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 static void free_compound_page(struct page *page);
226 compound_page_dtor * const compound_page_dtors[] = {
229 #ifdef CONFIG_HUGETLB_PAGE
234 int min_free_kbytes = 1024;
235 int user_min_free_kbytes = -1;
237 static unsigned long __meminitdata nr_kernel_pages;
238 static unsigned long __meminitdata nr_all_pages;
239 static unsigned long __meminitdata dma_reserve;
241 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
242 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
243 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
244 static unsigned long __initdata required_kernelcore;
245 static unsigned long __initdata required_movablecore;
246 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
248 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
250 EXPORT_SYMBOL(movable_zone);
251 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
254 int nr_node_ids __read_mostly = MAX_NUMNODES;
255 int nr_online_nodes __read_mostly = 1;
256 EXPORT_SYMBOL(nr_node_ids);
257 EXPORT_SYMBOL(nr_online_nodes);
260 int page_group_by_mobility_disabled __read_mostly;
262 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
263 static inline void reset_deferred_meminit(pg_data_t *pgdat)
265 pgdat->first_deferred_pfn = ULONG_MAX;
268 /* Returns true if the struct page for the pfn is uninitialised */
269 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
271 int nid = early_pfn_to_nid(pfn);
273 if (node_online(nid) && pfn >= NODE_DATA(nid)->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 static 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 set_compound_head(p, page);
475 #ifdef CONFIG_DEBUG_PAGEALLOC
476 unsigned int _debug_guardpage_minorder;
477 bool _debug_pagealloc_enabled __read_mostly;
478 bool _debug_guardpage_enabled __read_mostly;
480 static int __init early_debug_pagealloc(char *buf)
485 if (strcmp(buf, "on") == 0)
486 _debug_pagealloc_enabled = true;
490 early_param("debug_pagealloc", early_debug_pagealloc);
492 static bool need_debug_guardpage(void)
494 /* If we don't use debug_pagealloc, we don't need guard page */
495 if (!debug_pagealloc_enabled())
501 static void init_debug_guardpage(void)
503 if (!debug_pagealloc_enabled())
506 _debug_guardpage_enabled = true;
509 struct page_ext_operations debug_guardpage_ops = {
510 .need = need_debug_guardpage,
511 .init = init_debug_guardpage,
514 static int __init debug_guardpage_minorder_setup(char *buf)
518 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
519 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
522 _debug_guardpage_minorder = res;
523 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
526 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
528 static inline void set_page_guard(struct zone *zone, struct page *page,
529 unsigned int order, int migratetype)
531 struct page_ext *page_ext;
533 if (!debug_guardpage_enabled())
536 page_ext = lookup_page_ext(page);
537 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
539 INIT_LIST_HEAD(&page->lru);
540 set_page_private(page, order);
541 /* Guard pages are not available for any usage */
542 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
545 static inline void clear_page_guard(struct zone *zone, struct page *page,
546 unsigned int order, int migratetype)
548 struct page_ext *page_ext;
550 if (!debug_guardpage_enabled())
553 page_ext = lookup_page_ext(page);
554 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
556 set_page_private(page, 0);
557 if (!is_migrate_isolate(migratetype))
558 __mod_zone_freepage_state(zone, (1 << order), migratetype);
561 struct page_ext_operations debug_guardpage_ops = { NULL, };
562 static inline void set_page_guard(struct zone *zone, struct page *page,
563 unsigned int order, int migratetype) {}
564 static inline void clear_page_guard(struct zone *zone, struct page *page,
565 unsigned int order, int migratetype) {}
568 static inline void set_page_order(struct page *page, unsigned int order)
570 set_page_private(page, order);
571 __SetPageBuddy(page);
574 static inline void rmv_page_order(struct page *page)
576 __ClearPageBuddy(page);
577 set_page_private(page, 0);
581 * This function checks whether a page is free && is the buddy
582 * we can do coalesce a page and its buddy if
583 * (a) the buddy is not in a hole &&
584 * (b) the buddy is in the buddy system &&
585 * (c) a page and its buddy have the same order &&
586 * (d) a page and its buddy are in the same zone.
588 * For recording whether a page is in the buddy system, we set ->_mapcount
589 * PAGE_BUDDY_MAPCOUNT_VALUE.
590 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
591 * serialized by zone->lock.
593 * For recording page's order, we use page_private(page).
595 static inline int page_is_buddy(struct page *page, struct page *buddy,
598 if (!pfn_valid_within(page_to_pfn(buddy)))
601 if (page_is_guard(buddy) && page_order(buddy) == order) {
602 if (page_zone_id(page) != page_zone_id(buddy))
605 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
610 if (PageBuddy(buddy) && page_order(buddy) == order) {
612 * zone check is done late to avoid uselessly
613 * calculating zone/node ids for pages that could
616 if (page_zone_id(page) != page_zone_id(buddy))
619 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
627 * Freeing function for a buddy system allocator.
629 * The concept of a buddy system is to maintain direct-mapped table
630 * (containing bit values) for memory blocks of various "orders".
631 * The bottom level table contains the map for the smallest allocatable
632 * units of memory (here, pages), and each level above it describes
633 * pairs of units from the levels below, hence, "buddies".
634 * At a high level, all that happens here is marking the table entry
635 * at the bottom level available, and propagating the changes upward
636 * as necessary, plus some accounting needed to play nicely with other
637 * parts of the VM system.
638 * At each level, we keep a list of pages, which are heads of continuous
639 * free pages of length of (1 << order) and marked with _mapcount
640 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
642 * So when we are allocating or freeing one, we can derive the state of the
643 * other. That is, if we allocate a small block, and both were
644 * free, the remainder of the region must be split into blocks.
645 * If a block is freed, and its buddy is also free, then this
646 * triggers coalescing into a block of larger size.
651 static inline void __free_one_page(struct page *page,
653 struct zone *zone, unsigned int order,
656 unsigned long page_idx;
657 unsigned long combined_idx;
658 unsigned long uninitialized_var(buddy_idx);
660 unsigned int max_order;
662 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
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 (likely(!is_migrate_isolate(migratetype)))
669 __mod_zone_freepage_state(zone, 1 << order, migratetype);
671 page_idx = pfn & ((1 << MAX_ORDER) - 1);
673 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
674 VM_BUG_ON_PAGE(bad_range(zone, page), page);
677 while (order < max_order - 1) {
678 buddy_idx = __find_buddy_index(page_idx, order);
679 buddy = page + (buddy_idx - page_idx);
680 if (!page_is_buddy(page, buddy, order))
683 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
684 * merge with it and move up one order.
686 if (page_is_guard(buddy)) {
687 clear_page_guard(zone, buddy, order, migratetype);
689 list_del(&buddy->lru);
690 zone->free_area[order].nr_free--;
691 rmv_page_order(buddy);
693 combined_idx = buddy_idx & page_idx;
694 page = page + (combined_idx - page_idx);
695 page_idx = combined_idx;
698 if (max_order < MAX_ORDER) {
699 /* If we are here, it means order is >= pageblock_order.
700 * We want to prevent merge between freepages on isolate
701 * pageblock and normal pageblock. Without this, pageblock
702 * isolation could cause incorrect freepage or CMA accounting.
704 * We don't want to hit this code for the more frequent
707 if (unlikely(has_isolate_pageblock(zone))) {
710 buddy_idx = __find_buddy_index(page_idx, order);
711 buddy = page + (buddy_idx - page_idx);
712 buddy_mt = get_pageblock_migratetype(buddy);
714 if (migratetype != buddy_mt
715 && (is_migrate_isolate(migratetype) ||
716 is_migrate_isolate(buddy_mt)))
720 goto continue_merging;
724 set_page_order(page, order);
727 * If this is not the largest possible page, check if the buddy
728 * of the next-highest order is free. If it is, it's possible
729 * that pages are being freed that will coalesce soon. In case,
730 * that is happening, add the free page to the tail of the list
731 * so it's less likely to be used soon and more likely to be merged
732 * as a higher order page
734 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
735 struct page *higher_page, *higher_buddy;
736 combined_idx = buddy_idx & page_idx;
737 higher_page = page + (combined_idx - page_idx);
738 buddy_idx = __find_buddy_index(combined_idx, order + 1);
739 higher_buddy = higher_page + (buddy_idx - combined_idx);
740 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
741 list_add_tail(&page->lru,
742 &zone->free_area[order].free_list[migratetype]);
747 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
749 zone->free_area[order].nr_free++;
752 static inline int free_pages_check(struct page *page)
754 const char *bad_reason = NULL;
755 unsigned long bad_flags = 0;
757 if (unlikely(page_mapcount(page)))
758 bad_reason = "nonzero mapcount";
759 if (unlikely(page->mapping != NULL))
760 bad_reason = "non-NULL mapping";
761 if (unlikely(atomic_read(&page->_count) != 0))
762 bad_reason = "nonzero _count";
763 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
764 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
765 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
768 if (unlikely(page->mem_cgroup))
769 bad_reason = "page still charged to cgroup";
771 if (unlikely(bad_reason)) {
772 bad_page(page, bad_reason, bad_flags);
775 page_cpupid_reset_last(page);
776 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
777 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
782 * Frees a number of pages from the PCP lists
783 * Assumes all pages on list are in same zone, and of same order.
784 * count is the number of pages to free.
786 * If the zone was previously in an "all pages pinned" state then look to
787 * see if this freeing clears that state.
789 * And clear the zone's pages_scanned counter, to hold off the "all pages are
790 * pinned" detection logic.
792 static void free_pcppages_bulk(struct zone *zone, int count,
793 struct per_cpu_pages *pcp)
798 unsigned long nr_scanned;
800 spin_lock(&zone->lock);
801 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
803 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
807 struct list_head *list;
810 * Remove pages from lists in a round-robin fashion. A
811 * batch_free count is maintained that is incremented when an
812 * empty list is encountered. This is so more pages are freed
813 * off fuller lists instead of spinning excessively around empty
818 if (++migratetype == MIGRATE_PCPTYPES)
820 list = &pcp->lists[migratetype];
821 } while (list_empty(list));
823 /* This is the only non-empty list. Free them all. */
824 if (batch_free == MIGRATE_PCPTYPES)
825 batch_free = to_free;
828 int mt; /* migratetype of the to-be-freed page */
830 page = list_entry(list->prev, struct page, lru);
831 /* must delete as __free_one_page list manipulates */
832 list_del(&page->lru);
834 mt = get_pcppage_migratetype(page);
835 /* MIGRATE_ISOLATE page should not go to pcplists */
836 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
837 /* Pageblock could have been isolated meanwhile */
838 if (unlikely(has_isolate_pageblock(zone)))
839 mt = get_pageblock_migratetype(page);
841 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
842 trace_mm_page_pcpu_drain(page, 0, mt);
843 } while (--to_free && --batch_free && !list_empty(list));
845 spin_unlock(&zone->lock);
848 static void free_one_page(struct zone *zone,
849 struct page *page, unsigned long pfn,
853 unsigned long nr_scanned;
854 spin_lock(&zone->lock);
855 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
857 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
859 if (unlikely(has_isolate_pageblock(zone) ||
860 is_migrate_isolate(migratetype))) {
861 migratetype = get_pfnblock_migratetype(page, pfn);
863 __free_one_page(page, pfn, zone, order, migratetype);
864 spin_unlock(&zone->lock);
867 static int free_tail_pages_check(struct page *head_page, struct page *page)
872 * We rely page->lru.next never has bit 0 set, unless the page
873 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
875 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
877 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
881 if (unlikely(!PageTail(page))) {
882 bad_page(page, "PageTail not set", 0);
885 if (unlikely(compound_head(page) != head_page)) {
886 bad_page(page, "compound_head not consistent", 0);
891 clear_compound_head(page);
895 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
896 unsigned long zone, int nid)
898 set_page_links(page, zone, nid, pfn);
899 init_page_count(page);
900 page_mapcount_reset(page);
901 page_cpupid_reset_last(page);
903 INIT_LIST_HEAD(&page->lru);
904 #ifdef WANT_PAGE_VIRTUAL
905 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
906 if (!is_highmem_idx(zone))
907 set_page_address(page, __va(pfn << PAGE_SHIFT));
911 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
914 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
917 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
918 static void init_reserved_page(unsigned long pfn)
923 if (!early_page_uninitialised(pfn))
926 nid = early_pfn_to_nid(pfn);
927 pgdat = NODE_DATA(nid);
929 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
930 struct zone *zone = &pgdat->node_zones[zid];
932 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
935 __init_single_pfn(pfn, zid, nid);
938 static inline void init_reserved_page(unsigned long pfn)
941 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
944 * Initialised pages do not have PageReserved set. This function is
945 * called for each range allocated by the bootmem allocator and
946 * marks the pages PageReserved. The remaining valid pages are later
947 * sent to the buddy page allocator.
949 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
951 unsigned long start_pfn = PFN_DOWN(start);
952 unsigned long end_pfn = PFN_UP(end);
954 for (; start_pfn < end_pfn; start_pfn++) {
955 if (pfn_valid(start_pfn)) {
956 struct page *page = pfn_to_page(start_pfn);
958 init_reserved_page(start_pfn);
960 /* Avoid false-positive PageTail() */
961 INIT_LIST_HEAD(&page->lru);
963 SetPageReserved(page);
968 static bool free_pages_prepare(struct page *page, unsigned int order)
970 bool compound = PageCompound(page);
973 VM_BUG_ON_PAGE(PageTail(page), page);
974 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
976 trace_mm_page_free(page, order);
977 kmemcheck_free_shadow(page, order);
978 kasan_free_pages(page, order);
981 page->mapping = NULL;
982 bad += free_pages_check(page);
983 for (i = 1; i < (1 << order); i++) {
985 bad += free_tail_pages_check(page, page + i);
986 bad += free_pages_check(page + i);
991 reset_page_owner(page, order);
993 if (!PageHighMem(page)) {
994 debug_check_no_locks_freed(page_address(page),
996 debug_check_no_obj_freed(page_address(page),
999 arch_free_page(page, order);
1000 kernel_map_pages(page, 1 << order, 0);
1005 static void __free_pages_ok(struct page *page, unsigned int order)
1007 unsigned long flags;
1009 unsigned long pfn = page_to_pfn(page);
1011 if (!free_pages_prepare(page, order))
1014 migratetype = get_pfnblock_migratetype(page, pfn);
1015 local_irq_save(flags);
1016 __count_vm_events(PGFREE, 1 << order);
1017 free_one_page(page_zone(page), page, pfn, order, migratetype);
1018 local_irq_restore(flags);
1021 static void __init __free_pages_boot_core(struct page *page,
1022 unsigned long pfn, unsigned int order)
1024 unsigned int nr_pages = 1 << order;
1025 struct page *p = page;
1029 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1031 __ClearPageReserved(p);
1032 set_page_count(p, 0);
1034 __ClearPageReserved(p);
1035 set_page_count(p, 0);
1037 page_zone(page)->managed_pages += nr_pages;
1038 set_page_refcounted(page);
1039 __free_pages(page, order);
1042 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1043 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1045 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1047 int __meminit early_pfn_to_nid(unsigned long pfn)
1049 static DEFINE_SPINLOCK(early_pfn_lock);
1052 spin_lock(&early_pfn_lock);
1053 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1055 nid = first_online_node;
1056 spin_unlock(&early_pfn_lock);
1062 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1063 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1064 struct mminit_pfnnid_cache *state)
1068 nid = __early_pfn_to_nid(pfn, state);
1069 if (nid >= 0 && nid != node)
1074 /* Only safe to use early in boot when initialisation is single-threaded */
1075 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1077 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1082 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1086 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1087 struct mminit_pfnnid_cache *state)
1094 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1097 if (early_page_uninitialised(pfn))
1099 return __free_pages_boot_core(page, pfn, order);
1102 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1103 static void __init deferred_free_range(struct page *page,
1104 unsigned long pfn, int nr_pages)
1111 /* Free a large naturally-aligned chunk if possible */
1112 if (nr_pages == MAX_ORDER_NR_PAGES &&
1113 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1114 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1115 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1119 for (i = 0; i < nr_pages; i++, page++, pfn++)
1120 __free_pages_boot_core(page, pfn, 0);
1123 /* Completion tracking for deferred_init_memmap() threads */
1124 static atomic_t pgdat_init_n_undone __initdata;
1125 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1127 static inline void __init pgdat_init_report_one_done(void)
1129 if (atomic_dec_and_test(&pgdat_init_n_undone))
1130 complete(&pgdat_init_all_done_comp);
1133 /* Initialise remaining memory on a node */
1134 static int __init deferred_init_memmap(void *data)
1136 pg_data_t *pgdat = data;
1137 int nid = pgdat->node_id;
1138 struct mminit_pfnnid_cache nid_init_state = { };
1139 unsigned long start = jiffies;
1140 unsigned long nr_pages = 0;
1141 unsigned long walk_start, walk_end;
1144 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1145 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1147 if (first_init_pfn == ULONG_MAX) {
1148 pgdat_init_report_one_done();
1152 /* Bind memory initialisation thread to a local node if possible */
1153 if (!cpumask_empty(cpumask))
1154 set_cpus_allowed_ptr(current, cpumask);
1156 /* Sanity check boundaries */
1157 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1158 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1159 pgdat->first_deferred_pfn = ULONG_MAX;
1161 /* Only the highest zone is deferred so find it */
1162 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1163 zone = pgdat->node_zones + zid;
1164 if (first_init_pfn < zone_end_pfn(zone))
1168 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1169 unsigned long pfn, end_pfn;
1170 struct page *page = NULL;
1171 struct page *free_base_page = NULL;
1172 unsigned long free_base_pfn = 0;
1175 end_pfn = min(walk_end, zone_end_pfn(zone));
1176 pfn = first_init_pfn;
1177 if (pfn < walk_start)
1179 if (pfn < zone->zone_start_pfn)
1180 pfn = zone->zone_start_pfn;
1182 for (; pfn < end_pfn; pfn++) {
1183 if (!pfn_valid_within(pfn))
1187 * Ensure pfn_valid is checked every
1188 * MAX_ORDER_NR_PAGES for memory holes
1190 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1191 if (!pfn_valid(pfn)) {
1197 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1202 /* Minimise pfn page lookups and scheduler checks */
1203 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1206 nr_pages += nr_to_free;
1207 deferred_free_range(free_base_page,
1208 free_base_pfn, nr_to_free);
1209 free_base_page = NULL;
1210 free_base_pfn = nr_to_free = 0;
1212 page = pfn_to_page(pfn);
1217 VM_BUG_ON(page_zone(page) != zone);
1221 __init_single_page(page, pfn, zid, nid);
1222 if (!free_base_page) {
1223 free_base_page = page;
1224 free_base_pfn = pfn;
1229 /* Where possible, batch up pages for a single free */
1232 /* Free the current block of pages to allocator */
1233 nr_pages += nr_to_free;
1234 deferred_free_range(free_base_page, free_base_pfn,
1236 free_base_page = NULL;
1237 free_base_pfn = nr_to_free = 0;
1240 first_init_pfn = max(end_pfn, first_init_pfn);
1243 /* Sanity check that the next zone really is unpopulated */
1244 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1246 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1247 jiffies_to_msecs(jiffies - start));
1249 pgdat_init_report_one_done();
1253 void __init page_alloc_init_late(void)
1257 /* There will be num_node_state(N_MEMORY) threads */
1258 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1259 for_each_node_state(nid, N_MEMORY) {
1260 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1263 /* Block until all are initialised */
1264 wait_for_completion(&pgdat_init_all_done_comp);
1266 /* Reinit limits that are based on free pages after the kernel is up */
1267 files_maxfiles_init();
1269 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1272 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1273 void __init init_cma_reserved_pageblock(struct page *page)
1275 unsigned i = pageblock_nr_pages;
1276 struct page *p = page;
1279 __ClearPageReserved(p);
1280 set_page_count(p, 0);
1283 set_pageblock_migratetype(page, MIGRATE_CMA);
1285 if (pageblock_order >= MAX_ORDER) {
1286 i = pageblock_nr_pages;
1289 set_page_refcounted(p);
1290 __free_pages(p, MAX_ORDER - 1);
1291 p += MAX_ORDER_NR_PAGES;
1292 } while (i -= MAX_ORDER_NR_PAGES);
1294 set_page_refcounted(page);
1295 __free_pages(page, pageblock_order);
1298 adjust_managed_page_count(page, pageblock_nr_pages);
1303 * The order of subdivision here is critical for the IO subsystem.
1304 * Please do not alter this order without good reasons and regression
1305 * testing. Specifically, as large blocks of memory are subdivided,
1306 * the order in which smaller blocks are delivered depends on the order
1307 * they're subdivided in this function. This is the primary factor
1308 * influencing the order in which pages are delivered to the IO
1309 * subsystem according to empirical testing, and this is also justified
1310 * by considering the behavior of a buddy system containing a single
1311 * large block of memory acted on by a series of small allocations.
1312 * This behavior is a critical factor in sglist merging's success.
1316 static inline void expand(struct zone *zone, struct page *page,
1317 int low, int high, struct free_area *area,
1320 unsigned long size = 1 << high;
1322 while (high > low) {
1326 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1328 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1329 debug_guardpage_enabled() &&
1330 high < debug_guardpage_minorder()) {
1332 * Mark as guard pages (or page), that will allow to
1333 * merge back to allocator when buddy will be freed.
1334 * Corresponding page table entries will not be touched,
1335 * pages will stay not present in virtual address space
1337 set_page_guard(zone, &page[size], high, migratetype);
1340 list_add(&page[size].lru, &area->free_list[migratetype]);
1342 set_page_order(&page[size], high);
1347 * This page is about to be returned from the page allocator
1349 static inline int check_new_page(struct page *page)
1351 const char *bad_reason = NULL;
1352 unsigned long bad_flags = 0;
1354 if (unlikely(page_mapcount(page)))
1355 bad_reason = "nonzero mapcount";
1356 if (unlikely(page->mapping != NULL))
1357 bad_reason = "non-NULL mapping";
1358 if (unlikely(atomic_read(&page->_count) != 0))
1359 bad_reason = "nonzero _count";
1360 if (unlikely(page->flags & __PG_HWPOISON)) {
1361 bad_reason = "HWPoisoned (hardware-corrupted)";
1362 bad_flags = __PG_HWPOISON;
1364 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1365 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1366 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1369 if (unlikely(page->mem_cgroup))
1370 bad_reason = "page still charged to cgroup";
1372 if (unlikely(bad_reason)) {
1373 bad_page(page, bad_reason, bad_flags);
1379 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1384 for (i = 0; i < (1 << order); i++) {
1385 struct page *p = page + i;
1386 if (unlikely(check_new_page(p)))
1390 set_page_private(page, 0);
1391 set_page_refcounted(page);
1393 arch_alloc_page(page, order);
1394 kernel_map_pages(page, 1 << order, 1);
1395 kasan_alloc_pages(page, order);
1397 if (gfp_flags & __GFP_ZERO)
1398 for (i = 0; i < (1 << order); i++)
1399 clear_highpage(page + i);
1401 if (order && (gfp_flags & __GFP_COMP))
1402 prep_compound_page(page, order);
1404 set_page_owner(page, order, gfp_flags);
1407 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1408 * allocate the page. The expectation is that the caller is taking
1409 * steps that will free more memory. The caller should avoid the page
1410 * being used for !PFMEMALLOC purposes.
1412 if (alloc_flags & ALLOC_NO_WATERMARKS)
1413 set_page_pfmemalloc(page);
1415 clear_page_pfmemalloc(page);
1421 * Go through the free lists for the given migratetype and remove
1422 * the smallest available page from the freelists
1425 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1428 unsigned int current_order;
1429 struct free_area *area;
1432 /* Find a page of the appropriate size in the preferred list */
1433 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1434 area = &(zone->free_area[current_order]);
1435 if (list_empty(&area->free_list[migratetype]))
1438 page = list_entry(area->free_list[migratetype].next,
1440 list_del(&page->lru);
1441 rmv_page_order(page);
1443 expand(zone, page, order, current_order, area, migratetype);
1444 set_pcppage_migratetype(page, migratetype);
1453 * This array describes the order lists are fallen back to when
1454 * the free lists for the desirable migrate type are depleted
1456 static int fallbacks[MIGRATE_TYPES][4] = {
1457 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1458 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1459 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1461 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1463 #ifdef CONFIG_MEMORY_ISOLATION
1464 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1469 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1472 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1475 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1476 unsigned int order) { return NULL; }
1480 * Move the free pages in a range to the free lists of the requested type.
1481 * Note that start_page and end_pages are not aligned on a pageblock
1482 * boundary. If alignment is required, use move_freepages_block()
1484 int move_freepages(struct zone *zone,
1485 struct page *start_page, struct page *end_page,
1490 int pages_moved = 0;
1492 #ifndef CONFIG_HOLES_IN_ZONE
1494 * page_zone is not safe to call in this context when
1495 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1496 * anyway as we check zone boundaries in move_freepages_block().
1497 * Remove at a later date when no bug reports exist related to
1498 * grouping pages by mobility
1500 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1503 for (page = start_page; page <= end_page;) {
1504 /* Make sure we are not inadvertently changing nodes */
1505 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1507 if (!pfn_valid_within(page_to_pfn(page))) {
1512 if (!PageBuddy(page)) {
1517 order = page_order(page);
1518 list_move(&page->lru,
1519 &zone->free_area[order].free_list[migratetype]);
1521 pages_moved += 1 << order;
1527 int move_freepages_block(struct zone *zone, struct page *page,
1530 unsigned long start_pfn, end_pfn;
1531 struct page *start_page, *end_page;
1533 start_pfn = page_to_pfn(page);
1534 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1535 start_page = pfn_to_page(start_pfn);
1536 end_page = start_page + pageblock_nr_pages - 1;
1537 end_pfn = start_pfn + pageblock_nr_pages - 1;
1539 /* Do not cross zone boundaries */
1540 if (!zone_spans_pfn(zone, start_pfn))
1542 if (!zone_spans_pfn(zone, end_pfn))
1545 return move_freepages(zone, start_page, end_page, migratetype);
1548 static void change_pageblock_range(struct page *pageblock_page,
1549 int start_order, int migratetype)
1551 int nr_pageblocks = 1 << (start_order - pageblock_order);
1553 while (nr_pageblocks--) {
1554 set_pageblock_migratetype(pageblock_page, migratetype);
1555 pageblock_page += pageblock_nr_pages;
1560 * When we are falling back to another migratetype during allocation, try to
1561 * steal extra free pages from the same pageblocks to satisfy further
1562 * allocations, instead of polluting multiple pageblocks.
1564 * If we are stealing a relatively large buddy page, it is likely there will
1565 * be more free pages in the pageblock, so try to steal them all. For
1566 * reclaimable and unmovable allocations, we steal regardless of page size,
1567 * as fragmentation caused by those allocations polluting movable pageblocks
1568 * is worse than movable allocations stealing from unmovable and reclaimable
1571 static bool can_steal_fallback(unsigned int order, int start_mt)
1574 * Leaving this order check is intended, although there is
1575 * relaxed order check in next check. The reason is that
1576 * we can actually steal whole pageblock if this condition met,
1577 * but, below check doesn't guarantee it and that is just heuristic
1578 * so could be changed anytime.
1580 if (order >= pageblock_order)
1583 if (order >= pageblock_order / 2 ||
1584 start_mt == MIGRATE_RECLAIMABLE ||
1585 start_mt == MIGRATE_UNMOVABLE ||
1586 page_group_by_mobility_disabled)
1593 * This function implements actual steal behaviour. If order is large enough,
1594 * we can steal whole pageblock. If not, we first move freepages in this
1595 * pageblock and check whether half of pages are moved or not. If half of
1596 * pages are moved, we can change migratetype of pageblock and permanently
1597 * use it's pages as requested migratetype in the future.
1599 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1602 unsigned int current_order = page_order(page);
1605 /* Take ownership for orders >= pageblock_order */
1606 if (current_order >= pageblock_order) {
1607 change_pageblock_range(page, current_order, start_type);
1611 pages = move_freepages_block(zone, page, start_type);
1613 /* Claim the whole block if over half of it is free */
1614 if (pages >= (1 << (pageblock_order-1)) ||
1615 page_group_by_mobility_disabled)
1616 set_pageblock_migratetype(page, start_type);
1620 * Check whether there is a suitable fallback freepage with requested order.
1621 * If only_stealable is true, this function returns fallback_mt only if
1622 * we can steal other freepages all together. This would help to reduce
1623 * fragmentation due to mixed migratetype pages in one pageblock.
1625 int find_suitable_fallback(struct free_area *area, unsigned int order,
1626 int migratetype, bool only_stealable, bool *can_steal)
1631 if (area->nr_free == 0)
1636 fallback_mt = fallbacks[migratetype][i];
1637 if (fallback_mt == MIGRATE_TYPES)
1640 if (list_empty(&area->free_list[fallback_mt]))
1643 if (can_steal_fallback(order, migratetype))
1646 if (!only_stealable)
1657 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1658 * there are no empty page blocks that contain a page with a suitable order
1660 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1661 unsigned int alloc_order)
1664 unsigned long max_managed, flags;
1667 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1668 * Check is race-prone but harmless.
1670 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1671 if (zone->nr_reserved_highatomic >= max_managed)
1674 spin_lock_irqsave(&zone->lock, flags);
1676 /* Recheck the nr_reserved_highatomic limit under the lock */
1677 if (zone->nr_reserved_highatomic >= max_managed)
1681 mt = get_pageblock_migratetype(page);
1682 if (mt != MIGRATE_HIGHATOMIC &&
1683 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1684 zone->nr_reserved_highatomic += pageblock_nr_pages;
1685 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1686 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1690 spin_unlock_irqrestore(&zone->lock, flags);
1694 * Used when an allocation is about to fail under memory pressure. This
1695 * potentially hurts the reliability of high-order allocations when under
1696 * intense memory pressure but failed atomic allocations should be easier
1697 * to recover from than an OOM.
1699 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1701 struct zonelist *zonelist = ac->zonelist;
1702 unsigned long flags;
1708 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1710 /* Preserve at least one pageblock */
1711 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1714 spin_lock_irqsave(&zone->lock, flags);
1715 for (order = 0; order < MAX_ORDER; order++) {
1716 struct free_area *area = &(zone->free_area[order]);
1718 if (list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
1721 page = list_entry(area->free_list[MIGRATE_HIGHATOMIC].next,
1725 * It should never happen but changes to locking could
1726 * inadvertently allow a per-cpu drain to add pages
1727 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1728 * and watch for underflows.
1730 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
1731 zone->nr_reserved_highatomic);
1734 * Convert to ac->migratetype and avoid the normal
1735 * pageblock stealing heuristics. Minimally, the caller
1736 * is doing the work and needs the pages. More
1737 * importantly, if the block was always converted to
1738 * MIGRATE_UNMOVABLE or another type then the number
1739 * of pageblocks that cannot be completely freed
1742 set_pageblock_migratetype(page, ac->migratetype);
1743 move_freepages_block(zone, page, ac->migratetype);
1744 spin_unlock_irqrestore(&zone->lock, flags);
1747 spin_unlock_irqrestore(&zone->lock, flags);
1751 /* Remove an element from the buddy allocator from the fallback list */
1752 static inline struct page *
1753 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1755 struct free_area *area;
1756 unsigned int current_order;
1761 /* Find the largest possible block of pages in the other list */
1762 for (current_order = MAX_ORDER-1;
1763 current_order >= order && current_order <= MAX_ORDER-1;
1765 area = &(zone->free_area[current_order]);
1766 fallback_mt = find_suitable_fallback(area, current_order,
1767 start_migratetype, false, &can_steal);
1768 if (fallback_mt == -1)
1771 page = list_entry(area->free_list[fallback_mt].next,
1774 steal_suitable_fallback(zone, page, start_migratetype);
1776 /* Remove the page from the freelists */
1778 list_del(&page->lru);
1779 rmv_page_order(page);
1781 expand(zone, page, order, current_order, area,
1784 * The pcppage_migratetype may differ from pageblock's
1785 * migratetype depending on the decisions in
1786 * find_suitable_fallback(). This is OK as long as it does not
1787 * differ for MIGRATE_CMA pageblocks. Those can be used as
1788 * fallback only via special __rmqueue_cma_fallback() function
1790 set_pcppage_migratetype(page, start_migratetype);
1792 trace_mm_page_alloc_extfrag(page, order, current_order,
1793 start_migratetype, fallback_mt);
1802 * Do the hard work of removing an element from the buddy allocator.
1803 * Call me with the zone->lock already held.
1805 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1806 int migratetype, gfp_t gfp_flags)
1810 page = __rmqueue_smallest(zone, order, migratetype);
1811 if (unlikely(!page)) {
1812 if (migratetype == MIGRATE_MOVABLE)
1813 page = __rmqueue_cma_fallback(zone, order);
1816 page = __rmqueue_fallback(zone, order, migratetype);
1819 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1824 * Obtain a specified number of elements from the buddy allocator, all under
1825 * a single hold of the lock, for efficiency. Add them to the supplied list.
1826 * Returns the number of new pages which were placed at *list.
1828 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1829 unsigned long count, struct list_head *list,
1830 int migratetype, bool cold)
1834 spin_lock(&zone->lock);
1835 for (i = 0; i < count; ++i) {
1836 struct page *page = __rmqueue(zone, order, migratetype, 0);
1837 if (unlikely(page == NULL))
1841 * Split buddy pages returned by expand() are received here
1842 * in physical page order. The page is added to the callers and
1843 * list and the list head then moves forward. From the callers
1844 * perspective, the linked list is ordered by page number in
1845 * some conditions. This is useful for IO devices that can
1846 * merge IO requests if the physical pages are ordered
1850 list_add(&page->lru, list);
1852 list_add_tail(&page->lru, list);
1854 if (is_migrate_cma(get_pcppage_migratetype(page)))
1855 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1858 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1859 spin_unlock(&zone->lock);
1865 * Called from the vmstat counter updater to drain pagesets of this
1866 * currently executing processor on remote nodes after they have
1869 * Note that this function must be called with the thread pinned to
1870 * a single processor.
1872 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1874 unsigned long flags;
1875 int to_drain, batch;
1877 local_irq_save(flags);
1878 batch = READ_ONCE(pcp->batch);
1879 to_drain = min(pcp->count, batch);
1881 free_pcppages_bulk(zone, to_drain, pcp);
1882 pcp->count -= to_drain;
1884 local_irq_restore(flags);
1889 * Drain pcplists of the indicated processor and zone.
1891 * The processor must either be the current processor and the
1892 * thread pinned to the current processor or a processor that
1895 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1897 unsigned long flags;
1898 struct per_cpu_pageset *pset;
1899 struct per_cpu_pages *pcp;
1901 local_irq_save(flags);
1902 pset = per_cpu_ptr(zone->pageset, cpu);
1906 free_pcppages_bulk(zone, pcp->count, pcp);
1909 local_irq_restore(flags);
1913 * Drain pcplists of all zones on the indicated processor.
1915 * The processor must either be the current processor and the
1916 * thread pinned to the current processor or a processor that
1919 static void drain_pages(unsigned int cpu)
1923 for_each_populated_zone(zone) {
1924 drain_pages_zone(cpu, zone);
1929 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1931 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1932 * the single zone's pages.
1934 void drain_local_pages(struct zone *zone)
1936 int cpu = smp_processor_id();
1939 drain_pages_zone(cpu, zone);
1945 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1947 * When zone parameter is non-NULL, spill just the single zone's pages.
1949 * Note that this code is protected against sending an IPI to an offline
1950 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1951 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1952 * nothing keeps CPUs from showing up after we populated the cpumask and
1953 * before the call to on_each_cpu_mask().
1955 void drain_all_pages(struct zone *zone)
1960 * Allocate in the BSS so we wont require allocation in
1961 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1963 static cpumask_t cpus_with_pcps;
1966 * We don't care about racing with CPU hotplug event
1967 * as offline notification will cause the notified
1968 * cpu to drain that CPU pcps and on_each_cpu_mask
1969 * disables preemption as part of its processing
1971 for_each_online_cpu(cpu) {
1972 struct per_cpu_pageset *pcp;
1974 bool has_pcps = false;
1977 pcp = per_cpu_ptr(zone->pageset, cpu);
1981 for_each_populated_zone(z) {
1982 pcp = per_cpu_ptr(z->pageset, cpu);
1983 if (pcp->pcp.count) {
1991 cpumask_set_cpu(cpu, &cpus_with_pcps);
1993 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1995 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
1999 #ifdef CONFIG_HIBERNATION
2001 void mark_free_pages(struct zone *zone)
2003 unsigned long pfn, max_zone_pfn;
2004 unsigned long flags;
2005 unsigned int order, t;
2006 struct list_head *curr;
2008 if (zone_is_empty(zone))
2011 spin_lock_irqsave(&zone->lock, flags);
2013 max_zone_pfn = zone_end_pfn(zone);
2014 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2015 if (pfn_valid(pfn)) {
2016 struct page *page = pfn_to_page(pfn);
2018 if (!swsusp_page_is_forbidden(page))
2019 swsusp_unset_page_free(page);
2022 for_each_migratetype_order(order, t) {
2023 list_for_each(curr, &zone->free_area[order].free_list[t]) {
2026 pfn = page_to_pfn(list_entry(curr, struct page, lru));
2027 for (i = 0; i < (1UL << order); i++)
2028 swsusp_set_page_free(pfn_to_page(pfn + i));
2031 spin_unlock_irqrestore(&zone->lock, flags);
2033 #endif /* CONFIG_PM */
2036 * Free a 0-order page
2037 * cold == true ? free a cold page : free a hot page
2039 void free_hot_cold_page(struct page *page, bool cold)
2041 struct zone *zone = page_zone(page);
2042 struct per_cpu_pages *pcp;
2043 unsigned long flags;
2044 unsigned long pfn = page_to_pfn(page);
2047 if (!free_pages_prepare(page, 0))
2050 migratetype = get_pfnblock_migratetype(page, pfn);
2051 set_pcppage_migratetype(page, migratetype);
2052 local_irq_save(flags);
2053 __count_vm_event(PGFREE);
2056 * We only track unmovable, reclaimable and movable on pcp lists.
2057 * Free ISOLATE pages back to the allocator because they are being
2058 * offlined but treat RESERVE as movable pages so we can get those
2059 * areas back if necessary. Otherwise, we may have to free
2060 * excessively into the page allocator
2062 if (migratetype >= MIGRATE_PCPTYPES) {
2063 if (unlikely(is_migrate_isolate(migratetype))) {
2064 free_one_page(zone, page, pfn, 0, migratetype);
2067 migratetype = MIGRATE_MOVABLE;
2070 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2072 list_add(&page->lru, &pcp->lists[migratetype]);
2074 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2076 if (pcp->count >= pcp->high) {
2077 unsigned long batch = READ_ONCE(pcp->batch);
2078 free_pcppages_bulk(zone, batch, pcp);
2079 pcp->count -= batch;
2083 local_irq_restore(flags);
2087 * Free a list of 0-order pages
2089 void free_hot_cold_page_list(struct list_head *list, bool cold)
2091 struct page *page, *next;
2093 list_for_each_entry_safe(page, next, list, lru) {
2094 trace_mm_page_free_batched(page, cold);
2095 free_hot_cold_page(page, cold);
2100 * split_page takes a non-compound higher-order page, and splits it into
2101 * n (1<<order) sub-pages: page[0..n]
2102 * Each sub-page must be freed individually.
2104 * Note: this is probably too low level an operation for use in drivers.
2105 * Please consult with lkml before using this in your driver.
2107 void split_page(struct page *page, unsigned int order)
2112 VM_BUG_ON_PAGE(PageCompound(page), page);
2113 VM_BUG_ON_PAGE(!page_count(page), page);
2115 #ifdef CONFIG_KMEMCHECK
2117 * Split shadow pages too, because free(page[0]) would
2118 * otherwise free the whole shadow.
2120 if (kmemcheck_page_is_tracked(page))
2121 split_page(virt_to_page(page[0].shadow), order);
2124 gfp_mask = get_page_owner_gfp(page);
2125 set_page_owner(page, 0, gfp_mask);
2126 for (i = 1; i < (1 << order); i++) {
2127 set_page_refcounted(page + i);
2128 set_page_owner(page + i, 0, gfp_mask);
2131 EXPORT_SYMBOL_GPL(split_page);
2133 int __isolate_free_page(struct page *page, unsigned int order)
2135 unsigned long watermark;
2139 BUG_ON(!PageBuddy(page));
2141 zone = page_zone(page);
2142 mt = get_pageblock_migratetype(page);
2144 if (!is_migrate_isolate(mt)) {
2145 /* Obey watermarks as if the page was being allocated */
2146 watermark = low_wmark_pages(zone) + (1 << order);
2147 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2150 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2153 /* Remove page from free list */
2154 list_del(&page->lru);
2155 zone->free_area[order].nr_free--;
2156 rmv_page_order(page);
2158 set_page_owner(page, order, __GFP_MOVABLE);
2160 /* Set the pageblock if the isolated page is at least a pageblock */
2161 if (order >= pageblock_order - 1) {
2162 struct page *endpage = page + (1 << order) - 1;
2163 for (; page < endpage; page += pageblock_nr_pages) {
2164 int mt = get_pageblock_migratetype(page);
2165 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2166 set_pageblock_migratetype(page,
2172 return 1UL << order;
2176 * Similar to split_page except the page is already free. As this is only
2177 * being used for migration, the migratetype of the block also changes.
2178 * As this is called with interrupts disabled, the caller is responsible
2179 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2182 * Note: this is probably too low level an operation for use in drivers.
2183 * Please consult with lkml before using this in your driver.
2185 int split_free_page(struct page *page)
2190 order = page_order(page);
2192 nr_pages = __isolate_free_page(page, order);
2196 /* Split into individual pages */
2197 set_page_refcounted(page);
2198 split_page(page, order);
2203 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2206 struct page *buffered_rmqueue(struct zone *preferred_zone,
2207 struct zone *zone, unsigned int order,
2208 gfp_t gfp_flags, int alloc_flags, int migratetype)
2210 unsigned long flags;
2212 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2214 if (likely(order == 0)) {
2215 struct per_cpu_pages *pcp;
2216 struct list_head *list;
2218 local_irq_save(flags);
2219 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2220 list = &pcp->lists[migratetype];
2221 if (list_empty(list)) {
2222 pcp->count += rmqueue_bulk(zone, 0,
2225 if (unlikely(list_empty(list)))
2230 page = list_entry(list->prev, struct page, lru);
2232 page = list_entry(list->next, struct page, lru);
2234 list_del(&page->lru);
2237 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2239 * __GFP_NOFAIL is not to be used in new code.
2241 * All __GFP_NOFAIL callers should be fixed so that they
2242 * properly detect and handle allocation failures.
2244 * We most definitely don't want callers attempting to
2245 * allocate greater than order-1 page units with
2248 WARN_ON_ONCE(order > 1);
2250 spin_lock_irqsave(&zone->lock, flags);
2253 if (alloc_flags & ALLOC_HARDER) {
2254 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2256 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2259 page = __rmqueue(zone, order, migratetype, gfp_flags);
2260 spin_unlock(&zone->lock);
2263 __mod_zone_freepage_state(zone, -(1 << order),
2264 get_pcppage_migratetype(page));
2267 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2268 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2269 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2270 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2272 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2273 zone_statistics(preferred_zone, zone, gfp_flags);
2274 local_irq_restore(flags);
2276 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2280 local_irq_restore(flags);
2284 #ifdef CONFIG_FAIL_PAGE_ALLOC
2287 struct fault_attr attr;
2289 bool ignore_gfp_highmem;
2290 bool ignore_gfp_reclaim;
2292 } fail_page_alloc = {
2293 .attr = FAULT_ATTR_INITIALIZER,
2294 .ignore_gfp_reclaim = true,
2295 .ignore_gfp_highmem = true,
2299 static int __init setup_fail_page_alloc(char *str)
2301 return setup_fault_attr(&fail_page_alloc.attr, str);
2303 __setup("fail_page_alloc=", setup_fail_page_alloc);
2305 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2307 if (order < fail_page_alloc.min_order)
2309 if (gfp_mask & __GFP_NOFAIL)
2311 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2313 if (fail_page_alloc.ignore_gfp_reclaim &&
2314 (gfp_mask & __GFP_DIRECT_RECLAIM))
2317 return should_fail(&fail_page_alloc.attr, 1 << order);
2320 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2322 static int __init fail_page_alloc_debugfs(void)
2324 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2327 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2328 &fail_page_alloc.attr);
2330 return PTR_ERR(dir);
2332 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2333 &fail_page_alloc.ignore_gfp_reclaim))
2335 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2336 &fail_page_alloc.ignore_gfp_highmem))
2338 if (!debugfs_create_u32("min-order", mode, dir,
2339 &fail_page_alloc.min_order))
2344 debugfs_remove_recursive(dir);
2349 late_initcall(fail_page_alloc_debugfs);
2351 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2353 #else /* CONFIG_FAIL_PAGE_ALLOC */
2355 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2360 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2363 * Return true if free base pages are above 'mark'. For high-order checks it
2364 * will return true of the order-0 watermark is reached and there is at least
2365 * one free page of a suitable size. Checking now avoids taking the zone lock
2366 * to check in the allocation paths if no pages are free.
2368 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2369 unsigned long mark, int classzone_idx, int alloc_flags,
2374 const int alloc_harder = (alloc_flags & ALLOC_HARDER);
2376 /* free_pages may go negative - that's OK */
2377 free_pages -= (1 << order) - 1;
2379 if (alloc_flags & ALLOC_HIGH)
2383 * If the caller does not have rights to ALLOC_HARDER then subtract
2384 * the high-atomic reserves. This will over-estimate the size of the
2385 * atomic reserve but it avoids a search.
2387 if (likely(!alloc_harder))
2388 free_pages -= z->nr_reserved_highatomic;
2393 /* If allocation can't use CMA areas don't use free CMA pages */
2394 if (!(alloc_flags & ALLOC_CMA))
2395 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2399 * Check watermarks for an order-0 allocation request. If these
2400 * are not met, then a high-order request also cannot go ahead
2401 * even if a suitable page happened to be free.
2403 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2406 /* If this is an order-0 request then the watermark is fine */
2410 /* For a high-order request, check at least one suitable page is free */
2411 for (o = order; o < MAX_ORDER; o++) {
2412 struct free_area *area = &z->free_area[o];
2421 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2422 if (!list_empty(&area->free_list[mt]))
2427 if ((alloc_flags & ALLOC_CMA) &&
2428 !list_empty(&area->free_list[MIGRATE_CMA])) {
2436 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2437 int classzone_idx, int alloc_flags)
2439 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2440 zone_page_state(z, NR_FREE_PAGES));
2443 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2444 unsigned long mark, int classzone_idx)
2446 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2448 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2449 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2451 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2456 static bool zone_local(struct zone *local_zone, struct zone *zone)
2458 return local_zone->node == zone->node;
2461 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2463 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
2466 #else /* CONFIG_NUMA */
2467 static bool zone_local(struct zone *local_zone, struct zone *zone)
2472 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2476 #endif /* CONFIG_NUMA */
2478 static void reset_alloc_batches(struct zone *preferred_zone)
2480 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2483 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2484 high_wmark_pages(zone) - low_wmark_pages(zone) -
2485 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2486 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2487 } while (zone++ != preferred_zone);
2491 * get_page_from_freelist goes through the zonelist trying to allocate
2494 static struct page *
2495 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2496 const struct alloc_context *ac)
2498 struct zonelist *zonelist = ac->zonelist;
2500 struct page *page = NULL;
2502 int nr_fair_skipped = 0;
2503 bool zonelist_rescan;
2506 zonelist_rescan = false;
2509 * Scan zonelist, looking for a zone with enough free.
2510 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2512 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2516 if (cpusets_enabled() &&
2517 (alloc_flags & ALLOC_CPUSET) &&
2518 !cpuset_zone_allowed(zone, gfp_mask))
2521 * Distribute pages in proportion to the individual
2522 * zone size to ensure fair page aging. The zone a
2523 * page was allocated in should have no effect on the
2524 * time the page has in memory before being reclaimed.
2526 if (alloc_flags & ALLOC_FAIR) {
2527 if (!zone_local(ac->preferred_zone, zone))
2529 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2535 * When allocating a page cache page for writing, we
2536 * want to get it from a zone that is within its dirty
2537 * limit, such that no single zone holds more than its
2538 * proportional share of globally allowed dirty pages.
2539 * The dirty limits take into account the zone's
2540 * lowmem reserves and high watermark so that kswapd
2541 * should be able to balance it without having to
2542 * write pages from its LRU list.
2544 * This may look like it could increase pressure on
2545 * lower zones by failing allocations in higher zones
2546 * before they are full. But the pages that do spill
2547 * over are limited as the lower zones are protected
2548 * by this very same mechanism. It should not become
2549 * a practical burden to them.
2551 * XXX: For now, allow allocations to potentially
2552 * exceed the per-zone dirty limit in the slowpath
2553 * (spread_dirty_pages unset) before going into reclaim,
2554 * which is important when on a NUMA setup the allowed
2555 * zones are together not big enough to reach the
2556 * global limit. The proper fix for these situations
2557 * will require awareness of zones in the
2558 * dirty-throttling and the flusher threads.
2560 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2563 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2564 if (!zone_watermark_ok(zone, order, mark,
2565 ac->classzone_idx, alloc_flags)) {
2568 /* Checked here to keep the fast path fast */
2569 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2570 if (alloc_flags & ALLOC_NO_WATERMARKS)
2573 if (zone_reclaim_mode == 0 ||
2574 !zone_allows_reclaim(ac->preferred_zone, zone))
2577 ret = zone_reclaim(zone, gfp_mask, order);
2579 case ZONE_RECLAIM_NOSCAN:
2582 case ZONE_RECLAIM_FULL:
2583 /* scanned but unreclaimable */
2586 /* did we reclaim enough */
2587 if (zone_watermark_ok(zone, order, mark,
2588 ac->classzone_idx, alloc_flags))
2596 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2597 gfp_mask, alloc_flags, ac->migratetype);
2599 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2603 * If this is a high-order atomic allocation then check
2604 * if the pageblock should be reserved for the future
2606 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2607 reserve_highatomic_pageblock(page, zone, order);
2614 * The first pass makes sure allocations are spread fairly within the
2615 * local node. However, the local node might have free pages left
2616 * after the fairness batches are exhausted, and remote zones haven't
2617 * even been considered yet. Try once more without fairness, and
2618 * include remote zones now, before entering the slowpath and waking
2619 * kswapd: prefer spilling to a remote zone over swapping locally.
2621 if (alloc_flags & ALLOC_FAIR) {
2622 alloc_flags &= ~ALLOC_FAIR;
2623 if (nr_fair_skipped) {
2624 zonelist_rescan = true;
2625 reset_alloc_batches(ac->preferred_zone);
2627 if (nr_online_nodes > 1)
2628 zonelist_rescan = true;
2631 if (zonelist_rescan)
2638 * Large machines with many possible nodes should not always dump per-node
2639 * meminfo in irq context.
2641 static inline bool should_suppress_show_mem(void)
2646 ret = in_interrupt();
2651 static DEFINE_RATELIMIT_STATE(nopage_rs,
2652 DEFAULT_RATELIMIT_INTERVAL,
2653 DEFAULT_RATELIMIT_BURST);
2655 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
2657 unsigned int filter = SHOW_MEM_FILTER_NODES;
2659 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2660 debug_guardpage_minorder() > 0)
2664 * This documents exceptions given to allocations in certain
2665 * contexts that are allowed to allocate outside current's set
2668 if (!(gfp_mask & __GFP_NOMEMALLOC))
2669 if (test_thread_flag(TIF_MEMDIE) ||
2670 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2671 filter &= ~SHOW_MEM_FILTER_NODES;
2672 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2673 filter &= ~SHOW_MEM_FILTER_NODES;
2676 struct va_format vaf;
2679 va_start(args, fmt);
2684 pr_warn("%pV", &vaf);
2689 pr_warn("%s: page allocation failure: order:%u, mode:0x%x\n",
2690 current->comm, order, gfp_mask);
2693 if (!should_suppress_show_mem())
2697 static inline struct page *
2698 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2699 const struct alloc_context *ac, unsigned long *did_some_progress)
2701 struct oom_control oc = {
2702 .zonelist = ac->zonelist,
2703 .nodemask = ac->nodemask,
2704 .gfp_mask = gfp_mask,
2709 *did_some_progress = 0;
2712 * Acquire the oom lock. If that fails, somebody else is
2713 * making progress for us.
2715 if (!mutex_trylock(&oom_lock)) {
2716 *did_some_progress = 1;
2717 schedule_timeout_uninterruptible(1);
2722 * Go through the zonelist yet one more time, keep very high watermark
2723 * here, this is only to catch a parallel oom killing, we must fail if
2724 * we're still under heavy pressure.
2726 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2727 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2731 if (!(gfp_mask & __GFP_NOFAIL)) {
2732 /* Coredumps can quickly deplete all memory reserves */
2733 if (current->flags & PF_DUMPCORE)
2735 /* The OOM killer will not help higher order allocs */
2736 if (order > PAGE_ALLOC_COSTLY_ORDER)
2738 /* The OOM killer does not needlessly kill tasks for lowmem */
2739 if (ac->high_zoneidx < ZONE_NORMAL)
2741 /* The OOM killer does not compensate for IO-less reclaim */
2742 if (!(gfp_mask & __GFP_FS)) {
2744 * XXX: Page reclaim didn't yield anything,
2745 * and the OOM killer can't be invoked, but
2746 * keep looping as per tradition.
2748 *did_some_progress = 1;
2751 if (pm_suspended_storage())
2753 /* The OOM killer may not free memory on a specific node */
2754 if (gfp_mask & __GFP_THISNODE)
2757 /* Exhausted what can be done so it's blamo time */
2758 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL))
2759 *did_some_progress = 1;
2761 mutex_unlock(&oom_lock);
2765 #ifdef CONFIG_COMPACTION
2766 /* Try memory compaction for high-order allocations before reclaim */
2767 static struct page *
2768 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2769 int alloc_flags, const struct alloc_context *ac,
2770 enum migrate_mode mode, int *contended_compaction,
2771 bool *deferred_compaction)
2773 unsigned long compact_result;
2779 current->flags |= PF_MEMALLOC;
2780 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2781 mode, contended_compaction);
2782 current->flags &= ~PF_MEMALLOC;
2784 switch (compact_result) {
2785 case COMPACT_DEFERRED:
2786 *deferred_compaction = true;
2788 case COMPACT_SKIPPED:
2795 * At least in one zone compaction wasn't deferred or skipped, so let's
2796 * count a compaction stall
2798 count_vm_event(COMPACTSTALL);
2800 page = get_page_from_freelist(gfp_mask, order,
2801 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2804 struct zone *zone = page_zone(page);
2806 zone->compact_blockskip_flush = false;
2807 compaction_defer_reset(zone, order, true);
2808 count_vm_event(COMPACTSUCCESS);
2813 * It's bad if compaction run occurs and fails. The most likely reason
2814 * is that pages exist, but not enough to satisfy watermarks.
2816 count_vm_event(COMPACTFAIL);
2823 static inline struct page *
2824 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2825 int alloc_flags, const struct alloc_context *ac,
2826 enum migrate_mode mode, int *contended_compaction,
2827 bool *deferred_compaction)
2831 #endif /* CONFIG_COMPACTION */
2833 /* Perform direct synchronous page reclaim */
2835 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2836 const struct alloc_context *ac)
2838 struct reclaim_state reclaim_state;
2843 /* We now go into synchronous reclaim */
2844 cpuset_memory_pressure_bump();
2845 current->flags |= PF_MEMALLOC;
2846 lockdep_set_current_reclaim_state(gfp_mask);
2847 reclaim_state.reclaimed_slab = 0;
2848 current->reclaim_state = &reclaim_state;
2850 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2853 current->reclaim_state = NULL;
2854 lockdep_clear_current_reclaim_state();
2855 current->flags &= ~PF_MEMALLOC;
2862 /* The really slow allocator path where we enter direct reclaim */
2863 static inline struct page *
2864 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2865 int alloc_flags, const struct alloc_context *ac,
2866 unsigned long *did_some_progress)
2868 struct page *page = NULL;
2869 bool drained = false;
2871 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2872 if (unlikely(!(*did_some_progress)))
2876 page = get_page_from_freelist(gfp_mask, order,
2877 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2880 * If an allocation failed after direct reclaim, it could be because
2881 * pages are pinned on the per-cpu lists or in high alloc reserves.
2882 * Shrink them them and try again
2884 if (!page && !drained) {
2885 unreserve_highatomic_pageblock(ac);
2886 drain_all_pages(NULL);
2895 * This is called in the allocator slow-path if the allocation request is of
2896 * sufficient urgency to ignore watermarks and take other desperate measures
2898 static inline struct page *
2899 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2900 const struct alloc_context *ac)
2905 page = get_page_from_freelist(gfp_mask, order,
2906 ALLOC_NO_WATERMARKS, ac);
2908 if (!page && gfp_mask & __GFP_NOFAIL)
2909 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC,
2911 } while (!page && (gfp_mask & __GFP_NOFAIL));
2916 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2921 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2922 ac->high_zoneidx, ac->nodemask)
2923 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2927 gfp_to_alloc_flags(gfp_t gfp_mask)
2929 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2931 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2932 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2935 * The caller may dip into page reserves a bit more if the caller
2936 * cannot run direct reclaim, or if the caller has realtime scheduling
2937 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2938 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
2940 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2942 if (gfp_mask & __GFP_ATOMIC) {
2944 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2945 * if it can't schedule.
2947 if (!(gfp_mask & __GFP_NOMEMALLOC))
2948 alloc_flags |= ALLOC_HARDER;
2950 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2951 * comment for __cpuset_node_allowed().
2953 alloc_flags &= ~ALLOC_CPUSET;
2954 } else if (unlikely(rt_task(current)) && !in_interrupt())
2955 alloc_flags |= ALLOC_HARDER;
2957 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2958 if (gfp_mask & __GFP_MEMALLOC)
2959 alloc_flags |= ALLOC_NO_WATERMARKS;
2960 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2961 alloc_flags |= ALLOC_NO_WATERMARKS;
2962 else if (!in_interrupt() &&
2963 ((current->flags & PF_MEMALLOC) ||
2964 unlikely(test_thread_flag(TIF_MEMDIE))))
2965 alloc_flags |= ALLOC_NO_WATERMARKS;
2968 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2969 alloc_flags |= ALLOC_CMA;
2974 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2976 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2979 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
2981 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
2984 static inline struct page *
2985 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2986 struct alloc_context *ac)
2988 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
2989 struct page *page = NULL;
2991 unsigned long pages_reclaimed = 0;
2992 unsigned long did_some_progress;
2993 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2994 bool deferred_compaction = false;
2995 int contended_compaction = COMPACT_CONTENDED_NONE;
2998 * In the slowpath, we sanity check order to avoid ever trying to
2999 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3000 * be using allocators in order of preference for an area that is
3003 if (order >= MAX_ORDER) {
3004 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3009 * We also sanity check to catch abuse of atomic reserves being used by
3010 * callers that are not in atomic context.
3012 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3013 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3014 gfp_mask &= ~__GFP_ATOMIC;
3017 * If this allocation cannot block and it is for a specific node, then
3018 * fail early. There's no need to wakeup kswapd or retry for a
3019 * speculative node-specific allocation.
3021 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !can_direct_reclaim)
3025 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3026 wake_all_kswapds(order, ac);
3029 * OK, we're below the kswapd watermark and have kicked background
3030 * reclaim. Now things get more complex, so set up alloc_flags according
3031 * to how we want to proceed.
3033 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3036 * Find the true preferred zone if the allocation is unconstrained by
3039 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3040 struct zoneref *preferred_zoneref;
3041 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3042 ac->high_zoneidx, NULL, &ac->preferred_zone);
3043 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3046 /* This is the last chance, in general, before the goto nopage. */
3047 page = get_page_from_freelist(gfp_mask, order,
3048 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3052 /* Allocate without watermarks if the context allows */
3053 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3055 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3056 * the allocation is high priority and these type of
3057 * allocations are system rather than user orientated
3059 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3061 page = __alloc_pages_high_priority(gfp_mask, order, ac);
3068 /* Caller is not willing to reclaim, we can't balance anything */
3069 if (!can_direct_reclaim) {
3071 * All existing users of the deprecated __GFP_NOFAIL are
3072 * blockable, so warn of any new users that actually allow this
3073 * type of allocation to fail.
3075 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3079 /* Avoid recursion of direct reclaim */
3080 if (current->flags & PF_MEMALLOC)
3083 /* Avoid allocations with no watermarks from looping endlessly */
3084 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3088 * Try direct compaction. The first pass is asynchronous. Subsequent
3089 * attempts after direct reclaim are synchronous
3091 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3093 &contended_compaction,
3094 &deferred_compaction);
3098 /* Checks for THP-specific high-order allocations */
3099 if (is_thp_gfp_mask(gfp_mask)) {
3101 * If compaction is deferred for high-order allocations, it is
3102 * because sync compaction recently failed. If this is the case
3103 * and the caller requested a THP allocation, we do not want
3104 * to heavily disrupt the system, so we fail the allocation
3105 * instead of entering direct reclaim.
3107 if (deferred_compaction)
3111 * In all zones where compaction was attempted (and not
3112 * deferred or skipped), lock contention has been detected.
3113 * For THP allocation we do not want to disrupt the others
3114 * so we fallback to base pages instead.
3116 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3120 * If compaction was aborted due to need_resched(), we do not
3121 * want to further increase allocation latency, unless it is
3122 * khugepaged trying to collapse.
3124 if (contended_compaction == COMPACT_CONTENDED_SCHED
3125 && !(current->flags & PF_KTHREAD))
3130 * It can become very expensive to allocate transparent hugepages at
3131 * fault, so use asynchronous memory compaction for THP unless it is
3132 * khugepaged trying to collapse.
3134 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3135 migration_mode = MIGRATE_SYNC_LIGHT;
3137 /* Try direct reclaim and then allocating */
3138 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3139 &did_some_progress);
3143 /* Do not loop if specifically requested */
3144 if (gfp_mask & __GFP_NORETRY)
3147 /* Keep reclaiming pages as long as there is reasonable progress */
3148 pages_reclaimed += did_some_progress;
3149 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3150 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3151 /* Wait for some write requests to complete then retry */
3152 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3156 /* Reclaim has failed us, start killing things */
3157 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3161 /* Retry as long as the OOM killer is making progress */
3162 if (did_some_progress)
3167 * High-order allocations do not necessarily loop after
3168 * direct reclaim and reclaim/compaction depends on compaction
3169 * being called after reclaim so call directly if necessary
3171 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3173 &contended_compaction,
3174 &deferred_compaction);
3178 warn_alloc_failed(gfp_mask, order, NULL);
3184 * This is the 'heart' of the zoned buddy allocator.
3187 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3188 struct zonelist *zonelist, nodemask_t *nodemask)
3190 struct zoneref *preferred_zoneref;
3191 struct page *page = NULL;
3192 unsigned int cpuset_mems_cookie;
3193 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3194 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3195 struct alloc_context ac = {
3196 .high_zoneidx = gfp_zone(gfp_mask),
3197 .nodemask = nodemask,
3198 .migratetype = gfpflags_to_migratetype(gfp_mask),
3201 gfp_mask &= gfp_allowed_mask;
3203 lockdep_trace_alloc(gfp_mask);
3205 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3207 if (should_fail_alloc_page(gfp_mask, order))
3211 * Check the zones suitable for the gfp_mask contain at least one
3212 * valid zone. It's possible to have an empty zonelist as a result
3213 * of __GFP_THISNODE and a memoryless node
3215 if (unlikely(!zonelist->_zonerefs->zone))
3218 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3219 alloc_flags |= ALLOC_CMA;
3222 cpuset_mems_cookie = read_mems_allowed_begin();
3224 /* We set it here, as __alloc_pages_slowpath might have changed it */
3225 ac.zonelist = zonelist;
3227 /* Dirty zone balancing only done in the fast path */
3228 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3230 /* The preferred zone is used for statistics later */
3231 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3232 ac.nodemask ? : &cpuset_current_mems_allowed,
3233 &ac.preferred_zone);
3234 if (!ac.preferred_zone)
3236 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3238 /* First allocation attempt */
3239 alloc_mask = gfp_mask|__GFP_HARDWALL;
3240 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3241 if (unlikely(!page)) {
3243 * Runtime PM, block IO and its error handling path
3244 * can deadlock because I/O on the device might not
3247 alloc_mask = memalloc_noio_flags(gfp_mask);
3248 ac.spread_dirty_pages = false;
3250 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3253 if (kmemcheck_enabled && page)
3254 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3256 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3260 * When updating a task's mems_allowed, it is possible to race with
3261 * parallel threads in such a way that an allocation can fail while
3262 * the mask is being updated. If a page allocation is about to fail,
3263 * check if the cpuset changed during allocation and if so, retry.
3265 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3270 EXPORT_SYMBOL(__alloc_pages_nodemask);
3273 * Common helper functions.
3275 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3280 * __get_free_pages() returns a 32-bit address, which cannot represent
3283 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3285 page = alloc_pages(gfp_mask, order);
3288 return (unsigned long) page_address(page);
3290 EXPORT_SYMBOL(__get_free_pages);
3292 unsigned long get_zeroed_page(gfp_t gfp_mask)
3294 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3296 EXPORT_SYMBOL(get_zeroed_page);
3298 void __free_pages(struct page *page, unsigned int order)
3300 if (put_page_testzero(page)) {
3302 free_hot_cold_page(page, false);
3304 __free_pages_ok(page, order);
3308 EXPORT_SYMBOL(__free_pages);
3310 void free_pages(unsigned long addr, unsigned int order)
3313 VM_BUG_ON(!virt_addr_valid((void *)addr));
3314 __free_pages(virt_to_page((void *)addr), order);
3318 EXPORT_SYMBOL(free_pages);
3322 * An arbitrary-length arbitrary-offset area of memory which resides
3323 * within a 0 or higher order page. Multiple fragments within that page
3324 * are individually refcounted, in the page's reference counter.
3326 * The page_frag functions below provide a simple allocation framework for
3327 * page fragments. This is used by the network stack and network device
3328 * drivers to provide a backing region of memory for use as either an
3329 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3331 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3334 struct page *page = NULL;
3335 gfp_t gfp = gfp_mask;
3337 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3338 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3340 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3341 PAGE_FRAG_CACHE_MAX_ORDER);
3342 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3344 if (unlikely(!page))
3345 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3347 nc->va = page ? page_address(page) : NULL;
3352 void *__alloc_page_frag(struct page_frag_cache *nc,
3353 unsigned int fragsz, gfp_t gfp_mask)
3355 unsigned int size = PAGE_SIZE;
3359 if (unlikely(!nc->va)) {
3361 page = __page_frag_refill(nc, gfp_mask);
3365 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3366 /* if size can vary use size else just use PAGE_SIZE */
3369 /* Even if we own the page, we do not use atomic_set().
3370 * This would break get_page_unless_zero() users.
3372 atomic_add(size - 1, &page->_count);
3374 /* reset page count bias and offset to start of new frag */
3375 nc->pfmemalloc = page_is_pfmemalloc(page);
3376 nc->pagecnt_bias = size;
3380 offset = nc->offset - fragsz;
3381 if (unlikely(offset < 0)) {
3382 page = virt_to_page(nc->va);
3384 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3387 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3388 /* if size can vary use size else just use PAGE_SIZE */
3391 /* OK, page count is 0, we can safely set it */
3392 atomic_set(&page->_count, size);
3394 /* reset page count bias and offset to start of new frag */
3395 nc->pagecnt_bias = size;
3396 offset = size - fragsz;
3400 nc->offset = offset;
3402 return nc->va + offset;
3404 EXPORT_SYMBOL(__alloc_page_frag);
3407 * Frees a page fragment allocated out of either a compound or order 0 page.
3409 void __free_page_frag(void *addr)
3411 struct page *page = virt_to_head_page(addr);
3413 if (unlikely(put_page_testzero(page)))
3414 __free_pages_ok(page, compound_order(page));
3416 EXPORT_SYMBOL(__free_page_frag);
3419 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3420 * of the current memory cgroup.
3422 * It should be used when the caller would like to use kmalloc, but since the
3423 * allocation is large, it has to fall back to the page allocator.
3425 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3429 page = alloc_pages(gfp_mask, order);
3430 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3431 __free_pages(page, order);
3437 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3441 page = alloc_pages_node(nid, gfp_mask, order);
3442 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3443 __free_pages(page, order);
3450 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3453 void __free_kmem_pages(struct page *page, unsigned int order)
3455 memcg_kmem_uncharge(page, order);
3456 __free_pages(page, order);
3459 void free_kmem_pages(unsigned long addr, unsigned int order)
3462 VM_BUG_ON(!virt_addr_valid((void *)addr));
3463 __free_kmem_pages(virt_to_page((void *)addr), order);
3467 static void *make_alloc_exact(unsigned long addr, unsigned int order,
3471 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3472 unsigned long used = addr + PAGE_ALIGN(size);
3474 split_page(virt_to_page((void *)addr), order);
3475 while (used < alloc_end) {
3480 return (void *)addr;
3484 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3485 * @size: the number of bytes to allocate
3486 * @gfp_mask: GFP flags for the allocation
3488 * This function is similar to alloc_pages(), except that it allocates the
3489 * minimum number of pages to satisfy the request. alloc_pages() can only
3490 * allocate memory in power-of-two pages.
3492 * This function is also limited by MAX_ORDER.
3494 * Memory allocated by this function must be released by free_pages_exact().
3496 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3498 unsigned int order = get_order(size);
3501 addr = __get_free_pages(gfp_mask, order);
3502 return make_alloc_exact(addr, order, size);
3504 EXPORT_SYMBOL(alloc_pages_exact);
3507 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3509 * @nid: the preferred node ID where memory should be allocated
3510 * @size: the number of bytes to allocate
3511 * @gfp_mask: GFP flags for the allocation
3513 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3516 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3518 unsigned int order = get_order(size);
3519 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3522 return make_alloc_exact((unsigned long)page_address(p), order, size);
3526 * free_pages_exact - release memory allocated via alloc_pages_exact()
3527 * @virt: the value returned by alloc_pages_exact.
3528 * @size: size of allocation, same value as passed to alloc_pages_exact().
3530 * Release the memory allocated by a previous call to alloc_pages_exact.
3532 void free_pages_exact(void *virt, size_t size)
3534 unsigned long addr = (unsigned long)virt;
3535 unsigned long end = addr + PAGE_ALIGN(size);
3537 while (addr < end) {
3542 EXPORT_SYMBOL(free_pages_exact);
3545 * nr_free_zone_pages - count number of pages beyond high watermark
3546 * @offset: The zone index of the highest zone
3548 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3549 * high watermark within all zones at or below a given zone index. For each
3550 * zone, the number of pages is calculated as:
3551 * managed_pages - high_pages
3553 static unsigned long nr_free_zone_pages(int offset)
3558 /* Just pick one node, since fallback list is circular */
3559 unsigned long sum = 0;
3561 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3563 for_each_zone_zonelist(zone, z, zonelist, offset) {
3564 unsigned long size = zone->managed_pages;
3565 unsigned long high = high_wmark_pages(zone);
3574 * nr_free_buffer_pages - count number of pages beyond high watermark
3576 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3577 * watermark within ZONE_DMA and ZONE_NORMAL.
3579 unsigned long nr_free_buffer_pages(void)
3581 return nr_free_zone_pages(gfp_zone(GFP_USER));
3583 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3586 * nr_free_pagecache_pages - count number of pages beyond high watermark
3588 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3589 * high watermark within all zones.
3591 unsigned long nr_free_pagecache_pages(void)
3593 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3596 static inline void show_node(struct zone *zone)
3598 if (IS_ENABLED(CONFIG_NUMA))
3599 printk("Node %d ", zone_to_nid(zone));
3602 void si_meminfo(struct sysinfo *val)
3604 val->totalram = totalram_pages;
3605 val->sharedram = global_page_state(NR_SHMEM);
3606 val->freeram = global_page_state(NR_FREE_PAGES);
3607 val->bufferram = nr_blockdev_pages();
3608 val->totalhigh = totalhigh_pages;
3609 val->freehigh = nr_free_highpages();
3610 val->mem_unit = PAGE_SIZE;
3613 EXPORT_SYMBOL(si_meminfo);
3616 void si_meminfo_node(struct sysinfo *val, int nid)
3618 int zone_type; /* needs to be signed */
3619 unsigned long managed_pages = 0;
3620 pg_data_t *pgdat = NODE_DATA(nid);
3622 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3623 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3624 val->totalram = managed_pages;
3625 val->sharedram = node_page_state(nid, NR_SHMEM);
3626 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3627 #ifdef CONFIG_HIGHMEM
3628 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3629 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3635 val->mem_unit = PAGE_SIZE;
3640 * Determine whether the node should be displayed or not, depending on whether
3641 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3643 bool skip_free_areas_node(unsigned int flags, int nid)
3646 unsigned int cpuset_mems_cookie;
3648 if (!(flags & SHOW_MEM_FILTER_NODES))
3652 cpuset_mems_cookie = read_mems_allowed_begin();
3653 ret = !node_isset(nid, cpuset_current_mems_allowed);
3654 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3659 #define K(x) ((x) << (PAGE_SHIFT-10))
3661 static void show_migration_types(unsigned char type)
3663 static const char types[MIGRATE_TYPES] = {
3664 [MIGRATE_UNMOVABLE] = 'U',
3665 [MIGRATE_MOVABLE] = 'M',
3666 [MIGRATE_RECLAIMABLE] = 'E',
3667 [MIGRATE_HIGHATOMIC] = 'H',
3669 [MIGRATE_CMA] = 'C',
3671 #ifdef CONFIG_MEMORY_ISOLATION
3672 [MIGRATE_ISOLATE] = 'I',
3675 char tmp[MIGRATE_TYPES + 1];
3679 for (i = 0; i < MIGRATE_TYPES; i++) {
3680 if (type & (1 << i))
3685 printk("(%s) ", tmp);
3689 * Show free area list (used inside shift_scroll-lock stuff)
3690 * We also calculate the percentage fragmentation. We do this by counting the
3691 * memory on each free list with the exception of the first item on the list.
3694 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3697 void show_free_areas(unsigned int filter)
3699 unsigned long free_pcp = 0;
3703 for_each_populated_zone(zone) {
3704 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3707 for_each_online_cpu(cpu)
3708 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3711 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3712 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3713 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3714 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3715 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3716 " free:%lu free_pcp:%lu free_cma:%lu\n",
3717 global_page_state(NR_ACTIVE_ANON),
3718 global_page_state(NR_INACTIVE_ANON),
3719 global_page_state(NR_ISOLATED_ANON),
3720 global_page_state(NR_ACTIVE_FILE),
3721 global_page_state(NR_INACTIVE_FILE),
3722 global_page_state(NR_ISOLATED_FILE),
3723 global_page_state(NR_UNEVICTABLE),
3724 global_page_state(NR_FILE_DIRTY),
3725 global_page_state(NR_WRITEBACK),
3726 global_page_state(NR_UNSTABLE_NFS),
3727 global_page_state(NR_SLAB_RECLAIMABLE),
3728 global_page_state(NR_SLAB_UNRECLAIMABLE),
3729 global_page_state(NR_FILE_MAPPED),
3730 global_page_state(NR_SHMEM),
3731 global_page_state(NR_PAGETABLE),
3732 global_page_state(NR_BOUNCE),
3733 global_page_state(NR_FREE_PAGES),
3735 global_page_state(NR_FREE_CMA_PAGES));
3737 for_each_populated_zone(zone) {
3740 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3744 for_each_online_cpu(cpu)
3745 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3753 " active_anon:%lukB"
3754 " inactive_anon:%lukB"
3755 " active_file:%lukB"
3756 " inactive_file:%lukB"
3757 " unevictable:%lukB"
3758 " isolated(anon):%lukB"
3759 " isolated(file):%lukB"
3767 " slab_reclaimable:%lukB"
3768 " slab_unreclaimable:%lukB"
3769 " kernel_stack:%lukB"
3776 " writeback_tmp:%lukB"
3777 " pages_scanned:%lu"
3778 " all_unreclaimable? %s"
3781 K(zone_page_state(zone, NR_FREE_PAGES)),
3782 K(min_wmark_pages(zone)),
3783 K(low_wmark_pages(zone)),
3784 K(high_wmark_pages(zone)),
3785 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3786 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3787 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3788 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3789 K(zone_page_state(zone, NR_UNEVICTABLE)),
3790 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3791 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3792 K(zone->present_pages),
3793 K(zone->managed_pages),
3794 K(zone_page_state(zone, NR_MLOCK)),
3795 K(zone_page_state(zone, NR_FILE_DIRTY)),
3796 K(zone_page_state(zone, NR_WRITEBACK)),
3797 K(zone_page_state(zone, NR_FILE_MAPPED)),
3798 K(zone_page_state(zone, NR_SHMEM)),
3799 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3800 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3801 zone_page_state(zone, NR_KERNEL_STACK) *
3803 K(zone_page_state(zone, NR_PAGETABLE)),
3804 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3805 K(zone_page_state(zone, NR_BOUNCE)),
3807 K(this_cpu_read(zone->pageset->pcp.count)),
3808 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3809 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3810 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3811 (!zone_reclaimable(zone) ? "yes" : "no")
3813 printk("lowmem_reserve[]:");
3814 for (i = 0; i < MAX_NR_ZONES; i++)
3815 printk(" %ld", zone->lowmem_reserve[i]);
3819 for_each_populated_zone(zone) {
3821 unsigned long nr[MAX_ORDER], flags, total = 0;
3822 unsigned char types[MAX_ORDER];
3824 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3827 printk("%s: ", zone->name);
3829 spin_lock_irqsave(&zone->lock, flags);
3830 for (order = 0; order < MAX_ORDER; order++) {
3831 struct free_area *area = &zone->free_area[order];
3834 nr[order] = area->nr_free;
3835 total += nr[order] << order;
3838 for (type = 0; type < MIGRATE_TYPES; type++) {
3839 if (!list_empty(&area->free_list[type]))
3840 types[order] |= 1 << type;
3843 spin_unlock_irqrestore(&zone->lock, flags);
3844 for (order = 0; order < MAX_ORDER; order++) {
3845 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3847 show_migration_types(types[order]);
3849 printk("= %lukB\n", K(total));
3852 hugetlb_show_meminfo();
3854 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3856 show_swap_cache_info();
3859 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3861 zoneref->zone = zone;
3862 zoneref->zone_idx = zone_idx(zone);
3866 * Builds allocation fallback zone lists.
3868 * Add all populated zones of a node to the zonelist.
3870 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3874 enum zone_type zone_type = MAX_NR_ZONES;
3878 zone = pgdat->node_zones + zone_type;
3879 if (populated_zone(zone)) {
3880 zoneref_set_zone(zone,
3881 &zonelist->_zonerefs[nr_zones++]);
3882 check_highest_zone(zone_type);
3884 } while (zone_type);
3892 * 0 = automatic detection of better ordering.
3893 * 1 = order by ([node] distance, -zonetype)
3894 * 2 = order by (-zonetype, [node] distance)
3896 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3897 * the same zonelist. So only NUMA can configure this param.
3899 #define ZONELIST_ORDER_DEFAULT 0
3900 #define ZONELIST_ORDER_NODE 1
3901 #define ZONELIST_ORDER_ZONE 2
3903 /* zonelist order in the kernel.
3904 * set_zonelist_order() will set this to NODE or ZONE.
3906 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3907 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3911 /* The value user specified ....changed by config */
3912 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3913 /* string for sysctl */
3914 #define NUMA_ZONELIST_ORDER_LEN 16
3915 char numa_zonelist_order[16] = "default";
3918 * interface for configure zonelist ordering.
3919 * command line option "numa_zonelist_order"
3920 * = "[dD]efault - default, automatic configuration.
3921 * = "[nN]ode - order by node locality, then by zone within node
3922 * = "[zZ]one - order by zone, then by locality within zone
3925 static int __parse_numa_zonelist_order(char *s)
3927 if (*s == 'd' || *s == 'D') {
3928 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3929 } else if (*s == 'n' || *s == 'N') {
3930 user_zonelist_order = ZONELIST_ORDER_NODE;
3931 } else if (*s == 'z' || *s == 'Z') {
3932 user_zonelist_order = ZONELIST_ORDER_ZONE;
3935 "Ignoring invalid numa_zonelist_order value: "
3942 static __init int setup_numa_zonelist_order(char *s)
3949 ret = __parse_numa_zonelist_order(s);
3951 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3955 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3958 * sysctl handler for numa_zonelist_order
3960 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3961 void __user *buffer, size_t *length,
3964 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3966 static DEFINE_MUTEX(zl_order_mutex);
3968 mutex_lock(&zl_order_mutex);
3970 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3974 strcpy(saved_string, (char *)table->data);
3976 ret = proc_dostring(table, write, buffer, length, ppos);
3980 int oldval = user_zonelist_order;
3982 ret = __parse_numa_zonelist_order((char *)table->data);
3985 * bogus value. restore saved string
3987 strncpy((char *)table->data, saved_string,
3988 NUMA_ZONELIST_ORDER_LEN);
3989 user_zonelist_order = oldval;
3990 } else if (oldval != user_zonelist_order) {
3991 mutex_lock(&zonelists_mutex);
3992 build_all_zonelists(NULL, NULL);
3993 mutex_unlock(&zonelists_mutex);
3997 mutex_unlock(&zl_order_mutex);
4002 #define MAX_NODE_LOAD (nr_online_nodes)
4003 static int node_load[MAX_NUMNODES];
4006 * find_next_best_node - find the next node that should appear in a given node's fallback list
4007 * @node: node whose fallback list we're appending
4008 * @used_node_mask: nodemask_t of already used nodes
4010 * We use a number of factors to determine which is the next node that should
4011 * appear on a given node's fallback list. The node should not have appeared
4012 * already in @node's fallback list, and it should be the next closest node
4013 * according to the distance array (which contains arbitrary distance values
4014 * from each node to each node in the system), and should also prefer nodes
4015 * with no CPUs, since presumably they'll have very little allocation pressure
4016 * on them otherwise.
4017 * It returns -1 if no node is found.
4019 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4022 int min_val = INT_MAX;
4023 int best_node = NUMA_NO_NODE;
4024 const struct cpumask *tmp = cpumask_of_node(0);
4026 /* Use the local node if we haven't already */
4027 if (!node_isset(node, *used_node_mask)) {
4028 node_set(node, *used_node_mask);
4032 for_each_node_state(n, N_MEMORY) {
4034 /* Don't want a node to appear more than once */
4035 if (node_isset(n, *used_node_mask))
4038 /* Use the distance array to find the distance */
4039 val = node_distance(node, n);
4041 /* Penalize nodes under us ("prefer the next node") */
4044 /* Give preference to headless and unused nodes */
4045 tmp = cpumask_of_node(n);
4046 if (!cpumask_empty(tmp))
4047 val += PENALTY_FOR_NODE_WITH_CPUS;
4049 /* Slight preference for less loaded node */
4050 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4051 val += node_load[n];
4053 if (val < min_val) {
4060 node_set(best_node, *used_node_mask);
4067 * Build zonelists ordered by node and zones within node.
4068 * This results in maximum locality--normal zone overflows into local
4069 * DMA zone, if any--but risks exhausting DMA zone.
4071 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4074 struct zonelist *zonelist;
4076 zonelist = &pgdat->node_zonelists[0];
4077 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4079 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4080 zonelist->_zonerefs[j].zone = NULL;
4081 zonelist->_zonerefs[j].zone_idx = 0;
4085 * Build gfp_thisnode zonelists
4087 static void build_thisnode_zonelists(pg_data_t *pgdat)
4090 struct zonelist *zonelist;
4092 zonelist = &pgdat->node_zonelists[1];
4093 j = build_zonelists_node(pgdat, zonelist, 0);
4094 zonelist->_zonerefs[j].zone = NULL;
4095 zonelist->_zonerefs[j].zone_idx = 0;
4099 * Build zonelists ordered by zone and nodes within zones.
4100 * This results in conserving DMA zone[s] until all Normal memory is
4101 * exhausted, but results in overflowing to remote node while memory
4102 * may still exist in local DMA zone.
4104 static int node_order[MAX_NUMNODES];
4106 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4109 int zone_type; /* needs to be signed */
4111 struct zonelist *zonelist;
4113 zonelist = &pgdat->node_zonelists[0];
4115 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4116 for (j = 0; j < nr_nodes; j++) {
4117 node = node_order[j];
4118 z = &NODE_DATA(node)->node_zones[zone_type];
4119 if (populated_zone(z)) {
4121 &zonelist->_zonerefs[pos++]);
4122 check_highest_zone(zone_type);
4126 zonelist->_zonerefs[pos].zone = NULL;
4127 zonelist->_zonerefs[pos].zone_idx = 0;
4130 #if defined(CONFIG_64BIT)
4132 * Devices that require DMA32/DMA are relatively rare and do not justify a
4133 * penalty to every machine in case the specialised case applies. Default
4134 * to Node-ordering on 64-bit NUMA machines
4136 static int default_zonelist_order(void)
4138 return ZONELIST_ORDER_NODE;
4142 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4143 * by the kernel. If processes running on node 0 deplete the low memory zone
4144 * then reclaim will occur more frequency increasing stalls and potentially
4145 * be easier to OOM if a large percentage of the zone is under writeback or
4146 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4147 * Hence, default to zone ordering on 32-bit.
4149 static int default_zonelist_order(void)
4151 return ZONELIST_ORDER_ZONE;
4153 #endif /* CONFIG_64BIT */
4155 static void set_zonelist_order(void)
4157 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4158 current_zonelist_order = default_zonelist_order();
4160 current_zonelist_order = user_zonelist_order;
4163 static void build_zonelists(pg_data_t *pgdat)
4167 nodemask_t used_mask;
4168 int local_node, prev_node;
4169 struct zonelist *zonelist;
4170 unsigned int order = current_zonelist_order;
4172 /* initialize zonelists */
4173 for (i = 0; i < MAX_ZONELISTS; i++) {
4174 zonelist = pgdat->node_zonelists + i;
4175 zonelist->_zonerefs[0].zone = NULL;
4176 zonelist->_zonerefs[0].zone_idx = 0;
4179 /* NUMA-aware ordering of nodes */
4180 local_node = pgdat->node_id;
4181 load = nr_online_nodes;
4182 prev_node = local_node;
4183 nodes_clear(used_mask);
4185 memset(node_order, 0, sizeof(node_order));
4188 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4190 * We don't want to pressure a particular node.
4191 * So adding penalty to the first node in same
4192 * distance group to make it round-robin.
4194 if (node_distance(local_node, node) !=
4195 node_distance(local_node, prev_node))
4196 node_load[node] = load;
4200 if (order == ZONELIST_ORDER_NODE)
4201 build_zonelists_in_node_order(pgdat, node);
4203 node_order[j++] = node; /* remember order */
4206 if (order == ZONELIST_ORDER_ZONE) {
4207 /* calculate node order -- i.e., DMA last! */
4208 build_zonelists_in_zone_order(pgdat, j);
4211 build_thisnode_zonelists(pgdat);
4214 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4216 * Return node id of node used for "local" allocations.
4217 * I.e., first node id of first zone in arg node's generic zonelist.
4218 * Used for initializing percpu 'numa_mem', which is used primarily
4219 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4221 int local_memory_node(int node)
4225 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4226 gfp_zone(GFP_KERNEL),
4233 #else /* CONFIG_NUMA */
4235 static void set_zonelist_order(void)
4237 current_zonelist_order = ZONELIST_ORDER_ZONE;
4240 static void build_zonelists(pg_data_t *pgdat)
4242 int node, local_node;
4244 struct zonelist *zonelist;
4246 local_node = pgdat->node_id;
4248 zonelist = &pgdat->node_zonelists[0];
4249 j = build_zonelists_node(pgdat, zonelist, 0);
4252 * Now we build the zonelist so that it contains the zones
4253 * of all the other nodes.
4254 * We don't want to pressure a particular node, so when
4255 * building the zones for node N, we make sure that the
4256 * zones coming right after the local ones are those from
4257 * node N+1 (modulo N)
4259 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4260 if (!node_online(node))
4262 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4264 for (node = 0; node < local_node; node++) {
4265 if (!node_online(node))
4267 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4270 zonelist->_zonerefs[j].zone = NULL;
4271 zonelist->_zonerefs[j].zone_idx = 0;
4274 #endif /* CONFIG_NUMA */
4277 * Boot pageset table. One per cpu which is going to be used for all
4278 * zones and all nodes. The parameters will be set in such a way
4279 * that an item put on a list will immediately be handed over to
4280 * the buddy list. This is safe since pageset manipulation is done
4281 * with interrupts disabled.
4283 * The boot_pagesets must be kept even after bootup is complete for
4284 * unused processors and/or zones. They do play a role for bootstrapping
4285 * hotplugged processors.
4287 * zoneinfo_show() and maybe other functions do
4288 * not check if the processor is online before following the pageset pointer.
4289 * Other parts of the kernel may not check if the zone is available.
4291 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4292 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4293 static void setup_zone_pageset(struct zone *zone);
4296 * Global mutex to protect against size modification of zonelists
4297 * as well as to serialize pageset setup for the new populated zone.
4299 DEFINE_MUTEX(zonelists_mutex);
4301 /* return values int ....just for stop_machine() */
4302 static int __build_all_zonelists(void *data)
4306 pg_data_t *self = data;
4309 memset(node_load, 0, sizeof(node_load));
4312 if (self && !node_online(self->node_id)) {
4313 build_zonelists(self);
4316 for_each_online_node(nid) {
4317 pg_data_t *pgdat = NODE_DATA(nid);
4319 build_zonelists(pgdat);
4323 * Initialize the boot_pagesets that are going to be used
4324 * for bootstrapping processors. The real pagesets for
4325 * each zone will be allocated later when the per cpu
4326 * allocator is available.
4328 * boot_pagesets are used also for bootstrapping offline
4329 * cpus if the system is already booted because the pagesets
4330 * are needed to initialize allocators on a specific cpu too.
4331 * F.e. the percpu allocator needs the page allocator which
4332 * needs the percpu allocator in order to allocate its pagesets
4333 * (a chicken-egg dilemma).
4335 for_each_possible_cpu(cpu) {
4336 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4338 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4340 * We now know the "local memory node" for each node--
4341 * i.e., the node of the first zone in the generic zonelist.
4342 * Set up numa_mem percpu variable for on-line cpus. During
4343 * boot, only the boot cpu should be on-line; we'll init the
4344 * secondary cpus' numa_mem as they come on-line. During
4345 * node/memory hotplug, we'll fixup all on-line cpus.
4347 if (cpu_online(cpu))
4348 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4355 static noinline void __init
4356 build_all_zonelists_init(void)
4358 __build_all_zonelists(NULL);
4359 mminit_verify_zonelist();
4360 cpuset_init_current_mems_allowed();
4364 * Called with zonelists_mutex held always
4365 * unless system_state == SYSTEM_BOOTING.
4367 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4368 * [we're only called with non-NULL zone through __meminit paths] and
4369 * (2) call of __init annotated helper build_all_zonelists_init
4370 * [protected by SYSTEM_BOOTING].
4372 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4374 set_zonelist_order();
4376 if (system_state == SYSTEM_BOOTING) {
4377 build_all_zonelists_init();
4379 #ifdef CONFIG_MEMORY_HOTPLUG
4381 setup_zone_pageset(zone);
4383 /* we have to stop all cpus to guarantee there is no user
4385 stop_machine(__build_all_zonelists, pgdat, NULL);
4386 /* cpuset refresh routine should be here */
4388 vm_total_pages = nr_free_pagecache_pages();
4390 * Disable grouping by mobility if the number of pages in the
4391 * system is too low to allow the mechanism to work. It would be
4392 * more accurate, but expensive to check per-zone. This check is
4393 * made on memory-hotadd so a system can start with mobility
4394 * disabled and enable it later
4396 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4397 page_group_by_mobility_disabled = 1;
4399 page_group_by_mobility_disabled = 0;
4401 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4402 "Total pages: %ld\n",
4404 zonelist_order_name[current_zonelist_order],
4405 page_group_by_mobility_disabled ? "off" : "on",
4408 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4413 * Helper functions to size the waitqueue hash table.
4414 * Essentially these want to choose hash table sizes sufficiently
4415 * large so that collisions trying to wait on pages are rare.
4416 * But in fact, the number of active page waitqueues on typical
4417 * systems is ridiculously low, less than 200. So this is even
4418 * conservative, even though it seems large.
4420 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4421 * waitqueues, i.e. the size of the waitq table given the number of pages.
4423 #define PAGES_PER_WAITQUEUE 256
4425 #ifndef CONFIG_MEMORY_HOTPLUG
4426 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4428 unsigned long size = 1;
4430 pages /= PAGES_PER_WAITQUEUE;
4432 while (size < pages)
4436 * Once we have dozens or even hundreds of threads sleeping
4437 * on IO we've got bigger problems than wait queue collision.
4438 * Limit the size of the wait table to a reasonable size.
4440 size = min(size, 4096UL);
4442 return max(size, 4UL);
4446 * A zone's size might be changed by hot-add, so it is not possible to determine
4447 * a suitable size for its wait_table. So we use the maximum size now.
4449 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4451 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4452 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4453 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4455 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4456 * or more by the traditional way. (See above). It equals:
4458 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4459 * ia64(16K page size) : = ( 8G + 4M)byte.
4460 * powerpc (64K page size) : = (32G +16M)byte.
4462 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4469 * This is an integer logarithm so that shifts can be used later
4470 * to extract the more random high bits from the multiplicative
4471 * hash function before the remainder is taken.
4473 static inline unsigned long wait_table_bits(unsigned long size)
4479 * Initially all pages are reserved - free ones are freed
4480 * up by free_all_bootmem() once the early boot process is
4481 * done. Non-atomic initialization, single-pass.
4483 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4484 unsigned long start_pfn, enum memmap_context context)
4486 pg_data_t *pgdat = NODE_DATA(nid);
4487 unsigned long end_pfn = start_pfn + size;
4490 unsigned long nr_initialised = 0;
4492 if (highest_memmap_pfn < end_pfn - 1)
4493 highest_memmap_pfn = end_pfn - 1;
4495 z = &pgdat->node_zones[zone];
4496 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4498 * There can be holes in boot-time mem_map[]s
4499 * handed to this function. They do not
4500 * exist on hotplugged memory.
4502 if (context == MEMMAP_EARLY) {
4503 if (!early_pfn_valid(pfn))
4505 if (!early_pfn_in_nid(pfn, nid))
4507 if (!update_defer_init(pgdat, pfn, end_pfn,
4513 * Mark the block movable so that blocks are reserved for
4514 * movable at startup. This will force kernel allocations
4515 * to reserve their blocks rather than leaking throughout
4516 * the address space during boot when many long-lived
4517 * kernel allocations are made.
4519 * bitmap is created for zone's valid pfn range. but memmap
4520 * can be created for invalid pages (for alignment)
4521 * check here not to call set_pageblock_migratetype() against
4524 if (!(pfn & (pageblock_nr_pages - 1))) {
4525 struct page *page = pfn_to_page(pfn);
4527 __init_single_page(page, pfn, zone, nid);
4528 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4530 __init_single_pfn(pfn, zone, nid);
4535 static void __meminit zone_init_free_lists(struct zone *zone)
4537 unsigned int order, t;
4538 for_each_migratetype_order(order, t) {
4539 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4540 zone->free_area[order].nr_free = 0;
4544 #ifndef __HAVE_ARCH_MEMMAP_INIT
4545 #define memmap_init(size, nid, zone, start_pfn) \
4546 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4549 static int zone_batchsize(struct zone *zone)
4555 * The per-cpu-pages pools are set to around 1000th of the
4556 * size of the zone. But no more than 1/2 of a meg.
4558 * OK, so we don't know how big the cache is. So guess.
4560 batch = zone->managed_pages / 1024;
4561 if (batch * PAGE_SIZE > 512 * 1024)
4562 batch = (512 * 1024) / PAGE_SIZE;
4563 batch /= 4; /* We effectively *= 4 below */
4568 * Clamp the batch to a 2^n - 1 value. Having a power
4569 * of 2 value was found to be more likely to have
4570 * suboptimal cache aliasing properties in some cases.
4572 * For example if 2 tasks are alternately allocating
4573 * batches of pages, one task can end up with a lot
4574 * of pages of one half of the possible page colors
4575 * and the other with pages of the other colors.
4577 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4582 /* The deferral and batching of frees should be suppressed under NOMMU
4585 * The problem is that NOMMU needs to be able to allocate large chunks
4586 * of contiguous memory as there's no hardware page translation to
4587 * assemble apparent contiguous memory from discontiguous pages.
4589 * Queueing large contiguous runs of pages for batching, however,
4590 * causes the pages to actually be freed in smaller chunks. As there
4591 * can be a significant delay between the individual batches being
4592 * recycled, this leads to the once large chunks of space being
4593 * fragmented and becoming unavailable for high-order allocations.
4600 * pcp->high and pcp->batch values are related and dependent on one another:
4601 * ->batch must never be higher then ->high.
4602 * The following function updates them in a safe manner without read side
4605 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4606 * those fields changing asynchronously (acording the the above rule).
4608 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4609 * outside of boot time (or some other assurance that no concurrent updaters
4612 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4613 unsigned long batch)
4615 /* start with a fail safe value for batch */
4619 /* Update high, then batch, in order */
4626 /* a companion to pageset_set_high() */
4627 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4629 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4632 static void pageset_init(struct per_cpu_pageset *p)
4634 struct per_cpu_pages *pcp;
4637 memset(p, 0, sizeof(*p));
4641 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4642 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4645 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4648 pageset_set_batch(p, batch);
4652 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4653 * to the value high for the pageset p.
4655 static void pageset_set_high(struct per_cpu_pageset *p,
4658 unsigned long batch = max(1UL, high / 4);
4659 if ((high / 4) > (PAGE_SHIFT * 8))
4660 batch = PAGE_SHIFT * 8;
4662 pageset_update(&p->pcp, high, batch);
4665 static void pageset_set_high_and_batch(struct zone *zone,
4666 struct per_cpu_pageset *pcp)
4668 if (percpu_pagelist_fraction)
4669 pageset_set_high(pcp,
4670 (zone->managed_pages /
4671 percpu_pagelist_fraction));
4673 pageset_set_batch(pcp, zone_batchsize(zone));
4676 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4678 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4681 pageset_set_high_and_batch(zone, pcp);
4684 static void __meminit setup_zone_pageset(struct zone *zone)
4687 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4688 for_each_possible_cpu(cpu)
4689 zone_pageset_init(zone, cpu);
4693 * Allocate per cpu pagesets and initialize them.
4694 * Before this call only boot pagesets were available.
4696 void __init setup_per_cpu_pageset(void)
4700 for_each_populated_zone(zone)
4701 setup_zone_pageset(zone);
4704 static noinline __init_refok
4705 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4711 * The per-page waitqueue mechanism uses hashed waitqueues
4714 zone->wait_table_hash_nr_entries =
4715 wait_table_hash_nr_entries(zone_size_pages);
4716 zone->wait_table_bits =
4717 wait_table_bits(zone->wait_table_hash_nr_entries);
4718 alloc_size = zone->wait_table_hash_nr_entries
4719 * sizeof(wait_queue_head_t);
4721 if (!slab_is_available()) {
4722 zone->wait_table = (wait_queue_head_t *)
4723 memblock_virt_alloc_node_nopanic(
4724 alloc_size, zone->zone_pgdat->node_id);
4727 * This case means that a zone whose size was 0 gets new memory
4728 * via memory hot-add.
4729 * But it may be the case that a new node was hot-added. In
4730 * this case vmalloc() will not be able to use this new node's
4731 * memory - this wait_table must be initialized to use this new
4732 * node itself as well.
4733 * To use this new node's memory, further consideration will be
4736 zone->wait_table = vmalloc(alloc_size);
4738 if (!zone->wait_table)
4741 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4742 init_waitqueue_head(zone->wait_table + i);
4747 static __meminit void zone_pcp_init(struct zone *zone)
4750 * per cpu subsystem is not up at this point. The following code
4751 * relies on the ability of the linker to provide the
4752 * offset of a (static) per cpu variable into the per cpu area.
4754 zone->pageset = &boot_pageset;
4756 if (populated_zone(zone))
4757 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4758 zone->name, zone->present_pages,
4759 zone_batchsize(zone));
4762 int __meminit init_currently_empty_zone(struct zone *zone,
4763 unsigned long zone_start_pfn,
4766 struct pglist_data *pgdat = zone->zone_pgdat;
4768 ret = zone_wait_table_init(zone, size);
4771 pgdat->nr_zones = zone_idx(zone) + 1;
4773 zone->zone_start_pfn = zone_start_pfn;
4775 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4776 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4778 (unsigned long)zone_idx(zone),
4779 zone_start_pfn, (zone_start_pfn + size));
4781 zone_init_free_lists(zone);
4786 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4787 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4790 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4792 int __meminit __early_pfn_to_nid(unsigned long pfn,
4793 struct mminit_pfnnid_cache *state)
4795 unsigned long start_pfn, end_pfn;
4798 if (state->last_start <= pfn && pfn < state->last_end)
4799 return state->last_nid;
4801 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4803 state->last_start = start_pfn;
4804 state->last_end = end_pfn;
4805 state->last_nid = nid;
4810 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4813 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4814 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4815 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4817 * If an architecture guarantees that all ranges registered contain no holes
4818 * and may be freed, this this function may be used instead of calling
4819 * memblock_free_early_nid() manually.
4821 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4823 unsigned long start_pfn, end_pfn;
4826 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4827 start_pfn = min(start_pfn, max_low_pfn);
4828 end_pfn = min(end_pfn, max_low_pfn);
4830 if (start_pfn < end_pfn)
4831 memblock_free_early_nid(PFN_PHYS(start_pfn),
4832 (end_pfn - start_pfn) << PAGE_SHIFT,
4838 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4839 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4841 * If an architecture guarantees that all ranges registered contain no holes and may
4842 * be freed, this function may be used instead of calling memory_present() manually.
4844 void __init sparse_memory_present_with_active_regions(int nid)
4846 unsigned long start_pfn, end_pfn;
4849 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4850 memory_present(this_nid, start_pfn, end_pfn);
4854 * get_pfn_range_for_nid - Return the start and end page frames for a node
4855 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4856 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4857 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4859 * It returns the start and end page frame of a node based on information
4860 * provided by memblock_set_node(). If called for a node
4861 * with no available memory, a warning is printed and the start and end
4864 void __meminit get_pfn_range_for_nid(unsigned int nid,
4865 unsigned long *start_pfn, unsigned long *end_pfn)
4867 unsigned long this_start_pfn, this_end_pfn;
4873 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4874 *start_pfn = min(*start_pfn, this_start_pfn);
4875 *end_pfn = max(*end_pfn, this_end_pfn);
4878 if (*start_pfn == -1UL)
4883 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4884 * assumption is made that zones within a node are ordered in monotonic
4885 * increasing memory addresses so that the "highest" populated zone is used
4887 static void __init find_usable_zone_for_movable(void)
4890 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4891 if (zone_index == ZONE_MOVABLE)
4894 if (arch_zone_highest_possible_pfn[zone_index] >
4895 arch_zone_lowest_possible_pfn[zone_index])
4899 VM_BUG_ON(zone_index == -1);
4900 movable_zone = zone_index;
4904 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4905 * because it is sized independent of architecture. Unlike the other zones,
4906 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4907 * in each node depending on the size of each node and how evenly kernelcore
4908 * is distributed. This helper function adjusts the zone ranges
4909 * provided by the architecture for a given node by using the end of the
4910 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4911 * zones within a node are in order of monotonic increases memory addresses
4913 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4914 unsigned long zone_type,
4915 unsigned long node_start_pfn,
4916 unsigned long node_end_pfn,
4917 unsigned long *zone_start_pfn,
4918 unsigned long *zone_end_pfn)
4920 /* Only adjust if ZONE_MOVABLE is on this node */
4921 if (zone_movable_pfn[nid]) {
4922 /* Size ZONE_MOVABLE */
4923 if (zone_type == ZONE_MOVABLE) {
4924 *zone_start_pfn = zone_movable_pfn[nid];
4925 *zone_end_pfn = min(node_end_pfn,
4926 arch_zone_highest_possible_pfn[movable_zone]);
4928 /* Adjust for ZONE_MOVABLE starting within this range */
4929 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4930 *zone_end_pfn > zone_movable_pfn[nid]) {
4931 *zone_end_pfn = zone_movable_pfn[nid];
4933 /* Check if this whole range is within ZONE_MOVABLE */
4934 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4935 *zone_start_pfn = *zone_end_pfn;
4940 * Return the number of pages a zone spans in a node, including holes
4941 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4943 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4944 unsigned long zone_type,
4945 unsigned long node_start_pfn,
4946 unsigned long node_end_pfn,
4947 unsigned long *ignored)
4949 unsigned long zone_start_pfn, zone_end_pfn;
4951 /* When hotadd a new node from cpu_up(), the node should be empty */
4952 if (!node_start_pfn && !node_end_pfn)
4955 /* Get the start and end of the zone */
4956 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4957 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4958 adjust_zone_range_for_zone_movable(nid, zone_type,
4959 node_start_pfn, node_end_pfn,
4960 &zone_start_pfn, &zone_end_pfn);
4962 /* Check that this node has pages within the zone's required range */
4963 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4966 /* Move the zone boundaries inside the node if necessary */
4967 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4968 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4970 /* Return the spanned pages */
4971 return zone_end_pfn - zone_start_pfn;
4975 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4976 * then all holes in the requested range will be accounted for.
4978 unsigned long __meminit __absent_pages_in_range(int nid,
4979 unsigned long range_start_pfn,
4980 unsigned long range_end_pfn)
4982 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4983 unsigned long start_pfn, end_pfn;
4986 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4987 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4988 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4989 nr_absent -= end_pfn - start_pfn;
4995 * absent_pages_in_range - Return number of page frames in holes within a range
4996 * @start_pfn: The start PFN to start searching for holes
4997 * @end_pfn: The end PFN to stop searching for holes
4999 * It returns the number of pages frames in memory holes within a range.
5001 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5002 unsigned long end_pfn)
5004 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5007 /* Return the number of page frames in holes in a zone on a node */
5008 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5009 unsigned long zone_type,
5010 unsigned long node_start_pfn,
5011 unsigned long node_end_pfn,
5012 unsigned long *ignored)
5014 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5015 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5016 unsigned long zone_start_pfn, zone_end_pfn;
5018 /* When hotadd a new node from cpu_up(), the node should be empty */
5019 if (!node_start_pfn && !node_end_pfn)
5022 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5023 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5025 adjust_zone_range_for_zone_movable(nid, zone_type,
5026 node_start_pfn, node_end_pfn,
5027 &zone_start_pfn, &zone_end_pfn);
5028 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5031 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5032 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5033 unsigned long zone_type,
5034 unsigned long node_start_pfn,
5035 unsigned long node_end_pfn,
5036 unsigned long *zones_size)
5038 return zones_size[zone_type];
5041 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5042 unsigned long zone_type,
5043 unsigned long node_start_pfn,
5044 unsigned long node_end_pfn,
5045 unsigned long *zholes_size)
5050 return zholes_size[zone_type];
5053 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5055 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5056 unsigned long node_start_pfn,
5057 unsigned long node_end_pfn,
5058 unsigned long *zones_size,
5059 unsigned long *zholes_size)
5061 unsigned long realtotalpages = 0, totalpages = 0;
5064 for (i = 0; i < MAX_NR_ZONES; i++) {
5065 struct zone *zone = pgdat->node_zones + i;
5066 unsigned long size, real_size;
5068 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5072 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5073 node_start_pfn, node_end_pfn,
5075 zone->spanned_pages = size;
5076 zone->present_pages = real_size;
5079 realtotalpages += real_size;
5082 pgdat->node_spanned_pages = totalpages;
5083 pgdat->node_present_pages = realtotalpages;
5084 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5088 #ifndef CONFIG_SPARSEMEM
5090 * Calculate the size of the zone->blockflags rounded to an unsigned long
5091 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5092 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5093 * round what is now in bits to nearest long in bits, then return it in
5096 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5098 unsigned long usemapsize;
5100 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5101 usemapsize = roundup(zonesize, pageblock_nr_pages);
5102 usemapsize = usemapsize >> pageblock_order;
5103 usemapsize *= NR_PAGEBLOCK_BITS;
5104 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5106 return usemapsize / 8;
5109 static void __init setup_usemap(struct pglist_data *pgdat,
5111 unsigned long zone_start_pfn,
5112 unsigned long zonesize)
5114 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5115 zone->pageblock_flags = NULL;
5117 zone->pageblock_flags =
5118 memblock_virt_alloc_node_nopanic(usemapsize,
5122 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5123 unsigned long zone_start_pfn, unsigned long zonesize) {}
5124 #endif /* CONFIG_SPARSEMEM */
5126 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5128 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5129 void __paginginit set_pageblock_order(void)
5133 /* Check that pageblock_nr_pages has not already been setup */
5134 if (pageblock_order)
5137 if (HPAGE_SHIFT > PAGE_SHIFT)
5138 order = HUGETLB_PAGE_ORDER;
5140 order = MAX_ORDER - 1;
5143 * Assume the largest contiguous order of interest is a huge page.
5144 * This value may be variable depending on boot parameters on IA64 and
5147 pageblock_order = order;
5149 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5152 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5153 * is unused as pageblock_order is set at compile-time. See
5154 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5157 void __paginginit set_pageblock_order(void)
5161 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5163 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5164 unsigned long present_pages)
5166 unsigned long pages = spanned_pages;
5169 * Provide a more accurate estimation if there are holes within
5170 * the zone and SPARSEMEM is in use. If there are holes within the
5171 * zone, each populated memory region may cost us one or two extra
5172 * memmap pages due to alignment because memmap pages for each
5173 * populated regions may not naturally algined on page boundary.
5174 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5176 if (spanned_pages > present_pages + (present_pages >> 4) &&
5177 IS_ENABLED(CONFIG_SPARSEMEM))
5178 pages = present_pages;
5180 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5184 * Set up the zone data structures:
5185 * - mark all pages reserved
5186 * - mark all memory queues empty
5187 * - clear the memory bitmaps
5189 * NOTE: pgdat should get zeroed by caller.
5191 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5194 int nid = pgdat->node_id;
5195 unsigned long zone_start_pfn = pgdat->node_start_pfn;
5198 pgdat_resize_init(pgdat);
5199 #ifdef CONFIG_NUMA_BALANCING
5200 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5201 pgdat->numabalancing_migrate_nr_pages = 0;
5202 pgdat->numabalancing_migrate_next_window = jiffies;
5204 init_waitqueue_head(&pgdat->kswapd_wait);
5205 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5206 pgdat_page_ext_init(pgdat);
5208 for (j = 0; j < MAX_NR_ZONES; j++) {
5209 struct zone *zone = pgdat->node_zones + j;
5210 unsigned long size, realsize, freesize, memmap_pages;
5212 size = zone->spanned_pages;
5213 realsize = freesize = zone->present_pages;
5216 * Adjust freesize so that it accounts for how much memory
5217 * is used by this zone for memmap. This affects the watermark
5218 * and per-cpu initialisations
5220 memmap_pages = calc_memmap_size(size, realsize);
5221 if (!is_highmem_idx(j)) {
5222 if (freesize >= memmap_pages) {
5223 freesize -= memmap_pages;
5226 " %s zone: %lu pages used for memmap\n",
5227 zone_names[j], memmap_pages);
5230 " %s zone: %lu pages exceeds freesize %lu\n",
5231 zone_names[j], memmap_pages, freesize);
5234 /* Account for reserved pages */
5235 if (j == 0 && freesize > dma_reserve) {
5236 freesize -= dma_reserve;
5237 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5238 zone_names[0], dma_reserve);
5241 if (!is_highmem_idx(j))
5242 nr_kernel_pages += freesize;
5243 /* Charge for highmem memmap if there are enough kernel pages */
5244 else if (nr_kernel_pages > memmap_pages * 2)
5245 nr_kernel_pages -= memmap_pages;
5246 nr_all_pages += freesize;
5249 * Set an approximate value for lowmem here, it will be adjusted
5250 * when the bootmem allocator frees pages into the buddy system.
5251 * And all highmem pages will be managed by the buddy system.
5253 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5256 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5258 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5260 zone->name = zone_names[j];
5261 spin_lock_init(&zone->lock);
5262 spin_lock_init(&zone->lru_lock);
5263 zone_seqlock_init(zone);
5264 zone->zone_pgdat = pgdat;
5265 zone_pcp_init(zone);
5267 /* For bootup, initialized properly in watermark setup */
5268 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5270 lruvec_init(&zone->lruvec);
5274 set_pageblock_order();
5275 setup_usemap(pgdat, zone, zone_start_pfn, size);
5276 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5278 memmap_init(size, nid, j, zone_start_pfn);
5279 zone_start_pfn += size;
5283 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5285 unsigned long __maybe_unused start = 0;
5286 unsigned long __maybe_unused offset = 0;
5288 /* Skip empty nodes */
5289 if (!pgdat->node_spanned_pages)
5292 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5293 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5294 offset = pgdat->node_start_pfn - start;
5295 /* ia64 gets its own node_mem_map, before this, without bootmem */
5296 if (!pgdat->node_mem_map) {
5297 unsigned long size, end;
5301 * The zone's endpoints aren't required to be MAX_ORDER
5302 * aligned but the node_mem_map endpoints must be in order
5303 * for the buddy allocator to function correctly.
5305 end = pgdat_end_pfn(pgdat);
5306 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5307 size = (end - start) * sizeof(struct page);
5308 map = alloc_remap(pgdat->node_id, size);
5310 map = memblock_virt_alloc_node_nopanic(size,
5312 pgdat->node_mem_map = map + offset;
5314 #ifndef CONFIG_NEED_MULTIPLE_NODES
5316 * With no DISCONTIG, the global mem_map is just set as node 0's
5318 if (pgdat == NODE_DATA(0)) {
5319 mem_map = NODE_DATA(0)->node_mem_map;
5320 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5321 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5323 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5326 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5329 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5330 unsigned long node_start_pfn, unsigned long *zholes_size)
5332 pg_data_t *pgdat = NODE_DATA(nid);
5333 unsigned long start_pfn = 0;
5334 unsigned long end_pfn = 0;
5336 /* pg_data_t should be reset to zero when it's allocated */
5337 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5339 reset_deferred_meminit(pgdat);
5340 pgdat->node_id = nid;
5341 pgdat->node_start_pfn = node_start_pfn;
5342 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5343 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5344 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5345 (u64)start_pfn << PAGE_SHIFT,
5346 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5348 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5349 zones_size, zholes_size);
5351 alloc_node_mem_map(pgdat);
5352 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5353 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5354 nid, (unsigned long)pgdat,
5355 (unsigned long)pgdat->node_mem_map);
5358 free_area_init_core(pgdat);
5361 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5363 #if MAX_NUMNODES > 1
5365 * Figure out the number of possible node ids.
5367 void __init setup_nr_node_ids(void)
5369 unsigned int highest;
5371 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5372 nr_node_ids = highest + 1;
5377 * node_map_pfn_alignment - determine the maximum internode alignment
5379 * This function should be called after node map is populated and sorted.
5380 * It calculates the maximum power of two alignment which can distinguish
5383 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5384 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5385 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5386 * shifted, 1GiB is enough and this function will indicate so.
5388 * This is used to test whether pfn -> nid mapping of the chosen memory
5389 * model has fine enough granularity to avoid incorrect mapping for the
5390 * populated node map.
5392 * Returns the determined alignment in pfn's. 0 if there is no alignment
5393 * requirement (single node).
5395 unsigned long __init node_map_pfn_alignment(void)
5397 unsigned long accl_mask = 0, last_end = 0;
5398 unsigned long start, end, mask;
5402 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5403 if (!start || last_nid < 0 || last_nid == nid) {
5410 * Start with a mask granular enough to pin-point to the
5411 * start pfn and tick off bits one-by-one until it becomes
5412 * too coarse to separate the current node from the last.
5414 mask = ~((1 << __ffs(start)) - 1);
5415 while (mask && last_end <= (start & (mask << 1)))
5418 /* accumulate all internode masks */
5422 /* convert mask to number of pages */
5423 return ~accl_mask + 1;
5426 /* Find the lowest pfn for a node */
5427 static unsigned long __init find_min_pfn_for_node(int nid)
5429 unsigned long min_pfn = ULONG_MAX;
5430 unsigned long start_pfn;
5433 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5434 min_pfn = min(min_pfn, start_pfn);
5436 if (min_pfn == ULONG_MAX) {
5438 "Could not find start_pfn for node %d\n", nid);
5446 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5448 * It returns the minimum PFN based on information provided via
5449 * memblock_set_node().
5451 unsigned long __init find_min_pfn_with_active_regions(void)
5453 return find_min_pfn_for_node(MAX_NUMNODES);
5457 * early_calculate_totalpages()
5458 * Sum pages in active regions for movable zone.
5459 * Populate N_MEMORY for calculating usable_nodes.
5461 static unsigned long __init early_calculate_totalpages(void)
5463 unsigned long totalpages = 0;
5464 unsigned long start_pfn, end_pfn;
5467 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5468 unsigned long pages = end_pfn - start_pfn;
5470 totalpages += pages;
5472 node_set_state(nid, N_MEMORY);
5478 * Find the PFN the Movable zone begins in each node. Kernel memory
5479 * is spread evenly between nodes as long as the nodes have enough
5480 * memory. When they don't, some nodes will have more kernelcore than
5483 static void __init find_zone_movable_pfns_for_nodes(void)
5486 unsigned long usable_startpfn;
5487 unsigned long kernelcore_node, kernelcore_remaining;
5488 /* save the state before borrow the nodemask */
5489 nodemask_t saved_node_state = node_states[N_MEMORY];
5490 unsigned long totalpages = early_calculate_totalpages();
5491 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5492 struct memblock_region *r;
5494 /* Need to find movable_zone earlier when movable_node is specified. */
5495 find_usable_zone_for_movable();
5498 * If movable_node is specified, ignore kernelcore and movablecore
5501 if (movable_node_is_enabled()) {
5502 for_each_memblock(memory, r) {
5503 if (!memblock_is_hotpluggable(r))
5508 usable_startpfn = PFN_DOWN(r->base);
5509 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5510 min(usable_startpfn, zone_movable_pfn[nid]) :
5518 * If movablecore=nn[KMG] was specified, calculate what size of
5519 * kernelcore that corresponds so that memory usable for
5520 * any allocation type is evenly spread. If both kernelcore
5521 * and movablecore are specified, then the value of kernelcore
5522 * will be used for required_kernelcore if it's greater than
5523 * what movablecore would have allowed.
5525 if (required_movablecore) {
5526 unsigned long corepages;
5529 * Round-up so that ZONE_MOVABLE is at least as large as what
5530 * was requested by the user
5532 required_movablecore =
5533 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5534 required_movablecore = min(totalpages, required_movablecore);
5535 corepages = totalpages - required_movablecore;
5537 required_kernelcore = max(required_kernelcore, corepages);
5541 * If kernelcore was not specified or kernelcore size is larger
5542 * than totalpages, there is no ZONE_MOVABLE.
5544 if (!required_kernelcore || required_kernelcore >= totalpages)
5547 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5548 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5551 /* Spread kernelcore memory as evenly as possible throughout nodes */
5552 kernelcore_node = required_kernelcore / usable_nodes;
5553 for_each_node_state(nid, N_MEMORY) {
5554 unsigned long start_pfn, end_pfn;
5557 * Recalculate kernelcore_node if the division per node
5558 * now exceeds what is necessary to satisfy the requested
5559 * amount of memory for the kernel
5561 if (required_kernelcore < kernelcore_node)
5562 kernelcore_node = required_kernelcore / usable_nodes;
5565 * As the map is walked, we track how much memory is usable
5566 * by the kernel using kernelcore_remaining. When it is
5567 * 0, the rest of the node is usable by ZONE_MOVABLE
5569 kernelcore_remaining = kernelcore_node;
5571 /* Go through each range of PFNs within this node */
5572 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5573 unsigned long size_pages;
5575 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5576 if (start_pfn >= end_pfn)
5579 /* Account for what is only usable for kernelcore */
5580 if (start_pfn < usable_startpfn) {
5581 unsigned long kernel_pages;
5582 kernel_pages = min(end_pfn, usable_startpfn)
5585 kernelcore_remaining -= min(kernel_pages,
5586 kernelcore_remaining);
5587 required_kernelcore -= min(kernel_pages,
5588 required_kernelcore);
5590 /* Continue if range is now fully accounted */
5591 if (end_pfn <= usable_startpfn) {
5594 * Push zone_movable_pfn to the end so
5595 * that if we have to rebalance
5596 * kernelcore across nodes, we will
5597 * not double account here
5599 zone_movable_pfn[nid] = end_pfn;
5602 start_pfn = usable_startpfn;
5606 * The usable PFN range for ZONE_MOVABLE is from
5607 * start_pfn->end_pfn. Calculate size_pages as the
5608 * number of pages used as kernelcore
5610 size_pages = end_pfn - start_pfn;
5611 if (size_pages > kernelcore_remaining)
5612 size_pages = kernelcore_remaining;
5613 zone_movable_pfn[nid] = start_pfn + size_pages;
5616 * Some kernelcore has been met, update counts and
5617 * break if the kernelcore for this node has been
5620 required_kernelcore -= min(required_kernelcore,
5622 kernelcore_remaining -= size_pages;
5623 if (!kernelcore_remaining)
5629 * If there is still required_kernelcore, we do another pass with one
5630 * less node in the count. This will push zone_movable_pfn[nid] further
5631 * along on the nodes that still have memory until kernelcore is
5635 if (usable_nodes && required_kernelcore > usable_nodes)
5639 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5640 for (nid = 0; nid < MAX_NUMNODES; nid++)
5641 zone_movable_pfn[nid] =
5642 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5645 /* restore the node_state */
5646 node_states[N_MEMORY] = saved_node_state;
5649 /* Any regular or high memory on that node ? */
5650 static void check_for_memory(pg_data_t *pgdat, int nid)
5652 enum zone_type zone_type;
5654 if (N_MEMORY == N_NORMAL_MEMORY)
5657 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5658 struct zone *zone = &pgdat->node_zones[zone_type];
5659 if (populated_zone(zone)) {
5660 node_set_state(nid, N_HIGH_MEMORY);
5661 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5662 zone_type <= ZONE_NORMAL)
5663 node_set_state(nid, N_NORMAL_MEMORY);
5670 * free_area_init_nodes - Initialise all pg_data_t and zone data
5671 * @max_zone_pfn: an array of max PFNs for each zone
5673 * This will call free_area_init_node() for each active node in the system.
5674 * Using the page ranges provided by memblock_set_node(), the size of each
5675 * zone in each node and their holes is calculated. If the maximum PFN
5676 * between two adjacent zones match, it is assumed that the zone is empty.
5677 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5678 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5679 * starts where the previous one ended. For example, ZONE_DMA32 starts
5680 * at arch_max_dma_pfn.
5682 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5684 unsigned long start_pfn, end_pfn;
5687 /* Record where the zone boundaries are */
5688 memset(arch_zone_lowest_possible_pfn, 0,
5689 sizeof(arch_zone_lowest_possible_pfn));
5690 memset(arch_zone_highest_possible_pfn, 0,
5691 sizeof(arch_zone_highest_possible_pfn));
5693 start_pfn = find_min_pfn_with_active_regions();
5695 for (i = 0; i < MAX_NR_ZONES; i++) {
5696 if (i == ZONE_MOVABLE)
5699 end_pfn = max(max_zone_pfn[i], start_pfn);
5700 arch_zone_lowest_possible_pfn[i] = start_pfn;
5701 arch_zone_highest_possible_pfn[i] = end_pfn;
5703 start_pfn = end_pfn;
5705 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5706 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5708 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5709 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5710 find_zone_movable_pfns_for_nodes();
5712 /* Print out the zone ranges */
5713 pr_info("Zone ranges:\n");
5714 for (i = 0; i < MAX_NR_ZONES; i++) {
5715 if (i == ZONE_MOVABLE)
5717 pr_info(" %-8s ", zone_names[i]);
5718 if (arch_zone_lowest_possible_pfn[i] ==
5719 arch_zone_highest_possible_pfn[i])
5722 pr_cont("[mem %#018Lx-%#018Lx]\n",
5723 (u64)arch_zone_lowest_possible_pfn[i]
5725 ((u64)arch_zone_highest_possible_pfn[i]
5726 << PAGE_SHIFT) - 1);
5729 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5730 pr_info("Movable zone start for each node\n");
5731 for (i = 0; i < MAX_NUMNODES; i++) {
5732 if (zone_movable_pfn[i])
5733 pr_info(" Node %d: %#018Lx\n", i,
5734 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5737 /* Print out the early node map */
5738 pr_info("Early memory node ranges\n");
5739 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5740 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5741 (u64)start_pfn << PAGE_SHIFT,
5742 ((u64)end_pfn << PAGE_SHIFT) - 1);
5744 /* Initialise every node */
5745 mminit_verify_pageflags_layout();
5746 setup_nr_node_ids();
5747 for_each_online_node(nid) {
5748 pg_data_t *pgdat = NODE_DATA(nid);
5749 free_area_init_node(nid, NULL,
5750 find_min_pfn_for_node(nid), NULL);
5752 /* Any memory on that node */
5753 if (pgdat->node_present_pages)
5754 node_set_state(nid, N_MEMORY);
5755 check_for_memory(pgdat, nid);
5759 static int __init cmdline_parse_core(char *p, unsigned long *core)
5761 unsigned long long coremem;
5765 coremem = memparse(p, &p);
5766 *core = coremem >> PAGE_SHIFT;
5768 /* Paranoid check that UL is enough for the coremem value */
5769 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5775 * kernelcore=size sets the amount of memory for use for allocations that
5776 * cannot be reclaimed or migrated.
5778 static int __init cmdline_parse_kernelcore(char *p)
5780 return cmdline_parse_core(p, &required_kernelcore);
5784 * movablecore=size sets the amount of memory for use for allocations that
5785 * can be reclaimed or migrated.
5787 static int __init cmdline_parse_movablecore(char *p)
5789 return cmdline_parse_core(p, &required_movablecore);
5792 early_param("kernelcore", cmdline_parse_kernelcore);
5793 early_param("movablecore", cmdline_parse_movablecore);
5795 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5797 void adjust_managed_page_count(struct page *page, long count)
5799 spin_lock(&managed_page_count_lock);
5800 page_zone(page)->managed_pages += count;
5801 totalram_pages += count;
5802 #ifdef CONFIG_HIGHMEM
5803 if (PageHighMem(page))
5804 totalhigh_pages += count;
5806 spin_unlock(&managed_page_count_lock);
5808 EXPORT_SYMBOL(adjust_managed_page_count);
5810 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5813 unsigned long pages = 0;
5815 start = (void *)PAGE_ALIGN((unsigned long)start);
5816 end = (void *)((unsigned long)end & PAGE_MASK);
5817 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5818 if ((unsigned int)poison <= 0xFF)
5819 memset(pos, poison, PAGE_SIZE);
5820 free_reserved_page(virt_to_page(pos));
5824 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5825 s, pages << (PAGE_SHIFT - 10), start, end);
5829 EXPORT_SYMBOL(free_reserved_area);
5831 #ifdef CONFIG_HIGHMEM
5832 void free_highmem_page(struct page *page)
5834 __free_reserved_page(page);
5836 page_zone(page)->managed_pages++;
5842 void __init mem_init_print_info(const char *str)
5844 unsigned long physpages, codesize, datasize, rosize, bss_size;
5845 unsigned long init_code_size, init_data_size;
5847 physpages = get_num_physpages();
5848 codesize = _etext - _stext;
5849 datasize = _edata - _sdata;
5850 rosize = __end_rodata - __start_rodata;
5851 bss_size = __bss_stop - __bss_start;
5852 init_data_size = __init_end - __init_begin;
5853 init_code_size = _einittext - _sinittext;
5856 * Detect special cases and adjust section sizes accordingly:
5857 * 1) .init.* may be embedded into .data sections
5858 * 2) .init.text.* may be out of [__init_begin, __init_end],
5859 * please refer to arch/tile/kernel/vmlinux.lds.S.
5860 * 3) .rodata.* may be embedded into .text or .data sections.
5862 #define adj_init_size(start, end, size, pos, adj) \
5864 if (start <= pos && pos < end && size > adj) \
5868 adj_init_size(__init_begin, __init_end, init_data_size,
5869 _sinittext, init_code_size);
5870 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5871 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5872 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5873 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5875 #undef adj_init_size
5877 pr_info("Memory: %luK/%luK available "
5878 "(%luK kernel code, %luK rwdata, %luK rodata, "
5879 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5880 #ifdef CONFIG_HIGHMEM
5884 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5885 codesize >> 10, datasize >> 10, rosize >> 10,
5886 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5887 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5888 totalcma_pages << (PAGE_SHIFT-10),
5889 #ifdef CONFIG_HIGHMEM
5890 totalhigh_pages << (PAGE_SHIFT-10),
5892 str ? ", " : "", str ? str : "");
5896 * set_dma_reserve - set the specified number of pages reserved in the first zone
5897 * @new_dma_reserve: The number of pages to mark reserved
5899 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
5900 * In the DMA zone, a significant percentage may be consumed by kernel image
5901 * and other unfreeable allocations which can skew the watermarks badly. This
5902 * function may optionally be used to account for unfreeable pages in the
5903 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5904 * smaller per-cpu batchsize.
5906 void __init set_dma_reserve(unsigned long new_dma_reserve)
5908 dma_reserve = new_dma_reserve;
5911 void __init free_area_init(unsigned long *zones_size)
5913 free_area_init_node(0, zones_size,
5914 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5917 static int page_alloc_cpu_notify(struct notifier_block *self,
5918 unsigned long action, void *hcpu)
5920 int cpu = (unsigned long)hcpu;
5922 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5923 lru_add_drain_cpu(cpu);
5927 * Spill the event counters of the dead processor
5928 * into the current processors event counters.
5929 * This artificially elevates the count of the current
5932 vm_events_fold_cpu(cpu);
5935 * Zero the differential counters of the dead processor
5936 * so that the vm statistics are consistent.
5938 * This is only okay since the processor is dead and cannot
5939 * race with what we are doing.
5941 cpu_vm_stats_fold(cpu);
5946 void __init page_alloc_init(void)
5948 hotcpu_notifier(page_alloc_cpu_notify, 0);
5952 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5953 * or min_free_kbytes changes.
5955 static void calculate_totalreserve_pages(void)
5957 struct pglist_data *pgdat;
5958 unsigned long reserve_pages = 0;
5959 enum zone_type i, j;
5961 for_each_online_pgdat(pgdat) {
5962 for (i = 0; i < MAX_NR_ZONES; i++) {
5963 struct zone *zone = pgdat->node_zones + i;
5966 /* Find valid and maximum lowmem_reserve in the zone */
5967 for (j = i; j < MAX_NR_ZONES; j++) {
5968 if (zone->lowmem_reserve[j] > max)
5969 max = zone->lowmem_reserve[j];
5972 /* we treat the high watermark as reserved pages. */
5973 max += high_wmark_pages(zone);
5975 if (max > zone->managed_pages)
5976 max = zone->managed_pages;
5978 zone->totalreserve_pages = max;
5980 reserve_pages += max;
5983 totalreserve_pages = reserve_pages;
5987 * setup_per_zone_lowmem_reserve - called whenever
5988 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
5989 * has a correct pages reserved value, so an adequate number of
5990 * pages are left in the zone after a successful __alloc_pages().
5992 static void setup_per_zone_lowmem_reserve(void)
5994 struct pglist_data *pgdat;
5995 enum zone_type j, idx;
5997 for_each_online_pgdat(pgdat) {
5998 for (j = 0; j < MAX_NR_ZONES; j++) {
5999 struct zone *zone = pgdat->node_zones + j;
6000 unsigned long managed_pages = zone->managed_pages;
6002 zone->lowmem_reserve[j] = 0;
6006 struct zone *lower_zone;
6010 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6011 sysctl_lowmem_reserve_ratio[idx] = 1;
6013 lower_zone = pgdat->node_zones + idx;
6014 lower_zone->lowmem_reserve[j] = managed_pages /
6015 sysctl_lowmem_reserve_ratio[idx];
6016 managed_pages += lower_zone->managed_pages;
6021 /* update totalreserve_pages */
6022 calculate_totalreserve_pages();
6025 static void __setup_per_zone_wmarks(void)
6027 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6028 unsigned long lowmem_pages = 0;
6030 unsigned long flags;
6032 /* Calculate total number of !ZONE_HIGHMEM pages */
6033 for_each_zone(zone) {
6034 if (!is_highmem(zone))
6035 lowmem_pages += zone->managed_pages;
6038 for_each_zone(zone) {
6041 spin_lock_irqsave(&zone->lock, flags);
6042 tmp = (u64)pages_min * zone->managed_pages;
6043 do_div(tmp, lowmem_pages);
6044 if (is_highmem(zone)) {
6046 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6047 * need highmem pages, so cap pages_min to a small
6050 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6051 * deltas control asynch page reclaim, and so should
6052 * not be capped for highmem.
6054 unsigned long min_pages;
6056 min_pages = zone->managed_pages / 1024;
6057 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6058 zone->watermark[WMARK_MIN] = min_pages;
6061 * If it's a lowmem zone, reserve a number of pages
6062 * proportionate to the zone's size.
6064 zone->watermark[WMARK_MIN] = tmp;
6067 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
6068 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
6070 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6071 high_wmark_pages(zone) - low_wmark_pages(zone) -
6072 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6074 spin_unlock_irqrestore(&zone->lock, flags);
6077 /* update totalreserve_pages */
6078 calculate_totalreserve_pages();
6082 * setup_per_zone_wmarks - called when min_free_kbytes changes
6083 * or when memory is hot-{added|removed}
6085 * Ensures that the watermark[min,low,high] values for each zone are set
6086 * correctly with respect to min_free_kbytes.
6088 void setup_per_zone_wmarks(void)
6090 mutex_lock(&zonelists_mutex);
6091 __setup_per_zone_wmarks();
6092 mutex_unlock(&zonelists_mutex);
6096 * The inactive anon list should be small enough that the VM never has to
6097 * do too much work, but large enough that each inactive page has a chance
6098 * to be referenced again before it is swapped out.
6100 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6101 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6102 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6103 * the anonymous pages are kept on the inactive list.
6106 * memory ratio inactive anon
6107 * -------------------------------------
6116 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6118 unsigned int gb, ratio;
6120 /* Zone size in gigabytes */
6121 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6123 ratio = int_sqrt(10 * gb);
6127 zone->inactive_ratio = ratio;
6130 static void __meminit setup_per_zone_inactive_ratio(void)
6135 calculate_zone_inactive_ratio(zone);
6139 * Initialise min_free_kbytes.
6141 * For small machines we want it small (128k min). For large machines
6142 * we want it large (64MB max). But it is not linear, because network
6143 * bandwidth does not increase linearly with machine size. We use
6145 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6146 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6162 int __meminit init_per_zone_wmark_min(void)
6164 unsigned long lowmem_kbytes;
6165 int new_min_free_kbytes;
6167 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6168 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6170 if (new_min_free_kbytes > user_min_free_kbytes) {
6171 min_free_kbytes = new_min_free_kbytes;
6172 if (min_free_kbytes < 128)
6173 min_free_kbytes = 128;
6174 if (min_free_kbytes > 65536)
6175 min_free_kbytes = 65536;
6177 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6178 new_min_free_kbytes, user_min_free_kbytes);
6180 setup_per_zone_wmarks();
6181 refresh_zone_stat_thresholds();
6182 setup_per_zone_lowmem_reserve();
6183 setup_per_zone_inactive_ratio();
6186 core_initcall(init_per_zone_wmark_min)
6189 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6190 * that we can call two helper functions whenever min_free_kbytes
6193 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6194 void __user *buffer, size_t *length, loff_t *ppos)
6198 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6203 user_min_free_kbytes = min_free_kbytes;
6204 setup_per_zone_wmarks();
6210 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6211 void __user *buffer, size_t *length, loff_t *ppos)
6216 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6221 zone->min_unmapped_pages = (zone->managed_pages *
6222 sysctl_min_unmapped_ratio) / 100;
6226 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6227 void __user *buffer, size_t *length, loff_t *ppos)
6232 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6237 zone->min_slab_pages = (zone->managed_pages *
6238 sysctl_min_slab_ratio) / 100;
6244 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6245 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6246 * whenever sysctl_lowmem_reserve_ratio changes.
6248 * The reserve ratio obviously has absolutely no relation with the
6249 * minimum watermarks. The lowmem reserve ratio can only make sense
6250 * if in function of the boot time zone sizes.
6252 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6253 void __user *buffer, size_t *length, loff_t *ppos)
6255 proc_dointvec_minmax(table, write, buffer, length, ppos);
6256 setup_per_zone_lowmem_reserve();
6261 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6262 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6263 * pagelist can have before it gets flushed back to buddy allocator.
6265 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6266 void __user *buffer, size_t *length, loff_t *ppos)
6269 int old_percpu_pagelist_fraction;
6272 mutex_lock(&pcp_batch_high_lock);
6273 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6275 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6276 if (!write || ret < 0)
6279 /* Sanity checking to avoid pcp imbalance */
6280 if (percpu_pagelist_fraction &&
6281 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6282 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6288 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6291 for_each_populated_zone(zone) {
6294 for_each_possible_cpu(cpu)
6295 pageset_set_high_and_batch(zone,
6296 per_cpu_ptr(zone->pageset, cpu));
6299 mutex_unlock(&pcp_batch_high_lock);
6304 int hashdist = HASHDIST_DEFAULT;
6306 static int __init set_hashdist(char *str)
6310 hashdist = simple_strtoul(str, &str, 0);
6313 __setup("hashdist=", set_hashdist);
6317 * allocate a large system hash table from bootmem
6318 * - it is assumed that the hash table must contain an exact power-of-2
6319 * quantity of entries
6320 * - limit is the number of hash buckets, not the total allocation size
6322 void *__init alloc_large_system_hash(const char *tablename,
6323 unsigned long bucketsize,
6324 unsigned long numentries,
6327 unsigned int *_hash_shift,
6328 unsigned int *_hash_mask,
6329 unsigned long low_limit,
6330 unsigned long high_limit)
6332 unsigned long long max = high_limit;
6333 unsigned long log2qty, size;
6336 /* allow the kernel cmdline to have a say */
6338 /* round applicable memory size up to nearest megabyte */
6339 numentries = nr_kernel_pages;
6341 /* It isn't necessary when PAGE_SIZE >= 1MB */
6342 if (PAGE_SHIFT < 20)
6343 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6345 /* limit to 1 bucket per 2^scale bytes of low memory */
6346 if (scale > PAGE_SHIFT)
6347 numentries >>= (scale - PAGE_SHIFT);
6349 numentries <<= (PAGE_SHIFT - scale);
6351 /* Make sure we've got at least a 0-order allocation.. */
6352 if (unlikely(flags & HASH_SMALL)) {
6353 /* Makes no sense without HASH_EARLY */
6354 WARN_ON(!(flags & HASH_EARLY));
6355 if (!(numentries >> *_hash_shift)) {
6356 numentries = 1UL << *_hash_shift;
6357 BUG_ON(!numentries);
6359 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6360 numentries = PAGE_SIZE / bucketsize;
6362 numentries = roundup_pow_of_two(numentries);
6364 /* limit allocation size to 1/16 total memory by default */
6366 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6367 do_div(max, bucketsize);
6369 max = min(max, 0x80000000ULL);
6371 if (numentries < low_limit)
6372 numentries = low_limit;
6373 if (numentries > max)
6376 log2qty = ilog2(numentries);
6379 size = bucketsize << log2qty;
6380 if (flags & HASH_EARLY)
6381 table = memblock_virt_alloc_nopanic(size, 0);
6383 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6386 * If bucketsize is not a power-of-two, we may free
6387 * some pages at the end of hash table which
6388 * alloc_pages_exact() automatically does
6390 if (get_order(size) < MAX_ORDER) {
6391 table = alloc_pages_exact(size, GFP_ATOMIC);
6392 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6395 } while (!table && size > PAGE_SIZE && --log2qty);
6398 panic("Failed to allocate %s hash table\n", tablename);
6400 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6403 ilog2(size) - PAGE_SHIFT,
6407 *_hash_shift = log2qty;
6409 *_hash_mask = (1 << log2qty) - 1;
6414 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6415 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6418 #ifdef CONFIG_SPARSEMEM
6419 return __pfn_to_section(pfn)->pageblock_flags;
6421 return zone->pageblock_flags;
6422 #endif /* CONFIG_SPARSEMEM */
6425 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6427 #ifdef CONFIG_SPARSEMEM
6428 pfn &= (PAGES_PER_SECTION-1);
6429 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6431 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6432 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6433 #endif /* CONFIG_SPARSEMEM */
6437 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6438 * @page: The page within the block of interest
6439 * @pfn: The target page frame number
6440 * @end_bitidx: The last bit of interest to retrieve
6441 * @mask: mask of bits that the caller is interested in
6443 * Return: pageblock_bits flags
6445 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6446 unsigned long end_bitidx,
6450 unsigned long *bitmap;
6451 unsigned long bitidx, word_bitidx;
6454 zone = page_zone(page);
6455 bitmap = get_pageblock_bitmap(zone, pfn);
6456 bitidx = pfn_to_bitidx(zone, pfn);
6457 word_bitidx = bitidx / BITS_PER_LONG;
6458 bitidx &= (BITS_PER_LONG-1);
6460 word = bitmap[word_bitidx];
6461 bitidx += end_bitidx;
6462 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6466 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6467 * @page: The page within the block of interest
6468 * @flags: The flags to set
6469 * @pfn: The target page frame number
6470 * @end_bitidx: The last bit of interest
6471 * @mask: mask of bits that the caller is interested in
6473 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6475 unsigned long end_bitidx,
6479 unsigned long *bitmap;
6480 unsigned long bitidx, word_bitidx;
6481 unsigned long old_word, word;
6483 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6485 zone = page_zone(page);
6486 bitmap = get_pageblock_bitmap(zone, pfn);
6487 bitidx = pfn_to_bitidx(zone, pfn);
6488 word_bitidx = bitidx / BITS_PER_LONG;
6489 bitidx &= (BITS_PER_LONG-1);
6491 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6493 bitidx += end_bitidx;
6494 mask <<= (BITS_PER_LONG - bitidx - 1);
6495 flags <<= (BITS_PER_LONG - bitidx - 1);
6497 word = READ_ONCE(bitmap[word_bitidx]);
6499 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6500 if (word == old_word)
6507 * This function checks whether pageblock includes unmovable pages or not.
6508 * If @count is not zero, it is okay to include less @count unmovable pages
6510 * PageLRU check without isolation or lru_lock could race so that
6511 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6512 * expect this function should be exact.
6514 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6515 bool skip_hwpoisoned_pages)
6517 unsigned long pfn, iter, found;
6521 * For avoiding noise data, lru_add_drain_all() should be called
6522 * If ZONE_MOVABLE, the zone never contains unmovable pages
6524 if (zone_idx(zone) == ZONE_MOVABLE)
6526 mt = get_pageblock_migratetype(page);
6527 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6530 pfn = page_to_pfn(page);
6531 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6532 unsigned long check = pfn + iter;
6534 if (!pfn_valid_within(check))
6537 page = pfn_to_page(check);
6540 * Hugepages are not in LRU lists, but they're movable.
6541 * We need not scan over tail pages bacause we don't
6542 * handle each tail page individually in migration.
6544 if (PageHuge(page)) {
6545 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6550 * We can't use page_count without pin a page
6551 * because another CPU can free compound page.
6552 * This check already skips compound tails of THP
6553 * because their page->_count is zero at all time.
6555 if (!atomic_read(&page->_count)) {
6556 if (PageBuddy(page))
6557 iter += (1 << page_order(page)) - 1;
6562 * The HWPoisoned page may be not in buddy system, and
6563 * page_count() is not 0.
6565 if (skip_hwpoisoned_pages && PageHWPoison(page))
6571 * If there are RECLAIMABLE pages, we need to check
6572 * it. But now, memory offline itself doesn't call
6573 * shrink_node_slabs() and it still to be fixed.
6576 * If the page is not RAM, page_count()should be 0.
6577 * we don't need more check. This is an _used_ not-movable page.
6579 * The problematic thing here is PG_reserved pages. PG_reserved
6580 * is set to both of a memory hole page and a _used_ kernel
6589 bool is_pageblock_removable_nolock(struct page *page)
6595 * We have to be careful here because we are iterating over memory
6596 * sections which are not zone aware so we might end up outside of
6597 * the zone but still within the section.
6598 * We have to take care about the node as well. If the node is offline
6599 * its NODE_DATA will be NULL - see page_zone.
6601 if (!node_online(page_to_nid(page)))
6604 zone = page_zone(page);
6605 pfn = page_to_pfn(page);
6606 if (!zone_spans_pfn(zone, pfn))
6609 return !has_unmovable_pages(zone, page, 0, true);
6614 static unsigned long pfn_max_align_down(unsigned long pfn)
6616 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6617 pageblock_nr_pages) - 1);
6620 static unsigned long pfn_max_align_up(unsigned long pfn)
6622 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6623 pageblock_nr_pages));
6626 /* [start, end) must belong to a single zone. */
6627 static int __alloc_contig_migrate_range(struct compact_control *cc,
6628 unsigned long start, unsigned long end)
6630 /* This function is based on compact_zone() from compaction.c. */
6631 unsigned long nr_reclaimed;
6632 unsigned long pfn = start;
6633 unsigned int tries = 0;
6638 while (pfn < end || !list_empty(&cc->migratepages)) {
6639 if (fatal_signal_pending(current)) {
6644 if (list_empty(&cc->migratepages)) {
6645 cc->nr_migratepages = 0;
6646 pfn = isolate_migratepages_range(cc, pfn, end);
6652 } else if (++tries == 5) {
6653 ret = ret < 0 ? ret : -EBUSY;
6657 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6659 cc->nr_migratepages -= nr_reclaimed;
6661 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6662 NULL, 0, cc->mode, MR_CMA);
6665 putback_movable_pages(&cc->migratepages);
6672 * alloc_contig_range() -- tries to allocate given range of pages
6673 * @start: start PFN to allocate
6674 * @end: one-past-the-last PFN to allocate
6675 * @migratetype: migratetype of the underlaying pageblocks (either
6676 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6677 * in range must have the same migratetype and it must
6678 * be either of the two.
6680 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6681 * aligned, however it's the caller's responsibility to guarantee that
6682 * we are the only thread that changes migrate type of pageblocks the
6685 * The PFN range must belong to a single zone.
6687 * Returns zero on success or negative error code. On success all
6688 * pages which PFN is in [start, end) are allocated for the caller and
6689 * need to be freed with free_contig_range().
6691 int alloc_contig_range(unsigned long start, unsigned long end,
6692 unsigned migratetype)
6694 unsigned long outer_start, outer_end;
6698 struct compact_control cc = {
6699 .nr_migratepages = 0,
6701 .zone = page_zone(pfn_to_page(start)),
6702 .mode = MIGRATE_SYNC,
6703 .ignore_skip_hint = true,
6705 INIT_LIST_HEAD(&cc.migratepages);
6708 * What we do here is we mark all pageblocks in range as
6709 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6710 * have different sizes, and due to the way page allocator
6711 * work, we align the range to biggest of the two pages so
6712 * that page allocator won't try to merge buddies from
6713 * different pageblocks and change MIGRATE_ISOLATE to some
6714 * other migration type.
6716 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6717 * migrate the pages from an unaligned range (ie. pages that
6718 * we are interested in). This will put all the pages in
6719 * range back to page allocator as MIGRATE_ISOLATE.
6721 * When this is done, we take the pages in range from page
6722 * allocator removing them from the buddy system. This way
6723 * page allocator will never consider using them.
6725 * This lets us mark the pageblocks back as
6726 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6727 * aligned range but not in the unaligned, original range are
6728 * put back to page allocator so that buddy can use them.
6731 ret = start_isolate_page_range(pfn_max_align_down(start),
6732 pfn_max_align_up(end), migratetype,
6737 ret = __alloc_contig_migrate_range(&cc, start, end);
6742 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6743 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6744 * more, all pages in [start, end) are free in page allocator.
6745 * What we are going to do is to allocate all pages from
6746 * [start, end) (that is remove them from page allocator).
6748 * The only problem is that pages at the beginning and at the
6749 * end of interesting range may be not aligned with pages that
6750 * page allocator holds, ie. they can be part of higher order
6751 * pages. Because of this, we reserve the bigger range and
6752 * once this is done free the pages we are not interested in.
6754 * We don't have to hold zone->lock here because the pages are
6755 * isolated thus they won't get removed from buddy.
6758 lru_add_drain_all();
6759 drain_all_pages(cc.zone);
6762 outer_start = start;
6763 while (!PageBuddy(pfn_to_page(outer_start))) {
6764 if (++order >= MAX_ORDER) {
6768 outer_start &= ~0UL << order;
6771 /* Make sure the range is really isolated. */
6772 if (test_pages_isolated(outer_start, end, false)) {
6773 pr_info("%s: [%lx, %lx) PFNs busy\n",
6774 __func__, outer_start, end);
6779 /* Grab isolated pages from freelists. */
6780 outer_end = isolate_freepages_range(&cc, outer_start, end);
6786 /* Free head and tail (if any) */
6787 if (start != outer_start)
6788 free_contig_range(outer_start, start - outer_start);
6789 if (end != outer_end)
6790 free_contig_range(end, outer_end - end);
6793 undo_isolate_page_range(pfn_max_align_down(start),
6794 pfn_max_align_up(end), migratetype);
6798 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6800 unsigned int count = 0;
6802 for (; nr_pages--; pfn++) {
6803 struct page *page = pfn_to_page(pfn);
6805 count += page_count(page) != 1;
6808 WARN(count != 0, "%d pages are still in use!\n", count);
6812 #ifdef CONFIG_MEMORY_HOTPLUG
6814 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6815 * page high values need to be recalulated.
6817 void __meminit zone_pcp_update(struct zone *zone)
6820 mutex_lock(&pcp_batch_high_lock);
6821 for_each_possible_cpu(cpu)
6822 pageset_set_high_and_batch(zone,
6823 per_cpu_ptr(zone->pageset, cpu));
6824 mutex_unlock(&pcp_batch_high_lock);
6828 void zone_pcp_reset(struct zone *zone)
6830 unsigned long flags;
6832 struct per_cpu_pageset *pset;
6834 /* avoid races with drain_pages() */
6835 local_irq_save(flags);
6836 if (zone->pageset != &boot_pageset) {
6837 for_each_online_cpu(cpu) {
6838 pset = per_cpu_ptr(zone->pageset, cpu);
6839 drain_zonestat(zone, pset);
6841 free_percpu(zone->pageset);
6842 zone->pageset = &boot_pageset;
6844 local_irq_restore(flags);
6847 #ifdef CONFIG_MEMORY_HOTREMOVE
6849 * All pages in the range must be isolated before calling this.
6852 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6856 unsigned int order, i;
6858 unsigned long flags;
6859 /* find the first valid pfn */
6860 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6865 zone = page_zone(pfn_to_page(pfn));
6866 spin_lock_irqsave(&zone->lock, flags);
6868 while (pfn < end_pfn) {
6869 if (!pfn_valid(pfn)) {
6873 page = pfn_to_page(pfn);
6875 * The HWPoisoned page may be not in buddy system, and
6876 * page_count() is not 0.
6878 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6880 SetPageReserved(page);
6884 BUG_ON(page_count(page));
6885 BUG_ON(!PageBuddy(page));
6886 order = page_order(page);
6887 #ifdef CONFIG_DEBUG_VM
6888 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6889 pfn, 1 << order, end_pfn);
6891 list_del(&page->lru);
6892 rmv_page_order(page);
6893 zone->free_area[order].nr_free--;
6894 for (i = 0; i < (1 << order); i++)
6895 SetPageReserved((page+i));
6896 pfn += (1 << order);
6898 spin_unlock_irqrestore(&zone->lock, flags);
6902 #ifdef CONFIG_MEMORY_FAILURE
6903 bool is_free_buddy_page(struct page *page)
6905 struct zone *zone = page_zone(page);
6906 unsigned long pfn = page_to_pfn(page);
6907 unsigned long flags;
6910 spin_lock_irqsave(&zone->lock, flags);
6911 for (order = 0; order < MAX_ORDER; order++) {
6912 struct page *page_head = page - (pfn & ((1 << order) - 1));
6914 if (PageBuddy(page_head) && page_order(page_head) >= order)
6917 spin_unlock_irqrestore(&zone->lock, flags);
6919 return order < MAX_ORDER;