restore(flags);
6947 #ifdef CONFIG_MEMORY_HOTREMOVE
6949 * All pages in the range must be isolated before calling this.
6952 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6956 unsigned int order, i;
6958 unsigned long flags;
6959 /* find the first valid pfn */
6960 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6965 zone = page_zone(pfn_to_page(pfn));
6966 spin_lock_irqsave(&zone->lock, flags);
6968 while (pfn < end_pfn) {
6969 if (!pfn_valid(pfn)) {
6973 page = pfn_to_page(pfn);
6975 * The HWPoisoned page may be not in buddy system, and
6976 * page_count() is not 0.
6978 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6980 SetPageReserved(page);
6984 BUG_ON(page_count(page));
6985 BUG_ON(!PageBuddy(page));
6986 order = page_order(page);
6987 #ifdef CONFIG_DEBUG_VM
6988 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6989 pfn, 1 << order, end_pfn);
6991 list_del(&page->lru);
6992 rmv_page_order(page);
6993 zone->free_area[order].nr_free--;
6994 for (i = 0; i < (1 << order); i++)
6995 SetPageReserved((page+i));
6996 pfn += (1 << order);
6998 spin_unlock_irqrestore(&zone->lock, flags);
7002 #ifdef CONFIG_MEMORY_FAILURE
7003 bool is_free_buddy_page(struct page *page)
7005 struct zone *zone = page_zone(page);
7006 unsigned long pfn = page_to_pfn(page);
7007 unsigned long flags;
7010 spin_lock_irqsave(&zone->lock, flags);
7011 for (order = 0; order < MAX_ORDER; order++) {
7012 struct page *page_head = page - (pfn & ((1 << order) - 1));
7014 if (PageBuddy(page_head) && page_order(page_head) >= order)
7017 spin_unlock_irqrestore(&zone->lock, flags);
7019 return order < MAX_ORDER;