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/memremap.h>
47 #include <linux/stop_machine.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/page_ext.h>
54 #include <linux/debugobjects.h>
55 #include <linux/kmemleak.h>
56 #include <linux/compaction.h>
57 #include <trace/events/kmem.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/page_ext.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
67 #include <asm/sections.h>
68 #include <asm/tlbflush.h>
69 #include <asm/div64.h>
72 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
73 static DEFINE_MUTEX(pcp_batch_high_lock);
74 #define MIN_PERCPU_PAGELIST_FRACTION (8)
76 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
77 DEFINE_PER_CPU(int, numa_node);
78 EXPORT_PER_CPU_SYMBOL(numa_node);
81 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
83 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
84 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
85 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
86 * defined in <linux/topology.h>.
88 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
89 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
90 int _node_numa_mem_[MAX_NUMNODES];
94 * Array of node states.
96 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
97 [N_POSSIBLE] = NODE_MASK_ALL,
98 [N_ONLINE] = { { [0] = 1UL } },
100 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
101 #ifdef CONFIG_HIGHMEM
102 [N_HIGH_MEMORY] = { { [0] = 1UL } },
104 #ifdef CONFIG_MOVABLE_NODE
105 [N_MEMORY] = { { [0] = 1UL } },
107 [N_CPU] = { { [0] = 1UL } },
110 EXPORT_SYMBOL(node_states);
112 /* Protect totalram_pages and zone->managed_pages */
113 static DEFINE_SPINLOCK(managed_page_count_lock);
115 unsigned long totalram_pages __read_mostly;
116 unsigned long totalreserve_pages __read_mostly;
117 unsigned long totalcma_pages __read_mostly;
119 int percpu_pagelist_fraction;
120 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
123 * A cached value of the page's pageblock's migratetype, used when the page is
124 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
125 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
126 * Also the migratetype set in the page does not necessarily match the pcplist
127 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
128 * other index - this ensures that it will be put on the correct CMA freelist.
130 static inline int get_pcppage_migratetype(struct page *page)
135 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
137 page->index = migratetype;
140 #ifdef CONFIG_PM_SLEEP
142 * The following functions are used by the suspend/hibernate code to temporarily
143 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
144 * while devices are suspended. To avoid races with the suspend/hibernate code,
145 * they should always be called with pm_mutex held (gfp_allowed_mask also should
146 * only be modified with pm_mutex held, unless the suspend/hibernate code is
147 * guaranteed not to run in parallel with that modification).
150 static gfp_t saved_gfp_mask;
152 void pm_restore_gfp_mask(void)
154 WARN_ON(!mutex_is_locked(&pm_mutex));
155 if (saved_gfp_mask) {
156 gfp_allowed_mask = saved_gfp_mask;
161 void pm_restrict_gfp_mask(void)
163 WARN_ON(!mutex_is_locked(&pm_mutex));
164 WARN_ON(saved_gfp_mask);
165 saved_gfp_mask = gfp_allowed_mask;
166 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
169 bool pm_suspended_storage(void)
171 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
175 #endif /* CONFIG_PM_SLEEP */
177 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
178 unsigned int pageblock_order __read_mostly;
181 static void __free_pages_ok(struct page *page, unsigned int order);
184 * results with 256, 32 in the lowmem_reserve sysctl:
185 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
186 * 1G machine -> (16M dma, 784M normal, 224M high)
187 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
188 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
189 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
191 * TBD: should special case ZONE_DMA32 machines here - in those we normally
192 * don't need any ZONE_NORMAL reservation
194 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
195 #ifdef CONFIG_ZONE_DMA
198 #ifdef CONFIG_ZONE_DMA32
201 #ifdef CONFIG_HIGHMEM
207 EXPORT_SYMBOL(totalram_pages);
209 static char * const zone_names[MAX_NR_ZONES] = {
210 #ifdef CONFIG_ZONE_DMA
213 #ifdef CONFIG_ZONE_DMA32
217 #ifdef CONFIG_HIGHMEM
221 #ifdef CONFIG_ZONE_DEVICE
226 char * const migratetype_names[MIGRATE_TYPES] = {
234 #ifdef CONFIG_MEMORY_ISOLATION
239 compound_page_dtor * const compound_page_dtors[] = {
242 #ifdef CONFIG_HUGETLB_PAGE
245 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
250 int min_free_kbytes = 1024;
251 int user_min_free_kbytes = -1;
253 static unsigned long __meminitdata nr_kernel_pages;
254 static unsigned long __meminitdata nr_all_pages;
255 static unsigned long __meminitdata dma_reserve;
257 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
258 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
259 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
260 static unsigned long __initdata required_kernelcore;
261 static unsigned long __initdata required_movablecore;
262 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
263 static bool mirrored_kernelcore;
265 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
267 EXPORT_SYMBOL(movable_zone);
268 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
271 int nr_node_ids __read_mostly = MAX_NUMNODES;
272 int nr_online_nodes __read_mostly = 1;
273 EXPORT_SYMBOL(nr_node_ids);
274 EXPORT_SYMBOL(nr_online_nodes);
277 int page_group_by_mobility_disabled __read_mostly;
279 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
280 static inline void reset_deferred_meminit(pg_data_t *pgdat)
282 pgdat->first_deferred_pfn = ULONG_MAX;
285 /* Returns true if the struct page for the pfn is uninitialised */
286 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
288 if (pfn >= NODE_DATA(early_pfn_to_nid(pfn))->first_deferred_pfn)
294 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
296 if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
303 * Returns false when the remaining initialisation should be deferred until
304 * later in the boot cycle when it can be parallelised.
306 static inline bool update_defer_init(pg_data_t *pgdat,
307 unsigned long pfn, unsigned long zone_end,
308 unsigned long *nr_initialised)
310 /* Always populate low zones for address-contrained allocations */
311 if (zone_end < pgdat_end_pfn(pgdat))
314 /* Initialise at least 2G of the highest zone */
316 if (*nr_initialised > (2UL << (30 - PAGE_SHIFT)) &&
317 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
318 pgdat->first_deferred_pfn = pfn;
325 static inline void reset_deferred_meminit(pg_data_t *pgdat)
329 static inline bool early_page_uninitialised(unsigned long pfn)
334 static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
339 static inline bool update_defer_init(pg_data_t *pgdat,
340 unsigned long pfn, unsigned long zone_end,
341 unsigned long *nr_initialised)
348 void set_pageblock_migratetype(struct page *page, int migratetype)
350 if (unlikely(page_group_by_mobility_disabled &&
351 migratetype < MIGRATE_PCPTYPES))
352 migratetype = MIGRATE_UNMOVABLE;
354 set_pageblock_flags_group(page, (unsigned long)migratetype,
355 PB_migrate, PB_migrate_end);
358 #ifdef CONFIG_DEBUG_VM
359 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
363 unsigned long pfn = page_to_pfn(page);
364 unsigned long sp, start_pfn;
367 seq = zone_span_seqbegin(zone);
368 start_pfn = zone->zone_start_pfn;
369 sp = zone->spanned_pages;
370 if (!zone_spans_pfn(zone, pfn))
372 } while (zone_span_seqretry(zone, seq));
375 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
376 pfn, zone_to_nid(zone), zone->name,
377 start_pfn, start_pfn + sp);
382 static int page_is_consistent(struct zone *zone, struct page *page)
384 if (!pfn_valid_within(page_to_pfn(page)))
386 if (zone != page_zone(page))
392 * Temporary debugging check for pages not lying within a given zone.
394 static int bad_range(struct zone *zone, struct page *page)
396 if (page_outside_zone_boundaries(zone, page))
398 if (!page_is_consistent(zone, page))
404 static inline int bad_range(struct zone *zone, struct page *page)
410 static void bad_page(struct page *page, const char *reason,
411 unsigned long bad_flags)
413 static unsigned long resume;
414 static unsigned long nr_shown;
415 static unsigned long nr_unshown;
417 /* Don't complain about poisoned pages */
418 if (PageHWPoison(page)) {
419 page_mapcount_reset(page); /* remove PageBuddy */
424 * Allow a burst of 60 reports, then keep quiet for that minute;
425 * or allow a steady drip of one report per second.
427 if (nr_shown == 60) {
428 if (time_before(jiffies, resume)) {
434 "BUG: Bad page state: %lu messages suppressed\n",
441 resume = jiffies + 60 * HZ;
443 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
444 current->comm, page_to_pfn(page));
445 __dump_page(page, reason);
446 bad_flags &= page->flags;
448 pr_alert("bad because of flags: %#lx(%pGp)\n",
449 bad_flags, &bad_flags);
450 dump_page_owner(page);
455 /* Leave bad fields for debug, except PageBuddy could make trouble */
456 page_mapcount_reset(page); /* remove PageBuddy */
457 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
461 * Higher-order pages are called "compound pages". They are structured thusly:
463 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
465 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
466 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
468 * The first tail page's ->compound_dtor holds the offset in array of compound
469 * page destructors. See compound_page_dtors.
471 * The first tail page's ->compound_order holds the order of allocation.
472 * This usage means that zero-order pages may not be compound.
475 void free_compound_page(struct page *page)
477 __free_pages_ok(page, compound_order(page));
480 void prep_compound_page(struct page *page, unsigned int order)
483 int nr_pages = 1 << order;
485 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
486 set_compound_order(page, order);
488 for (i = 1; i < nr_pages; i++) {
489 struct page *p = page + i;
490 set_page_count(p, 0);
491 p->mapping = TAIL_MAPPING;
492 set_compound_head(p, page);
494 atomic_set(compound_mapcount_ptr(page), -1);
497 #ifdef CONFIG_DEBUG_PAGEALLOC
498 unsigned int _debug_guardpage_minorder;
499 bool _debug_pagealloc_enabled __read_mostly
500 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
501 EXPORT_SYMBOL(_debug_pagealloc_enabled);
502 bool _debug_guardpage_enabled __read_mostly;
504 static int __init early_debug_pagealloc(char *buf)
509 if (strcmp(buf, "on") == 0)
510 _debug_pagealloc_enabled = true;
512 if (strcmp(buf, "off") == 0)
513 _debug_pagealloc_enabled = false;
517 early_param("debug_pagealloc", early_debug_pagealloc);
519 static bool need_debug_guardpage(void)
521 /* If we don't use debug_pagealloc, we don't need guard page */
522 if (!debug_pagealloc_enabled())
528 static void init_debug_guardpage(void)
530 if (!debug_pagealloc_enabled())
533 _debug_guardpage_enabled = true;
536 struct page_ext_operations debug_guardpage_ops = {
537 .need = need_debug_guardpage,
538 .init = init_debug_guardpage,
541 static int __init debug_guardpage_minorder_setup(char *buf)
545 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
546 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
549 _debug_guardpage_minorder = res;
550 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
553 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
555 static inline void set_page_guard(struct zone *zone, struct page *page,
556 unsigned int order, int migratetype)
558 struct page_ext *page_ext;
560 if (!debug_guardpage_enabled())
563 page_ext = lookup_page_ext(page);
564 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
566 INIT_LIST_HEAD(&page->lru);
567 set_page_private(page, order);
568 /* Guard pages are not available for any usage */
569 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
572 static inline void clear_page_guard(struct zone *zone, struct page *page,
573 unsigned int order, int migratetype)
575 struct page_ext *page_ext;
577 if (!debug_guardpage_enabled())
580 page_ext = lookup_page_ext(page);
581 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
583 set_page_private(page, 0);
584 if (!is_migrate_isolate(migratetype))
585 __mod_zone_freepage_state(zone, (1 << order), migratetype);
588 struct page_ext_operations debug_guardpage_ops = { NULL, };
589 static inline void set_page_guard(struct zone *zone, struct page *page,
590 unsigned int order, int migratetype) {}
591 static inline void clear_page_guard(struct zone *zone, struct page *page,
592 unsigned int order, int migratetype) {}
595 static inline void set_page_order(struct page *page, unsigned int order)
597 set_page_private(page, order);
598 __SetPageBuddy(page);
601 static inline void rmv_page_order(struct page *page)
603 __ClearPageBuddy(page);
604 set_page_private(page, 0);
608 * This function checks whether a page is free && is the buddy
609 * we can do coalesce a page and its buddy if
610 * (a) the buddy is not in a hole &&
611 * (b) the buddy is in the buddy system &&
612 * (c) a page and its buddy have the same order &&
613 * (d) a page and its buddy are in the same zone.
615 * For recording whether a page is in the buddy system, we set ->_mapcount
616 * PAGE_BUDDY_MAPCOUNT_VALUE.
617 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
618 * serialized by zone->lock.
620 * For recording page's order, we use page_private(page).
622 static inline int page_is_buddy(struct page *page, struct page *buddy,
625 if (!pfn_valid_within(page_to_pfn(buddy)))
628 if (page_is_guard(buddy) && page_order(buddy) == order) {
629 if (page_zone_id(page) != page_zone_id(buddy))
632 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
637 if (PageBuddy(buddy) && page_order(buddy) == order) {
639 * zone check is done late to avoid uselessly
640 * calculating zone/node ids for pages that could
643 if (page_zone_id(page) != page_zone_id(buddy))
646 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
654 * Freeing function for a buddy system allocator.
656 * The concept of a buddy system is to maintain direct-mapped table
657 * (containing bit values) for memory blocks of various "orders".
658 * The bottom level table contains the map for the smallest allocatable
659 * units of memory (here, pages), and each level above it describes
660 * pairs of units from the levels below, hence, "buddies".
661 * At a high level, all that happens here is marking the table entry
662 * at the bottom level available, and propagating the changes upward
663 * as necessary, plus some accounting needed to play nicely with other
664 * parts of the VM system.
665 * At each level, we keep a list of pages, which are heads of continuous
666 * free pages of length of (1 << order) and marked with _mapcount
667 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
669 * So when we are allocating or freeing one, we can derive the state of the
670 * other. That is, if we allocate a small block, and both were
671 * free, the remainder of the region must be split into blocks.
672 * If a block is freed, and its buddy is also free, then this
673 * triggers coalescing into a block of larger size.
678 static inline void __free_one_page(struct page *page,
680 struct zone *zone, unsigned int order,
683 unsigned long page_idx;
684 unsigned long combined_idx;
685 unsigned long uninitialized_var(buddy_idx);
687 unsigned int max_order = MAX_ORDER;
689 VM_BUG_ON(!zone_is_initialized(zone));
690 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
692 VM_BUG_ON(migratetype == -1);
693 if (is_migrate_isolate(migratetype)) {
695 * We restrict max order of merging to prevent merge
696 * between freepages on isolate pageblock and normal
697 * pageblock. Without this, pageblock isolation
698 * could cause incorrect freepage accounting.
700 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
702 __mod_zone_freepage_state(zone, 1 << order, migratetype);
705 page_idx = pfn & ((1 << max_order) - 1);
707 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
708 VM_BUG_ON_PAGE(bad_range(zone, page), page);
710 while (order < max_order - 1) {
711 buddy_idx = __find_buddy_index(page_idx, order);
712 buddy = page + (buddy_idx - page_idx);
713 if (!page_is_buddy(page, buddy, order))
716 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
717 * merge with it and move up one order.
719 if (page_is_guard(buddy)) {
720 clear_page_guard(zone, buddy, order, migratetype);
722 list_del(&buddy->lru);
723 zone->free_area[order].nr_free--;
724 rmv_page_order(buddy);
726 combined_idx = buddy_idx & page_idx;
727 page = page + (combined_idx - page_idx);
728 page_idx = combined_idx;
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(atomic_read(&page->_mapcount) != -1))
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_last_entry(list, 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 switch (page - head_page) {
890 /* the first tail page: ->mapping is compound_mapcount() */
891 if (unlikely(compound_mapcount(page))) {
892 bad_page(page, "nonzero compound_mapcount", 0);
898 * the second tail page: ->mapping is
899 * page_deferred_list().next -- ignore value.
903 if (page->mapping != TAIL_MAPPING) {
904 bad_page(page, "corrupted mapping in tail page", 0);
909 if (unlikely(!PageTail(page))) {
910 bad_page(page, "PageTail not set", 0);
913 if (unlikely(compound_head(page) != head_page)) {
914 bad_page(page, "compound_head not consistent", 0);
919 page->mapping = NULL;
920 clear_compound_head(page);
924 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
925 unsigned long zone, int nid)
927 set_page_links(page, zone, nid, pfn);
928 init_page_count(page);
929 page_mapcount_reset(page);
930 page_cpupid_reset_last(page);
932 INIT_LIST_HEAD(&page->lru);
933 #ifdef WANT_PAGE_VIRTUAL
934 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
935 if (!is_highmem_idx(zone))
936 set_page_address(page, __va(pfn << PAGE_SHIFT));
940 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
943 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
946 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
947 static void init_reserved_page(unsigned long pfn)
952 if (!early_page_uninitialised(pfn))
955 nid = early_pfn_to_nid(pfn);
956 pgdat = NODE_DATA(nid);
958 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
959 struct zone *zone = &pgdat->node_zones[zid];
961 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
964 __init_single_pfn(pfn, zid, nid);
967 static inline void init_reserved_page(unsigned long pfn)
970 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
973 * Initialised pages do not have PageReserved set. This function is
974 * called for each range allocated by the bootmem allocator and
975 * marks the pages PageReserved. The remaining valid pages are later
976 * sent to the buddy page allocator.
978 void __meminit reserve_bootmem_region(unsigned long start, unsigned long end)
980 unsigned long start_pfn = PFN_DOWN(start);
981 unsigned long end_pfn = PFN_UP(end);
983 for (; start_pfn < end_pfn; start_pfn++) {
984 if (pfn_valid(start_pfn)) {
985 struct page *page = pfn_to_page(start_pfn);
987 init_reserved_page(start_pfn);
989 /* Avoid false-positive PageTail() */
990 INIT_LIST_HEAD(&page->lru);
992 SetPageReserved(page);
997 static bool free_pages_prepare(struct page *page, unsigned int order)
999 bool compound = PageCompound(page);
1002 VM_BUG_ON_PAGE(PageTail(page), page);
1003 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1005 trace_mm_page_free(page, order);
1006 kmemcheck_free_shadow(page, order);
1007 kasan_free_pages(page, order);
1010 page->mapping = NULL;
1011 bad += free_pages_check(page);
1012 for (i = 1; i < (1 << order); i++) {
1014 bad += free_tail_pages_check(page, page + i);
1015 bad += free_pages_check(page + i);
1020 reset_page_owner(page, order);
1022 if (!PageHighMem(page)) {
1023 debug_check_no_locks_freed(page_address(page),
1024 PAGE_SIZE << order);
1025 debug_check_no_obj_freed(page_address(page),
1026 PAGE_SIZE << order);
1028 arch_free_page(page, order);
1029 kernel_poison_pages(page, 1 << order, 0);
1030 kernel_map_pages(page, 1 << order, 0);
1035 static void __free_pages_ok(struct page *page, unsigned int order)
1037 unsigned long flags;
1039 unsigned long pfn = page_to_pfn(page);
1041 if (!free_pages_prepare(page, order))
1044 migratetype = get_pfnblock_migratetype(page, pfn);
1045 local_irq_save(flags);
1046 __count_vm_events(PGFREE, 1 << order);
1047 free_one_page(page_zone(page), page, pfn, order, migratetype);
1048 local_irq_restore(flags);
1051 static void __init __free_pages_boot_core(struct page *page,
1052 unsigned long pfn, unsigned int order)
1054 unsigned int nr_pages = 1 << order;
1055 struct page *p = page;
1059 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1061 __ClearPageReserved(p);
1062 set_page_count(p, 0);
1064 __ClearPageReserved(p);
1065 set_page_count(p, 0);
1067 page_zone(page)->managed_pages += nr_pages;
1068 set_page_refcounted(page);
1069 __free_pages(page, order);
1072 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1073 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1075 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1077 int __meminit early_pfn_to_nid(unsigned long pfn)
1079 static DEFINE_SPINLOCK(early_pfn_lock);
1082 spin_lock(&early_pfn_lock);
1083 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1086 spin_unlock(&early_pfn_lock);
1092 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1093 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1094 struct mminit_pfnnid_cache *state)
1098 nid = __early_pfn_to_nid(pfn, state);
1099 if (nid >= 0 && nid != node)
1104 /* Only safe to use early in boot when initialisation is single-threaded */
1105 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1107 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1112 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1116 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1117 struct mminit_pfnnid_cache *state)
1124 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1127 if (early_page_uninitialised(pfn))
1129 return __free_pages_boot_core(page, pfn, order);
1133 * Check that the whole (or subset of) a pageblock given by the interval of
1134 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1135 * with the migration of free compaction scanner. The scanners then need to
1136 * use only pfn_valid_within() check for arches that allow holes within
1139 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1141 * It's possible on some configurations to have a setup like node0 node1 node0
1142 * i.e. it's possible that all pages within a zones range of pages do not
1143 * belong to a single zone. We assume that a border between node0 and node1
1144 * can occur within a single pageblock, but not a node0 node1 node0
1145 * interleaving within a single pageblock. It is therefore sufficient to check
1146 * the first and last page of a pageblock and avoid checking each individual
1147 * page in a pageblock.
1149 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1150 unsigned long end_pfn, struct zone *zone)
1152 struct page *start_page;
1153 struct page *end_page;
1155 /* end_pfn is one past the range we are checking */
1158 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1161 start_page = pfn_to_page(start_pfn);
1163 if (page_zone(start_page) != zone)
1166 end_page = pfn_to_page(end_pfn);
1168 /* This gives a shorter code than deriving page_zone(end_page) */
1169 if (page_zone_id(start_page) != page_zone_id(end_page))
1175 void set_zone_contiguous(struct zone *zone)
1177 unsigned long block_start_pfn = zone->zone_start_pfn;
1178 unsigned long block_end_pfn;
1180 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1181 for (; block_start_pfn < zone_end_pfn(zone);
1182 block_start_pfn = block_end_pfn,
1183 block_end_pfn += pageblock_nr_pages) {
1185 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1187 if (!__pageblock_pfn_to_page(block_start_pfn,
1188 block_end_pfn, zone))
1192 /* We confirm that there is no hole */
1193 zone->contiguous = true;
1196 void clear_zone_contiguous(struct zone *zone)
1198 zone->contiguous = false;
1201 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1202 static void __init deferred_free_range(struct page *page,
1203 unsigned long pfn, int nr_pages)
1210 /* Free a large naturally-aligned chunk if possible */
1211 if (nr_pages == MAX_ORDER_NR_PAGES &&
1212 (pfn & (MAX_ORDER_NR_PAGES-1)) == 0) {
1213 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1214 __free_pages_boot_core(page, pfn, MAX_ORDER-1);
1218 for (i = 0; i < nr_pages; i++, page++, pfn++)
1219 __free_pages_boot_core(page, pfn, 0);
1222 /* Completion tracking for deferred_init_memmap() threads */
1223 static atomic_t pgdat_init_n_undone __initdata;
1224 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1226 static inline void __init pgdat_init_report_one_done(void)
1228 if (atomic_dec_and_test(&pgdat_init_n_undone))
1229 complete(&pgdat_init_all_done_comp);
1232 /* Initialise remaining memory on a node */
1233 static int __init deferred_init_memmap(void *data)
1235 pg_data_t *pgdat = data;
1236 int nid = pgdat->node_id;
1237 struct mminit_pfnnid_cache nid_init_state = { };
1238 unsigned long start = jiffies;
1239 unsigned long nr_pages = 0;
1240 unsigned long walk_start, walk_end;
1243 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1244 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1246 if (first_init_pfn == ULONG_MAX) {
1247 pgdat_init_report_one_done();
1251 /* Bind memory initialisation thread to a local node if possible */
1252 if (!cpumask_empty(cpumask))
1253 set_cpus_allowed_ptr(current, cpumask);
1255 /* Sanity check boundaries */
1256 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1257 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1258 pgdat->first_deferred_pfn = ULONG_MAX;
1260 /* Only the highest zone is deferred so find it */
1261 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1262 zone = pgdat->node_zones + zid;
1263 if (first_init_pfn < zone_end_pfn(zone))
1267 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1268 unsigned long pfn, end_pfn;
1269 struct page *page = NULL;
1270 struct page *free_base_page = NULL;
1271 unsigned long free_base_pfn = 0;
1274 end_pfn = min(walk_end, zone_end_pfn(zone));
1275 pfn = first_init_pfn;
1276 if (pfn < walk_start)
1278 if (pfn < zone->zone_start_pfn)
1279 pfn = zone->zone_start_pfn;
1281 for (; pfn < end_pfn; pfn++) {
1282 if (!pfn_valid_within(pfn))
1286 * Ensure pfn_valid is checked every
1287 * MAX_ORDER_NR_PAGES for memory holes
1289 if ((pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
1290 if (!pfn_valid(pfn)) {
1296 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1301 /* Minimise pfn page lookups and scheduler checks */
1302 if (page && (pfn & (MAX_ORDER_NR_PAGES - 1)) != 0) {
1305 nr_pages += nr_to_free;
1306 deferred_free_range(free_base_page,
1307 free_base_pfn, nr_to_free);
1308 free_base_page = NULL;
1309 free_base_pfn = nr_to_free = 0;
1311 page = pfn_to_page(pfn);
1316 VM_BUG_ON(page_zone(page) != zone);
1320 __init_single_page(page, pfn, zid, nid);
1321 if (!free_base_page) {
1322 free_base_page = page;
1323 free_base_pfn = pfn;
1328 /* Where possible, batch up pages for a single free */
1331 /* Free the current block of pages to allocator */
1332 nr_pages += nr_to_free;
1333 deferred_free_range(free_base_page, free_base_pfn,
1335 free_base_page = NULL;
1336 free_base_pfn = nr_to_free = 0;
1339 first_init_pfn = max(end_pfn, first_init_pfn);
1342 /* Sanity check that the next zone really is unpopulated */
1343 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1345 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1346 jiffies_to_msecs(jiffies - start));
1348 pgdat_init_report_one_done();
1351 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1353 void __init page_alloc_init_late(void)
1357 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1360 /* There will be num_node_state(N_MEMORY) threads */
1361 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1362 for_each_node_state(nid, N_MEMORY) {
1363 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1366 /* Block until all are initialised */
1367 wait_for_completion(&pgdat_init_all_done_comp);
1369 /* Reinit limits that are based on free pages after the kernel is up */
1370 files_maxfiles_init();
1373 for_each_populated_zone(zone)
1374 set_zone_contiguous(zone);
1378 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1379 void __init init_cma_reserved_pageblock(struct page *page)
1381 unsigned i = pageblock_nr_pages;
1382 struct page *p = page;
1385 __ClearPageReserved(p);
1386 set_page_count(p, 0);
1389 set_pageblock_migratetype(page, MIGRATE_CMA);
1391 if (pageblock_order >= MAX_ORDER) {
1392 i = pageblock_nr_pages;
1395 set_page_refcounted(p);
1396 __free_pages(p, MAX_ORDER - 1);
1397 p += MAX_ORDER_NR_PAGES;
1398 } while (i -= MAX_ORDER_NR_PAGES);
1400 set_page_refcounted(page);
1401 __free_pages(page, pageblock_order);
1404 adjust_managed_page_count(page, pageblock_nr_pages);
1409 * The order of subdivision here is critical for the IO subsystem.
1410 * Please do not alter this order without good reasons and regression
1411 * testing. Specifically, as large blocks of memory are subdivided,
1412 * the order in which smaller blocks are delivered depends on the order
1413 * they're subdivided in this function. This is the primary factor
1414 * influencing the order in which pages are delivered to the IO
1415 * subsystem according to empirical testing, and this is also justified
1416 * by considering the behavior of a buddy system containing a single
1417 * large block of memory acted on by a series of small allocations.
1418 * This behavior is a critical factor in sglist merging's success.
1422 static inline void expand(struct zone *zone, struct page *page,
1423 int low, int high, struct free_area *area,
1426 unsigned long size = 1 << high;
1428 while (high > low) {
1432 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1434 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
1435 debug_guardpage_enabled() &&
1436 high < debug_guardpage_minorder()) {
1438 * Mark as guard pages (or page), that will allow to
1439 * merge back to allocator when buddy will be freed.
1440 * Corresponding page table entries will not be touched,
1441 * pages will stay not present in virtual address space
1443 set_page_guard(zone, &page[size], high, migratetype);
1446 list_add(&page[size].lru, &area->free_list[migratetype]);
1448 set_page_order(&page[size], high);
1453 * This page is about to be returned from the page allocator
1455 static inline int check_new_page(struct page *page)
1457 const char *bad_reason = NULL;
1458 unsigned long bad_flags = 0;
1460 if (unlikely(atomic_read(&page->_mapcount) != -1))
1461 bad_reason = "nonzero mapcount";
1462 if (unlikely(page->mapping != NULL))
1463 bad_reason = "non-NULL mapping";
1464 if (unlikely(atomic_read(&page->_count) != 0))
1465 bad_reason = "nonzero _count";
1466 if (unlikely(page->flags & __PG_HWPOISON)) {
1467 bad_reason = "HWPoisoned (hardware-corrupted)";
1468 bad_flags = __PG_HWPOISON;
1470 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1471 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1472 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1475 if (unlikely(page->mem_cgroup))
1476 bad_reason = "page still charged to cgroup";
1478 if (unlikely(bad_reason)) {
1479 bad_page(page, bad_reason, bad_flags);
1485 static inline bool free_pages_prezeroed(bool poisoned)
1487 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1488 page_poisoning_enabled() && poisoned;
1491 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1495 bool poisoned = true;
1497 for (i = 0; i < (1 << order); i++) {
1498 struct page *p = page + i;
1499 if (unlikely(check_new_page(p)))
1502 poisoned &= page_is_poisoned(p);
1505 set_page_private(page, 0);
1506 set_page_refcounted(page);
1508 arch_alloc_page(page, order);
1509 kernel_map_pages(page, 1 << order, 1);
1510 kernel_poison_pages(page, 1 << order, 1);
1511 kasan_alloc_pages(page, order);
1513 if (!free_pages_prezeroed(poisoned) && (gfp_flags & __GFP_ZERO))
1514 for (i = 0; i < (1 << order); i++)
1515 clear_highpage(page + i);
1517 if (order && (gfp_flags & __GFP_COMP))
1518 prep_compound_page(page, order);
1520 set_page_owner(page, order, gfp_flags);
1523 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1524 * allocate the page. The expectation is that the caller is taking
1525 * steps that will free more memory. The caller should avoid the page
1526 * being used for !PFMEMALLOC purposes.
1528 if (alloc_flags & ALLOC_NO_WATERMARKS)
1529 set_page_pfmemalloc(page);
1531 clear_page_pfmemalloc(page);
1537 * Go through the free lists for the given migratetype and remove
1538 * the smallest available page from the freelists
1541 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1544 unsigned int current_order;
1545 struct free_area *area;
1548 /* Find a page of the appropriate size in the preferred list */
1549 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1550 area = &(zone->free_area[current_order]);
1551 page = list_first_entry_or_null(&area->free_list[migratetype],
1555 list_del(&page->lru);
1556 rmv_page_order(page);
1558 expand(zone, page, order, current_order, area, migratetype);
1559 set_pcppage_migratetype(page, migratetype);
1568 * This array describes the order lists are fallen back to when
1569 * the free lists for the desirable migrate type are depleted
1571 static int fallbacks[MIGRATE_TYPES][4] = {
1572 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1573 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1574 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1576 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1578 #ifdef CONFIG_MEMORY_ISOLATION
1579 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1584 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1587 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1590 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1591 unsigned int order) { return NULL; }
1595 * Move the free pages in a range to the free lists of the requested type.
1596 * Note that start_page and end_pages are not aligned on a pageblock
1597 * boundary. If alignment is required, use move_freepages_block()
1599 int move_freepages(struct zone *zone,
1600 struct page *start_page, struct page *end_page,
1605 int pages_moved = 0;
1607 #ifndef CONFIG_HOLES_IN_ZONE
1609 * page_zone is not safe to call in this context when
1610 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1611 * anyway as we check zone boundaries in move_freepages_block().
1612 * Remove at a later date when no bug reports exist related to
1613 * grouping pages by mobility
1615 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1618 for (page = start_page; page <= end_page;) {
1619 /* Make sure we are not inadvertently changing nodes */
1620 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1622 if (!pfn_valid_within(page_to_pfn(page))) {
1627 if (!PageBuddy(page)) {
1632 order = page_order(page);
1633 list_move(&page->lru,
1634 &zone->free_area[order].free_list[migratetype]);
1636 pages_moved += 1 << order;
1642 int move_freepages_block(struct zone *zone, struct page *page,
1645 unsigned long start_pfn, end_pfn;
1646 struct page *start_page, *end_page;
1648 start_pfn = page_to_pfn(page);
1649 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1650 start_page = pfn_to_page(start_pfn);
1651 end_page = start_page + pageblock_nr_pages - 1;
1652 end_pfn = start_pfn + pageblock_nr_pages - 1;
1654 /* Do not cross zone boundaries */
1655 if (!zone_spans_pfn(zone, start_pfn))
1657 if (!zone_spans_pfn(zone, end_pfn))
1660 return move_freepages(zone, start_page, end_page, migratetype);
1663 static void change_pageblock_range(struct page *pageblock_page,
1664 int start_order, int migratetype)
1666 int nr_pageblocks = 1 << (start_order - pageblock_order);
1668 while (nr_pageblocks--) {
1669 set_pageblock_migratetype(pageblock_page, migratetype);
1670 pageblock_page += pageblock_nr_pages;
1675 * When we are falling back to another migratetype during allocation, try to
1676 * steal extra free pages from the same pageblocks to satisfy further
1677 * allocations, instead of polluting multiple pageblocks.
1679 * If we are stealing a relatively large buddy page, it is likely there will
1680 * be more free pages in the pageblock, so try to steal them all. For
1681 * reclaimable and unmovable allocations, we steal regardless of page size,
1682 * as fragmentation caused by those allocations polluting movable pageblocks
1683 * is worse than movable allocations stealing from unmovable and reclaimable
1686 static bool can_steal_fallback(unsigned int order, int start_mt)
1689 * Leaving this order check is intended, although there is
1690 * relaxed order check in next check. The reason is that
1691 * we can actually steal whole pageblock if this condition met,
1692 * but, below check doesn't guarantee it and that is just heuristic
1693 * so could be changed anytime.
1695 if (order >= pageblock_order)
1698 if (order >= pageblock_order / 2 ||
1699 start_mt == MIGRATE_RECLAIMABLE ||
1700 start_mt == MIGRATE_UNMOVABLE ||
1701 page_group_by_mobility_disabled)
1708 * This function implements actual steal behaviour. If order is large enough,
1709 * we can steal whole pageblock. If not, we first move freepages in this
1710 * pageblock and check whether half of pages are moved or not. If half of
1711 * pages are moved, we can change migratetype of pageblock and permanently
1712 * use it's pages as requested migratetype in the future.
1714 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1717 unsigned int current_order = page_order(page);
1720 /* Take ownership for orders >= pageblock_order */
1721 if (current_order >= pageblock_order) {
1722 change_pageblock_range(page, current_order, start_type);
1726 pages = move_freepages_block(zone, page, start_type);
1728 /* Claim the whole block if over half of it is free */
1729 if (pages >= (1 << (pageblock_order-1)) ||
1730 page_group_by_mobility_disabled)
1731 set_pageblock_migratetype(page, start_type);
1735 * Check whether there is a suitable fallback freepage with requested order.
1736 * If only_stealable is true, this function returns fallback_mt only if
1737 * we can steal other freepages all together. This would help to reduce
1738 * fragmentation due to mixed migratetype pages in one pageblock.
1740 int find_suitable_fallback(struct free_area *area, unsigned int order,
1741 int migratetype, bool only_stealable, bool *can_steal)
1746 if (area->nr_free == 0)
1751 fallback_mt = fallbacks[migratetype][i];
1752 if (fallback_mt == MIGRATE_TYPES)
1755 if (list_empty(&area->free_list[fallback_mt]))
1758 if (can_steal_fallback(order, migratetype))
1761 if (!only_stealable)
1772 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1773 * there are no empty page blocks that contain a page with a suitable order
1775 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
1776 unsigned int alloc_order)
1779 unsigned long max_managed, flags;
1782 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1783 * Check is race-prone but harmless.
1785 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
1786 if (zone->nr_reserved_highatomic >= max_managed)
1789 spin_lock_irqsave(&zone->lock, flags);
1791 /* Recheck the nr_reserved_highatomic limit under the lock */
1792 if (zone->nr_reserved_highatomic >= max_managed)
1796 mt = get_pageblock_migratetype(page);
1797 if (mt != MIGRATE_HIGHATOMIC &&
1798 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
1799 zone->nr_reserved_highatomic += pageblock_nr_pages;
1800 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
1801 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
1805 spin_unlock_irqrestore(&zone->lock, flags);
1809 * Used when an allocation is about to fail under memory pressure. This
1810 * potentially hurts the reliability of high-order allocations when under
1811 * intense memory pressure but failed atomic allocations should be easier
1812 * to recover from than an OOM.
1814 static void unreserve_highatomic_pageblock(const struct alloc_context *ac)
1816 struct zonelist *zonelist = ac->zonelist;
1817 unsigned long flags;
1823 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
1825 /* Preserve at least one pageblock */
1826 if (zone->nr_reserved_highatomic <= pageblock_nr_pages)
1829 spin_lock_irqsave(&zone->lock, flags);
1830 for (order = 0; order < MAX_ORDER; order++) {
1831 struct free_area *area = &(zone->free_area[order]);
1833 page = list_first_entry_or_null(
1834 &area->free_list[MIGRATE_HIGHATOMIC],
1840 * It should never happen but changes to locking could
1841 * inadvertently allow a per-cpu drain to add pages
1842 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1843 * and watch for underflows.
1845 zone->nr_reserved_highatomic -= min(pageblock_nr_pages,
1846 zone->nr_reserved_highatomic);
1849 * Convert to ac->migratetype and avoid the normal
1850 * pageblock stealing heuristics. Minimally, the caller
1851 * is doing the work and needs the pages. More
1852 * importantly, if the block was always converted to
1853 * MIGRATE_UNMOVABLE or another type then the number
1854 * of pageblocks that cannot be completely freed
1857 set_pageblock_migratetype(page, ac->migratetype);
1858 move_freepages_block(zone, page, ac->migratetype);
1859 spin_unlock_irqrestore(&zone->lock, flags);
1862 spin_unlock_irqrestore(&zone->lock, flags);
1866 /* Remove an element from the buddy allocator from the fallback list */
1867 static inline struct page *
1868 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1870 struct free_area *area;
1871 unsigned int current_order;
1876 /* Find the largest possible block of pages in the other list */
1877 for (current_order = MAX_ORDER-1;
1878 current_order >= order && current_order <= MAX_ORDER-1;
1880 area = &(zone->free_area[current_order]);
1881 fallback_mt = find_suitable_fallback(area, current_order,
1882 start_migratetype, false, &can_steal);
1883 if (fallback_mt == -1)
1886 page = list_first_entry(&area->free_list[fallback_mt],
1889 steal_suitable_fallback(zone, page, start_migratetype);
1891 /* Remove the page from the freelists */
1893 list_del(&page->lru);
1894 rmv_page_order(page);
1896 expand(zone, page, order, current_order, area,
1899 * The pcppage_migratetype may differ from pageblock's
1900 * migratetype depending on the decisions in
1901 * find_suitable_fallback(). This is OK as long as it does not
1902 * differ for MIGRATE_CMA pageblocks. Those can be used as
1903 * fallback only via special __rmqueue_cma_fallback() function
1905 set_pcppage_migratetype(page, start_migratetype);
1907 trace_mm_page_alloc_extfrag(page, order, current_order,
1908 start_migratetype, fallback_mt);
1917 * Do the hard work of removing an element from the buddy allocator.
1918 * Call me with the zone->lock already held.
1920 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1925 page = __rmqueue_smallest(zone, order, migratetype);
1926 if (unlikely(!page)) {
1927 if (migratetype == MIGRATE_MOVABLE)
1928 page = __rmqueue_cma_fallback(zone, order);
1931 page = __rmqueue_fallback(zone, order, migratetype);
1934 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1939 * Obtain a specified number of elements from the buddy allocator, all under
1940 * a single hold of the lock, for efficiency. Add them to the supplied list.
1941 * Returns the number of new pages which were placed at *list.
1943 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1944 unsigned long count, struct list_head *list,
1945 int migratetype, bool cold)
1949 spin_lock(&zone->lock);
1950 for (i = 0; i < count; ++i) {
1951 struct page *page = __rmqueue(zone, order, migratetype);
1952 if (unlikely(page == NULL))
1956 * Split buddy pages returned by expand() are received here
1957 * in physical page order. The page is added to the callers and
1958 * list and the list head then moves forward. From the callers
1959 * perspective, the linked list is ordered by page number in
1960 * some conditions. This is useful for IO devices that can
1961 * merge IO requests if the physical pages are ordered
1965 list_add(&page->lru, list);
1967 list_add_tail(&page->lru, list);
1969 if (is_migrate_cma(get_pcppage_migratetype(page)))
1970 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1973 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1974 spin_unlock(&zone->lock);
1980 * Called from the vmstat counter updater to drain pagesets of this
1981 * currently executing processor on remote nodes after they have
1984 * Note that this function must be called with the thread pinned to
1985 * a single processor.
1987 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1989 unsigned long flags;
1990 int to_drain, batch;
1992 local_irq_save(flags);
1993 batch = READ_ONCE(pcp->batch);
1994 to_drain = min(pcp->count, batch);
1996 free_pcppages_bulk(zone, to_drain, pcp);
1997 pcp->count -= to_drain;
1999 local_irq_restore(flags);
2004 * Drain pcplists of the indicated processor and zone.
2006 * The processor must either be the current processor and the
2007 * thread pinned to the current processor or a processor that
2010 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2012 unsigned long flags;
2013 struct per_cpu_pageset *pset;
2014 struct per_cpu_pages *pcp;
2016 local_irq_save(flags);
2017 pset = per_cpu_ptr(zone->pageset, cpu);
2021 free_pcppages_bulk(zone, pcp->count, pcp);
2024 local_irq_restore(flags);
2028 * Drain pcplists of all zones on the indicated processor.
2030 * The processor must either be the current processor and the
2031 * thread pinned to the current processor or a processor that
2034 static void drain_pages(unsigned int cpu)
2038 for_each_populated_zone(zone) {
2039 drain_pages_zone(cpu, zone);
2044 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2046 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2047 * the single zone's pages.
2049 void drain_local_pages(struct zone *zone)
2051 int cpu = smp_processor_id();
2054 drain_pages_zone(cpu, zone);
2060 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2062 * When zone parameter is non-NULL, spill just the single zone's pages.
2064 * Note that this code is protected against sending an IPI to an offline
2065 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
2066 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
2067 * nothing keeps CPUs from showing up after we populated the cpumask and
2068 * before the call to on_each_cpu_mask().
2070 void drain_all_pages(struct zone *zone)
2075 * Allocate in the BSS so we wont require allocation in
2076 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2078 static cpumask_t cpus_with_pcps;
2081 * We don't care about racing with CPU hotplug event
2082 * as offline notification will cause the notified
2083 * cpu to drain that CPU pcps and on_each_cpu_mask
2084 * disables preemption as part of its processing
2086 for_each_online_cpu(cpu) {
2087 struct per_cpu_pageset *pcp;
2089 bool has_pcps = false;
2092 pcp = per_cpu_ptr(zone->pageset, cpu);
2096 for_each_populated_zone(z) {
2097 pcp = per_cpu_ptr(z->pageset, cpu);
2098 if (pcp->pcp.count) {
2106 cpumask_set_cpu(cpu, &cpus_with_pcps);
2108 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2110 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
2114 #ifdef CONFIG_HIBERNATION
2116 void mark_free_pages(struct zone *zone)
2118 unsigned long pfn, max_zone_pfn;
2119 unsigned long flags;
2120 unsigned int order, t;
2123 if (zone_is_empty(zone))
2126 spin_lock_irqsave(&zone->lock, flags);
2128 max_zone_pfn = zone_end_pfn(zone);
2129 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2130 if (pfn_valid(pfn)) {
2131 page = pfn_to_page(pfn);
2132 if (!swsusp_page_is_forbidden(page))
2133 swsusp_unset_page_free(page);
2136 for_each_migratetype_order(order, t) {
2137 list_for_each_entry(page,
2138 &zone->free_area[order].free_list[t], lru) {
2141 pfn = page_to_pfn(page);
2142 for (i = 0; i < (1UL << order); i++)
2143 swsusp_set_page_free(pfn_to_page(pfn + i));
2146 spin_unlock_irqrestore(&zone->lock, flags);
2148 #endif /* CONFIG_PM */
2151 * Free a 0-order page
2152 * cold == true ? free a cold page : free a hot page
2154 void free_hot_cold_page(struct page *page, bool cold)
2156 struct zone *zone = page_zone(page);
2157 struct per_cpu_pages *pcp;
2158 unsigned long flags;
2159 unsigned long pfn = page_to_pfn(page);
2162 if (!free_pages_prepare(page, 0))
2165 migratetype = get_pfnblock_migratetype(page, pfn);
2166 set_pcppage_migratetype(page, migratetype);
2167 local_irq_save(flags);
2168 __count_vm_event(PGFREE);
2171 * We only track unmovable, reclaimable and movable on pcp lists.
2172 * Free ISOLATE pages back to the allocator because they are being
2173 * offlined but treat RESERVE as movable pages so we can get those
2174 * areas back if necessary. Otherwise, we may have to free
2175 * excessively into the page allocator
2177 if (migratetype >= MIGRATE_PCPTYPES) {
2178 if (unlikely(is_migrate_isolate(migratetype))) {
2179 free_one_page(zone, page, pfn, 0, migratetype);
2182 migratetype = MIGRATE_MOVABLE;
2185 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2187 list_add(&page->lru, &pcp->lists[migratetype]);
2189 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2191 if (pcp->count >= pcp->high) {
2192 unsigned long batch = READ_ONCE(pcp->batch);
2193 free_pcppages_bulk(zone, batch, pcp);
2194 pcp->count -= batch;
2198 local_irq_restore(flags);
2202 * Free a list of 0-order pages
2204 void free_hot_cold_page_list(struct list_head *list, bool cold)
2206 struct page *page, *next;
2208 list_for_each_entry_safe(page, next, list, lru) {
2209 trace_mm_page_free_batched(page, cold);
2210 free_hot_cold_page(page, cold);
2215 * split_page takes a non-compound higher-order page, and splits it into
2216 * n (1<<order) sub-pages: page[0..n]
2217 * Each sub-page must be freed individually.
2219 * Note: this is probably too low level an operation for use in drivers.
2220 * Please consult with lkml before using this in your driver.
2222 void split_page(struct page *page, unsigned int order)
2227 VM_BUG_ON_PAGE(PageCompound(page), page);
2228 VM_BUG_ON_PAGE(!page_count(page), page);
2230 #ifdef CONFIG_KMEMCHECK
2232 * Split shadow pages too, because free(page[0]) would
2233 * otherwise free the whole shadow.
2235 if (kmemcheck_page_is_tracked(page))
2236 split_page(virt_to_page(page[0].shadow), order);
2239 gfp_mask = get_page_owner_gfp(page);
2240 set_page_owner(page, 0, gfp_mask);
2241 for (i = 1; i < (1 << order); i++) {
2242 set_page_refcounted(page + i);
2243 set_page_owner(page + i, 0, gfp_mask);
2246 EXPORT_SYMBOL_GPL(split_page);
2248 int __isolate_free_page(struct page *page, unsigned int order)
2250 unsigned long watermark;
2254 BUG_ON(!PageBuddy(page));
2256 zone = page_zone(page);
2257 mt = get_pageblock_migratetype(page);
2259 if (!is_migrate_isolate(mt)) {
2260 /* Obey watermarks as if the page was being allocated */
2261 watermark = low_wmark_pages(zone) + (1 << order);
2262 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
2265 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2268 /* Remove page from free list */
2269 list_del(&page->lru);
2270 zone->free_area[order].nr_free--;
2271 rmv_page_order(page);
2273 set_page_owner(page, order, __GFP_MOVABLE);
2275 /* Set the pageblock if the isolated page is at least a pageblock */
2276 if (order >= pageblock_order - 1) {
2277 struct page *endpage = page + (1 << order) - 1;
2278 for (; page < endpage; page += pageblock_nr_pages) {
2279 int mt = get_pageblock_migratetype(page);
2280 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
2281 set_pageblock_migratetype(page,
2287 return 1UL << order;
2291 * Similar to split_page except the page is already free. As this is only
2292 * being used for migration, the migratetype of the block also changes.
2293 * As this is called with interrupts disabled, the caller is responsible
2294 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2297 * Note: this is probably too low level an operation for use in drivers.
2298 * Please consult with lkml before using this in your driver.
2300 int split_free_page(struct page *page)
2305 order = page_order(page);
2307 nr_pages = __isolate_free_page(page, order);
2311 /* Split into individual pages */
2312 set_page_refcounted(page);
2313 split_page(page, order);
2318 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2321 struct page *buffered_rmqueue(struct zone *preferred_zone,
2322 struct zone *zone, unsigned int order,
2323 gfp_t gfp_flags, int alloc_flags, int migratetype)
2325 unsigned long flags;
2327 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2329 if (likely(order == 0)) {
2330 struct per_cpu_pages *pcp;
2331 struct list_head *list;
2333 local_irq_save(flags);
2334 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2335 list = &pcp->lists[migratetype];
2336 if (list_empty(list)) {
2337 pcp->count += rmqueue_bulk(zone, 0,
2340 if (unlikely(list_empty(list)))
2345 page = list_last_entry(list, struct page, lru);
2347 page = list_first_entry(list, struct page, lru);
2349 list_del(&page->lru);
2352 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
2354 * __GFP_NOFAIL is not to be used in new code.
2356 * All __GFP_NOFAIL callers should be fixed so that they
2357 * properly detect and handle allocation failures.
2359 * We most definitely don't want callers attempting to
2360 * allocate greater than order-1 page units with
2363 WARN_ON_ONCE(order > 1);
2365 spin_lock_irqsave(&zone->lock, flags);
2368 if (alloc_flags & ALLOC_HARDER) {
2369 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2371 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2374 page = __rmqueue(zone, order, migratetype);
2375 spin_unlock(&zone->lock);
2378 __mod_zone_freepage_state(zone, -(1 << order),
2379 get_pcppage_migratetype(page));
2382 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
2383 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
2384 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
2385 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2387 __count_zone_vm_events(PGALLOC, zone, 1 << order);
2388 zone_statistics(preferred_zone, zone, gfp_flags);
2389 local_irq_restore(flags);
2391 VM_BUG_ON_PAGE(bad_range(zone, page), page);
2395 local_irq_restore(flags);
2399 #ifdef CONFIG_FAIL_PAGE_ALLOC
2402 struct fault_attr attr;
2404 bool ignore_gfp_highmem;
2405 bool ignore_gfp_reclaim;
2407 } fail_page_alloc = {
2408 .attr = FAULT_ATTR_INITIALIZER,
2409 .ignore_gfp_reclaim = true,
2410 .ignore_gfp_highmem = true,
2414 static int __init setup_fail_page_alloc(char *str)
2416 return setup_fault_attr(&fail_page_alloc.attr, str);
2418 __setup("fail_page_alloc=", setup_fail_page_alloc);
2420 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2422 if (order < fail_page_alloc.min_order)
2424 if (gfp_mask & __GFP_NOFAIL)
2426 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2428 if (fail_page_alloc.ignore_gfp_reclaim &&
2429 (gfp_mask & __GFP_DIRECT_RECLAIM))
2432 return should_fail(&fail_page_alloc.attr, 1 << order);
2435 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2437 static int __init fail_page_alloc_debugfs(void)
2439 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2442 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2443 &fail_page_alloc.attr);
2445 return PTR_ERR(dir);
2447 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2448 &fail_page_alloc.ignore_gfp_reclaim))
2450 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2451 &fail_page_alloc.ignore_gfp_highmem))
2453 if (!debugfs_create_u32("min-order", mode, dir,
2454 &fail_page_alloc.min_order))
2459 debugfs_remove_recursive(dir);
2464 late_initcall(fail_page_alloc_debugfs);
2466 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2468 #else /* CONFIG_FAIL_PAGE_ALLOC */
2470 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2475 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2478 * Return true if free base pages are above 'mark'. For high-order checks it
2479 * will return true of the order-0 watermark is reached and there is at least
2480 * one free page of a suitable size. Checking now avoids taking the zone lock
2481 * to check in the allocation paths if no pages are free.
2483 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
2484 unsigned long mark, int classzone_idx, int alloc_flags,
2489 const int alloc_harder = (alloc_flags & ALLOC_HARDER);
2491 /* free_pages may go negative - that's OK */
2492 free_pages -= (1 << order) - 1;
2494 if (alloc_flags & ALLOC_HIGH)
2498 * If the caller does not have rights to ALLOC_HARDER then subtract
2499 * the high-atomic reserves. This will over-estimate the size of the
2500 * atomic reserve but it avoids a search.
2502 if (likely(!alloc_harder))
2503 free_pages -= z->nr_reserved_highatomic;
2508 /* If allocation can't use CMA areas don't use free CMA pages */
2509 if (!(alloc_flags & ALLOC_CMA))
2510 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2514 * Check watermarks for an order-0 allocation request. If these
2515 * are not met, then a high-order request also cannot go ahead
2516 * even if a suitable page happened to be free.
2518 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2521 /* If this is an order-0 request then the watermark is fine */
2525 /* For a high-order request, check at least one suitable page is free */
2526 for (o = order; o < MAX_ORDER; o++) {
2527 struct free_area *area = &z->free_area[o];
2536 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2537 if (!list_empty(&area->free_list[mt]))
2542 if ((alloc_flags & ALLOC_CMA) &&
2543 !list_empty(&area->free_list[MIGRATE_CMA])) {
2551 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2552 int classzone_idx, int alloc_flags)
2554 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2555 zone_page_state(z, NR_FREE_PAGES));
2558 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2559 unsigned long mark, int classzone_idx)
2561 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2563 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2564 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2566 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2571 static bool zone_local(struct zone *local_zone, struct zone *zone)
2573 return local_zone->node == zone->node;
2576 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2578 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2581 #else /* CONFIG_NUMA */
2582 static bool zone_local(struct zone *local_zone, struct zone *zone)
2587 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2591 #endif /* CONFIG_NUMA */
2593 static void reset_alloc_batches(struct zone *preferred_zone)
2595 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2598 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2599 high_wmark_pages(zone) - low_wmark_pages(zone) -
2600 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2601 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2602 } while (zone++ != preferred_zone);
2606 * get_page_from_freelist goes through the zonelist trying to allocate
2609 static struct page *
2610 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2611 const struct alloc_context *ac)
2613 struct zonelist *zonelist = ac->zonelist;
2615 struct page *page = NULL;
2617 int nr_fair_skipped = 0;
2618 bool zonelist_rescan;
2621 zonelist_rescan = false;
2624 * Scan zonelist, looking for a zone with enough free.
2625 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2627 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2631 if (cpusets_enabled() &&
2632 (alloc_flags & ALLOC_CPUSET) &&
2633 !cpuset_zone_allowed(zone, gfp_mask))
2636 * Distribute pages in proportion to the individual
2637 * zone size to ensure fair page aging. The zone a
2638 * page was allocated in should have no effect on the
2639 * time the page has in memory before being reclaimed.
2641 if (alloc_flags & ALLOC_FAIR) {
2642 if (!zone_local(ac->preferred_zone, zone))
2644 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2650 * When allocating a page cache page for writing, we
2651 * want to get it from a zone that is within its dirty
2652 * limit, such that no single zone holds more than its
2653 * proportional share of globally allowed dirty pages.
2654 * The dirty limits take into account the zone's
2655 * lowmem reserves and high watermark so that kswapd
2656 * should be able to balance it without having to
2657 * write pages from its LRU list.
2659 * This may look like it could increase pressure on
2660 * lower zones by failing allocations in higher zones
2661 * before they are full. But the pages that do spill
2662 * over are limited as the lower zones are protected
2663 * by this very same mechanism. It should not become
2664 * a practical burden to them.
2666 * XXX: For now, allow allocations to potentially
2667 * exceed the per-zone dirty limit in the slowpath
2668 * (spread_dirty_pages unset) before going into reclaim,
2669 * which is important when on a NUMA setup the allowed
2670 * zones are together not big enough to reach the
2671 * global limit. The proper fix for these situations
2672 * will require awareness of zones in the
2673 * dirty-throttling and the flusher threads.
2675 if (ac->spread_dirty_pages && !zone_dirty_ok(zone))
2678 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2679 if (!zone_watermark_ok(zone, order, mark,
2680 ac->classzone_idx, alloc_flags)) {
2683 /* Checked here to keep the fast path fast */
2684 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2685 if (alloc_flags & ALLOC_NO_WATERMARKS)
2688 if (zone_reclaim_mode == 0 ||
2689 !zone_allows_reclaim(ac->preferred_zone, zone))
2692 ret = zone_reclaim(zone, gfp_mask, order);
2694 case ZONE_RECLAIM_NOSCAN:
2697 case ZONE_RECLAIM_FULL:
2698 /* scanned but unreclaimable */
2701 /* did we reclaim enough */
2702 if (zone_watermark_ok(zone, order, mark,
2703 ac->classzone_idx, alloc_flags))
2711 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2712 gfp_mask, alloc_flags, ac->migratetype);
2714 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2718 * If this is a high-order atomic allocation then check
2719 * if the pageblock should be reserved for the future
2721 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
2722 reserve_highatomic_pageblock(page, zone, order);
2729 * The first pass makes sure allocations are spread fairly within the
2730 * local node. However, the local node might have free pages left
2731 * after the fairness batches are exhausted, and remote zones haven't
2732 * even been considered yet. Try once more without fairness, and
2733 * include remote zones now, before entering the slowpath and waking
2734 * kswapd: prefer spilling to a remote zone over swapping locally.
2736 if (alloc_flags & ALLOC_FAIR) {
2737 alloc_flags &= ~ALLOC_FAIR;
2738 if (nr_fair_skipped) {
2739 zonelist_rescan = true;
2740 reset_alloc_batches(ac->preferred_zone);
2742 if (nr_online_nodes > 1)
2743 zonelist_rescan = true;
2746 if (zonelist_rescan)
2753 * Large machines with many possible nodes should not always dump per-node
2754 * meminfo in irq context.
2756 static inline bool should_suppress_show_mem(void)
2761 ret = in_interrupt();
2766 static DEFINE_RATELIMIT_STATE(nopage_rs,
2767 DEFAULT_RATELIMIT_INTERVAL,
2768 DEFAULT_RATELIMIT_BURST);
2770 void warn_alloc_failed(gfp_t gfp_mask, unsigned int order, const char *fmt, ...)
2772 unsigned int filter = SHOW_MEM_FILTER_NODES;
2774 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2775 debug_guardpage_minorder() > 0)
2779 * This documents exceptions given to allocations in certain
2780 * contexts that are allowed to allocate outside current's set
2783 if (!(gfp_mask & __GFP_NOMEMALLOC))
2784 if (test_thread_flag(TIF_MEMDIE) ||
2785 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2786 filter &= ~SHOW_MEM_FILTER_NODES;
2787 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
2788 filter &= ~SHOW_MEM_FILTER_NODES;
2791 struct va_format vaf;
2794 va_start(args, fmt);
2799 pr_warn("%pV", &vaf);
2804 pr_warn("%s: page allocation failure: order:%u, mode:%#x(%pGg)\n",
2805 current->comm, order, gfp_mask, &gfp_mask);
2807 if (!should_suppress_show_mem())
2811 static inline struct page *
2812 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2813 const struct alloc_context *ac, unsigned long *did_some_progress)
2815 struct oom_control oc = {
2816 .zonelist = ac->zonelist,
2817 .nodemask = ac->nodemask,
2818 .gfp_mask = gfp_mask,
2823 *did_some_progress = 0;
2826 * Acquire the oom lock. If that fails, somebody else is
2827 * making progress for us.
2829 if (!mutex_trylock(&oom_lock)) {
2830 *did_some_progress = 1;
2831 schedule_timeout_uninterruptible(1);
2836 * Go through the zonelist yet one more time, keep very high watermark
2837 * here, this is only to catch a parallel oom killing, we must fail if
2838 * we're still under heavy pressure.
2840 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2841 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2845 if (!(gfp_mask & __GFP_NOFAIL)) {
2846 /* Coredumps can quickly deplete all memory reserves */
2847 if (current->flags & PF_DUMPCORE)
2849 /* The OOM killer will not help higher order allocs */
2850 if (order > PAGE_ALLOC_COSTLY_ORDER)
2852 /* The OOM killer does not needlessly kill tasks for lowmem */
2853 if (ac->high_zoneidx < ZONE_NORMAL)
2855 /* The OOM killer does not compensate for IO-less reclaim */
2856 if (!(gfp_mask & __GFP_FS)) {
2858 * XXX: Page reclaim didn't yield anything,
2859 * and the OOM killer can't be invoked, but
2860 * keep looping as per tradition.
2862 *did_some_progress = 1;
2865 if (pm_suspended_storage())
2867 /* The OOM killer may not free memory on a specific node */
2868 if (gfp_mask & __GFP_THISNODE)
2871 /* Exhausted what can be done so it's blamo time */
2872 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
2873 *did_some_progress = 1;
2875 if (gfp_mask & __GFP_NOFAIL) {
2876 page = get_page_from_freelist(gfp_mask, order,
2877 ALLOC_NO_WATERMARKS|ALLOC_CPUSET, ac);
2879 * fallback to ignore cpuset restriction if our nodes
2883 page = get_page_from_freelist(gfp_mask, order,
2884 ALLOC_NO_WATERMARKS, ac);
2888 mutex_unlock(&oom_lock);
2892 #ifdef CONFIG_COMPACTION
2893 /* Try memory compaction for high-order allocations before reclaim */
2894 static struct page *
2895 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2896 int alloc_flags, const struct alloc_context *ac,
2897 enum migrate_mode mode, int *contended_compaction,
2898 bool *deferred_compaction)
2900 unsigned long compact_result;
2906 current->flags |= PF_MEMALLOC;
2907 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2908 mode, contended_compaction);
2909 current->flags &= ~PF_MEMALLOC;
2911 switch (compact_result) {
2912 case COMPACT_DEFERRED:
2913 *deferred_compaction = true;
2915 case COMPACT_SKIPPED:
2922 * At least in one zone compaction wasn't deferred or skipped, so let's
2923 * count a compaction stall
2925 count_vm_event(COMPACTSTALL);
2927 page = get_page_from_freelist(gfp_mask, order,
2928 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2931 struct zone *zone = page_zone(page);
2933 zone->compact_blockskip_flush = false;
2934 compaction_defer_reset(zone, order, true);
2935 count_vm_event(COMPACTSUCCESS);
2940 * It's bad if compaction run occurs and fails. The most likely reason
2941 * is that pages exist, but not enough to satisfy watermarks.
2943 count_vm_event(COMPACTFAIL);
2950 static inline struct page *
2951 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2952 int alloc_flags, const struct alloc_context *ac,
2953 enum migrate_mode mode, int *contended_compaction,
2954 bool *deferred_compaction)
2958 #endif /* CONFIG_COMPACTION */
2960 /* Perform direct synchronous page reclaim */
2962 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2963 const struct alloc_context *ac)
2965 struct reclaim_state reclaim_state;
2970 /* We now go into synchronous reclaim */
2971 cpuset_memory_pressure_bump();
2972 current->flags |= PF_MEMALLOC;
2973 lockdep_set_current_reclaim_state(gfp_mask);
2974 reclaim_state.reclaimed_slab = 0;
2975 current->reclaim_state = &reclaim_state;
2977 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2980 current->reclaim_state = NULL;
2981 lockdep_clear_current_reclaim_state();
2982 current->flags &= ~PF_MEMALLOC;
2989 /* The really slow allocator path where we enter direct reclaim */
2990 static inline struct page *
2991 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2992 int alloc_flags, const struct alloc_context *ac,
2993 unsigned long *did_some_progress)
2995 struct page *page = NULL;
2996 bool drained = false;
2998 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2999 if (unlikely(!(*did_some_progress)))
3003 page = get_page_from_freelist(gfp_mask, order,
3004 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3007 * If an allocation failed after direct reclaim, it could be because
3008 * pages are pinned on the per-cpu lists or in high alloc reserves.
3009 * Shrink them them and try again
3011 if (!page && !drained) {
3012 unreserve_highatomic_pageblock(ac);
3013 drain_all_pages(NULL);
3021 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3026 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3027 ac->high_zoneidx, ac->nodemask)
3028 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
3032 gfp_to_alloc_flags(gfp_t gfp_mask)
3034 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3036 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3037 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3040 * The caller may dip into page reserves a bit more if the caller
3041 * cannot run direct reclaim, or if the caller has realtime scheduling
3042 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3043 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3045 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3047 if (gfp_mask & __GFP_ATOMIC) {
3049 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3050 * if it can't schedule.
3052 if (!(gfp_mask & __GFP_NOMEMALLOC))
3053 alloc_flags |= ALLOC_HARDER;
3055 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3056 * comment for __cpuset_node_allowed().
3058 alloc_flags &= ~ALLOC_CPUSET;
3059 } else if (unlikely(rt_task(current)) && !in_interrupt())
3060 alloc_flags |= ALLOC_HARDER;
3062 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
3063 if (gfp_mask & __GFP_MEMALLOC)
3064 alloc_flags |= ALLOC_NO_WATERMARKS;
3065 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3066 alloc_flags |= ALLOC_NO_WATERMARKS;
3067 else if (!in_interrupt() &&
3068 ((current->flags & PF_MEMALLOC) ||
3069 unlikely(test_thread_flag(TIF_MEMDIE))))
3070 alloc_flags |= ALLOC_NO_WATERMARKS;
3073 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3074 alloc_flags |= ALLOC_CMA;
3079 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3081 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
3084 static inline bool is_thp_gfp_mask(gfp_t gfp_mask)
3086 return (gfp_mask & (GFP_TRANSHUGE | __GFP_KSWAPD_RECLAIM)) == GFP_TRANSHUGE;
3089 static inline struct page *
3090 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3091 struct alloc_context *ac)
3093 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3094 struct page *page = NULL;
3096 unsigned long pages_reclaimed = 0;
3097 unsigned long did_some_progress;
3098 enum migrate_mode migration_mode = MIGRATE_ASYNC;
3099 bool deferred_compaction = false;
3100 int contended_compaction = COMPACT_CONTENDED_NONE;
3103 * In the slowpath, we sanity check order to avoid ever trying to
3104 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3105 * be using allocators in order of preference for an area that is
3108 if (order >= MAX_ORDER) {
3109 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3114 * We also sanity check to catch abuse of atomic reserves being used by
3115 * callers that are not in atomic context.
3117 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3118 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3119 gfp_mask &= ~__GFP_ATOMIC;
3122 * If this allocation cannot block and it is for a specific node, then
3123 * fail early. There's no need to wakeup kswapd or retry for a
3124 * speculative node-specific allocation.
3126 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !can_direct_reclaim)
3130 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3131 wake_all_kswapds(order, ac);
3134 * OK, we're below the kswapd watermark and have kicked background
3135 * reclaim. Now things get more complex, so set up alloc_flags according
3136 * to how we want to proceed.
3138 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3141 * Find the true preferred zone if the allocation is unconstrained by
3144 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
3145 struct zoneref *preferred_zoneref;
3146 preferred_zoneref = first_zones_zonelist(ac->zonelist,
3147 ac->high_zoneidx, NULL, &ac->preferred_zone);
3148 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
3151 /* This is the last chance, in general, before the goto nopage. */
3152 page = get_page_from_freelist(gfp_mask, order,
3153 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
3157 /* Allocate without watermarks if the context allows */
3158 if (alloc_flags & ALLOC_NO_WATERMARKS) {
3160 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3161 * the allocation is high priority and these type of
3162 * allocations are system rather than user orientated
3164 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3165 page = get_page_from_freelist(gfp_mask, order,
3166 ALLOC_NO_WATERMARKS, ac);
3171 /* Caller is not willing to reclaim, we can't balance anything */
3172 if (!can_direct_reclaim) {
3174 * All existing users of the __GFP_NOFAIL are blockable, so warn
3175 * of any new users that actually allow this type of allocation
3178 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
3182 /* Avoid recursion of direct reclaim */
3183 if (current->flags & PF_MEMALLOC) {
3185 * __GFP_NOFAIL request from this context is rather bizarre
3186 * because we cannot reclaim anything and only can loop waiting
3187 * for somebody to do a work for us.
3189 if (WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3196 /* Avoid allocations with no watermarks from looping endlessly */
3197 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
3201 * Try direct compaction. The first pass is asynchronous. Subsequent
3202 * attempts after direct reclaim are synchronous
3204 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3206 &contended_compaction,
3207 &deferred_compaction);
3211 /* Checks for THP-specific high-order allocations */
3212 if (is_thp_gfp_mask(gfp_mask)) {
3214 * If compaction is deferred for high-order allocations, it is
3215 * because sync compaction recently failed. If this is the case
3216 * and the caller requested a THP allocation, we do not want
3217 * to heavily disrupt the system, so we fail the allocation
3218 * instead of entering direct reclaim.
3220 if (deferred_compaction)
3224 * In all zones where compaction was attempted (and not
3225 * deferred or skipped), lock contention has been detected.
3226 * For THP allocation we do not want to disrupt the others
3227 * so we fallback to base pages instead.
3229 if (contended_compaction == COMPACT_CONTENDED_LOCK)
3233 * If compaction was aborted due to need_resched(), we do not
3234 * want to further increase allocation latency, unless it is
3235 * khugepaged trying to collapse.
3237 if (contended_compaction == COMPACT_CONTENDED_SCHED
3238 && !(current->flags & PF_KTHREAD))
3243 * It can become very expensive to allocate transparent hugepages at
3244 * fault, so use asynchronous memory compaction for THP unless it is
3245 * khugepaged trying to collapse.
3247 if (!is_thp_gfp_mask(gfp_mask) || (current->flags & PF_KTHREAD))
3248 migration_mode = MIGRATE_SYNC_LIGHT;
3250 /* Try direct reclaim and then allocating */
3251 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3252 &did_some_progress);
3256 /* Do not loop if specifically requested */
3257 if (gfp_mask & __GFP_NORETRY)
3260 /* Keep reclaiming pages as long as there is reasonable progress */
3261 pages_reclaimed += did_some_progress;
3262 if ((did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER) ||
3263 ((gfp_mask & __GFP_REPEAT) && pages_reclaimed < (1 << order))) {
3264 /* Wait for some write requests to complete then retry */
3265 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
3269 /* Reclaim has failed us, start killing things */
3270 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3274 /* Retry as long as the OOM killer is making progress */
3275 if (did_some_progress)
3280 * High-order allocations do not necessarily loop after
3281 * direct reclaim and reclaim/compaction depends on compaction
3282 * being called after reclaim so call directly if necessary
3284 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags,
3286 &contended_compaction,
3287 &deferred_compaction);
3291 warn_alloc_failed(gfp_mask, order, NULL);
3297 * This is the 'heart' of the zoned buddy allocator.
3300 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3301 struct zonelist *zonelist, nodemask_t *nodemask)
3303 struct zoneref *preferred_zoneref;
3304 struct page *page = NULL;
3305 unsigned int cpuset_mems_cookie;
3306 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
3307 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
3308 struct alloc_context ac = {
3309 .high_zoneidx = gfp_zone(gfp_mask),
3310 .nodemask = nodemask,
3311 .migratetype = gfpflags_to_migratetype(gfp_mask),
3314 gfp_mask &= gfp_allowed_mask;
3316 lockdep_trace_alloc(gfp_mask);
3318 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3320 if (should_fail_alloc_page(gfp_mask, order))
3324 * Check the zones suitable for the gfp_mask contain at least one
3325 * valid zone. It's possible to have an empty zonelist as a result
3326 * of __GFP_THISNODE and a memoryless node
3328 if (unlikely(!zonelist->_zonerefs->zone))
3331 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
3332 alloc_flags |= ALLOC_CMA;
3335 cpuset_mems_cookie = read_mems_allowed_begin();
3337 /* We set it here, as __alloc_pages_slowpath might have changed it */
3338 ac.zonelist = zonelist;
3340 /* Dirty zone balancing only done in the fast path */
3341 ac.spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3343 /* The preferred zone is used for statistics later */
3344 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
3345 ac.nodemask ? : &cpuset_current_mems_allowed,
3346 &ac.preferred_zone);
3347 if (!ac.preferred_zone)
3349 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
3351 /* First allocation attempt */
3352 alloc_mask = gfp_mask|__GFP_HARDWALL;
3353 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3354 if (unlikely(!page)) {
3356 * Runtime PM, block IO and its error handling path
3357 * can deadlock because I/O on the device might not
3360 alloc_mask = memalloc_noio_flags(gfp_mask);
3361 ac.spread_dirty_pages = false;
3363 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3366 if (kmemcheck_enabled && page)
3367 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
3369 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
3373 * When updating a task's mems_allowed, it is possible to race with
3374 * parallel threads in such a way that an allocation can fail while
3375 * the mask is being updated. If a page allocation is about to fail,
3376 * check if the cpuset changed during allocation and if so, retry.
3378 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
3383 EXPORT_SYMBOL(__alloc_pages_nodemask);
3386 * Common helper functions.
3388 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
3393 * __get_free_pages() returns a 32-bit address, which cannot represent
3396 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
3398 page = alloc_pages(gfp_mask, order);
3401 return (unsigned long) page_address(page);
3403 EXPORT_SYMBOL(__get_free_pages);
3405 unsigned long get_zeroed_page(gfp_t gfp_mask)
3407 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
3409 EXPORT_SYMBOL(get_zeroed_page);
3411 void __free_pages(struct page *page, unsigned int order)
3413 if (put_page_testzero(page)) {
3415 free_hot_cold_page(page, false);
3417 __free_pages_ok(page, order);
3421 EXPORT_SYMBOL(__free_pages);
3423 void free_pages(unsigned long addr, unsigned int order)
3426 VM_BUG_ON(!virt_addr_valid((void *)addr));
3427 __free_pages(virt_to_page((void *)addr), order);
3431 EXPORT_SYMBOL(free_pages);
3435 * An arbitrary-length arbitrary-offset area of memory which resides
3436 * within a 0 or higher order page. Multiple fragments within that page
3437 * are individually refcounted, in the page's reference counter.
3439 * The page_frag functions below provide a simple allocation framework for
3440 * page fragments. This is used by the network stack and network device
3441 * drivers to provide a backing region of memory for use as either an
3442 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3444 static struct page *__page_frag_refill(struct page_frag_cache *nc,
3447 struct page *page = NULL;
3448 gfp_t gfp = gfp_mask;
3450 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3451 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
3453 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
3454 PAGE_FRAG_CACHE_MAX_ORDER);
3455 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
3457 if (unlikely(!page))
3458 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
3460 nc->va = page ? page_address(page) : NULL;
3465 void *__alloc_page_frag(struct page_frag_cache *nc,
3466 unsigned int fragsz, gfp_t gfp_mask)
3468 unsigned int size = PAGE_SIZE;
3472 if (unlikely(!nc->va)) {
3474 page = __page_frag_refill(nc, gfp_mask);
3478 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3479 /* if size can vary use size else just use PAGE_SIZE */
3482 /* Even if we own the page, we do not use atomic_set().
3483 * This would break get_page_unless_zero() users.
3485 atomic_add(size - 1, &page->_count);
3487 /* reset page count bias and offset to start of new frag */
3488 nc->pfmemalloc = page_is_pfmemalloc(page);
3489 nc->pagecnt_bias = size;
3493 offset = nc->offset - fragsz;
3494 if (unlikely(offset < 0)) {
3495 page = virt_to_page(nc->va);
3497 if (!atomic_sub_and_test(nc->pagecnt_bias, &page->_count))
3500 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3501 /* if size can vary use size else just use PAGE_SIZE */
3504 /* OK, page count is 0, we can safely set it */
3505 atomic_set(&page->_count, size);
3507 /* reset page count bias and offset to start of new frag */
3508 nc->pagecnt_bias = size;
3509 offset = size - fragsz;
3513 nc->offset = offset;
3515 return nc->va + offset;
3517 EXPORT_SYMBOL(__alloc_page_frag);
3520 * Frees a page fragment allocated out of either a compound or order 0 page.
3522 void __free_page_frag(void *addr)
3524 struct page *page = virt_to_head_page(addr);
3526 if (unlikely(put_page_testzero(page)))
3527 __free_pages_ok(page, compound_order(page));
3529 EXPORT_SYMBOL(__free_page_frag);
3532 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3533 * of the current memory cgroup if __GFP_ACCOUNT is set, other than that it is
3534 * equivalent to alloc_pages.
3536 * It should be used when the caller would like to use kmalloc, but since the
3537 * allocation is large, it has to fall back to the page allocator.
3539 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
3543 page = alloc_pages(gfp_mask, order);
3544 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3545 __free_pages(page, order);
3551 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
3555 page = alloc_pages_node(nid, gfp_mask, order);
3556 if (page && memcg_kmem_charge(page, gfp_mask, order) != 0) {
3557 __free_pages(page, order);
3564 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3567 void __free_kmem_pages(struct page *page, unsigned int order)
3569 memcg_kmem_uncharge(page, order);
3570 __free_pages(page, order);
3573 void free_kmem_pages(unsigned long addr, unsigned int order)
3576 VM_BUG_ON(!virt_addr_valid((void *)addr));
3577 __free_kmem_pages(virt_to_page((void *)addr), order);
3581 static void *make_alloc_exact(unsigned long addr, unsigned int order,
3585 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3586 unsigned long used = addr + PAGE_ALIGN(size);
3588 split_page(virt_to_page((void *)addr), order);
3589 while (used < alloc_end) {
3594 return (void *)addr;
3598 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3599 * @size: the number of bytes to allocate
3600 * @gfp_mask: GFP flags for the allocation
3602 * This function is similar to alloc_pages(), except that it allocates the
3603 * minimum number of pages to satisfy the request. alloc_pages() can only
3604 * allocate memory in power-of-two pages.
3606 * This function is also limited by MAX_ORDER.
3608 * Memory allocated by this function must be released by free_pages_exact().
3610 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3612 unsigned int order = get_order(size);
3615 addr = __get_free_pages(gfp_mask, order);
3616 return make_alloc_exact(addr, order, size);
3618 EXPORT_SYMBOL(alloc_pages_exact);
3621 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3623 * @nid: the preferred node ID where memory should be allocated
3624 * @size: the number of bytes to allocate
3625 * @gfp_mask: GFP flags for the allocation
3627 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3630 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3632 unsigned int order = get_order(size);
3633 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3636 return make_alloc_exact((unsigned long)page_address(p), order, size);
3640 * free_pages_exact - release memory allocated via alloc_pages_exact()
3641 * @virt: the value returned by alloc_pages_exact.
3642 * @size: size of allocation, same value as passed to alloc_pages_exact().
3644 * Release the memory allocated by a previous call to alloc_pages_exact.
3646 void free_pages_exact(void *virt, size_t size)
3648 unsigned long addr = (unsigned long)virt;
3649 unsigned long end = addr + PAGE_ALIGN(size);
3651 while (addr < end) {
3656 EXPORT_SYMBOL(free_pages_exact);
3659 * nr_free_zone_pages - count number of pages beyond high watermark
3660 * @offset: The zone index of the highest zone
3662 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3663 * high watermark within all zones at or below a given zone index. For each
3664 * zone, the number of pages is calculated as:
3665 * managed_pages - high_pages
3667 static unsigned long nr_free_zone_pages(int offset)
3672 /* Just pick one node, since fallback list is circular */
3673 unsigned long sum = 0;
3675 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3677 for_each_zone_zonelist(zone, z, zonelist, offset) {
3678 unsigned long size = zone->managed_pages;
3679 unsigned long high = high_wmark_pages(zone);
3688 * nr_free_buffer_pages - count number of pages beyond high watermark
3690 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3691 * watermark within ZONE_DMA and ZONE_NORMAL.
3693 unsigned long nr_free_buffer_pages(void)
3695 return nr_free_zone_pages(gfp_zone(GFP_USER));
3697 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3700 * nr_free_pagecache_pages - count number of pages beyond high watermark
3702 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3703 * high watermark within all zones.
3705 unsigned long nr_free_pagecache_pages(void)
3707 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3710 static inline void show_node(struct zone *zone)
3712 if (IS_ENABLED(CONFIG_NUMA))
3713 printk("Node %d ", zone_to_nid(zone));
3716 void si_meminfo(struct sysinfo *val)
3718 val->totalram = totalram_pages;
3719 val->sharedram = global_page_state(NR_SHMEM);
3720 val->freeram = global_page_state(NR_FREE_PAGES);
3721 val->bufferram = nr_blockdev_pages();
3722 val->totalhigh = totalhigh_pages;
3723 val->freehigh = nr_free_highpages();
3724 val->mem_unit = PAGE_SIZE;
3727 EXPORT_SYMBOL(si_meminfo);
3730 void si_meminfo_node(struct sysinfo *val, int nid)
3732 int zone_type; /* needs to be signed */
3733 unsigned long managed_pages = 0;
3734 pg_data_t *pgdat = NODE_DATA(nid);
3736 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3737 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3738 val->totalram = managed_pages;
3739 val->sharedram = node_page_state(nid, NR_SHMEM);
3740 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3741 #ifdef CONFIG_HIGHMEM
3742 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3743 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3749 val->mem_unit = PAGE_SIZE;
3754 * Determine whether the node should be displayed or not, depending on whether
3755 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3757 bool skip_free_areas_node(unsigned int flags, int nid)
3760 unsigned int cpuset_mems_cookie;
3762 if (!(flags & SHOW_MEM_FILTER_NODES))
3766 cpuset_mems_cookie = read_mems_allowed_begin();
3767 ret = !node_isset(nid, cpuset_current_mems_allowed);
3768 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3773 #define K(x) ((x) << (PAGE_SHIFT-10))
3775 static void show_migration_types(unsigned char type)
3777 static const char types[MIGRATE_TYPES] = {
3778 [MIGRATE_UNMOVABLE] = 'U',
3779 [MIGRATE_MOVABLE] = 'M',
3780 [MIGRATE_RECLAIMABLE] = 'E',
3781 [MIGRATE_HIGHATOMIC] = 'H',
3783 [MIGRATE_CMA] = 'C',
3785 #ifdef CONFIG_MEMORY_ISOLATION
3786 [MIGRATE_ISOLATE] = 'I',
3789 char tmp[MIGRATE_TYPES + 1];
3793 for (i = 0; i < MIGRATE_TYPES; i++) {
3794 if (type & (1 << i))
3799 printk("(%s) ", tmp);
3803 * Show free area list (used inside shift_scroll-lock stuff)
3804 * We also calculate the percentage fragmentation. We do this by counting the
3805 * memory on each free list with the exception of the first item on the list.
3808 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3811 void show_free_areas(unsigned int filter)
3813 unsigned long free_pcp = 0;
3817 for_each_populated_zone(zone) {
3818 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3821 for_each_online_cpu(cpu)
3822 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3825 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3826 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3827 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3828 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3829 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3830 " free:%lu free_pcp:%lu free_cma:%lu\n",
3831 global_page_state(NR_ACTIVE_ANON),
3832 global_page_state(NR_INACTIVE_ANON),
3833 global_page_state(NR_ISOLATED_ANON),
3834 global_page_state(NR_ACTIVE_FILE),
3835 global_page_state(NR_INACTIVE_FILE),
3836 global_page_state(NR_ISOLATED_FILE),
3837 global_page_state(NR_UNEVICTABLE),
3838 global_page_state(NR_FILE_DIRTY),
3839 global_page_state(NR_WRITEBACK),
3840 global_page_state(NR_UNSTABLE_NFS),
3841 global_page_state(NR_SLAB_RECLAIMABLE),
3842 global_page_state(NR_SLAB_UNRECLAIMABLE),
3843 global_page_state(NR_FILE_MAPPED),
3844 global_page_state(NR_SHMEM),
3845 global_page_state(NR_PAGETABLE),
3846 global_page_state(NR_BOUNCE),
3847 global_page_state(NR_FREE_PAGES),
3849 global_page_state(NR_FREE_CMA_PAGES));
3851 for_each_populated_zone(zone) {
3854 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3858 for_each_online_cpu(cpu)
3859 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3867 " active_anon:%lukB"
3868 " inactive_anon:%lukB"
3869 " active_file:%lukB"
3870 " inactive_file:%lukB"
3871 " unevictable:%lukB"
3872 " isolated(anon):%lukB"
3873 " isolated(file):%lukB"
3881 " slab_reclaimable:%lukB"
3882 " slab_unreclaimable:%lukB"
3883 " kernel_stack:%lukB"
3890 " writeback_tmp:%lukB"
3891 " pages_scanned:%lu"
3892 " all_unreclaimable? %s"
3895 K(zone_page_state(zone, NR_FREE_PAGES)),
3896 K(min_wmark_pages(zone)),
3897 K(low_wmark_pages(zone)),
3898 K(high_wmark_pages(zone)),
3899 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3900 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3901 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3902 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3903 K(zone_page_state(zone, NR_UNEVICTABLE)),
3904 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3905 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3906 K(zone->present_pages),
3907 K(zone->managed_pages),
3908 K(zone_page_state(zone, NR_MLOCK)),
3909 K(zone_page_state(zone, NR_FILE_DIRTY)),
3910 K(zone_page_state(zone, NR_WRITEBACK)),
3911 K(zone_page_state(zone, NR_FILE_MAPPED)),
3912 K(zone_page_state(zone, NR_SHMEM)),
3913 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3914 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3915 zone_page_state(zone, NR_KERNEL_STACK) *
3917 K(zone_page_state(zone, NR_PAGETABLE)),
3918 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3919 K(zone_page_state(zone, NR_BOUNCE)),
3921 K(this_cpu_read(zone->pageset->pcp.count)),
3922 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3923 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3924 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3925 (!zone_reclaimable(zone) ? "yes" : "no")
3927 printk("lowmem_reserve[]:");
3928 for (i = 0; i < MAX_NR_ZONES; i++)
3929 printk(" %ld", zone->lowmem_reserve[i]);
3933 for_each_populated_zone(zone) {
3935 unsigned long nr[MAX_ORDER], flags, total = 0;
3936 unsigned char types[MAX_ORDER];
3938 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3941 printk("%s: ", zone->name);
3943 spin_lock_irqsave(&zone->lock, flags);
3944 for (order = 0; order < MAX_ORDER; order++) {
3945 struct free_area *area = &zone->free_area[order];
3948 nr[order] = area->nr_free;
3949 total += nr[order] << order;
3952 for (type = 0; type < MIGRATE_TYPES; type++) {
3953 if (!list_empty(&area->free_list[type]))
3954 types[order] |= 1 << type;
3957 spin_unlock_irqrestore(&zone->lock, flags);
3958 for (order = 0; order < MAX_ORDER; order++) {
3959 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3961 show_migration_types(types[order]);
3963 printk("= %lukB\n", K(total));
3966 hugetlb_show_meminfo();
3968 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3970 show_swap_cache_info();
3973 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3975 zoneref->zone = zone;
3976 zoneref->zone_idx = zone_idx(zone);
3980 * Builds allocation fallback zone lists.
3982 * Add all populated zones of a node to the zonelist.
3984 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3988 enum zone_type zone_type = MAX_NR_ZONES;
3992 zone = pgdat->node_zones + zone_type;
3993 if (populated_zone(zone)) {
3994 zoneref_set_zone(zone,
3995 &zonelist->_zonerefs[nr_zones++]);
3996 check_highest_zone(zone_type);
3998 } while (zone_type);
4006 * 0 = automatic detection of better ordering.
4007 * 1 = order by ([node] distance, -zonetype)
4008 * 2 = order by (-zonetype, [node] distance)
4010 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4011 * the same zonelist. So only NUMA can configure this param.
4013 #define ZONELIST_ORDER_DEFAULT 0
4014 #define ZONELIST_ORDER_NODE 1
4015 #define ZONELIST_ORDER_ZONE 2
4017 /* zonelist order in the kernel.
4018 * set_zonelist_order() will set this to NODE or ZONE.
4020 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4021 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4025 /* The value user specified ....changed by config */
4026 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4027 /* string for sysctl */
4028 #define NUMA_ZONELIST_ORDER_LEN 16
4029 char numa_zonelist_order[16] = "default";
4032 * interface for configure zonelist ordering.
4033 * command line option "numa_zonelist_order"
4034 * = "[dD]efault - default, automatic configuration.
4035 * = "[nN]ode - order by node locality, then by zone within node
4036 * = "[zZ]one - order by zone, then by locality within zone
4039 static int __parse_numa_zonelist_order(char *s)
4041 if (*s == 'd' || *s == 'D') {
4042 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4043 } else if (*s == 'n' || *s == 'N') {
4044 user_zonelist_order = ZONELIST_ORDER_NODE;
4045 } else if (*s == 'z' || *s == 'Z') {
4046 user_zonelist_order = ZONELIST_ORDER_ZONE;
4049 "Ignoring invalid numa_zonelist_order value: "
4056 static __init int setup_numa_zonelist_order(char *s)
4063 ret = __parse_numa_zonelist_order(s);
4065 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4069 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4072 * sysctl handler for numa_zonelist_order
4074 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4075 void __user *buffer, size_t *length,
4078 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4080 static DEFINE_MUTEX(zl_order_mutex);
4082 mutex_lock(&zl_order_mutex);
4084 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4088 strcpy(saved_string, (char *)table->data);
4090 ret = proc_dostring(table, write, buffer, length, ppos);
4094 int oldval = user_zonelist_order;
4096 ret = __parse_numa_zonelist_order((char *)table->data);
4099 * bogus value. restore saved string
4101 strncpy((char *)table->data, saved_string,
4102 NUMA_ZONELIST_ORDER_LEN);
4103 user_zonelist_order = oldval;
4104 } else if (oldval != user_zonelist_order) {
4105 mutex_lock(&zonelists_mutex);
4106 build_all_zonelists(NULL, NULL);
4107 mutex_unlock(&zonelists_mutex);
4111 mutex_unlock(&zl_order_mutex);
4116 #define MAX_NODE_LOAD (nr_online_nodes)
4117 static int node_load[MAX_NUMNODES];
4120 * find_next_best_node - find the next node that should appear in a given node's fallback list
4121 * @node: node whose fallback list we're appending
4122 * @used_node_mask: nodemask_t of already used nodes
4124 * We use a number of factors to determine which is the next node that should
4125 * appear on a given node's fallback list. The node should not have appeared
4126 * already in @node's fallback list, and it should be the next closest node
4127 * according to the distance array (which contains arbitrary distance values
4128 * from each node to each node in the system), and should also prefer nodes
4129 * with no CPUs, since presumably they'll have very little allocation pressure
4130 * on them otherwise.
4131 * It returns -1 if no node is found.
4133 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4136 int min_val = INT_MAX;
4137 int best_node = NUMA_NO_NODE;
4138 const struct cpumask *tmp = cpumask_of_node(0);
4140 /* Use the local node if we haven't already */
4141 if (!node_isset(node, *used_node_mask)) {
4142 node_set(node, *used_node_mask);
4146 for_each_node_state(n, N_MEMORY) {
4148 /* Don't want a node to appear more than once */
4149 if (node_isset(n, *used_node_mask))
4152 /* Use the distance array to find the distance */
4153 val = node_distance(node, n);
4155 /* Penalize nodes under us ("prefer the next node") */
4158 /* Give preference to headless and unused nodes */
4159 tmp = cpumask_of_node(n);
4160 if (!cpumask_empty(tmp))
4161 val += PENALTY_FOR_NODE_WITH_CPUS;
4163 /* Slight preference for less loaded node */
4164 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4165 val += node_load[n];
4167 if (val < min_val) {
4174 node_set(best_node, *used_node_mask);
4181 * Build zonelists ordered by node and zones within node.
4182 * This results in maximum locality--normal zone overflows into local
4183 * DMA zone, if any--but risks exhausting DMA zone.
4185 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4188 struct zonelist *zonelist;
4190 zonelist = &pgdat->node_zonelists[0];
4191 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4193 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4194 zonelist->_zonerefs[j].zone = NULL;
4195 zonelist->_zonerefs[j].zone_idx = 0;
4199 * Build gfp_thisnode zonelists
4201 static void build_thisnode_zonelists(pg_data_t *pgdat)
4204 struct zonelist *zonelist;
4206 zonelist = &pgdat->node_zonelists[1];
4207 j = build_zonelists_node(pgdat, zonelist, 0);
4208 zonelist->_zonerefs[j].zone = NULL;
4209 zonelist->_zonerefs[j].zone_idx = 0;
4213 * Build zonelists ordered by zone and nodes within zones.
4214 * This results in conserving DMA zone[s] until all Normal memory is
4215 * exhausted, but results in overflowing to remote node while memory
4216 * may still exist in local DMA zone.
4218 static int node_order[MAX_NUMNODES];
4220 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4223 int zone_type; /* needs to be signed */
4225 struct zonelist *zonelist;
4227 zonelist = &pgdat->node_zonelists[0];
4229 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4230 for (j = 0; j < nr_nodes; j++) {
4231 node = node_order[j];
4232 z = &NODE_DATA(node)->node_zones[zone_type];
4233 if (populated_zone(z)) {
4235 &zonelist->_zonerefs[pos++]);
4236 check_highest_zone(zone_type);
4240 zonelist->_zonerefs[pos].zone = NULL;
4241 zonelist->_zonerefs[pos].zone_idx = 0;
4244 #if defined(CONFIG_64BIT)
4246 * Devices that require DMA32/DMA are relatively rare and do not justify a
4247 * penalty to every machine in case the specialised case applies. Default
4248 * to Node-ordering on 64-bit NUMA machines
4250 static int default_zonelist_order(void)
4252 return ZONELIST_ORDER_NODE;
4256 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4257 * by the kernel. If processes running on node 0 deplete the low memory zone
4258 * then reclaim will occur more frequency increasing stalls and potentially
4259 * be easier to OOM if a large percentage of the zone is under writeback or
4260 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4261 * Hence, default to zone ordering on 32-bit.
4263 static int default_zonelist_order(void)
4265 return ZONELIST_ORDER_ZONE;
4267 #endif /* CONFIG_64BIT */
4269 static void set_zonelist_order(void)
4271 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4272 current_zonelist_order = default_zonelist_order();
4274 current_zonelist_order = user_zonelist_order;
4277 static void build_zonelists(pg_data_t *pgdat)
4280 nodemask_t used_mask;
4281 int local_node, prev_node;
4282 struct zonelist *zonelist;
4283 unsigned int order = current_zonelist_order;
4285 /* initialize zonelists */
4286 for (i = 0; i < MAX_ZONELISTS; i++) {
4287 zonelist = pgdat->node_zonelists + i;
4288 zonelist->_zonerefs[0].zone = NULL;
4289 zonelist->_zonerefs[0].zone_idx = 0;
4292 /* NUMA-aware ordering of nodes */
4293 local_node = pgdat->node_id;
4294 load = nr_online_nodes;
4295 prev_node = local_node;
4296 nodes_clear(used_mask);
4298 memset(node_order, 0, sizeof(node_order));
4301 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4303 * We don't want to pressure a particular node.
4304 * So adding penalty to the first node in same
4305 * distance group to make it round-robin.
4307 if (node_distance(local_node, node) !=
4308 node_distance(local_node, prev_node))
4309 node_load[node] = load;
4313 if (order == ZONELIST_ORDER_NODE)
4314 build_zonelists_in_node_order(pgdat, node);
4316 node_order[i++] = node; /* remember order */
4319 if (order == ZONELIST_ORDER_ZONE) {
4320 /* calculate node order -- i.e., DMA last! */
4321 build_zonelists_in_zone_order(pgdat, i);
4324 build_thisnode_zonelists(pgdat);
4327 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4329 * Return node id of node used for "local" allocations.
4330 * I.e., first node id of first zone in arg node's generic zonelist.
4331 * Used for initializing percpu 'numa_mem', which is used primarily
4332 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4334 int local_memory_node(int node)
4338 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
4339 gfp_zone(GFP_KERNEL),
4346 #else /* CONFIG_NUMA */
4348 static void set_zonelist_order(void)
4350 current_zonelist_order = ZONELIST_ORDER_ZONE;
4353 static void build_zonelists(pg_data_t *pgdat)
4355 int node, local_node;
4357 struct zonelist *zonelist;
4359 local_node = pgdat->node_id;
4361 zonelist = &pgdat->node_zonelists[0];
4362 j = build_zonelists_node(pgdat, zonelist, 0);
4365 * Now we build the zonelist so that it contains the zones
4366 * of all the other nodes.
4367 * We don't want to pressure a particular node, so when
4368 * building the zones for node N, we make sure that the
4369 * zones coming right after the local ones are those from
4370 * node N+1 (modulo N)
4372 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
4373 if (!node_online(node))
4375 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4377 for (node = 0; node < local_node; node++) {
4378 if (!node_online(node))
4380 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4383 zonelist->_zonerefs[j].zone = NULL;
4384 zonelist->_zonerefs[j].zone_idx = 0;
4387 #endif /* CONFIG_NUMA */
4390 * Boot pageset table. One per cpu which is going to be used for all
4391 * zones and all nodes. The parameters will be set in such a way
4392 * that an item put on a list will immediately be handed over to
4393 * the buddy list. This is safe since pageset manipulation is done
4394 * with interrupts disabled.
4396 * The boot_pagesets must be kept even after bootup is complete for
4397 * unused processors and/or zones. They do play a role for bootstrapping
4398 * hotplugged processors.
4400 * zoneinfo_show() and maybe other functions do
4401 * not check if the processor is online before following the pageset pointer.
4402 * Other parts of the kernel may not check if the zone is available.
4404 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
4405 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
4406 static void setup_zone_pageset(struct zone *zone);
4409 * Global mutex to protect against size modification of zonelists
4410 * as well as to serialize pageset setup for the new populated zone.
4412 DEFINE_MUTEX(zonelists_mutex);
4414 /* return values int ....just for stop_machine() */
4415 static int __build_all_zonelists(void *data)
4419 pg_data_t *self = data;
4422 memset(node_load, 0, sizeof(node_load));
4425 if (self && !node_online(self->node_id)) {
4426 build_zonelists(self);
4429 for_each_online_node(nid) {
4430 pg_data_t *pgdat = NODE_DATA(nid);
4432 build_zonelists(pgdat);
4436 * Initialize the boot_pagesets that are going to be used
4437 * for bootstrapping processors. The real pagesets for
4438 * each zone will be allocated later when the per cpu
4439 * allocator is available.
4441 * boot_pagesets are used also for bootstrapping offline
4442 * cpus if the system is already booted because the pagesets
4443 * are needed to initialize allocators on a specific cpu too.
4444 * F.e. the percpu allocator needs the page allocator which
4445 * needs the percpu allocator in order to allocate its pagesets
4446 * (a chicken-egg dilemma).
4448 for_each_possible_cpu(cpu) {
4449 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
4451 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4453 * We now know the "local memory node" for each node--
4454 * i.e., the node of the first zone in the generic zonelist.
4455 * Set up numa_mem percpu variable for on-line cpus. During
4456 * boot, only the boot cpu should be on-line; we'll init the
4457 * secondary cpus' numa_mem as they come on-line. During
4458 * node/memory hotplug, we'll fixup all on-line cpus.
4460 if (cpu_online(cpu))
4461 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
4468 static noinline void __init
4469 build_all_zonelists_init(void)
4471 __build_all_zonelists(NULL);
4472 mminit_verify_zonelist();
4473 cpuset_init_current_mems_allowed();
4477 * Called with zonelists_mutex held always
4478 * unless system_state == SYSTEM_BOOTING.
4480 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4481 * [we're only called with non-NULL zone through __meminit paths] and
4482 * (2) call of __init annotated helper build_all_zonelists_init
4483 * [protected by SYSTEM_BOOTING].
4485 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
4487 set_zonelist_order();
4489 if (system_state == SYSTEM_BOOTING) {
4490 build_all_zonelists_init();
4492 #ifdef CONFIG_MEMORY_HOTPLUG
4494 setup_zone_pageset(zone);
4496 /* we have to stop all cpus to guarantee there is no user
4498 stop_machine(__build_all_zonelists, pgdat, NULL);
4499 /* cpuset refresh routine should be here */
4501 vm_total_pages = nr_free_pagecache_pages();
4503 * Disable grouping by mobility if the number of pages in the
4504 * system is too low to allow the mechanism to work. It would be
4505 * more accurate, but expensive to check per-zone. This check is
4506 * made on memory-hotadd so a system can start with mobility
4507 * disabled and enable it later
4509 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
4510 page_group_by_mobility_disabled = 1;
4512 page_group_by_mobility_disabled = 0;
4514 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4515 "Total pages: %ld\n",
4517 zonelist_order_name[current_zonelist_order],
4518 page_group_by_mobility_disabled ? "off" : "on",
4521 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
4526 * Helper functions to size the waitqueue hash table.
4527 * Essentially these want to choose hash table sizes sufficiently
4528 * large so that collisions trying to wait on pages are rare.
4529 * But in fact, the number of active page waitqueues on typical
4530 * systems is ridiculously low, less than 200. So this is even
4531 * conservative, even though it seems large.
4533 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4534 * waitqueues, i.e. the size of the waitq table given the number of pages.
4536 #define PAGES_PER_WAITQUEUE 256
4538 #ifndef CONFIG_MEMORY_HOTPLUG
4539 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4541 unsigned long size = 1;
4543 pages /= PAGES_PER_WAITQUEUE;
4545 while (size < pages)
4549 * Once we have dozens or even hundreds of threads sleeping
4550 * on IO we've got bigger problems than wait queue collision.
4551 * Limit the size of the wait table to a reasonable size.
4553 size = min(size, 4096UL);
4555 return max(size, 4UL);
4559 * A zone's size might be changed by hot-add, so it is not possible to determine
4560 * a suitable size for its wait_table. So we use the maximum size now.
4562 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4564 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4565 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4566 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4568 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4569 * or more by the traditional way. (See above). It equals:
4571 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4572 * ia64(16K page size) : = ( 8G + 4M)byte.
4573 * powerpc (64K page size) : = (32G +16M)byte.
4575 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4582 * This is an integer logarithm so that shifts can be used later
4583 * to extract the more random high bits from the multiplicative
4584 * hash function before the remainder is taken.
4586 static inline unsigned long wait_table_bits(unsigned long size)
4592 * Initially all pages are reserved - free ones are freed
4593 * up by free_all_bootmem() once the early boot process is
4594 * done. Non-atomic initialization, single-pass.
4596 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4597 unsigned long start_pfn, enum memmap_context context)
4599 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
4600 unsigned long end_pfn = start_pfn + size;
4601 pg_data_t *pgdat = NODE_DATA(nid);
4603 unsigned long nr_initialised = 0;
4604 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4605 struct memblock_region *r = NULL, *tmp;
4608 if (highest_memmap_pfn < end_pfn - 1)
4609 highest_memmap_pfn = end_pfn - 1;
4612 * Honor reservation requested by the driver for this ZONE_DEVICE
4615 if (altmap && start_pfn == altmap->base_pfn)
4616 start_pfn += altmap->reserve;
4618 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4620 * There can be holes in boot-time mem_map[]s handed to this
4621 * function. They do not exist on hotplugged memory.
4623 if (context != MEMMAP_EARLY)
4626 if (!early_pfn_valid(pfn))
4628 if (!early_pfn_in_nid(pfn, nid))
4630 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
4633 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4635 * If not mirrored_kernelcore and ZONE_MOVABLE exists, range
4636 * from zone_movable_pfn[nid] to end of each node should be
4637 * ZONE_MOVABLE not ZONE_NORMAL. skip it.
4639 if (!mirrored_kernelcore && zone_movable_pfn[nid])
4640 if (zone == ZONE_NORMAL && pfn >= zone_movable_pfn[nid])
4644 * Check given memblock attribute by firmware which can affect
4645 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
4646 * mirrored, it's an overlapped memmap init. skip it.
4648 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
4649 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
4650 for_each_memblock(memory, tmp)
4651 if (pfn < memblock_region_memory_end_pfn(tmp))
4655 if (pfn >= memblock_region_memory_base_pfn(r) &&
4656 memblock_is_mirror(r)) {
4657 /* already initialized as NORMAL */
4658 pfn = memblock_region_memory_end_pfn(r);
4666 * Mark the block movable so that blocks are reserved for
4667 * movable at startup. This will force kernel allocations
4668 * to reserve their blocks rather than leaking throughout
4669 * the address space during boot when many long-lived
4670 * kernel allocations are made.
4672 * bitmap is created for zone's valid pfn range. but memmap
4673 * can be created for invalid pages (for alignment)
4674 * check here not to call set_pageblock_migratetype() against
4677 if (!(pfn & (pageblock_nr_pages - 1))) {
4678 struct page *page = pfn_to_page(pfn);
4680 __init_single_page(page, pfn, zone, nid);
4681 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4683 __init_single_pfn(pfn, zone, nid);
4688 static void __meminit zone_init_free_lists(struct zone *zone)
4690 unsigned int order, t;
4691 for_each_migratetype_order(order, t) {
4692 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4693 zone->free_area[order].nr_free = 0;
4697 #ifndef __HAVE_ARCH_MEMMAP_INIT
4698 #define memmap_init(size, nid, zone, start_pfn) \
4699 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4702 static int zone_batchsize(struct zone *zone)
4708 * The per-cpu-pages pools are set to around 1000th of the
4709 * size of the zone. But no more than 1/2 of a meg.
4711 * OK, so we don't know how big the cache is. So guess.
4713 batch = zone->managed_pages / 1024;
4714 if (batch * PAGE_SIZE > 512 * 1024)
4715 batch = (512 * 1024) / PAGE_SIZE;
4716 batch /= 4; /* We effectively *= 4 below */
4721 * Clamp the batch to a 2^n - 1 value. Having a power
4722 * of 2 value was found to be more likely to have
4723 * suboptimal cache aliasing properties in some cases.
4725 * For example if 2 tasks are alternately allocating
4726 * batches of pages, one task can end up with a lot
4727 * of pages of one half of the possible page colors
4728 * and the other with pages of the other colors.
4730 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4735 /* The deferral and batching of frees should be suppressed under NOMMU
4738 * The problem is that NOMMU needs to be able to allocate large chunks
4739 * of contiguous memory as there's no hardware page translation to
4740 * assemble apparent contiguous memory from discontiguous pages.
4742 * Queueing large contiguous runs of pages for batching, however,
4743 * causes the pages to actually be freed in smaller chunks. As there
4744 * can be a significant delay between the individual batches being
4745 * recycled, this leads to the once large chunks of space being
4746 * fragmented and becoming unavailable for high-order allocations.
4753 * pcp->high and pcp->batch values are related and dependent on one another:
4754 * ->batch must never be higher then ->high.
4755 * The following function updates them in a safe manner without read side
4758 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4759 * those fields changing asynchronously (acording the the above rule).
4761 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4762 * outside of boot time (or some other assurance that no concurrent updaters
4765 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4766 unsigned long batch)
4768 /* start with a fail safe value for batch */
4772 /* Update high, then batch, in order */
4779 /* a companion to pageset_set_high() */
4780 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4782 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4785 static void pageset_init(struct per_cpu_pageset *p)
4787 struct per_cpu_pages *pcp;
4790 memset(p, 0, sizeof(*p));
4794 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4795 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4798 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4801 pageset_set_batch(p, batch);
4805 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4806 * to the value high for the pageset p.
4808 static void pageset_set_high(struct per_cpu_pageset *p,
4811 unsigned long batch = max(1UL, high / 4);
4812 if ((high / 4) > (PAGE_SHIFT * 8))
4813 batch = PAGE_SHIFT * 8;
4815 pageset_update(&p->pcp, high, batch);
4818 static void pageset_set_high_and_batch(struct zone *zone,
4819 struct per_cpu_pageset *pcp)
4821 if (percpu_pagelist_fraction)
4822 pageset_set_high(pcp,
4823 (zone->managed_pages /
4824 percpu_pagelist_fraction));
4826 pageset_set_batch(pcp, zone_batchsize(zone));
4829 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4831 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4834 pageset_set_high_and_batch(zone, pcp);
4837 static void __meminit setup_zone_pageset(struct zone *zone)
4840 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4841 for_each_possible_cpu(cpu)
4842 zone_pageset_init(zone, cpu);
4846 * Allocate per cpu pagesets and initialize them.
4847 * Before this call only boot pagesets were available.
4849 void __init setup_per_cpu_pageset(void)
4853 for_each_populated_zone(zone)
4854 setup_zone_pageset(zone);
4857 static noinline __init_refok
4858 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4864 * The per-page waitqueue mechanism uses hashed waitqueues
4867 zone->wait_table_hash_nr_entries =
4868 wait_table_hash_nr_entries(zone_size_pages);
4869 zone->wait_table_bits =
4870 wait_table_bits(zone->wait_table_hash_nr_entries);
4871 alloc_size = zone->wait_table_hash_nr_entries
4872 * sizeof(wait_queue_head_t);
4874 if (!slab_is_available()) {
4875 zone->wait_table = (wait_queue_head_t *)
4876 memblock_virt_alloc_node_nopanic(
4877 alloc_size, zone->zone_pgdat->node_id);
4880 * This case means that a zone whose size was 0 gets new memory
4881 * via memory hot-add.
4882 * But it may be the case that a new node was hot-added. In
4883 * this case vmalloc() will not be able to use this new node's
4884 * memory - this wait_table must be initialized to use this new
4885 * node itself as well.
4886 * To use this new node's memory, further consideration will be
4889 zone->wait_table = vmalloc(alloc_size);
4891 if (!zone->wait_table)
4894 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4895 init_waitqueue_head(zone->wait_table + i);
4900 static __meminit void zone_pcp_init(struct zone *zone)
4903 * per cpu subsystem is not up at this point. The following code
4904 * relies on the ability of the linker to provide the
4905 * offset of a (static) per cpu variable into the per cpu area.
4907 zone->pageset = &boot_pageset;
4909 if (populated_zone(zone))
4910 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4911 zone->name, zone->present_pages,
4912 zone_batchsize(zone));
4915 int __meminit init_currently_empty_zone(struct zone *zone,
4916 unsigned long zone_start_pfn,
4919 struct pglist_data *pgdat = zone->zone_pgdat;
4921 ret = zone_wait_table_init(zone, size);
4924 pgdat->nr_zones = zone_idx(zone) + 1;
4926 zone->zone_start_pfn = zone_start_pfn;
4928 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4929 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4931 (unsigned long)zone_idx(zone),
4932 zone_start_pfn, (zone_start_pfn + size));
4934 zone_init_free_lists(zone);
4939 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4940 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4943 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4945 int __meminit __early_pfn_to_nid(unsigned long pfn,
4946 struct mminit_pfnnid_cache *state)
4948 unsigned long start_pfn, end_pfn;
4951 if (state->last_start <= pfn && pfn < state->last_end)
4952 return state->last_nid;
4954 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4956 state->last_start = start_pfn;
4957 state->last_end = end_pfn;
4958 state->last_nid = nid;
4963 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4966 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4967 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4968 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4970 * If an architecture guarantees that all ranges registered contain no holes
4971 * and may be freed, this this function may be used instead of calling
4972 * memblock_free_early_nid() manually.
4974 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4976 unsigned long start_pfn, end_pfn;
4979 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4980 start_pfn = min(start_pfn, max_low_pfn);
4981 end_pfn = min(end_pfn, max_low_pfn);
4983 if (start_pfn < end_pfn)
4984 memblock_free_early_nid(PFN_PHYS(start_pfn),
4985 (end_pfn - start_pfn) << PAGE_SHIFT,
4991 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4992 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4994 * If an architecture guarantees that all ranges registered contain no holes and may
4995 * be freed, this function may be used instead of calling memory_present() manually.
4997 void __init sparse_memory_present_with_active_regions(int nid)
4999 unsigned long start_pfn, end_pfn;
5002 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5003 memory_present(this_nid, start_pfn, end_pfn);
5007 * get_pfn_range_for_nid - Return the start and end page frames for a node
5008 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5009 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5010 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5012 * It returns the start and end page frame of a node based on information
5013 * provided by memblock_set_node(). If called for a node
5014 * with no available memory, a warning is printed and the start and end
5017 void __meminit get_pfn_range_for_nid(unsigned int nid,
5018 unsigned long *start_pfn, unsigned long *end_pfn)
5020 unsigned long this_start_pfn, this_end_pfn;
5026 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5027 *start_pfn = min(*start_pfn, this_start_pfn);
5028 *end_pfn = max(*end_pfn, this_end_pfn);
5031 if (*start_pfn == -1UL)
5036 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5037 * assumption is made that zones within a node are ordered in monotonic
5038 * increasing memory addresses so that the "highest" populated zone is used
5040 static void __init find_usable_zone_for_movable(void)
5043 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5044 if (zone_index == ZONE_MOVABLE)
5047 if (arch_zone_highest_possible_pfn[zone_index] >
5048 arch_zone_lowest_possible_pfn[zone_index])
5052 VM_BUG_ON(zone_index == -1);
5053 movable_zone = zone_index;
5057 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5058 * because it is sized independent of architecture. Unlike the other zones,
5059 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5060 * in each node depending on the size of each node and how evenly kernelcore
5061 * is distributed. This helper function adjusts the zone ranges
5062 * provided by the architecture for a given node by using the end of the
5063 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5064 * zones within a node are in order of monotonic increases memory addresses
5066 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5067 unsigned long zone_type,
5068 unsigned long node_start_pfn,
5069 unsigned long node_end_pfn,
5070 unsigned long *zone_start_pfn,
5071 unsigned long *zone_end_pfn)
5073 /* Only adjust if ZONE_MOVABLE is on this node */
5074 if (zone_movable_pfn[nid]) {
5075 /* Size ZONE_MOVABLE */
5076 if (zone_type == ZONE_MOVABLE) {
5077 *zone_start_pfn = zone_movable_pfn[nid];
5078 *zone_end_pfn = min(node_end_pfn,
5079 arch_zone_highest_possible_pfn[movable_zone]);
5081 /* Check if this whole range is within ZONE_MOVABLE */
5082 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5083 *zone_start_pfn = *zone_end_pfn;
5088 * Return the number of pages a zone spans in a node, including holes
5089 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5091 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5092 unsigned long zone_type,
5093 unsigned long node_start_pfn,
5094 unsigned long node_end_pfn,
5095 unsigned long *zone_start_pfn,
5096 unsigned long *zone_end_pfn,
5097 unsigned long *ignored)
5099 /* When hotadd a new node from cpu_up(), the node should be empty */
5100 if (!node_start_pfn && !node_end_pfn)
5103 /* Get the start and end of the zone */
5104 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5105 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5106 adjust_zone_range_for_zone_movable(nid, zone_type,
5107 node_start_pfn, node_end_pfn,
5108 zone_start_pfn, zone_end_pfn);
5110 /* Check that this node has pages within the zone's required range */
5111 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5114 /* Move the zone boundaries inside the node if necessary */
5115 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5116 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5118 /* Return the spanned pages */
5119 return *zone_end_pfn - *zone_start_pfn;
5123 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5124 * then all holes in the requested range will be accounted for.
5126 unsigned long __meminit __absent_pages_in_range(int nid,
5127 unsigned long range_start_pfn,
5128 unsigned long range_end_pfn)
5130 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5131 unsigned long start_pfn, end_pfn;
5134 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5135 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5136 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5137 nr_absent -= end_pfn - start_pfn;
5143 * absent_pages_in_range - Return number of page frames in holes within a range
5144 * @start_pfn: The start PFN to start searching for holes
5145 * @end_pfn: The end PFN to stop searching for holes
5147 * It returns the number of pages frames in memory holes within a range.
5149 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5150 unsigned long end_pfn)
5152 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5155 /* Return the number of page frames in holes in a zone on a node */
5156 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5157 unsigned long zone_type,
5158 unsigned long node_start_pfn,
5159 unsigned long node_end_pfn,
5160 unsigned long *ignored)
5162 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5163 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5164 unsigned long zone_start_pfn, zone_end_pfn;
5165 unsigned long nr_absent;
5167 /* When hotadd a new node from cpu_up(), the node should be empty */
5168 if (!node_start_pfn && !node_end_pfn)
5171 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5172 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5174 adjust_zone_range_for_zone_movable(nid, zone_type,
5175 node_start_pfn, node_end_pfn,
5176 &zone_start_pfn, &zone_end_pfn);
5177 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5180 * ZONE_MOVABLE handling.
5181 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5184 if (zone_movable_pfn[nid]) {
5185 if (mirrored_kernelcore) {
5186 unsigned long start_pfn, end_pfn;
5187 struct memblock_region *r;
5189 for_each_memblock(memory, r) {
5190 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5191 zone_start_pfn, zone_end_pfn);
5192 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5193 zone_start_pfn, zone_end_pfn);
5195 if (zone_type == ZONE_MOVABLE &&
5196 memblock_is_mirror(r))
5197 nr_absent += end_pfn - start_pfn;
5199 if (zone_type == ZONE_NORMAL &&
5200 !memblock_is_mirror(r))
5201 nr_absent += end_pfn - start_pfn;
5204 if (zone_type == ZONE_NORMAL)
5205 nr_absent += node_end_pfn - zone_movable_pfn[nid];
5212 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5213 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5214 unsigned long zone_type,
5215 unsigned long node_start_pfn,
5216 unsigned long node_end_pfn,
5217 unsigned long *zone_start_pfn,
5218 unsigned long *zone_end_pfn,
5219 unsigned long *zones_size)
5223 *zone_start_pfn = node_start_pfn;
5224 for (zone = 0; zone < zone_type; zone++)
5225 *zone_start_pfn += zones_size[zone];
5227 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5229 return zones_size[zone_type];
5232 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5233 unsigned long zone_type,
5234 unsigned long node_start_pfn,
5235 unsigned long node_end_pfn,
5236 unsigned long *zholes_size)
5241 return zholes_size[zone_type];
5244 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5246 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5247 unsigned long node_start_pfn,
5248 unsigned long node_end_pfn,
5249 unsigned long *zones_size,
5250 unsigned long *zholes_size)
5252 unsigned long realtotalpages = 0, totalpages = 0;
5255 for (i = 0; i < MAX_NR_ZONES; i++) {
5256 struct zone *zone = pgdat->node_zones + i;
5257 unsigned long zone_start_pfn, zone_end_pfn;
5258 unsigned long size, real_size;
5260 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5266 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5267 node_start_pfn, node_end_pfn,
5270 zone->zone_start_pfn = zone_start_pfn;
5272 zone->zone_start_pfn = 0;
5273 zone->spanned_pages = size;
5274 zone->present_pages = real_size;
5277 realtotalpages += real_size;
5280 pgdat->node_spanned_pages = totalpages;
5281 pgdat->node_present_pages = realtotalpages;
5282 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5286 #ifndef CONFIG_SPARSEMEM
5288 * Calculate the size of the zone->blockflags rounded to an unsigned long
5289 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5290 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5291 * round what is now in bits to nearest long in bits, then return it in
5294 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5296 unsigned long usemapsize;
5298 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5299 usemapsize = roundup(zonesize, pageblock_nr_pages);
5300 usemapsize = usemapsize >> pageblock_order;
5301 usemapsize *= NR_PAGEBLOCK_BITS;
5302 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5304 return usemapsize / 8;
5307 static void __init setup_usemap(struct pglist_data *pgdat,
5309 unsigned long zone_start_pfn,
5310 unsigned long zonesize)
5312 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5313 zone->pageblock_flags = NULL;
5315 zone->pageblock_flags =
5316 memblock_virt_alloc_node_nopanic(usemapsize,
5320 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5321 unsigned long zone_start_pfn, unsigned long zonesize) {}
5322 #endif /* CONFIG_SPARSEMEM */
5324 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5326 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5327 void __paginginit set_pageblock_order(void)
5331 /* Check that pageblock_nr_pages has not already been setup */
5332 if (pageblock_order)
5335 if (HPAGE_SHIFT > PAGE_SHIFT)
5336 order = HUGETLB_PAGE_ORDER;
5338 order = MAX_ORDER - 1;
5341 * Assume the largest contiguous order of interest is a huge page.
5342 * This value may be variable depending on boot parameters on IA64 and
5345 pageblock_order = order;
5347 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5350 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5351 * is unused as pageblock_order is set at compile-time. See
5352 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5355 void __paginginit set_pageblock_order(void)
5359 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5361 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5362 unsigned long present_pages)
5364 unsigned long pages = spanned_pages;
5367 * Provide a more accurate estimation if there are holes within
5368 * the zone and SPARSEMEM is in use. If there are holes within the
5369 * zone, each populated memory region may cost us one or two extra
5370 * memmap pages due to alignment because memmap pages for each
5371 * populated regions may not naturally algined on page boundary.
5372 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5374 if (spanned_pages > present_pages + (present_pages >> 4) &&
5375 IS_ENABLED(CONFIG_SPARSEMEM))
5376 pages = present_pages;
5378 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5382 * Set up the zone data structures:
5383 * - mark all pages reserved
5384 * - mark all memory queues empty
5385 * - clear the memory bitmaps
5387 * NOTE: pgdat should get zeroed by caller.
5389 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5392 int nid = pgdat->node_id;
5395 pgdat_resize_init(pgdat);
5396 #ifdef CONFIG_NUMA_BALANCING
5397 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5398 pgdat->numabalancing_migrate_nr_pages = 0;
5399 pgdat->numabalancing_migrate_next_window = jiffies;
5401 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5402 spin_lock_init(&pgdat->split_queue_lock);
5403 INIT_LIST_HEAD(&pgdat->split_queue);
5404 pgdat->split_queue_len = 0;
5406 init_waitqueue_head(&pgdat->kswapd_wait);
5407 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5408 #ifdef CONFIG_COMPACTION
5409 init_waitqueue_head(&pgdat->kcompactd_wait);
5411 pgdat_page_ext_init(pgdat);
5413 for (j = 0; j < MAX_NR_ZONES; j++) {
5414 struct zone *zone = pgdat->node_zones + j;
5415 unsigned long size, realsize, freesize, memmap_pages;
5416 unsigned long zone_start_pfn = zone->zone_start_pfn;
5418 size = zone->spanned_pages;
5419 realsize = freesize = zone->present_pages;
5422 * Adjust freesize so that it accounts for how much memory
5423 * is used by this zone for memmap. This affects the watermark
5424 * and per-cpu initialisations
5426 memmap_pages = calc_memmap_size(size, realsize);
5427 if (!is_highmem_idx(j)) {
5428 if (freesize >= memmap_pages) {
5429 freesize -= memmap_pages;
5432 " %s zone: %lu pages used for memmap\n",
5433 zone_names[j], memmap_pages);
5436 " %s zone: %lu pages exceeds freesize %lu\n",
5437 zone_names[j], memmap_pages, freesize);
5440 /* Account for reserved pages */
5441 if (j == 0 && freesize > dma_reserve) {
5442 freesize -= dma_reserve;
5443 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
5444 zone_names[0], dma_reserve);
5447 if (!is_highmem_idx(j))
5448 nr_kernel_pages += freesize;
5449 /* Charge for highmem memmap if there are enough kernel pages */
5450 else if (nr_kernel_pages > memmap_pages * 2)
5451 nr_kernel_pages -= memmap_pages;
5452 nr_all_pages += freesize;
5455 * Set an approximate value for lowmem here, it will be adjusted
5456 * when the bootmem allocator frees pages into the buddy system.
5457 * And all highmem pages will be managed by the buddy system.
5459 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
5462 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
5464 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
5466 zone->name = zone_names[j];
5467 spin_lock_init(&zone->lock);
5468 spin_lock_init(&zone->lru_lock);
5469 zone_seqlock_init(zone);
5470 zone->zone_pgdat = pgdat;
5471 zone_pcp_init(zone);
5473 /* For bootup, initialized properly in watermark setup */
5474 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
5476 lruvec_init(&zone->lruvec);
5480 set_pageblock_order();
5481 setup_usemap(pgdat, zone, zone_start_pfn, size);
5482 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
5484 memmap_init(size, nid, j, zone_start_pfn);
5488 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
5490 unsigned long __maybe_unused start = 0;
5491 unsigned long __maybe_unused offset = 0;
5493 /* Skip empty nodes */
5494 if (!pgdat->node_spanned_pages)
5497 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5498 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5499 offset = pgdat->node_start_pfn - start;
5500 /* ia64 gets its own node_mem_map, before this, without bootmem */
5501 if (!pgdat->node_mem_map) {
5502 unsigned long size, end;
5506 * The zone's endpoints aren't required to be MAX_ORDER
5507 * aligned but the node_mem_map endpoints must be in order
5508 * for the buddy allocator to function correctly.
5510 end = pgdat_end_pfn(pgdat);
5511 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5512 size = (end - start) * sizeof(struct page);
5513 map = alloc_remap(pgdat->node_id, size);
5515 map = memblock_virt_alloc_node_nopanic(size,
5517 pgdat->node_mem_map = map + offset;
5519 #ifndef CONFIG_NEED_MULTIPLE_NODES
5521 * With no DISCONTIG, the global mem_map is just set as node 0's
5523 if (pgdat == NODE_DATA(0)) {
5524 mem_map = NODE_DATA(0)->node_mem_map;
5525 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5526 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5528 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5531 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5534 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5535 unsigned long node_start_pfn, unsigned long *zholes_size)
5537 pg_data_t *pgdat = NODE_DATA(nid);
5538 unsigned long start_pfn = 0;
5539 unsigned long end_pfn = 0;
5541 /* pg_data_t should be reset to zero when it's allocated */
5542 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5544 reset_deferred_meminit(pgdat);
5545 pgdat->node_id = nid;
5546 pgdat->node_start_pfn = node_start_pfn;
5547 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5548 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5549 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5550 (u64)start_pfn << PAGE_SHIFT,
5551 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
5553 start_pfn = node_start_pfn;
5555 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5556 zones_size, zholes_size);
5558 alloc_node_mem_map(pgdat);
5559 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5560 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5561 nid, (unsigned long)pgdat,
5562 (unsigned long)pgdat->node_mem_map);
5565 free_area_init_core(pgdat);
5568 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5570 #if MAX_NUMNODES > 1
5572 * Figure out the number of possible node ids.
5574 void __init setup_nr_node_ids(void)
5576 unsigned int highest;
5578 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
5579 nr_node_ids = highest + 1;
5584 * node_map_pfn_alignment - determine the maximum internode alignment
5586 * This function should be called after node map is populated and sorted.
5587 * It calculates the maximum power of two alignment which can distinguish
5590 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5591 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5592 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5593 * shifted, 1GiB is enough and this function will indicate so.
5595 * This is used to test whether pfn -> nid mapping of the chosen memory
5596 * model has fine enough granularity to avoid incorrect mapping for the
5597 * populated node map.
5599 * Returns the determined alignment in pfn's. 0 if there is no alignment
5600 * requirement (single node).
5602 unsigned long __init node_map_pfn_alignment(void)
5604 unsigned long accl_mask = 0, last_end = 0;
5605 unsigned long start, end, mask;
5609 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5610 if (!start || last_nid < 0 || last_nid == nid) {
5617 * Start with a mask granular enough to pin-point to the
5618 * start pfn and tick off bits one-by-one until it becomes
5619 * too coarse to separate the current node from the last.
5621 mask = ~((1 << __ffs(start)) - 1);
5622 while (mask && last_end <= (start & (mask << 1)))
5625 /* accumulate all internode masks */
5629 /* convert mask to number of pages */
5630 return ~accl_mask + 1;
5633 /* Find the lowest pfn for a node */
5634 static unsigned long __init find_min_pfn_for_node(int nid)
5636 unsigned long min_pfn = ULONG_MAX;
5637 unsigned long start_pfn;
5640 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5641 min_pfn = min(min_pfn, start_pfn);
5643 if (min_pfn == ULONG_MAX) {
5645 "Could not find start_pfn for node %d\n", nid);
5653 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5655 * It returns the minimum PFN based on information provided via
5656 * memblock_set_node().
5658 unsigned long __init find_min_pfn_with_active_regions(void)
5660 return find_min_pfn_for_node(MAX_NUMNODES);
5664 * early_calculate_totalpages()
5665 * Sum pages in active regions for movable zone.
5666 * Populate N_MEMORY for calculating usable_nodes.
5668 static unsigned long __init early_calculate_totalpages(void)
5670 unsigned long totalpages = 0;
5671 unsigned long start_pfn, end_pfn;
5674 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5675 unsigned long pages = end_pfn - start_pfn;
5677 totalpages += pages;
5679 node_set_state(nid, N_MEMORY);
5685 * Find the PFN the Movable zone begins in each node. Kernel memory
5686 * is spread evenly between nodes as long as the nodes have enough
5687 * memory. When they don't, some nodes will have more kernelcore than
5690 static void __init find_zone_movable_pfns_for_nodes(void)
5693 unsigned long usable_startpfn;
5694 unsigned long kernelcore_node, kernelcore_remaining;
5695 /* save the state before borrow the nodemask */
5696 nodemask_t saved_node_state = node_states[N_MEMORY];
5697 unsigned long totalpages = early_calculate_totalpages();
5698 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5699 struct memblock_region *r;
5701 /* Need to find movable_zone earlier when movable_node is specified. */
5702 find_usable_zone_for_movable();
5705 * If movable_node is specified, ignore kernelcore and movablecore
5708 if (movable_node_is_enabled()) {
5709 for_each_memblock(memory, r) {
5710 if (!memblock_is_hotpluggable(r))
5715 usable_startpfn = PFN_DOWN(r->base);
5716 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5717 min(usable_startpfn, zone_movable_pfn[nid]) :
5725 * If kernelcore=mirror is specified, ignore movablecore option
5727 if (mirrored_kernelcore) {
5728 bool mem_below_4gb_not_mirrored = false;
5730 for_each_memblock(memory, r) {
5731 if (memblock_is_mirror(r))
5736 usable_startpfn = memblock_region_memory_base_pfn(r);
5738 if (usable_startpfn < 0x100000) {
5739 mem_below_4gb_not_mirrored = true;
5743 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5744 min(usable_startpfn, zone_movable_pfn[nid]) :
5748 if (mem_below_4gb_not_mirrored)
5749 pr_warn("This configuration results in unmirrored kernel memory.");
5755 * If movablecore=nn[KMG] was specified, calculate what size of
5756 * kernelcore that corresponds so that memory usable for
5757 * any allocation type is evenly spread. If both kernelcore
5758 * and movablecore are specified, then the value of kernelcore
5759 * will be used for required_kernelcore if it's greater than
5760 * what movablecore would have allowed.
5762 if (required_movablecore) {
5763 unsigned long corepages;
5766 * Round-up so that ZONE_MOVABLE is at least as large as what
5767 * was requested by the user
5769 required_movablecore =
5770 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5771 required_movablecore = min(totalpages, required_movablecore);
5772 corepages = totalpages - required_movablecore;
5774 required_kernelcore = max(required_kernelcore, corepages);
5778 * If kernelcore was not specified or kernelcore size is larger
5779 * than totalpages, there is no ZONE_MOVABLE.
5781 if (!required_kernelcore || required_kernelcore >= totalpages)
5784 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5785 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5788 /* Spread kernelcore memory as evenly as possible throughout nodes */
5789 kernelcore_node = required_kernelcore / usable_nodes;
5790 for_each_node_state(nid, N_MEMORY) {
5791 unsigned long start_pfn, end_pfn;
5794 * Recalculate kernelcore_node if the division per node
5795 * now exceeds what is necessary to satisfy the requested
5796 * amount of memory for the kernel
5798 if (required_kernelcore < kernelcore_node)
5799 kernelcore_node = required_kernelcore / usable_nodes;
5802 * As the map is walked, we track how much memory is usable
5803 * by the kernel using kernelcore_remaining. When it is
5804 * 0, the rest of the node is usable by ZONE_MOVABLE
5806 kernelcore_remaining = kernelcore_node;
5808 /* Go through each range of PFNs within this node */
5809 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5810 unsigned long size_pages;
5812 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5813 if (start_pfn >= end_pfn)
5816 /* Account for what is only usable for kernelcore */
5817 if (start_pfn < usable_startpfn) {
5818 unsigned long kernel_pages;
5819 kernel_pages = min(end_pfn, usable_startpfn)
5822 kernelcore_remaining -= min(kernel_pages,
5823 kernelcore_remaining);
5824 required_kernelcore -= min(kernel_pages,
5825 required_kernelcore);
5827 /* Continue if range is now fully accounted */
5828 if (end_pfn <= usable_startpfn) {
5831 * Push zone_movable_pfn to the end so
5832 * that if we have to rebalance
5833 * kernelcore across nodes, we will
5834 * not double account here
5836 zone_movable_pfn[nid] = end_pfn;
5839 start_pfn = usable_startpfn;
5843 * The usable PFN range for ZONE_MOVABLE is from
5844 * start_pfn->end_pfn. Calculate size_pages as the
5845 * number of pages used as kernelcore
5847 size_pages = end_pfn - start_pfn;
5848 if (size_pages > kernelcore_remaining)
5849 size_pages = kernelcore_remaining;
5850 zone_movable_pfn[nid] = start_pfn + size_pages;
5853 * Some kernelcore has been met, update counts and
5854 * break if the kernelcore for this node has been
5857 required_kernelcore -= min(required_kernelcore,
5859 kernelcore_remaining -= size_pages;
5860 if (!kernelcore_remaining)
5866 * If there is still required_kernelcore, we do another pass with one
5867 * less node in the count. This will push zone_movable_pfn[nid] further
5868 * along on the nodes that still have memory until kernelcore is
5872 if (usable_nodes && required_kernelcore > usable_nodes)
5876 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5877 for (nid = 0; nid < MAX_NUMNODES; nid++)
5878 zone_movable_pfn[nid] =
5879 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5882 /* restore the node_state */
5883 node_states[N_MEMORY] = saved_node_state;
5886 /* Any regular or high memory on that node ? */
5887 static void check_for_memory(pg_data_t *pgdat, int nid)
5889 enum zone_type zone_type;
5891 if (N_MEMORY == N_NORMAL_MEMORY)
5894 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5895 struct zone *zone = &pgdat->node_zones[zone_type];
5896 if (populated_zone(zone)) {
5897 node_set_state(nid, N_HIGH_MEMORY);
5898 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5899 zone_type <= ZONE_NORMAL)
5900 node_set_state(nid, N_NORMAL_MEMORY);
5907 * free_area_init_nodes - Initialise all pg_data_t and zone data
5908 * @max_zone_pfn: an array of max PFNs for each zone
5910 * This will call free_area_init_node() for each active node in the system.
5911 * Using the page ranges provided by memblock_set_node(), the size of each
5912 * zone in each node and their holes is calculated. If the maximum PFN
5913 * between two adjacent zones match, it is assumed that the zone is empty.
5914 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5915 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5916 * starts where the previous one ended. For example, ZONE_DMA32 starts
5917 * at arch_max_dma_pfn.
5919 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5921 unsigned long start_pfn, end_pfn;
5924 /* Record where the zone boundaries are */
5925 memset(arch_zone_lowest_possible_pfn, 0,
5926 sizeof(arch_zone_lowest_possible_pfn));
5927 memset(arch_zone_highest_possible_pfn, 0,
5928 sizeof(arch_zone_highest_possible_pfn));
5929 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5930 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5931 for (i = 1; i < MAX_NR_ZONES; i++) {
5932 if (i == ZONE_MOVABLE)
5934 arch_zone_lowest_possible_pfn[i] =
5935 arch_zone_highest_possible_pfn[i-1];
5936 arch_zone_highest_possible_pfn[i] =
5937 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5939 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5940 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5942 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5943 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5944 find_zone_movable_pfns_for_nodes();
5946 /* Print out the zone ranges */
5947 pr_info("Zone ranges:\n");
5948 for (i = 0; i < MAX_NR_ZONES; i++) {
5949 if (i == ZONE_MOVABLE)
5951 pr_info(" %-8s ", zone_names[i]);
5952 if (arch_zone_lowest_possible_pfn[i] ==
5953 arch_zone_highest_possible_pfn[i])
5956 pr_cont("[mem %#018Lx-%#018Lx]\n",
5957 (u64)arch_zone_lowest_possible_pfn[i]
5959 ((u64)arch_zone_highest_possible_pfn[i]
5960 << PAGE_SHIFT) - 1);
5963 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5964 pr_info("Movable zone start for each node\n");
5965 for (i = 0; i < MAX_NUMNODES; i++) {
5966 if (zone_movable_pfn[i])
5967 pr_info(" Node %d: %#018Lx\n", i,
5968 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5971 /* Print out the early node map */
5972 pr_info("Early memory node ranges\n");
5973 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5974 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5975 (u64)start_pfn << PAGE_SHIFT,
5976 ((u64)end_pfn << PAGE_SHIFT) - 1);
5978 /* Initialise every node */
5979 mminit_verify_pageflags_layout();
5980 setup_nr_node_ids();
5981 for_each_online_node(nid) {
5982 pg_data_t *pgdat = NODE_DATA(nid);
5983 free_area_init_node(nid, NULL,
5984 find_min_pfn_for_node(nid), NULL);
5986 /* Any memory on that node */
5987 if (pgdat->node_present_pages)
5988 node_set_state(nid, N_MEMORY);
5989 check_for_memory(pgdat, nid);
5993 static int __init cmdline_parse_core(char *p, unsigned long *core)
5995 unsigned long long coremem;
5999 coremem = memparse(p, &p);
6000 *core = coremem >> PAGE_SHIFT;
6002 /* Paranoid check that UL is enough for the coremem value */
6003 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6009 * kernelcore=size sets the amount of memory for use for allocations that
6010 * cannot be reclaimed or migrated.
6012 static int __init cmdline_parse_kernelcore(char *p)
6014 /* parse kernelcore=mirror */
6015 if (parse_option_str(p, "mirror")) {
6016 mirrored_kernelcore = true;
6020 return cmdline_parse_core(p, &required_kernelcore);
6024 * movablecore=size sets the amount of memory for use for allocations that
6025 * can be reclaimed or migrated.
6027 static int __init cmdline_parse_movablecore(char *p)
6029 return cmdline_parse_core(p, &required_movablecore);
6032 early_param("kernelcore", cmdline_parse_kernelcore);
6033 early_param("movablecore", cmdline_parse_movablecore);
6035 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6037 void adjust_managed_page_count(struct page *page, long count)
6039 spin_lock(&managed_page_count_lock);
6040 page_zone(page)->managed_pages += count;
6041 totalram_pages += count;
6042 #ifdef CONFIG_HIGHMEM
6043 if (PageHighMem(page))
6044 totalhigh_pages += count;
6046 spin_unlock(&managed_page_count_lock);
6048 EXPORT_SYMBOL(adjust_managed_page_count);
6050 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6053 unsigned long pages = 0;
6055 start = (void *)PAGE_ALIGN((unsigned long)start);
6056 end = (void *)((unsigned long)end & PAGE_MASK);
6057 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6058 if ((unsigned int)poison <= 0xFF)
6059 memset(pos, poison, PAGE_SIZE);
6060 free_reserved_page(virt_to_page(pos));
6064 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
6065 s, pages << (PAGE_SHIFT - 10), start, end);
6069 EXPORT_SYMBOL(free_reserved_area);
6071 #ifdef CONFIG_HIGHMEM
6072 void free_highmem_page(struct page *page)
6074 __free_reserved_page(page);
6076 page_zone(page)->managed_pages++;
6082 void __init mem_init_print_info(const char *str)
6084 unsigned long physpages, codesize, datasize, rosize, bss_size;
6085 unsigned long init_code_size, init_data_size;
6087 physpages = get_num_physpages();
6088 codesize = _etext - _stext;
6089 datasize = _edata - _sdata;
6090 rosize = __end_rodata - __start_rodata;
6091 bss_size = __bss_stop - __bss_start;
6092 init_data_size = __init_end - __init_begin;
6093 init_code_size = _einittext - _sinittext;
6096 * Detect special cases and adjust section sizes accordingly:
6097 * 1) .init.* may be embedded into .data sections
6098 * 2) .init.text.* may be out of [__init_begin, __init_end],
6099 * please refer to arch/tile/kernel/vmlinux.lds.S.
6100 * 3) .rodata.* may be embedded into .text or .data sections.
6102 #define adj_init_size(start, end, size, pos, adj) \
6104 if (start <= pos && pos < end && size > adj) \
6108 adj_init_size(__init_begin, __init_end, init_data_size,
6109 _sinittext, init_code_size);
6110 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6111 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6112 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6113 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6115 #undef adj_init_size
6117 pr_info("Memory: %luK/%luK available "
6118 "(%luK kernel code, %luK rwdata, %luK rodata, "
6119 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
6120 #ifdef CONFIG_HIGHMEM
6124 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
6125 codesize >> 10, datasize >> 10, rosize >> 10,
6126 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6127 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
6128 totalcma_pages << (PAGE_SHIFT-10),
6129 #ifdef CONFIG_HIGHMEM
6130 totalhigh_pages << (PAGE_SHIFT-10),
6132 str ? ", " : "", str ? str : "");
6136 * set_dma_reserve - set the specified number of pages reserved in the first zone
6137 * @new_dma_reserve: The number of pages to mark reserved
6139 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6140 * In the DMA zone, a significant percentage may be consumed by kernel image
6141 * and other unfreeable allocations which can skew the watermarks badly. This
6142 * function may optionally be used to account for unfreeable pages in the
6143 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6144 * smaller per-cpu batchsize.
6146 void __init set_dma_reserve(unsigned long new_dma_reserve)
6148 dma_reserve = new_dma_reserve;
6151 void __init free_area_init(unsigned long *zones_size)
6153 free_area_init_node(0, zones_size,
6154 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6157 static int page_alloc_cpu_notify(struct notifier_block *self,
6158 unsigned long action, void *hcpu)
6160 int cpu = (unsigned long)hcpu;
6162 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
6163 lru_add_drain_cpu(cpu);
6167 * Spill the event counters of the dead processor
6168 * into the current processors event counters.
6169 * This artificially elevates the count of the current
6172 vm_events_fold_cpu(cpu);
6175 * Zero the differential counters of the dead processor
6176 * so that the vm statistics are consistent.
6178 * This is only okay since the processor is dead and cannot
6179 * race with what we are doing.
6181 cpu_vm_stats_fold(cpu);
6186 void __init page_alloc_init(void)
6188 hotcpu_notifier(page_alloc_cpu_notify, 0);
6192 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6193 * or min_free_kbytes changes.
6195 static void calculate_totalreserve_pages(void)
6197 struct pglist_data *pgdat;
6198 unsigned long reserve_pages = 0;
6199 enum zone_type i, j;
6201 for_each_online_pgdat(pgdat) {
6202 for (i = 0; i < MAX_NR_ZONES; i++) {
6203 struct zone *zone = pgdat->node_zones + i;
6206 /* Find valid and maximum lowmem_reserve in the zone */
6207 for (j = i; j < MAX_NR_ZONES; j++) {
6208 if (zone->lowmem_reserve[j] > max)
6209 max = zone->lowmem_reserve[j];
6212 /* we treat the high watermark as reserved pages. */
6213 max += high_wmark_pages(zone);
6215 if (max > zone->managed_pages)
6216 max = zone->managed_pages;
6218 zone->totalreserve_pages = max;
6220 reserve_pages += max;
6223 totalreserve_pages = reserve_pages;
6227 * setup_per_zone_lowmem_reserve - called whenever
6228 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6229 * has a correct pages reserved value, so an adequate number of
6230 * pages are left in the zone after a successful __alloc_pages().
6232 static void setup_per_zone_lowmem_reserve(void)
6234 struct pglist_data *pgdat;
6235 enum zone_type j, idx;
6237 for_each_online_pgdat(pgdat) {
6238 for (j = 0; j < MAX_NR_ZONES; j++) {
6239 struct zone *zone = pgdat->node_zones + j;
6240 unsigned long managed_pages = zone->managed_pages;
6242 zone->lowmem_reserve[j] = 0;
6246 struct zone *lower_zone;
6250 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6251 sysctl_lowmem_reserve_ratio[idx] = 1;
6253 lower_zone = pgdat->node_zones + idx;
6254 lower_zone->lowmem_reserve[j] = managed_pages /
6255 sysctl_lowmem_reserve_ratio[idx];
6256 managed_pages += lower_zone->managed_pages;
6261 /* update totalreserve_pages */
6262 calculate_totalreserve_pages();
6265 static void __setup_per_zone_wmarks(void)
6267 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6268 unsigned long lowmem_pages = 0;
6270 unsigned long flags;
6272 /* Calculate total number of !ZONE_HIGHMEM pages */
6273 for_each_zone(zone) {
6274 if (!is_highmem(zone))
6275 lowmem_pages += zone->managed_pages;
6278 for_each_zone(zone) {
6281 spin_lock_irqsave(&zone->lock, flags);
6282 tmp = (u64)pages_min * zone->managed_pages;
6283 do_div(tmp, lowmem_pages);
6284 if (is_highmem(zone)) {
6286 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6287 * need highmem pages, so cap pages_min to a small
6290 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6291 * deltas control asynch page reclaim, and so should
6292 * not be capped for highmem.
6294 unsigned long min_pages;
6296 min_pages = zone->managed_pages / 1024;
6297 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6298 zone->watermark[WMARK_MIN] = min_pages;
6301 * If it's a lowmem zone, reserve a number of pages
6302 * proportionate to the zone's size.
6304 zone->watermark[WMARK_MIN] = tmp;
6307 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
6308 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
6310 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
6311 high_wmark_pages(zone) - low_wmark_pages(zone) -
6312 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
6314 spin_unlock_irqrestore(&zone->lock, flags);
6317 /* update totalreserve_pages */
6318 calculate_totalreserve_pages();
6322 * setup_per_zone_wmarks - called when min_free_kbytes changes
6323 * or when memory is hot-{added|removed}
6325 * Ensures that the watermark[min,low,high] values for each zone are set
6326 * correctly with respect to min_free_kbytes.
6328 void setup_per_zone_wmarks(void)
6330 mutex_lock(&zonelists_mutex);
6331 __setup_per_zone_wmarks();
6332 mutex_unlock(&zonelists_mutex);
6336 * The inactive anon list should be small enough that the VM never has to
6337 * do too much work, but large enough that each inactive page has a chance
6338 * to be referenced again before it is swapped out.
6340 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6341 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6342 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6343 * the anonymous pages are kept on the inactive list.
6346 * memory ratio inactive anon
6347 * -------------------------------------
6356 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
6358 unsigned int gb, ratio;
6360 /* Zone size in gigabytes */
6361 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
6363 ratio = int_sqrt(10 * gb);
6367 zone->inactive_ratio = ratio;
6370 static void __meminit setup_per_zone_inactive_ratio(void)
6375 calculate_zone_inactive_ratio(zone);
6379 * Initialise min_free_kbytes.
6381 * For small machines we want it small (128k min). For large machines
6382 * we want it large (64MB max). But it is not linear, because network
6383 * bandwidth does not increase linearly with machine size. We use
6385 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6386 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6402 int __meminit init_per_zone_wmark_min(void)
6404 unsigned long lowmem_kbytes;
6405 int new_min_free_kbytes;
6407 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6408 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6410 if (new_min_free_kbytes > user_min_free_kbytes) {
6411 min_free_kbytes = new_min_free_kbytes;
6412 if (min_free_kbytes < 128)
6413 min_free_kbytes = 128;
6414 if (min_free_kbytes > 65536)
6415 min_free_kbytes = 65536;
6417 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6418 new_min_free_kbytes, user_min_free_kbytes);
6420 setup_per_zone_wmarks();
6421 refresh_zone_stat_thresholds();
6422 setup_per_zone_lowmem_reserve();
6423 setup_per_zone_inactive_ratio();
6426 module_init(init_per_zone_wmark_min)
6429 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6430 * that we can call two helper functions whenever min_free_kbytes
6433 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6434 void __user *buffer, size_t *length, loff_t *ppos)
6438 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6443 user_min_free_kbytes = min_free_kbytes;
6444 setup_per_zone_wmarks();
6450 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
6451 void __user *buffer, size_t *length, loff_t *ppos)
6456 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6461 zone->min_unmapped_pages = (zone->managed_pages *
6462 sysctl_min_unmapped_ratio) / 100;
6466 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
6467 void __user *buffer, size_t *length, loff_t *ppos)
6472 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6477 zone->min_slab_pages = (zone->managed_pages *
6478 sysctl_min_slab_ratio) / 100;
6484 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6485 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6486 * whenever sysctl_lowmem_reserve_ratio changes.
6488 * The reserve ratio obviously has absolutely no relation with the
6489 * minimum watermarks. The lowmem reserve ratio can only make sense
6490 * if in function of the boot time zone sizes.
6492 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
6493 void __user *buffer, size_t *length, loff_t *ppos)
6495 proc_dointvec_minmax(table, write, buffer, length, ppos);
6496 setup_per_zone_lowmem_reserve();
6501 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6502 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6503 * pagelist can have before it gets flushed back to buddy allocator.
6505 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
6506 void __user *buffer, size_t *length, loff_t *ppos)
6509 int old_percpu_pagelist_fraction;
6512 mutex_lock(&pcp_batch_high_lock);
6513 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
6515 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6516 if (!write || ret < 0)
6519 /* Sanity checking to avoid pcp imbalance */
6520 if (percpu_pagelist_fraction &&
6521 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
6522 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6528 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6531 for_each_populated_zone(zone) {
6534 for_each_possible_cpu(cpu)
6535 pageset_set_high_and_batch(zone,
6536 per_cpu_ptr(zone->pageset, cpu));
6539 mutex_unlock(&pcp_batch_high_lock);
6544 int hashdist = HASHDIST_DEFAULT;
6546 static int __init set_hashdist(char *str)
6550 hashdist = simple_strtoul(str, &str, 0);
6553 __setup("hashdist=", set_hashdist);
6557 * allocate a large system hash table from bootmem
6558 * - it is assumed that the hash table must contain an exact power-of-2
6559 * quantity of entries
6560 * - limit is the number of hash buckets, not the total allocation size
6562 void *__init alloc_large_system_hash(const char *tablename,
6563 unsigned long bucketsize,
6564 unsigned long numentries,
6567 unsigned int *_hash_shift,
6568 unsigned int *_hash_mask,
6569 unsigned long low_limit,
6570 unsigned long high_limit)
6572 unsigned long long max = high_limit;
6573 unsigned long log2qty, size;
6576 /* allow the kernel cmdline to have a say */
6578 /* round applicable memory size up to nearest megabyte */
6579 numentries = nr_kernel_pages;
6581 /* It isn't necessary when PAGE_SIZE >= 1MB */
6582 if (PAGE_SHIFT < 20)
6583 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6585 /* limit to 1 bucket per 2^scale bytes of low memory */
6586 if (scale > PAGE_SHIFT)
6587 numentries >>= (scale - PAGE_SHIFT);
6589 numentries <<= (PAGE_SHIFT - scale);
6591 /* Make sure we've got at least a 0-order allocation.. */
6592 if (unlikely(flags & HASH_SMALL)) {
6593 /* Makes no sense without HASH_EARLY */
6594 WARN_ON(!(flags & HASH_EARLY));
6595 if (!(numentries >> *_hash_shift)) {
6596 numentries = 1UL << *_hash_shift;
6597 BUG_ON(!numentries);
6599 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6600 numentries = PAGE_SIZE / bucketsize;
6602 numentries = roundup_pow_of_two(numentries);
6604 /* limit allocation size to 1/16 total memory by default */
6606 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6607 do_div(max, bucketsize);
6609 max = min(max, 0x80000000ULL);
6611 if (numentries < low_limit)
6612 numentries = low_limit;
6613 if (numentries > max)
6616 log2qty = ilog2(numentries);
6619 size = bucketsize << log2qty;
6620 if (flags & HASH_EARLY)
6621 table = memblock_virt_alloc_nopanic(size, 0);
6623 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6626 * If bucketsize is not a power-of-two, we may free
6627 * some pages at the end of hash table which
6628 * alloc_pages_exact() automatically does
6630 if (get_order(size) < MAX_ORDER) {
6631 table = alloc_pages_exact(size, GFP_ATOMIC);
6632 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6635 } while (!table && size > PAGE_SIZE && --log2qty);
6638 panic("Failed to allocate %s hash table\n", tablename);
6640 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6643 ilog2(size) - PAGE_SHIFT,
6647 *_hash_shift = log2qty;
6649 *_hash_mask = (1 << log2qty) - 1;
6654 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6655 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6658 #ifdef CONFIG_SPARSEMEM
6659 return __pfn_to_section(pfn)->pageblock_flags;
6661 return zone->pageblock_flags;
6662 #endif /* CONFIG_SPARSEMEM */
6665 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6667 #ifdef CONFIG_SPARSEMEM
6668 pfn &= (PAGES_PER_SECTION-1);
6669 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6671 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6672 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6673 #endif /* CONFIG_SPARSEMEM */
6677 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6678 * @page: The page within the block of interest
6679 * @pfn: The target page frame number
6680 * @end_bitidx: The last bit of interest to retrieve
6681 * @mask: mask of bits that the caller is interested in
6683 * Return: pageblock_bits flags
6685 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6686 unsigned long end_bitidx,
6690 unsigned long *bitmap;
6691 unsigned long bitidx, word_bitidx;
6694 zone = page_zone(page);
6695 bitmap = get_pageblock_bitmap(zone, pfn);
6696 bitidx = pfn_to_bitidx(zone, pfn);
6697 word_bitidx = bitidx / BITS_PER_LONG;
6698 bitidx &= (BITS_PER_LONG-1);
6700 word = bitmap[word_bitidx];
6701 bitidx += end_bitidx;
6702 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6706 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6707 * @page: The page within the block of interest
6708 * @flags: The flags to set
6709 * @pfn: The target page frame number
6710 * @end_bitidx: The last bit of interest
6711 * @mask: mask of bits that the caller is interested in
6713 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6715 unsigned long end_bitidx,
6719 unsigned long *bitmap;
6720 unsigned long bitidx, word_bitidx;
6721 unsigned long old_word, word;
6723 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6725 zone = page_zone(page);
6726 bitmap = get_pageblock_bitmap(zone, pfn);
6727 bitidx = pfn_to_bitidx(zone, pfn);
6728 word_bitidx = bitidx / BITS_PER_LONG;
6729 bitidx &= (BITS_PER_LONG-1);
6731 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6733 bitidx += end_bitidx;
6734 mask <<= (BITS_PER_LONG - bitidx - 1);
6735 flags <<= (BITS_PER_LONG - bitidx - 1);
6737 word = READ_ONCE(bitmap[word_bitidx]);
6739 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6740 if (word == old_word)
6747 * This function checks whether pageblock includes unmovable pages or not.
6748 * If @count is not zero, it is okay to include less @count unmovable pages
6750 * PageLRU check without isolation or lru_lock could race so that
6751 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6752 * expect this function should be exact.
6754 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6755 bool skip_hwpoisoned_pages)
6757 unsigned long pfn, iter, found;
6761 * For avoiding noise data, lru_add_drain_all() should be called
6762 * If ZONE_MOVABLE, the zone never contains unmovable pages
6764 if (zone_idx(zone) == ZONE_MOVABLE)
6766 mt = get_pageblock_migratetype(page);
6767 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6770 pfn = page_to_pfn(page);
6771 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6772 unsigned long check = pfn + iter;
6774 if (!pfn_valid_within(check))
6777 page = pfn_to_page(check);
6780 * Hugepages are not in LRU lists, but they're movable.
6781 * We need not scan over tail pages bacause we don't
6782 * handle each tail page individually in migration.
6784 if (PageHuge(page)) {
6785 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6790 * We can't use page_count without pin a page
6791 * because another CPU can free compound page.
6792 * This check already skips compound tails of THP
6793 * because their page->_count is zero at all time.
6795 if (!atomic_read(&page->_count)) {
6796 if (PageBuddy(page))
6797 iter += (1 << page_order(page)) - 1;
6802 * The HWPoisoned page may be not in buddy system, and
6803 * page_count() is not 0.
6805 if (skip_hwpoisoned_pages && PageHWPoison(page))
6811 * If there are RECLAIMABLE pages, we need to check
6812 * it. But now, memory offline itself doesn't call
6813 * shrink_node_slabs() and it still to be fixed.
6816 * If the page is not RAM, page_count()should be 0.
6817 * we don't need more check. This is an _used_ not-movable page.
6819 * The problematic thing here is PG_reserved pages. PG_reserved
6820 * is set to both of a memory hole page and a _used_ kernel
6829 bool is_pageblock_removable_nolock(struct page *page)
6835 * We have to be careful here because we are iterating over memory
6836 * sections which are not zone aware so we might end up outside of
6837 * the zone but still within the section.
6838 * We have to take care about the node as well. If the node is offline
6839 * its NODE_DATA will be NULL - see page_zone.
6841 if (!node_online(page_to_nid(page)))
6844 zone = page_zone(page);
6845 pfn = page_to_pfn(page);
6846 if (!zone_spans_pfn(zone, pfn))
6849 return !has_unmovable_pages(zone, page, 0, true);
6852 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
6854 static unsigned long pfn_max_align_down(unsigned long pfn)
6856 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6857 pageblock_nr_pages) - 1);
6860 static unsigned long pfn_max_align_up(unsigned long pfn)
6862 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6863 pageblock_nr_pages));
6866 /* [start, end) must belong to a single zone. */
6867 static int __alloc_contig_migrate_range(struct compact_control *cc,
6868 unsigned long start, unsigned long end)
6870 /* This function is based on compact_zone() from compaction.c. */
6871 unsigned long nr_reclaimed;
6872 unsigned long pfn = start;
6873 unsigned int tries = 0;
6878 while (pfn < end || !list_empty(&cc->migratepages)) {
6879 if (fatal_signal_pending(current)) {
6884 if (list_empty(&cc->migratepages)) {
6885 cc->nr_migratepages = 0;
6886 pfn = isolate_migratepages_range(cc, pfn, end);
6892 } else if (++tries == 5) {
6893 ret = ret < 0 ? ret : -EBUSY;
6897 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6899 cc->nr_migratepages -= nr_reclaimed;
6901 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6902 NULL, 0, cc->mode, MR_CMA);
6905 putback_movable_pages(&cc->migratepages);
6912 * alloc_contig_range() -- tries to allocate given range of pages
6913 * @start: start PFN to allocate
6914 * @end: one-past-the-last PFN to allocate
6915 * @migratetype: migratetype of the underlaying pageblocks (either
6916 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6917 * in range must have the same migratetype and it must
6918 * be either of the two.
6920 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6921 * aligned, however it's the caller's responsibility to guarantee that
6922 * we are the only thread that changes migrate type of pageblocks the
6925 * The PFN range must belong to a single zone.
6927 * Returns zero on success or negative error code. On success all
6928 * pages which PFN is in [start, end) are allocated for the caller and
6929 * need to be freed with free_contig_range().
6931 int alloc_contig_range(unsigned long start, unsigned long end,
6932 unsigned migratetype)
6934 unsigned long outer_start, outer_end;
6938 struct compact_control cc = {
6939 .nr_migratepages = 0,
6941 .zone = page_zone(pfn_to_page(start)),
6942 .mode = MIGRATE_SYNC,
6943 .ignore_skip_hint = true,
6945 INIT_LIST_HEAD(&cc.migratepages);
6948 * What we do here is we mark all pageblocks in range as
6949 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6950 * have different sizes, and due to the way page allocator
6951 * work, we align the range to biggest of the two pages so
6952 * that page allocator won't try to merge buddies from
6953 * different pageblocks and change MIGRATE_ISOLATE to some
6954 * other migration type.
6956 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6957 * migrate the pages from an unaligned range (ie. pages that
6958 * we are interested in). This will put all the pages in
6959 * range back to page allocator as MIGRATE_ISOLATE.
6961 * When this is done, we take the pages in range from page
6962 * allocator removing them from the buddy system. This way
6963 * page allocator will never consider using them.
6965 * This lets us mark the pageblocks back as
6966 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6967 * aligned range but not in the unaligned, original range are
6968 * put back to page allocator so that buddy can use them.
6971 ret = start_isolate_page_range(pfn_max_align_down(start),
6972 pfn_max_align_up(end), migratetype,
6978 * In case of -EBUSY, we'd like to know which page causes problem.
6979 * So, just fall through. We will check it in test_pages_isolated().
6981 ret = __alloc_contig_migrate_range(&cc, start, end);
6982 if (ret && ret != -EBUSY)
6986 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6987 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6988 * more, all pages in [start, end) are free in page allocator.
6989 * What we are going to do is to allocate all pages from
6990 * [start, end) (that is remove them from page allocator).
6992 * The only problem is that pages at the beginning and at the
6993 * end of interesting range may be not aligned with pages that
6994 * page allocator holds, ie. they can be part of higher order
6995 * pages. Because of this, we reserve the bigger range and
6996 * once this is done free the pages we are not interested in.
6998 * We don't have to hold zone->lock here because the pages are
6999 * isolated thus they won't get removed from buddy.
7002 lru_add_drain_all();
7003 drain_all_pages(cc.zone);
7006 outer_start = start;
7007 while (!PageBuddy(pfn_to_page(outer_start))) {
7008 if (++order >= MAX_ORDER) {
7009 outer_start = start;
7012 outer_start &= ~0UL << order;
7015 if (outer_start != start) {
7016 order = page_order(pfn_to_page(outer_start));
7019 * outer_start page could be small order buddy page and
7020 * it doesn't include start page. Adjust outer_start
7021 * in this case to report failed page properly
7022 * on tracepoint in test_pages_isolated()
7024 if (outer_start + (1UL << order) <= start)
7025 outer_start = start;
7028 /* Make sure the range is really isolated. */
7029 if (test_pages_isolated(outer_start, end, false)) {
7030 pr_info("%s: [%lx, %lx) PFNs busy\n",
7031 __func__, outer_start, end);
7036 /* Grab isolated pages from freelists. */
7037 outer_end = isolate_freepages_range(&cc, outer_start, end);
7043 /* Free head and tail (if any) */
7044 if (start != outer_start)
7045 free_contig_range(outer_start, start - outer_start);
7046 if (end != outer_end)
7047 free_contig_range(end, outer_end - end);
7050 undo_isolate_page_range(pfn_max_align_down(start),
7051 pfn_max_align_up(end), migratetype);
7055 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7057 unsigned int count = 0;
7059 for (; nr_pages--; pfn++) {
7060 struct page *page = pfn_to_page(pfn);
7062 count += page_count(page) != 1;
7065 WARN(count != 0, "%d pages are still in use!\n", count);
7069 #ifdef CONFIG_MEMORY_HOTPLUG
7071 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7072 * page high values need to be recalulated.
7074 void __meminit zone_pcp_update(struct zone *zone)
7077 mutex_lock(&pcp_batch_high_lock);
7078 for_each_possible_cpu(cpu)
7079 pageset_set_high_and_batch(zone,
7080 per_cpu_ptr(zone->pageset, cpu));
7081 mutex_unlock(&pcp_batch_high_lock);
7085 void zone_pcp_reset(struct zone *zone)
7087 unsigned long flags;
7089 struct per_cpu_pageset *pset;
7091 /* avoid races with drain_pages() */
7092 local_irq_save(flags);
7093 if (zone->pageset != &boot_pageset) {
7094 for_each_online_cpu(cpu) {
7095 pset = per_cpu_ptr(zone->pageset, cpu);
7096 drain_zonestat(zone, pset);
7098 free_percpu(zone->pageset);
7099 zone->pageset = &boot_pageset;
7101 local_irq_restore(flags);
7104 #ifdef CONFIG_MEMORY_HOTREMOVE
7106 * All pages in the range must be isolated before calling this.
7109 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7113 unsigned int order, i;
7115 unsigned long flags;
7116 /* find the first valid pfn */
7117 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7122 zone = page_zone(pfn_to_page(pfn));
7123 spin_lock_irqsave(&zone->lock, flags);
7125 while (pfn < end_pfn) {
7126 if (!pfn_valid(pfn)) {
7130 page = pfn_to_page(pfn);
7132 * The HWPoisoned page may be not in buddy system, and
7133 * page_count() is not 0.
7135 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7137 SetPageReserved(page);
7141 BUG_ON(page_count(page));
7142 BUG_ON(!PageBuddy(page));
7143 order = page_order(page);
7144 #ifdef CONFIG_DEBUG_VM
7145 printk(KERN_INFO "remove from free list %lx %d %lx\n",
7146 pfn, 1 << order, end_pfn);
7148 list_del(&page->lru);
7149 rmv_page_order(page);
7150 zone->free_area[order].nr_free--;
7151 for (i = 0; i < (1 << order); i++)
7152 SetPageReserved((page+i));
7153 pfn += (1 << order);
7155 spin_unlock_irqrestore(&zone->lock, flags);
7159 bool is_free_buddy_page(struct page *page)
7161 struct zone *zone = page_zone(page);
7162 unsigned long pfn = page_to_pfn(page);
7163 unsigned long flags;
7166 spin_lock_irqsave(&zone->lock, flags);
7167 for (order = 0; order < MAX_ORDER; order++) {
7168 struct page *page_head = page - (pfn & ((1 << order) - 1));
7170 if (PageBuddy(page_head) && page_order(page_head) >= order)
7173 spin_unlock_irqrestore(&zone->lock, flags);
7175 return order < MAX_ORDER;