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 int percpu_pagelist_fraction;
119 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
122 * A cached value of the page's pageblock's migratetype, used when the page is
123 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
124 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
125 * Also the migratetype set in the page does not necessarily match the pcplist
126 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
127 * other index - this ensures that it will be put on the correct CMA freelist.
129 static inline int get_pcppage_migratetype(struct page *page)
134 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
136 page->index = migratetype;
139 #ifdef CONFIG_PM_SLEEP
141 * The following functions are used by the suspend/hibernate code to temporarily
142 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
143 * while devices are suspended. To avoid races with the suspend/hibernate code,
144 * they should always be called with pm_mutex held (gfp_allowed_mask also should
145 * only be modified with pm_mutex held, unless the suspend/hibernate code is
146 * guaranteed not to run in parallel with that modification).
149 static gfp_t saved_gfp_mask;
151 void pm_restore_gfp_mask(void)
153 WARN_ON(!mutex_is_locked(&pm_mutex));
154 if (saved_gfp_mask) {
155 gfp_allowed_mask = saved_gfp_mask;
160 void pm_restrict_gfp_mask(void)
162 WARN_ON(!mutex_is_locked(&pm_mutex));
163 WARN_ON(saved_gfp_mask);
164 saved_gfp_mask = gfp_allowed_mask;
165 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
168 bool pm_suspended_storage(void)
170 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
174 #endif /* CONFIG_PM_SLEEP */
176 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
177 unsigned int pageblock_order __read_mostly;
180 static void __free_pages_ok(struct page *page, unsigned int order);
183 * results with 256, 32 in the lowmem_reserve sysctl:
184 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
185 * 1G machine -> (16M dma, 784M normal, 224M high)
186 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
187 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
188 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
190 * TBD: should special case ZONE_DMA32 machines here - in those we normally
191 * don't need any ZONE_NORMAL reservation
193 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
194 #ifdef CONFIG_ZONE_DMA
197 #ifdef CONFIG_ZONE_DMA32
200 #ifdef CONFIG_HIGHMEM
206 EXPORT_SYMBOL(totalram_pages);
208 static char * const zone_names[MAX_NR_ZONES] = {
209 #ifdef CONFIG_ZONE_DMA
212 #ifdef CONFIG_ZONE_DMA32
216 #ifdef CONFIG_HIGHMEM
220 #ifdef CONFIG_ZONE_DEVICE
225 compound_page_dtor * const compound_page_dtors[] = {
228 #ifdef CONFIG_HUGETLB_PAGE
231 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
236 int min_free_kbytes = 1024;
237 int user_min_free_kbytes = -1;
239 static unsigned long __meminitdata nr_kernel_pages;
240 static unsigned long __meminitdata nr_all_pages;
241 static unsigned long __meminitdata dma_reserve;
243 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
244 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
245 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
246 static unsigned long __initdata required_kernelcore;
247 static unsigned long __initdata required_movablecore;
248 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
250 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
252 EXPORT_SYMBOL(movable_zone);
253 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
256 int nr_node_ids __read_mostly = MAX_NUMNODES;
257 int nr_online_nodes __read_mostly = 1;
258 EXPORT_SYMBOL(nr_node_ids);
259 EXPORT_SYMBOL(nr_online_nodes);
262 int page_group_by_mobility_disabled __read_mostly;
264 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
265 static inline void reset_deferred_meminit(pg_data_t *pgdat)
267 pgdat->first_deferred_pfn = ULONG_MAX;
270 /* Returns true if the struct page for the pfn is uninitialised */
271 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
273 if (pfn >= NODE_DATA(early_pfn_to_nid(pfn))->first_deferred_pfn)
279 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
281 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
288 * Returns false when the remaining initialisation should be deferred until
289 * later in the boot cycle when it can be parallelised.
291 static inline bool update_defer_init(pg_data_t *pgdat,
292 unsigned long pfn, unsigned long zone_end,
293 unsigned long *nr_initialised)
295 /* Always populate low zones for address-contrained allocations */
296 if (zone_end < pgdat_end_pfn(pgdat))
299 /* Initialise at least 2G of the highest zone */
301 if (*nr_initialised > (2UL << (30 - PAGE_SHIFT)) &&
302 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
303 pgdat->first_deferred_pfn = pfn;
310 static inline void reset_deferred_meminit(pg_data_t *pgdat)
314 static inline bool early_page_uninitialised(unsigned long pfn)
319 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
324 static inline bool update_defer_init(pg_data_t *pgdat,
325 unsigned long pfn, unsigned long zone_end,
326 unsigned long *nr_initialised)
333 void set_pageblock_migratetype(struct page *page, int migratetype)
335 if (unlikely(page_group_by_mobility_disabled &&
336 migratetype < MIGRATE_PCPTYPES))
337 migratetype = MIGRATE_UNMOVABLE;
339 set_pageblock_flags_group(page, (unsigned long)migratetype,
340 PB_migrate, PB_migrate_end);
343 #ifdef CONFIG_DEBUG_VM
344 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
348 unsigned long pfn = page_to_pfn(page);
349 unsigned long sp, start_pfn;
352 seq = zone_span_seqbegin(zone);
353 start_pfn = zone->zone_start_pfn;
354 sp = zone->spanned_pages;
355 if (!zone_spans_pfn(zone, pfn))
357 } while (zone_span_seqretry(zone, seq));
360 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
361 pfn, zone_to_nid(zone), zone->name,
362 start_pfn, start_pfn + sp);
367 static int page_is_consistent(struct zone *zone, struct page *page)
369 if (!pfn_valid_within(page_to_pfn(page)))
371 if (zone != page_zone(page))
377 * Temporary debugging check for pages not lying within a given zone.
379 static int bad_range(struct zone *zone, struct page *page)
381 if (page_outside_zone_boundaries(zone, page))
383 if (!page_is_consistent(zone, page))
389 static inline int bad_range(struct zone *zone, struct page *page)
395 static void bad_page(struct page *page, const char *reason,
396 unsigned long bad_flags)
398 static unsigned long resume;
399 static unsigned long nr_shown;
400 static unsigned long nr_unshown;
402 /* Don't complain about poisoned pages */
403 if (PageHWPoison(page)) {
404 page_mapcount_reset(page); /* remove PageBuddy */
409 * Allow a burst of 60 reports, then keep quiet for that minute;
410 * or allow a steady drip of one report per second.
412 if (nr_shown == 60) {
413 if (time_before(jiffies, resume)) {
419 "BUG: Bad page state: %lu messages suppressed\n",
426 resume = jiffies + 60 * HZ;
428 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
429 current->comm, page_to_pfn(page));
430 dump_page_badflags(page, reason, bad_flags);
435 /* Leave bad fields for debug, except PageBuddy could make trouble */
436 page_mapcount_reset(page); /* remove PageBuddy */
437 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
441 * Higher-order pages are called "compound pages". They are structured thusly:
443 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
445 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
446 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
448 * The first tail page's ->compound_dtor holds the offset in array of compound
449 * page destructors. See compound_page_dtors.
451 * The first tail page's ->compound_order holds the order of allocation.
452 * This usage means that zero-order pages may not be compound.
455 void free_compound_page(struct page *page)
457 __free_pages_ok(page, compound_order(page));
460 void prep_compound_page(struct page *page, unsigned int order)
463 int nr_pages = 1 << order;
465 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
466 set_compound_order(page, order);
468 for (i = 1; i < nr_pages; i++) {
469 struct page *p = page + i;
470 set_page_count(p, 0);
471 p->mapping = TAIL_MAPPING;
472 set_compound_head(p, page);
474 atomic_set(compound_mapcount_ptr(page), -1);
477 #ifdef CONFIG_DEBUG_PAGEALLOC
478 unsigned int _debug_guardpage_minorder;
479 bool _debug_pagealloc_enabled __read_mostly;
480 bool _debug_guardpage_enabled __read_mostly;
482 static int __init early_debug_pagealloc(char *buf)
487 if (strcmp(buf, "on") == 0)
488 _debug_pagealloc_enabled = true;
492 early_param("debug_pagealloc", early_debug_pagealloc);
494 static bool need_debug_guardpage(void)
496 /* If we don't use debug_pagealloc, we don't need guard page */
497 if (!debug_pagealloc_enabled())
503 static void init_debug_guardpage(void)
505 if (!debug_pagealloc_enabled())
508 _debug_guardpage_enabled = true;
511 struct page_ext_operations debug_guardpage_ops = {
512 .need = need_debug_guardpage,
513 .init = init_debug_guardpage,
516 static int __init debug_guardpage_minorder_setup(char *buf)
520 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
521 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
524 _debug_guardpage_minorder = res;
525 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
528 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
530 static inline void set_page_guard(struct zone *zone, struct page *page,
531 unsigned int order, int migratetype)
533 struct page_ext *page_ext;
535 if (!debug_guardpage_enabled())
538 page_ext = lookup_page_ext(page);
539 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
541 INIT_LIST_HEAD(&page->lru);
542 set_page_private(page, order);
543 /* Guard pages are not available for any usage */
544 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
547 static inline void clear_page_guard(struct zone *zone, struct page *page,
548 unsigned int order, int migratetype)
550 struct page_ext *page_ext;
552 if (!debug_guardpage_enabled())
555 page_ext = lookup_page_ext(page);
556 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
558 set_page_private(page, 0);
559 if (!is_migrate_isolate(migratetype))
560 __mod_zone_freepage_state(zone, (1 << order), migratetype);
563 struct page_ext_operations debug_guardpage_ops = { NULL, };
564 static inline void set_page_guard(struct zone *zone, struct page *page,
565 unsigned int order, int migratetype) {}
566 static inline void clear_page_guard(struct zone *zone, struct page *page,
567 unsigned int order, int migratetype) {}
570 static inline void set_page_order(struct page *page, unsigned int order)
572 set_page_private(page, order);
573 __SetPageBuddy(page);
576 static inline void rmv_page_order(struct page *page)
578 __ClearPageBuddy(page);
579 set_page_private(page, 0);
583 * This function checks whether a page is free && is the buddy
584 * we can do coalesce a page and its buddy if
585 * (a) the buddy is not in a hole &&
586 * (b) the buddy is in the buddy system &&
587 * (c) a page and its buddy have the same order &&
588 * (d) a page and its buddy are in the same zone.
590 * For recording whether a page is in the buddy system, we set ->_mapcount
591 * PAGE_BUDDY_MAPCOUNT_VALUE.
592 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
593 * serialized by zone->lock.
595 * For recording page's order, we use page_private(page).
597 static inline int page_is_buddy(struct page *page, struct page *buddy,
600 if (!pfn_valid_within(page_to_pfn(buddy)))
603 if (page_is_guard(buddy) && page_order(buddy) == order) {
604 if (page_zone_id(page) != page_zone_id(buddy))
607 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
612 if (PageBuddy(buddy) && page_order(buddy) == order) {
614 * zone check is done late to avoid uselessly
615 * calculating zone/node ids for pages that could
618 if (page_zone_id(page) != page_zone_id(buddy))
621 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
629 * Freeing function for a buddy system allocator.
631 * The concept of a buddy system is to maintain direct-mapped table
632 * (containing bit values) for memory blocks of various "orders".
633 * The bottom level table contains the map for the smallest allocatable
634 * units of memory (here, pages), and each level above it describes
635 * pairs of units from the levels below, hence, "buddies".
636 * At a high level, all that happens here is marking the table entry
637 * at the bottom level available, and propagating the changes upward
638 * as necessary, plus some accounting needed to play nicely with other
639 * parts of the VM system.
640 * At each level, we keep a list of pages, which are heads of continuous
641 * free pages of length of (1 << order) and marked with _mapcount
642 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
644 * So when we are allocating or freeing one, we can derive the state of the
645 * other. That is, if we allocate a small block, and both were
646 * free, the remainder of the region must be split into blocks.
647 * If a block is freed, and its buddy is also free, then this
648 * triggers coalescing into a block of larger size.
653 static inline void __free_one_page(struct page *page,
655 struct zone *zone, unsigned int order,
658 unsigned long page_idx;
659 unsigned long combined_idx;
660 unsigned long uninitialized_var(buddy_idx);
662 unsigned int max_order = MAX_ORDER;
664 VM_BUG_ON(!zone_is_initialized(zone));
665 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
667 VM_BUG_ON(migratetype == -1);
668 if (is_migrate_isolate(migratetype)) {
670 * We restrict max order of merging to prevent merge
671 * between freepages on isolate pageblock and normal
672 * pageblock. Without this, pageblock isolation
673 * could cause incorrect freepage accounting.
675 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
677 __mod_zone_freepage_state(zone, 1 << order, migratetype);
680 page_idx = pfn & ((1 << max_order) - 1);
682 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
683 VM_BUG_ON_PAGE(bad_range(zone, page), page);
685 while (order < max_order - 1) {
686 buddy_idx = __find_buddy_index(page_idx, order);
687 buddy = page + (buddy_idx - page_idx);
688 if (!page_is_buddy(page, buddy, order))
691 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
692 * merge with it and move up one order.
694 if (page_is_guard(buddy)) {
695 clear_page_guard(zone, buddy, order, migratetype);
697 list_del(&buddy->lru);
698 zone->free_area[order].nr_free--;
699 rmv_page_order(buddy);
701 combined_idx = buddy_idx & page_idx;
702 page = page + (combined_idx - page_idx);
703 page_idx = combined_idx;
706 set_page_order(page, order);
709 * If this is not the largest possible page, check if the buddy
710 * of the next-highest order is free. If it is, it's possible
711 * that pages are being freed that will coalesce soon. In case,
712 * that is happening, add the free page to the tail of the list
713 * so it's less likely to be used soon and more likely to be merged
714 * as a higher order page
716 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
717 struct page *higher_page, *higher_buddy;
718 combined_idx = buddy_idx & page_idx;
719 higher_page = page + (combined_idx - page_idx);
720 buddy_idx = __find_buddy_index(combined_idx, order + 1);
721 higher_buddy = higher_page + (buddy_idx - combined_idx);
722 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
723 list_add_tail(&page->lru,
724 &zone->free_area[order].free_list[migratetype]);
729 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
731 zone->free_area[order].nr_free++;
734 static inline int free_pages_check(struct page *page)
736 const char *bad_reason = NULL;
737 unsigned long bad_flags = 0;
739 if (unlikely(atomic_read(&page->_mapcount) != -1))
740 bad_reason = "nonzero mapcount";
741 if (unlikely(page->mapping != NULL))
742 bad_reason = "non-NULL mapping";
743 if (unlikely(atomic_read(&page->_count) != 0))
744 bad_reason = "nonzero _count";
745 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
746 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
747 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
750 if (unlikely(page->mem_cgroup))
751 bad_reason = "page still charged to cgroup";
753 if (unlikely(bad_reason)) {
754 bad_page(page, bad_reason, bad_flags);
757 page_cpupid_reset_last(page);
758 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
759 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
764 * Frees a number of pages from the PCP lists
765 * Assumes all pages on list are in same zone, and of same order.
766 * count is the number of pages to free.
768 * If the zone was previously in an "all pages pinned" state then look to
769 * see if this freeing clears that state.
771 * And clear the zone's pages_scanned counter, to hold off the "all pages are
772 * pinned" detection logic.
774 static void free_pcppages_bulk(struct zone *zone, int count,
775 struct per_cpu_pages *pcp)
780 unsigned long nr_scanned;
782 spin_lock(&zone->lock);
783 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
785 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
789 struct list_head *list;
792 * Remove pages from lists in a round-robin fashion. A
793 * batch_free count is maintained that is incremented when an
794 * empty list is encountered. This is so more pages are freed
795 * off fuller lists instead of spinning excessively around empty
800 if (++migratetype == MIGRATE_PCPTYPES)
802 list = &pcp->lists[migratetype];
803 } while (list_empty(list));
805 /* This is the only non-empty list. Free them all. */
806 if (batch_free == MIGRATE_PCPTYPES)
807 batch_free = to_free;
810 int mt; /* migratetype of the to-be-freed page */
812 page = list_last_entry(list, struct page, lru);
813 /* must delete as __free_one_page list manipulates */
814 list_del(&page->lru);
816 mt = get_pcppage_migratetype(page);
817 /* MIGRATE_ISOLATE page should not go to pcplists */
818 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
819 /* Pageblock could have been isolated meanwhile */
820 if (unlikely(has_isolate_pageblock(zone)))
821 mt = get_pageblock_migratetype(page);
823 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
824 trace_mm_page_pcpu_drain(page, 0, mt);
825 } while (--to_free && --batch_free && !list_empty(list));
827 spin_unlock(&zone->lock);
830 static void free_one_page(struct zone *zone,
831 struct page *page, unsigned long pfn,
835 unsigned long nr_scanned;
836 spin_lock(&zone->lock);
837 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
839 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
841 if (unlikely(has_isolate_pageblock(zone) ||
842 is_migrate_isolate(migratetype))) {
843 migratetype = get_pfnblock_migratetype(page, pfn);
845 __free_one_page(page, pfn, zone, order, migratetype);
846 spin_unlock(&zone->lock);
849 static int free_tail_pages_check(struct page *head_page, struct page *page)
854 * We rely page->lru.next never has bit 0 set, unless the page
855 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
857 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
859 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
863 switch (page - head_page) {
865 /* the first tail page: ->mapping is compound_mapcount() */
866 if (unlikely(compound_mapcount(page))) {
867 bad_page(page, "nonzero compound_mapcount", 0);
873 * the second tail page: ->mapping is
874 * page_deferred_list().next -- ignore value.
878 if (page->mapping != TAIL_MAPPING) {
879 bad_page(page, "corrupted mapping in tail page", 0);
884 if (unlikely(!PageTail(page))) {
885 bad_page(page, "PageTail not set", 0);
888 if (unlikely(compound_head(page) != head_page)) {
889 bad_page(page, "compound_head not consistent", 0);
894 page->mapping = NULL;
895 clear_compound_head(page);
899 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
900 unsigned long zone, int nid)
902 set_page_links(page, zone, nid, pfn);
903 init_page_count(page);
904 page_mapcount_reset(page);
905 page_cpupid_reset_last(page);
907 INIT_LIST_HEAD(&page->lru);
908 #ifdef WANT_PAGE_VIRTUAL
909 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
910 if (!is_highmem_idx(zone))
911 set_page_address(page, __va(pfn << PAGE_SHIFT));
915 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
918 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
921 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
922 static void init_reserved_page(unsigned long pfn)
927 if (!early_page_uninitialised(pfn))
930 nid = early_pfn_to_nid(pfn);
931 pgdat = NODE_DATA(nid);
933 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
934 struct zone *zone = &pgdat->node_zones[zid];
936 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
939 __init_single_pfn(pfn, zid, nid);
942 static inline void init_reserved_page(unsigned long pfn)
945 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
948 * Initialised pages do not have PageReserved set. This function is
949 * called for each range allocated by the bootmem allocator and
950 * marks the pages PageReserved. The remaining valid pages are later
951 * sent to the buddy page allocator.
953 void __meminit reserve_bootmem_region(unsigned long start, unsigned long end)
955 unsigned long start_pfn = PFN_DOWN(start);
956 unsigned long end_pfn = PFN_UP(end);
958 for (; start_pfn < end_pfn; start_pfn++) {
959 if (pfn_valid(start_pfn)) {
960 struct page *page = pfn_to_page(start_pfn);
962 init_reserved_page(start_pfn);
964 /* Avoid false-positive PageTail() */
965 INIT_LIST_HEAD(&page->lru);
967 SetPageReserved(page);
972 static bool free_pages_prepare(struct page *page, unsigned int order)
974 bool compound = PageCompound(page);
977 VM_BUG_ON_PAGE(PageTail(page), page);
978 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
980 trace_mm_page_free(page, order);
981 kmemcheck_free_shadow(page, order);
982 kasan_free_pages(page, order);
985 page->mapping = NULL;
986 bad += free_pages_check(page);
987 for (i = 1; i < (1 << order); i++) {
989 bad += free_tail_pages_check(page, page + i);
990 bad += free_pages_check(page + i);
995 reset_page_owner(page, order);
997 if (!PageHighMem(page)) {
998 debug_check_no_locks_freed(page_address(page),
1000 debug_check_no_obj_freed(page_address(page),
1001 PAGE_SIZE << order);
1003 arch_free_page(page, order);
1004 kernel_map_pages(page, 1 << order, 0);
1009 static void __free_pages_ok(struct page *page, unsigned int order)
1011 unsigned long flags;
1013 unsigned long pfn = page_to_pfn(page);
1015 if (!free_pages_prepare(page, order))
1018 migratetype = get_pfnblock_migratetype(page, pfn);
1019 local_irq_save(flags);
1020 __count_vm_events(PGFREE, 1 << order);
1021 free_one_page(page_zone(page), page, pfn, order, migratetype);
1022 local_irq_restore(flags);
1025 static void __init __free_pages_boot_core(struct page *page,
1026 unsigned long pfn, unsigned int order)
1028 unsigned int nr_pages = 1 << order;
1029 struct page *p = page;
1033 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1035 __ClearPageReserved(p);
1036 set_page_count(p, 0);
1038 __ClearPageReserved(p);
1039 set_page_count(p, 0);
1041 page_zone(page)->managed_pages += nr_pages;
1042 set_page_refcounted(page);
1043 __free_pages(page, order);
1046 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1047 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1049 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1051 int __meminit early_pfn_to_nid(unsigned long pfn)
1053 static DEFINE_SPINLOCK(early_pfn_lock);
1056 spin_lock(&early_pfn_lock);
1057 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1060 spin_unlock(&early_pfn_lock);
1066 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1067 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1068 struct mminit_pfnnid_cache *state)
1072 nid = __early_pfn_to_nid(pfn, state);
1073 if (nid >= 0 && nid != node)
1078 /* Only safe to use early in boot when initialisation is single-threaded */
1079 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1081 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1086 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1090 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1091 struct mminit_pfnnid_cache *state)
1098 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1101 if (early_page_uninitialised(pfn))
1103 return __free_pages_boot_core(page, pfn, order);
1106 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1107 static void __init deferred_free_range(struct page *page,
1108 unsigned long pfn, int nr_pages)
1115 /* Free a large naturally-aligned chunk if possible */
1116 if (nr_pages == MAX_ORDER_NR_PAGES &&
1117 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1118 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1119 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1123 for (i = 0; i < nr_pages; i++, page++, pfn++)
1124 __free_pages_boot_core(page, pfn, 0);
1127 /* Completion tracking for deferred_init_memmap() threads */
1128 static atomic_t pgdat_init_n_undone __initdata;
1129 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1131 static inline void __init pgdat_init_report_one_done(void)
1133 if (atomic_dec_and_test(&pgdat_init_n_undone))
1134 complete(&pgdat_init_all_done_comp);
1137 /* Initialise remaining memory on a node */
1138 static int __init deferred_init_memmap(void *data)
1140 pg_data_t *pgdat = data;
1141 int nid = pgdat->node_id;
1142 struct mminit_pfnnid_cache nid_init_state = { };
1143 unsigned long start = jiffies;
1144 unsigned long nr_pages = 0;
1145 unsigned long walk_start, walk_end;
1148 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1149 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1151 if (first_init_pfn == ULONG_MAX) {
1152 pgdat_init_report_one_done();
1156 /* Bind memory initialisation thread to a local node if possible */
1157 if (!cpumask_empty(cpumask))
1158 set_cpus_allowed_ptr(current, cpumask);
1160 /* Sanity check boundaries */
1161 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1162 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1163 pgdat->first_deferred_pfn = ULONG_MAX;
1165 /* Only the highest zone is deferred so find it */
1166 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1167 zone = pgdat->node_zones + zid;
1168 if (first_init_pfn < zone_end_pfn(zone))
1172 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1173 unsigned long pfn, end_pfn;
1174 struct page *page = NULL;
1175 struct page *free_base_page = NULL;
1176 unsigned long free_base_pfn = 0;
1179 end_pfn = min(walk_end, zone_end_pfn(zone));
1180 pfn = first_init_pfn;
1181 if (pfn < walk_start)
1183 if (pfn < zone->zone_start_pfn)
1184 pfn = zone->zone_start_pfn;
1186 for (; pfn < end_pfn; pfn++) {
1187 if (!pfn_valid_within(pfn))
1191 * Ensure pfn_valid is checked every
1192 * MAX_ORDER_NR_PAGES for memory holes
1194 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1195 if (!pfn_valid(pfn)) {
1201 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1206 /* Minimise pfn page lookups and scheduler checks */
1207 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1210 nr_pages += nr_to_free;
1211 deferred_free_range(free_base_page,
1212 free_base_pfn, nr_to_free);
1213 free_base_page = NULL;
1214 free_base_pfn = nr_to_free = 0;
1216 page = pfn_to_page(pfn);
1221 VM_BUG_ON(page_zone(page) != zone);
1225 __init_single_page(page, pfn, zid, nid);
1226 if (!free_base_page) {
1227 free_base_page = page;
1228 free_base_pfn = pfn;
1233 /* Where possible, batch up pages for a single free */
1236 /* Free the current block of pages to allocator */
1237 nr_pages += nr_to_free;
1238 deferred_free_range(free_base_page, free_base_pfn,
1240 free_base_page = NULL;
1241 free_base_pfn = nr_to_free = 0;
1244 first_init_pfn = max(end_pfn, first_init_pfn);
1247 /* Sanity check that the next zone really is unpopulated */
1248 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1250 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1251 jiffies_to_msecs(jiffies - start));
1253 pgdat_init_report_one_done();
1257 void __init page_alloc_init_late(void)
1261 /* There will be num_node_state(N_MEMORY) threads */
1262 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1263 for_each_node_state(nid, N_MEMORY) {
1264 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1267 /* Block until all are initialised */
1268 wait_for_completion(&pgdat_init_all_done_comp);
1270 /* Reinit limits that are based on free pages after the kernel is up */
1271 files_maxfiles_init();
1273 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1276 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1277 void __init init_cma_reserved_pageblock(struct page *page)
1279 unsigned i = pageblock_nr_pages;
1280 struct page *p = page;
1283 __ClearPageReserved(p);
1284 set_page_count(p, 0);
1287 set_pageblock_migratetype(page, MIGRATE_CMA);
1289 if (pageblock_order >= MAX_ORDER) {
1290 i = pageblock_nr_pages;
1293 set_page_refcounted(p);
1294 __free_pages(p, MAX_ORDER - 1);
1295 p += MAX_ORDER_NR_PAGES;
1296 } while (i -= MAX_ORDER_NR_PAGES);
1298 set_page_refcounted(page);
1299 __free_pages(page, pageblock_order);
1302 adjust_managed_page_count(page, pageblock_nr_pages);
1307 * The order of subdivision here is critical for the IO subsystem.
1308 * Please do not alter this order without good reasons and regression
1309 * testing. Specifically, as large blocks of memory are subdivided,
1310 * the order in which smaller blocks are delivered depends on the order
1311 * they're subdivided in this function. This is the primary factor
1312 * influencing the order in which pages are delivered to the IO
1313 * subsystem according to empirical testing, and this is also justified
1314 * by considering the behavior of a buddy system containing a single
1315 * large block of memory acted on by a series of small allocations.
1316 * This behavior is a critical factor in sglist merging's success.
1320 static inline void expand(struct zone *zone, struct page *page,
1321 int low, int high, struct free_area *area,
1324 unsigned long size = 1 << high;
1326 while (high > low) {
1330 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1332 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1333 debug_guardpage_enabled() &&
1334 high < debug_guardpage_minorder()) {
1336 * Mark as guard pages (or page), that will allow to
1337 * merge back to allocator when buddy will be freed.
1338 * Corresponding page table entries will not be touched,
1339 * pages will stay not present in virtual address space
1341 set_page_guard(zone, &page[size], high, migratetype);
1344 list_add(&page[size].lru, &area->free_list[migratetype]);
1346 set_page_order(&page[size], high);
1351 * This page is about to be returned from the page allocator
1353 static inline int check_new_page(struct page *page)
1355 const char *bad_reason = NULL;
1356 unsigned long bad_flags = 0;
1358 if (unlikely(atomic_read(&page->_mapcount) != -1))
1359 bad_reason = "nonzero mapcount";
1360 if (unlikely(page->mapping != NULL))
1361 bad_reason = "non-NULL mapping";
1362 if (unlikely(atomic_read(&page->_count) != 0))
1363 bad_reason = "nonzero _count";
1364 if (unlikely(page->flags & __PG_HWPOISON)) {
1365 bad_reason = "HWPoisoned (hardware-corrupted)";
1366 bad_flags = __PG_HWPOISON;
1368 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1369 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1370 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1373 if (unlikely(page->mem_cgroup))
1374 bad_reason = "page still charged to cgroup";
1376 if (unlikely(bad_reason)) {
1377 bad_page(page, bad_reason, bad_flags);
1383 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1388 for (i = 0; i < (1 << order); i++) {
1389 struct page *p = page + i;
1390 if (unlikely(check_new_page(p)))
1394 set_page_private(page, 0);
1395 set_page_refcounted(page);
1397 arch_alloc_page(page, order);
1398 kernel_map_pages(page, 1 << order, 1);
1399 kasan_alloc_pages(page, order);
1401 if (gfp_flags & __GFP_ZERO)
1402 for (i = 0; i < (1 << order); i++)
1403 clear_highpage(page + i);
1405 if (order && (gfp_flags & __GFP_COMP))
1406 prep_compound_page(page, order);
1408 set_page_owner(page, order, gfp_flags);
1411 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1412 * allocate the page. The expectation is that the caller is taking
1413 * steps that will free more memory. The caller should avoid the page
1414 * being used for !PFMEMALLOC purposes.
1416 if (alloc_flags & ALLOC_NO_WATERMARKS)
1417 set_page_pfmemalloc(page);
1419 clear_page_pfmemalloc(page);
1425 * Go through the free lists for the given migratetype and remove
1426 * the smallest available page from the freelists
1429 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1432 unsigned int current_order;
1433 struct free_area *area;
1436 /* Find a page of the appropriate size in the preferred list */
1437 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1438 area = &(zone->free_area[current_order]);
1439 page = list_first_entry_or_null(&area->free_list[migratetype],
1443 list_del(&page->lru);
1444 rmv_page_order(page);
1446 expand(zone, page, order, current_order, area, migratetype);
1447 set_pcppage_migratetype(page, migratetype);
1456 * This array describes the order lists are fallen back to when
1457 * the free lists for the desirable migrate type are depleted
1459 static int fallbacks[MIGRATE_TYPES][4] = {
1460 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1461 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1462 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1464 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1466 #ifdef CONFIG_MEMORY_ISOLATION
1467 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1472 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1475 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1478 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1479 unsigned int order) { return NULL; }
1483 * Move the free pages in a range to the free lists of the requested type.
1484 * Note that start_page and end_pages are not aligned on a pageblock
1485 * boundary. If alignment is required, use move_freepages_block()
1487 int move_freepages(struct zone *zone,
1488 struct page *start_page, struct page *end_page,
1493 int pages_moved = 0;
1495 #ifndef CONFIG_HOLES_IN_ZONE
1497 * page_zone is not safe to call in this context when
1498 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1499 * anyway as we check zone boundaries in move_freepages_block().
1500 * Remove at a later date when no bug reports exist related to
1501 * grouping pages by mobility
1503 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1506 for (page = start_page; page <= end_page;) {
1507 /* Make sure we are not inadvertently changing nodes */
1508 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1510 if (!pfn_valid_within(page_to_pfn(page))) {
1515 if (!PageBuddy(page)) {
1520 order = page_order(page);
1521 list_move(&page->lru,
1522 &zone->free_area[order].free_list[migratetype]);
1524 pages_moved += 1 << order;
1530 int move_freepages_block(struct zone *zone, struct page *page,
1533 unsigned long start_pfn, end_pfn;
1534 struct page *start_page, *end_page;
1536 start_pfn = page_to_pfn(page);
1537 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1538 start_page = pfn_to_page(start_pfn);
1539 end_page = start_page + pageblock_nr_pages - 1;
1540 end_pfn = start_pfn + pageblock_nr_pages - 1;
1542 /* Do not cross zone boundaries */
1543 if (!zone_spans_pfn(zone, start_pfn))
1545 if (!zone_spans_pfn(zone, end_pfn))
1548 return move_freepages(zone, start_page, end_page, migratetype);
1551 static void change_pageblock_range(struct page *pageblock_page,
1552 int start_order, int migratetype)
1554 int nr_pageblocks = 1 << (start_order - pageblock_order);
1556 while (nr_pageblocks--) {
1557 set_pageblock_migratetype(pageblock_page, migratetype);
1558 pageblock_page += pageblock_nr_pages;
1563 * When we are falling back to another migratetype during allocation, try to
1564 * steal extra free pages from the same pageblocks to satisfy further
1565 * allocations, instead of polluting multiple pageblocks.
1567 * If we are stealing a relatively large buddy page, it is likely there will
1568 * be more free pages in the pageblock, so try to steal them all. For
1569 * reclaimable and unmovable allocations, we steal regardless of page size,
1570 * as fragmentation caused by those allocations polluting movable pageblocks
1571 * is worse than movable allocations stealing from unmovable and reclaimable
1574 static bool can_steal_fallback(unsigned int order, int start_mt)
1577 * Leaving this order check is intended, although there is
1578 * relaxed order check in next check. The reason is that
1579 * we can actually steal whole pageblock if this condition met,
1580 * but, below check doesn't guarantee it and that is just heuristic
1581 * so could be changed anytime.
1583 if (order >= pageblock_order)
1586 if (order >= pageblock_order / 2 ||
1587 start_mt == MIGRATE_RECLAIMABLE ||
1588 start_mt == MIGRATE_UNMOVABLE ||
1589 page_group_by_mobility_disabled)
1596 * This function implements actual steal behaviour. If order is large enough,
1597 * we can steal whole pageblock. If not, we first move freepages in this
1598 * pageblock and check whether half of pages are moved or not. If half of
1599 * pages are moved, we can change migratetype of pageblock and permanently
1600 * use it's pages as requested migratetype in the future.
1602 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1605 unsigned int current_order = page_order(page);
1608 /* Take ownership for orders >= pageblock_order */
1609 if (current_order >= pageblock_order) {
1610 change_pageblock_range(page, current_order, start_type);
1614 pages = move_freepages_block(zone, page, start_type);
1616 /* Claim the whole block if over half of it is free */
1617 if (pages >= (1 << (pageblock_order-1)) ||
1618 page_group_by_mobility_disabled)
1619 set_pageblock_migratetype(page, start_type);
1623 * Check whether there is a suitable fallback freepage with requested order.
1624 * If only_stealable is true, this function returns fallback_mt only if
1625 * we can steal other freepages all together. This would help to reduce
1626 * fragmentation due to mixed migratetype pages in one pageblock.
1628 int find_suitable_fallback(struct free_area *area, unsigned int order,
1629 int migratetype, bool only_stealable, bool *can_steal)
1634 if (area->nr_free == 0)
1639 fallback_mt = fallbacks[migratetype][i];
1640 if (fallback_mt == MIGRATE_TYPES)
1643 if (list_empty(&area->free_list[fallback_mt]))
1646 if (can_steal_fallback(order, migratetype))
1649 if (!only_stealable)
1660 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1661 * there are no empty page blocks that contain a page with a suitable order
1663 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1664 unsigned int alloc_order)
1667 unsigned long max_managed, flags;
1670 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1671 * Check is race-prone but harmless.
1673 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1674 if (zone->nr_reserved_highatomic >= max_managed)
1677 spin_lock_irqsave(&zone->lock, flags);
1679 /* Recheck the nr_reserved_highatomic limit under the lock */
1680 if (zone->nr_reserved_highatomic >= max_managed)
1684 mt = get_pageblock_migratetype(page);
1685 if (mt != MIGRATE_HIGHATOMIC &&
1686 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1687 zone->nr_reserved_highatomic += pageblock_nr_pages;
1688 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1689 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1693 spin_unlock_irqrestore(&zone->lock, flags);
1697 * Used when an allocation is about to fail under memory pressure. This
1698 * potentially hurts the reliability of high-order allocations when under
1699 * intense memory pressure but failed atomic allocations should be easier
1700 * to recover from than an OOM.
1702 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1704 struct zonelist *zonelist = ac->zonelist;
1705 unsigned long flags;
1711 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1713 /* Preserve at least one pageblock */
1714 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1717 spin_lock_irqsave(&zone->lock, flags);
1718 for (order = 0; order < MAX_ORDER; order++) {
1719 struct free_area *area = &(zone->free_area[order]);
1721 page = list_first_entry_or_null(
1722 &area->free_list[MIGRATE_HIGHATOMIC],
1728 * It should never happen but changes to locking could
1729 * inadvertently allow a per-cpu drain to add pages
1730 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1731 * and watch for underflows.
1733 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
1734 zone->nr_reserved_highatomic);
1737 * Convert to ac->migratetype and avoid the normal
1738 * pageblock stealing heuristics. Minimally, the caller
1739 * is doing the work and needs the pages. More
1740 * importantly, if the block was always converted to
1741 * MIGRATE_UNMOVABLE or another type then the number
1742 * of pageblocks that cannot be completely freed
1745 set_pageblock_migratetype(page, ac->migratetype);
1746 move_freepages_block(zone, page, ac->migratetype);
1747 spin_unlock_irqrestore(&zone->lock, flags);
1750 spin_unlock_irqrestore(&zone->lock, flags);
1754 /* Remove an element from the buddy allocator from the fallback list */
1755 static inline struct page *
1756 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1758 struct free_area *area;
1759 unsigned int current_order;
1764 /* Find the largest possible block of pages in the other list */
1765 for (current_order = MAX_ORDER-1;
1766 current_order >= order && current_order <= MAX_ORDER-1;
1768 area = &(zone->free_area[current_order]);
1769 fallback_mt = find_suitable_fallback(area, current_order,
1770 start_migratetype, false, &can_steal);
1771 if (fallback_mt == -1)
1774 page = list_first_entry(&area->free_list[fallback_mt],
1777 steal_suitable_fallback(zone, page, start_migratetype);
1779 /* Remove the page from the freelists */
1781 list_del(&page->lru);
1782 rmv_page_order(page);
1784 expand(zone, page, order, current_order, area,
1787 * The pcppage_migratetype may differ from pageblock's
1788 * migratetype depending on the decisions in
1789 * find_suitable_fallback(). This is OK as long as it does not
1790 * differ for MIGRATE_CMA pageblocks. Those can be used as
1791 * fallback only via special __rmqueue_cma_fallback() function
1793 set_pcppage_migratetype(page, start_migratetype);
1795 trace_mm_page_alloc_extfrag(page, order, current_order,
1796 start_migratetype, fallback_mt);
1805 * Do the hard work of removing an element from the buddy allocator.
1806 * Call me with the zone->lock already held.
1808 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1813 page = __rmqueue_smallest(zone, order, migratetype);
1814 if (unlikely(!page)) {
1815 if (migratetype == MIGRATE_MOVABLE)
1816 page = __rmqueue_cma_fallback(zone, order);
1819 page = __rmqueue_fallback(zone, order, migratetype);
1822 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1827 * Obtain a specified number of elements from the buddy allocator, all under
1828 * a single hold of the lock, for efficiency. Add them to the supplied list.
1829 * Returns the number of new pages which were placed at *list.
1831 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1832 unsigned long count, struct list_head *list,
1833 int migratetype, bool cold)
1837 spin_lock(&zone->lock);
1838 for (i = 0; i < count; ++i) {
1839 struct page *page = __rmqueue(zone, order, migratetype);
1840 if (unlikely(page == NULL))
1844 * Split buddy pages returned by expand() are received here
1845 * in physical page order. The page is added to the callers and
1846 * list and the list head then moves forward. From the callers
1847 * perspective, the linked list is ordered by page number in
1848 * some conditions. This is useful for IO devices that can
1849 * merge IO requests if the physical pages are ordered
1853 list_add(&page->lru, list);
1855 list_add_tail(&page->lru, list);
1857 if (is_migrate_cma(get_pcppage_migratetype(page)))
1858 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1861 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1862 spin_unlock(&zone->lock);
1868 * Called from the vmstat counter updater to drain pagesets of this
1869 * currently executing processor on remote nodes after they have
1872 * Note that this function must be called with the thread pinned to
1873 * a single processor.
1875 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1877 unsigned long flags;
1878 int to_drain, batch;
1880 local_irq_save(flags);
1881 batch = READ_ONCE(pcp->batch);
1882 to_drain = min(pcp->count, batch);
1884 free_pcppages_bulk(zone, to_drain, pcp);
1885 pcp->count -= to_drain;
1887 local_irq_restore(flags);
1892 * Drain pcplists of the indicated processor and zone.
1894 * The processor must either be the current processor and the
1895 * thread pinned to the current processor or a processor that
1898 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1900 unsigned long flags;
1901 struct per_cpu_pageset *pset;
1902 struct per_cpu_pages *pcp;
1904 local_irq_save(flags);
1905 pset = per_cpu_ptr(zone->pageset, cpu);
1909 free_pcppages_bulk(zone, pcp->count, pcp);
1912 local_irq_restore(flags);
1916 * Drain pcplists of all zones on the indicated processor.
1918 * The processor must either be the current processor and the
1919 * thread pinned to the current processor or a processor that
1922 static void drain_pages(unsigned int cpu)
1926 for_each_populated_zone(zone) {
1927 drain_pages_zone(cpu, zone);
1932 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1934 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1935 * the single zone's pages.
1937 void drain_local_pages(struct zone *zone)
1939 int cpu = smp_processor_id();
1942 drain_pages_zone(cpu, zone);
1948 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1950 * When zone parameter is non-NULL, spill just the single zone's pages.
1952 * Note that this code is protected against sending an IPI to an offline
1953 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1954 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1955 * nothing keeps CPUs from showing up after we populated the cpumask and
1956 * before the call to on_each_cpu_mask().
1958 void drain_all_pages(struct zone *zone)
1963 * Allocate in the BSS so we wont require allocation in
1964 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1966 static cpumask_t cpus_with_pcps;
1969 * We don't care about racing with CPU hotplug event
1970 * as offline notification will cause the notified
1971 * cpu to drain that CPU pcps and on_each_cpu_mask
1972 * disables preemption as part of its processing
1974 for_each_online_cpu(cpu) {
1975 struct per_cpu_pageset *pcp;
1977 bool has_pcps = false;
1980 pcp = per_cpu_ptr(zone->pageset, cpu);
1984 for_each_populated_zone(z) {
1985 pcp = per_cpu_ptr(z->pageset, cpu);
1986 if (pcp->pcp.count) {
1994 cpumask_set_cpu(cpu, &cpus_with_pcps);
1996 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1998 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2002 #ifdef CONFIG_HIBERNATION
2004 void mark_free_pages(struct zone *zone)
2006 unsigned long pfn, max_zone_pfn;
2007 unsigned long flags;
2008 unsigned int order, t;
2011 if (zone_is_empty(zone))
2014 spin_lock_irqsave(&zone->lock, flags);
2016 max_zone_pfn = zone_end_pfn(zone);
2017 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2018 if (pfn_valid(pfn)) {
2019 page = pfn_to_page(pfn);
2020 if (!swsusp_page_is_forbidden(page))
2021 swsusp_unset_page_free(page);
2024 for_each_migratetype_order(order, t) {
2025 list_for_each_entry(page,
2026 &zone->free_area[order].free_list[t], lru) {
2029 pfn = page_to_pfn(page);
2030 for (i = 0; i < (1UL << order); i++)
2031 swsusp_set_page_free(pfn_to_page(pfn + i));
2034 spin_unlock_irqrestore(&zone->lock, flags);
2036 #endif /* CONFIG_PM */
2039 * Free a 0-order page
2040 * cold == true ? free a cold page : free a hot page
2042 void free_hot_cold_page(struct page *page, bool cold)
2044 struct zone *zone = page_zone(page);
2045 struct per_cpu_pages *pcp;
2046 unsigned long flags;
2047 unsigned long pfn = page_to_pfn(page);
2050 if (!free_pages_prepare(page, 0))
2053 migratetype = get_pfnblock_migratetype(page, pfn);
2054 set_pcppage_migratetype(page, migratetype);
2055 local_irq_save(flags);
2056 __count_vm_event(PGFREE);
2059 * We only track unmovable, reclaimable and movable on pcp lists.
2060 * Free ISOLATE pages back to the allocator because they are being
2061 * offlined but treat RESERVE as movable pages so we can get those
2062 * areas back if necessary. Otherwise, we may have to free
2063 * excessively into the page allocator
2065 if (migratetype >= MIGRATE_PCPTYPES) {
2066 if (unlikely(is_migrate_isolate(migratetype))) {
2067 free_one_page(zone, page, pfn, 0, migratetype);
2070 migratetype = MIGRATE_MOVABLE;
2073 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2075 list_add(&page->lru, &pcp->lists[migratetype]);
2077 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2079 if (pcp->count >= pcp->high) {
2080 unsigned long batch = READ_ONCE(pcp->batch);
2081 free_pcppages_bulk(zone, batch, pcp);
2082 pcp->count -= batch;
2086 local_irq_restore(flags);
2090 * Free a list of 0-order pages
2092 void free_hot_cold_page_list(struct list_head *list, bool cold)
2094 struct page *page, *next;
2096 list_for_each_entry_safe(page, next, list, lru) {
2097 trace_mm_page_free_batched(page, cold);
2098 free_hot_cold_page(page, cold);
2103 * split_page takes a non-compound higher-order page, and splits it into
2104 * n (1<<order) sub-pages: page[0..n]
2105 * Each sub-page must be freed individually.
2107 * Note: this is probably too low level an operation for use in drivers.
2108 * Please consult with lkml before using this in your driver.
2110 void split_page(struct page *page, unsigned int order)
2115 VM_BUG_ON_PAGE(PageCompound(page), page);
2116 VM_BUG_ON_PAGE(!page_count(page), page);
2118 #ifdef CONFIG_KMEMCHECK
2120 * Split shadow pages too, because free(page[0]) would
2121 * otherwise free the whole shadow.
2123 if (kmemcheck_page_is_tracked(page))
2124 split_page(virt_to_page(page[0].shadow), order);
2127 gfp_mask = get_page_owner_gfp(page);
2128 set_page_owner(page, 0, gfp_mask);
2129 for (i = 1; i < (1 << order); i++) {
2130 set_page_refcounted(page + i);
2131 set_page_owner(page + i, 0, gfp_mask);
2134 EXPORT_SYMBOL_GPL(split_page);
2136 int __isolate_free_page(struct page *page, unsigned int order)
2138 unsigned long watermark;
2142 BUG_ON(!PageBuddy(page));
2144 zone = page_zone(page);
2145 mt = get_pageblock_migratetype(page);
2147 if (!is_migrate_isolate(mt)) {
2148 /* Obey watermarks as if the page was being allocated */
2149 watermark = low_wmark_pages(zone) + (1 << order);
2150 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2153 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2156 /* Remove page from free list */
2157 list_del(&page->lru);
2158 zone->free_area[order].nr_free--;
2159 rmv_page_order(page);
2161 set_page_owner(page, order, __GFP_MOVABLE);
2163 /* Set the pageblock if the isolated page is at least a pageblock */
2164 if (order >= pageblock_order - 1) {
2165 struct page *endpage = page + (1 << order) - 1;
2166 for (; page < endpage; page += pageblock_nr_pages) {
2167 int mt = get_pageblock_migratetype(page);
2168 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2169 set_pageblock_migratetype(page,
2175 return 1UL << order;
2179 * Similar to split_page except the page is already free. As this is only
2180 * being used for migration, the migratetype of the block also changes.
2181 * As this is called with interrupts disabled, the caller is responsible
2182 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2185 * Note: this is probably too low level an operation for use in drivers.
2186 * Please consult with lkml before using this in your driver.
2188 int split_free_page(struct page *page)
2193 order = page_order(page);
2195 nr_pages = __isolate_free_page(page, order);
2199 /* Split into individual pages */
2200 set_page_refcounted(page);
2201 split_page(page, order);
2206 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2209 struct page *buffered_rmqueue(struct zone *preferred_zone,
2210 struct zone *zone, unsigned int order,
2211 gfp_t gfp_flags, int alloc_flags, int migratetype)
2213 unsigned long flags;
2215 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2217 if (likely(order == 0)) {
2218 struct per_cpu_pages *pcp;
2219 struct list_head *list;
2221 local_irq_save(flags);
2222 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2223 list = &pcp->lists[migratetype];
2224 if (list_empty(list)) {
2225 pcp->count += rmqueue_bulk(zone, 0,
2228 if (unlikely(list_empty(list)))
2233 page = list_last_entry(list, struct page, lru);
2235 page = list_first_entry(list, struct page, lru);
2237 list_del(&page->lru);
2240 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2242 * __GFP_NOFAIL is not to be used in new code.
2244 * All __GFP_NOFAIL callers should be fixed so that they
2245 * properly detect and handle allocation failures.
2247 * We most definitely don't want callers attempting to
2248 * allocate greater than order-1 page units with
2251 WARN_ON_ONCE(order > 1);
2253 spin_lock_irqsave(&zone->lock, flags);
2256 if (alloc_flags & ALLOC_HARDER) {
2257 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2259 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2262 page = __rmqueue(zone, order, migratetype);
2263 spin_unlock(&zone->lock);
2266 __mod_zone_freepage_state(zone, -(1 << order),
2267 get_pcppage_migratetype(page));
2270 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2271 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2272 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2273 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2275 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2276 zone_statistics(preferred_zone, zone, gfp_flags);
2277 local_irq_restore(flags);
2279 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2283 local_irq_restore(flags);
2287 #ifdef CONFIG_FAIL_PAGE_ALLOC
2290 struct fault_attr attr;
2292 bool ignore_gfp_highmem;
2293 bool ignore_gfp_reclaim;
2295 } fail_page_alloc = {
2296 .attr = FAULT_ATTR_INITIALIZER,
2297 .ignore_gfp_reclaim = true,
2298 .ignore_gfp_highmem = true,
2302 static int __init setup_fail_page_alloc(char *str)
2304 return setup_fault_attr(&fail_page_alloc.attr, str);
2306 __setup("fail_page_alloc=", setup_fail_page_alloc);
2308 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2310 if (order < fail_page_alloc.min_order)
2312 if (gfp_mask & __GFP_NOFAIL)
2314 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2316 if (fail_page_alloc.ignore_gfp_reclaim &&
2317 (gfp_mask & __GFP_DIRECT_RECLAIM))
2320 return should_fail(&fail_page_alloc.attr, 1 << order);
2323 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2325 static int __init fail_page_alloc_debugfs(void)
2327 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2330 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2331 &fail_page_alloc.attr);
2333 return PTR_ERR(dir);
2335 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2336 &fail_page_alloc.ignore_gfp_reclaim))
2338 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2339 &fail_page_alloc.ignore_gfp_highmem))
2341 if (!debugfs_create_u32("min-order", mode, dir,
2342 &fail_page_alloc.min_order))
2347 debugfs_remove_recursive(dir);
2352 late_initcall(fail_page_alloc_debugfs);
2354 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2356 #else /* CONFIG_FAIL_PAGE_ALLOC */
2358 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2363 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2366 * Return true if free base pages are above 'mark'. For high-order checks it
2367 * will return true of the order-0 watermark is reached and there is at least
2368 * one free page of a suitable size. Checking now avoids taking the zone lock
2369 * to check in the allocation paths if no pages are free.
2371 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2372 unsigned long mark, int classzone_idx, int alloc_flags,
2377 const int alloc_harder = (alloc_flags & ALLOC_HARDER);
2379 /* free_pages may go negative - that's OK */
2380 free_pages -= (1 << order) - 1;
2382 if (alloc_flags & ALLOC_HIGH)
2386 * If the caller does not have rights to ALLOC_HARDER then subtract
2387 * the high-atomic reserves. This will over-estimate the size of the
2388 * atomic reserve but it avoids a search.
2390 if (likely(!alloc_harder))
2391 free_pages -= z->nr_reserved_highatomic;
2396 /* If allocation can't use CMA areas don't use free CMA pages */
2397 if (!(alloc_flags & ALLOC_CMA))
2398 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2402 * Check watermarks for an order-0 allocation request. If these
2403 * are not met, then a high-order request also cannot go ahead
2404 * even if a suitable page happened to be free.
2406 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2409 /* If this is an order-0 request then the watermark is fine */
2413 /* For a high-order request, check at least one suitable page is free */
2414 for (o = order; o < MAX_ORDER; o++) {
2415 struct free_area *area = &z->free_area[o];
2424 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2425 if (!list_empty(&area->free_list[mt]))
2430 if ((alloc_flags & ALLOC_CMA) &&
2431 !list_empty(&area->free_list[MIGRATE_CMA])) {
2439 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2440 int classzone_idx, int alloc_flags)
2442 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2443 zone_page_state(z, NR_FREE_PAGES));
2446 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2447 unsigned long mark, int classzone_idx)
2449 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2451 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2452 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2454 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2459 static bool zone_local(struct zone *local_zone, struct zone *zone)
2461 return local_zone->node == zone->node;
2464 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2466 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2469 #else /* CONFIG_NUMA */
2470 static bool zone_local(struct zone *local_zone, struct zone *zone)
2475 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2479 #endif /* CONFIG_NUMA */
2481 static void reset_alloc_batches(struct zone *preferred_zone)
2483 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2486 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2487 high_wmark_pages(zone) - low_wmark_pages(zone) -
2488 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2489 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2490 } while (zone++ != preferred_zone);
2494 * get_page_from_freelist goes through the zonelist trying to allocate
2497 static struct page *
2498 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2499 const struct alloc_context *ac)
2501 struct zonelist *zonelist = ac->zonelist;
2503 struct page *page = NULL;
2505 int nr_fair_skipped = 0;
2506 bool zonelist_rescan;
2509 zonelist_rescan = false;
2512 * Scan zonelist, looking for a zone with enough free.
2513 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2515 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2519 if (cpusets_enabled() &&
2520 (alloc_flags & ALLOC_CPUSET) &&
2521 !cpuset_zone_allowed(zone, gfp_mask))
2524 * Distribute pages in proportion to the individual
2525 * zone size to ensure fair page aging. The zone a
2526 * page was allocated in should have no effect on the
2527 * time the page has in memory before being reclaimed.
2529 if (alloc_flags & ALLOC_FAIR) {
2530 if (!zone_local(ac->preferred_zone, zone))
2532 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2538 * When allocating a page cache page for writing, we
2539 * want to get it from a zone that is within its dirty
2540 * limit, such that no single zone holds more than its
2541 * proportional share of globally allowed dirty pages.
2542 * The dirty limits take into account the zone's
2543 * lowmem reserves and high watermark so that kswapd
2544 * should be able to balance it without having to
2545 * write pages from its LRU list.
2547 * This may look like it could increase pressure on
2548 * lower zones by failing allocations in higher zones
2549 * before they are full. But the pages that do spill
2550 * over are limited as the lower zones are protected
2551 * by this very same mechanism. It should not become
2552 * a practical burden to them.
2554 * XXX: For now, allow allocations to potentially
2555 * exceed the per-zone dirty limit in the slowpath
2556 * (spread_dirty_pages unset) before going into reclaim,
2557 * which is important when on a NUMA setup the allowed
2558 * zones are together not big enough to reach the
2559 * global limit. The proper fix for these situations
2560 * will require awareness of zones in the
2561 * dirty-throttling and the flusher threads.
2563 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2566 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2567 if (!zone_watermark_ok(zone, order, mark,
2568 ac->classzone_idx, alloc_flags)) {
2571 /* Checked here to keep the fast path fast */
2572 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2573 if (alloc_flags & ALLOC_NO_WATERMARKS)
2576 if (zone_reclaim_mode == 0 ||
2577 !zone_allows_reclaim(ac->preferred_zone, zone))
2580 ret = zone_reclaim(zone, gfp_mask, order);
2582 case ZONE_RECLAIM_NOSCAN:
2585 case ZONE_RECLAIM_FULL:
2586 /* scanned but unreclaimable */
2589 /* did we reclaim enough */
2590 if (zone_watermark_ok(zone, order, mark,
2591 ac->classzone_idx, alloc_flags))
2599 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2600 gfp_mask, alloc_flags, ac->migratetype);
2602 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2606 * If this is a high-order atomic allocation then check
2607 * if the pageblock should be reserved for the future
2609 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2610 reserve_highatomic_pageblock(page, zone, order);
2617 * The first pass makes sure allocations are spread fairly within the
2618 * local node. However, the local node might have free pages left
2619 * after the fairness batches are exhausted, and remote zones haven't
2620 * even been considered yet. Try once more without fairness, and
2621 * include remote zones now, before entering the slowpath and waking
2622 * kswapd: prefer spilling to a remote zone over swapping locally.
2624 if (alloc_flags & ALLOC_FAIR) {
2625 alloc_flags &= ~ALLOC_FAIR;
2626 if (nr_fair_skipped) {
2627 zonelist_rescan = true;
2628 reset_alloc_batches(ac->preferred_zone);
2630 if (nr_online_nodes > 1)
2631 zonelist_rescan = true;
2634 if (zonelist_rescan)
2641 * Large machines with many possible nodes should not always dump per-node
2642 * meminfo in irq context.
2644 static inline bool should_suppress_show_mem(void)
2649 ret = in_interrupt();
2654 static DEFINE_RATELIMIT_STATE(nopage_rs,
2655 DEFAULT_RATELIMIT_INTERVAL,
2656 DEFAULT_RATELIMIT_BURST);
2658 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
2660 unsigned int filter = SHOW_MEM_FILTER_NODES;
2662 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2663 debug_guardpage_minorder() > 0)
2667 * This documents exceptions given to allocations in certain
2668 * contexts that are allowed to allocate outside current's set
2671 if (!(gfp_mask & __GFP_NOMEMALLOC))
2672 if (test_thread_flag(TIF_MEMDIE) ||
2673 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2674 filter &= ~SHOW_MEM_FILTER_NODES;
2675 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2676 filter &= ~SHOW_MEM_FILTER_NODES;
2679 struct va_format vaf;
2682 va_start(args, fmt);
2687 pr_warn("%pV", &vaf);
2692 pr_warn("%s: page allocation failure: order:%u, mode:0x%x\n",
2693 current->comm, order, gfp_mask);
2696 if (!should_suppress_show_mem())
2700 static inline struct page *
2701 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2702 const struct alloc_context *ac, unsigned long *did_some_progress)
2704 struct oom_control oc = {
2705 .zonelist = ac->zonelist,
2706 .nodemask = ac->nodemask,
2707 .gfp_mask = gfp_mask,
2712 *did_some_progress = 0;
2715 * Acquire the oom lock. If that fails, somebody else is
2716 * making progress for us.
2718 if (!mutex_trylock(&oom_lock)) {
2719 *did_some_progress = 1;
2720 schedule_timeout_uninterruptible(1);
2725 * Go through the zonelist yet one more time, keep very high watermark
2726 * here, this is only to catch a parallel oom killing, we must fail if
2727 * we're still under heavy pressure.
2729 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2730 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2734 if (!(gfp_mask & __GFP_NOFAIL)) {
2735 /* Coredumps can quickly deplete all memory reserves */
2736 if (current->flags & PF_DUMPCORE)
2738 /* The OOM killer will not help higher order allocs */
2739 if (order > PAGE_ALLOC_COSTLY_ORDER)
2741 /* The OOM killer does not needlessly kill tasks for lowmem */
2742 if (ac->high_zoneidx < ZONE_NORMAL)
2744 /* The OOM killer does not compensate for IO-less reclaim */
2745 if (!(gfp_mask & __GFP_FS)) {
2747 * XXX: Page reclaim didn't yield anything,
2748 * and the OOM killer can't be invoked, but
2749 * keep looping as per tradition.
2751 *did_some_progress = 1;
2754 if (pm_suspended_storage())
2756 /* The OOM killer may not free memory on a specific node */
2757 if (gfp_mask & __GFP_THISNODE)
2760 /* Exhausted what can be done so it's blamo time */
2761 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
2762 *did_some_progress = 1;
2764 if (gfp_mask & __GFP_NOFAIL) {
2765 page = get_page_from_freelist(gfp_mask, order,
2766 ALLOC_NO_WATERMARKS|ALLOC_CPUSET, ac);
2768 * fallback to ignore cpuset restriction if our nodes
2772 page = get_page_from_freelist(gfp_mask, order,
2773 ALLOC_NO_WATERMARKS, ac);
2777 mutex_unlock(&oom_lock);
2781 #ifdef CONFIG_COMPACTION
2782 /* Try memory compaction for high-order allocations before reclaim */
2783 static struct page *
2784 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2785 int alloc_flags, const struct alloc_context *ac,
2786 enum migrate_mode mode, int *contended_compaction,
2787 bool *deferred_compaction)
2789 unsigned long compact_result;
2795 current->flags |= PF_MEMALLOC;
2796 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2797 mode, contended_compaction);
2798 current->flags &= ~PF_MEMALLOC;
2800 switch (compact_result) {
2801 case COMPACT_DEFERRED:
2802 *deferred_compaction = true;
2804 case COMPACT_SKIPPED:
2811 * At least in one zone compaction wasn't deferred or skipped, so let's
2812 * count a compaction stall
2814 count_vm_event(COMPACTSTALL);
2816 page = get_page_from_freelist(gfp_mask, order,
2817 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2820 struct zone *zone = page_zone(page);
2822 zone->compact_blockskip_flush = false;
2823 compaction_defer_reset(zone, order, true);
2824 count_vm_event(COMPACTSUCCESS);
2829 * It's bad if compaction run occurs and fails. The most likely reason
2830 * is that pages exist, but not enough to satisfy watermarks.
2832 count_vm_event(COMPACTFAIL);
2839 static inline struct page *
2840 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2841 int alloc_flags, const struct alloc_context *ac,
2842 enum migrate_mode mode, int *contended_compaction,
2843 bool *deferred_compaction)
2847 #endif /* CONFIG_COMPACTION */
2849 /* Perform direct synchronous page reclaim */
2851 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2852 const struct alloc_context *ac)
2854 struct reclaim_state reclaim_state;
2859 /* We now go into synchronous reclaim */
2860 cpuset_memory_pressure_bump();
2861 current->flags |= PF_MEMALLOC;
2862 lockdep_set_current_reclaim_state(gfp_mask);
2863 reclaim_state.reclaimed_slab = 0;
2864 current->reclaim_state = &reclaim_state;
2866 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2869 current->reclaim_state = NULL;
2870 lockdep_clear_current_reclaim_state();
2871 current->flags &= ~PF_MEMALLOC;
2878 /* The really slow allocator path where we enter direct reclaim */
2879 static inline struct page *
2880 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2881 int alloc_flags, const struct alloc_context *ac,
2882 unsigned long *did_some_progress)
2884 struct page *page = NULL;
2885 bool drained = false;
2887 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2888 if (unlikely(!(*did_some_progress)))
2892 page = get_page_from_freelist(gfp_mask, order,
2893 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2896 * If an allocation failed after direct reclaim, it could be because
2897 * pages are pinned on the per-cpu lists or in high alloc reserves.
2898 * Shrink them them and try again
2900 if (!page && !drained) {
2901 unreserve_highatomic_pageblock(ac);
2902 drain_all_pages(NULL);
2910 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2915 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2916 ac->high_zoneidx, ac->nodemask)
2917 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2921 gfp_to_alloc_flags(gfp_t gfp_mask)
2923 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2925 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2926 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2929 * The caller may dip into page reserves a bit more if the caller
2930 * cannot run direct reclaim, or if the caller has realtime scheduling
2931 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2932 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
2934 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2936 if (gfp_mask & __GFP_ATOMIC) {
2938 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2939 * if it can't schedule.
2941 if (!(gfp_mask & __GFP_NOMEMALLOC))
2942 alloc_flags |= ALLOC_HARDER;
2944 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2945 * comment for __cpuset_node_allowed().
2947 alloc_flags &= ~ALLOC_CPUSET;
2948 } else if (unlikely(rt_task(current)) && !in_interrupt())
2949 alloc_flags |= ALLOC_HARDER;
2951 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2952 if (gfp_mask & __GFP_MEMALLOC)
2953 alloc_flags |= ALLOC_NO_WATERMARKS;
2954 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2955 alloc_flags |= ALLOC_NO_WATERMARKS;
2956 else if (!in_interrupt() &&
2957 ((current->flags & PF_MEMALLOC) ||
2958 unlikely(test_thread_flag(TIF_MEMDIE))))
2959 alloc_flags |= ALLOC_NO_WATERMARKS;
2962 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2963 alloc_flags |= ALLOC_CMA;
2968 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2970 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2973 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
2975 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
2978 static inline struct page *
2979 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2980 struct alloc_context *ac)
2982 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
2983 struct page *page = NULL;
2985 unsigned long pages_reclaimed = 0;
2986 unsigned long did_some_progress;
2987 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2988 bool deferred_compaction = false;
2989 int contended_compaction = COMPACT_CONTENDED_NONE;
2992 * In the slowpath, we sanity check order to avoid ever trying to
2993 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2994 * be using allocators in order of preference for an area that is
2997 if (order >= MAX_ORDER) {
2998 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3003 * We also sanity check to catch abuse of atomic reserves being used by
3004 * callers that are not in atomic context.
3006 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3007 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3008 gfp_mask &= ~__GFP_ATOMIC;
3011 * If this allocation cannot block and it is for a specific node, then
3012 * fail early. There's no need to wakeup kswapd or retry for a
3013 * speculative node-specific allocation.
3015 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !can_direct_reclaim)
3019 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3020 wake_all_kswapds(order, ac);
3023 * OK, we're below the kswapd watermark and have kicked background
3024 * reclaim. Now things get more complex, so set up alloc_flags according
3025 * to how we want to proceed.
3027 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3030 * Find the true preferred zone if the allocation is unconstrained by
3033 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3034 struct zoneref *preferred_zoneref;
3035 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3036 ac->high_zoneidx, NULL, &ac->preferred_zone);
3037 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3040 /* This is the last chance, in general, before the goto nopage. */
3041 page = get_page_from_freelist(gfp_mask, order,
3042 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3046 /* Allocate without watermarks if the context allows */
3047 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3049 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3050 * the allocation is high priority and these type of
3051 * allocations are system rather than user orientated
3053 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3054 page = get_page_from_freelist(gfp_mask, order,
3055 ALLOC_NO_WATERMARKS, ac);
3060 /* Caller is not willing to reclaim, we can't balance anything */
3061 if (!can_direct_reclaim) {
3063 * All existing users of the __GFP_NOFAIL are blockable, so warn
3064 * of any new users that actually allow this type of allocation
3067 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3071 /* Avoid recursion of direct reclaim */
3072 if (current->flags & PF_MEMALLOC) {
3074 * __GFP_NOFAIL request from this context is rather bizarre
3075 * because we cannot reclaim anything and only can loop waiting
3076 * for somebody to do a work for us.
3078 if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3085 /* Avoid allocations with no watermarks from looping endlessly */
3086 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3090 * Try direct compaction. The first pass is asynchronous. Subsequent
3091 * attempts after direct reclaim are synchronous
3093 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3095 &contended_compaction,
3096 &deferred_compaction);
3100 /* Checks for THP-specific high-order allocations */
3101 if (is_thp_gfp_mask(gfp_mask)) {
3103 * If compaction is deferred for high-order allocations, it is
3104 * because sync compaction recently failed. If this is the case
3105 * and the caller requested a THP allocation, we do not want
3106 * to heavily disrupt the system, so we fail the allocation
3107 * instead of entering direct reclaim.
3109 if (deferred_compaction)
3113 * In all zones where compaction was attempted (and not
3114 * deferred or skipped), lock contention has been detected.
3115 * For THP allocation we do not want to disrupt the others
3116 * so we fallback to base pages instead.
3118 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3122 * If compaction was aborted due to need_resched(), we do not
3123 * want to further increase allocation latency, unless it is
3124 * khugepaged trying to collapse.
3126 if (contended_compaction == COMPACT_CONTENDED_SCHED
3127 && !(current->flags & PF_KTHREAD))
3132 * It can become very expensive to allocate transparent hugepages at
3133 * fault, so use asynchronous memory compaction for THP unless it is
3134 * khugepaged trying to collapse.
3136 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3137 migration_mode = MIGRATE_SYNC_LIGHT;
3139 /* Try direct reclaim and then allocating */
3140 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3141 &did_some_progress);
3145 /* Do not loop if specifically requested */
3146 if (gfp_mask & __GFP_NORETRY)
3149 /* Keep reclaiming pages as long as there is reasonable progress */
3150 pages_reclaimed += did_some_progress;
3151 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3152 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3153 /* Wait for some write requests to complete then retry */
3154 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3158 /* Reclaim has failed us, start killing things */
3159 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3163 /* Retry as long as the OOM killer is making progress */
3164 if (did_some_progress)
3169 * High-order allocations do not necessarily loop after
3170 * direct reclaim and reclaim/compaction depends on compaction
3171 * being called after reclaim so call directly if necessary
3173 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3175 &contended_compaction,
3176 &deferred_compaction);
3180 warn_alloc_failed(gfp_mask, order, NULL);
3186 * This is the 'heart' of the zoned buddy allocator.
3189 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3190 struct zonelist *zonelist, nodemask_t *nodemask)
3192 struct zoneref *preferred_zoneref;
3193 struct page *page = NULL;
3194 unsigned int cpuset_mems_cookie;
3195 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3196 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3197 struct alloc_context ac = {
3198 .high_zoneidx = gfp_zone(gfp_mask),
3199 .nodemask = nodemask,
3200 .migratetype = gfpflags_to_migratetype(gfp_mask),
3203 gfp_mask &= gfp_allowed_mask;
3205 lockdep_trace_alloc(gfp_mask);
3207 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3209 if (should_fail_alloc_page(gfp_mask, order))
3213 * Check the zones suitable for the gfp_mask contain at least one
3214 * valid zone. It's possible to have an empty zonelist as a result
3215 * of __GFP_THISNODE and a memoryless node
3217 if (unlikely(!zonelist->_zonerefs->zone))
3220 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3221 alloc_flags |= ALLOC_CMA;
3224 cpuset_mems_cookie = read_mems_allowed_begin();
3226 /* We set it here, as __alloc_pages_slowpath might have changed it */
3227 ac.zonelist = zonelist;
3229 /* Dirty zone balancing only done in the fast path */
3230 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3232 /* The preferred zone is used for statistics later */
3233 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3234 ac.nodemask ? : &cpuset_current_mems_allowed,
3235 &ac.preferred_zone);
3236 if (!ac.preferred_zone)
3238 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3240 /* First allocation attempt */
3241 alloc_mask = gfp_mask|__GFP_HARDWALL;
3242 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3243 if (unlikely(!page)) {
3245 * Runtime PM, block IO and its error handling path
3246 * can deadlock because I/O on the device might not
3249 alloc_mask = memalloc_noio_flags(gfp_mask);
3250 ac.spread_dirty_pages = false;
3252 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3255 if (kmemcheck_enabled && page)
3256 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3258 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3262 * When updating a task's mems_allowed, it is possible to race with
3263 * parallel threads in such a way that an allocation can fail while
3264 * the mask is being updated. If a page allocation is about to fail,
3265 * check if the cpuset changed during allocation and if so, retry.
3267 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3272 EXPORT_SYMBOL(__alloc_pages_nodemask);
3275 * Common helper functions.
3277 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3282 * __get_free_pages() returns a 32-bit address, which cannot represent
3285 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3287 page = alloc_pages(gfp_mask, order);
3290 return (unsigned long) page_address(page);
3292 EXPORT_SYMBOL(__get_free_pages);
3294 unsigned long get_zeroed_page(gfp_t gfp_mask)
3296 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3298 EXPORT_SYMBOL(get_zeroed_page);
3300 void __free_pages(struct page *page, unsigned int order)
3302 if (put_page_testzero(page)) {
3304 free_hot_cold_page(page, false);
3306 __free_pages_ok(page, order);
3310 EXPORT_SYMBOL(__free_pages);
3312 void free_pages(unsigned long addr, unsigned int order)
3315 VM_BUG_ON(!virt_addr_valid((void *)addr));
3316 __free_pages(virt_to_page((void *)addr), order);
3320 EXPORT_SYMBOL(free_pages);
3324 * An arbitrary-length arbitrary-offset area of memory which resides
3325 * within a 0 or higher order page. Multiple fragments within that page
3326 * are individually refcounted, in the page's reference counter.
3328 * The page_frag functions below provide a simple allocation framework for
3329 * page fragments. This is used by the network stack and network device
3330 * drivers to provide a backing region of memory for use as either an
3331 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3333 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3336 struct page *page = NULL;
3337 gfp_t gfp = gfp_mask;
3339 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3340 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3342 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3343 PAGE_FRAG_CACHE_MAX_ORDER);
3344 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3346 if (unlikely(!page))
3347 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3349 nc->va = page ? page_address(page) : NULL;
3354 void *__alloc_page_frag(struct page_frag_cache *nc,
3355 unsigned int fragsz, gfp_t gfp_mask)
3357 unsigned int size = PAGE_SIZE;
3361 if (unlikely(!nc->va)) {
3363 page = __page_frag_refill(nc, gfp_mask);
3367 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3368 /* if size can vary use size else just use PAGE_SIZE */
3371 /* Even if we own the page, we do not use atomic_set().
3372 * This would break get_page_unless_zero() users.
3374 atomic_add(size - 1, &page->_count);
3376 /* reset page count bias and offset to start of new frag */
3377 nc->pfmemalloc = page_is_pfmemalloc(page);
3378 nc->pagecnt_bias = size;
3382 offset = nc->offset - fragsz;
3383 if (unlikely(offset < 0)) {
3384 page = virt_to_page(nc->va);
3386 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3389 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3390 /* if size can vary use size else just use PAGE_SIZE */
3393 /* OK, page count is 0, we can safely set it */
3394 atomic_set(&page->_count, size);
3396 /* reset page count bias and offset to start of new frag */
3397 nc->pagecnt_bias = size;
3398 offset = size - fragsz;
3402 nc->offset = offset;
3404 return nc->va + offset;
3406 EXPORT_SYMBOL(__alloc_page_frag);
3409 * Frees a page fragment allocated out of either a compound or order 0 page.
3411 void __free_page_frag(void *addr)
3413 struct page *page = virt_to_head_page(addr);
3415 if (unlikely(put_page_testzero(page)))
3416 __free_pages_ok(page, compound_order(page));
3418 EXPORT_SYMBOL(__free_page_frag);
3421 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3422 * of the current memory cgroup if __GFP_ACCOUNT is set, other than that it is
3423 * equivalent to alloc_pages.
3425 * It should be used when the caller would like to use kmalloc, but since the
3426 * allocation is large, it has to fall back to the page allocator.
3428 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3432 page = alloc_pages(gfp_mask, order);
3433 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3434 __free_pages(page, order);
3440 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3444 page = alloc_pages_node(nid, gfp_mask, order);
3445 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3446 __free_pages(page, order);
3453 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3456 void __free_kmem_pages(struct page *page, unsigned int order)
3458 memcg_kmem_uncharge(page, order);
3459 __free_pages(page, order);
3462 void free_kmem_pages(unsigned long addr, unsigned int order)
3465 VM_BUG_ON(!virt_addr_valid((void *)addr));
3466 __free_kmem_pages(virt_to_page((void *)addr), order);
3470 static void *make_alloc_exact(unsigned long addr, unsigned int order,
3474 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3475 unsigned long used = addr + PAGE_ALIGN(size);
3477 split_page(virt_to_page((void *)addr), order);
3478 while (used < alloc_end) {
3483 return (void *)addr;
3487 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3488 * @size: the number of bytes to allocate
3489 * @gfp_mask: GFP flags for the allocation
3491 * This function is similar to alloc_pages(), except that it allocates the
3492 * minimum number of pages to satisfy the request. alloc_pages() can only
3493 * allocate memory in power-of-two pages.
3495 * This function is also limited by MAX_ORDER.
3497 * Memory allocated by this function must be released by free_pages_exact().
3499 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3501 unsigned int order = get_order(size);
3504 addr = __get_free_pages(gfp_mask, order);
3505 return make_alloc_exact(addr, order, size);
3507 EXPORT_SYMBOL(alloc_pages_exact);
3510 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3512 * @nid: the preferred node ID where memory should be allocated
3513 * @size: the number of bytes to allocate
3514 * @gfp_mask: GFP flags for the allocation
3516 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3519 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3521 unsigned int order = get_order(size);
3522 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3525 return make_alloc_exact((unsigned long)page_address(p), order, size);
3529 * free_pages_exact - release memory allocated via alloc_pages_exact()
3530 * @virt: the value returned by alloc_pages_exact.
3531 * @size: size of allocation, same value as passed to alloc_pages_exact().
3533 * Release the memory allocated by a previous call to alloc_pages_exact.
3535 void free_pages_exact(void *virt, size_t size)
3537 unsigned long addr = (unsigned long)virt;
3538 unsigned long end = addr + PAGE_ALIGN(size);
3540 while (addr < end) {
3545 EXPORT_SYMBOL(free_pages_exact);
3548 * nr_free_zone_pages - count number of pages beyond high watermark
3549 * @offset: The zone index of the highest zone
3551 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3552 * high watermark within all zones at or below a given zone index. For each
3553 * zone, the number of pages is calculated as:
3554 * managed_pages - high_pages
3556 static unsigned long nr_free_zone_pages(int offset)
3561 /* Just pick one node, since fallback list is circular */
3562 unsigned long sum = 0;
3564 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3566 for_each_zone_zonelist(zone, z, zonelist, offset) {
3567 unsigned long size = zone->managed_pages;
3568 unsigned long high = high_wmark_pages(zone);
3577 * nr_free_buffer_pages - count number of pages beyond high watermark
3579 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3580 * watermark within ZONE_DMA and ZONE_NORMAL.
3582 unsigned long nr_free_buffer_pages(void)
3584 return nr_free_zone_pages(gfp_zone(GFP_USER));
3586 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3589 * nr_free_pagecache_pages - count number of pages beyond high watermark
3591 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3592 * high watermark within all zones.
3594 unsigned long nr_free_pagecache_pages(void)
3596 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3599 static inline void show_node(struct zone *zone)
3601 if (IS_ENABLED(CONFIG_NUMA))
3602 printk("Node %d ", zone_to_nid(zone));
3605 void si_meminfo(struct sysinfo *val)
3607 val->totalram = totalram_pages;
3608 val->sharedram = global_page_state(NR_SHMEM);
3609 val->freeram = global_page_state(NR_FREE_PAGES);
3610 val->bufferram = nr_blockdev_pages();
3611 val->totalhigh = totalhigh_pages;
3612 val->freehigh = nr_free_highpages();
3613 val->mem_unit = PAGE_SIZE;
3616 EXPORT_SYMBOL(si_meminfo);
3619 void si_meminfo_node(struct sysinfo *val, int nid)
3621 int zone_type; /* needs to be signed */
3622 unsigned long managed_pages = 0;
3623 pg_data_t *pgdat = NODE_DATA(nid);
3625 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3626 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3627 val->totalram = managed_pages;
3628 val->sharedram = node_page_state(nid, NR_SHMEM);
3629 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3630 #ifdef CONFIG_HIGHMEM
3631 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3632 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3638 val->mem_unit = PAGE_SIZE;
3643 * Determine whether the node should be displayed or not, depending on whether
3644 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3646 bool skip_free_areas_node(unsigned int flags, int nid)
3649 unsigned int cpuset_mems_cookie;
3651 if (!(flags & SHOW_MEM_FILTER_NODES))
3655 cpuset_mems_cookie = read_mems_allowed_begin();
3656 ret = !node_isset(nid, cpuset_current_mems_allowed);
3657 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3662 #define K(x) ((x) << (PAGE_SHIFT-10))
3664 static void show_migration_types(unsigned char type)
3666 static const char types[MIGRATE_TYPES] = {
3667 [MIGRATE_UNMOVABLE] = 'U',
3668 [MIGRATE_MOVABLE] = 'M',
3669 [MIGRATE_RECLAIMABLE] = 'E',
3670 [MIGRATE_HIGHATOMIC] = 'H',
3672 [MIGRATE_CMA] = 'C',
3674 #ifdef CONFIG_MEMORY_ISOLATION
3675 [MIGRATE_ISOLATE] = 'I',
3678 char tmp[MIGRATE_TYPES + 1];
3682 for (i = 0; i < MIGRATE_TYPES; i++) {
3683 if (type & (1 << i))
3688 printk("(%s) ", tmp);
3692 * Show free area list (used inside shift_scroll-lock stuff)
3693 * We also calculate the percentage fragmentation. We do this by counting the
3694 * memory on each free list with the exception of the first item on the list.
3697 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3700 void show_free_areas(unsigned int filter)
3702 unsigned long free_pcp = 0;
3706 for_each_populated_zone(zone) {
3707 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3710 for_each_online_cpu(cpu)
3711 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3714 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3715 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3716 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3717 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3718 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3719 " free:%lu free_pcp:%lu free_cma:%lu\n",
3720 global_page_state(NR_ACTIVE_ANON),
3721 global_page_state(NR_INACTIVE_ANON),
3722 global_page_state(NR_ISOLATED_ANON),
3723 global_page_state(NR_ACTIVE_FILE),
3724 global_page_state(NR_INACTIVE_FILE),
3725 global_page_state(NR_ISOLATED_FILE),
3726 global_page_state(NR_UNEVICTABLE),
3727 global_page_state(NR_FILE_DIRTY),
3728 global_page_state(NR_WRITEBACK),
3729 global_page_state(NR_UNSTABLE_NFS),
3730 global_page_state(NR_SLAB_RECLAIMABLE),
3731 global_page_state(NR_SLAB_UNRECLAIMABLE),
3732 global_page_state(NR_FILE_MAPPED),
3733 global_page_state(NR_SHMEM),
3734 global_page_state(NR_PAGETABLE),
3735 global_page_state(NR_BOUNCE),
3736 global_page_state(NR_FREE_PAGES),
3738 global_page_state(NR_FREE_CMA_PAGES));
3740 for_each_populated_zone(zone) {
3743 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3747 for_each_online_cpu(cpu)
3748 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3756 " active_anon:%lukB"
3757 " inactive_anon:%lukB"
3758 " active_file:%lukB"
3759 " inactive_file:%lukB"
3760 " unevictable:%lukB"
3761 " isolated(anon):%lukB"
3762 " isolated(file):%lukB"
3770 " slab_reclaimable:%lukB"
3771 " slab_unreclaimable:%lukB"
3772 " kernel_stack:%lukB"
3779 " writeback_tmp:%lukB"
3780 " pages_scanned:%lu"
3781 " all_unreclaimable? %s"
3784 K(zone_page_state(zone, NR_FREE_PAGES)),
3785 K(min_wmark_pages(zone)),
3786 K(low_wmark_pages(zone)),
3787 K(high_wmark_pages(zone)),
3788 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3789 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3790 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3791 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3792 K(zone_page_state(zone, NR_UNEVICTABLE)),
3793 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3794 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3795 K(zone->present_pages),
3796 K(zone->managed_pages),
3797 K(zone_page_state(zone, NR_MLOCK)),
3798 K(zone_page_state(zone, NR_FILE_DIRTY)),
3799 K(zone_page_state(zone, NR_WRITEBACK)),
3800 K(zone_page_state(zone, NR_FILE_MAPPED)),
3801 K(zone_page_state(zone, NR_SHMEM)),
3802 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3803 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3804 zone_page_state(zone, NR_KERNEL_STACK) *
3806 K(zone_page_state(zone, NR_PAGETABLE)),
3807 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3808 K(zone_page_state(zone, NR_BOUNCE)),
3810 K(this_cpu_read(zone->pageset->pcp.count)),
3811 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3812 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3813 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3814 (!zone_reclaimable(zone) ? "yes" : "no")
3816 printk("lowmem_reserve[]:");
3817 for (i = 0; i < MAX_NR_ZONES; i++)
3818 printk(" %ld", zone->lowmem_reserve[i]);
3822 for_each_populated_zone(zone) {
3824 unsigned long nr[MAX_ORDER], flags, total = 0;
3825 unsigned char types[MAX_ORDER];
3827 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3830 printk("%s: ", zone->name);
3832 spin_lock_irqsave(&zone->lock, flags);
3833 for (order = 0; order < MAX_ORDER; order++) {
3834 struct free_area *area = &zone->free_area[order];
3837 nr[order] = area->nr_free;
3838 total += nr[order] << order;
3841 for (type = 0; type < MIGRATE_TYPES; type++) {
3842 if (!list_empty(&area->free_list[type]))
3843 types[order] |= 1 << type;
3846 spin_unlock_irqrestore(&zone->lock, flags);
3847 for (order = 0; order < MAX_ORDER; order++) {
3848 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3850 show_migration_types(types[order]);
3852 printk("= %lukB\n", K(total));
3855 hugetlb_show_meminfo();
3857 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3859 show_swap_cache_info();
3862 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3864 zoneref->zone = zone;
3865 zoneref->zone_idx = zone_idx(zone);
3869 * Builds allocation fallback zone lists.
3871 * Add all populated zones of a node to the zonelist.
3873 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3877 enum zone_type zone_type = MAX_NR_ZONES;
3881 zone = pgdat->node_zones + zone_type;
3882 if (populated_zone(zone)) {
3883 zoneref_set_zone(zone,
3884 &zonelist->_zonerefs[nr_zones++]);
3885 check_highest_zone(zone_type);
3887 } while (zone_type);
3895 * 0 = automatic detection of better ordering.
3896 * 1 = order by ([node] distance, -zonetype)
3897 * 2 = order by (-zonetype, [node] distance)
3899 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3900 * the same zonelist. So only NUMA can configure this param.
3902 #define ZONELIST_ORDER_DEFAULT 0
3903 #define ZONELIST_ORDER_NODE 1
3904 #define ZONELIST_ORDER_ZONE 2
3906 /* zonelist order in the kernel.
3907 * set_zonelist_order() will set this to NODE or ZONE.
3909 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3910 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3914 /* The value user specified ....changed by config */
3915 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3916 /* string for sysctl */
3917 #define NUMA_ZONELIST_ORDER_LEN 16
3918 char numa_zonelist_order[16] = "default";
3921 * interface for configure zonelist ordering.
3922 * command line option "numa_zonelist_order"
3923 * = "[dD]efault - default, automatic configuration.
3924 * = "[nN]ode - order by node locality, then by zone within node
3925 * = "[zZ]one - order by zone, then by locality within zone
3928 static int __parse_numa_zonelist_order(char *s)
3930 if (*s == 'd' || *s == 'D') {
3931 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3932 } else if (*s == 'n' || *s == 'N') {
3933 user_zonelist_order = ZONELIST_ORDER_NODE;
3934 } else if (*s == 'z' || *s == 'Z') {
3935 user_zonelist_order = ZONELIST_ORDER_ZONE;
3938 "Ignoring invalid numa_zonelist_order value: "
3945 static __init int setup_numa_zonelist_order(char *s)
3952 ret = __parse_numa_zonelist_order(s);
3954 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3958 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3961 * sysctl handler for numa_zonelist_order
3963 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3964 void __user *buffer, size_t *length,
3967 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3969 static DEFINE_MUTEX(zl_order_mutex);
3971 mutex_lock(&zl_order_mutex);
3973 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3977 strcpy(saved_string, (char *)table->data);
3979 ret = proc_dostring(table, write, buffer, length, ppos);
3983 int oldval = user_zonelist_order;
3985 ret = __parse_numa_zonelist_order((char *)table->data);
3988 * bogus value. restore saved string
3990 strncpy((char *)table->data, saved_string,
3991 NUMA_ZONELIST_ORDER_LEN);
3992 user_zonelist_order = oldval;
3993 } else if (oldval != user_zonelist_order) {
3994 mutex_lock(&zonelists_mutex);
3995 build_all_zonelists(NULL, NULL);
3996 mutex_unlock(&zonelists_mutex);
4000 mutex_unlock(&zl_order_mutex);
4005 #define MAX_NODE_LOAD (nr_online_nodes)
4006 static int node_load[MAX_NUMNODES];
4009 * find_next_best_node - find the next node that should appear in a given node's fallback list
4010 * @node: node whose fallback list we're appending
4011 * @used_node_mask: nodemask_t of already used nodes
4013 * We use a number of factors to determine which is the next node that should
4014 * appear on a given node's fallback list. The node should not have appeared
4015 * already in @node's fallback list, and it should be the next closest node
4016 * according to the distance array (which contains arbitrary distance values
4017 * from each node to each node in the system), and should also prefer nodes
4018 * with no CPUs, since presumably they'll have very little allocation pressure
4019 * on them otherwise.
4020 * It returns -1 if no node is found.
4022 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4025 int min_val = INT_MAX;
4026 int best_node = NUMA_NO_NODE;
4027 const struct cpumask *tmp = cpumask_of_node(0);
4029 /* Use the local node if we haven't already */
4030 if (!node_isset(node, *used_node_mask)) {
4031 node_set(node, *used_node_mask);
4035 for_each_node_state(n, N_MEMORY) {
4037 /* Don't want a node to appear more than once */
4038 if (node_isset(n, *used_node_mask))
4041 /* Use the distance array to find the distance */
4042 val = node_distance(node, n);
4044 /* Penalize nodes under us ("prefer the next node") */
4047 /* Give preference to headless and unused nodes */
4048 tmp = cpumask_of_node(n);
4049 if (!cpumask_empty(tmp))
4050 val += PENALTY_FOR_NODE_WITH_CPUS;
4052 /* Slight preference for less loaded node */
4053 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4054 val += node_load[n];
4056 if (val < min_val) {
4063 node_set(best_node, *used_node_mask);
4070 * Build zonelists ordered by node and zones within node.
4071 * This results in maximum locality--normal zone overflows into local
4072 * DMA zone, if any--but risks exhausting DMA zone.
4074 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4077 struct zonelist *zonelist;
4079 zonelist = &pgdat->node_zonelists[0];
4080 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4082 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4083 zonelist->_zonerefs[j].zone = NULL;
4084 zonelist->_zonerefs[j].zone_idx = 0;
4088 * Build gfp_thisnode zonelists
4090 static void build_thisnode_zonelists(pg_data_t *pgdat)
4093 struct zonelist *zonelist;
4095 zonelist = &pgdat->node_zonelists[1];
4096 j = build_zonelists_node(pgdat, zonelist, 0);
4097 zonelist->_zonerefs[j].zone = NULL;
4098 zonelist->_zonerefs[j].zone_idx = 0;
4102 * Build zonelists ordered by zone and nodes within zones.
4103 * This results in conserving DMA zone[s] until all Normal memory is
4104 * exhausted, but results in overflowing to remote node while memory
4105 * may still exist in local DMA zone.
4107 static int node_order[MAX_NUMNODES];
4109 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4112 int zone_type; /* needs to be signed */
4114 struct zonelist *zonelist;
4116 zonelist = &pgdat->node_zonelists[0];
4118 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4119 for (j = 0; j < nr_nodes; j++) {
4120 node = node_order[j];
4121 z = &NODE_DATA(node)->node_zones[zone_type];
4122 if (populated_zone(z)) {
4124 &zonelist->_zonerefs[pos++]);
4125 check_highest_zone(zone_type);
4129 zonelist->_zonerefs[pos].zone = NULL;
4130 zonelist->_zonerefs[pos].zone_idx = 0;
4133 #if defined(CONFIG_64BIT)
4135 * Devices that require DMA32/DMA are relatively rare and do not justify a
4136 * penalty to every machine in case the specialised case applies. Default
4137 * to Node-ordering on 64-bit NUMA machines
4139 static int default_zonelist_order(void)
4141 return ZONELIST_ORDER_NODE;
4145 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4146 * by the kernel. If processes running on node 0 deplete the low memory zone
4147 * then reclaim will occur more frequency increasing stalls and potentially
4148 * be easier to OOM if a large percentage of the zone is under writeback or
4149 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4150 * Hence, default to zone ordering on 32-bit.
4152 static int default_zonelist_order(void)
4154 return ZONELIST_ORDER_ZONE;
4156 #endif /* CONFIG_64BIT */
4158 static void set_zonelist_order(void)
4160 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4161 current_zonelist_order = default_zonelist_order();
4163 current_zonelist_order = user_zonelist_order;
4166 static void build_zonelists(pg_data_t *pgdat)
4169 nodemask_t used_mask;
4170 int local_node, prev_node;
4171 struct zonelist *zonelist;
4172 unsigned int order = current_zonelist_order;
4174 /* initialize zonelists */
4175 for (i = 0; i < MAX_ZONELISTS; i++) {
4176 zonelist = pgdat->node_zonelists + i;
4177 zonelist->_zonerefs[0].zone = NULL;
4178 zonelist->_zonerefs[0].zone_idx = 0;
4181 /* NUMA-aware ordering of nodes */
4182 local_node = pgdat->node_id;
4183 load = nr_online_nodes;
4184 prev_node = local_node;
4185 nodes_clear(used_mask);
4187 memset(node_order, 0, sizeof(node_order));
4190 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4192 * We don't want to pressure a particular node.
4193 * So adding penalty to the first node in same
4194 * distance group to make it round-robin.
4196 if (node_distance(local_node, node) !=
4197 node_distance(local_node, prev_node))
4198 node_load[node] = load;
4202 if (order == ZONELIST_ORDER_NODE)
4203 build_zonelists_in_node_order(pgdat, node);
4205 node_order[i++] = node; /* remember order */
4208 if (order == ZONELIST_ORDER_ZONE) {
4209 /* calculate node order -- i.e., DMA last! */
4210 build_zonelists_in_zone_order(pgdat, i);
4213 build_thisnode_zonelists(pgdat);
4216 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4218 * Return node id of node used for "local" allocations.
4219 * I.e., first node id of first zone in arg node's generic zonelist.
4220 * Used for initializing percpu 'numa_mem', which is used primarily
4221 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4223 int local_memory_node(int node)
4227 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4228 gfp_zone(GFP_KERNEL),
4235 #else /* CONFIG_NUMA */
4237 static void set_zonelist_order(void)
4239 current_zonelist_order = ZONELIST_ORDER_ZONE;
4242 static void build_zonelists(pg_data_t *pgdat)
4244 int node, local_node;
4246 struct zonelist *zonelist;
4248 local_node = pgdat->node_id;
4250 zonelist = &pgdat->node_zonelists[0];
4251 j = build_zonelists_node(pgdat, zonelist, 0);
4254 * Now we build the zonelist so that it contains the zones
4255 * of all the other nodes.
4256 * We don't want to pressure a particular node, so when
4257 * building the zones for node N, we make sure that the
4258 * zones coming right after the local ones are those from
4259 * node N+1 (modulo N)
4261 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4262 if (!node_online(node))
4264 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4266 for (node = 0; node < local_node; node++) {
4267 if (!node_online(node))
4269 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4272 zonelist->_zonerefs[j].zone = NULL;
4273 zonelist->_zonerefs[j].zone_idx = 0;
4276 #endif /* CONFIG_NUMA */
4279 * Boot pageset table. One per cpu which is going to be used for all
4280 * zones and all nodes. The parameters will be set in such a way
4281 * that an item put on a list will immediately be handed over to
4282 * the buddy list. This is safe since pageset manipulation is done
4283 * with interrupts disabled.
4285 * The boot_pagesets must be kept even after bootup is complete for
4286 * unused processors and/or zones. They do play a role for bootstrapping
4287 * hotplugged processors.
4289 * zoneinfo_show() and maybe other functions do
4290 * not check if the processor is online before following the pageset pointer.
4291 * Other parts of the kernel may not check if the zone is available.
4293 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4294 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4295 static void setup_zone_pageset(struct zone *zone);
4298 * Global mutex to protect against size modification of zonelists
4299 * as well as to serialize pageset setup for the new populated zone.
4301 DEFINE_MUTEX(zonelists_mutex);
4303 /* return values int ....just for stop_machine() */
4304 static int __build_all_zonelists(void *data)
4308 pg_data_t *self = data;
4311 memset(node_load, 0, sizeof(node_load));
4314 if (self && !node_online(self->node_id)) {
4315 build_zonelists(self);
4318 for_each_online_node(nid) {
4319 pg_data_t *pgdat = NODE_DATA(nid);
4321 build_zonelists(pgdat);
4325 * Initialize the boot_pagesets that are going to be used
4326 * for bootstrapping processors. The real pagesets for
4327 * each zone will be allocated later when the per cpu
4328 * allocator is available.
4330 * boot_pagesets are used also for bootstrapping offline
4331 * cpus if the system is already booted because the pagesets
4332 * are needed to initialize allocators on a specific cpu too.
4333 * F.e. the percpu allocator needs the page allocator which
4334 * needs the percpu allocator in order to allocate its pagesets
4335 * (a chicken-egg dilemma).
4337 for_each_possible_cpu(cpu) {
4338 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4340 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4342 * We now know the "local memory node" for each node--
4343 * i.e., the node of the first zone in the generic zonelist.
4344 * Set up numa_mem percpu variable for on-line cpus. During
4345 * boot, only the boot cpu should be on-line; we'll init the
4346 * secondary cpus' numa_mem as they come on-line. During
4347 * node/memory hotplug, we'll fixup all on-line cpus.
4349 if (cpu_online(cpu))
4350 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4357 static noinline void __init
4358 build_all_zonelists_init(void)
4360 __build_all_zonelists(NULL);
4361 mminit_verify_zonelist();
4362 cpuset_init_current_mems_allowed();
4366 * Called with zonelists_mutex held always
4367 * unless system_state == SYSTEM_BOOTING.
4369 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4370 * [we're only called with non-NULL zone through __meminit paths] and
4371 * (2) call of __init annotated helper build_all_zonelists_init
4372 * [protected by SYSTEM_BOOTING].
4374 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4376 set_zonelist_order();
4378 if (system_state == SYSTEM_BOOTING) {
4379 build_all_zonelists_init();
4381 #ifdef CONFIG_MEMORY_HOTPLUG
4383 setup_zone_pageset(zone);
4385 /* we have to stop all cpus to guarantee there is no user
4387 stop_machine(__build_all_zonelists, pgdat, NULL);
4388 /* cpuset refresh routine should be here */
4390 vm_total_pages = nr_free_pagecache_pages();
4392 * Disable grouping by mobility if the number of pages in the
4393 * system is too low to allow the mechanism to work. It would be
4394 * more accurate, but expensive to check per-zone. This check is
4395 * made on memory-hotadd so a system can start with mobility
4396 * disabled and enable it later
4398 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4399 page_group_by_mobility_disabled = 1;
4401 page_group_by_mobility_disabled = 0;
4403 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4404 "Total pages: %ld\n",
4406 zonelist_order_name[current_zonelist_order],
4407 page_group_by_mobility_disabled ? "off" : "on",
4410 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4415 * Helper functions to size the waitqueue hash table.
4416 * Essentially these want to choose hash table sizes sufficiently
4417 * large so that collisions trying to wait on pages are rare.
4418 * But in fact, the number of active page waitqueues on typical
4419 * systems is ridiculously low, less than 200. So this is even
4420 * conservative, even though it seems large.
4422 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4423 * waitqueues, i.e. the size of the waitq table given the number of pages.
4425 #define PAGES_PER_WAITQUEUE 256
4427 #ifndef CONFIG_MEMORY_HOTPLUG
4428 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4430 unsigned long size = 1;
4432 pages /= PAGES_PER_WAITQUEUE;
4434 while (size < pages)
4438 * Once we have dozens or even hundreds of threads sleeping
4439 * on IO we've got bigger problems than wait queue collision.
4440 * Limit the size of the wait table to a reasonable size.
4442 size = min(size, 4096UL);
4444 return max(size, 4UL);
4448 * A zone's size might be changed by hot-add, so it is not possible to determine
4449 * a suitable size for its wait_table. So we use the maximum size now.
4451 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4453 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4454 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4455 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4457 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4458 * or more by the traditional way. (See above). It equals:
4460 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4461 * ia64(16K page size) : = ( 8G + 4M)byte.
4462 * powerpc (64K page size) : = (32G +16M)byte.
4464 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4471 * This is an integer logarithm so that shifts can be used later
4472 * to extract the more random high bits from the multiplicative
4473 * hash function before the remainder is taken.
4475 static inline unsigned long wait_table_bits(unsigned long size)
4481 * Initially all pages are reserved - free ones are freed
4482 * up by free_all_bootmem() once the early boot process is
4483 * done. Non-atomic initialization, single-pass.
4485 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4486 unsigned long start_pfn, enum memmap_context context)
4488 pg_data_t *pgdat = NODE_DATA(nid);
4489 unsigned long end_pfn = start_pfn + size;
4492 unsigned long nr_initialised = 0;
4494 if (highest_memmap_pfn < end_pfn - 1)
4495 highest_memmap_pfn = end_pfn - 1;
4497 z = &pgdat->node_zones[zone];
4498 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4500 * There can be holes in boot-time mem_map[]s
4501 * handed to this function. They do not
4502 * exist on hotplugged memory.
4504 if (context == MEMMAP_EARLY) {
4505 if (!early_pfn_valid(pfn))
4507 if (!early_pfn_in_nid(pfn, nid))
4509 if (!update_defer_init(pgdat, pfn, end_pfn,
4515 * Mark the block movable so that blocks are reserved for
4516 * movable at startup. This will force kernel allocations
4517 * to reserve their blocks rather than leaking throughout
4518 * the address space during boot when many long-lived
4519 * kernel allocations are made.
4521 * bitmap is created for zone's valid pfn range. but memmap
4522 * can be created for invalid pages (for alignment)
4523 * check here not to call set_pageblock_migratetype() against
4526 if (!(pfn & (pageblock_nr_pages - 1))) {
4527 struct page *page = pfn_to_page(pfn);
4529 __init_single_page(page, pfn, zone, nid);
4530 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4532 __init_single_pfn(pfn, zone, nid);
4537 static void __meminit zone_init_free_lists(struct zone *zone)
4539 unsigned int order, t;
4540 for_each_migratetype_order(order, t) {
4541 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4542 zone->free_area[order].nr_free = 0;
4546 #ifndef __HAVE_ARCH_MEMMAP_INIT
4547 #define memmap_init(size, nid, zone, start_pfn) \
4548 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4551 static int zone_batchsize(struct zone *zone)
4557 * The per-cpu-pages pools are set to around 1000th of the
4558 * size of the zone. But no more than 1/2 of a meg.
4560 * OK, so we don't know how big the cache is. So guess.
4562 batch = zone->managed_pages / 1024;
4563 if (batch * PAGE_SIZE > 512 * 1024)
4564 batch = (512 * 1024) / PAGE_SIZE;
4565 batch /= 4; /* We effectively *= 4 below */
4570 * Clamp the batch to a 2^n - 1 value. Having a power
4571 * of 2 value was found to be more likely to have
4572 * suboptimal cache aliasing properties in some cases.
4574 * For example if 2 tasks are alternately allocating
4575 * batches of pages, one task can end up with a lot
4576 * of pages of one half of the possible page colors
4577 * and the other with pages of the other colors.
4579 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4584 /* The deferral and batching of frees should be suppressed under NOMMU
4587 * The problem is that NOMMU needs to be able to allocate large chunks
4588 * of contiguous memory as there's no hardware page translation to
4589 * assemble apparent contiguous memory from discontiguous pages.
4591 * Queueing large contiguous runs of pages for batching, however,
4592 * causes the pages to actually be freed in smaller chunks. As there
4593 * can be a significant delay between the individual batches being
4594 * recycled, this leads to the once large chunks of space being
4595 * fragmented and becoming unavailable for high-order allocations.
4602 * pcp->high and pcp->batch values are related and dependent on one another:
4603 * ->batch must never be higher then ->high.
4604 * The following function updates them in a safe manner without read side
4607 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4608 * those fields changing asynchronously (acording the the above rule).
4610 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4611 * outside of boot time (or some other assurance that no concurrent updaters
4614 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4615 unsigned long batch)
4617 /* start with a fail safe value for batch */
4621 /* Update high, then batch, in order */
4628 /* a companion to pageset_set_high() */
4629 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4631 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4634 static void pageset_init(struct per_cpu_pageset *p)
4636 struct per_cpu_pages *pcp;
4639 memset(p, 0, sizeof(*p));
4643 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4644 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4647 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4650 pageset_set_batch(p, batch);
4654 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4655 * to the value high for the pageset p.
4657 static void pageset_set_high(struct per_cpu_pageset *p,
4660 unsigned long batch = max(1UL, high / 4);
4661 if ((high / 4) > (PAGE_SHIFT * 8))
4662 batch = PAGE_SHIFT * 8;
4664 pageset_update(&p->pcp, high, batch);
4667 static void pageset_set_high_and_batch(struct zone *zone,
4668 struct per_cpu_pageset *pcp)
4670 if (percpu_pagelist_fraction)
4671 pageset_set_high(pcp,
4672 (zone->managed_pages /
4673 percpu_pagelist_fraction));
4675 pageset_set_batch(pcp, zone_batchsize(zone));
4678 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4680 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4683 pageset_set_high_and_batch(zone, pcp);
4686 static void __meminit setup_zone_pageset(struct zone *zone)
4689 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4690 for_each_possible_cpu(cpu)
4691 zone_pageset_init(zone, cpu);
4695 * Allocate per cpu pagesets and initialize them.
4696 * Before this call only boot pagesets were available.
4698 void __init setup_per_cpu_pageset(void)
4702 for_each_populated_zone(zone)
4703 setup_zone_pageset(zone);
4706 static noinline __init_refok
4707 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4713 * The per-page waitqueue mechanism uses hashed waitqueues
4716 zone->wait_table_hash_nr_entries =
4717 wait_table_hash_nr_entries(zone_size_pages);
4718 zone->wait_table_bits =
4719 wait_table_bits(zone->wait_table_hash_nr_entries);
4720 alloc_size = zone->wait_table_hash_nr_entries
4721 * sizeof(wait_queue_head_t);
4723 if (!slab_is_available()) {
4724 zone->wait_table = (wait_queue_head_t *)
4725 memblock_virt_alloc_node_nopanic(
4726 alloc_size, zone->zone_pgdat->node_id);
4729 * This case means that a zone whose size was 0 gets new memory
4730 * via memory hot-add.
4731 * But it may be the case that a new node was hot-added. In
4732 * this case vmalloc() will not be able to use this new node's
4733 * memory - this wait_table must be initialized to use this new
4734 * node itself as well.
4735 * To use this new node's memory, further consideration will be
4738 zone->wait_table = vmalloc(alloc_size);
4740 if (!zone->wait_table)
4743 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4744 init_waitqueue_head(zone->wait_table + i);
4749 static __meminit void zone_pcp_init(struct zone *zone)
4752 * per cpu subsystem is not up at this point. The following code
4753 * relies on the ability of the linker to provide the
4754 * offset of a (static) per cpu variable into the per cpu area.
4756 zone->pageset = &boot_pageset;
4758 if (populated_zone(zone))
4759 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4760 zone->name, zone->present_pages,
4761 zone_batchsize(zone));
4764 int __meminit init_currently_empty_zone(struct zone *zone,
4765 unsigned long zone_start_pfn,
4768 struct pglist_data *pgdat = zone->zone_pgdat;
4770 ret = zone_wait_table_init(zone, size);
4773 pgdat->nr_zones = zone_idx(zone) + 1;
4775 zone->zone_start_pfn = zone_start_pfn;
4777 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4778 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4780 (unsigned long)zone_idx(zone),
4781 zone_start_pfn, (zone_start_pfn + size));
4783 zone_init_free_lists(zone);
4788 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4789 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4792 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4794 int __meminit __early_pfn_to_nid(unsigned long pfn,
4795 struct mminit_pfnnid_cache *state)
4797 unsigned long start_pfn, end_pfn;
4800 if (state->last_start <= pfn && pfn < state->last_end)
4801 return state->last_nid;
4803 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4805 state->last_start = start_pfn;
4806 state->last_end = end_pfn;
4807 state->last_nid = nid;
4812 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4815 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4816 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4817 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4819 * If an architecture guarantees that all ranges registered contain no holes
4820 * and may be freed, this this function may be used instead of calling
4821 * memblock_free_early_nid() manually.
4823 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4825 unsigned long start_pfn, end_pfn;
4828 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4829 start_pfn = min(start_pfn, max_low_pfn);
4830 end_pfn = min(end_pfn, max_low_pfn);
4832 if (start_pfn < end_pfn)
4833 memblock_free_early_nid(PFN_PHYS(start_pfn),
4834 (end_pfn - start_pfn) << PAGE_SHIFT,
4840 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4841 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4843 * If an architecture guarantees that all ranges registered contain no holes and may
4844 * be freed, this function may be used instead of calling memory_present() manually.
4846 void __init sparse_memory_present_with_active_regions(int nid)
4848 unsigned long start_pfn, end_pfn;
4851 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4852 memory_present(this_nid, start_pfn, end_pfn);
4856 * get_pfn_range_for_nid - Return the start and end page frames for a node
4857 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4858 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4859 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4861 * It returns the start and end page frame of a node based on information
4862 * provided by memblock_set_node(). If called for a node
4863 * with no available memory, a warning is printed and the start and end
4866 void __meminit get_pfn_range_for_nid(unsigned int nid,
4867 unsigned long *start_pfn, unsigned long *end_pfn)
4869 unsigned long this_start_pfn, this_end_pfn;
4875 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4876 *start_pfn = min(*start_pfn, this_start_pfn);
4877 *end_pfn = max(*end_pfn, this_end_pfn);
4880 if (*start_pfn == -1UL)
4885 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4886 * assumption is made that zones within a node are ordered in monotonic
4887 * increasing memory addresses so that the "highest" populated zone is used
4889 static void __init find_usable_zone_for_movable(void)
4892 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4893 if (zone_index == ZONE_MOVABLE)
4896 if (arch_zone_highest_possible_pfn[zone_index] >
4897 arch_zone_lowest_possible_pfn[zone_index])
4901 VM_BUG_ON(zone_index == -1);
4902 movable_zone = zone_index;
4906 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4907 * because it is sized independent of architecture. Unlike the other zones,
4908 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4909 * in each node depending on the size of each node and how evenly kernelcore
4910 * is distributed. This helper function adjusts the zone ranges
4911 * provided by the architecture for a given node by using the end of the
4912 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4913 * zones within a node are in order of monotonic increases memory addresses
4915 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4916 unsigned long zone_type,
4917 unsigned long node_start_pfn,
4918 unsigned long node_end_pfn,
4919 unsigned long *zone_start_pfn,
4920 unsigned long *zone_end_pfn)
4922 /* Only adjust if ZONE_MOVABLE is on this node */
4923 if (zone_movable_pfn[nid]) {
4924 /* Size ZONE_MOVABLE */
4925 if (zone_type == ZONE_MOVABLE) {
4926 *zone_start_pfn = zone_movable_pfn[nid];
4927 *zone_end_pfn = min(node_end_pfn,
4928 arch_zone_highest_possible_pfn[movable_zone]);
4930 /* Adjust for ZONE_MOVABLE starting within this range */
4931 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4932 *zone_end_pfn > zone_movable_pfn[nid]) {
4933 *zone_end_pfn = zone_movable_pfn[nid];
4935 /* Check if this whole range is within ZONE_MOVABLE */
4936 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4937 *zone_start_pfn = *zone_end_pfn;
4942 * Return the number of pages a zone spans in a node, including holes
4943 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4945 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4946 unsigned long zone_type,
4947 unsigned long node_start_pfn,
4948 unsigned long node_end_pfn,
4949 unsigned long *ignored)
4951 unsigned long zone_start_pfn, zone_end_pfn;
4953 /* When hotadd a new node from cpu_up(), the node should be empty */
4954 if (!node_start_pfn && !node_end_pfn)
4957 /* Get the start and end of the zone */
4958 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4959 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4960 adjust_zone_range_for_zone_movable(nid, zone_type,
4961 node_start_pfn, node_end_pfn,
4962 &zone_start_pfn, &zone_end_pfn);
4964 /* Check that this node has pages within the zone's required range */
4965 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4968 /* Move the zone boundaries inside the node if necessary */
4969 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4970 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4972 /* Return the spanned pages */
4973 return zone_end_pfn - zone_start_pfn;
4977 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4978 * then all holes in the requested range will be accounted for.
4980 unsigned long __meminit __absent_pages_in_range(int nid,
4981 unsigned long range_start_pfn,
4982 unsigned long range_end_pfn)
4984 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4985 unsigned long start_pfn, end_pfn;
4988 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4989 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4990 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4991 nr_absent -= end_pfn - start_pfn;
4997 * absent_pages_in_range - Return number of page frames in holes within a range
4998 * @start_pfn: The start PFN to start searching for holes
4999 * @end_pfn: The end PFN to stop searching for holes
5001 * It returns the number of pages frames in memory holes within a range.
5003 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5004 unsigned long end_pfn)
5006 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5009 /* Return the number of page frames in holes in a zone on a node */
5010 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5011 unsigned long zone_type,
5012 unsigned long node_start_pfn,
5013 unsigned long node_end_pfn,
5014 unsigned long *ignored)
5016 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5017 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5018 unsigned long zone_start_pfn, zone_end_pfn;
5020 /* When hotadd a new node from cpu_up(), the node should be empty */
5021 if (!node_start_pfn && !node_end_pfn)
5024 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5025 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5027 adjust_zone_range_for_zone_movable(nid, zone_type,
5028 node_start_pfn, node_end_pfn,
5029 &zone_start_pfn, &zone_end_pfn);
5030 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5033 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5034 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5035 unsigned long zone_type,
5036 unsigned long node_start_pfn,
5037 unsigned long node_end_pfn,
5038 unsigned long *zones_size)
5040 return zones_size[zone_type];
5043 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5044 unsigned long zone_type,
5045 unsigned long node_start_pfn,
5046 unsigned long node_end_pfn,
5047 unsigned long *zholes_size)
5052 return zholes_size[zone_type];
5055 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5057 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5058 unsigned long node_start_pfn,
5059 unsigned long node_end_pfn,
5060 unsigned long *zones_size,
5061 unsigned long *zholes_size)
5063 unsigned long realtotalpages = 0, totalpages = 0;
5066 for (i = 0; i < MAX_NR_ZONES; i++) {
5067 struct zone *zone = pgdat->node_zones + i;
5068 unsigned long size, real_size;
5070 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5074 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5075 node_start_pfn, node_end_pfn,
5077 zone->spanned_pages = size;
5078 zone->present_pages = real_size;
5081 realtotalpages += real_size;
5084 pgdat->node_spanned_pages = totalpages;
5085 pgdat->node_present_pages = realtotalpages;
5086 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5090 #ifndef CONFIG_SPARSEMEM
5092 * Calculate the size of the zone->blockflags rounded to an unsigned long
5093 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5094 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5095 * round what is now in bits to nearest long in bits, then return it in
5098 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5100 unsigned long usemapsize;
5102 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5103 usemapsize = roundup(zonesize, pageblock_nr_pages);
5104 usemapsize = usemapsize >> pageblock_order;
5105 usemapsize *= NR_PAGEBLOCK_BITS;
5106 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5108 return usemapsize / 8;
5111 static void __init setup_usemap(struct pglist_data *pgdat,
5113 unsigned long zone_start_pfn,
5114 unsigned long zonesize)
5116 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5117 zone->pageblock_flags = NULL;
5119 zone->pageblock_flags =
5120 memblock_virt_alloc_node_nopanic(usemapsize,
5124 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5125 unsigned long zone_start_pfn, unsigned long zonesize) {}
5126 #endif /* CONFIG_SPARSEMEM */
5128 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5130 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5131 void __paginginit set_pageblock_order(void)
5135 /* Check that pageblock_nr_pages has not already been setup */
5136 if (pageblock_order)
5139 if (HPAGE_SHIFT > PAGE_SHIFT)
5140 order = HUGETLB_PAGE_ORDER;
5142 order = MAX_ORDER - 1;
5145 * Assume the largest contiguous order of interest is a huge page.
5146 * This value may be variable depending on boot parameters on IA64 and
5149 pageblock_order = order;
5151 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5154 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5155 * is unused as pageblock_order is set at compile-time. See
5156 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5159 void __paginginit set_pageblock_order(void)
5163 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5165 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5166 unsigned long present_pages)
5168 unsigned long pages = spanned_pages;
5171 * Provide a more accurate estimation if there are holes within
5172 * the zone and SPARSEMEM is in use. If there are holes within the
5173 * zone, each populated memory region may cost us one or two extra
5174 * memmap pages due to alignment because memmap pages for each
5175 * populated regions may not naturally algined on page boundary.
5176 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5178 if (spanned_pages > present_pages + (present_pages >> 4) &&
5179 IS_ENABLED(CONFIG_SPARSEMEM))
5180 pages = present_pages;
5182 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5186 * Set up the zone data structures:
5187 * - mark all pages reserved
5188 * - mark all memory queues empty
5189 * - clear the memory bitmaps
5191 * NOTE: pgdat should get zeroed by caller.
5193 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5196 int nid = pgdat->node_id;
5197 unsigned long zone_start_pfn = pgdat->node_start_pfn;
5200 pgdat_resize_init(pgdat);
5201 #ifdef CONFIG_NUMA_BALANCING
5202 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5203 pgdat->numabalancing_migrate_nr_pages = 0;
5204 pgdat->numabalancing_migrate_next_window = jiffies;
5206 init_waitqueue_head(&pgdat->kswapd_wait);
5207 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5208 pgdat_page_ext_init(pgdat);
5210 for (j = 0; j < MAX_NR_ZONES; j++) {
5211 struct zone *zone = pgdat->node_zones + j;
5212 unsigned long size, realsize, freesize, memmap_pages;
5214 size = zone->spanned_pages;
5215 realsize = freesize = zone->present_pages;
5218 * Adjust freesize so that it accounts for how much memory
5219 * is used by this zone for memmap. This affects the watermark
5220 * and per-cpu initialisations
5222 memmap_pages = calc_memmap_size(size, realsize);
5223 if (!is_highmem_idx(j)) {
5224 if (freesize >= memmap_pages) {
5225 freesize -= memmap_pages;
5228 " %s zone: %lu pages used for memmap\n",
5229 zone_names[j], memmap_pages);
5232 " %s zone: %lu pages exceeds freesize %lu\n",
5233 zone_names[j], memmap_pages, freesize);
5236 /* Account for reserved pages */
5237 if (j == 0 && freesize > dma_reserve) {
5238 freesize -= dma_reserve;
5239 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5240 zone_names[0], dma_reserve);
5243 if (!is_highmem_idx(j))
5244 nr_kernel_pages += freesize;
5245 /* Charge for highmem memmap if there are enough kernel pages */
5246 else if (nr_kernel_pages > memmap_pages * 2)
5247 nr_kernel_pages -= memmap_pages;
5248 nr_all_pages += freesize;
5251 * Set an approximate value for lowmem here, it will be adjusted
5252 * when the bootmem allocator frees pages into the buddy system.
5253 * And all highmem pages will be managed by the buddy system.
5255 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5258 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5260 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5262 zone->name = zone_names[j];
5263 spin_lock_init(&zone->lock);
5264 spin_lock_init(&zone->lru_lock);
5265 zone_seqlock_init(zone);
5266 zone->zone_pgdat = pgdat;
5267 zone_pcp_init(zone);
5269 /* For bootup, initialized properly in watermark setup */
5270 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5272 lruvec_init(&zone->lruvec);
5276 set_pageblock_order();
5277 setup_usemap(pgdat, zone, zone_start_pfn, size);
5278 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5280 memmap_init(size, nid, j, zone_start_pfn);
5281 zone_start_pfn += size;
5285 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5287 unsigned long __maybe_unused start = 0;
5288 unsigned long __maybe_unused offset = 0;
5290 /* Skip empty nodes */
5291 if (!pgdat->node_spanned_pages)
5294 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5295 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5296 offset = pgdat->node_start_pfn - start;
5297 /* ia64 gets its own node_mem_map, before this, without bootmem */
5298 if (!pgdat->node_mem_map) {
5299 unsigned long size, end;
5303 * The zone's endpoints aren't required to be MAX_ORDER
5304 * aligned but the node_mem_map endpoints must be in order
5305 * for the buddy allocator to function correctly.
5307 end = pgdat_end_pfn(pgdat);
5308 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5309 size = (end - start) * sizeof(struct page);
5310 map = alloc_remap(pgdat->node_id, size);
5312 map = memblock_virt_alloc_node_nopanic(size,
5314 pgdat->node_mem_map = map + offset;
5316 #ifndef CONFIG_NEED_MULTIPLE_NODES
5318 * With no DISCONTIG, the global mem_map is just set as node 0's
5320 if (pgdat == NODE_DATA(0)) {
5321 mem_map = NODE_DATA(0)->node_mem_map;
5322 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5323 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5325 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5328 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5331 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5332 unsigned long node_start_pfn, unsigned long *zholes_size)
5334 pg_data_t *pgdat = NODE_DATA(nid);
5335 unsigned long start_pfn = 0;
5336 unsigned long end_pfn = 0;
5338 /* pg_data_t should be reset to zero when it's allocated */
5339 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5341 reset_deferred_meminit(pgdat);
5342 pgdat->node_id = nid;
5343 pgdat->node_start_pfn = node_start_pfn;
5344 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5345 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5346 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5347 (u64)start_pfn << PAGE_SHIFT,
5348 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5350 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5351 zones_size, zholes_size);
5353 alloc_node_mem_map(pgdat);
5354 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5355 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5356 nid, (unsigned long)pgdat,
5357 (unsigned long)pgdat->node_mem_map);
5360 free_area_init_core(pgdat);
5363 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5365 #if MAX_NUMNODES > 1
5367 * Figure out the number of possible node ids.
5369 void __init setup_nr_node_ids(void)
5371 unsigned int highest;
5373 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5374 nr_node_ids = highest + 1;
5379 * node_map_pfn_alignment - determine the maximum internode alignment
5381 * This function should be called after node map is populated and sorted.
5382 * It calculates the maximum power of two alignment which can distinguish
5385 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5386 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5387 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5388 * shifted, 1GiB is enough and this function will indicate so.
5390 * This is used to test whether pfn -> nid mapping of the chosen memory
5391 * model has fine enough granularity to avoid incorrect mapping for the
5392 * populated node map.
5394 * Returns the determined alignment in pfn's. 0 if there is no alignment
5395 * requirement (single node).
5397 unsigned long __init node_map_pfn_alignment(void)
5399 unsigned long accl_mask = 0, last_end = 0;
5400 unsigned long start, end, mask;
5404 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5405 if (!start || last_nid < 0 || last_nid == nid) {
5412 * Start with a mask granular enough to pin-point to the
5413 * start pfn and tick off bits one-by-one until it becomes
5414 * too coarse to separate the current node from the last.
5416 mask = ~((1 << __ffs(start)) - 1);
5417 while (mask && last_end <= (start & (mask << 1)))
5420 /* accumulate all internode masks */
5424 /* convert mask to number of pages */
5425 return ~accl_mask + 1;
5428 /* Find the lowest pfn for a node */
5429 static unsigned long __init find_min_pfn_for_node(int nid)
5431 unsigned long min_pfn = ULONG_MAX;
5432 unsigned long start_pfn;
5435 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5436 min_pfn = min(min_pfn, start_pfn);
5438 if (min_pfn == ULONG_MAX) {
5440 "Could not find start_pfn for node %d\n", nid);
5448 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5450 * It returns the minimum PFN based on information provided via
5451 * memblock_set_node().
5453 unsigned long __init find_min_pfn_with_active_regions(void)
5455 return find_min_pfn_for_node(MAX_NUMNODES);
5459 * early_calculate_totalpages()
5460 * Sum pages in active regions for movable zone.
5461 * Populate N_MEMORY for calculating usable_nodes.
5463 static unsigned long __init early_calculate_totalpages(void)
5465 unsigned long totalpages = 0;
5466 unsigned long start_pfn, end_pfn;
5469 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5470 unsigned long pages = end_pfn - start_pfn;
5472 totalpages += pages;
5474 node_set_state(nid, N_MEMORY);
5480 * Find the PFN the Movable zone begins in each node. Kernel memory
5481 * is spread evenly between nodes as long as the nodes have enough
5482 * memory. When they don't, some nodes will have more kernelcore than
5485 static void __init find_zone_movable_pfns_for_nodes(void)
5488 unsigned long usable_startpfn;
5489 unsigned long kernelcore_node, kernelcore_remaining;
5490 /* save the state before borrow the nodemask */
5491 nodemask_t saved_node_state = node_states[N_MEMORY];
5492 unsigned long totalpages = early_calculate_totalpages();
5493 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5494 struct memblock_region *r;
5496 /* Need to find movable_zone earlier when movable_node is specified. */
5497 find_usable_zone_for_movable();
5500 * If movable_node is specified, ignore kernelcore and movablecore
5503 if (movable_node_is_enabled()) {
5504 for_each_memblock(memory, r) {
5505 if (!memblock_is_hotpluggable(r))
5510 usable_startpfn = PFN_DOWN(r->base);
5511 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5512 min(usable_startpfn, zone_movable_pfn[nid]) :
5520 * If movablecore=nn[KMG] was specified, calculate what size of
5521 * kernelcore that corresponds so that memory usable for
5522 * any allocation type is evenly spread. If both kernelcore
5523 * and movablecore are specified, then the value of kernelcore
5524 * will be used for required_kernelcore if it's greater than
5525 * what movablecore would have allowed.
5527 if (required_movablecore) {
5528 unsigned long corepages;
5531 * Round-up so that ZONE_MOVABLE is at least as large as what
5532 * was requested by the user
5534 required_movablecore =
5535 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5536 required_movablecore = min(totalpages, required_movablecore);
5537 corepages = totalpages - required_movablecore;
5539 required_kernelcore = max(required_kernelcore, corepages);
5543 * If kernelcore was not specified or kernelcore size is larger
5544 * than totalpages, there is no ZONE_MOVABLE.
5546 if (!required_kernelcore || required_kernelcore >= totalpages)
5549 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5550 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5553 /* Spread kernelcore memory as evenly as possible throughout nodes */
5554 kernelcore_node = required_kernelcore / usable_nodes;
5555 for_each_node_state(nid, N_MEMORY) {
5556 unsigned long start_pfn, end_pfn;
5559 * Recalculate kernelcore_node if the division per node
5560 * now exceeds what is necessary to satisfy the requested
5561 * amount of memory for the kernel
5563 if (required_kernelcore < kernelcore_node)
5564 kernelcore_node = required_kernelcore / usable_nodes;
5567 * As the map is walked, we track how much memory is usable
5568 * by the kernel using kernelcore_remaining. When it is
5569 * 0, the rest of the node is usable by ZONE_MOVABLE
5571 kernelcore_remaining = kernelcore_node;
5573 /* Go through each range of PFNs within this node */
5574 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5575 unsigned long size_pages;
5577 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5578 if (start_pfn >= end_pfn)
5581 /* Account for what is only usable for kernelcore */
5582 if (start_pfn < usable_startpfn) {
5583 unsigned long kernel_pages;
5584 kernel_pages = min(end_pfn, usable_startpfn)
5587 kernelcore_remaining -= min(kernel_pages,
5588 kernelcore_remaining);
5589 required_kernelcore -= min(kernel_pages,
5590 required_kernelcore);
5592 /* Continue if range is now fully accounted */
5593 if (end_pfn <= usable_startpfn) {
5596 * Push zone_movable_pfn to the end so
5597 * that if we have to rebalance
5598 * kernelcore across nodes, we will
5599 * not double account here
5601 zone_movable_pfn[nid] = end_pfn;
5604 start_pfn = usable_startpfn;
5608 * The usable PFN range for ZONE_MOVABLE is from
5609 * start_pfn->end_pfn. Calculate size_pages as the
5610 * number of pages used as kernelcore
5612 size_pages = end_pfn - start_pfn;
5613 if (size_pages > kernelcore_remaining)
5614 size_pages = kernelcore_remaining;
5615 zone_movable_pfn[nid] = start_pfn + size_pages;
5618 * Some kernelcore has been met, update counts and
5619 * break if the kernelcore for this node has been
5622 required_kernelcore -= min(required_kernelcore,
5624 kernelcore_remaining -= size_pages;
5625 if (!kernelcore_remaining)
5631 * If there is still required_kernelcore, we do another pass with one
5632 * less node in the count. This will push zone_movable_pfn[nid] further
5633 * along on the nodes that still have memory until kernelcore is
5637 if (usable_nodes && required_kernelcore > usable_nodes)
5641 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5642 for (nid = 0; nid < MAX_NUMNODES; nid++)
5643 zone_movable_pfn[nid] =
5644 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5647 /* restore the node_state */
5648 node_states[N_MEMORY] = saved_node_state;
5651 /* Any regular or high memory on that node ? */
5652 static void check_for_memory(pg_data_t *pgdat, int nid)
5654 enum zone_type zone_type;
5656 if (N_MEMORY == N_NORMAL_MEMORY)
5659 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5660 struct zone *zone = &pgdat->node_zones[zone_type];
5661 if (populated_zone(zone)) {
5662 node_set_state(nid, N_HIGH_MEMORY);
5663 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5664 zone_type <= ZONE_NORMAL)
5665 node_set_state(nid, N_NORMAL_MEMORY);
5672 * free_area_init_nodes - Initialise all pg_data_t and zone data
5673 * @max_zone_pfn: an array of max PFNs for each zone
5675 * This will call free_area_init_node() for each active node in the system.
5676 * Using the page ranges provided by memblock_set_node(), the size of each
5677 * zone in each node and their holes is calculated. If the maximum PFN
5678 * between two adjacent zones match, it is assumed that the zone is empty.
5679 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5680 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5681 * starts where the previous one ended. For example, ZONE_DMA32 starts
5682 * at arch_max_dma_pfn.
5684 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5686 unsigned long start_pfn, end_pfn;
5689 /* Record where the zone boundaries are */
5690 memset(arch_zone_lowest_possible_pfn, 0,
5691 sizeof(arch_zone_lowest_possible_pfn));
5692 memset(arch_zone_highest_possible_pfn, 0,
5693 sizeof(arch_zone_highest_possible_pfn));
5694 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5695 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5696 for (i = 1; i < MAX_NR_ZONES; i++) {
5697 if (i == ZONE_MOVABLE)
5699 arch_zone_lowest_possible_pfn[i] =
5700 arch_zone_highest_possible_pfn[i-1];
5701 arch_zone_highest_possible_pfn[i] =
5702 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5704 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5705 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5707 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5708 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5709 find_zone_movable_pfns_for_nodes();
5711 /* Print out the zone ranges */
5712 pr_info("Zone ranges:\n");
5713 for (i = 0; i < MAX_NR_ZONES; i++) {
5714 if (i == ZONE_MOVABLE)
5716 pr_info(" %-8s ", zone_names[i]);
5717 if (arch_zone_lowest_possible_pfn[i] ==
5718 arch_zone_highest_possible_pfn[i])
5721 pr_cont("[mem %#018Lx-%#018Lx]\n",
5722 (u64)arch_zone_lowest_possible_pfn[i]
5724 ((u64)arch_zone_highest_possible_pfn[i]
5725 << PAGE_SHIFT) - 1);
5728 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5729 pr_info("Movable zone start for each node\n");
5730 for (i = 0; i < MAX_NUMNODES; i++) {
5731 if (zone_movable_pfn[i])
5732 pr_info(" Node %d: %#018Lx\n", i,
5733 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5736 /* Print out the early node map */
5737 pr_info("Early memory node ranges\n");
5738 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5739 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5740 (u64)start_pfn << PAGE_SHIFT,
5741 ((u64)end_pfn << PAGE_SHIFT) - 1);
5743 /* Initialise every node */
5744 mminit_verify_pageflags_layout();
5745 setup_nr_node_ids();
5746 for_each_online_node(nid) {
5747 pg_data_t *pgdat = NODE_DATA(nid);
5748 free_area_init_node(nid, NULL,
5749 find_min_pfn_for_node(nid), NULL);
5751 /* Any memory on that node */
5752 if (pgdat->node_present_pages)
5753 node_set_state(nid, N_MEMORY);
5754 check_for_memory(pgdat, nid);
5758 static int __init cmdline_parse_core(char *p, unsigned long *core)
5760 unsigned long long coremem;
5764 coremem = memparse(p, &p);
5765 *core = coremem >> PAGE_SHIFT;
5767 /* Paranoid check that UL is enough for the coremem value */
5768 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5774 * kernelcore=size sets the amount of memory for use for allocations that
5775 * cannot be reclaimed or migrated.
5777 static int __init cmdline_parse_kernelcore(char *p)
5779 return cmdline_parse_core(p, &required_kernelcore);
5783 * movablecore=size sets the amount of memory for use for allocations that
5784 * can be reclaimed or migrated.
5786 static int __init cmdline_parse_movablecore(char *p)
5788 return cmdline_parse_core(p, &required_movablecore);
5791 early_param("kernelcore", cmdline_parse_kernelcore);
5792 early_param("movablecore", cmdline_parse_movablecore);
5794 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5796 void adjust_managed_page_count(struct page *page, long count)
5798 spin_lock(&managed_page_count_lock);
5799 page_zone(page)->managed_pages += count;
5800 totalram_pages += count;
5801 #ifdef CONFIG_HIGHMEM
5802 if (PageHighMem(page))
5803 totalhigh_pages += count;
5805 spin_unlock(&managed_page_count_lock);
5807 EXPORT_SYMBOL(adjust_managed_page_count);
5809 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5812 unsigned long pages = 0;
5814 start = (void *)PAGE_ALIGN((unsigned long)start);
5815 end = (void *)((unsigned long)end & PAGE_MASK);
5816 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5817 if ((unsigned int)poison <= 0xFF)
5818 memset(pos, poison, PAGE_SIZE);
5819 free_reserved_page(virt_to_page(pos));
5823 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5824 s, pages << (PAGE_SHIFT - 10), start, end);
5828 EXPORT_SYMBOL(free_reserved_area);
5830 #ifdef CONFIG_HIGHMEM
5831 void free_highmem_page(struct page *page)
5833 __free_reserved_page(page);
5835 page_zone(page)->managed_pages++;
5841 void __init mem_init_print_info(const char *str)
5843 unsigned long physpages, codesize, datasize, rosize, bss_size;
5844 unsigned long init_code_size, init_data_size;
5846 physpages = get_num_physpages();
5847 codesize = _etext - _stext;
5848 datasize = _edata - _sdata;
5849 rosize = __end_rodata - __start_rodata;
5850 bss_size = __bss_stop - __bss_start;
5851 init_data_size = __init_end - __init_begin;
5852 init_code_size = _einittext - _sinittext;
5855 * Detect special cases and adjust section sizes accordingly:
5856 * 1) .init.* may be embedded into .data sections
5857 * 2) .init.text.* may be out of [__init_begin, __init_end],
5858 * please refer to arch/tile/kernel/vmlinux.lds.S.
5859 * 3) .rodata.* may be embedded into .text or .data sections.
5861 #define adj_init_size(start, end, size, pos, adj) \
5863 if (start <= pos && pos < end && size > adj) \
5867 adj_init_size(__init_begin, __init_end, init_data_size,
5868 _sinittext, init_code_size);
5869 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5870 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5871 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5872 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5874 #undef adj_init_size
5876 pr_info("Memory: %luK/%luK available "
5877 "(%luK kernel code, %luK rwdata, %luK rodata, "
5878 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5879 #ifdef CONFIG_HIGHMEM
5883 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5884 codesize >> 10, datasize >> 10, rosize >> 10,
5885 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5886 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5887 totalcma_pages << (PAGE_SHIFT-10),
5888 #ifdef CONFIG_HIGHMEM
5889 totalhigh_pages << (PAGE_SHIFT-10),
5891 str ? ", " : "", str ? str : "");
5895 * set_dma_reserve - set the specified number of pages reserved in the first zone
5896 * @new_dma_reserve: The number of pages to mark reserved
5898 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
5899 * In the DMA zone, a significant percentage may be consumed by kernel image
5900 * and other unfreeable allocations which can skew the watermarks badly. This
5901 * function may optionally be used to account for unfreeable pages in the
5902 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5903 * smaller per-cpu batchsize.
5905 void __init set_dma_reserve(unsigned long new_dma_reserve)
5907 dma_reserve = new_dma_reserve;
5910 void __init free_area_init(unsigned long *zones_size)
5912 free_area_init_node(0, zones_size,
5913 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5916 static int page_alloc_cpu_notify(struct notifier_block *self,
5917 unsigned long action, void *hcpu)
5919 int cpu = (unsigned long)hcpu;
5921 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5922 lru_add_drain_cpu(cpu);
5926 * Spill the event counters of the dead processor
5927 * into the current processors event counters.
5928 * This artificially elevates the count of the current
5931 vm_events_fold_cpu(cpu);
5934 * Zero the differential counters of the dead processor
5935 * so that the vm statistics are consistent.
5937 * This is only okay since the processor is dead and cannot
5938 * race with what we are doing.
5940 cpu_vm_stats_fold(cpu);
5945 void __init page_alloc_init(void)
5947 hotcpu_notifier(page_alloc_cpu_notify, 0);
5951 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5952 * or min_free_kbytes changes.
5954 static void calculate_totalreserve_pages(void)
5956 struct pglist_data *pgdat;
5957 unsigned long reserve_pages = 0;
5958 enum zone_type i, j;
5960 for_each_online_pgdat(pgdat) {
5961 for (i = 0; i < MAX_NR_ZONES; i++) {
5962 struct zone *zone = pgdat->node_zones + i;
5965 /* Find valid and maximum lowmem_reserve in the zone */
5966 for (j = i; j < MAX_NR_ZONES; j++) {
5967 if (zone->lowmem_reserve[j] > max)
5968 max = zone->lowmem_reserve[j];
5971 /* we treat the high watermark as reserved pages. */
5972 max += high_wmark_pages(zone);
5974 if (max > zone->managed_pages)
5975 max = zone->managed_pages;
5977 zone->totalreserve_pages = max;
5979 reserve_pages += max;
5982 totalreserve_pages = reserve_pages;
5986 * setup_per_zone_lowmem_reserve - called whenever
5987 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
5988 * has a correct pages reserved value, so an adequate number of
5989 * pages are left in the zone after a successful __alloc_pages().
5991 static void setup_per_zone_lowmem_reserve(void)
5993 struct pglist_data *pgdat;
5994 enum zone_type j, idx;
5996 for_each_online_pgdat(pgdat) {
5997 for (j = 0; j < MAX_NR_ZONES; j++) {
5998 struct zone *zone = pgdat->node_zones + j;
5999 unsigned long managed_pages = zone->managed_pages;
6001 zone->lowmem_reserve[j] = 0;
6005 struct zone *lower_zone;
6009 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6010 sysctl_lowmem_reserve_ratio[idx] = 1;
6012 lower_zone = pgdat->node_zones + idx;
6013 lower_zone->lowmem_reserve[j] = managed_pages /
6014 sysctl_lowmem_reserve_ratio[idx];
6015 managed_pages += lower_zone->managed_pages;
6020 /* update totalreserve_pages */
6021 calculate_totalreserve_pages();
6024 static void __setup_per_zone_wmarks(void)
6026 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6027 unsigned long lowmem_pages = 0;
6029 unsigned long flags;
6031 /* Calculate total number of !ZONE_HIGHMEM pages */
6032 for_each_zone(zone) {
6033 if (!is_highmem(zone))
6034 lowmem_pages += zone->managed_pages;
6037 for_each_zone(zone) {
6040 spin_lock_irqsave(&zone->lock, flags);
6041 tmp = (u64)pages_min * zone->managed_pages;
6042 do_div(tmp, lowmem_pages);
6043 if (is_highmem(zone)) {
6045 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6046 * need highmem pages, so cap pages_min to a small
6049 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6050 * deltas control asynch page reclaim, and so should
6051 * not be capped for highmem.
6053 unsigned long min_pages;
6055 min_pages = zone->managed_pages / 1024;
6056 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6057 zone->watermark[WMARK_MIN] = min_pages;
6060 * If it's a lowmem zone, reserve a number of pages
6061 * proportionate to the zone's size.
6063 zone->watermark[WMARK_MIN] = tmp;
6066 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
6067 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
6069 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6070 high_wmark_pages(zone) - low_wmark_pages(zone) -
6071 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6073 spin_unlock_irqrestore(&zone->lock, flags);
6076 /* update totalreserve_pages */
6077 calculate_totalreserve_pages();
6081 * setup_per_zone_wmarks - called when min_free_kbytes changes
6082 * or when memory is hot-{added|removed}
6084 * Ensures that the watermark[min,low,high] values for each zone are set
6085 * correctly with respect to min_free_kbytes.
6087 void setup_per_zone_wmarks(void)
6089 mutex_lock(&zonelists_mutex);
6090 __setup_per_zone_wmarks();
6091 mutex_unlock(&zonelists_mutex);
6095 * The inactive anon list should be small enough that the VM never has to
6096 * do too much work, but large enough that each inactive page has a chance
6097 * to be referenced again before it is swapped out.
6099 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6100 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6101 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6102 * the anonymous pages are kept on the inactive list.
6105 * memory ratio inactive anon
6106 * -------------------------------------
6115 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6117 unsigned int gb, ratio;
6119 /* Zone size in gigabytes */
6120 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6122 ratio = int_sqrt(10 * gb);
6126 zone->inactive_ratio = ratio;
6129 static void __meminit setup_per_zone_inactive_ratio(void)
6134 calculate_zone_inactive_ratio(zone);
6138 * Initialise min_free_kbytes.
6140 * For small machines we want it small (128k min). For large machines
6141 * we want it large (64MB max). But it is not linear, because network
6142 * bandwidth does not increase linearly with machine size. We use
6144 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6145 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6161 int __meminit init_per_zone_wmark_min(void)
6163 unsigned long lowmem_kbytes;
6164 int new_min_free_kbytes;
6166 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6167 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6169 if (new_min_free_kbytes > user_min_free_kbytes) {
6170 min_free_kbytes = new_min_free_kbytes;
6171 if (min_free_kbytes < 128)
6172 min_free_kbytes = 128;
6173 if (min_free_kbytes > 65536)
6174 min_free_kbytes = 65536;
6176 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6177 new_min_free_kbytes, user_min_free_kbytes);
6179 setup_per_zone_wmarks();
6180 refresh_zone_stat_thresholds();
6181 setup_per_zone_lowmem_reserve();
6182 setup_per_zone_inactive_ratio();
6185 module_init(init_per_zone_wmark_min)
6188 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6189 * that we can call two helper functions whenever min_free_kbytes
6192 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6193 void __user *buffer, size_t *length, loff_t *ppos)
6197 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6202 user_min_free_kbytes = min_free_kbytes;
6203 setup_per_zone_wmarks();
6209 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6210 void __user *buffer, size_t *length, loff_t *ppos)
6215 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6220 zone->min_unmapped_pages = (zone->managed_pages *
6221 sysctl_min_unmapped_ratio) / 100;
6225 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6226 void __user *buffer, size_t *length, loff_t *ppos)
6231 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6236 zone->min_slab_pages = (zone->managed_pages *
6237 sysctl_min_slab_ratio) / 100;
6243 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6244 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6245 * whenever sysctl_lowmem_reserve_ratio changes.
6247 * The reserve ratio obviously has absolutely no relation with the
6248 * minimum watermarks. The lowmem reserve ratio can only make sense
6249 * if in function of the boot time zone sizes.
6251 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6252 void __user *buffer, size_t *length, loff_t *ppos)
6254 proc_dointvec_minmax(table, write, buffer, length, ppos);
6255 setup_per_zone_lowmem_reserve();
6260 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6261 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6262 * pagelist can have before it gets flushed back to buddy allocator.
6264 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6265 void __user *buffer, size_t *length, loff_t *ppos)
6268 int old_percpu_pagelist_fraction;
6271 mutex_lock(&pcp_batch_high_lock);
6272 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6274 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6275 if (!write || ret < 0)
6278 /* Sanity checking to avoid pcp imbalance */
6279 if (percpu_pagelist_fraction &&
6280 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6281 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6287 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6290 for_each_populated_zone(zone) {
6293 for_each_possible_cpu(cpu)
6294 pageset_set_high_and_batch(zone,
6295 per_cpu_ptr(zone->pageset, cpu));
6298 mutex_unlock(&pcp_batch_high_lock);
6303 int hashdist = HASHDIST_DEFAULT;
6305 static int __init set_hashdist(char *str)
6309 hashdist = simple_strtoul(str, &str, 0);
6312 __setup("hashdist=", set_hashdist);
6316 * allocate a large system hash table from bootmem
6317 * - it is assumed that the hash table must contain an exact power-of-2
6318 * quantity of entries
6319 * - limit is the number of hash buckets, not the total allocation size
6321 void *__init alloc_large_system_hash(const char *tablename,
6322 unsigned long bucketsize,
6323 unsigned long numentries,
6326 unsigned int *_hash_shift,
6327 unsigned int *_hash_mask,
6328 unsigned long low_limit,
6329 unsigned long high_limit)
6331 unsigned long long max = high_limit;
6332 unsigned long log2qty, size;
6335 /* allow the kernel cmdline to have a say */
6337 /* round applicable memory size up to nearest megabyte */
6338 numentries = nr_kernel_pages;
6340 /* It isn't necessary when PAGE_SIZE >= 1MB */
6341 if (PAGE_SHIFT < 20)
6342 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6344 /* limit to 1 bucket per 2^scale bytes of low memory */
6345 if (scale > PAGE_SHIFT)
6346 numentries >>= (scale - PAGE_SHIFT);
6348 numentries <<= (PAGE_SHIFT - scale);
6350 /* Make sure we've got at least a 0-order allocation.. */
6351 if (unlikely(flags & HASH_SMALL)) {
6352 /* Makes no sense without HASH_EARLY */
6353 WARN_ON(!(flags & HASH_EARLY));
6354 if (!(numentries >> *_hash_shift)) {
6355 numentries = 1UL << *_hash_shift;
6356 BUG_ON(!numentries);
6358 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6359 numentries = PAGE_SIZE / bucketsize;
6361 numentries = roundup_pow_of_two(numentries);
6363 /* limit allocation size to 1/16 total memory by default */
6365 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6366 do_div(max, bucketsize);
6368 max = min(max, 0x80000000ULL);
6370 if (numentries < low_limit)
6371 numentries = low_limit;
6372 if (numentries > max)
6375 log2qty = ilog2(numentries);
6378 size = bucketsize << log2qty;
6379 if (flags & HASH_EARLY)
6380 table = memblock_virt_alloc_nopanic(size, 0);
6382 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6385 * If bucketsize is not a power-of-two, we may free
6386 * some pages at the end of hash table which
6387 * alloc_pages_exact() automatically does
6389 if (get_order(size) < MAX_ORDER) {
6390 table = alloc_pages_exact(size, GFP_ATOMIC);
6391 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6394 } while (!table && size > PAGE_SIZE && --log2qty);
6397 panic("Failed to allocate %s hash table\n", tablename);
6399 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6402 ilog2(size) - PAGE_SHIFT,
6406 *_hash_shift = log2qty;
6408 *_hash_mask = (1 << log2qty) - 1;
6413 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6414 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6417 #ifdef CONFIG_SPARSEMEM
6418 return __pfn_to_section(pfn)->pageblock_flags;
6420 return zone->pageblock_flags;
6421 #endif /* CONFIG_SPARSEMEM */
6424 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6426 #ifdef CONFIG_SPARSEMEM
6427 pfn &= (PAGES_PER_SECTION-1);
6428 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6430 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6431 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6432 #endif /* CONFIG_SPARSEMEM */
6436 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6437 * @page: The page within the block of interest
6438 * @pfn: The target page frame number
6439 * @end_bitidx: The last bit of interest to retrieve
6440 * @mask: mask of bits that the caller is interested in
6442 * Return: pageblock_bits flags
6444 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6445 unsigned long end_bitidx,
6449 unsigned long *bitmap;
6450 unsigned long bitidx, word_bitidx;
6453 zone = page_zone(page);
6454 bitmap = get_pageblock_bitmap(zone, pfn);
6455 bitidx = pfn_to_bitidx(zone, pfn);
6456 word_bitidx = bitidx / BITS_PER_LONG;
6457 bitidx &= (BITS_PER_LONG-1);
6459 word = bitmap[word_bitidx];
6460 bitidx += end_bitidx;
6461 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6465 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6466 * @page: The page within the block of interest
6467 * @flags: The flags to set
6468 * @pfn: The target page frame number
6469 * @end_bitidx: The last bit of interest
6470 * @mask: mask of bits that the caller is interested in
6472 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6474 unsigned long end_bitidx,
6478 unsigned long *bitmap;
6479 unsigned long bitidx, word_bitidx;
6480 unsigned long old_word, word;
6482 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6484 zone = page_zone(page);
6485 bitmap = get_pageblock_bitmap(zone, pfn);
6486 bitidx = pfn_to_bitidx(zone, pfn);
6487 word_bitidx = bitidx / BITS_PER_LONG;
6488 bitidx &= (BITS_PER_LONG-1);
6490 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6492 bitidx += end_bitidx;
6493 mask <<= (BITS_PER_LONG - bitidx - 1);
6494 flags <<= (BITS_PER_LONG - bitidx - 1);
6496 word = READ_ONCE(bitmap[word_bitidx]);
6498 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6499 if (word == old_word)
6506 * This function checks whether pageblock includes unmovable pages or not.
6507 * If @count is not zero, it is okay to include less @count unmovable pages
6509 * PageLRU check without isolation or lru_lock could race so that
6510 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6511 * expect this function should be exact.
6513 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6514 bool skip_hwpoisoned_pages)
6516 unsigned long pfn, iter, found;
6520 * For avoiding noise data, lru_add_drain_all() should be called
6521 * If ZONE_MOVABLE, the zone never contains unmovable pages
6523 if (zone_idx(zone) == ZONE_MOVABLE)
6525 mt = get_pageblock_migratetype(page);
6526 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6529 pfn = page_to_pfn(page);
6530 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6531 unsigned long check = pfn + iter;
6533 if (!pfn_valid_within(check))
6536 page = pfn_to_page(check);
6539 * Hugepages are not in LRU lists, but they're movable.
6540 * We need not scan over tail pages bacause we don't
6541 * handle each tail page individually in migration.
6543 if (PageHuge(page)) {
6544 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6549 * We can't use page_count without pin a page
6550 * because another CPU can free compound page.
6551 * This check already skips compound tails of THP
6552 * because their page->_count is zero at all time.
6554 if (!atomic_read(&page->_count)) {
6555 if (PageBuddy(page))
6556 iter += (1 << page_order(page)) - 1;
6561 * The HWPoisoned page may be not in buddy system, and
6562 * page_count() is not 0.
6564 if (skip_hwpoisoned_pages && PageHWPoison(page))
6570 * If there are RECLAIMABLE pages, we need to check
6571 * it. But now, memory offline itself doesn't call
6572 * shrink_node_slabs() and it still to be fixed.
6575 * If the page is not RAM, page_count()should be 0.
6576 * we don't need more check. This is an _used_ not-movable page.
6578 * The problematic thing here is PG_reserved pages. PG_reserved
6579 * is set to both of a memory hole page and a _used_ kernel
6588 bool is_pageblock_removable_nolock(struct page *page)
6594 * We have to be careful here because we are iterating over memory
6595 * sections which are not zone aware so we might end up outside of
6596 * the zone but still within the section.
6597 * We have to take care about the node as well. If the node is offline
6598 * its NODE_DATA will be NULL - see page_zone.
6600 if (!node_online(page_to_nid(page)))
6603 zone = page_zone(page);
6604 pfn = page_to_pfn(page);
6605 if (!zone_spans_pfn(zone, pfn))
6608 return !has_unmovable_pages(zone, page, 0, true);
6613 static unsigned long pfn_max_align_down(unsigned long pfn)
6615 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6616 pageblock_nr_pages) - 1);
6619 static unsigned long pfn_max_align_up(unsigned long pfn)
6621 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6622 pageblock_nr_pages));
6625 /* [start, end) must belong to a single zone. */
6626 static int __alloc_contig_migrate_range(struct compact_control *cc,
6627 unsigned long start, unsigned long end)
6629 /* This function is based on compact_zone() from compaction.c. */
6630 unsigned long nr_reclaimed;
6631 unsigned long pfn = start;
6632 unsigned int tries = 0;
6637 while (pfn < end || !list_empty(&cc->migratepages)) {
6638 if (fatal_signal_pending(current)) {
6643 if (list_empty(&cc->migratepages)) {
6644 cc->nr_migratepages = 0;
6645 pfn = isolate_migratepages_range(cc, pfn, end);
6651 } else if (++tries == 5) {
6652 ret = ret < 0 ? ret : -EBUSY;
6656 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6658 cc->nr_migratepages -= nr_reclaimed;
6660 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6661 NULL, 0, cc->mode, MR_CMA);
6664 putback_movable_pages(&cc->migratepages);
6671 * alloc_contig_range() -- tries to allocate given range of pages
6672 * @start: start PFN to allocate
6673 * @end: one-past-the-last PFN to allocate
6674 * @migratetype: migratetype of the underlaying pageblocks (either
6675 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6676 * in range must have the same migratetype and it must
6677 * be either of the two.
6679 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6680 * aligned, however it's the caller's responsibility to guarantee that
6681 * we are the only thread that changes migrate type of pageblocks the
6684 * The PFN range must belong to a single zone.
6686 * Returns zero on success or negative error code. On success all
6687 * pages which PFN is in [start, end) are allocated for the caller and
6688 * need to be freed with free_contig_range().
6690 int alloc_contig_range(unsigned long start, unsigned long end,
6691 unsigned migratetype)
6693 unsigned long outer_start, outer_end;
6697 struct compact_control cc = {
6698 .nr_migratepages = 0,
6700 .zone = page_zone(pfn_to_page(start)),
6701 .mode = MIGRATE_SYNC,
6702 .ignore_skip_hint = true,
6704 INIT_LIST_HEAD(&cc.migratepages);
6707 * What we do here is we mark all pageblocks in range as
6708 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6709 * have different sizes, and due to the way page allocator
6710 * work, we align the range to biggest of the two pages so
6711 * that page allocator won't try to merge buddies from
6712 * different pageblocks and change MIGRATE_ISOLATE to some
6713 * other migration type.
6715 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6716 * migrate the pages from an unaligned range (ie. pages that
6717 * we are interested in). This will put all the pages in
6718 * range back to page allocator as MIGRATE_ISOLATE.
6720 * When this is done, we take the pages in range from page
6721 * allocator removing them from the buddy system. This way
6722 * page allocator will never consider using them.
6724 * This lets us mark the pageblocks back as
6725 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6726 * aligned range but not in the unaligned, original range are
6727 * put back to page allocator so that buddy can use them.
6730 ret = start_isolate_page_range(pfn_max_align_down(start),
6731 pfn_max_align_up(end), migratetype,
6737 * In case of -EBUSY, we'd like to know which page causes problem.
6738 * So, just fall through. We will check it in test_pages_isolated().
6740 ret = __alloc_contig_migrate_range(&cc, start, end);
6741 if (ret && ret != -EBUSY)
6745 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6746 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6747 * more, all pages in [start, end) are free in page allocator.
6748 * What we are going to do is to allocate all pages from
6749 * [start, end) (that is remove them from page allocator).
6751 * The only problem is that pages at the beginning and at the
6752 * end of interesting range may be not aligned with pages that
6753 * page allocator holds, ie. they can be part of higher order
6754 * pages. Because of this, we reserve the bigger range and
6755 * once this is done free the pages we are not interested in.
6757 * We don't have to hold zone->lock here because the pages are
6758 * isolated thus they won't get removed from buddy.
6761 lru_add_drain_all();
6762 drain_all_pages(cc.zone);
6765 outer_start = start;
6766 while (!PageBuddy(pfn_to_page(outer_start))) {
6767 if (++order >= MAX_ORDER) {
6768 outer_start = start;
6771 outer_start &= ~0UL << order;
6774 if (outer_start != start) {
6775 order = page_order(pfn_to_page(outer_start));
6778 * outer_start page could be small order buddy page and
6779 * it doesn't include start page. Adjust outer_start
6780 * in this case to report failed page properly
6781 * on tracepoint in test_pages_isolated()
6783 if (outer_start + (1UL << order) <= start)
6784 outer_start = start;
6787 /* Make sure the range is really isolated. */
6788 if (test_pages_isolated(outer_start, end, false)) {
6789 pr_info("%s: [%lx, %lx) PFNs busy\n",
6790 __func__, outer_start, end);
6795 /* Grab isolated pages from freelists. */
6796 outer_end = isolate_freepages_range(&cc, outer_start, end);
6802 /* Free head and tail (if any) */
6803 if (start != outer_start)
6804 free_contig_range(outer_start, start - outer_start);
6805 if (end != outer_end)
6806 free_contig_range(end, outer_end - end);
6809 undo_isolate_page_range(pfn_max_align_down(start),
6810 pfn_max_align_up(end), migratetype);
6814 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6816 unsigned int count = 0;
6818 for (; nr_pages--; pfn++) {
6819 struct page *page = pfn_to_page(pfn);
6821 count += page_count(page) != 1;
6824 WARN(count != 0, "%d pages are still in use!\n", count);
6828 #ifdef CONFIG_MEMORY_HOTPLUG
6830 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6831 * page high values need to be recalulated.
6833 void __meminit zone_pcp_update(struct zone *zone)
6836 mutex_lock(&pcp_batch_high_lock);
6837 for_each_possible_cpu(cpu)
6838 pageset_set_high_and_batch(zone,
6839 per_cpu_ptr(zone->pageset, cpu));
6840 mutex_unlock(&pcp_batch_high_lock);
6844 void zone_pcp_reset(struct zone *zone)
6846 unsigned long flags;
6848 struct per_cpu_pageset *pset;
6850 /* avoid races with drain_pages() */
6851 local_irq_save(flags);
6852 if (zone->pageset != &boot_pageset) {
6853 for_each_online_cpu(cpu) {
6854 pset = per_cpu_ptr(zone->pageset, cpu);
6855 drain_zonestat(zone, pset);
6857 free_percpu(zone->pageset);
6858 zone->pageset = &boot_pageset;
6860 local_irq_restore(flags);
6863 #ifdef CONFIG_MEMORY_HOTREMOVE
6865 * All pages in the range must be isolated before calling this.
6868 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6872 unsigned int order, i;
6874 unsigned long flags;
6875 /* find the first valid pfn */
6876 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6881 zone = page_zone(pfn_to_page(pfn));
6882 spin_lock_irqsave(&zone->lock, flags);
6884 while (pfn < end_pfn) {
6885 if (!pfn_valid(pfn)) {
6889 page = pfn_to_page(pfn);
6891 * The HWPoisoned page may be not in buddy system, and
6892 * page_count() is not 0.
6894 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6896 SetPageReserved(page);
6900 BUG_ON(page_count(page));
6901 BUG_ON(!PageBuddy(page));
6902 order = page_order(page);
6903 #ifdef CONFIG_DEBUG_VM
6904 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6905 pfn, 1 << order, end_pfn);
6907 list_del(&page->lru);
6908 rmv_page_order(page);
6909 zone->free_area[order].nr_free--;
6910 for (i = 0; i < (1 << order); i++)
6911 SetPageReserved((page+i));
6912 pfn += (1 << order);
6914 spin_unlock_irqrestore(&zone->lock, flags);
6918 #ifdef CONFIG_MEMORY_FAILURE
6919 bool is_free_buddy_page(struct page *page)
6921 struct zone *zone = page_zone(page);
6922 unsigned long pfn = page_to_pfn(page);
6923 unsigned long flags;
6926 spin_lock_irqsave(&zone->lock, flags);
6927 for (order = 0; order < MAX_ORDER; order++) {
6928 struct page *page_head = page - (pfn & ((1 << order) - 1));
6930 if (PageBuddy(page_head) && page_order(page_head) >= order)
6933 spin_unlock_irqrestore(&zone->lock, flags);
6935 return order < MAX_ORDER;