2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/notifier.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/stop_machine.h>
47 #include <linux/sort.h>
48 #include <linux/pfn.h>
49 #include <linux/backing-dev.h>
50 #include <linux/fault-inject.h>
51 #include <linux/page-isolation.h>
52 #include <linux/page_ext.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <linux/prefetch.h>
58 #include <linux/mm_inline.h>
59 #include <linux/migrate.h>
60 #include <linux/page_ext.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/page_owner.h>
64 #include <linux/kthread.h>
66 #include <asm/sections.h>
67 #include <asm/tlbflush.h>
68 #include <asm/div64.h>
71 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
72 static DEFINE_MUTEX(pcp_batch_high_lock);
73 #define MIN_PERCPU_PAGELIST_FRACTION (8)
75 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
76 DEFINE_PER_CPU(int, numa_node);
77 EXPORT_PER_CPU_SYMBOL(numa_node);
80 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
82 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
83 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
84 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
85 * defined in <linux/topology.h>.
87 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
88 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
89 int _node_numa_mem_[MAX_NUMNODES];
93 * Array of node states.
95 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
96 [N_POSSIBLE] = NODE_MASK_ALL,
97 [N_ONLINE] = { { [0] = 1UL } },
99 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
100 #ifdef CONFIG_HIGHMEM
101 [N_HIGH_MEMORY] = { { [0] = 1UL } },
103 #ifdef CONFIG_MOVABLE_NODE
104 [N_MEMORY] = { { [0] = 1UL } },
106 [N_CPU] = { { [0] = 1UL } },
109 EXPORT_SYMBOL(node_states);
111 /* Protect totalram_pages and zone->managed_pages */
112 static DEFINE_SPINLOCK(managed_page_count_lock);
114 unsigned long totalram_pages __read_mostly;
115 unsigned long totalreserve_pages __read_mostly;
116 unsigned long totalcma_pages __read_mostly;
118 * When calculating the number of globally allowed dirty pages, there
119 * is a certain number of per-zone reserves that should not be
120 * considered dirtyable memory. This is the sum of those reserves
121 * over all existing zones that contribute dirtyable memory.
123 unsigned long dirty_balance_reserve __read_mostly;
125 int percpu_pagelist_fraction;
126 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
129 * A cached value of the page's pageblock's migratetype, used when the page is
130 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
131 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
132 * Also the migratetype set in the page does not necessarily match the pcplist
133 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
134 * other index - this ensures that it will be put on the correct CMA freelist.
136 static inline int get_pcppage_migratetype(struct page *page)
141 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
143 page->index = migratetype;
146 #ifdef CONFIG_PM_SLEEP
148 * The following functions are used by the suspend/hibernate code to temporarily
149 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
150 * while devices are suspended. To avoid races with the suspend/hibernate code,
151 * they should always be called with pm_mutex held (gfp_allowed_mask also should
152 * only be modified with pm_mutex held, unless the suspend/hibernate code is
153 * guaranteed not to run in parallel with that modification).
156 static gfp_t saved_gfp_mask;
158 void pm_restore_gfp_mask(void)
160 WARN_ON(!mutex_is_locked(&pm_mutex));
161 if (saved_gfp_mask) {
162 gfp_allowed_mask = saved_gfp_mask;
167 void pm_restrict_gfp_mask(void)
169 WARN_ON(!mutex_is_locked(&pm_mutex));
170 WARN_ON(saved_gfp_mask);
171 saved_gfp_mask = gfp_allowed_mask;
172 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
175 bool pm_suspended_storage(void)
177 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
181 #endif /* CONFIG_PM_SLEEP */
183 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
184 unsigned int pageblock_order __read_mostly;
187 static void __free_pages_ok(struct page *page, unsigned int order);
190 * results with 256, 32 in the lowmem_reserve sysctl:
191 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
192 * 1G machine -> (16M dma, 784M normal, 224M high)
193 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
194 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
195 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
197 * TBD: should special case ZONE_DMA32 machines here - in those we normally
198 * don't need any ZONE_NORMAL reservation
200 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
201 #ifdef CONFIG_ZONE_DMA
204 #ifdef CONFIG_ZONE_DMA32
207 #ifdef CONFIG_HIGHMEM
213 EXPORT_SYMBOL(totalram_pages);
215 static char * const zone_names[MAX_NR_ZONES] = {
216 #ifdef CONFIG_ZONE_DMA
219 #ifdef CONFIG_ZONE_DMA32
223 #ifdef CONFIG_HIGHMEM
227 #ifdef CONFIG_ZONE_DEVICE
232 static void free_compound_page(struct page *page);
233 compound_page_dtor * const compound_page_dtors[] = {
236 #ifdef CONFIG_HUGETLB_PAGE
241 int min_free_kbytes = 1024;
242 int user_min_free_kbytes = -1;
244 static unsigned long __meminitdata nr_kernel_pages;
245 static unsigned long __meminitdata nr_all_pages;
246 static unsigned long __meminitdata dma_reserve;
248 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
249 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
250 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
251 static unsigned long __initdata required_kernelcore;
252 static unsigned long __initdata required_movablecore;
253 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
255 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
257 EXPORT_SYMBOL(movable_zone);
258 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
261 int nr_node_ids __read_mostly = MAX_NUMNODES;
262 int nr_online_nodes __read_mostly = 1;
263 EXPORT_SYMBOL(nr_node_ids);
264 EXPORT_SYMBOL(nr_online_nodes);
267 int page_group_by_mobility_disabled __read_mostly;
269 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
270 static inline void reset_deferred_meminit(pg_data_t *pgdat)
272 pgdat->first_deferred_pfn = ULONG_MAX;
275 /* Returns true if the struct page for the pfn is uninitialised */
276 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
278 int nid = early_pfn_to_nid(pfn);
280 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
286 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
288 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
295 * Returns false when the remaining initialisation should be deferred until
296 * later in the boot cycle when it can be parallelised.
298 static inline bool update_defer_init(pg_data_t *pgdat,
299 unsigned long pfn, unsigned long zone_end,
300 unsigned long *nr_initialised)
302 /* Always populate low zones for address-contrained allocations */
303 if (zone_end < pgdat_end_pfn(pgdat))
306 /* Initialise at least 2G of the highest zone */
308 if (*nr_initialised > (2UL << (30 - PAGE_SHIFT)) &&
309 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
310 pgdat->first_deferred_pfn = pfn;
317 static inline void reset_deferred_meminit(pg_data_t *pgdat)
321 static inline bool early_page_uninitialised(unsigned long pfn)
326 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
331 static inline bool update_defer_init(pg_data_t *pgdat,
332 unsigned long pfn, unsigned long zone_end,
333 unsigned long *nr_initialised)
340 void set_pageblock_migratetype(struct page *page, int migratetype)
342 if (unlikely(page_group_by_mobility_disabled &&
343 migratetype < MIGRATE_PCPTYPES))
344 migratetype = MIGRATE_UNMOVABLE;
346 set_pageblock_flags_group(page, (unsigned long)migratetype,
347 PB_migrate, PB_migrate_end);
350 #ifdef CONFIG_DEBUG_VM
351 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
355 unsigned long pfn = page_to_pfn(page);
356 unsigned long sp, start_pfn;
359 seq = zone_span_seqbegin(zone);
360 start_pfn = zone->zone_start_pfn;
361 sp = zone->spanned_pages;
362 if (!zone_spans_pfn(zone, pfn))
364 } while (zone_span_seqretry(zone, seq));
367 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
368 pfn, zone_to_nid(zone), zone->name,
369 start_pfn, start_pfn + sp);
374 static int page_is_consistent(struct zone *zone, struct page *page)
376 if (!pfn_valid_within(page_to_pfn(page)))
378 if (zone != page_zone(page))
384 * Temporary debugging check for pages not lying within a given zone.
386 static int bad_range(struct zone *zone, struct page *page)
388 if (page_outside_zone_boundaries(zone, page))
390 if (!page_is_consistent(zone, page))
396 static inline int bad_range(struct zone *zone, struct page *page)
402 static void bad_page(struct page *page, const char *reason,
403 unsigned long bad_flags)
405 static unsigned long resume;
406 static unsigned long nr_shown;
407 static unsigned long nr_unshown;
409 /* Don't complain about poisoned pages */
410 if (PageHWPoison(page)) {
411 page_mapcount_reset(page); /* remove PageBuddy */
416 * Allow a burst of 60 reports, then keep quiet for that minute;
417 * or allow a steady drip of one report per second.
419 if (nr_shown == 60) {
420 if (time_before(jiffies, resume)) {
426 "BUG: Bad page state: %lu messages suppressed\n",
433 resume = jiffies + 60 * HZ;
435 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
436 current->comm, page_to_pfn(page));
437 dump_page_badflags(page, reason, bad_flags);
442 /* Leave bad fields for debug, except PageBuddy could make trouble */
443 page_mapcount_reset(page); /* remove PageBuddy */
444 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
448 * Higher-order pages are called "compound pages". They are structured thusly:
450 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
452 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
453 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
455 * The first tail page's ->compound_dtor holds the offset in array of compound
456 * page destructors. See compound_page_dtors.
458 * The first tail page's ->compound_order holds the order of allocation.
459 * This usage means that zero-order pages may not be compound.
462 static void free_compound_page(struct page *page)
464 __free_pages_ok(page, compound_order(page));
467 void prep_compound_page(struct page *page, unsigned int order)
470 int nr_pages = 1 << order;
472 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
473 set_compound_order(page, order);
475 for (i = 1; i < nr_pages; i++) {
476 struct page *p = page + i;
477 set_page_count(p, 0);
478 set_compound_head(p, page);
482 #ifdef CONFIG_DEBUG_PAGEALLOC
483 unsigned int _debug_guardpage_minorder;
484 bool _debug_pagealloc_enabled __read_mostly;
485 bool _debug_guardpage_enabled __read_mostly;
487 static int __init early_debug_pagealloc(char *buf)
492 if (strcmp(buf, "on") == 0)
493 _debug_pagealloc_enabled = true;
497 early_param("debug_pagealloc", early_debug_pagealloc);
499 static bool need_debug_guardpage(void)
501 /* If we don't use debug_pagealloc, we don't need guard page */
502 if (!debug_pagealloc_enabled())
508 static void init_debug_guardpage(void)
510 if (!debug_pagealloc_enabled())
513 _debug_guardpage_enabled = true;
516 struct page_ext_operations debug_guardpage_ops = {
517 .need = need_debug_guardpage,
518 .init = init_debug_guardpage,
521 static int __init debug_guardpage_minorder_setup(char *buf)
525 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
526 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
529 _debug_guardpage_minorder = res;
530 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
533 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
535 static inline void set_page_guard(struct zone *zone, struct page *page,
536 unsigned int order, int migratetype)
538 struct page_ext *page_ext;
540 if (!debug_guardpage_enabled())
543 page_ext = lookup_page_ext(page);
544 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
546 INIT_LIST_HEAD(&page->lru);
547 set_page_private(page, order);
548 /* Guard pages are not available for any usage */
549 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
552 static inline void clear_page_guard(struct zone *zone, struct page *page,
553 unsigned int order, int migratetype)
555 struct page_ext *page_ext;
557 if (!debug_guardpage_enabled())
560 page_ext = lookup_page_ext(page);
561 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
563 set_page_private(page, 0);
564 if (!is_migrate_isolate(migratetype))
565 __mod_zone_freepage_state(zone, (1 << order), migratetype);
568 struct page_ext_operations debug_guardpage_ops = { NULL, };
569 static inline void set_page_guard(struct zone *zone, struct page *page,
570 unsigned int order, int migratetype) {}
571 static inline void clear_page_guard(struct zone *zone, struct page *page,
572 unsigned int order, int migratetype) {}
575 static inline void set_page_order(struct page *page, unsigned int order)
577 set_page_private(page, order);
578 __SetPageBuddy(page);
581 static inline void rmv_page_order(struct page *page)
583 __ClearPageBuddy(page);
584 set_page_private(page, 0);
588 * This function checks whether a page is free && is the buddy
589 * we can do coalesce a page and its buddy if
590 * (a) the buddy is not in a hole &&
591 * (b) the buddy is in the buddy system &&
592 * (c) a page and its buddy have the same order &&
593 * (d) a page and its buddy are in the same zone.
595 * For recording whether a page is in the buddy system, we set ->_mapcount
596 * PAGE_BUDDY_MAPCOUNT_VALUE.
597 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
598 * serialized by zone->lock.
600 * For recording page's order, we use page_private(page).
602 static inline int page_is_buddy(struct page *page, struct page *buddy,
605 if (!pfn_valid_within(page_to_pfn(buddy)))
608 if (page_is_guard(buddy) && page_order(buddy) == order) {
609 if (page_zone_id(page) != page_zone_id(buddy))
612 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
617 if (PageBuddy(buddy) && page_order(buddy) == order) {
619 * zone check is done late to avoid uselessly
620 * calculating zone/node ids for pages that could
623 if (page_zone_id(page) != page_zone_id(buddy))
626 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
634 * Freeing function for a buddy system allocator.
636 * The concept of a buddy system is to maintain direct-mapped table
637 * (containing bit values) for memory blocks of various "orders".
638 * The bottom level table contains the map for the smallest allocatable
639 * units of memory (here, pages), and each level above it describes
640 * pairs of units from the levels below, hence, "buddies".
641 * At a high level, all that happens here is marking the table entry
642 * at the bottom level available, and propagating the changes upward
643 * as necessary, plus some accounting needed to play nicely with other
644 * parts of the VM system.
645 * At each level, we keep a list of pages, which are heads of continuous
646 * free pages of length of (1 << order) and marked with _mapcount
647 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
649 * So when we are allocating or freeing one, we can derive the state of the
650 * other. That is, if we allocate a small block, and both were
651 * free, the remainder of the region must be split into blocks.
652 * If a block is freed, and its buddy is also free, then this
653 * triggers coalescing into a block of larger size.
658 static inline void __free_one_page(struct page *page,
660 struct zone *zone, unsigned int order,
663 unsigned long page_idx;
664 unsigned long combined_idx;
665 unsigned long uninitialized_var(buddy_idx);
667 unsigned int max_order;
669 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
671 VM_BUG_ON(!zone_is_initialized(zone));
672 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
674 VM_BUG_ON(migratetype == -1);
675 if (likely(!is_migrate_isolate(migratetype)))
676 __mod_zone_freepage_state(zone, 1 << order, migratetype);
678 page_idx = pfn & ((1 << MAX_ORDER) - 1);
680 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
681 VM_BUG_ON_PAGE(bad_range(zone, page), page);
684 while (order < max_order - 1) {
685 buddy_idx = __find_buddy_index(page_idx, order);
686 buddy = page + (buddy_idx - page_idx);
687 if (!page_is_buddy(page, buddy, order))
690 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
691 * merge with it and move up one order.
693 if (page_is_guard(buddy)) {
694 clear_page_guard(zone, buddy, order, migratetype);
696 list_del(&buddy->lru);
697 zone->free_area[order].nr_free--;
698 rmv_page_order(buddy);
700 combined_idx = buddy_idx & page_idx;
701 page = page + (combined_idx - page_idx);
702 page_idx = combined_idx;
705 if (max_order < MAX_ORDER) {
706 /* If we are here, it means order is >= pageblock_order.
707 * We want to prevent merge between freepages on isolate
708 * pageblock and normal pageblock. Without this, pageblock
709 * isolation could cause incorrect freepage or CMA accounting.
711 * We don't want to hit this code for the more frequent
714 if (unlikely(has_isolate_pageblock(zone))) {
717 buddy_idx = __find_buddy_index(page_idx, order);
718 buddy = page + (buddy_idx - page_idx);
719 buddy_mt = get_pageblock_migratetype(buddy);
721 if (migratetype != buddy_mt
722 && (is_migrate_isolate(migratetype) ||
723 is_migrate_isolate(buddy_mt)))
727 goto continue_merging;
731 set_page_order(page, order);
734 * If this is not the largest possible page, check if the buddy
735 * of the next-highest order is free. If it is, it's possible
736 * that pages are being freed that will coalesce soon. In case,
737 * that is happening, add the free page to the tail of the list
738 * so it's less likely to be used soon and more likely to be merged
739 * as a higher order page
741 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
742 struct page *higher_page, *higher_buddy;
743 combined_idx = buddy_idx & page_idx;
744 higher_page = page + (combined_idx - page_idx);
745 buddy_idx = __find_buddy_index(combined_idx, order + 1);
746 higher_buddy = higher_page + (buddy_idx - combined_idx);
747 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
748 list_add_tail(&page->lru,
749 &zone->free_area[order].free_list[migratetype]);
754 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
756 zone->free_area[order].nr_free++;
759 static inline int free_pages_check(struct page *page)
761 const char *bad_reason = NULL;
762 unsigned long bad_flags = 0;
764 if (unlikely(page_mapcount(page)))
765 bad_reason = "nonzero mapcount";
766 if (unlikely(page->mapping != NULL))
767 bad_reason = "non-NULL mapping";
768 if (unlikely(atomic_read(&page->_count) != 0))
769 bad_reason = "nonzero _count";
770 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
771 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
772 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
775 if (unlikely(page->mem_cgroup))
776 bad_reason = "page still charged to cgroup";
778 if (unlikely(bad_reason)) {
779 bad_page(page, bad_reason, bad_flags);
782 page_cpupid_reset_last(page);
783 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
784 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
789 * Frees a number of pages from the PCP lists
790 * Assumes all pages on list are in same zone, and of same order.
791 * count is the number of pages to free.
793 * If the zone was previously in an "all pages pinned" state then look to
794 * see if this freeing clears that state.
796 * And clear the zone's pages_scanned counter, to hold off the "all pages are
797 * pinned" detection logic.
799 static void free_pcppages_bulk(struct zone *zone, int count,
800 struct per_cpu_pages *pcp)
805 unsigned long nr_scanned;
807 spin_lock(&zone->lock);
808 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
810 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
814 struct list_head *list;
817 * Remove pages from lists in a round-robin fashion. A
818 * batch_free count is maintained that is incremented when an
819 * empty list is encountered. This is so more pages are freed
820 * off fuller lists instead of spinning excessively around empty
825 if (++migratetype == MIGRATE_PCPTYPES)
827 list = &pcp->lists[migratetype];
828 } while (list_empty(list));
830 /* This is the only non-empty list. Free them all. */
831 if (batch_free == MIGRATE_PCPTYPES)
832 batch_free = to_free;
835 int mt; /* migratetype of the to-be-freed page */
837 page = list_entry(list->prev, struct page, lru);
838 /* must delete as __free_one_page list manipulates */
839 list_del(&page->lru);
841 mt = get_pcppage_migratetype(page);
842 /* MIGRATE_ISOLATE page should not go to pcplists */
843 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
844 /* Pageblock could have been isolated meanwhile */
845 if (unlikely(has_isolate_pageblock(zone)))
846 mt = get_pageblock_migratetype(page);
848 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
849 trace_mm_page_pcpu_drain(page, 0, mt);
850 } while (--to_free && --batch_free && !list_empty(list));
852 spin_unlock(&zone->lock);
855 static void free_one_page(struct zone *zone,
856 struct page *page, unsigned long pfn,
860 unsigned long nr_scanned;
861 spin_lock(&zone->lock);
862 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
864 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
866 if (unlikely(has_isolate_pageblock(zone) ||
867 is_migrate_isolate(migratetype))) {
868 migratetype = get_pfnblock_migratetype(page, pfn);
870 __free_one_page(page, pfn, zone, order, migratetype);
871 spin_unlock(&zone->lock);
874 static int free_tail_pages_check(struct page *head_page, struct page *page)
879 * We rely page->lru.next never has bit 0 set, unless the page
880 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
882 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
884 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
888 if (unlikely(!PageTail(page))) {
889 bad_page(page, "PageTail not set", 0);
892 if (unlikely(compound_head(page) != head_page)) {
893 bad_page(page, "compound_head not consistent", 0);
898 clear_compound_head(page);
902 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
903 unsigned long zone, int nid)
905 set_page_links(page, zone, nid, pfn);
906 init_page_count(page);
907 page_mapcount_reset(page);
908 page_cpupid_reset_last(page);
910 INIT_LIST_HEAD(&page->lru);
911 #ifdef WANT_PAGE_VIRTUAL
912 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
913 if (!is_highmem_idx(zone))
914 set_page_address(page, __va(pfn << PAGE_SHIFT));
918 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
921 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
924 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
925 static void init_reserved_page(unsigned long pfn)
930 if (!early_page_uninitialised(pfn))
933 nid = early_pfn_to_nid(pfn);
934 pgdat = NODE_DATA(nid);
936 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
937 struct zone *zone = &pgdat->node_zones[zid];
939 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
942 __init_single_pfn(pfn, zid, nid);
945 static inline void init_reserved_page(unsigned long pfn)
948 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
951 * Initialised pages do not have PageReserved set. This function is
952 * called for each range allocated by the bootmem allocator and
953 * marks the pages PageReserved. The remaining valid pages are later
954 * sent to the buddy page allocator.
956 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
958 unsigned long start_pfn = PFN_DOWN(start);
959 unsigned long end_pfn = PFN_UP(end);
961 for (; start_pfn < end_pfn; start_pfn++) {
962 if (pfn_valid(start_pfn)) {
963 struct page *page = pfn_to_page(start_pfn);
965 init_reserved_page(start_pfn);
967 /* Avoid false-positive PageTail() */
968 INIT_LIST_HEAD(&page->lru);
970 SetPageReserved(page);
975 static bool free_pages_prepare(struct page *page, unsigned int order)
977 bool compound = PageCompound(page);
980 VM_BUG_ON_PAGE(PageTail(page), page);
981 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
983 trace_mm_page_free(page, order);
984 kmemcheck_free_shadow(page, order);
985 kasan_free_pages(page, order);
988 page->mapping = NULL;
989 bad += free_pages_check(page);
990 for (i = 1; i < (1 << order); i++) {
992 bad += free_tail_pages_check(page, page + i);
993 bad += free_pages_check(page + i);
998 reset_page_owner(page, order);
1000 if (!PageHighMem(page)) {
1001 debug_check_no_locks_freed(page_address(page),
1002 PAGE_SIZE << order);
1003 debug_check_no_obj_freed(page_address(page),
1004 PAGE_SIZE << order);
1006 arch_free_page(page, order);
1007 kernel_map_pages(page, 1 << order, 0);
1012 static void __free_pages_ok(struct page *page, unsigned int order)
1014 unsigned long flags;
1016 unsigned long pfn = page_to_pfn(page);
1018 if (!free_pages_prepare(page, order))
1021 migratetype = get_pfnblock_migratetype(page, pfn);
1022 local_irq_save(flags);
1023 __count_vm_events(PGFREE, 1 << order);
1024 free_one_page(page_zone(page), page, pfn, order, migratetype);
1025 local_irq_restore(flags);
1028 static void __init __free_pages_boot_core(struct page *page,
1029 unsigned long pfn, unsigned int order)
1031 unsigned int nr_pages = 1 << order;
1032 struct page *p = page;
1036 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1038 __ClearPageReserved(p);
1039 set_page_count(p, 0);
1041 __ClearPageReserved(p);
1042 set_page_count(p, 0);
1044 page_zone(page)->managed_pages += nr_pages;
1045 set_page_refcounted(page);
1046 __free_pages(page, order);
1049 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1050 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1052 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1054 int __meminit early_pfn_to_nid(unsigned long pfn)
1056 static DEFINE_SPINLOCK(early_pfn_lock);
1059 spin_lock(&early_pfn_lock);
1060 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1062 nid = first_online_node;
1063 spin_unlock(&early_pfn_lock);
1069 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1070 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1071 struct mminit_pfnnid_cache *state)
1075 nid = __early_pfn_to_nid(pfn, state);
1076 if (nid >= 0 && nid != node)
1081 /* Only safe to use early in boot when initialisation is single-threaded */
1082 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1084 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1089 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1093 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1094 struct mminit_pfnnid_cache *state)
1101 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1104 if (early_page_uninitialised(pfn))
1106 return __free_pages_boot_core(page, pfn, order);
1109 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1110 static void __init deferred_free_range(struct page *page,
1111 unsigned long pfn, int nr_pages)
1118 /* Free a large naturally-aligned chunk if possible */
1119 if (nr_pages == MAX_ORDER_NR_PAGES &&
1120 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1121 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1122 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1126 for (i = 0; i < nr_pages; i++, page++, pfn++)
1127 __free_pages_boot_core(page, pfn, 0);
1130 /* Completion tracking for deferred_init_memmap() threads */
1131 static atomic_t pgdat_init_n_undone __initdata;
1132 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1134 static inline void __init pgdat_init_report_one_done(void)
1136 if (atomic_dec_and_test(&pgdat_init_n_undone))
1137 complete(&pgdat_init_all_done_comp);
1140 /* Initialise remaining memory on a node */
1141 static int __init deferred_init_memmap(void *data)
1143 pg_data_t *pgdat = data;
1144 int nid = pgdat->node_id;
1145 struct mminit_pfnnid_cache nid_init_state = { };
1146 unsigned long start = jiffies;
1147 unsigned long nr_pages = 0;
1148 unsigned long walk_start, walk_end;
1151 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1152 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1154 if (first_init_pfn == ULONG_MAX) {
1155 pgdat_init_report_one_done();
1159 /* Bind memory initialisation thread to a local node if possible */
1160 if (!cpumask_empty(cpumask))
1161 set_cpus_allowed_ptr(current, cpumask);
1163 /* Sanity check boundaries */
1164 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1165 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1166 pgdat->first_deferred_pfn = ULONG_MAX;
1168 /* Only the highest zone is deferred so find it */
1169 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1170 zone = pgdat->node_zones + zid;
1171 if (first_init_pfn < zone_end_pfn(zone))
1175 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1176 unsigned long pfn, end_pfn;
1177 struct page *page = NULL;
1178 struct page *free_base_page = NULL;
1179 unsigned long free_base_pfn = 0;
1182 end_pfn = min(walk_end, zone_end_pfn(zone));
1183 pfn = first_init_pfn;
1184 if (pfn < walk_start)
1186 if (pfn < zone->zone_start_pfn)
1187 pfn = zone->zone_start_pfn;
1189 for (; pfn < end_pfn; pfn++) {
1190 if (!pfn_valid_within(pfn))
1194 * Ensure pfn_valid is checked every
1195 * MAX_ORDER_NR_PAGES for memory holes
1197 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1198 if (!pfn_valid(pfn)) {
1204 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1209 /* Minimise pfn page lookups and scheduler checks */
1210 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1213 nr_pages += nr_to_free;
1214 deferred_free_range(free_base_page,
1215 free_base_pfn, nr_to_free);
1216 free_base_page = NULL;
1217 free_base_pfn = nr_to_free = 0;
1219 page = pfn_to_page(pfn);
1224 VM_BUG_ON(page_zone(page) != zone);
1228 __init_single_page(page, pfn, zid, nid);
1229 if (!free_base_page) {
1230 free_base_page = page;
1231 free_base_pfn = pfn;
1236 /* Where possible, batch up pages for a single free */
1239 /* Free the current block of pages to allocator */
1240 nr_pages += nr_to_free;
1241 deferred_free_range(free_base_page, free_base_pfn,
1243 free_base_page = NULL;
1244 free_base_pfn = nr_to_free = 0;
1247 first_init_pfn = max(end_pfn, first_init_pfn);
1250 /* Sanity check that the next zone really is unpopulated */
1251 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1253 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1254 jiffies_to_msecs(jiffies - start));
1256 pgdat_init_report_one_done();
1260 void __init page_alloc_init_late(void)
1264 /* There will be num_node_state(N_MEMORY) threads */
1265 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1266 for_each_node_state(nid, N_MEMORY) {
1267 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1270 /* Block until all are initialised */
1271 wait_for_completion(&pgdat_init_all_done_comp);
1273 /* Reinit limits that are based on free pages after the kernel is up */
1274 files_maxfiles_init();
1276 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1279 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1280 void __init init_cma_reserved_pageblock(struct page *page)
1282 unsigned i = pageblock_nr_pages;
1283 struct page *p = page;
1286 __ClearPageReserved(p);
1287 set_page_count(p, 0);
1290 set_pageblock_migratetype(page, MIGRATE_CMA);
1292 if (pageblock_order >= MAX_ORDER) {
1293 i = pageblock_nr_pages;
1296 set_page_refcounted(p);
1297 __free_pages(p, MAX_ORDER - 1);
1298 p += MAX_ORDER_NR_PAGES;
1299 } while (i -= MAX_ORDER_NR_PAGES);
1301 set_page_refcounted(page);
1302 __free_pages(page, pageblock_order);
1305 adjust_managed_page_count(page, pageblock_nr_pages);
1310 * The order of subdivision here is critical for the IO subsystem.
1311 * Please do not alter this order without good reasons and regression
1312 * testing. Specifically, as large blocks of memory are subdivided,
1313 * the order in which smaller blocks are delivered depends on the order
1314 * they're subdivided in this function. This is the primary factor
1315 * influencing the order in which pages are delivered to the IO
1316 * subsystem according to empirical testing, and this is also justified
1317 * by considering the behavior of a buddy system containing a single
1318 * large block of memory acted on by a series of small allocations.
1319 * This behavior is a critical factor in sglist merging's success.
1323 static inline void expand(struct zone *zone, struct page *page,
1324 int low, int high, struct free_area *area,
1327 unsigned long size = 1 << high;
1329 while (high > low) {
1333 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1335 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1336 debug_guardpage_enabled() &&
1337 high < debug_guardpage_minorder()) {
1339 * Mark as guard pages (or page), that will allow to
1340 * merge back to allocator when buddy will be freed.
1341 * Corresponding page table entries will not be touched,
1342 * pages will stay not present in virtual address space
1344 set_page_guard(zone, &page[size], high, migratetype);
1347 list_add(&page[size].lru, &area->free_list[migratetype]);
1349 set_page_order(&page[size], high);
1354 * This page is about to be returned from the page allocator
1356 static inline int check_new_page(struct page *page)
1358 const char *bad_reason = NULL;
1359 unsigned long bad_flags = 0;
1361 if (unlikely(page_mapcount(page)))
1362 bad_reason = "nonzero mapcount";
1363 if (unlikely(page->mapping != NULL))
1364 bad_reason = "non-NULL mapping";
1365 if (unlikely(atomic_read(&page->_count) != 0))
1366 bad_reason = "nonzero _count";
1367 if (unlikely(page->flags & __PG_HWPOISON)) {
1368 bad_reason = "HWPoisoned (hardware-corrupted)";
1369 bad_flags = __PG_HWPOISON;
1371 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1372 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1373 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1376 if (unlikely(page->mem_cgroup))
1377 bad_reason = "page still charged to cgroup";
1379 if (unlikely(bad_reason)) {
1380 bad_page(page, bad_reason, bad_flags);
1386 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1391 for (i = 0; i < (1 << order); i++) {
1392 struct page *p = page + i;
1393 if (unlikely(check_new_page(p)))
1397 set_page_private(page, 0);
1398 set_page_refcounted(page);
1400 arch_alloc_page(page, order);
1401 kernel_map_pages(page, 1 << order, 1);
1402 kasan_alloc_pages(page, order);
1404 if (gfp_flags & __GFP_ZERO)
1405 for (i = 0; i < (1 << order); i++)
1406 clear_highpage(page + i);
1408 if (order && (gfp_flags & __GFP_COMP))
1409 prep_compound_page(page, order);
1411 set_page_owner(page, order, gfp_flags);
1414 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1415 * allocate the page. The expectation is that the caller is taking
1416 * steps that will free more memory. The caller should avoid the page
1417 * being used for !PFMEMALLOC purposes.
1419 if (alloc_flags & ALLOC_NO_WATERMARKS)
1420 set_page_pfmemalloc(page);
1422 clear_page_pfmemalloc(page);
1428 * Go through the free lists for the given migratetype and remove
1429 * the smallest available page from the freelists
1432 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1435 unsigned int current_order;
1436 struct free_area *area;
1439 /* Find a page of the appropriate size in the preferred list */
1440 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1441 area = &(zone->free_area[current_order]);
1442 if (list_empty(&area->free_list[migratetype]))
1445 page = list_entry(area->free_list[migratetype].next,
1447 list_del(&page->lru);
1448 rmv_page_order(page);
1450 expand(zone, page, order, current_order, area, migratetype);
1451 set_pcppage_migratetype(page, migratetype);
1460 * This array describes the order lists are fallen back to when
1461 * the free lists for the desirable migrate type are depleted
1463 static int fallbacks[MIGRATE_TYPES][4] = {
1464 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1465 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1466 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1468 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1470 #ifdef CONFIG_MEMORY_ISOLATION
1471 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1476 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1479 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1482 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1483 unsigned int order) { return NULL; }
1487 * Move the free pages in a range to the free lists of the requested type.
1488 * Note that start_page and end_pages are not aligned on a pageblock
1489 * boundary. If alignment is required, use move_freepages_block()
1491 int move_freepages(struct zone *zone,
1492 struct page *start_page, struct page *end_page,
1497 int pages_moved = 0;
1499 #ifndef CONFIG_HOLES_IN_ZONE
1501 * page_zone is not safe to call in this context when
1502 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1503 * anyway as we check zone boundaries in move_freepages_block().
1504 * Remove at a later date when no bug reports exist related to
1505 * grouping pages by mobility
1507 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1510 for (page = start_page; page <= end_page;) {
1511 /* Make sure we are not inadvertently changing nodes */
1512 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1514 if (!pfn_valid_within(page_to_pfn(page))) {
1519 if (!PageBuddy(page)) {
1524 order = page_order(page);
1525 list_move(&page->lru,
1526 &zone->free_area[order].free_list[migratetype]);
1528 pages_moved += 1 << order;
1534 int move_freepages_block(struct zone *zone, struct page *page,
1537 unsigned long start_pfn, end_pfn;
1538 struct page *start_page, *end_page;
1540 start_pfn = page_to_pfn(page);
1541 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1542 start_page = pfn_to_page(start_pfn);
1543 end_page = start_page + pageblock_nr_pages - 1;
1544 end_pfn = start_pfn + pageblock_nr_pages - 1;
1546 /* Do not cross zone boundaries */
1547 if (!zone_spans_pfn(zone, start_pfn))
1549 if (!zone_spans_pfn(zone, end_pfn))
1552 return move_freepages(zone, start_page, end_page, migratetype);
1555 static void change_pageblock_range(struct page *pageblock_page,
1556 int start_order, int migratetype)
1558 int nr_pageblocks = 1 << (start_order - pageblock_order);
1560 while (nr_pageblocks--) {
1561 set_pageblock_migratetype(pageblock_page, migratetype);
1562 pageblock_page += pageblock_nr_pages;
1567 * When we are falling back to another migratetype during allocation, try to
1568 * steal extra free pages from the same pageblocks to satisfy further
1569 * allocations, instead of polluting multiple pageblocks.
1571 * If we are stealing a relatively large buddy page, it is likely there will
1572 * be more free pages in the pageblock, so try to steal them all. For
1573 * reclaimable and unmovable allocations, we steal regardless of page size,
1574 * as fragmentation caused by those allocations polluting movable pageblocks
1575 * is worse than movable allocations stealing from unmovable and reclaimable
1578 static bool can_steal_fallback(unsigned int order, int start_mt)
1581 * Leaving this order check is intended, although there is
1582 * relaxed order check in next check. The reason is that
1583 * we can actually steal whole pageblock if this condition met,
1584 * but, below check doesn't guarantee it and that is just heuristic
1585 * so could be changed anytime.
1587 if (order >= pageblock_order)
1590 if (order >= pageblock_order / 2 ||
1591 start_mt == MIGRATE_RECLAIMABLE ||
1592 start_mt == MIGRATE_UNMOVABLE ||
1593 page_group_by_mobility_disabled)
1600 * This function implements actual steal behaviour. If order is large enough,
1601 * we can steal whole pageblock. If not, we first move freepages in this
1602 * pageblock and check whether half of pages are moved or not. If half of
1603 * pages are moved, we can change migratetype of pageblock and permanently
1604 * use it's pages as requested migratetype in the future.
1606 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1609 unsigned int current_order = page_order(page);
1612 /* Take ownership for orders >= pageblock_order */
1613 if (current_order >= pageblock_order) {
1614 change_pageblock_range(page, current_order, start_type);
1618 pages = move_freepages_block(zone, page, start_type);
1620 /* Claim the whole block if over half of it is free */
1621 if (pages >= (1 << (pageblock_order-1)) ||
1622 page_group_by_mobility_disabled)
1623 set_pageblock_migratetype(page, start_type);
1627 * Check whether there is a suitable fallback freepage with requested order.
1628 * If only_stealable is true, this function returns fallback_mt only if
1629 * we can steal other freepages all together. This would help to reduce
1630 * fragmentation due to mixed migratetype pages in one pageblock.
1632 int find_suitable_fallback(struct free_area *area, unsigned int order,
1633 int migratetype, bool only_stealable, bool *can_steal)
1638 if (area->nr_free == 0)
1643 fallback_mt = fallbacks[migratetype][i];
1644 if (fallback_mt == MIGRATE_TYPES)
1647 if (list_empty(&area->free_list[fallback_mt]))
1650 if (can_steal_fallback(order, migratetype))
1653 if (!only_stealable)
1664 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1665 * there are no empty page blocks that contain a page with a suitable order
1667 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1668 unsigned int alloc_order)
1671 unsigned long max_managed, flags;
1674 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1675 * Check is race-prone but harmless.
1677 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1678 if (zone->nr_reserved_highatomic >= max_managed)
1681 spin_lock_irqsave(&zone->lock, flags);
1683 /* Recheck the nr_reserved_highatomic limit under the lock */
1684 if (zone->nr_reserved_highatomic >= max_managed)
1688 mt = get_pageblock_migratetype(page);
1689 if (mt != MIGRATE_HIGHATOMIC &&
1690 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1691 zone->nr_reserved_highatomic += pageblock_nr_pages;
1692 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1693 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1697 spin_unlock_irqrestore(&zone->lock, flags);
1701 * Used when an allocation is about to fail under memory pressure. This
1702 * potentially hurts the reliability of high-order allocations when under
1703 * intense memory pressure but failed atomic allocations should be easier
1704 * to recover from than an OOM.
1706 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1708 struct zonelist *zonelist = ac->zonelist;
1709 unsigned long flags;
1715 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1717 /* Preserve at least one pageblock */
1718 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1721 spin_lock_irqsave(&zone->lock, flags);
1722 for (order = 0; order < MAX_ORDER; order++) {
1723 struct free_area *area = &(zone->free_area[order]);
1725 if (list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
1728 page = list_entry(area->free_list[MIGRATE_HIGHATOMIC].next,
1732 * It should never happen but changes to locking could
1733 * inadvertently allow a per-cpu drain to add pages
1734 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1735 * and watch for underflows.
1737 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
1738 zone->nr_reserved_highatomic);
1741 * Convert to ac->migratetype and avoid the normal
1742 * pageblock stealing heuristics. Minimally, the caller
1743 * is doing the work and needs the pages. More
1744 * importantly, if the block was always converted to
1745 * MIGRATE_UNMOVABLE or another type then the number
1746 * of pageblocks that cannot be completely freed
1749 set_pageblock_migratetype(page, ac->migratetype);
1750 move_freepages_block(zone, page, ac->migratetype);
1751 spin_unlock_irqrestore(&zone->lock, flags);
1754 spin_unlock_irqrestore(&zone->lock, flags);
1758 /* Remove an element from the buddy allocator from the fallback list */
1759 static inline struct page *
1760 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1762 struct free_area *area;
1763 unsigned int current_order;
1768 /* Find the largest possible block of pages in the other list */
1769 for (current_order = MAX_ORDER-1;
1770 current_order >= order && current_order <= MAX_ORDER-1;
1772 area = &(zone->free_area[current_order]);
1773 fallback_mt = find_suitable_fallback(area, current_order,
1774 start_migratetype, false, &can_steal);
1775 if (fallback_mt == -1)
1778 page = list_entry(area->free_list[fallback_mt].next,
1781 steal_suitable_fallback(zone, page, start_migratetype);
1783 /* Remove the page from the freelists */
1785 list_del(&page->lru);
1786 rmv_page_order(page);
1788 expand(zone, page, order, current_order, area,
1791 * The pcppage_migratetype may differ from pageblock's
1792 * migratetype depending on the decisions in
1793 * find_suitable_fallback(). This is OK as long as it does not
1794 * differ for MIGRATE_CMA pageblocks. Those can be used as
1795 * fallback only via special __rmqueue_cma_fallback() function
1797 set_pcppage_migratetype(page, start_migratetype);
1799 trace_mm_page_alloc_extfrag(page, order, current_order,
1800 start_migratetype, fallback_mt);
1809 * Do the hard work of removing an element from the buddy allocator.
1810 * Call me with the zone->lock already held.
1812 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1813 int migratetype, gfp_t gfp_flags)
1817 page = __rmqueue_smallest(zone, order, migratetype);
1818 if (unlikely(!page)) {
1819 if (migratetype == MIGRATE_MOVABLE)
1820 page = __rmqueue_cma_fallback(zone, order);
1823 page = __rmqueue_fallback(zone, order, migratetype);
1826 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1831 * Obtain a specified number of elements from the buddy allocator, all under
1832 * a single hold of the lock, for efficiency. Add them to the supplied list.
1833 * Returns the number of new pages which were placed at *list.
1835 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1836 unsigned long count, struct list_head *list,
1837 int migratetype, bool cold)
1841 spin_lock(&zone->lock);
1842 for (i = 0; i < count; ++i) {
1843 struct page *page = __rmqueue(zone, order, migratetype, 0);
1844 if (unlikely(page == NULL))
1848 * Split buddy pages returned by expand() are received here
1849 * in physical page order. The page is added to the callers and
1850 * list and the list head then moves forward. From the callers
1851 * perspective, the linked list is ordered by page number in
1852 * some conditions. This is useful for IO devices that can
1853 * merge IO requests if the physical pages are ordered
1857 list_add(&page->lru, list);
1859 list_add_tail(&page->lru, list);
1861 if (is_migrate_cma(get_pcppage_migratetype(page)))
1862 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1865 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1866 spin_unlock(&zone->lock);
1872 * Called from the vmstat counter updater to drain pagesets of this
1873 * currently executing processor on remote nodes after they have
1876 * Note that this function must be called with the thread pinned to
1877 * a single processor.
1879 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1881 unsigned long flags;
1882 int to_drain, batch;
1884 local_irq_save(flags);
1885 batch = READ_ONCE(pcp->batch);
1886 to_drain = min(pcp->count, batch);
1888 free_pcppages_bulk(zone, to_drain, pcp);
1889 pcp->count -= to_drain;
1891 local_irq_restore(flags);
1896 * Drain pcplists of the indicated processor and zone.
1898 * The processor must either be the current processor and the
1899 * thread pinned to the current processor or a processor that
1902 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1904 unsigned long flags;
1905 struct per_cpu_pageset *pset;
1906 struct per_cpu_pages *pcp;
1908 local_irq_save(flags);
1909 pset = per_cpu_ptr(zone->pageset, cpu);
1913 free_pcppages_bulk(zone, pcp->count, pcp);
1916 local_irq_restore(flags);
1920 * Drain pcplists of all zones on the indicated processor.
1922 * The processor must either be the current processor and the
1923 * thread pinned to the current processor or a processor that
1926 static void drain_pages(unsigned int cpu)
1930 for_each_populated_zone(zone) {
1931 drain_pages_zone(cpu, zone);
1936 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1938 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1939 * the single zone's pages.
1941 void drain_local_pages(struct zone *zone)
1943 int cpu = smp_processor_id();
1946 drain_pages_zone(cpu, zone);
1952 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1954 * When zone parameter is non-NULL, spill just the single zone's pages.
1956 * Note that this code is protected against sending an IPI to an offline
1957 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1958 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1959 * nothing keeps CPUs from showing up after we populated the cpumask and
1960 * before the call to on_each_cpu_mask().
1962 void drain_all_pages(struct zone *zone)
1967 * Allocate in the BSS so we wont require allocation in
1968 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1970 static cpumask_t cpus_with_pcps;
1973 * We don't care about racing with CPU hotplug event
1974 * as offline notification will cause the notified
1975 * cpu to drain that CPU pcps and on_each_cpu_mask
1976 * disables preemption as part of its processing
1978 for_each_online_cpu(cpu) {
1979 struct per_cpu_pageset *pcp;
1981 bool has_pcps = false;
1984 pcp = per_cpu_ptr(zone->pageset, cpu);
1988 for_each_populated_zone(z) {
1989 pcp = per_cpu_ptr(z->pageset, cpu);
1990 if (pcp->pcp.count) {
1998 cpumask_set_cpu(cpu, &cpus_with_pcps);
2000 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2002 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2006 #ifdef CONFIG_HIBERNATION
2008 void mark_free_pages(struct zone *zone)
2010 unsigned long pfn, max_zone_pfn;
2011 unsigned long flags;
2012 unsigned int order, t;
2013 struct list_head *curr;
2015 if (zone_is_empty(zone))
2018 spin_lock_irqsave(&zone->lock, flags);
2020 max_zone_pfn = zone_end_pfn(zone);
2021 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2022 if (pfn_valid(pfn)) {
2023 struct page *page = pfn_to_page(pfn);
2025 if (!swsusp_page_is_forbidden(page))
2026 swsusp_unset_page_free(page);
2029 for_each_migratetype_order(order, t) {
2030 list_for_each(curr, &zone->free_area[order].free_list[t]) {
2033 pfn = page_to_pfn(list_entry(curr, struct page, lru));
2034 for (i = 0; i < (1UL << order); i++)
2035 swsusp_set_page_free(pfn_to_page(pfn + i));
2038 spin_unlock_irqrestore(&zone->lock, flags);
2040 #endif /* CONFIG_PM */
2043 * Free a 0-order page
2044 * cold == true ? free a cold page : free a hot page
2046 void free_hot_cold_page(struct page *page, bool cold)
2048 struct zone *zone = page_zone(page);
2049 struct per_cpu_pages *pcp;
2050 unsigned long flags;
2051 unsigned long pfn = page_to_pfn(page);
2054 if (!free_pages_prepare(page, 0))
2057 migratetype = get_pfnblock_migratetype(page, pfn);
2058 set_pcppage_migratetype(page, migratetype);
2059 local_irq_save(flags);
2060 __count_vm_event(PGFREE);
2063 * We only track unmovable, reclaimable and movable on pcp lists.
2064 * Free ISOLATE pages back to the allocator because they are being
2065 * offlined but treat RESERVE as movable pages so we can get those
2066 * areas back if necessary. Otherwise, we may have to free
2067 * excessively into the page allocator
2069 if (migratetype >= MIGRATE_PCPTYPES) {
2070 if (unlikely(is_migrate_isolate(migratetype))) {
2071 free_one_page(zone, page, pfn, 0, migratetype);
2074 migratetype = MIGRATE_MOVABLE;
2077 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2079 list_add(&page->lru, &pcp->lists[migratetype]);
2081 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2083 if (pcp->count >= pcp->high) {
2084 unsigned long batch = READ_ONCE(pcp->batch);
2085 free_pcppages_bulk(zone, batch, pcp);
2086 pcp->count -= batch;
2090 local_irq_restore(flags);
2094 * Free a list of 0-order pages
2096 void free_hot_cold_page_list(struct list_head *list, bool cold)
2098 struct page *page, *next;
2100 list_for_each_entry_safe(page, next, list, lru) {
2101 trace_mm_page_free_batched(page, cold);
2102 free_hot_cold_page(page, cold);
2107 * split_page takes a non-compound higher-order page, and splits it into
2108 * n (1<<order) sub-pages: page[0..n]
2109 * Each sub-page must be freed individually.
2111 * Note: this is probably too low level an operation for use in drivers.
2112 * Please consult with lkml before using this in your driver.
2114 void split_page(struct page *page, unsigned int order)
2119 VM_BUG_ON_PAGE(PageCompound(page), page);
2120 VM_BUG_ON_PAGE(!page_count(page), page);
2122 #ifdef CONFIG_KMEMCHECK
2124 * Split shadow pages too, because free(page[0]) would
2125 * otherwise free the whole shadow.
2127 if (kmemcheck_page_is_tracked(page))
2128 split_page(virt_to_page(page[0].shadow), order);
2131 gfp_mask = get_page_owner_gfp(page);
2132 set_page_owner(page, 0, gfp_mask);
2133 for (i = 1; i < (1 << order); i++) {
2134 set_page_refcounted(page + i);
2135 set_page_owner(page + i, 0, gfp_mask);
2138 EXPORT_SYMBOL_GPL(split_page);
2140 int __isolate_free_page(struct page *page, unsigned int order)
2142 unsigned long watermark;
2146 BUG_ON(!PageBuddy(page));
2148 zone = page_zone(page);
2149 mt = get_pageblock_migratetype(page);
2151 if (!is_migrate_isolate(mt)) {
2152 /* Obey watermarks as if the page was being allocated */
2153 watermark = low_wmark_pages(zone) + (1 << order);
2154 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2157 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2160 /* Remove page from free list */
2161 list_del(&page->lru);
2162 zone->free_area[order].nr_free--;
2163 rmv_page_order(page);
2165 set_page_owner(page, order, __GFP_MOVABLE);
2167 /* Set the pageblock if the isolated page is at least a pageblock */
2168 if (order >= pageblock_order - 1) {
2169 struct page *endpage = page + (1 << order) - 1;
2170 for (; page < endpage; page += pageblock_nr_pages) {
2171 int mt = get_pageblock_migratetype(page);
2172 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2173 set_pageblock_migratetype(page,
2179 return 1UL << order;
2183 * Similar to split_page except the page is already free. As this is only
2184 * being used for migration, the migratetype of the block also changes.
2185 * As this is called with interrupts disabled, the caller is responsible
2186 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2189 * Note: this is probably too low level an operation for use in drivers.
2190 * Please consult with lkml before using this in your driver.
2192 int split_free_page(struct page *page)
2197 order = page_order(page);
2199 nr_pages = __isolate_free_page(page, order);
2203 /* Split into individual pages */
2204 set_page_refcounted(page);
2205 split_page(page, order);
2210 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2213 struct page *buffered_rmqueue(struct zone *preferred_zone,
2214 struct zone *zone, unsigned int order,
2215 gfp_t gfp_flags, int alloc_flags, int migratetype)
2217 unsigned long flags;
2219 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2221 if (likely(order == 0)) {
2222 struct per_cpu_pages *pcp;
2223 struct list_head *list;
2225 local_irq_save(flags);
2226 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2227 list = &pcp->lists[migratetype];
2228 if (list_empty(list)) {
2229 pcp->count += rmqueue_bulk(zone, 0,
2232 if (unlikely(list_empty(list)))
2237 page = list_entry(list->prev, struct page, lru);
2239 page = list_entry(list->next, struct page, lru);
2241 list_del(&page->lru);
2244 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2246 * __GFP_NOFAIL is not to be used in new code.
2248 * All __GFP_NOFAIL callers should be fixed so that they
2249 * properly detect and handle allocation failures.
2251 * We most definitely don't want callers attempting to
2252 * allocate greater than order-1 page units with
2255 WARN_ON_ONCE(order > 1);
2257 spin_lock_irqsave(&zone->lock, flags);
2260 if (alloc_flags & ALLOC_HARDER) {
2261 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2263 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2266 page = __rmqueue(zone, order, migratetype, gfp_flags);
2267 spin_unlock(&zone->lock);
2270 __mod_zone_freepage_state(zone, -(1 << order),
2271 get_pcppage_migratetype(page));
2274 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2275 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2276 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2277 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2279 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2280 zone_statistics(preferred_zone, zone, gfp_flags);
2281 local_irq_restore(flags);
2283 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2287 local_irq_restore(flags);
2291 #ifdef CONFIG_FAIL_PAGE_ALLOC
2294 struct fault_attr attr;
2296 bool ignore_gfp_highmem;
2297 bool ignore_gfp_reclaim;
2299 } fail_page_alloc = {
2300 .attr = FAULT_ATTR_INITIALIZER,
2301 .ignore_gfp_reclaim = true,
2302 .ignore_gfp_highmem = true,
2306 static int __init setup_fail_page_alloc(char *str)
2308 return setup_fault_attr(&fail_page_alloc.attr, str);
2310 __setup("fail_page_alloc=", setup_fail_page_alloc);
2312 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2314 if (order < fail_page_alloc.min_order)
2316 if (gfp_mask & __GFP_NOFAIL)
2318 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2320 if (fail_page_alloc.ignore_gfp_reclaim &&
2321 (gfp_mask & __GFP_DIRECT_RECLAIM))
2324 return should_fail(&fail_page_alloc.attr, 1 << order);
2327 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2329 static int __init fail_page_alloc_debugfs(void)
2331 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2334 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2335 &fail_page_alloc.attr);
2337 return PTR_ERR(dir);
2339 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2340 &fail_page_alloc.ignore_gfp_reclaim))
2342 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2343 &fail_page_alloc.ignore_gfp_highmem))
2345 if (!debugfs_create_u32("min-order", mode, dir,
2346 &fail_page_alloc.min_order))
2351 debugfs_remove_recursive(dir);
2356 late_initcall(fail_page_alloc_debugfs);
2358 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2360 #else /* CONFIG_FAIL_PAGE_ALLOC */
2362 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2367 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2370 * Return true if free base pages are above 'mark'. For high-order checks it
2371 * will return true of the order-0 watermark is reached and there is at least
2372 * one free page of a suitable size. Checking now avoids taking the zone lock
2373 * to check in the allocation paths if no pages are free.
2375 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2376 unsigned long mark, int classzone_idx, int alloc_flags,
2381 const int alloc_harder = (alloc_flags & ALLOC_HARDER);
2383 /* free_pages may go negative - that's OK */
2384 free_pages -= (1 << order) - 1;
2386 if (alloc_flags & ALLOC_HIGH)
2390 * If the caller does not have rights to ALLOC_HARDER then subtract
2391 * the high-atomic reserves. This will over-estimate the size of the
2392 * atomic reserve but it avoids a search.
2394 if (likely(!alloc_harder))
2395 free_pages -= z->nr_reserved_highatomic;
2400 /* If allocation can't use CMA areas don't use free CMA pages */
2401 if (!(alloc_flags & ALLOC_CMA))
2402 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2406 * Check watermarks for an order-0 allocation request. If these
2407 * are not met, then a high-order request also cannot go ahead
2408 * even if a suitable page happened to be free.
2410 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2413 /* If this is an order-0 request then the watermark is fine */
2417 /* For a high-order request, check at least one suitable page is free */
2418 for (o = order; o < MAX_ORDER; o++) {
2419 struct free_area *area = &z->free_area[o];
2428 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2429 if (!list_empty(&area->free_list[mt]))
2434 if ((alloc_flags & ALLOC_CMA) &&
2435 !list_empty(&area->free_list[MIGRATE_CMA])) {
2443 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2444 int classzone_idx, int alloc_flags)
2446 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2447 zone_page_state(z, NR_FREE_PAGES));
2450 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2451 unsigned long mark, int classzone_idx)
2453 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2455 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2456 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2458 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2463 static bool zone_local(struct zone *local_zone, struct zone *zone)
2465 return local_zone->node == zone->node;
2468 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2470 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
2473 #else /* CONFIG_NUMA */
2474 static bool zone_local(struct zone *local_zone, struct zone *zone)
2479 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2483 #endif /* CONFIG_NUMA */
2485 static void reset_alloc_batches(struct zone *preferred_zone)
2487 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2490 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2491 high_wmark_pages(zone) - low_wmark_pages(zone) -
2492 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2493 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2494 } while (zone++ != preferred_zone);
2498 * get_page_from_freelist goes through the zonelist trying to allocate
2501 static struct page *
2502 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2503 const struct alloc_context *ac)
2505 struct zonelist *zonelist = ac->zonelist;
2507 struct page *page = NULL;
2509 int nr_fair_skipped = 0;
2510 bool zonelist_rescan;
2513 zonelist_rescan = false;
2516 * Scan zonelist, looking for a zone with enough free.
2517 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2519 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2523 if (cpusets_enabled() &&
2524 (alloc_flags & ALLOC_CPUSET) &&
2525 !cpuset_zone_allowed(zone, gfp_mask))
2528 * Distribute pages in proportion to the individual
2529 * zone size to ensure fair page aging. The zone a
2530 * page was allocated in should have no effect on the
2531 * time the page has in memory before being reclaimed.
2533 if (alloc_flags & ALLOC_FAIR) {
2534 if (!zone_local(ac->preferred_zone, zone))
2536 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2542 * When allocating a page cache page for writing, we
2543 * want to get it from a zone that is within its dirty
2544 * limit, such that no single zone holds more than its
2545 * proportional share of globally allowed dirty pages.
2546 * The dirty limits take into account the zone's
2547 * lowmem reserves and high watermark so that kswapd
2548 * should be able to balance it without having to
2549 * write pages from its LRU list.
2551 * This may look like it could increase pressure on
2552 * lower zones by failing allocations in higher zones
2553 * before they are full. But the pages that do spill
2554 * over are limited as the lower zones are protected
2555 * by this very same mechanism. It should not become
2556 * a practical burden to them.
2558 * XXX: For now, allow allocations to potentially
2559 * exceed the per-zone dirty limit in the slowpath
2560 * (spread_dirty_pages unset) before going into reclaim,
2561 * which is important when on a NUMA setup the allowed
2562 * zones are together not big enough to reach the
2563 * global limit. The proper fix for these situations
2564 * will require awareness of zones in the
2565 * dirty-throttling and the flusher threads.
2567 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2570 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2571 if (!zone_watermark_ok(zone, order, mark,
2572 ac->classzone_idx, alloc_flags)) {
2575 /* Checked here to keep the fast path fast */
2576 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2577 if (alloc_flags & ALLOC_NO_WATERMARKS)
2580 if (zone_reclaim_mode == 0 ||
2581 !zone_allows_reclaim(ac->preferred_zone, zone))
2584 ret = zone_reclaim(zone, gfp_mask, order);
2586 case ZONE_RECLAIM_NOSCAN:
2589 case ZONE_RECLAIM_FULL:
2590 /* scanned but unreclaimable */
2593 /* did we reclaim enough */
2594 if (zone_watermark_ok(zone, order, mark,
2595 ac->classzone_idx, alloc_flags))
2603 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2604 gfp_mask, alloc_flags, ac->migratetype);
2606 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2610 * If this is a high-order atomic allocation then check
2611 * if the pageblock should be reserved for the future
2613 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2614 reserve_highatomic_pageblock(page, zone, order);
2621 * The first pass makes sure allocations are spread fairly within the
2622 * local node. However, the local node might have free pages left
2623 * after the fairness batches are exhausted, and remote zones haven't
2624 * even been considered yet. Try once more without fairness, and
2625 * include remote zones now, before entering the slowpath and waking
2626 * kswapd: prefer spilling to a remote zone over swapping locally.
2628 if (alloc_flags & ALLOC_FAIR) {
2629 alloc_flags &= ~ALLOC_FAIR;
2630 if (nr_fair_skipped) {
2631 zonelist_rescan = true;
2632 reset_alloc_batches(ac->preferred_zone);
2634 if (nr_online_nodes > 1)
2635 zonelist_rescan = true;
2638 if (zonelist_rescan)
2645 * Large machines with many possible nodes should not always dump per-node
2646 * meminfo in irq context.
2648 static inline bool should_suppress_show_mem(void)
2653 ret = in_interrupt();
2658 static DEFINE_RATELIMIT_STATE(nopage_rs,
2659 DEFAULT_RATELIMIT_INTERVAL,
2660 DEFAULT_RATELIMIT_BURST);
2662 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
2664 unsigned int filter = SHOW_MEM_FILTER_NODES;
2666 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2667 debug_guardpage_minorder() > 0)
2671 * This documents exceptions given to allocations in certain
2672 * contexts that are allowed to allocate outside current's set
2675 if (!(gfp_mask & __GFP_NOMEMALLOC))
2676 if (test_thread_flag(TIF_MEMDIE) ||
2677 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2678 filter &= ~SHOW_MEM_FILTER_NODES;
2679 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2680 filter &= ~SHOW_MEM_FILTER_NODES;
2683 struct va_format vaf;
2686 va_start(args, fmt);
2691 pr_warn("%pV", &vaf);
2696 pr_warn("%s: page allocation failure: order:%u, mode:0x%x\n",
2697 current->comm, order, gfp_mask);
2700 if (!should_suppress_show_mem())
2704 static inline struct page *
2705 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2706 const struct alloc_context *ac, unsigned long *did_some_progress)
2708 struct oom_control oc = {
2709 .zonelist = ac->zonelist,
2710 .nodemask = ac->nodemask,
2711 .gfp_mask = gfp_mask,
2716 *did_some_progress = 0;
2719 * Acquire the oom lock. If that fails, somebody else is
2720 * making progress for us.
2722 if (!mutex_trylock(&oom_lock)) {
2723 *did_some_progress = 1;
2724 schedule_timeout_uninterruptible(1);
2729 * Go through the zonelist yet one more time, keep very high watermark
2730 * here, this is only to catch a parallel oom killing, we must fail if
2731 * we're still under heavy pressure.
2733 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2734 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2738 if (!(gfp_mask & __GFP_NOFAIL)) {
2739 /* Coredumps can quickly deplete all memory reserves */
2740 if (current->flags & PF_DUMPCORE)
2742 /* The OOM killer will not help higher order allocs */
2743 if (order > PAGE_ALLOC_COSTLY_ORDER)
2745 /* The OOM killer does not needlessly kill tasks for lowmem */
2746 if (ac->high_zoneidx < ZONE_NORMAL)
2748 /* The OOM killer does not compensate for IO-less reclaim */
2749 if (!(gfp_mask & __GFP_FS)) {
2751 * XXX: Page reclaim didn't yield anything,
2752 * and the OOM killer can't be invoked, but
2753 * keep looping as per tradition.
2755 *did_some_progress = 1;
2758 if (pm_suspended_storage())
2760 /* The OOM killer may not free memory on a specific node */
2761 if (gfp_mask & __GFP_THISNODE)
2764 /* Exhausted what can be done so it's blamo time */
2765 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL))
2766 *did_some_progress = 1;
2768 mutex_unlock(&oom_lock);
2772 #ifdef CONFIG_COMPACTION
2773 /* Try memory compaction for high-order allocations before reclaim */
2774 static struct page *
2775 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2776 int alloc_flags, const struct alloc_context *ac,
2777 enum migrate_mode mode, int *contended_compaction,
2778 bool *deferred_compaction)
2780 unsigned long compact_result;
2786 current->flags |= PF_MEMALLOC;
2787 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2788 mode, contended_compaction);
2789 current->flags &= ~PF_MEMALLOC;
2791 switch (compact_result) {
2792 case COMPACT_DEFERRED:
2793 *deferred_compaction = true;
2795 case COMPACT_SKIPPED:
2802 * At least in one zone compaction wasn't deferred or skipped, so let's
2803 * count a compaction stall
2805 count_vm_event(COMPACTSTALL);
2807 page = get_page_from_freelist(gfp_mask, order,
2808 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2811 struct zone *zone = page_zone(page);
2813 zone->compact_blockskip_flush = false;
2814 compaction_defer_reset(zone, order, true);
2815 count_vm_event(COMPACTSUCCESS);
2820 * It's bad if compaction run occurs and fails. The most likely reason
2821 * is that pages exist, but not enough to satisfy watermarks.
2823 count_vm_event(COMPACTFAIL);
2830 static inline struct page *
2831 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2832 int alloc_flags, const struct alloc_context *ac,
2833 enum migrate_mode mode, int *contended_compaction,
2834 bool *deferred_compaction)
2838 #endif /* CONFIG_COMPACTION */
2840 /* Perform direct synchronous page reclaim */
2842 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2843 const struct alloc_context *ac)
2845 struct reclaim_state reclaim_state;
2850 /* We now go into synchronous reclaim */
2851 cpuset_memory_pressure_bump();
2852 current->flags |= PF_MEMALLOC;
2853 lockdep_set_current_reclaim_state(gfp_mask);
2854 reclaim_state.reclaimed_slab = 0;
2855 current->reclaim_state = &reclaim_state;
2857 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2860 current->reclaim_state = NULL;
2861 lockdep_clear_current_reclaim_state();
2862 current->flags &= ~PF_MEMALLOC;
2869 /* The really slow allocator path where we enter direct reclaim */
2870 static inline struct page *
2871 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2872 int alloc_flags, const struct alloc_context *ac,
2873 unsigned long *did_some_progress)
2875 struct page *page = NULL;
2876 bool drained = false;
2878 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2879 if (unlikely(!(*did_some_progress)))
2883 page = get_page_from_freelist(gfp_mask, order,
2884 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2887 * If an allocation failed after direct reclaim, it could be because
2888 * pages are pinned on the per-cpu lists or in high alloc reserves.
2889 * Shrink them them and try again
2891 if (!page && !drained) {
2892 unreserve_highatomic_pageblock(ac);
2893 drain_all_pages(NULL);
2902 * This is called in the allocator slow-path if the allocation request is of
2903 * sufficient urgency to ignore watermarks and take other desperate measures
2905 static inline struct page *
2906 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2907 const struct alloc_context *ac)
2912 page = get_page_from_freelist(gfp_mask, order,
2913 ALLOC_NO_WATERMARKS, ac);
2915 if (!page && gfp_mask & __GFP_NOFAIL)
2916 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC,
2918 } while (!page && (gfp_mask & __GFP_NOFAIL));
2923 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2928 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2929 ac->high_zoneidx, ac->nodemask)
2930 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2934 gfp_to_alloc_flags(gfp_t gfp_mask)
2936 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2938 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2939 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2942 * The caller may dip into page reserves a bit more if the caller
2943 * cannot run direct reclaim, or if the caller has realtime scheduling
2944 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2945 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
2947 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2949 if (gfp_mask & __GFP_ATOMIC) {
2951 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2952 * if it can't schedule.
2954 if (!(gfp_mask & __GFP_NOMEMALLOC))
2955 alloc_flags |= ALLOC_HARDER;
2957 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2958 * comment for __cpuset_node_allowed().
2960 alloc_flags &= ~ALLOC_CPUSET;
2961 } else if (unlikely(rt_task(current)) && !in_interrupt())
2962 alloc_flags |= ALLOC_HARDER;
2964 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2965 if (gfp_mask & __GFP_MEMALLOC)
2966 alloc_flags |= ALLOC_NO_WATERMARKS;
2967 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2968 alloc_flags |= ALLOC_NO_WATERMARKS;
2969 else if (!in_interrupt() &&
2970 ((current->flags & PF_MEMALLOC) ||
2971 unlikely(test_thread_flag(TIF_MEMDIE))))
2972 alloc_flags |= ALLOC_NO_WATERMARKS;
2975 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2976 alloc_flags |= ALLOC_CMA;
2981 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2983 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2986 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
2988 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
2991 static inline struct page *
2992 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2993 struct alloc_context *ac)
2995 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
2996 struct page *page = NULL;
2998 unsigned long pages_reclaimed = 0;
2999 unsigned long did_some_progress;
3000 enum migrate_mode migration_mode = MIGRATE_ASYNC;
3001 bool deferred_compaction = false;
3002 int contended_compaction = COMPACT_CONTENDED_NONE;
3005 * In the slowpath, we sanity check order to avoid ever trying to
3006 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3007 * be using allocators in order of preference for an area that is
3010 if (order >= MAX_ORDER) {
3011 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3016 * We also sanity check to catch abuse of atomic reserves being used by
3017 * callers that are not in atomic context.
3019 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3020 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3021 gfp_mask &= ~__GFP_ATOMIC;
3024 * If this allocation cannot block and it is for a specific node, then
3025 * fail early. There's no need to wakeup kswapd or retry for a
3026 * speculative node-specific allocation.
3028 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !can_direct_reclaim)
3032 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3033 wake_all_kswapds(order, ac);
3036 * OK, we're below the kswapd watermark and have kicked background
3037 * reclaim. Now things get more complex, so set up alloc_flags according
3038 * to how we want to proceed.
3040 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3043 * Find the true preferred zone if the allocation is unconstrained by
3046 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3047 struct zoneref *preferred_zoneref;
3048 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3049 ac->high_zoneidx, NULL, &ac->preferred_zone);
3050 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3053 /* This is the last chance, in general, before the goto nopage. */
3054 page = get_page_from_freelist(gfp_mask, order,
3055 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3059 /* Allocate without watermarks if the context allows */
3060 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3062 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3063 * the allocation is high priority and these type of
3064 * allocations are system rather than user orientated
3066 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3068 page = __alloc_pages_high_priority(gfp_mask, order, ac);
3075 /* Caller is not willing to reclaim, we can't balance anything */
3076 if (!can_direct_reclaim) {
3078 * All existing users of the deprecated __GFP_NOFAIL are
3079 * blockable, so warn of any new users that actually allow this
3080 * type of allocation to fail.
3082 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3086 /* Avoid recursion of direct reclaim */
3087 if (current->flags & PF_MEMALLOC)
3090 /* Avoid allocations with no watermarks from looping endlessly */
3091 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3095 * Try direct compaction. The first pass is asynchronous. Subsequent
3096 * attempts after direct reclaim are synchronous
3098 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3100 &contended_compaction,
3101 &deferred_compaction);
3105 /* Checks for THP-specific high-order allocations */
3106 if (is_thp_gfp_mask(gfp_mask)) {
3108 * If compaction is deferred for high-order allocations, it is
3109 * because sync compaction recently failed. If this is the case
3110 * and the caller requested a THP allocation, we do not want
3111 * to heavily disrupt the system, so we fail the allocation
3112 * instead of entering direct reclaim.
3114 if (deferred_compaction)
3118 * In all zones where compaction was attempted (and not
3119 * deferred or skipped), lock contention has been detected.
3120 * For THP allocation we do not want to disrupt the others
3121 * so we fallback to base pages instead.
3123 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3127 * If compaction was aborted due to need_resched(), we do not
3128 * want to further increase allocation latency, unless it is
3129 * khugepaged trying to collapse.
3131 if (contended_compaction == COMPACT_CONTENDED_SCHED
3132 && !(current->flags & PF_KTHREAD))
3137 * It can become very expensive to allocate transparent hugepages at
3138 * fault, so use asynchronous memory compaction for THP unless it is
3139 * khugepaged trying to collapse.
3141 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3142 migration_mode = MIGRATE_SYNC_LIGHT;
3144 /* Try direct reclaim and then allocating */
3145 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3146 &did_some_progress);
3150 /* Do not loop if specifically requested */
3151 if (gfp_mask & __GFP_NORETRY)
3154 /* Keep reclaiming pages as long as there is reasonable progress */
3155 pages_reclaimed += did_some_progress;
3156 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3157 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3158 /* Wait for some write requests to complete then retry */
3159 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3163 /* Reclaim has failed us, start killing things */
3164 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3168 /* Retry as long as the OOM killer is making progress */
3169 if (did_some_progress)
3174 * High-order allocations do not necessarily loop after
3175 * direct reclaim and reclaim/compaction depends on compaction
3176 * being called after reclaim so call directly if necessary
3178 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3180 &contended_compaction,
3181 &deferred_compaction);
3185 warn_alloc_failed(gfp_mask, order, NULL);
3191 * This is the 'heart' of the zoned buddy allocator.
3194 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3195 struct zonelist *zonelist, nodemask_t *nodemask)
3197 struct zoneref *preferred_zoneref;
3198 struct page *page = NULL;
3199 unsigned int cpuset_mems_cookie;
3200 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3201 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3202 struct alloc_context ac = {
3203 .high_zoneidx = gfp_zone(gfp_mask),
3204 .nodemask = nodemask,
3205 .migratetype = gfpflags_to_migratetype(gfp_mask),
3208 gfp_mask &= gfp_allowed_mask;
3210 lockdep_trace_alloc(gfp_mask);
3212 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3214 if (should_fail_alloc_page(gfp_mask, order))
3218 * Check the zones suitable for the gfp_mask contain at least one
3219 * valid zone. It's possible to have an empty zonelist as a result
3220 * of __GFP_THISNODE and a memoryless node
3222 if (unlikely(!zonelist->_zonerefs->zone))
3225 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3226 alloc_flags |= ALLOC_CMA;
3229 cpuset_mems_cookie = read_mems_allowed_begin();
3231 /* We set it here, as __alloc_pages_slowpath might have changed it */
3232 ac.zonelist = zonelist;
3234 /* Dirty zone balancing only done in the fast path */
3235 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3237 /* The preferred zone is used for statistics later */
3238 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3239 ac.nodemask ? : &cpuset_current_mems_allowed,
3240 &ac.preferred_zone);
3241 if (!ac.preferred_zone)
3243 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3245 /* First allocation attempt */
3246 alloc_mask = gfp_mask|__GFP_HARDWALL;
3247 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3248 if (unlikely(!page)) {
3250 * Runtime PM, block IO and its error handling path
3251 * can deadlock because I/O on the device might not
3254 alloc_mask = memalloc_noio_flags(gfp_mask);
3255 ac.spread_dirty_pages = false;
3257 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3260 if (kmemcheck_enabled && page)
3261 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3263 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3267 * When updating a task's mems_allowed, it is possible to race with
3268 * parallel threads in such a way that an allocation can fail while
3269 * the mask is being updated. If a page allocation is about to fail,
3270 * check if the cpuset changed during allocation and if so, retry.
3272 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3277 EXPORT_SYMBOL(__alloc_pages_nodemask);
3280 * Common helper functions.
3282 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3287 * __get_free_pages() returns a 32-bit address, which cannot represent
3290 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3292 page = alloc_pages(gfp_mask, order);
3295 return (unsigned long) page_address(page);
3297 EXPORT_SYMBOL(__get_free_pages);
3299 unsigned long get_zeroed_page(gfp_t gfp_mask)
3301 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3303 EXPORT_SYMBOL(get_zeroed_page);
3305 void __free_pages(struct page *page, unsigned int order)
3307 if (put_page_testzero(page)) {
3309 free_hot_cold_page(page, false);
3311 __free_pages_ok(page, order);
3315 EXPORT_SYMBOL(__free_pages);
3317 void free_pages(unsigned long addr, unsigned int order)
3320 VM_BUG_ON(!virt_addr_valid((void *)addr));
3321 __free_pages(virt_to_page((void *)addr), order);
3325 EXPORT_SYMBOL(free_pages);
3329 * An arbitrary-length arbitrary-offset area of memory which resides
3330 * within a 0 or higher order page. Multiple fragments within that page
3331 * are individually refcounted, in the page's reference counter.
3333 * The page_frag functions below provide a simple allocation framework for
3334 * page fragments. This is used by the network stack and network device
3335 * drivers to provide a backing region of memory for use as either an
3336 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3338 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3341 struct page *page = NULL;
3342 gfp_t gfp = gfp_mask;
3344 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3345 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3347 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3348 PAGE_FRAG_CACHE_MAX_ORDER);
3349 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3351 if (unlikely(!page))
3352 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3354 nc->va = page ? page_address(page) : NULL;
3359 void *__alloc_page_frag(struct page_frag_cache *nc,
3360 unsigned int fragsz, gfp_t gfp_mask)
3362 unsigned int size = PAGE_SIZE;
3366 if (unlikely(!nc->va)) {
3368 page = __page_frag_refill(nc, gfp_mask);
3372 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3373 /* if size can vary use size else just use PAGE_SIZE */
3376 /* Even if we own the page, we do not use atomic_set().
3377 * This would break get_page_unless_zero() users.
3379 atomic_add(size - 1, &page->_count);
3381 /* reset page count bias and offset to start of new frag */
3382 nc->pfmemalloc = page_is_pfmemalloc(page);
3383 nc->pagecnt_bias = size;
3387 offset = nc->offset - fragsz;
3388 if (unlikely(offset < 0)) {
3389 page = virt_to_page(nc->va);
3391 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3394 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3395 /* if size can vary use size else just use PAGE_SIZE */
3398 /* OK, page count is 0, we can safely set it */
3399 atomic_set(&page->_count, size);
3401 /* reset page count bias and offset to start of new frag */
3402 nc->pagecnt_bias = size;
3403 offset = size - fragsz;
3407 nc->offset = offset;
3409 return nc->va + offset;
3411 EXPORT_SYMBOL(__alloc_page_frag);
3414 * Frees a page fragment allocated out of either a compound or order 0 page.
3416 void __free_page_frag(void *addr)
3418 struct page *page = virt_to_head_page(addr);
3420 if (unlikely(put_page_testzero(page)))
3421 __free_pages_ok(page, compound_order(page));
3423 EXPORT_SYMBOL(__free_page_frag);
3426 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3427 * of the current memory cgroup.
3429 * It should be used when the caller would like to use kmalloc, but since the
3430 * allocation is large, it has to fall back to the page allocator.
3432 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3436 page = alloc_pages(gfp_mask, order);
3437 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3438 __free_pages(page, order);
3444 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3448 page = alloc_pages_node(nid, gfp_mask, order);
3449 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3450 __free_pages(page, order);
3457 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3460 void __free_kmem_pages(struct page *page, unsigned int order)
3462 memcg_kmem_uncharge(page, order);
3463 __free_pages(page, order);
3466 void free_kmem_pages(unsigned long addr, unsigned int order)
3469 VM_BUG_ON(!virt_addr_valid((void *)addr));
3470 __free_kmem_pages(virt_to_page((void *)addr), order);
3474 static void *make_alloc_exact(unsigned long addr, unsigned int order,
3478 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3479 unsigned long used = addr + PAGE_ALIGN(size);
3481 split_page(virt_to_page((void *)addr), order);
3482 while (used < alloc_end) {
3487 return (void *)addr;
3491 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3492 * @size: the number of bytes to allocate
3493 * @gfp_mask: GFP flags for the allocation
3495 * This function is similar to alloc_pages(), except that it allocates the
3496 * minimum number of pages to satisfy the request. alloc_pages() can only
3497 * allocate memory in power-of-two pages.
3499 * This function is also limited by MAX_ORDER.
3501 * Memory allocated by this function must be released by free_pages_exact().
3503 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3505 unsigned int order = get_order(size);
3508 addr = __get_free_pages(gfp_mask, order);
3509 return make_alloc_exact(addr, order, size);
3511 EXPORT_SYMBOL(alloc_pages_exact);
3514 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3516 * @nid: the preferred node ID where memory should be allocated
3517 * @size: the number of bytes to allocate
3518 * @gfp_mask: GFP flags for the allocation
3520 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3523 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3525 unsigned int order = get_order(size);
3526 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3529 return make_alloc_exact((unsigned long)page_address(p), order, size);
3533 * free_pages_exact - release memory allocated via alloc_pages_exact()
3534 * @virt: the value returned by alloc_pages_exact.
3535 * @size: size of allocation, same value as passed to alloc_pages_exact().
3537 * Release the memory allocated by a previous call to alloc_pages_exact.
3539 void free_pages_exact(void *virt, size_t size)
3541 unsigned long addr = (unsigned long)virt;
3542 unsigned long end = addr + PAGE_ALIGN(size);
3544 while (addr < end) {
3549 EXPORT_SYMBOL(free_pages_exact);
3552 * nr_free_zone_pages - count number of pages beyond high watermark
3553 * @offset: The zone index of the highest zone
3555 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3556 * high watermark within all zones at or below a given zone index. For each
3557 * zone, the number of pages is calculated as:
3558 * managed_pages - high_pages
3560 static unsigned long nr_free_zone_pages(int offset)
3565 /* Just pick one node, since fallback list is circular */
3566 unsigned long sum = 0;
3568 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3570 for_each_zone_zonelist(zone, z, zonelist, offset) {
3571 unsigned long size = zone->managed_pages;
3572 unsigned long high = high_wmark_pages(zone);
3581 * nr_free_buffer_pages - count number of pages beyond high watermark
3583 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3584 * watermark within ZONE_DMA and ZONE_NORMAL.
3586 unsigned long nr_free_buffer_pages(void)
3588 return nr_free_zone_pages(gfp_zone(GFP_USER));
3590 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3593 * nr_free_pagecache_pages - count number of pages beyond high watermark
3595 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3596 * high watermark within all zones.
3598 unsigned long nr_free_pagecache_pages(void)
3600 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3603 static inline void show_node(struct zone *zone)
3605 if (IS_ENABLED(CONFIG_NUMA))
3606 printk("Node %d ", zone_to_nid(zone));
3609 void si_meminfo(struct sysinfo *val)
3611 val->totalram = totalram_pages;
3612 val->sharedram = global_page_state(NR_SHMEM);
3613 val->freeram = global_page_state(NR_FREE_PAGES);
3614 val->bufferram = nr_blockdev_pages();
3615 val->totalhigh = totalhigh_pages;
3616 val->freehigh = nr_free_highpages();
3617 val->mem_unit = PAGE_SIZE;
3620 EXPORT_SYMBOL(si_meminfo);
3623 void si_meminfo_node(struct sysinfo *val, int nid)
3625 int zone_type; /* needs to be signed */
3626 unsigned long managed_pages = 0;
3627 pg_data_t *pgdat = NODE_DATA(nid);
3629 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3630 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3631 val->totalram = managed_pages;
3632 val->sharedram = node_page_state(nid, NR_SHMEM);
3633 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3634 #ifdef CONFIG_HIGHMEM
3635 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3636 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3642 val->mem_unit = PAGE_SIZE;
3647 * Determine whether the node should be displayed or not, depending on whether
3648 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3650 bool skip_free_areas_node(unsigned int flags, int nid)
3653 unsigned int cpuset_mems_cookie;
3655 if (!(flags & SHOW_MEM_FILTER_NODES))
3659 cpuset_mems_cookie = read_mems_allowed_begin();
3660 ret = !node_isset(nid, cpuset_current_mems_allowed);
3661 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3666 #define K(x) ((x) << (PAGE_SHIFT-10))
3668 static void show_migration_types(unsigned char type)
3670 static const char types[MIGRATE_TYPES] = {
3671 [MIGRATE_UNMOVABLE] = 'U',
3672 [MIGRATE_MOVABLE] = 'M',
3673 [MIGRATE_RECLAIMABLE] = 'E',
3674 [MIGRATE_HIGHATOMIC] = 'H',
3676 [MIGRATE_CMA] = 'C',
3678 #ifdef CONFIG_MEMORY_ISOLATION
3679 [MIGRATE_ISOLATE] = 'I',
3682 char tmp[MIGRATE_TYPES + 1];
3686 for (i = 0; i < MIGRATE_TYPES; i++) {
3687 if (type & (1 << i))
3692 printk("(%s) ", tmp);
3696 * Show free area list (used inside shift_scroll-lock stuff)
3697 * We also calculate the percentage fragmentation. We do this by counting the
3698 * memory on each free list with the exception of the first item on the list.
3701 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3704 void show_free_areas(unsigned int filter)
3706 unsigned long free_pcp = 0;
3710 for_each_populated_zone(zone) {
3711 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3714 for_each_online_cpu(cpu)
3715 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3718 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3719 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3720 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3721 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3722 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3723 " free:%lu free_pcp:%lu free_cma:%lu\n",
3724 global_page_state(NR_ACTIVE_ANON),
3725 global_page_state(NR_INACTIVE_ANON),
3726 global_page_state(NR_ISOLATED_ANON),
3727 global_page_state(NR_ACTIVE_FILE),
3728 global_page_state(NR_INACTIVE_FILE),
3729 global_page_state(NR_ISOLATED_FILE),
3730 global_page_state(NR_UNEVICTABLE),
3731 global_page_state(NR_FILE_DIRTY),
3732 global_page_state(NR_WRITEBACK),
3733 global_page_state(NR_UNSTABLE_NFS),
3734 global_page_state(NR_SLAB_RECLAIMABLE),
3735 global_page_state(NR_SLAB_UNRECLAIMABLE),
3736 global_page_state(NR_FILE_MAPPED),
3737 global_page_state(NR_SHMEM),
3738 global_page_state(NR_PAGETABLE),
3739 global_page_state(NR_BOUNCE),
3740 global_page_state(NR_FREE_PAGES),
3742 global_page_state(NR_FREE_CMA_PAGES));
3744 for_each_populated_zone(zone) {
3747 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3751 for_each_online_cpu(cpu)
3752 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3760 " active_anon:%lukB"
3761 " inactive_anon:%lukB"
3762 " active_file:%lukB"
3763 " inactive_file:%lukB"
3764 " unevictable:%lukB"
3765 " isolated(anon):%lukB"
3766 " isolated(file):%lukB"
3774 " slab_reclaimable:%lukB"
3775 " slab_unreclaimable:%lukB"
3776 " kernel_stack:%lukB"
3783 " writeback_tmp:%lukB"
3784 " pages_scanned:%lu"
3785 " all_unreclaimable? %s"
3788 K(zone_page_state(zone, NR_FREE_PAGES)),
3789 K(min_wmark_pages(zone)),
3790 K(low_wmark_pages(zone)),
3791 K(high_wmark_pages(zone)),
3792 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3793 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3794 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3795 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3796 K(zone_page_state(zone, NR_UNEVICTABLE)),
3797 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3798 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3799 K(zone->present_pages),
3800 K(zone->managed_pages),
3801 K(zone_page_state(zone, NR_MLOCK)),
3802 K(zone_page_state(zone, NR_FILE_DIRTY)),
3803 K(zone_page_state(zone, NR_WRITEBACK)),
3804 K(zone_page_state(zone, NR_FILE_MAPPED)),
3805 K(zone_page_state(zone, NR_SHMEM)),
3806 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3807 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3808 zone_page_state(zone, NR_KERNEL_STACK) *
3810 K(zone_page_state(zone, NR_PAGETABLE)),
3811 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3812 K(zone_page_state(zone, NR_BOUNCE)),
3814 K(this_cpu_read(zone->pageset->pcp.count)),
3815 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3816 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3817 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3818 (!zone_reclaimable(zone) ? "yes" : "no")
3820 printk("lowmem_reserve[]:");
3821 for (i = 0; i < MAX_NR_ZONES; i++)
3822 printk(" %ld", zone->lowmem_reserve[i]);
3826 for_each_populated_zone(zone) {
3828 unsigned long nr[MAX_ORDER], flags, total = 0;
3829 unsigned char types[MAX_ORDER];
3831 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3834 printk("%s: ", zone->name);
3836 spin_lock_irqsave(&zone->lock, flags);
3837 for (order = 0; order < MAX_ORDER; order++) {
3838 struct free_area *area = &zone->free_area[order];
3841 nr[order] = area->nr_free;
3842 total += nr[order] << order;
3845 for (type = 0; type < MIGRATE_TYPES; type++) {
3846 if (!list_empty(&area->free_list[type]))
3847 types[order] |= 1 << type;
3850 spin_unlock_irqrestore(&zone->lock, flags);
3851 for (order = 0; order < MAX_ORDER; order++) {
3852 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3854 show_migration_types(types[order]);
3856 printk("= %lukB\n", K(total));
3859 hugetlb_show_meminfo();
3861 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3863 show_swap_cache_info();
3866 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3868 zoneref->zone = zone;
3869 zoneref->zone_idx = zone_idx(zone);
3873 * Builds allocation fallback zone lists.
3875 * Add all populated zones of a node to the zonelist.
3877 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3881 enum zone_type zone_type = MAX_NR_ZONES;
3885 zone = pgdat->node_zones + zone_type;
3886 if (populated_zone(zone)) {
3887 zoneref_set_zone(zone,
3888 &zonelist->_zonerefs[nr_zones++]);
3889 check_highest_zone(zone_type);
3891 } while (zone_type);
3899 * 0 = automatic detection of better ordering.
3900 * 1 = order by ([node] distance, -zonetype)
3901 * 2 = order by (-zonetype, [node] distance)
3903 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3904 * the same zonelist. So only NUMA can configure this param.
3906 #define ZONELIST_ORDER_DEFAULT 0
3907 #define ZONELIST_ORDER_NODE 1
3908 #define ZONELIST_ORDER_ZONE 2
3910 /* zonelist order in the kernel.
3911 * set_zonelist_order() will set this to NODE or ZONE.
3913 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3914 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3918 /* The value user specified ....changed by config */
3919 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3920 /* string for sysctl */
3921 #define NUMA_ZONELIST_ORDER_LEN 16
3922 char numa_zonelist_order[16] = "default";
3925 * interface for configure zonelist ordering.
3926 * command line option "numa_zonelist_order"
3927 * = "[dD]efault - default, automatic configuration.
3928 * = "[nN]ode - order by node locality, then by zone within node
3929 * = "[zZ]one - order by zone, then by locality within zone
3932 static int __parse_numa_zonelist_order(char *s)
3934 if (*s == 'd' || *s == 'D') {
3935 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3936 } else if (*s == 'n' || *s == 'N') {
3937 user_zonelist_order = ZONELIST_ORDER_NODE;
3938 } else if (*s == 'z' || *s == 'Z') {
3939 user_zonelist_order = ZONELIST_ORDER_ZONE;
3942 "Ignoring invalid numa_zonelist_order value: "
3949 static __init int setup_numa_zonelist_order(char *s)
3956 ret = __parse_numa_zonelist_order(s);
3958 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3962 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3965 * sysctl handler for numa_zonelist_order
3967 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3968 void __user *buffer, size_t *length,
3971 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3973 static DEFINE_MUTEX(zl_order_mutex);
3975 mutex_lock(&zl_order_mutex);
3977 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3981 strcpy(saved_string, (char *)table->data);
3983 ret = proc_dostring(table, write, buffer, length, ppos);
3987 int oldval = user_zonelist_order;
3989 ret = __parse_numa_zonelist_order((char *)table->data);
3992 * bogus value. restore saved string
3994 strncpy((char *)table->data, saved_string,
3995 NUMA_ZONELIST_ORDER_LEN);
3996 user_zonelist_order = oldval;
3997 } else if (oldval != user_zonelist_order) {
3998 mutex_lock(&zonelists_mutex);
3999 build_all_zonelists(NULL, NULL);
4000 mutex_unlock(&zonelists_mutex);
4004 mutex_unlock(&zl_order_mutex);
4009 #define MAX_NODE_LOAD (nr_online_nodes)
4010 static int node_load[MAX_NUMNODES];
4013 * find_next_best_node - find the next node that should appear in a given node's fallback list
4014 * @node: node whose fallback list we're appending
4015 * @used_node_mask: nodemask_t of already used nodes
4017 * We use a number of factors to determine which is the next node that should
4018 * appear on a given node's fallback list. The node should not have appeared
4019 * already in @node's fallback list, and it should be the next closest node
4020 * according to the distance array (which contains arbitrary distance values
4021 * from each node to each node in the system), and should also prefer nodes
4022 * with no CPUs, since presumably they'll have very little allocation pressure
4023 * on them otherwise.
4024 * It returns -1 if no node is found.
4026 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4029 int min_val = INT_MAX;
4030 int best_node = NUMA_NO_NODE;
4031 const struct cpumask *tmp = cpumask_of_node(0);
4033 /* Use the local node if we haven't already */
4034 if (!node_isset(node, *used_node_mask)) {
4035 node_set(node, *used_node_mask);
4039 for_each_node_state(n, N_MEMORY) {
4041 /* Don't want a node to appear more than once */
4042 if (node_isset(n, *used_node_mask))
4045 /* Use the distance array to find the distance */
4046 val = node_distance(node, n);
4048 /* Penalize nodes under us ("prefer the next node") */
4051 /* Give preference to headless and unused nodes */
4052 tmp = cpumask_of_node(n);
4053 if (!cpumask_empty(tmp))
4054 val += PENALTY_FOR_NODE_WITH_CPUS;
4056 /* Slight preference for less loaded node */
4057 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4058 val += node_load[n];
4060 if (val < min_val) {
4067 node_set(best_node, *used_node_mask);
4074 * Build zonelists ordered by node and zones within node.
4075 * This results in maximum locality--normal zone overflows into local
4076 * DMA zone, if any--but risks exhausting DMA zone.
4078 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4081 struct zonelist *zonelist;
4083 zonelist = &pgdat->node_zonelists[0];
4084 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4086 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4087 zonelist->_zonerefs[j].zone = NULL;
4088 zonelist->_zonerefs[j].zone_idx = 0;
4092 * Build gfp_thisnode zonelists
4094 static void build_thisnode_zonelists(pg_data_t *pgdat)
4097 struct zonelist *zonelist;
4099 zonelist = &pgdat->node_zonelists[1];
4100 j = build_zonelists_node(pgdat, zonelist, 0);
4101 zonelist->_zonerefs[j].zone = NULL;
4102 zonelist->_zonerefs[j].zone_idx = 0;
4106 * Build zonelists ordered by zone and nodes within zones.
4107 * This results in conserving DMA zone[s] until all Normal memory is
4108 * exhausted, but results in overflowing to remote node while memory
4109 * may still exist in local DMA zone.
4111 static int node_order[MAX_NUMNODES];
4113 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4116 int zone_type; /* needs to be signed */
4118 struct zonelist *zonelist;
4120 zonelist = &pgdat->node_zonelists[0];
4122 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4123 for (j = 0; j < nr_nodes; j++) {
4124 node = node_order[j];
4125 z = &NODE_DATA(node)->node_zones[zone_type];
4126 if (populated_zone(z)) {
4128 &zonelist->_zonerefs[pos++]);
4129 check_highest_zone(zone_type);
4133 zonelist->_zonerefs[pos].zone = NULL;
4134 zonelist->_zonerefs[pos].zone_idx = 0;
4137 #if defined(CONFIG_64BIT)
4139 * Devices that require DMA32/DMA are relatively rare and do not justify a
4140 * penalty to every machine in case the specialised case applies. Default
4141 * to Node-ordering on 64-bit NUMA machines
4143 static int default_zonelist_order(void)
4145 return ZONELIST_ORDER_NODE;
4149 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4150 * by the kernel. If processes running on node 0 deplete the low memory zone
4151 * then reclaim will occur more frequency increasing stalls and potentially
4152 * be easier to OOM if a large percentage of the zone is under writeback or
4153 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4154 * Hence, default to zone ordering on 32-bit.
4156 static int default_zonelist_order(void)
4158 return ZONELIST_ORDER_ZONE;
4160 #endif /* CONFIG_64BIT */
4162 static void set_zonelist_order(void)
4164 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4165 current_zonelist_order = default_zonelist_order();
4167 current_zonelist_order = user_zonelist_order;
4170 static void build_zonelists(pg_data_t *pgdat)
4174 nodemask_t used_mask;
4175 int local_node, prev_node;
4176 struct zonelist *zonelist;
4177 unsigned int order = current_zonelist_order;
4179 /* initialize zonelists */
4180 for (i = 0; i < MAX_ZONELISTS; i++) {
4181 zonelist = pgdat->node_zonelists + i;
4182 zonelist->_zonerefs[0].zone = NULL;
4183 zonelist->_zonerefs[0].zone_idx = 0;
4186 /* NUMA-aware ordering of nodes */
4187 local_node = pgdat->node_id;
4188 load = nr_online_nodes;
4189 prev_node = local_node;
4190 nodes_clear(used_mask);
4192 memset(node_order, 0, sizeof(node_order));
4195 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4197 * We don't want to pressure a particular node.
4198 * So adding penalty to the first node in same
4199 * distance group to make it round-robin.
4201 if (node_distance(local_node, node) !=
4202 node_distance(local_node, prev_node))
4203 node_load[node] = load;
4207 if (order == ZONELIST_ORDER_NODE)
4208 build_zonelists_in_node_order(pgdat, node);
4210 node_order[j++] = node; /* remember order */
4213 if (order == ZONELIST_ORDER_ZONE) {
4214 /* calculate node order -- i.e., DMA last! */
4215 build_zonelists_in_zone_order(pgdat, j);
4218 build_thisnode_zonelists(pgdat);
4221 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4223 * Return node id of node used for "local" allocations.
4224 * I.e., first node id of first zone in arg node's generic zonelist.
4225 * Used for initializing percpu 'numa_mem', which is used primarily
4226 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4228 int local_memory_node(int node)
4232 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4233 gfp_zone(GFP_KERNEL),
4240 #else /* CONFIG_NUMA */
4242 static void set_zonelist_order(void)
4244 current_zonelist_order = ZONELIST_ORDER_ZONE;
4247 static void build_zonelists(pg_data_t *pgdat)
4249 int node, local_node;
4251 struct zonelist *zonelist;
4253 local_node = pgdat->node_id;
4255 zonelist = &pgdat->node_zonelists[0];
4256 j = build_zonelists_node(pgdat, zonelist, 0);
4259 * Now we build the zonelist so that it contains the zones
4260 * of all the other nodes.
4261 * We don't want to pressure a particular node, so when
4262 * building the zones for node N, we make sure that the
4263 * zones coming right after the local ones are those from
4264 * node N+1 (modulo N)
4266 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4267 if (!node_online(node))
4269 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4271 for (node = 0; node < local_node; node++) {
4272 if (!node_online(node))
4274 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4277 zonelist->_zonerefs[j].zone = NULL;
4278 zonelist->_zonerefs[j].zone_idx = 0;
4281 #endif /* CONFIG_NUMA */
4284 * Boot pageset table. One per cpu which is going to be used for all
4285 * zones and all nodes. The parameters will be set in such a way
4286 * that an item put on a list will immediately be handed over to
4287 * the buddy list. This is safe since pageset manipulation is done
4288 * with interrupts disabled.
4290 * The boot_pagesets must be kept even after bootup is complete for
4291 * unused processors and/or zones. They do play a role for bootstrapping
4292 * hotplugged processors.
4294 * zoneinfo_show() and maybe other functions do
4295 * not check if the processor is online before following the pageset pointer.
4296 * Other parts of the kernel may not check if the zone is available.
4298 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4299 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4300 static void setup_zone_pageset(struct zone *zone);
4303 * Global mutex to protect against size modification of zonelists
4304 * as well as to serialize pageset setup for the new populated zone.
4306 DEFINE_MUTEX(zonelists_mutex);
4308 /* return values int ....just for stop_machine() */
4309 static int __build_all_zonelists(void *data)
4313 pg_data_t *self = data;
4316 memset(node_load, 0, sizeof(node_load));
4319 if (self && !node_online(self->node_id)) {
4320 build_zonelists(self);
4323 for_each_online_node(nid) {
4324 pg_data_t *pgdat = NODE_DATA(nid);
4326 build_zonelists(pgdat);
4330 * Initialize the boot_pagesets that are going to be used
4331 * for bootstrapping processors. The real pagesets for
4332 * each zone will be allocated later when the per cpu
4333 * allocator is available.
4335 * boot_pagesets are used also for bootstrapping offline
4336 * cpus if the system is already booted because the pagesets
4337 * are needed to initialize allocators on a specific cpu too.
4338 * F.e. the percpu allocator needs the page allocator which
4339 * needs the percpu allocator in order to allocate its pagesets
4340 * (a chicken-egg dilemma).
4342 for_each_possible_cpu(cpu) {
4343 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4345 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4347 * We now know the "local memory node" for each node--
4348 * i.e., the node of the first zone in the generic zonelist.
4349 * Set up numa_mem percpu variable for on-line cpus. During
4350 * boot, only the boot cpu should be on-line; we'll init the
4351 * secondary cpus' numa_mem as they come on-line. During
4352 * node/memory hotplug, we'll fixup all on-line cpus.
4354 if (cpu_online(cpu))
4355 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4362 static noinline void __init
4363 build_all_zonelists_init(void)
4365 __build_all_zonelists(NULL);
4366 mminit_verify_zonelist();
4367 cpuset_init_current_mems_allowed();
4371 * Called with zonelists_mutex held always
4372 * unless system_state == SYSTEM_BOOTING.
4374 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4375 * [we're only called with non-NULL zone through __meminit paths] and
4376 * (2) call of __init annotated helper build_all_zonelists_init
4377 * [protected by SYSTEM_BOOTING].
4379 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4381 set_zonelist_order();
4383 if (system_state == SYSTEM_BOOTING) {
4384 build_all_zonelists_init();
4386 #ifdef CONFIG_MEMORY_HOTPLUG
4388 setup_zone_pageset(zone);
4390 /* we have to stop all cpus to guarantee there is no user
4392 stop_machine(__build_all_zonelists, pgdat, NULL);
4393 /* cpuset refresh routine should be here */
4395 vm_total_pages = nr_free_pagecache_pages();
4397 * Disable grouping by mobility if the number of pages in the
4398 * system is too low to allow the mechanism to work. It would be
4399 * more accurate, but expensive to check per-zone. This check is
4400 * made on memory-hotadd so a system can start with mobility
4401 * disabled and enable it later
4403 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4404 page_group_by_mobility_disabled = 1;
4406 page_group_by_mobility_disabled = 0;
4408 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4409 "Total pages: %ld\n",
4411 zonelist_order_name[current_zonelist_order],
4412 page_group_by_mobility_disabled ? "off" : "on",
4415 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4420 * Helper functions to size the waitqueue hash table.
4421 * Essentially these want to choose hash table sizes sufficiently
4422 * large so that collisions trying to wait on pages are rare.
4423 * But in fact, the number of active page waitqueues on typical
4424 * systems is ridiculously low, less than 200. So this is even
4425 * conservative, even though it seems large.
4427 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4428 * waitqueues, i.e. the size of the waitq table given the number of pages.
4430 #define PAGES_PER_WAITQUEUE 256
4432 #ifndef CONFIG_MEMORY_HOTPLUG
4433 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4435 unsigned long size = 1;
4437 pages /= PAGES_PER_WAITQUEUE;
4439 while (size < pages)
4443 * Once we have dozens or even hundreds of threads sleeping
4444 * on IO we've got bigger problems than wait queue collision.
4445 * Limit the size of the wait table to a reasonable size.
4447 size = min(size, 4096UL);
4449 return max(size, 4UL);
4453 * A zone's size might be changed by hot-add, so it is not possible to determine
4454 * a suitable size for its wait_table. So we use the maximum size now.
4456 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4458 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4459 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4460 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4462 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4463 * or more by the traditional way. (See above). It equals:
4465 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4466 * ia64(16K page size) : = ( 8G + 4M)byte.
4467 * powerpc (64K page size) : = (32G +16M)byte.
4469 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4476 * This is an integer logarithm so that shifts can be used later
4477 * to extract the more random high bits from the multiplicative
4478 * hash function before the remainder is taken.
4480 static inline unsigned long wait_table_bits(unsigned long size)
4486 * Initially all pages are reserved - free ones are freed
4487 * up by free_all_bootmem() once the early boot process is
4488 * done. Non-atomic initialization, single-pass.
4490 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4491 unsigned long start_pfn, enum memmap_context context)
4493 pg_data_t *pgdat = NODE_DATA(nid);
4494 unsigned long end_pfn = start_pfn + size;
4497 unsigned long nr_initialised = 0;
4499 if (highest_memmap_pfn < end_pfn - 1)
4500 highest_memmap_pfn = end_pfn - 1;
4502 z = &pgdat->node_zones[zone];
4503 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4505 * There can be holes in boot-time mem_map[]s
4506 * handed to this function. They do not
4507 * exist on hotplugged memory.
4509 if (context == MEMMAP_EARLY) {
4510 if (!early_pfn_valid(pfn))
4512 if (!early_pfn_in_nid(pfn, nid))
4514 if (!update_defer_init(pgdat, pfn, end_pfn,
4520 * Mark the block movable so that blocks are reserved for
4521 * movable at startup. This will force kernel allocations
4522 * to reserve their blocks rather than leaking throughout
4523 * the address space during boot when many long-lived
4524 * kernel allocations are made.
4526 * bitmap is created for zone's valid pfn range. but memmap
4527 * can be created for invalid pages (for alignment)
4528 * check here not to call set_pageblock_migratetype() against
4531 if (!(pfn & (pageblock_nr_pages - 1))) {
4532 struct page *page = pfn_to_page(pfn);
4534 __init_single_page(page, pfn, zone, nid);
4535 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4537 __init_single_pfn(pfn, zone, nid);
4542 static void __meminit zone_init_free_lists(struct zone *zone)
4544 unsigned int order, t;
4545 for_each_migratetype_order(order, t) {
4546 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4547 zone->free_area[order].nr_free = 0;
4551 #ifndef __HAVE_ARCH_MEMMAP_INIT
4552 #define memmap_init(size, nid, zone, start_pfn) \
4553 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4556 static int zone_batchsize(struct zone *zone)
4562 * The per-cpu-pages pools are set to around 1000th of the
4563 * size of the zone. But no more than 1/2 of a meg.
4565 * OK, so we don't know how big the cache is. So guess.
4567 batch = zone->managed_pages / 1024;
4568 if (batch * PAGE_SIZE > 512 * 1024)
4569 batch = (512 * 1024) / PAGE_SIZE;
4570 batch /= 4; /* We effectively *= 4 below */
4575 * Clamp the batch to a 2^n - 1 value. Having a power
4576 * of 2 value was found to be more likely to have
4577 * suboptimal cache aliasing properties in some cases.
4579 * For example if 2 tasks are alternately allocating
4580 * batches of pages, one task can end up with a lot
4581 * of pages of one half of the possible page colors
4582 * and the other with pages of the other colors.
4584 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4589 /* The deferral and batching of frees should be suppressed under NOMMU
4592 * The problem is that NOMMU needs to be able to allocate large chunks
4593 * of contiguous memory as there's no hardware page translation to
4594 * assemble apparent contiguous memory from discontiguous pages.
4596 * Queueing large contiguous runs of pages for batching, however,
4597 * causes the pages to actually be freed in smaller chunks. As there
4598 * can be a significant delay between the individual batches being
4599 * recycled, this leads to the once large chunks of space being
4600 * fragmented and becoming unavailable for high-order allocations.
4607 * pcp->high and pcp->batch values are related and dependent on one another:
4608 * ->batch must never be higher then ->high.
4609 * The following function updates them in a safe manner without read side
4612 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4613 * those fields changing asynchronously (acording the the above rule).
4615 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4616 * outside of boot time (or some other assurance that no concurrent updaters
4619 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4620 unsigned long batch)
4622 /* start with a fail safe value for batch */
4626 /* Update high, then batch, in order */
4633 /* a companion to pageset_set_high() */
4634 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4636 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4639 static void pageset_init(struct per_cpu_pageset *p)
4641 struct per_cpu_pages *pcp;
4644 memset(p, 0, sizeof(*p));
4648 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4649 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4652 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4655 pageset_set_batch(p, batch);
4659 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4660 * to the value high for the pageset p.
4662 static void pageset_set_high(struct per_cpu_pageset *p,
4665 unsigned long batch = max(1UL, high / 4);
4666 if ((high / 4) > (PAGE_SHIFT * 8))
4667 batch = PAGE_SHIFT * 8;
4669 pageset_update(&p->pcp, high, batch);
4672 static void pageset_set_high_and_batch(struct zone *zone,
4673 struct per_cpu_pageset *pcp)
4675 if (percpu_pagelist_fraction)
4676 pageset_set_high(pcp,
4677 (zone->managed_pages /
4678 percpu_pagelist_fraction));
4680 pageset_set_batch(pcp, zone_batchsize(zone));
4683 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4685 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4688 pageset_set_high_and_batch(zone, pcp);
4691 static void __meminit setup_zone_pageset(struct zone *zone)
4694 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4695 for_each_possible_cpu(cpu)
4696 zone_pageset_init(zone, cpu);
4700 * Allocate per cpu pagesets and initialize them.
4701 * Before this call only boot pagesets were available.
4703 void __init setup_per_cpu_pageset(void)
4707 for_each_populated_zone(zone)
4708 setup_zone_pageset(zone);
4711 static noinline __init_refok
4712 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4718 * The per-page waitqueue mechanism uses hashed waitqueues
4721 zone->wait_table_hash_nr_entries =
4722 wait_table_hash_nr_entries(zone_size_pages);
4723 zone->wait_table_bits =
4724 wait_table_bits(zone->wait_table_hash_nr_entries);
4725 alloc_size = zone->wait_table_hash_nr_entries
4726 * sizeof(wait_queue_head_t);
4728 if (!slab_is_available()) {
4729 zone->wait_table = (wait_queue_head_t *)
4730 memblock_virt_alloc_node_nopanic(
4731 alloc_size, zone->zone_pgdat->node_id);
4734 * This case means that a zone whose size was 0 gets new memory
4735 * via memory hot-add.
4736 * But it may be the case that a new node was hot-added. In
4737 * this case vmalloc() will not be able to use this new node's
4738 * memory - this wait_table must be initialized to use this new
4739 * node itself as well.
4740 * To use this new node's memory, further consideration will be
4743 zone->wait_table = vmalloc(alloc_size);
4745 if (!zone->wait_table)
4748 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4749 init_waitqueue_head(zone->wait_table + i);
4754 static __meminit void zone_pcp_init(struct zone *zone)
4757 * per cpu subsystem is not up at this point. The following code
4758 * relies on the ability of the linker to provide the
4759 * offset of a (static) per cpu variable into the per cpu area.
4761 zone->pageset = &boot_pageset;
4763 if (populated_zone(zone))
4764 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4765 zone->name, zone->present_pages,
4766 zone_batchsize(zone));
4769 int __meminit init_currently_empty_zone(struct zone *zone,
4770 unsigned long zone_start_pfn,
4773 struct pglist_data *pgdat = zone->zone_pgdat;
4775 ret = zone_wait_table_init(zone, size);
4778 pgdat->nr_zones = zone_idx(zone) + 1;
4780 zone->zone_start_pfn = zone_start_pfn;
4782 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4783 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4785 (unsigned long)zone_idx(zone),
4786 zone_start_pfn, (zone_start_pfn + size));
4788 zone_init_free_lists(zone);
4793 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4794 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4797 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4799 int __meminit __early_pfn_to_nid(unsigned long pfn,
4800 struct mminit_pfnnid_cache *state)
4802 unsigned long start_pfn, end_pfn;
4805 if (state->last_start <= pfn && pfn < state->last_end)
4806 return state->last_nid;
4808 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4810 state->last_start = start_pfn;
4811 state->last_end = end_pfn;
4812 state->last_nid = nid;
4817 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4820 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4821 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4822 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4824 * If an architecture guarantees that all ranges registered contain no holes
4825 * and may be freed, this this function may be used instead of calling
4826 * memblock_free_early_nid() manually.
4828 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4830 unsigned long start_pfn, end_pfn;
4833 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4834 start_pfn = min(start_pfn, max_low_pfn);
4835 end_pfn = min(end_pfn, max_low_pfn);
4837 if (start_pfn < end_pfn)
4838 memblock_free_early_nid(PFN_PHYS(start_pfn),
4839 (end_pfn - start_pfn) << PAGE_SHIFT,
4845 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4846 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4848 * If an architecture guarantees that all ranges registered contain no holes and may
4849 * be freed, this function may be used instead of calling memory_present() manually.
4851 void __init sparse_memory_present_with_active_regions(int nid)
4853 unsigned long start_pfn, end_pfn;
4856 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4857 memory_present(this_nid, start_pfn, end_pfn);
4861 * get_pfn_range_for_nid - Return the start and end page frames for a node
4862 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4863 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4864 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4866 * It returns the start and end page frame of a node based on information
4867 * provided by memblock_set_node(). If called for a node
4868 * with no available memory, a warning is printed and the start and end
4871 void __meminit get_pfn_range_for_nid(unsigned int nid,
4872 unsigned long *start_pfn, unsigned long *end_pfn)
4874 unsigned long this_start_pfn, this_end_pfn;
4880 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4881 *start_pfn = min(*start_pfn, this_start_pfn);
4882 *end_pfn = max(*end_pfn, this_end_pfn);
4885 if (*start_pfn == -1UL)
4890 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4891 * assumption is made that zones within a node are ordered in monotonic
4892 * increasing memory addresses so that the "highest" populated zone is used
4894 static void __init find_usable_zone_for_movable(void)
4897 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4898 if (zone_index == ZONE_MOVABLE)
4901 if (arch_zone_highest_possible_pfn[zone_index] >
4902 arch_zone_lowest_possible_pfn[zone_index])
4906 VM_BUG_ON(zone_index == -1);
4907 movable_zone = zone_index;
4911 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4912 * because it is sized independent of architecture. Unlike the other zones,
4913 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4914 * in each node depending on the size of each node and how evenly kernelcore
4915 * is distributed. This helper function adjusts the zone ranges
4916 * provided by the architecture for a given node by using the end of the
4917 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4918 * zones within a node are in order of monotonic increases memory addresses
4920 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4921 unsigned long zone_type,
4922 unsigned long node_start_pfn,
4923 unsigned long node_end_pfn,
4924 unsigned long *zone_start_pfn,
4925 unsigned long *zone_end_pfn)
4927 /* Only adjust if ZONE_MOVABLE is on this node */
4928 if (zone_movable_pfn[nid]) {
4929 /* Size ZONE_MOVABLE */
4930 if (zone_type == ZONE_MOVABLE) {
4931 *zone_start_pfn = zone_movable_pfn[nid];
4932 *zone_end_pfn = min(node_end_pfn,
4933 arch_zone_highest_possible_pfn[movable_zone]);
4935 /* Adjust for ZONE_MOVABLE starting within this range */
4936 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4937 *zone_end_pfn > zone_movable_pfn[nid]) {
4938 *zone_end_pfn = zone_movable_pfn[nid];
4940 /* Check if this whole range is within ZONE_MOVABLE */
4941 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4942 *zone_start_pfn = *zone_end_pfn;
4947 * Return the number of pages a zone spans in a node, including holes
4948 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4950 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4951 unsigned long zone_type,
4952 unsigned long node_start_pfn,
4953 unsigned long node_end_pfn,
4954 unsigned long *ignored)
4956 unsigned long zone_start_pfn, zone_end_pfn;
4958 /* When hotadd a new node from cpu_up(), the node should be empty */
4959 if (!node_start_pfn && !node_end_pfn)
4962 /* Get the start and end of the zone */
4963 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4964 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4965 adjust_zone_range_for_zone_movable(nid, zone_type,
4966 node_start_pfn, node_end_pfn,
4967 &zone_start_pfn, &zone_end_pfn);
4969 /* Check that this node has pages within the zone's required range */
4970 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4973 /* Move the zone boundaries inside the node if necessary */
4974 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4975 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4977 /* Return the spanned pages */
4978 return zone_end_pfn - zone_start_pfn;
4982 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4983 * then all holes in the requested range will be accounted for.
4985 unsigned long __meminit __absent_pages_in_range(int nid,
4986 unsigned long range_start_pfn,
4987 unsigned long range_end_pfn)
4989 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4990 unsigned long start_pfn, end_pfn;
4993 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4994 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4995 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4996 nr_absent -= end_pfn - start_pfn;
5002 * absent_pages_in_range - Return number of page frames in holes within a range
5003 * @start_pfn: The start PFN to start searching for holes
5004 * @end_pfn: The end PFN to stop searching for holes
5006 * It returns the number of pages frames in memory holes within a range.
5008 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5009 unsigned long end_pfn)
5011 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5014 /* Return the number of page frames in holes in a zone on a node */
5015 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5016 unsigned long zone_type,
5017 unsigned long node_start_pfn,
5018 unsigned long node_end_pfn,
5019 unsigned long *ignored)
5021 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5022 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5023 unsigned long zone_start_pfn, zone_end_pfn;
5025 /* When hotadd a new node from cpu_up(), the node should be empty */
5026 if (!node_start_pfn && !node_end_pfn)
5029 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5030 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5032 adjust_zone_range_for_zone_movable(nid, zone_type,
5033 node_start_pfn, node_end_pfn,
5034 &zone_start_pfn, &zone_end_pfn);
5035 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5038 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5039 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5040 unsigned long zone_type,
5041 unsigned long node_start_pfn,
5042 unsigned long node_end_pfn,
5043 unsigned long *zones_size)
5045 return zones_size[zone_type];
5048 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5049 unsigned long zone_type,
5050 unsigned long node_start_pfn,
5051 unsigned long node_end_pfn,
5052 unsigned long *zholes_size)
5057 return zholes_size[zone_type];
5060 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5062 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5063 unsigned long node_start_pfn,
5064 unsigned long node_end_pfn,
5065 unsigned long *zones_size,
5066 unsigned long *zholes_size)
5068 unsigned long realtotalpages = 0, totalpages = 0;
5071 for (i = 0; i < MAX_NR_ZONES; i++) {
5072 struct zone *zone = pgdat->node_zones + i;
5073 unsigned long size, real_size;
5075 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5079 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5080 node_start_pfn, node_end_pfn,
5082 zone->spanned_pages = size;
5083 zone->present_pages = real_size;
5086 realtotalpages += real_size;
5089 pgdat->node_spanned_pages = totalpages;
5090 pgdat->node_present_pages = realtotalpages;
5091 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5095 #ifndef CONFIG_SPARSEMEM
5097 * Calculate the size of the zone->blockflags rounded to an unsigned long
5098 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5099 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5100 * round what is now in bits to nearest long in bits, then return it in
5103 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5105 unsigned long usemapsize;
5107 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5108 usemapsize = roundup(zonesize, pageblock_nr_pages);
5109 usemapsize = usemapsize >> pageblock_order;
5110 usemapsize *= NR_PAGEBLOCK_BITS;
5111 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5113 return usemapsize / 8;
5116 static void __init setup_usemap(struct pglist_data *pgdat,
5118 unsigned long zone_start_pfn,
5119 unsigned long zonesize)
5121 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5122 zone->pageblock_flags = NULL;
5124 zone->pageblock_flags =
5125 memblock_virt_alloc_node_nopanic(usemapsize,
5129 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5130 unsigned long zone_start_pfn, unsigned long zonesize) {}
5131 #endif /* CONFIG_SPARSEMEM */
5133 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5135 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5136 void __paginginit set_pageblock_order(void)
5140 /* Check that pageblock_nr_pages has not already been setup */
5141 if (pageblock_order)
5144 if (HPAGE_SHIFT > PAGE_SHIFT)
5145 order = HUGETLB_PAGE_ORDER;
5147 order = MAX_ORDER - 1;
5150 * Assume the largest contiguous order of interest is a huge page.
5151 * This value may be variable depending on boot parameters on IA64 and
5154 pageblock_order = order;
5156 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5159 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5160 * is unused as pageblock_order is set at compile-time. See
5161 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5164 void __paginginit set_pageblock_order(void)
5168 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5170 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5171 unsigned long present_pages)
5173 unsigned long pages = spanned_pages;
5176 * Provide a more accurate estimation if there are holes within
5177 * the zone and SPARSEMEM is in use. If there are holes within the
5178 * zone, each populated memory region may cost us one or two extra
5179 * memmap pages due to alignment because memmap pages for each
5180 * populated regions may not naturally algined on page boundary.
5181 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5183 if (spanned_pages > present_pages + (present_pages >> 4) &&
5184 IS_ENABLED(CONFIG_SPARSEMEM))
5185 pages = present_pages;
5187 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5191 * Set up the zone data structures:
5192 * - mark all pages reserved
5193 * - mark all memory queues empty
5194 * - clear the memory bitmaps
5196 * NOTE: pgdat should get zeroed by caller.
5198 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5201 int nid = pgdat->node_id;
5202 unsigned long zone_start_pfn = pgdat->node_start_pfn;
5205 pgdat_resize_init(pgdat);
5206 #ifdef CONFIG_NUMA_BALANCING
5207 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5208 pgdat->numabalancing_migrate_nr_pages = 0;
5209 pgdat->numabalancing_migrate_next_window = jiffies;
5211 init_waitqueue_head(&pgdat->kswapd_wait);
5212 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5213 pgdat_page_ext_init(pgdat);
5215 for (j = 0; j < MAX_NR_ZONES; j++) {
5216 struct zone *zone = pgdat->node_zones + j;
5217 unsigned long size, realsize, freesize, memmap_pages;
5219 size = zone->spanned_pages;
5220 realsize = freesize = zone->present_pages;
5223 * Adjust freesize so that it accounts for how much memory
5224 * is used by this zone for memmap. This affects the watermark
5225 * and per-cpu initialisations
5227 memmap_pages = calc_memmap_size(size, realsize);
5228 if (!is_highmem_idx(j)) {
5229 if (freesize >= memmap_pages) {
5230 freesize -= memmap_pages;
5233 " %s zone: %lu pages used for memmap\n",
5234 zone_names[j], memmap_pages);
5237 " %s zone: %lu pages exceeds freesize %lu\n",
5238 zone_names[j], memmap_pages, freesize);
5241 /* Account for reserved pages */
5242 if (j == 0 && freesize > dma_reserve) {
5243 freesize -= dma_reserve;
5244 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5245 zone_names[0], dma_reserve);
5248 if (!is_highmem_idx(j))
5249 nr_kernel_pages += freesize;
5250 /* Charge for highmem memmap if there are enough kernel pages */
5251 else if (nr_kernel_pages > memmap_pages * 2)
5252 nr_kernel_pages -= memmap_pages;
5253 nr_all_pages += freesize;
5256 * Set an approximate value for lowmem here, it will be adjusted
5257 * when the bootmem allocator frees pages into the buddy system.
5258 * And all highmem pages will be managed by the buddy system.
5260 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5263 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5265 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5267 zone->name = zone_names[j];
5268 spin_lock_init(&zone->lock);
5269 spin_lock_init(&zone->lru_lock);
5270 zone_seqlock_init(zone);
5271 zone->zone_pgdat = pgdat;
5272 zone_pcp_init(zone);
5274 /* For bootup, initialized properly in watermark setup */
5275 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5277 lruvec_init(&zone->lruvec);
5281 set_pageblock_order();
5282 setup_usemap(pgdat, zone, zone_start_pfn, size);
5283 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5285 memmap_init(size, nid, j, zone_start_pfn);
5286 zone_start_pfn += size;
5290 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5292 unsigned long __maybe_unused start = 0;
5293 unsigned long __maybe_unused offset = 0;
5295 /* Skip empty nodes */
5296 if (!pgdat->node_spanned_pages)
5299 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5300 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5301 offset = pgdat->node_start_pfn - start;
5302 /* ia64 gets its own node_mem_map, before this, without bootmem */
5303 if (!pgdat->node_mem_map) {
5304 unsigned long size, end;
5308 * The zone's endpoints aren't required to be MAX_ORDER
5309 * aligned but the node_mem_map endpoints must be in order
5310 * for the buddy allocator to function correctly.
5312 end = pgdat_end_pfn(pgdat);
5313 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5314 size = (end - start) * sizeof(struct page);
5315 map = alloc_remap(pgdat->node_id, size);
5317 map = memblock_virt_alloc_node_nopanic(size,
5319 pgdat->node_mem_map = map + offset;
5321 #ifndef CONFIG_NEED_MULTIPLE_NODES
5323 * With no DISCONTIG, the global mem_map is just set as node 0's
5325 if (pgdat == NODE_DATA(0)) {
5326 mem_map = NODE_DATA(0)->node_mem_map;
5327 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5328 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5330 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5333 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5336 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5337 unsigned long node_start_pfn, unsigned long *zholes_size)
5339 pg_data_t *pgdat = NODE_DATA(nid);
5340 unsigned long start_pfn = 0;
5341 unsigned long end_pfn = 0;
5343 /* pg_data_t should be reset to zero when it's allocated */
5344 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5346 reset_deferred_meminit(pgdat);
5347 pgdat->node_id = nid;
5348 pgdat->node_start_pfn = node_start_pfn;
5349 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5350 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5351 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5352 (u64)start_pfn << PAGE_SHIFT,
5353 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5355 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5356 zones_size, zholes_size);
5358 alloc_node_mem_map(pgdat);
5359 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5360 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5361 nid, (unsigned long)pgdat,
5362 (unsigned long)pgdat->node_mem_map);
5365 free_area_init_core(pgdat);
5368 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5370 #if MAX_NUMNODES > 1
5372 * Figure out the number of possible node ids.
5374 void __init setup_nr_node_ids(void)
5376 unsigned int highest;
5378 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5379 nr_node_ids = highest + 1;
5384 * node_map_pfn_alignment - determine the maximum internode alignment
5386 * This function should be called after node map is populated and sorted.
5387 * It calculates the maximum power of two alignment which can distinguish
5390 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5391 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5392 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5393 * shifted, 1GiB is enough and this function will indicate so.
5395 * This is used to test whether pfn -> nid mapping of the chosen memory
5396 * model has fine enough granularity to avoid incorrect mapping for the
5397 * populated node map.
5399 * Returns the determined alignment in pfn's. 0 if there is no alignment
5400 * requirement (single node).
5402 unsigned long __init node_map_pfn_alignment(void)
5404 unsigned long accl_mask = 0, last_end = 0;
5405 unsigned long start, end, mask;
5409 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5410 if (!start || last_nid < 0 || last_nid == nid) {
5417 * Start with a mask granular enough to pin-point to the
5418 * start pfn and tick off bits one-by-one until it becomes
5419 * too coarse to separate the current node from the last.
5421 mask = ~((1 << __ffs(start)) - 1);
5422 while (mask && last_end <= (start & (mask << 1)))
5425 /* accumulate all internode masks */
5429 /* convert mask to number of pages */
5430 return ~accl_mask + 1;
5433 /* Find the lowest pfn for a node */
5434 static unsigned long __init find_min_pfn_for_node(int nid)
5436 unsigned long min_pfn = ULONG_MAX;
5437 unsigned long start_pfn;
5440 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5441 min_pfn = min(min_pfn, start_pfn);
5443 if (min_pfn == ULONG_MAX) {
5445 "Could not find start_pfn for node %d\n", nid);
5453 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5455 * It returns the minimum PFN based on information provided via
5456 * memblock_set_node().
5458 unsigned long __init find_min_pfn_with_active_regions(void)
5460 return find_min_pfn_for_node(MAX_NUMNODES);
5464 * early_calculate_totalpages()
5465 * Sum pages in active regions for movable zone.
5466 * Populate N_MEMORY for calculating usable_nodes.
5468 static unsigned long __init early_calculate_totalpages(void)
5470 unsigned long totalpages = 0;
5471 unsigned long start_pfn, end_pfn;
5474 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5475 unsigned long pages = end_pfn - start_pfn;
5477 totalpages += pages;
5479 node_set_state(nid, N_MEMORY);
5485 * Find the PFN the Movable zone begins in each node. Kernel memory
5486 * is spread evenly between nodes as long as the nodes have enough
5487 * memory. When they don't, some nodes will have more kernelcore than
5490 static void __init find_zone_movable_pfns_for_nodes(void)
5493 unsigned long usable_startpfn;
5494 unsigned long kernelcore_node, kernelcore_remaining;
5495 /* save the state before borrow the nodemask */
5496 nodemask_t saved_node_state = node_states[N_MEMORY];
5497 unsigned long totalpages = early_calculate_totalpages();
5498 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5499 struct memblock_region *r;
5501 /* Need to find movable_zone earlier when movable_node is specified. */
5502 find_usable_zone_for_movable();
5505 * If movable_node is specified, ignore kernelcore and movablecore
5508 if (movable_node_is_enabled()) {
5509 for_each_memblock(memory, r) {
5510 if (!memblock_is_hotpluggable(r))
5515 usable_startpfn = PFN_DOWN(r->base);
5516 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5517 min(usable_startpfn, zone_movable_pfn[nid]) :
5525 * If movablecore=nn[KMG] was specified, calculate what size of
5526 * kernelcore that corresponds so that memory usable for
5527 * any allocation type is evenly spread. If both kernelcore
5528 * and movablecore are specified, then the value of kernelcore
5529 * will be used for required_kernelcore if it's greater than
5530 * what movablecore would have allowed.
5532 if (required_movablecore) {
5533 unsigned long corepages;
5536 * Round-up so that ZONE_MOVABLE is at least as large as what
5537 * was requested by the user
5539 required_movablecore =
5540 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5541 required_movablecore = min(totalpages, required_movablecore);
5542 corepages = totalpages - required_movablecore;
5544 required_kernelcore = max(required_kernelcore, corepages);
5548 * If kernelcore was not specified or kernelcore size is larger
5549 * than totalpages, there is no ZONE_MOVABLE.
5551 if (!required_kernelcore || required_kernelcore >= totalpages)
5554 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5555 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5558 /* Spread kernelcore memory as evenly as possible throughout nodes */
5559 kernelcore_node = required_kernelcore / usable_nodes;
5560 for_each_node_state(nid, N_MEMORY) {
5561 unsigned long start_pfn, end_pfn;
5564 * Recalculate kernelcore_node if the division per node
5565 * now exceeds what is necessary to satisfy the requested
5566 * amount of memory for the kernel
5568 if (required_kernelcore < kernelcore_node)
5569 kernelcore_node = required_kernelcore / usable_nodes;
5572 * As the map is walked, we track how much memory is usable
5573 * by the kernel using kernelcore_remaining. When it is
5574 * 0, the rest of the node is usable by ZONE_MOVABLE
5576 kernelcore_remaining = kernelcore_node;
5578 /* Go through each range of PFNs within this node */
5579 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5580 unsigned long size_pages;
5582 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5583 if (start_pfn >= end_pfn)
5586 /* Account for what is only usable for kernelcore */
5587 if (start_pfn < usable_startpfn) {
5588 unsigned long kernel_pages;
5589 kernel_pages = min(end_pfn, usable_startpfn)
5592 kernelcore_remaining -= min(kernel_pages,
5593 kernelcore_remaining);
5594 required_kernelcore -= min(kernel_pages,
5595 required_kernelcore);
5597 /* Continue if range is now fully accounted */
5598 if (end_pfn <= usable_startpfn) {
5601 * Push zone_movable_pfn to the end so
5602 * that if we have to rebalance
5603 * kernelcore across nodes, we will
5604 * not double account here
5606 zone_movable_pfn[nid] = end_pfn;
5609 start_pfn = usable_startpfn;
5613 * The usable PFN range for ZONE_MOVABLE is from
5614 * start_pfn->end_pfn. Calculate size_pages as the
5615 * number of pages used as kernelcore
5617 size_pages = end_pfn - start_pfn;
5618 if (size_pages > kernelcore_remaining)
5619 size_pages = kernelcore_remaining;
5620 zone_movable_pfn[nid] = start_pfn + size_pages;
5623 * Some kernelcore has been met, update counts and
5624 * break if the kernelcore for this node has been
5627 required_kernelcore -= min(required_kernelcore,
5629 kernelcore_remaining -= size_pages;
5630 if (!kernelcore_remaining)
5636 * If there is still required_kernelcore, we do another pass with one
5637 * less node in the count. This will push zone_movable_pfn[nid] further
5638 * along on the nodes that still have memory until kernelcore is
5642 if (usable_nodes && required_kernelcore > usable_nodes)
5646 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5647 for (nid = 0; nid < MAX_NUMNODES; nid++)
5648 zone_movable_pfn[nid] =
5649 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5652 /* restore the node_state */
5653 node_states[N_MEMORY] = saved_node_state;
5656 /* Any regular or high memory on that node ? */
5657 static void check_for_memory(pg_data_t *pgdat, int nid)
5659 enum zone_type zone_type;
5661 if (N_MEMORY == N_NORMAL_MEMORY)
5664 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5665 struct zone *zone = &pgdat->node_zones[zone_type];
5666 if (populated_zone(zone)) {
5667 node_set_state(nid, N_HIGH_MEMORY);
5668 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5669 zone_type <= ZONE_NORMAL)
5670 node_set_state(nid, N_NORMAL_MEMORY);
5677 * free_area_init_nodes - Initialise all pg_data_t and zone data
5678 * @max_zone_pfn: an array of max PFNs for each zone
5680 * This will call free_area_init_node() for each active node in the system.
5681 * Using the page ranges provided by memblock_set_node(), the size of each
5682 * zone in each node and their holes is calculated. If the maximum PFN
5683 * between two adjacent zones match, it is assumed that the zone is empty.
5684 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5685 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5686 * starts where the previous one ended. For example, ZONE_DMA32 starts
5687 * at arch_max_dma_pfn.
5689 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5691 unsigned long start_pfn, end_pfn;
5694 /* Record where the zone boundaries are */
5695 memset(arch_zone_lowest_possible_pfn, 0,
5696 sizeof(arch_zone_lowest_possible_pfn));
5697 memset(arch_zone_highest_possible_pfn, 0,
5698 sizeof(arch_zone_highest_possible_pfn));
5700 start_pfn = find_min_pfn_with_active_regions();
5702 for (i = 0; i < MAX_NR_ZONES; i++) {
5703 if (i == ZONE_MOVABLE)
5706 end_pfn = max(max_zone_pfn[i], start_pfn);
5707 arch_zone_lowest_possible_pfn[i] = start_pfn;
5708 arch_zone_highest_possible_pfn[i] = end_pfn;
5710 start_pfn = end_pfn;
5712 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5713 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5715 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5716 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5717 find_zone_movable_pfns_for_nodes();
5719 /* Print out the zone ranges */
5720 pr_info("Zone ranges:\n");
5721 for (i = 0; i < MAX_NR_ZONES; i++) {
5722 if (i == ZONE_MOVABLE)
5724 pr_info(" %-8s ", zone_names[i]);
5725 if (arch_zone_lowest_possible_pfn[i] ==
5726 arch_zone_highest_possible_pfn[i])
5729 pr_cont("[mem %#018Lx-%#018Lx]\n",
5730 (u64)arch_zone_lowest_possible_pfn[i]
5732 ((u64)arch_zone_highest_possible_pfn[i]
5733 << PAGE_SHIFT) - 1);
5736 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5737 pr_info("Movable zone start for each node\n");
5738 for (i = 0; i < MAX_NUMNODES; i++) {
5739 if (zone_movable_pfn[i])
5740 pr_info(" Node %d: %#018Lx\n", i,
5741 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5744 /* Print out the early node map */
5745 pr_info("Early memory node ranges\n");
5746 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5747 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5748 (u64)start_pfn << PAGE_SHIFT,
5749 ((u64)end_pfn << PAGE_SHIFT) - 1);
5751 /* Initialise every node */
5752 mminit_verify_pageflags_layout();
5753 setup_nr_node_ids();
5754 for_each_online_node(nid) {
5755 pg_data_t *pgdat = NODE_DATA(nid);
5756 free_area_init_node(nid, NULL,
5757 find_min_pfn_for_node(nid), NULL);
5759 /* Any memory on that node */
5760 if (pgdat->node_present_pages)
5761 node_set_state(nid, N_MEMORY);
5762 check_for_memory(pgdat, nid);
5766 static int __init cmdline_parse_core(char *p, unsigned long *core)
5768 unsigned long long coremem;
5772 coremem = memparse(p, &p);
5773 *core = coremem >> PAGE_SHIFT;
5775 /* Paranoid check that UL is enough for the coremem value */
5776 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5782 * kernelcore=size sets the amount of memory for use for allocations that
5783 * cannot be reclaimed or migrated.
5785 static int __init cmdline_parse_kernelcore(char *p)
5787 return cmdline_parse_core(p, &required_kernelcore);
5791 * movablecore=size sets the amount of memory for use for allocations that
5792 * can be reclaimed or migrated.
5794 static int __init cmdline_parse_movablecore(char *p)
5796 return cmdline_parse_core(p, &required_movablecore);
5799 early_param("kernelcore", cmdline_parse_kernelcore);
5800 early_param("movablecore", cmdline_parse_movablecore);
5802 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5804 void adjust_managed_page_count(struct page *page, long count)
5806 spin_lock(&managed_page_count_lock);
5807 page_zone(page)->managed_pages += count;
5808 totalram_pages += count;
5809 #ifdef CONFIG_HIGHMEM
5810 if (PageHighMem(page))
5811 totalhigh_pages += count;
5813 spin_unlock(&managed_page_count_lock);
5815 EXPORT_SYMBOL(adjust_managed_page_count);
5817 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5820 unsigned long pages = 0;
5822 start = (void *)PAGE_ALIGN((unsigned long)start);
5823 end = (void *)((unsigned long)end & PAGE_MASK);
5824 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5825 if ((unsigned int)poison <= 0xFF)
5826 memset(pos, poison, PAGE_SIZE);
5827 free_reserved_page(virt_to_page(pos));
5831 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5832 s, pages << (PAGE_SHIFT - 10), start, end);
5836 EXPORT_SYMBOL(free_reserved_area);
5838 #ifdef CONFIG_HIGHMEM
5839 void free_highmem_page(struct page *page)
5841 __free_reserved_page(page);
5843 page_zone(page)->managed_pages++;
5849 void __init mem_init_print_info(const char *str)
5851 unsigned long physpages, codesize, datasize, rosize, bss_size;
5852 unsigned long init_code_size, init_data_size;
5854 physpages = get_num_physpages();
5855 codesize = _etext - _stext;
5856 datasize = _edata - _sdata;
5857 rosize = __end_rodata - __start_rodata;
5858 bss_size = __bss_stop - __bss_start;
5859 init_data_size = __init_end - __init_begin;
5860 init_code_size = _einittext - _sinittext;
5863 * Detect special cases and adjust section sizes accordingly:
5864 * 1) .init.* may be embedded into .data sections
5865 * 2) .init.text.* may be out of [__init_begin, __init_end],
5866 * please refer to arch/tile/kernel/vmlinux.lds.S.
5867 * 3) .rodata.* may be embedded into .text or .data sections.
5869 #define adj_init_size(start, end, size, pos, adj) \
5871 if (start <= pos && pos < end && size > adj) \
5875 adj_init_size(__init_begin, __init_end, init_data_size,
5876 _sinittext, init_code_size);
5877 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5878 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5879 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5880 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5882 #undef adj_init_size
5884 pr_info("Memory: %luK/%luK available "
5885 "(%luK kernel code, %luK rwdata, %luK rodata, "
5886 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5887 #ifdef CONFIG_HIGHMEM
5891 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5892 codesize >> 10, datasize >> 10, rosize >> 10,
5893 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5894 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5895 totalcma_pages << (PAGE_SHIFT-10),
5896 #ifdef CONFIG_HIGHMEM
5897 totalhigh_pages << (PAGE_SHIFT-10),
5899 str ? ", " : "", str ? str : "");
5903 * set_dma_reserve - set the specified number of pages reserved in the first zone
5904 * @new_dma_reserve: The number of pages to mark reserved
5906 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
5907 * In the DMA zone, a significant percentage may be consumed by kernel image
5908 * and other unfreeable allocations which can skew the watermarks badly. This
5909 * function may optionally be used to account for unfreeable pages in the
5910 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5911 * smaller per-cpu batchsize.
5913 void __init set_dma_reserve(unsigned long new_dma_reserve)
5915 dma_reserve = new_dma_reserve;
5918 void __init free_area_init(unsigned long *zones_size)
5920 free_area_init_node(0, zones_size,
5921 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5924 static int page_alloc_cpu_notify(struct notifier_block *self,
5925 unsigned long action, void *hcpu)
5927 int cpu = (unsigned long)hcpu;
5929 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5930 lru_add_drain_cpu(cpu);
5934 * Spill the event counters of the dead processor
5935 * into the current processors event counters.
5936 * This artificially elevates the count of the current
5939 vm_events_fold_cpu(cpu);
5942 * Zero the differential counters of the dead processor
5943 * so that the vm statistics are consistent.
5945 * This is only okay since the processor is dead and cannot
5946 * race with what we are doing.
5948 cpu_vm_stats_fold(cpu);
5953 void __init page_alloc_init(void)
5955 hotcpu_notifier(page_alloc_cpu_notify, 0);
5959 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5960 * or min_free_kbytes changes.
5962 static void calculate_totalreserve_pages(void)
5964 struct pglist_data *pgdat;
5965 unsigned long reserve_pages = 0;
5966 enum zone_type i, j;
5968 for_each_online_pgdat(pgdat) {
5969 for (i = 0; i < MAX_NR_ZONES; i++) {
5970 struct zone *zone = pgdat->node_zones + i;
5973 /* Find valid and maximum lowmem_reserve in the zone */
5974 for (j = i; j < MAX_NR_ZONES; j++) {
5975 if (zone->lowmem_reserve[j] > max)
5976 max = zone->lowmem_reserve[j];
5979 /* we treat the high watermark as reserved pages. */
5980 max += high_wmark_pages(zone);
5982 if (max > zone->managed_pages)
5983 max = zone->managed_pages;
5984 reserve_pages += max;
5986 * Lowmem reserves are not available to
5987 * GFP_HIGHUSER page cache allocations and
5988 * kswapd tries to balance zones to their high
5989 * watermark. As a result, neither should be
5990 * regarded as dirtyable memory, to prevent a
5991 * situation where reclaim has to clean pages
5992 * in order to balance the zones.
5994 zone->dirty_balance_reserve = max;
5997 dirty_balance_reserve = reserve_pages;
5998 totalreserve_pages = reserve_pages;
6002 * setup_per_zone_lowmem_reserve - called whenever
6003 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6004 * has a correct pages reserved value, so an adequate number of
6005 * pages are left in the zone after a successful __alloc_pages().
6007 static void setup_per_zone_lowmem_reserve(void)
6009 struct pglist_data *pgdat;
6010 enum zone_type j, idx;
6012 for_each_online_pgdat(pgdat) {
6013 for (j = 0; j < MAX_NR_ZONES; j++) {
6014 struct zone *zone = pgdat->node_zones + j;
6015 unsigned long managed_pages = zone->managed_pages;
6017 zone->lowmem_reserve[j] = 0;
6021 struct zone *lower_zone;
6025 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6026 sysctl_lowmem_reserve_ratio[idx] = 1;
6028 lower_zone = pgdat->node_zones + idx;
6029 lower_zone->lowmem_reserve[j] = managed_pages /
6030 sysctl_lowmem_reserve_ratio[idx];
6031 managed_pages += lower_zone->managed_pages;
6036 /* update totalreserve_pages */
6037 calculate_totalreserve_pages();
6040 static void __setup_per_zone_wmarks(void)
6042 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6043 unsigned long lowmem_pages = 0;
6045 unsigned long flags;
6047 /* Calculate total number of !ZONE_HIGHMEM pages */
6048 for_each_zone(zone) {
6049 if (!is_highmem(zone))
6050 lowmem_pages += zone->managed_pages;
6053 for_each_zone(zone) {
6056 spin_lock_irqsave(&zone->lock, flags);
6057 tmp = (u64)pages_min * zone->managed_pages;
6058 do_div(tmp, lowmem_pages);
6059 if (is_highmem(zone)) {
6061 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6062 * need highmem pages, so cap pages_min to a small
6065 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6066 * deltas control asynch page reclaim, and so should
6067 * not be capped for highmem.
6069 unsigned long min_pages;
6071 min_pages = zone->managed_pages / 1024;
6072 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6073 zone->watermark[WMARK_MIN] = min_pages;
6076 * If it's a lowmem zone, reserve a number of pages
6077 * proportionate to the zone's size.
6079 zone->watermark[WMARK_MIN] = tmp;
6082 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
6083 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
6085 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6086 high_wmark_pages(zone) - low_wmark_pages(zone) -
6087 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6089 spin_unlock_irqrestore(&zone->lock, flags);
6092 /* update totalreserve_pages */
6093 calculate_totalreserve_pages();
6097 * setup_per_zone_wmarks - called when min_free_kbytes changes
6098 * or when memory is hot-{added|removed}
6100 * Ensures that the watermark[min,low,high] values for each zone are set
6101 * correctly with respect to min_free_kbytes.
6103 void setup_per_zone_wmarks(void)
6105 mutex_lock(&zonelists_mutex);
6106 __setup_per_zone_wmarks();
6107 mutex_unlock(&zonelists_mutex);
6111 * The inactive anon list should be small enough that the VM never has to
6112 * do too much work, but large enough that each inactive page has a chance
6113 * to be referenced again before it is swapped out.
6115 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6116 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6117 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6118 * the anonymous pages are kept on the inactive list.
6121 * memory ratio inactive anon
6122 * -------------------------------------
6131 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6133 unsigned int gb, ratio;
6135 /* Zone size in gigabytes */
6136 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6138 ratio = int_sqrt(10 * gb);
6142 zone->inactive_ratio = ratio;
6145 static void __meminit setup_per_zone_inactive_ratio(void)
6150 calculate_zone_inactive_ratio(zone);
6154 * Initialise min_free_kbytes.
6156 * For small machines we want it small (128k min). For large machines
6157 * we want it large (64MB max). But it is not linear, because network
6158 * bandwidth does not increase linearly with machine size. We use
6160 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6161 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6177 int __meminit init_per_zone_wmark_min(void)
6179 unsigned long lowmem_kbytes;
6180 int new_min_free_kbytes;
6182 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6183 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6185 if (new_min_free_kbytes > user_min_free_kbytes) {
6186 min_free_kbytes = new_min_free_kbytes;
6187 if (min_free_kbytes < 128)
6188 min_free_kbytes = 128;
6189 if (min_free_kbytes > 65536)
6190 min_free_kbytes = 65536;
6192 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6193 new_min_free_kbytes, user_min_free_kbytes);
6195 setup_per_zone_wmarks();
6196 refresh_zone_stat_thresholds();
6197 setup_per_zone_lowmem_reserve();
6198 setup_per_zone_inactive_ratio();
6201 core_initcall(init_per_zone_wmark_min)
6204 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6205 * that we can call two helper functions whenever min_free_kbytes
6208 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6209 void __user *buffer, size_t *length, loff_t *ppos)
6213 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6218 user_min_free_kbytes = min_free_kbytes;
6219 setup_per_zone_wmarks();
6225 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6226 void __user *buffer, size_t *length, loff_t *ppos)
6231 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6236 zone->min_unmapped_pages = (zone->managed_pages *
6237 sysctl_min_unmapped_ratio) / 100;
6241 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6242 void __user *buffer, size_t *length, loff_t *ppos)
6247 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6252 zone->min_slab_pages = (zone->managed_pages *
6253 sysctl_min_slab_ratio) / 100;
6259 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6260 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6261 * whenever sysctl_lowmem_reserve_ratio changes.
6263 * The reserve ratio obviously has absolutely no relation with the
6264 * minimum watermarks. The lowmem reserve ratio can only make sense
6265 * if in function of the boot time zone sizes.
6267 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6268 void __user *buffer, size_t *length, loff_t *ppos)
6270 proc_dointvec_minmax(table, write, buffer, length, ppos);
6271 setup_per_zone_lowmem_reserve();
6276 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6277 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6278 * pagelist can have before it gets flushed back to buddy allocator.
6280 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6281 void __user *buffer, size_t *length, loff_t *ppos)
6284 int old_percpu_pagelist_fraction;
6287 mutex_lock(&pcp_batch_high_lock);
6288 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6290 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6291 if (!write || ret < 0)
6294 /* Sanity checking to avoid pcp imbalance */
6295 if (percpu_pagelist_fraction &&
6296 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6297 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6303 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6306 for_each_populated_zone(zone) {
6309 for_each_possible_cpu(cpu)
6310 pageset_set_high_and_batch(zone,
6311 per_cpu_ptr(zone->pageset, cpu));
6314 mutex_unlock(&pcp_batch_high_lock);
6319 int hashdist = HASHDIST_DEFAULT;
6321 static int __init set_hashdist(char *str)
6325 hashdist = simple_strtoul(str, &str, 0);
6328 __setup("hashdist=", set_hashdist);
6332 * allocate a large system hash table from bootmem
6333 * - it is assumed that the hash table must contain an exact power-of-2
6334 * quantity of entries
6335 * - limit is the number of hash buckets, not the total allocation size
6337 void *__init alloc_large_system_hash(const char *tablename,
6338 unsigned long bucketsize,
6339 unsigned long numentries,
6342 unsigned int *_hash_shift,
6343 unsigned int *_hash_mask,
6344 unsigned long low_limit,
6345 unsigned long high_limit)
6347 unsigned long long max = high_limit;
6348 unsigned long log2qty, size;
6351 /* allow the kernel cmdline to have a say */
6353 /* round applicable memory size up to nearest megabyte */
6354 numentries = nr_kernel_pages;
6356 /* It isn't necessary when PAGE_SIZE >= 1MB */
6357 if (PAGE_SHIFT < 20)
6358 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6360 /* limit to 1 bucket per 2^scale bytes of low memory */
6361 if (scale > PAGE_SHIFT)
6362 numentries >>= (scale - PAGE_SHIFT);
6364 numentries <<= (PAGE_SHIFT - scale);
6366 /* Make sure we've got at least a 0-order allocation.. */
6367 if (unlikely(flags & HASH_SMALL)) {
6368 /* Makes no sense without HASH_EARLY */
6369 WARN_ON(!(flags & HASH_EARLY));
6370 if (!(numentries >> *_hash_shift)) {
6371 numentries = 1UL << *_hash_shift;
6372 BUG_ON(!numentries);
6374 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6375 numentries = PAGE_SIZE / bucketsize;
6377 numentries = roundup_pow_of_two(numentries);
6379 /* limit allocation size to 1/16 total memory by default */
6381 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6382 do_div(max, bucketsize);
6384 max = min(max, 0x80000000ULL);
6386 if (numentries < low_limit)
6387 numentries = low_limit;
6388 if (numentries > max)
6391 log2qty = ilog2(numentries);
6394 size = bucketsize << log2qty;
6395 if (flags & HASH_EARLY)
6396 table = memblock_virt_alloc_nopanic(size, 0);
6398 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6401 * If bucketsize is not a power-of-two, we may free
6402 * some pages at the end of hash table which
6403 * alloc_pages_exact() automatically does
6405 if (get_order(size) < MAX_ORDER) {
6406 table = alloc_pages_exact(size, GFP_ATOMIC);
6407 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6410 } while (!table && size > PAGE_SIZE && --log2qty);
6413 panic("Failed to allocate %s hash table\n", tablename);
6415 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6418 ilog2(size) - PAGE_SHIFT,
6422 *_hash_shift = log2qty;
6424 *_hash_mask = (1 << log2qty) - 1;
6429 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6430 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6433 #ifdef CONFIG_SPARSEMEM
6434 return __pfn_to_section(pfn)->pageblock_flags;
6436 return zone->pageblock_flags;
6437 #endif /* CONFIG_SPARSEMEM */
6440 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6442 #ifdef CONFIG_SPARSEMEM
6443 pfn &= (PAGES_PER_SECTION-1);
6444 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6446 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6447 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6448 #endif /* CONFIG_SPARSEMEM */
6452 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6453 * @page: The page within the block of interest
6454 * @pfn: The target page frame number
6455 * @end_bitidx: The last bit of interest to retrieve
6456 * @mask: mask of bits that the caller is interested in
6458 * Return: pageblock_bits flags
6460 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6461 unsigned long end_bitidx,
6465 unsigned long *bitmap;
6466 unsigned long bitidx, word_bitidx;
6469 zone = page_zone(page);
6470 bitmap = get_pageblock_bitmap(zone, pfn);
6471 bitidx = pfn_to_bitidx(zone, pfn);
6472 word_bitidx = bitidx / BITS_PER_LONG;
6473 bitidx &= (BITS_PER_LONG-1);
6475 word = bitmap[word_bitidx];
6476 bitidx += end_bitidx;
6477 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6481 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6482 * @page: The page within the block of interest
6483 * @flags: The flags to set
6484 * @pfn: The target page frame number
6485 * @end_bitidx: The last bit of interest
6486 * @mask: mask of bits that the caller is interested in
6488 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6490 unsigned long end_bitidx,
6494 unsigned long *bitmap;
6495 unsigned long bitidx, word_bitidx;
6496 unsigned long old_word, word;
6498 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6500 zone = page_zone(page);
6501 bitmap = get_pageblock_bitmap(zone, pfn);
6502 bitidx = pfn_to_bitidx(zone, pfn);
6503 word_bitidx = bitidx / BITS_PER_LONG;
6504 bitidx &= (BITS_PER_LONG-1);
6506 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6508 bitidx += end_bitidx;
6509 mask <<= (BITS_PER_LONG - bitidx - 1);
6510 flags <<= (BITS_PER_LONG - bitidx - 1);
6512 word = READ_ONCE(bitmap[word_bitidx]);
6514 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6515 if (word == old_word)
6522 * This function checks whether pageblock includes unmovable pages or not.
6523 * If @count is not zero, it is okay to include less @count unmovable pages
6525 * PageLRU check without isolation or lru_lock could race so that
6526 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6527 * expect this function should be exact.
6529 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6530 bool skip_hwpoisoned_pages)
6532 unsigned long pfn, iter, found;
6536 * For avoiding noise data, lru_add_drain_all() should be called
6537 * If ZONE_MOVABLE, the zone never contains unmovable pages
6539 if (zone_idx(zone) == ZONE_MOVABLE)
6541 mt = get_pageblock_migratetype(page);
6542 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6545 pfn = page_to_pfn(page);
6546 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6547 unsigned long check = pfn + iter;
6549 if (!pfn_valid_within(check))
6552 page = pfn_to_page(check);
6555 * Hugepages are not in LRU lists, but they're movable.
6556 * We need not scan over tail pages bacause we don't
6557 * handle each tail page individually in migration.
6559 if (PageHuge(page)) {
6560 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6565 * We can't use page_count without pin a page
6566 * because another CPU can free compound page.
6567 * This check already skips compound tails of THP
6568 * because their page->_count is zero at all time.
6570 if (!atomic_read(&page->_count)) {
6571 if (PageBuddy(page))
6572 iter += (1 << page_order(page)) - 1;
6577 * The HWPoisoned page may be not in buddy system, and
6578 * page_count() is not 0.
6580 if (skip_hwpoisoned_pages && PageHWPoison(page))
6586 * If there are RECLAIMABLE pages, we need to check
6587 * it. But now, memory offline itself doesn't call
6588 * shrink_node_slabs() and it still to be fixed.
6591 * If the page is not RAM, page_count()should be 0.
6592 * we don't need more check. This is an _used_ not-movable page.
6594 * The problematic thing here is PG_reserved pages. PG_reserved
6595 * is set to both of a memory hole page and a _used_ kernel
6604 bool is_pageblock_removable_nolock(struct page *page)
6610 * We have to be careful here because we are iterating over memory
6611 * sections which are not zone aware so we might end up outside of
6612 * the zone but still within the section.
6613 * We have to take care about the node as well. If the node is offline
6614 * its NODE_DATA will be NULL - see page_zone.
6616 if (!node_online(page_to_nid(page)))
6619 zone = page_zone(page);
6620 pfn = page_to_pfn(page);
6621 if (!zone_spans_pfn(zone, pfn))
6624 return !has_unmovable_pages(zone, page, 0, true);
6629 static unsigned long pfn_max_align_down(unsigned long pfn)
6631 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6632 pageblock_nr_pages) - 1);
6635 static unsigned long pfn_max_align_up(unsigned long pfn)
6637 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6638 pageblock_nr_pages));
6641 /* [start, end) must belong to a single zone. */
6642 static int __alloc_contig_migrate_range(struct compact_control *cc,
6643 unsigned long start, unsigned long end)
6645 /* This function is based on compact_zone() from compaction.c. */
6646 unsigned long nr_reclaimed;
6647 unsigned long pfn = start;
6648 unsigned int tries = 0;
6653 while (pfn < end || !list_empty(&cc->migratepages)) {
6654 if (fatal_signal_pending(current)) {
6659 if (list_empty(&cc->migratepages)) {
6660 cc->nr_migratepages = 0;
6661 pfn = isolate_migratepages_range(cc, pfn, end);
6667 } else if (++tries == 5) {
6668 ret = ret < 0 ? ret : -EBUSY;
6672 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6674 cc->nr_migratepages -= nr_reclaimed;
6676 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6677 NULL, 0, cc->mode, MR_CMA);
6680 putback_movable_pages(&cc->migratepages);
6687 * alloc_contig_range() -- tries to allocate given range of pages
6688 * @start: start PFN to allocate
6689 * @end: one-past-the-last PFN to allocate
6690 * @migratetype: migratetype of the underlaying pageblocks (either
6691 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6692 * in range must have the same migratetype and it must
6693 * be either of the two.
6695 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6696 * aligned, however it's the caller's responsibility to guarantee that
6697 * we are the only thread that changes migrate type of pageblocks the
6700 * The PFN range must belong to a single zone.
6702 * Returns zero on success or negative error code. On success all
6703 * pages which PFN is in [start, end) are allocated for the caller and
6704 * need to be freed with free_contig_range().
6706 int alloc_contig_range(unsigned long start, unsigned long end,
6707 unsigned migratetype)
6709 unsigned long outer_start, outer_end;
6713 struct compact_control cc = {
6714 .nr_migratepages = 0,
6716 .zone = page_zone(pfn_to_page(start)),
6717 .mode = MIGRATE_SYNC,
6718 .ignore_skip_hint = true,
6720 INIT_LIST_HEAD(&cc.migratepages);
6723 * What we do here is we mark all pageblocks in range as
6724 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6725 * have different sizes, and due to the way page allocator
6726 * work, we align the range to biggest of the two pages so
6727 * that page allocator won't try to merge buddies from
6728 * different pageblocks and change MIGRATE_ISOLATE to some
6729 * other migration type.
6731 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6732 * migrate the pages from an unaligned range (ie. pages that
6733 * we are interested in). This will put all the pages in
6734 * range back to page allocator as MIGRATE_ISOLATE.
6736 * When this is done, we take the pages in range from page
6737 * allocator removing them from the buddy system. This way
6738 * page allocator will never consider using them.
6740 * This lets us mark the pageblocks back as
6741 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6742 * aligned range but not in the unaligned, original range are
6743 * put back to page allocator so that buddy can use them.
6746 ret = start_isolate_page_range(pfn_max_align_down(start),
6747 pfn_max_align_up(end), migratetype,
6752 ret = __alloc_contig_migrate_range(&cc, start, end);
6757 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6758 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6759 * more, all pages in [start, end) are free in page allocator.
6760 * What we are going to do is to allocate all pages from
6761 * [start, end) (that is remove them from page allocator).
6763 * The only problem is that pages at the beginning and at the
6764 * end of interesting range may be not aligned with pages that
6765 * page allocator holds, ie. they can be part of higher order
6766 * pages. Because of this, we reserve the bigger range and
6767 * once this is done free the pages we are not interested in.
6769 * We don't have to hold zone->lock here because the pages are
6770 * isolated thus they won't get removed from buddy.
6773 lru_add_drain_all();
6774 drain_all_pages(cc.zone);
6777 outer_start = start;
6778 while (!PageBuddy(pfn_to_page(outer_start))) {
6779 if (++order >= MAX_ORDER) {
6783 outer_start &= ~0UL << order;
6786 /* Make sure the range is really isolated. */
6787 if (test_pages_isolated(outer_start, end, false)) {
6788 pr_info("%s: [%lx, %lx) PFNs busy\n",
6789 __func__, outer_start, end);
6794 /* Grab isolated pages from freelists. */
6795 outer_end = isolate_freepages_range(&cc, outer_start, end);
6801 /* Free head and tail (if any) */
6802 if (start != outer_start)
6803 free_contig_range(outer_start, start - outer_start);
6804 if (end != outer_end)
6805 free_contig_range(end, outer_end - end);
6808 undo_isolate_page_range(pfn_max_align_down(start),
6809 pfn_max_align_up(end), migratetype);
6813 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6815 unsigned int count = 0;
6817 for (; nr_pages--; pfn++) {
6818 struct page *page = pfn_to_page(pfn);
6820 count += page_count(page) != 1;
6823 WARN(count != 0, "%d pages are still in use!\n", count);
6827 #ifdef CONFIG_MEMORY_HOTPLUG
6829 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6830 * page high values need to be recalulated.
6832 void __meminit zone_pcp_update(struct zone *zone)
6835 mutex_lock(&pcp_batch_high_lock);
6836 for_each_possible_cpu(cpu)
6837 pageset_set_high_and_batch(zone,
6838 per_cpu_ptr(zone->pageset, cpu));
6839 mutex_unlock(&pcp_batch_high_lock);
6843 void zone_pcp_reset(struct zone *zone)
6845 unsigned long flags;
6847 struct per_cpu_pageset *pset;
6849 /* avoid races with drain_pages() */
6850 local_irq_save(flags);
6851 if (zone->pageset != &boot_pageset) {
6852 for_each_online_cpu(cpu) {
6853 pset = per_cpu_ptr(zone->pageset, cpu);
6854 drain_zonestat(zone, pset);
6856 free_percpu(zone->pageset);
6857 zone->pageset = &boot_pageset;
6859 local_irq_restore(flags);
6862 #ifdef CONFIG_MEMORY_HOTREMOVE
6864 * All pages in the range must be isolated before calling this.
6867 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6871 unsigned int order, i;
6873 unsigned long flags;
6874 /* find the first valid pfn */
6875 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6880 zone = page_zone(pfn_to_page(pfn));
6881 spin_lock_irqsave(&zone->lock, flags);
6883 while (pfn < end_pfn) {
6884 if (!pfn_valid(pfn)) {
6888 page = pfn_to_page(pfn);
6890 * The HWPoisoned page may be not in buddy system, and
6891 * page_count() is not 0.
6893 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6895 SetPageReserved(page);
6899 BUG_ON(page_count(page));
6900 BUG_ON(!PageBuddy(page));
6901 order = page_order(page);
6902 #ifdef CONFIG_DEBUG_VM
6903 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6904 pfn, 1 << order, end_pfn);
6906 list_del(&page->lru);
6907 rmv_page_order(page);
6908 zone->free_area[order].nr_free--;
6909 for (i = 0; i < (1 << order); i++)
6910 SetPageReserved((page+i));
6911 pfn += (1 << order);
6913 spin_unlock_irqrestore(&zone->lock, flags);
6917 #ifdef CONFIG_MEMORY_FAILURE
6918 bool is_free_buddy_page(struct page *page)
6920 struct zone *zone = page_zone(page);
6921 unsigned long pfn = page_to_pfn(page);
6922 unsigned long flags;
6925 spin_lock_irqsave(&zone->lock, flags);
6926 for (order = 0; order < MAX_ORDER; order++) {
6927 struct page *page_head = page - (pfn & ((1 << order) - 1));
6929 if (PageBuddy(page_head) && page_order(page_head) >= order)
6932 spin_unlock_irqrestore(&zone->lock, flags);
6934 return order < MAX_ORDER;